{"id": "2606.19652", "title": "Convex training of Lipschitz-regularized shallow neural networks", "abstract": "In this work, we introduce a training procedure for shallow neural networks that promotes robustness against adversarial attacks. We solve a non-convex Lipschitz-regularized training program by introducing a convex restriction that can be efficiently solved to global optimality. Our approach can be employed as a post-processing step by taking a pre-trained network as an initial solution to then solving the convex program whose optimal network is guaranteed to be no worse than the initial one. We illustrate the improvements of our training procedure with experiments using real world datasets for regression tasks under an adversarial setting. We show numerically that solving our proposed convex program yields networks with lower objective values on the Lipschitz-regularized program compared to existing methods. Additionally, we show that on certain datasets, networks obtained using our convex training program are both more accurate and robust with respect to adversarial attacks.", "date": "2026-06-17", "categories": []}
{"id": "2606.19736", "title": "VFACamou: View-Fused Adversarial Camouflage for Environment-Adaptive Physical Evasion", "abstract": "Adversarial camouflage in the physical world remains highly challenging, particularly under UAV reconnaissance where targets undergo continuous geometric changes and extreme illumination variations. Existing methods either optimize 2D digital perturbations that fail to generalize to dynamic viewpoints or produce visually unnatural textures that cannot be deployed in real scenarios. Therefore, we propose an end-to-end framework for adversarial camouflage generation that automatically produces wearable adversarial patterns and maintains stable attack performance in real physical environments with changing viewpoints, poses, and lighting conditions. Our method integrates UV-volume rendering with a diffusion-based texture generator, enabling consistent appearance under varying scales, poses, and lighting conditions. To ensure environmental realism, we propose an illumination color consistency estimator that extracts dominant background attributes and guides a natural texture loss to align the generated UV texture with the surrounding environment. A multi-scale dynamic training strategy further enhances robustness against viewpoint shifts and body deformation. Extensive experiments across multiple mainstream detectors demonstrate that our method achieves strong and stable physical attack performance while maintaining high perceptual naturalness, reducing human detection rates without introducing unnatural artifacts.", "date": "2026-06-18", "categories": []}
{"id": "2505.03646", "title": "Revealing Hidden Vulnerabilities in Autoencoders through Gradient Signal Restoration", "abstract": "Adversarial robustness of deep autoencoders (AEs) has received less attention than that of discriminative models, although their compressed latent representations induce ill-conditioned mappings that can amplify small input perturbations and destabilize reconstructions. Existing white-box attacks for AEs, which optimize norm-bounded adversarial perturbations to maximize reconstruction damage, often converge to suboptimal perturbations, thereby potentially overstating AE robustness. We show that this limitation is linked to vanishing adversarial loss gradients during backpropagation through ill-conditioned layers, associated with near-zero singular values in their intermediate weight matrices. To address this, we propose GRILL (Gradient Signal Restoration in Ill-Conditioned Layers), a framework designed to mitigate gradient degradation and improve the reliability of adversarial robustness evaluation in encoder-decoder architectures. GRILL is designed to mitigate adversarial gradient degradation during optimization, enabling attacks to better approximate high-distortion perturbations under fixed norm constraints. Through extensive experiments across multiple AE architectures, under both sample-specific and universal attacks, as well as standard and adaptive attack settings, we show that GRILL significantly increases attack effectiveness, thereby exposing vulnerabilities hidden by existing attack limitations. Beyond AEs, we provide preliminary evidence that modern multimodal encoder-decoder architectures exhibit similar vulnerabilities.", "date": "2025-05-06", "categories": []}
{"id": "2511.20002", "title": "Semantic Router: On the Feasibility of Hijacking MLLMs via a Single Adversarial Perturbation", "abstract": "Multimodal Large Language Models (MLLMs) are increasingly deployed in stateless systems, such as autonomous driving and robotics. This paper investigates a novel threat: Semantic-Aware Hijacking. We explore the feasibility of hijacking multiple stateless decisions simultaneously using a single universal perturbation. We introduce the Semantic-Aware Universal Perturbation (SAUP), which acts as a semantic router, \"actively\" perceiving input semantics and routing them to distinct, attacker-defined targets. To achieve this, we conduct theoretical and empirical analysis on the geometric properties in the latent space. Guided by these insights, we propose the Semantic-Oriented (SORT) optimization strategy and annotate a new dataset with fine-grained semantics to evaluate performance. Extensive experiments on three representative MLLMs demonstrate the fundamental feasibility of this attack, achieving a 66% attack success rate over five targets using a single frame against Qwen.", "date": "2025-11-25", "categories": []}
{"id": "2606.18318", "title": "Budget-Aware Adaptive Adversarial Patches for Black-Box Object Detection", "abstract": "Adversarial patches pose a practical threat to modern object detectors. Prior work shows vulnerability, but three gaps limit actionable insight: (i) few \\emph{score-based black-box} attacks \\emph{jointly} optimize patch \\emph{location, texture, and size} under tight query budgets; (ii) success is rarely tied to the patch's \\emph{visual footprint}; and (iii) evaluations often conflate EOT robustness with plain-view suppression. We present \\method{}, a query-efficient, budget-adaptive black-box attack that couples a lightweight \\emph{Contextual Thompson-Sampling} placer with NES-style pixel updates, growing the patch only when progress stalls. Reporting is anchored by a \\emph{strict plain-image} suppression test; EOT is audited but never used as a substitute for success, and optional appearance/printability weights expose strength--visibility trade-offs. Across YOLOv5, Faster R-CNN, and YOLOS, \\method{} achieves strong suppression on CNN-based detectors and substantial suppression on the transformer-based detector, using compact patches and exposing clear query--footprint trade-offs relative to fixed-size and heuristic baselines. A print--capture pilot further shows transfer across unseen physical objects and viewpoints.", "date": "2026-06-16", "categories": []}
{"id": "2606.18697", "title": "Stealthy World Model Manipulation via Data Poisoning", "abstract": "Model-based learning agents use learned world models to predict future states, plan actions, and adapt to new environments. However, the process of updating world models from collected experience creates a training-time attack surface: adversarially poisoned fine-tuning trajectories can manipulate the learned dynamics and thereby corrupt downstream planning. In this paper, we propose SWAAP, the first two-stage data poisoning framework for learned world models. In the first stage, SWAAP identifies a harmful target world model that induces low-return behavior under planning while remaining close to clean dynamics, using first-order bilevel optimization enabled by a transition-gradient theorem. In the second stage, SWAAP realizes this target through stealth-constrained gradient matching, modifying only a limited fraction of fine-tuning transition targets so that the induced training gradients steer the victim model toward the adversarial target, while a prediction-error regularizer encourages the poisoned targets to remain close to the world model's natural approximation error. To assess attack stealthiness, we evaluate defenses and detectability across three stages of the poisoning pipeline: pre-training detection of poisoned transitions, robust training during fine-tuning, and test-time monitoring of the resulting world model. Across diverse continuous-control tasks, SWAAP causes substantial performance degradation while keeping poisoned transitions close to clean data and evading the evaluated non-adaptive residual/CUSUM/TRIM-style defenses. These results reveal a practical vulnerability in world-model adaptation pipelines and highlight the need for robustness methods that protect both world-model training data and learned dynamics.", "date": "2026-06-17", "categories": []}
{"id": "2606.19212", "title": "Generalised Eigenvalue Geometry of Semantic Adversarial Attacks", "abstract": "Recent empirical work shows that semantically equivalent paraphrases can fool financial sentiment classifiers: although a paraphrase remains close to the original under a strong reference embedding, it may shift the target model's representation enough to change the predicted class. Existing robustness theory either assumes a single-model threat model or focuses mainly on empirical attack algorithms. We develop a continuous local model of semantic paraphrase perturbations that captures this two-model structure. We show that the worst-case local displacement of the target representation, subject to a proxy-model budget, is governed by the largest generalised eigenvalue of a matrix pencil $(A,B)$ constructed from the Jacobians of the two embedding maps. The resulting attackability index $\\lambda^*(x)$ is intrinsic to the local paraphrase geometry and the chosen embedders, yields a closed-form prediction-flip condition for affine readouts, and supports conservative population and finite-sample attackability certificates. For uniform control over classes of affine readouts, we derive a distribution-free VC bound for binary attackability indicators and a scale-sensitive margin bound based on an attackability-adjusted margin that subtracts a local geometric penalty from the standard classifier margin. We also connect the continuous theory to discrete paraphrase search, identify an asymmetry between successful and unsuccessful finite searches, and give a covering condition under which the discrete and continuous settings agree. Finally, we propose an empirical verification framework using soft-token relaxations and generated paraphrase sets to assess the local eigenvalue geometry, prediction-flip condition, and finite-search approximation on a deployed financial-text classifier.", "date": "2026-06-17", "categories": []}
{"id": "2408.12099", "title": "Query-Efficient Video Adversarial Attack with Stylized Logo on Service Computing", "abstract": "In service computing, video classification has become fundamental to many intelligent applications. While Deep Neural Networks (DNNs) have demonstrated excellent performance in recognizing video content, recent studies have shown that DNNs are highly vulnerable to adversarial examples. Thus, understanding adversarial attacks can better respond to emergency situations. In order to improve attack performance, many style-transfer-based attacks and patch-based attacks have been proposed. However, the global perturbation of the former will bring unnatural global colors, while the latter is difficult to achieve success in targeted attacks due to the limited perturbation space. Moreover, compared to a plethora of methods targeting image classifiers, video adversarial attacks remain relatively underexplored. Therefore, to generate adversarial examples with a low budget and to provide them with a higher verisimilitude, we propose a novel black-box video attack framework, called Stylized Logo Attack (SLA). SLA is conducted through three stages. The first stage involves building a style reference set for logos, which can not only make the generated examples more natural, but also carry more target class features in targeted attacks. Then, Reinforcement Learning is employed to determine the style reference and position parameters of the logo within the video, which ensures that the stylized logo is placed in the video with optimal attributes. Finally, perturbations are optimized in a step-by-step manner so as to improve the fooling rate. Experimental results indicate that SLA can achieve better performance than state-of-the-art methods and still maintain good deception effects when facing various defense methods. We believe SLA can raise awareness among the security community about the reliability and security of video classification systems and serve as a memorandum of possible attack methods.", "date": "2024-08-22", "categories": []}
{"id": "2510.11709", "title": "Adversarial Attacks Leverage Interference Between Features in Superposition", "abstract": "Why do adversarial examples exist, and why do they transfer between models? Existing explanations appeal to high-dimensional geometry, non-robust patterns in the input, and decision boundary structure, but none provides a representation-level mechanism that explains why specific perturbations succeed and why attacks transfer between models. In this paper, we show that adversarial vulnerability can stem from efficient information encoding in neural networks. Specifically, vulnerability can arise from superposition - the phenomenon where networks represent more concepts than they have dimensions, forcing non-orthogonal representation and thus interference. This interference causes perturbations targeting one representation to affect others, creating vulnerabilities determined by interference patterns. In synthetic settings with precisely controlled superposition, we establish that superposition suffices to create adversarial vulnerability. The resulting attacks are predictable: PGD-discovered perturbations align with theoretically optimal perturbations derived from the interference geometry. Models trained on similar data develop similar interference patterns, explaining attack transferability. We then show that successful attacks on image classifiers exhibit the structure predicted by our proposed mechanism. These findings reveal that adversarial vulnerability can be a byproduct of networks' representational compression, complementing existing explanations based on data properties or architectural factors.", "date": "2025-10-13", "categories": []}
{"id": "2511.01352", "title": "MiniFool -- Physics-Constraint-Aware Minimizer-Based Adversarial Attacks in Deep Neural Networks", "abstract": "In this paper, we present a new algorithm, MiniFool, that implements physics-inspired adversarial attacks for testing neural network-based classification tasks in particle and astroparticle physics. While we initially developed the algorithm for the search for astrophysical tau neutrinos with the IceCube Neutrino Observatory, we apply it to further data from other science domains, thus demonstrating its general applicability. Here, we apply the algorithm to the well-known MNIST data set and furthermore, to Open Data data from the CMS experiment at the Large Hadron Collider. The algorithm is based on minimizing a cost function that combines a $\\chi^2$ based test-statistic with the deviation from the desired target score. The test statistic quantifies the probability of the perturbations applied to the data based on the experimental uncertainties. For our studied use cases, we find that the likelihood of a flipped classification differs for both the initially correctly and incorrectly classified events. When testing changes of the classifications as a function of an attack parameter that scales the experimental uncertainties, the robustness of the network decision can be quantified. Furthermore, this allows testing the robustness of the classification of unlabeled experimental data.", "date": "2025-11-03", "categories": []}
{"id": "2604.00605", "title": "Fluently Lying: Adversarial Robustness Can Be Substrate-Dependent", "abstract": "The primary tools used to monitor and defend object detectors under adversarial attack assume that when accuracy degrades, detection count drops in tandem. This coupling was assumed, not measured. We report a counterexample observed on a single model: under standard PGD, EMS-YOLO, a spiking neural network (SNN) object detector, retains more than 70% of its detections while mAP collapses from 0.528 to 0.042. We term this count-preserving accuracy collapse Quality Corruption (QC), to distinguish it from the suppression that dominates untargeted evaluation. Across four SNN architectures and two threat models (l-infinity and l-2), QC appears only in one of the four detectors tested (EMS-YOLO). On this model, all five standard defense components fail to detect or mitigate QC, suggesting the defense ecosystem may rely on a shared assumption calibrated on a single substrate. These results provide, to our knowledge, the first evidence that adversarial failure modes can be substrate-dependent.", "date": "2026-04-01", "categories": []}
{"id": "2606.15531", "title": "Greedy Coordinate Diffusion: Effective and Semantically Coherent Adversarial Attacks via Diffusion Guidance", "abstract": "Adversarial attacks on large language models have limited practical impact despite extensive research. Optimization-based attacks such as Greedy Coordinate Gradient (GCG) (Zou et al., 2023) produce high-perplexity, incoherent suffixes that existing defenses easily detect (Bengio et al., 2024).\n  Moreover, attempting to enforce coherence constraints during optimization often prevents the attack from successfully eliciting the specific targeted response, resulting in low success rates against robust models. Conversely, attacks that maintain coherence often alter the semantic intent of queries; when the model complies with these altered queries, responses fail to address the adversary's original goal. In this work, we introduce Greedy Coordinate Diffusion (GCD), a novel framework that efficiently generates adversarial attacks against safety-aligned models while maintaining low perplexity and high semantic adherence to the adversary's original intent. GCD leverages the generative priors of discrete diffusion language models to guide the search for adversarial suffixes that achieve semantic coherence and adherence.\n  Unlike GCG, GCD does not require direct gradient access, allowing it to operate in a gray-box setting. We show GCD achieves highest ASR while remaining competitive on response-quality scores, and that the constructed adversarial prompts are detected at lower rates than other methods by perplexity-based and guard-model filters.", "date": "2026-06-14", "categories": []}
{"id": "2606.17435", "title": "MorphStrata: Layer-Specific Perturbations for Generating Morphence Students in Time-Series Moving Target Defense", "abstract": "Time-series forecasting models remain vulnerable to gradient-based adversarial attacks while existing defense mechanisms typically incur a trade-off in robustness for bounded response and compute cost. The problem is pronounced in Moving Target Defense where maintaining multiple randomized model instances substantially exacerbates the training overhead. In this work, we introduce MorphStrata, a student generation strategy with selective, layer-specific stochastic noise injection that extends the traditional Morphence defense. MorphStrata uses a Transformer backbone as the teacher and perturbs randomly selected architectural blocks to create structured heterogeneity across student models in response to varied data distributions and threat models. We evaluate against vanilla Transformer and Morphence backbones on a suite of benchmarks including the Jena Climate, Electricity Load Diagrams, and Appliances Energy Prediction using FGSM, BIM and PGD attacks across multiple attack strengths. Across datasets and attack regimes, the proposed ensemble maintains comparable adversarial RMSE. Specifically, for high entropy, periodic datasets as in the case of the AEP data, MorphStrata achieves the lowest RMSE across all attacks and perturbation budgets, improving over the static baseline by up to 24.11% and 97.97% under FGSM and BIM respectively at an epsilon value of 0.5 over 30 randomized trials. Targeting the layers to generate MorphStrata students accounts for less than 1% increase in train-times over the Morphence MTD baseline for most of the experiments, while accounting for double digit gains in adversarial RMSE reduction. We also observe a positive correlation between higher pairwise L2 distance (among generated students) and overall defense effectiveness. In summary, MorphStrata maintains adversarial robustness as an MTD defense at marginal cost deltas when compared to existing baselines.", "date": "2026-06-16", "categories": []}
{"id": "2606.17540", "title": "TaFD: Threat-Aware Frequency Decoupling for Adversarial Robustness against Heterogeneous Attacks", "abstract": "Multi-threat robustness remains a fundamental challenge in deep learning. Although joint adversarial training (JAT) is widely adopted, it suffers from negative transfer under heterogeneous threats, particularly between $\\ell_p$-bounded and semantic attacks. Through first-order gradient analysis, we formalize this as gradient incompatibility and theoretically establish the necessity of decoupled optimization. We further reveal that these conflicting threats exhibit separable spectral characteristics in the frequency domain. Motivated by this observation, we propose Threat-aware Frequency Decoupling (TaFD), a two-stage defense framework that reformulates JAT as a frequency-domain divide-and-conquer paradigm. TaFD first discovers latent threat domains via unsupervised clustering of attack spectral prototypes and trains a lightweight classifier for inference-time threat domain identification. Conditioned on the prediction, TaFD employs a Frequency-Conditional Convolution that learns threat-domain-specific spectral masks and routes each sample to the corresponding expert, enforcing structural parameter separation and alleviating optimization conflicts.\n  We validate TaFD on three representative image-classification benchmarks (CIFAR-10, CIFAR-100, and Tiny-ImageNet) and on two representative architectures (the convolutional ResNet and the hybrid-transformer MobileViT). Extensive results demonstrate that TaFD achieves more balanced robustness against heterogeneous attacks than existing JAT and frequency-domain baselines, improving average robust accuracy by approximately 11\\% over the strongest baseline while maintaining leading clean accuracy.", "date": "2026-06-16", "categories": []}
{"id": "2606.17711", "title": "Structured Adversarial Camouflage via Voronoi Diagrams", "abstract": "Pixel-wise adversarial patches are computationally heavy and often visually detectable, limiting utility in security-critical systems. We present adversarial Voronoi camouflage that optimizes only seed-point locations under fixed, printable palettes using a soft assignment, producing structured, splinter camouflage-like patterns without additional regularization. Evaluated on person detection with COCO-style AP@[.5:.95], naive placement (Inria -> COCO) performs comparably bad, while garment-level application via segmentation mask (3DPeople) results in a significant AP drop. The attack transfers to out-of-domain backgrounds and across detector families (YOLOv9/10/11/12), indicating robustness in black-box settings. Repainting with different palettes largely nullifies the effect, and single-color tweaks show limited tolerance (<=0.17), highlighting a structure-palette coupling. The parameter-efficient, palette-constrained design improves visual plausibility while degrading real-time detector performance. Physical validation and color calibration are left for future work.\n  Code: https://github.com/JensBayer/Voronoi\n  This paper was originally presented at the International Conference on Military Communication and Information Systems (ICMCIS), organized by the Information Systems Technology (IST) Scientific and Technical Committee, IST-224-RSY - the ICMCIS, held in Bath, United Kingdom, 12-13 May 2026.", "date": "2026-06-16", "categories": []}
{"id": "2606.17815", "title": "Beyond Native Success: Auditing Deployment-Interface Exposure of CLIP Backdoors", "abstract": "Contrastive Language-Image Pre-training models are widely reused across downstream interfaces, including feature extraction, retrieval, reranking, and selection. Existing CLIP backdoor, however, usually validate attacks on a small attack-native task, leaving unclear whether the same poisoned checkpoint remains exposed, weakens, or becomes not applicable when reused through other interfaces. We introduce DIFE, a Deployment-Interface Footprint Evaluation framework that audits backdoored CLIP checkpoints across deployment interfaces. DIFE makes various evaluations comparable by specifying each interface's component readout, trigger channel, target event, reference condition, and metric. DIFE also introduces effective-footprint diagnosis to identify the reusable CLIP component or component combination that carries exposure and explains where risk transfers. Auditing reproduced CLIP backdoors with DIFE reveals a structured landscape: native success is not a checkpoint-level risk certificate, exposure follows component footprints, text-side poisoning does not yield textual-encoder control, and some coupled attacks remain mechanism-bound. This audit reveals a import gapin existing CLIP backdoors: a textual encoder that itself becomes a reusable carrier of adversarial behavior. We therefore introduce BadTextTower to fill this gap. BadTextTower produces strong text-conditioned retrieval, reranking, and selection exposure while leaving visual-only reuse nearly clean.", "date": "2026-06-16", "categories": []}
{"id": "2606.18193", "title": "A Red-Team Study of Anthropic Fable 5 & Opus 4.8 Models", "abstract": "We evaluate the adversarial robustness of two frontier large language models (LLMs) developed by Anthropic, Fable 5 and Opus 4.8, against four families of automated jailbreak attack across 7 826 harmful intents spanning a ten-category harm taxonomy. Using the HackAgent red-teaming framework, hundreds of thousands of adversarial attempts were generated and every apparent success was independently re-adjudicated by a panel of three judge models (majority vote). Both models resist the majority of attacks, but the residual surface is larger than aggregate framing suggests: it is dominated by adaptive iterative attacks, while static obfuscation is near-fully neutralised. The strongest adaptive search (tree-of-attacks) breaks Opus 4.8 on 11.5% of intents overall, whereas Fable 5 stays in the single digits (6.1% worst-case). Aggregate rates therefore should not be read as reassurance. Even in these hardened configurations, the two models produced 1 620 (Opus 4.8) and 702 (Fable 5) panel-confirmed harmful completions spanning every harm category, located automatically, cheaply, and within the first one or two refinement steps by an attacker model with no human expert in the loop. The reasonable conclusion is that even the best, most-tested frontier models remain reliably breakable under sustained automated pressure.", "date": "2026-06-16", "categories": []}
{"id": "2502.05214", "title": "CoRPA: Adversarial Image Generation for Chest X-rays Using Concept Vector Perturbations and Generative Models", "abstract": "Deep learning models for medical image classification tasks are becoming widely implemented in AI-assisted diagnostic tools, aiming to enhance diagnostic accuracy, reduce clinician workloads, and improve patient outcomes. However, their vulnerability to adversarial attacks poses significant risks to patient safety. Current attack methodologies use general techniques such as model querying or pixel value perturbations to generate adversarial examples designed to fool a model. These approaches may not adequately address the unique characteristics of clinical errors stemming from missed or incorrectly identified clinical features. We propose the Concept-based Report Perturbation Attack (CoRPA), a clinically-focused black-box adversarial attack framework tailored to the medical imaging domain. CoRPA leverages clinical concepts to generate adversarial radiological reports and images that closely mirror realistic clinical misdiagnosis scenarios. We demonstrate the utility of CoRPA using the MIMIC-CXR-JPG dataset of chest X-rays and radiological reports. Our evaluation reveals that deep learning models exhibiting strong resilience to conventional adversarial attacks are significantly less robust when subjected to CoRPA's clinically-focused perturbations. This underscores the importance of addressing domain-specific vulnerabilities in medical AI systems. By introducing a specialized adversarial attack framework, this study provides a foundation for developing robust, real-world-ready AI models in healthcare, ensuring their safe and reliable deployment in high-stakes clinical environments.", "date": "2025-02-04", "categories": []}
{"id": "2505.14418", "title": "Hidden Ghost Hand: Unveiling Backdoor Vulnerabilities in MLLM-Powered Mobile GUI Agents", "abstract": "Graphical user interface (GUI) agents powered by multimodal large language models (MLLMs) have shown greater promise for human-interaction. However, due to the high fine-tuning cost, users often rely on open-source GUI agents or APIs offered by AI providers, which introduces a critical but underexplored supply chain threat: backdoor attacks. In this work, we first unveil that MLLM-powered GUI agents naturally expose multiple interaction-level triggers, such as historical steps, environment states, and task progress. Based on this observation, we introduce AgentGhost, an effective and stealthy framework for red-teaming backdoor attacks. Specifically, we first construct composite triggers by combining goal and interaction levels, allowing GUI agents to unintentionally activate backdoors while ensuring task utility. Then, we formulate backdoor injection as a Min-Max optimization problem that uses supervised contrastive learning to maximize the feature difference across sample classes at the representation space, improving flexibility of the backdoor. Meanwhile, it adopts supervised fine-tuning to minimize the discrepancy between backdoor and clean behavior generation, enhancing effectiveness and utility. Extensive evaluations of various agent models in two established mobile benchmarks show that AgentGhost is effective and generic, with attack accuracy that reaches 99.7\\% on three attack objectives, and shows stealthiness with only 1\\% utility degradation. Furthermore, we tailor a defense method against AgentGhost that reduces the attack accuracy to 22.1\\%. Our code is available at \\texttt{anonymous}.", "date": "2025-05-20", "categories": []}
{"id": "2506.18756", "title": "Semantic-Preserving Prompt Hijacking: A Black-Box Adversarial Attack on Auto-Prompt Optimization", "abstract": "LLMs increasingly integrate auto-suggestion optimization modules, enabling them to rewrite and display user input before generating the final response. While this design aims to enhance transparency and trust, its process of autonomously selecting a single best result from multiple candidate solutions allows attackers to hijack this optimization process by inducing subtle, imperceptible semantic shifts. To address this, we propose a semantic preservation hijacking attack method based on black-box conditions: Adaptive Greedy Local Search. This method hierarchically decomposes the input text, masks key language units, and dynamically adjusts candidate replacement words at predefined semantic checkpoints. This maximizes the deviation between the model output and the original intent while strictly maintaining semantic similarity to the original text. Experimental results on commercial and open-source LLMs demonstrate that, under the same semantic similarity constraints, this method achieves a higher attack success rate than existing attack methods in over 2400 test cases. Code is available at: https://github.com/franz-chang/DOBS", "date": "2025-05-26", "categories": []}
{"id": "2509.14959", "title": "Discrete optimal transport is a strong audio adversarial attack", "abstract": "In this paper, we investigate discrete optimal transport (DOT) as a black-box attack against modern automatic speaker verification (ASV) and anti-spoofing countermeasure (CM) systems.\n  Our attack operates as a post-processing distribution-alignment step. Frame-level WavLM embeddings of generated speech (or another person speech) are aligned to an unpaired bona fide speech pool using entropic optimal transport and a top-k barycentric projection, followed by neural vocoding. Unlike gradient-based attacks, the proposed method requires no access to model parameters, gradients, or training data.\n  Experiments on ASVspoof2019 and ASVspoof5 demonstrate that DOT attack substantially increases CM EER and substantially degrades ASV performance across multiple spoofing attacks. The attack transfers across datasets and remains effective after CM fine-tuning. Analysis using speaker similarity, Fr\\'echet Audio Distance, and visualization of embedding distributions suggests that DOT succeeds by shifting source speech toward bona fide regions of the representation space rather than by maximizing speaker similarity.\n  These results indicate that optimal-transport-based distribution alignment represents a previously underexplored attack vector for contemporary ASV and anti-spoofing systems.", "date": "2025-09-18", "categories": []}
{"id": "2509.19197", "title": "A Validation Strategy for Deep Learning Models: Evaluating and Enhancing Robustness", "abstract": "Data-driven models, especially deep learning classifiers often demonstrate great success on clean datasets. Yet, they remain vulnerable to common data distortions such as adversarial and common corruption perturbations. These perturbations can significantly degrade performance, thereby challenging the overall reliability of the models. Traditional robustness validation typically relies on perturbed test datasets to assess and improve model performance. In our framework, however, we propose a validation approach that extracts \"weak robust\" samples directly from the training dataset via local robustness analysis. These samples, being the most susceptible to perturbations, serve as an early and sensitive indicator of the model's vulnerabilities. By evaluating models on these challenging training instances, we gain a more nuanced understanding of its robustness, which informs targeted performance enhancement. We demonstrate the effectiveness of our approach on models trained with CIFAR-10, CIFAR-100, and ImageNet, highlighting how robustness validation guided by weak robust samples can drive meaningful improvements in model reliability under adversarial and common corruption scenarios.", "date": "2025-09-23", "categories": []}
{"id": "2604.17805", "title": "Ranking Abuse via Strategic Pairwise Data Perturbations", "abstract": "Pairwise ranking systems based on Maximum Likelihood Estimation (MLE), such as the Bradley-Terry model, are widely used to aggregate preferences from pairwise comparisons. However, their robustness under strategic data manipulation remains insufficiently understood.\n  In this paper, we study the vulnerability of MLE-based ranking systems to adversarial perturbations. We formulate the manipulation task as a constrained combinatorial optimization problem and propose an Adaptive Subset Selection Attack (ASSA) to efficiently identify high-impact perturbations.\n  Experimental results on both synthetic data and real-world election datasets show that MLE-based rankings exhibit a sharp phase-transition behavior: beyond a small perturbation budget, a limited number of strategic voters can significantly alter the global ranking. In particular, our method consistently outperforms random and greedy baselines under constrained budgets.\n  These findings reveal a fundamental sensitivity of MLE-based ranking mechanisms to structured perturbations and highlight the need for more robust aggregation methods in collective decision-making systems.", "date": "2026-04-20", "categories": []}
{"id": "2605.26595", "title": "Cordyceps: Covert Control Attacks on LLMs via Data Poisoning", "abstract": "Large language models (LLMs) are often fine-tuned on uncurated text datasets that adversaries can poison. Existing poisoning attacks primarily rely on fixed trigger phrases that defenses such as outlier detection, clean-data regularization, or online monitoring can neutralize. In this paper, we propose a data poisoning method that teaches an LLM an information hiding scheme reliably and stealthily through semantic associations between shared knowledge such as facts or concepts and attacker-chosen phrases. The induced hiding scheme can encode and decode arbitrary malicious instructions, thus revealing a new and subtle poisoning-induced vulnerability: covert control attacks.\n  We precisely characterize covert control attacks and evaluate them across $5$ LLMs, $3$ backdoor defenses, and $4$ prompt injection defenses. With a small poisoned fraction, covert control attacks outperform heuristic-based prompt injection attacks in average attack success rate by about $40\\%$ relative to clean fine-tuned models. They also circumvent defenses based on detection and fine-tuning, maintaining up to $93\\%$ attack success rate after backdoor defenses and up to $98\\%$ after prompt injection defenses.", "date": "2026-05-26", "categories": []}
{"id": "2606.14865", "title": "GRAPE: Guided Parameter-Space Evolution for Compact Adversarial Robustness", "abstract": "Adversarial Training (AT) improves neural network robustness, but most methods train a fixed parameter space from the start. This paper asks whether the order in which parameters become optimizable can affect the final robust solution, even when the final architecture or computation budget is controlled.\n  We propose GRAPE, Guided Parameter-Space Evolution, a training framework for compact adversarial robustness. GRAPE combines parameter-space stabilization with progressive hidden expansion: it stabilizes robust optimization in the currently exposed space, gradually releases new optimizable dimensions, and uses an adversarial spectral utilization score to guide newly released capacity toward high-pressure modules. In contrast to fixed-structure AT, GRAPE treats robust model learning as a process of progressive parameter-space exposure and evolution.\n  Under the standard $\\ell_\\infty$ threat model on CIFAR-10, with fixed-structure ResNet-18 AT as a controlled reference, GRAPE improves PGD-20 robust accuracy from 51.70% to 56.94% at a nearly matched computation budget with a FLOPs ratio of 1.009x, while reducing parameter count by about 21.4%. A sequential grow variant with the same final ResNet-18 architecture reaches 56.52% PGD-20 robust accuracy, indicating that the gain is not only due to final architecture differences but also to the parameter-space exposure path. These results suggest that guided parameter-space evolution can yield compact and robust parameter configurations under matched computation.", "date": "2026-06-12", "categories": []}
{"id": "2606.14987", "title": "Continual Backdoor Training in IoT/CPS", "abstract": "Internet of Things (IoT) and Cyber-physical systems (CPS) increasingly rely on continual learning (CL) to adapt to evolving environments, device heterogeneity, and concept drift, thereby improving overall utility. While continual adaptation is essential for long-lived IoT deployments where data patterns evolve, it also introduces new security vulnerabilities. In particular, backdoor attacks can exploit incremental updates, replay buffers, and representation reuse to implant persistent malicious behaviors that remain dormant during normal operation but activate upon specific triggers. In this paper, we present a backdoor attack in continual learning used in IoT/CPS systems. To this end, we formalize an IoT/CPS-specific threat model, analyze why continual learning amplifies backdoor persistence in IoT pipelines, and evaluate our technique under varying conditions. Our analysis highlights critical open challenges in securing lifelong learning in IoT/CPS and industrial IoT (IIoT) environments, as well as the need for heightened security controls.", "date": "2026-06-12", "categories": []}
{"id": "2606.15308", "title": "Forced Deferral: Manipulating Routing Decisions in Multimodal LLM Cascades", "abstract": "While multimodal large language models (MLLMs) have shown strong visual reasoning abilities, serving a large model for every query is computationally expensive. MLLM cascades mitigate this cost by first querying a weak but cheaper model and deferring to a strong model when the weak model's output is unconfident. However, since the weak model's confidence directly controls compute allocation, these systems expose a new attack surface: an adversary can manipulate confidence so that their queries are consistently deferred to the strong model. Motivated by this vulnerability, we introduce the Forced Deferral Attack (FDA), an adversarial image attack that lowers the weak model's confidence and causes cascades to route queries to the strong model. FDA learns a universal border trigger by optimizing a temperature-flattened objective. This objective pushes the weak model's token distribution on triggered inputs toward less concentrated targets constructed from its clean responses. Across datasets, model families, and deferral metrics, FDA consistently increases strong-model routing while outperforming image-perturbation and prompt-injection baselines. These results show that MLLM cascades are vulnerable to attacks that manipulate compute allocation, forcing unintended strong-model usage without directly targeting answer correctness.", "date": "2026-06-13", "categories": []}
{"id": "2606.15493", "title": "Model Stealing Through the Lens of Model Multiplicity", "abstract": "Model stealing attacks, where adversaries create high-fidelity surrogate models, are a significant threat to the intellectual property of machine learning services. Conventional wisdom suggests these surrogates could provide adversaries with economic leverage comparable to the original service providers. This paper challenges this assumption by evaluating model stealing attacks beyond mere fidelity to the target model. Because query-based extraction provides only partial supervision of the target's input-output behavior, the surrogate is not uniquely identified: many near-optimal surrogates can achieve comparable fidelity while differing in deployment-relevant properties. Instead of performing a classic learning-based model stealing attack, we compute the Rashomon Set (i.e., the set of almost-equally-accurate models) of surrogate models, and evaluate its diversity using multiplicity metrics (ambiguity, discrepancy, and Rashomon Capacity) and group fairness metrics. Across tabular, medical imaging, and NLP tasks, our experiments on real-world datasets reveal that despite exhibiting similar fidelity to the target model, surrogate models can display significant variances in other critical performance metrics. These findings cast doubt on the presumed equivalence between high-fidelity surrogates and the target model in practical deployment scenarios.", "date": "2026-06-13", "categories": []}
{"id": "2606.15788", "title": "GAS-Leak-LLM: Genetic Algorithm-Based Suffix Optimization for Black-Box LLM Jailbreaking", "abstract": "Large Language Models (LLMs) constitute pivotal components within the AI-dominated information technology ecosystem. To mitigate risks associated with harmful or policy-violating outputs, commercial systems employ advanced alignment strategies and multi-layered content moderation mechanisms. Despite these safeguards, recent research has demonstrated that LLMs remain vulnerable to adversarial manipulation, particularly through jailbreaking and prompt injection techniques. In this work, we propose GAS-Leak-LLM a novel jailbreaking attack based on a genetic algorithm that systematically evolves adversarial suffix to bypass safety constraints. Operating in a strict black-box setting, our method requires no access to model parameters or internals, thereby reflecting realistic threat scenarios in deployed systems. Through the iterative application of selection, mutation, and crossover heuristics, the framework systematically explores the discrete prompt space to identify high-fitness adversarial suffixes. Empirical findings reveal critical shortcomings in existing safety enforcement mechanisms and confirm the effectiveness and practical viability of the proposed attack.", "date": "2026-06-14", "categories": []}
{"id": "2606.16100", "title": "Your \"Pro\" LLM Subscription May Actually Be \"Free\": Exposing Fingerprint Spoofing Risks in LLM Inference Services", "abstract": "As Large Language Model (LLM) APIs become ubiquitous, users increasingly rely on black-box fingerprinting to verify that providers are serving the advertised premium models. However, these methods may overlook adversarial providers who manipulate model weights to cheat the fingerprint process. We introduce a novel threat termed fingerprint spoofing, where a malicious provider stealthily serves a weaker model that has been parameter-efficiently fine-tuned to mimic a stronger model, thereby evading user-side fingerprinting. We first formally prove that user-side resource constraints (i.e., finite query budgets and weak fingerprinting classifiers) make current fingerprinting vulnerable to fingerprint spoofing. Guided by this theoretical analysis, we propose GhostPrint, a cost-effective attack framework leveraging surrogate modeling, reward-ranked fine-tuning, and knowledge distillation. Extensive evaluations in both static and continual fingerprinting settings demonstrate that GhostPrint allows weak models to consistently bypass representative fingerprint methods while maintaining utility at a low fine-tuning cost, exposing a critical vulnerability in current LLM fingerprinting pipelines.", "date": "2026-06-15", "categories": []}
{"id": "2606.16242", "title": "Rapid Poison: Practical Poisoning Attacks Against the Rapid Response Framework", "abstract": "The Rapid Response (RR) framework, deployed in production systems, including Anthropic's ASL-3 safeguards, continuously improves jailbreak-detection classifiers. When new jailbreaks emerge that bypass these classifiers, Rapid Response generates synthetic variants for training, helping the model generalize from the new attacks and quickly adapt. We reveal that prompt injection can infiltrate this pipeline to deliver poisoned samples into the classifier's training set, enabling two attack objectives: (I) targeted poisoning attacks that create false positives on harmless samples by categorizing them as a jailbreak, with a specific desired feature (e.g., certain formatting, subject, or keyword), (II) concept-based backdoor attacks that induce false negatives on jailbreak inputs, generalizing even to jailbreaks from attack strategies the defender explicitly trained against, when the backdoor trigger is present. Importantly, our threat model restricts adversaries to modifying only jailbreak samples (not benign data or labels), a constraint unexplored by prior work that makes the second objective particularly challenging. We address this with Omission Attack, which exploits a new phenomenon: when training on concept-absent unsafe samples, the classifier misassociates that concept's presence with the safe label. Both attacks cause substantial and in some cases near-complete label flipping at only a 1% poisoning rate, achieving up to 100% false positive rates and up to 96% false negative rates.", "date": "2026-06-15", "categories": []}
{"id": "2606.16519", "title": "BadWorld: Adversarial Attacks on World Models", "abstract": "Visual world models (VWMs) synthesize interactive, action-conditioned rollouts from a single context image. However, it remains an open question how robust these models are to adversarial perturbations. Standard adversarial attacks fail to assess this vulnerability because attackers lack ground-truth future videos and cannot predict subsequent user controls. We introduce BadWorld, a label-free adversarial framework tailored for autoregressive VWMs that systematically overcomes both constraints. First, to bypass the need for future supervision, we propose a self-supervised velocity attack that directly disrupts the early denoising dynamics of the model. Second, to ensure the attack generalizes across unpredictable user actions, we formulate a trajectory-adaptive bi-level optimization that actively mines hard control sequences to forge control-agnostic perturbations. Evaluated on representative VWMs with continuous and discrete controls, BadWorld exposes severe structural fragility. Visually indistinguishable adversarial images reliably trigger catastrophic degradation in future rollouts, leading to incomplete denoising, structural collapse, and control inconsistency. These findings reveal critical risks for deploying VWMs in safety-critical systems while highlighting a practical mechanism for privacy protection.", "date": "2026-06-15", "categories": []}
{"id": "2606.16605", "title": "ARB4WM: An Adversarial Robustness Benchmark for World Models in Continuous Control", "abstract": "World models are widely used in robotic and agentic engineering control systems due to their ability to learn latent dynamics for planning and decision-making. As these systems are increasingly deployed in safety-critical settings, understanding their robustness under adversarial conditions has become essential. However, existing evaluations lack a unified benchmark for testing adversarial threats across the policy, value, and latent-dynamics levels of world-model agents. To fill this gap, we present ARB4WM, a unified evaluation framework for pre-deployment robustness and risk assessment of world-model agents under visual perturbations. ARB4WM defines five white-box loss objectives across these three levels and studies their effects when combined with single-step or multi-step perturbation strategies and temporal attack modes, including full-frame, half-sequence, and sparse-frame exposure. Specifically, we evaluate four Dreamer-style agents across 20 tasks from MetaWorld and the DeepMind Control Suite under different loss objectives, perturbation strategies, and temporal attack modes. Results show that attacks targeting value estimation, latent representations, and RSSM dynamics can be as damaging as direct policy disruption, and that early or frequent perturbations are especially harmful, while input-level defenses provide limited recovery under adaptive attacks. These findings suggest that safety, risk, and reliability assessment for world models should cover multiple component-oriented attack objectives and temporal exposure protocols rather than relying solely on action-space robustness. Source code is available at https://github.com/zaoanguai/ARB4WM.", "date": "2026-06-15", "categories": []}
{"id": "2606.16751", "title": "Automated jailbreak attack targeting multiple defense strategies", "abstract": "Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks. However, their safety remains a critical concern due to their susceptibility to adversarial prompt-based attacks. In this paper, we present UNIATTACK, an adversarial testing framework designed from a defense-oriented perspective to systematically construct effective black-box attack prompts. Unlike prior approaches that rely on static templates or iterative model-specific tuning, UNIATTACK extracts minimal but high-impact attack features from diverse existing attacks, optimizes them via a specialized attacker LLM, and composes them into flexible templates through automated refinement process. This feature-centric construction enables one-shot attacks that generalize across multiple models and safety categories, providing a practical tool for assessing LLM robustness. Our evaluation results shows that compared to the baselines, UNIATTACK achieves an average attack success rate (ASR) improvement of 64.63\\%-248.82\\% on models deployed with multi-layered defense mechanisms and it only takes 0.03\\%-4.96\\% cost of the baselines. UNIATTACK artifact is available at https://anonymous.4open.science/r/UniAttack-Artifact-30F1.", "date": "2026-06-15", "categories": []}
{"id": "2604.00887", "title": "Towards Physically Realizable Adversarial Attenuation Patch against SAR Object Detection", "abstract": "Deep neural networks have demonstrated excellent performance in SAR target detection tasks but remain susceptible to adversarial attacks. Existing SAR-specific attack methods can effectively deceive detectors; however, they often introduce noticeable perturbations and are largely confined to digital domain, neglecting physical implementation constrains for attacking SAR systems. In this paper, a novel Adversarial Attenuation Patch (AAP) method is proposed that employs energy-constrained optimization strategy coupled with an attenuation-based deployment framework to achieve a seamless balance between attack effectiveness and stealthiness. More importantly, AAP exhibits strong potential for physical realization by aligning with signal-level electronic jamming mechanisms. Experimental results show that AAP effectively degrades detection performance while preserving high imperceptibility, and shows favorable transferability across different models. This study provides a physical grounded perspective for adversarial attacks on SAR target detection systems and facilitates the design of more covert and practically deployable attack strategies. The source code is made available at https://github.com/boremycin/SAAP.", "date": "2026-04-01", "categories": []}
{"id": "2606.13801", "title": "Neural Variability Enhances Artificial Network Robustness", "abstract": "Neural responses in cortex exhibit substantial trial-to-trial variability in response to repeated stimuli, while peripheral sensory neurons respond far more consistently, leading many to wonder whether stochasticity may carry meaning. Existing work has argued that noise and signal correlations may be optimized for discrimination in animals, whereas artificial neural network (ANN) studies have shown similar benefits of noise in machine learning tasks, although most ANN work has neglected the effects of correlations. Here we investigate whether correlated noise improves the robustness of artificial neural networks to adversarial attacks and naturalistic image modifications. Using the covariance of activations under modified versus clean inputs, we find that structured noise may significantly improve network robustness. Robustness to naturalistic image modifications benefits most from structure, but this structure transfers poorly across modification types. In contrast, noise structure from adversarial attacks can generalize to other kinds of attacks. These results suggest that structured noise in ANN activations generally improves robustness, establishing a biologically plausible strategy for creating robust artificial neural networks that only relies on local information.", "date": "2026-06-11", "categories": []}
{"id": "2606.14078", "title": "Rethinking Backdoor Adversarial Unlearning through the Lens of Catastrophic Forgetting in Continual Learning", "abstract": "Existing studies reveal that current backdoor defenses exhibit limited robustness and often fail against specific types of attacks. More concerningly, prevailing safety tuning strategies tend to provide only superficial safety protection, as they fall short of completely eliminating the backdoor effects. In this work, we present a novel formulation of backdoor learning and unlearning as a sequential, three-stage process from a continual learning perspective. Within this framework, we formally define complete backdoor unlearning and further derive the necessary conditions for achieving it based on the mechanism of catastrophic forgetting. Guided by these insights, we propose Blind Inversion-Backdoor Adversarial Unlearning (BI-BAU), which formulates the generation of adversarial examples satisfying the unlearning conditions as a blind inversion problem. We solve this by integrating the bi-level optimization process of adversarial training into an Expectation-Maximization (EM) algorithm framework to optimize the maximum a posteriori (MAP) objective. Furthermore, BI-BAU is extended to untargeted adversarial scenarios with unknown target classes, as well as to multi-modal contrastive learning tasks, enhancing its applicability to real-world deployment scenarios where pre-trained models may be compromised. Extensive experiments demonstrate that our method exhibits general applicability across a wide spectrum of backdoor attacks and can effectively and thoroughly eliminate the backdoor effects from a backdoor model.", "date": "2026-06-12", "categories": []}
{"id": "2606.14504", "title": "Scratched Lenses, Shifted Depth: Passive Camera-Side Optical Attacks", "abstract": "Physical adversarial attacks on vision systems are typically studied through scene manipulation, such as adversarial patches or projections, where the adversary controls what the camera observes. Camera-side attacks using stickers or auxiliary optics have also been explored, but they treat attacks as image-space perturbations from designed patterns. This misses how physical imperfections interact with scene-dependent lighting and optics. We identify a threat: passive lens-side damage that is persistent yet trigger-conditioned, producing optical artifacts that bias geometric inference under particular visual conditions.\n  We instantiate this threat through Scratch-induced Lens Adversarial Streak Hijacking SLASH, a physical-world attack caused by small scratches on a camera lens or protective cover. Scratches interact with bright light sources and specular reflections to create structured streak artifacts that distort depth cues. Since the perturbation is fixed in the optical path but triggered by the scene, it is both persistent and selective. We formulate the attack in optical space, model the scratch pattern as a trigger-conditioned optical channel, and optimize one fixed configuration across diverse viewing conditions. We evaluate SLASH on monocular depth estimation and monocular 3D object detection in digital and real-world settings. Under the fixed-scratch constraint, directional depth shifts reach up to 32% relative error for monocular depth estimation, with consistent effects on monocular 3D object detection. Physical experiments confirm transfer to real camera recordings, inducing depth shifts above the model's natural prediction baseline. These findings reveal an attack surface where benign-looking hardware imperfections act as latent, scene-triggered adversarial mechanisms, challenging assumptions about physical robustness and motivating defenses for secure vision systems.", "date": "2026-06-12", "categories": []}
{"id": "2606.14658", "title": "Giving AI a Headache: Acoustic Adversarial Attacks to Computer Vision Applications", "abstract": "Artificial Intelligence (AI) is increasingly used to automate a variety of real-world computer vision (CV) applications, such as autonomous vehicle control, facial recognition, and security cameras. Recent research has shown that acoustic vibration can induce real physical motion in cameras, interfering with their internal stabilization mechanisms. Because the motion falls outside the conditions the stabilization system was designed to handle, the system introduces artifacts into the frame, causing AI-based CV models to misclassify, miss targets, or hallucinate objects.\n  Previous work used ultrasonic frequencies (>20 kHz) to perform short-range attacks, which limits them to short distances due to the attenuation exhibited by high frequencies. In this work, we investigate acoustic attacks using lower frequencies in the audible range (<20 kHz), and we further expand our analysis to include how various image and object features are affected by the attacks.\n  Specifically, we performed physical experiments to demonstrate the viability of our attacks on an off-the-shelf object detection model (YOLO11) by resonating a commercially available camera with various frequencies. Based on our results, we provide insights into several factors that make an AI CV system more vulnerable to these attacks, which could help inform the development of future mitigation strategies.", "date": "2026-06-12", "categories": []}
{"id": "2605.18231", "title": "Attacking the First-Principle: A Black-Box, Query-Free Targeted Mimicry Attack on Binary Function Classifiers", "abstract": "Binary function classifiers play a crucial role in maintaining the security and integrity of software systems by detecting malicious code and unauthorized modifications. However, machine learning-based classifiers are vulnerable to adversarial attacks that can evade detection. In this study, we present Kelpie, a novel framework for executing mimicry attacks, a stronger type of targeted evasion attacks, on binary function classifiers in a black-box, zero-query setting. Unlike previous approaches that rely on querying the target classifier to refine untargeted evasion attacks, Kelpie leverages code transformations that preserve the functionality of malicious payloads while causing them to be misclassified as we want. Through extensive experimentation, we demonstrate that Kelpie can successfully execute mimicry attacks against six state-of-the-art binary function classifiers representing different model architectures without requiring direct interaction with them. We further validate our approach with a practical demonstration, involving a keylogger and a wiper concealed within benign-looking functions embedded in an application. This work, to our best knowledge, is the first to demonstrate such a mimicry attack in a black-box, zero-query context, raising important questions about the reliability and security of existing machine learning-based binary function classifiers.", "date": "2026-05-18", "categories": []}
{"id": "2606.12263", "title": "VOID: Defeating Unauthorized Mimicry in Latent Diffusion Models", "abstract": "While Latent Diffusion Models (LDMs) have revolutionized visual synthesis, they are increasingly exploited for unauthorized mimicry of individuals. Existing defenses inject deceptive perturbations to steer the generated images toward irrelevant targets. However, this approach hinges on an ungrounded assumption: subtle perturbations can maintain their deceptive efficacy throughout an LDM's extensive generation process. In reality, the model's innate restoration mechanism will remove such perturbations and cause individual identities to re-emerge in the images generated.\n  We propose VOID, a defense framework that overcomes this conundrum by manipulating an LDM's intrinsic stochasticity. VOID perturbs the diffusion pipeline in two novel ways: 1) amplifying the latent encoding errors to shatter an image's semantic structure, and 2) counteracting the target guidance signals to suppress the model's restoration capabilities. This results in a semantic corruption that thwarts any unauthorized mimicry. Notably, the security gain does not come at the price of visual utility, as VOID simultaneously manages to confine perturbations to human-imperceptible regions of protected images. Our comprehensive evaluation of 24 state-of-the-art defenses against 10 mimicry attacks on 5 datasets demonstrates VOID's unprecedented protection power: it increases the average Frechet Inception Distance (FID) from 113 to 365, a 223% improvement over the strongest defense to date.", "date": "2026-06-10", "categories": []}
{"id": "2606.12655", "title": "Amnesia: A Stealthy Replay Attack on Continual Learning Dreams", "abstract": "Continual learning (CL) models often use experience replay to reduce catastrophic forgetting, but their robustness to replay sampling interference remains underexplored. Existing CL attacks alter inputs or training pipelines (poisoning/backdoors) and rarely include explicit auditable constraints, limiting realism. Here, auditability means a monitor can verify compliance from sampler-visible telemetry - e.g., logged replay index/label statistics - by checking that the realized replay class histogram stays close to a nominal baseline and that replay rate is unchanged per batch and/or over a rolling window. We study a limited-privilege insider who controls only replay index selection, not pixels, labels, or model parameters, while staying within auditable limits such as queue priorities. We introduce Amnesia, a replay composition attack that maximizes degradation under two budgets: a visibility budget delta bounding the TV/KL divergence from a nominal class histogram p0, and a mass budget f fixing the replay rate. Amnesia has two steps: (i) compute lightweight class utilities, such as EMA loss or confidence, to tilt p0 toward harmful classes; and (ii) project the tilt back into the delta-ball using efficient KL (exponential tilt) or TV (balanced mass redistribution) optimizers. A windowed scheduler enforces rolling audits. Across challenging CL benchmarks and strong replay baselines, Amnesia consistently lowers final accuracy (ACC) and worsens backward transfer (-BWT). The KL variant delivers high impact while remaining largely undetected under multiple audit schemes, including per-batch and rolling-window checks. The TV variant is more damaging but easier to detect, especially under tight per-class constraints. These results expose index-only replay control as a practical, auditable threat surface in CL systems and establish a principled impact-visibility trade-off.", "date": "2026-06-10", "categories": []}
{"id": "2606.13022", "title": "Quality-Preserving Imperceptible Adversarial Attack on Skeleton-based Human Action Recognition", "abstract": "Adversarial attacks on skeletal human action recognition have received significant attention. However, existing methods typically introduce noise-like perturbations that degrade motion quality post-attack, and thereby are inherently perceptible with recent advancements in S-HAR systems. We discover that this degradation stems from the gap between empirical and true risks during the optimization process of previous adversarial attacks. To address this issue, we propose an attack where adversarial motions are obtained without compromising their motion quality. To minimize the risk gap and preserve motion quality, we propose a distribution-based adversarial attack method without introducing noise-like perturbations. To faithfully evaluate the motion quality, we propose a new metric that aligns with human perception on real-world naturalness. Experiments have been conducted on the state-of-the-art S-HAR methods across two datasets, demonstrating the superiority of our method in both the attack success rate and the post-attack motion quality through qualitative and quantitative analyses. The success of our quality-preserving attack application and distribution-based method raises serious concerns about the robustness of action recognizers, highlighting the need for further enhancements in this domain.", "date": "2026-06-11", "categories": []}
{"id": "2606.13282", "title": "ERTS: Adversarial Robustness Testing of Ethical AI via Semantic Perturbation in a Bounded Consequence Space", "abstract": "As AI systems are deployed in high-stakes ethical contexts such as healthcare triage, autonomous vehicle control, and employment screening, formal methods for evaluating their robustness against adversarial manipulation of ethical reasoning remain underdeveloped. This paper introduces the Ethical Robustness Testing System (ERTS), a closed-pipeline framework that: (1) encodes ethical dilemmas into a 22-dimensional Ethical Consequence Space (ECS) grounded in established ethical theory; (2) applies 17 semantic perturbation functions subject to 6 validity constraint classes including a novel semantic coherence constraint; (3) measures decision deviation via a 4-component Ethical Instability Index (EII); and (4) produces domain-adaptive pre-deployment robustness assessment verdicts. We evaluate 4 structured baseline models and 2 production LLMs (Gemini 2.0 Flash and Llama 3.2) across 50 ethical scenarios spanning 8 deployment domains, generating 1,500 adversarial test cases. Results demonstrate that only 33% of models achieve assessment clearance, with the local Llama-3.2 model proving particularly vulnerable to fairness corruption and information degradation attacks (ERS = 0.737). To the best of our knowledge, no existing framework combines a bounded ethical consequence space, semantic coherence constraints, and domain-adaptive assessment in a single adversarial testing pipeline.", "date": "2026-06-11", "categories": []}
{"id": "2606.13439", "title": "S-GBT: Smooth Growth Bound Tensor for Certified Robustness Against Word Substitution Attacks in NLP", "abstract": "Despite recent progress in Natural Language Processing (NLP), models remain vulnerable to word substitution attacks. Most existing defenses focus on first order sensitivity and measure how much the output changes when the input is slightly perturbed. However, they ignore how this sensitivity evolves, which is described by curvature. When gradients vary sharply, models can still fail. This paper introduces the Smooth Growth Bound Tensor (S-GBT), a second order method that bounds the Hessian element-wise, for which we provide formal theoretical proofs on the resulting robustness bounds. A regularization term is added during training to minimize these bounds. This yields tighter certified robustness against word substitution attacks. The change in the output under word substitution is bounded by both a linear term and a quadratic term. S-GBT is derived for two architectures: Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN). The method is integrated directly into the training objective. Its effectiveness is evaluated on multiple benchmark datasets. The results show that combining first and second order regularization improves certified robust accuracy by up to 23.4% compared to prior methods, while clean accuracy remains competitive. These findings indicate that controlling both the gradient and its variation is a promising direction for building more robust models.", "date": "2026-06-11", "categories": []}
{"id": "2506.03933", "title": "Diffusion-based Cumulative Adversarial Purification for Vision Language Models", "abstract": "Vision Language Models (VLMs) have shown remarkable capabilities in multimodal understanding, yet their susceptibility to adversarial perturbations poses a significant threat to their reliability in real-world applications. Despite often being imperceptible to humans, these perturbations can drastically alter model outputs, leading to erroneous interpretations and decisions. This paper introduces DiffCAP, a novel diffusion-based purification strategy that can effectively neutralize adversarial corruptions in VLMs. We theoretically establish a provable recovery region in the forward diffusion process and meanwhile quantify the convergence rate of semantic variation with respect to VLMs. These findings manifest that adversarial effects monotonically fade as diffusion unfolds. Guided by this principle, DiffCAP leverages noise injection with a similarity threshold of VLM embeddings as an adaptive criterion, before reverse diffusion restores a clean and reliable representation for VLM inference. Through extensive experiments across six datasets with three VLMs under varying attack strengths in three task scenarios, we show that DiffCAP outperforms existing defense techniques by a substantial margin. Notably, DiffCAP significantly reduces both hyperparameter tuning complexity and the required diffusion time, thereby accelerating the denoising process. Equipped with theorems and empirical support, DiffCAP provides a robust and practical solution for securely deploying VLMs in adversarial environments. The source code is available at https://github.com/JasonFu1998/DiffCAP.", "date": "2025-06-04", "categories": []}
{"id": "2602.05746", "title": "Learning to Inject: Automated Prompt Injection via Reinforcement Learning", "abstract": "Prompt injection is a critical vulnerability in LLM agents, yet the strongest methods still rely on human red-teamers and hand-crafted prompts. Adapting automated jailbreak optimizers does not close this gap: jailbreaks shape models toward generic compliance, while prompt injection requires emitting specific tool calls with correct parameters. The success signal is binary, and randomly sampled suffixes almost never trigger it, so standard optimizers have no gradient to follow. We present AutoInject, a black-box reinforcement learning (RL) framework that learns adversarial suffixes for prompt injection. A learned comparison-based reward scores each candidate against the best suffix seen so far, turning the binary signal into a dense reward suitable for RL optimization. The framework supports both online query-based attacks and offline-trained transferable suffixes that need no utility access at deployment, and incorporates a utility objective when task-completion feedback is available. On AgentDojo, AutoInject outperforms template attacks, GCG, TAP, and adaptive attack across production models, with statistically significant improvements under McNemar's test with p<0.05. Suffixes learned by AutoInject also break Meta-SecAlign-70B, a model fine-tuned specifically to resist prompt injection, where template attacks fail outright. The results establish an automated baseline for prompt injection and expose a gap between preference-based defenses and adaptive optimization-based attackers.", "date": "2026-02-05", "categories": []}
{"id": "2605.31219", "title": "Latent Geometric Chords for Query-Efficient Decision-Based Adversarial Attacks", "abstract": "While decision-based black-box adversarial attacks present a severe security threat, current methodologies suffer from fundamental limitations. Pixel-wise attacks frequently introduce unnatural, high-frequency visual artifacts, while latent-space frameworks are confined by the limited search space of low-dimensional manifolds and inherent reconstruction flaws. To resolve these limitations, we propose Latent Geometric Chords (LGC) for Query-Efficient Decision-Based Adversarial Attacks alongside a variant, LGC-H. At its core, LGC navigates decision boundaries by executing a curvature-aware geometric search within a compressed semantic manifold. To guarantee high visual fidelity and circumvent dimensionality bottlenecks, we introduce a Residual-based Adversarial Generation (RAG) mechanism. RAG isolates semantic perturbations as geometric chords and superimposes them directly onto the original source image. RAG substantially resolves baseline reconstruction flaws and effectively doubles the permissible search space dimensions. Experimental results demonstrate that LGC achieves robust cross-dataset transferability and substantially outperforms state-of-the-art baselines. Notably, our method, LGC, minimizes perturbation magnitudes while achieving state-of-the-art visual fidelity--with a Structural Similarity Index Measure (SSIM) exceeding 0.99 and a Learned Perceptual Image Patch Similarity (LPIPS) below 0.01 at 5000 queries--and sustaining high attack success rates under stringent perceptual constraints, successfully compromising adversarially trained robust models. The source code is available at: https://github.com/eihmuekhine/Latent-Geometric-Chords.", "date": "2026-05-29", "categories": []}
{"id": "2606.11265", "title": "When Poison Fails After Retrieval: Revisiting Corpus Poisoning under Chunking and Reranking Pipelines", "abstract": "Retrieval-Augmented Generation (RAG) systems are vulnerable to corpus poisoning attacks that manipulate downstream model outputs through malicious knowledge injection. Existing studies mainly evaluate poisoning under simplified retrieval settings, overlooking practical RAG pipelines involving document chunking, dense retrieval, reranking, and grounded generation. In this paper, we revisit corpus poisoning under realistic multi-stage retrieval pipelines and show that many existing attacks substantially degrade after reranking despite achieving high retrieval-stage relevance. We identify retrieval granularity mismatch as a key reason for this failure: document-level adversarial signals are often fragmented during chunking, while rerankers favor locally coherent and answer-bearing passages rather than globally optimized semantic similarity. Based on this observation, we propose Chunk-aware and Rerank-Consistent Poisoning (CRCP), a poisoning framework that jointly optimizes retrieval relevance, reranker consistency, and chunk-boundary robustness. CRCP explicitly models chunking transformations during optimization to generate locally self-contained adversarial passages that remain effective under varying chunking configurations. Experiments on standard RAG benchmarks with multiple retrievers and rerankers show that existing poisoning methods are highly sensitive to chunk size and reranking strategies, whereas CRCP achieves substantially higher attack success rates and stronger robustness across realistic retrieval pipelines. Our findings highlight an important realism gap in current RAG security evaluation and suggest that poisoning in modern RAG systems should be studied as a multi-stage retrieval consistency problem rather than a retrieval-only problem.", "date": "2026-06-09", "categories": []}
{"id": "2606.11409", "title": "Risk Under Pressure: Compute-Aware Evaluation of Adversarial Robustness in Language Models", "abstract": "Adversarial robustness evaluations of large language models (LLMs) typically report attack success rate (ASR) under fixed query budgets, implicitly treating all attacks as equally costly. In practice, the computational expense of different attack strategies can vary by orders of magnitude. Consequently, ASR at a fixed budget can obscure the true effort required to jailbreak a model, thereby making it hard to determine whether an attack's cost justifies its payoff to the attacker. We propose a compute-aware evaluation framework based on computational pressure, measured in cumulative floating-point operations (FLOPs), as a proxy for adversarial effort. We introduce risk-compute curves, which map compute budgets to attack risk, and derive two metrics that summarize the average pressure required for a given attack to succeed. Across ten models spanning three families and four different stages in language model training and alignment, evaluated with three attack strategies (gradient-based, iterative refinement, and template-based) on two jailbreak robustness benchmarks, we find: (1) alignment training has non-monotonic effects on compute-space robustness; (2) scaling model size reduces gradient-based attack effectiveness but has limited impact on cheaper template-based attacks; (3) gradient-based attacks optimized on a surrogate model can transfer to a separate target model, providing a way to reduce attacker costs; (4) compute cost varies by up to ${\\approx}5{\\times}$ across harm categories within a single model; and (5) safety-aligned RL increases aggregate cost while leaving some categories disproportionately accessible. We release our framework to enable compute-aware risk assessment and evaluation.", "date": "2026-06-09", "categories": []}
{"id": "2606.11425", "title": "JailbreakOPT: Tool-Assisted Iterative Jailbreak Prompt Optimization", "abstract": "Jailbreak attacks expose persistent safety weaknesses in large language models (LLMs), but existing stateless single-turn methods face a trade-off: hand-crafted prompts are expressive but static, while iterative prompt optimization can adapt but often relies on low-level mutations that require many target queries. We propose JailbreakOPT, a tool-assisted framework for improving iterative single-turn jailbreak prompt optimization. JailbreakOPT organizes diverse atomic jailbreak prompts into an attack tool library and composes them through a unified intra-episode optimization abstraction to generate stronger standalone attack prompts. To reuse experience across attack episodes, JailbreakOPT further frames tool selection as a contextual bandit problem and applies contextual Thompson sampling to guide exploration and exploitation based on past outcomes. Experiments across multiple target LLMs and attack goals show that JailbreakOPT improves attack success rate (ASR) while reducing the number of attacks until success (No.A) compared with atomic single-turn attacks and existing iterative optimization baselines. This paper may contain offensive or harmful content.", "date": "2026-06-09", "categories": []}
{"id": "2606.11804", "title": "Toward Trustworthy AI: Multi-Target Adversarial Attacks and Robust Defenses for Continuous Data Summarization", "abstract": "Trustworthy AI requires reliable data-processing pipelines, not only robust downstream predictive models. As an upstream component, data summarization determines which information is retained and passed to subsequent learning or decision modules. Therefore, adversarial perturbations to the summarization process can compromise trustworthy AI in an upstream manner: they may alter the selected summary, reduce its representativeness, and further degrade the utility of subsequent learning tasks. In this paper, we study adversarial attacks on continuous data summarization under similarity-level perturbations through DR-submodular optimization. We show that a class of multi-resolution image summarization objectives can be formulated as multilinear extensions of non-negative submodular set functions and satisfy DR-submodularity with $m$-weak monotonicity. We then formulate multi-target attack generation as a min-max problem, where one admissible perturbation of the similarity structure is optimized to degrade multiple target summarization models. To mitigate such perturbations, we formulate robust defense against mixed attack types as a regularized max-min problem. For both problems, we develop approximation algorithms with theoretical guarantees. Experiments on real-data and controlled clustered benchmarks show that the proposed attack is effective in representative low-to-moderate budget regimes and can induce downstream task-performance loss. The proposed defense improves the robustness--mitigation trade-off in structured settings, while also revealing the parameter sensitivity of robust protection on real data.", "date": "2026-06-10", "categories": []}
{"id": "2606.12251", "title": "Reinforcement Learning Disrupts Gradient-Based Adversarial Optimization", "abstract": "Gradient-based adversarial attacks remain a dominant threat to deep neural networks (DNNs), as they exploit gradient information to efficiently optimize adversarial perturbations. To address this, we investigate whether reinforcement learning (RL) training can disrupt the gradient structure used by attackers by training image classifiers with policy-gradient objectives and epsilon-greedy exploration. Through systematic experiments across CIFAR-10, CIFAR-100, and ImageNet-100 with multiple architectures, we find that RL-trained classifiers significantly disrupt gradient-based adversarial optimization. To explain this, we conduct a comprehensive mechanism analysis using loss landscape visualization, static and dynamic gradient indicators, and predictive entropy. Our analysis reveals that RL acts as an implicit regularizer, producing models with highly unstable gradient directions and smaller gradient magnitudes. This combination makes each PGD step both unreliable in direction and limited in magnitude, causing gradient-based attacks to fail within practical iteration budgets. We further show that combining RL with adversarial training (RL-adv) provides a dual-layer defense operating at two complementary levels: RL degrades gradient information available to attackers (gradient-level defense), while adversarial training strengthens decision boundaries (boundary-level defense). RL-adv achieves the highest robustness across all major attack types evaluated, including gradient-based (PGD, AutoAttack), transfer-based, and query-based attacks, outperforming SL-adv by a significant margin. These findings identify RL-induced gradient disruption as a complementary robustness mechanism and motivate future research on hybrid SL-RL training schedules that combine SL's efficiency with RL's gradient-regularization properties.", "date": "2026-06-10", "categories": []}
{"id": "2403.00420", "title": "Robust Deep Reinforcement Learning Through Adversarial Attacks and Training : A Survey", "abstract": "Deep Reinforcement Learning (DRL) is a subfield of machine learning for training autonomous agents that take sequential actions across complex environments. Despite its significant performance in well-known environments, it remains susceptible to minor condition variations, raising concerns about its reliability in real-world applications. To improve usability, DRL must demonstrate trustworthiness and robustness. A way to improve the robustness of DRL to unknown changes in the environmental conditions and possible perturbations is through Adversarial Training, by training the agent against well-suited adversarial attacks on the observations and the dynamics of the environment. Addressing this critical issue, our work presents an in-depth analysis of contemporary adversarial attack and training methodologies, systematically categorizing them and comparing their objectives and operational mechanisms.", "date": "2024-03-01", "categories": []}
{"id": "2407.20242", "title": "BadRobot: Jailbreaking Embodied LLM Agents in the Physical World", "abstract": "Embodied AI represents systems where AI is integrated into physical entities. Large Language Model (LLM), which exhibits powerful language understanding abilities, has been extensively employed in embodied AI by facilitating sophisticated task planning. However, a critical safety issue remains overlooked: could these embodied LLMs perpetrate harmful behaviors? In response, we introduce BadRobot, a novel attack paradigm aiming to make embodied LLMs violate safety and ethical constraints through typical voice-based user-system interactions. Specifically, three vulnerabilities are exploited to achieve this type of attack: (i) manipulation of LLMs within robotic systems, (ii) misalignment between linguistic outputs and physical actions, and (iii) unintentional hazardous behaviors caused by world knowledge's flaws. Furthermore, we construct a benchmark of various malicious physical action queries to evaluate BadRobot's attack performance. Based on this benchmark, extensive experiments against existing prominent embodied LLM frameworks (e.g., Voxposer, Code as Policies, and ProgPrompt) demonstrate the effectiveness of our BadRobot. Our code is available at https://github.com/Rookie143/BadRobot.", "date": "2024-07-16", "categories": []}
{"id": "2501.00745", "title": "Dynamics of Adversarial Attacks on Large Language Model-Based Search Engines", "abstract": "The increasing integration of Large Language Model (LLM) based search engines has transformed the landscape of information retrieval. However, these systems are vulnerable to adversarial attacks, especially ranking manipulation attacks, where attackers craft webpage content to manipulate the LLM's ranking and promote specific content, gaining an unfair advantage over competitors. In this paper, we study the dynamics of ranking manipulation attacks. We frame this problem as an Infinitely Repeated Prisoners' Dilemma, where multiple players strategically decide whether to cooperate or attack. We analyze the conditions under which cooperation can be sustained, identifying key factors such as attack costs, discount rates, attack success rates, and trigger strategies that influence player behavior. We identify tipping points in the system dynamics, demonstrating that cooperation is more likely to be sustained when players are forward-looking. However, from a defense perspective, we find that simply reducing attack success probabilities can, paradoxically, incentivize attacks under certain conditions. Furthermore, defensive measures to cap the upper bound of attack success rates may prove futile in some scenarios. These insights highlight the complexity of securing LLM-based systems. Our work provides a theoretical foundation and practical insights for understanding and mitigating their vulnerabilities, while emphasizing the importance of adaptive security strategies and thoughtful ecosystem design.", "date": "2025-01-01", "categories": []}
{"id": "2601.19210", "title": "Contrastive Spectral Rectification: Test-Time Defense towards Zero-shot Adversarial Robustness of CLIP", "abstract": "Vision-language models (VLMs) such as CLIP have demonstrated remarkable zero-shot generalization, yet remain highly vulnerable to adversarial examples (AEs). While test-time defenses are promising, existing methods fail to provide sufficient robustness against strong attacks and are often hampered by high inference latency and task-specific applicability. To address these limitations, we start by investigating the intrinsic properties of AEs, which reveals that AEs exhibit severe feature inconsistency under progressive frequency attenuation. We further attribute this to the model's inherent spectral bias. Leveraging this insight, we propose an efficient test-time defense named Contrastive Spectral Rectification (CSR). CSR optimizes a rectification perturbation to realign the input with the natural manifold under a spectral-guided contrastive objective, which is applied input-adaptively. Extensive experiments across 16 classification benchmarks demonstrate that CSR outperforms the SOTA by an average of 18.1% against strong APGD with modest inference overhead. Furthermore, CSR exhibits broad applicability across diverse visual tasks. Code is available at https://github.com/Summu77/CSR.", "date": "2026-01-27", "categories": []}
{"id": "2606.09909", "title": "Bypassing Copyright Protection in Diffusion-based Customization via Two-Stage Latent Feature Optimization", "abstract": "With the growing concerns over copyright infringement in diffusion-based customization, adversarial attacks have emerged as a prominent defense strategy to prevent malicious content forgery in personalized image generation. However, current defenses typically introduce persistent perturbations in the latent space of Latent Diffusion Models (LDMs), which remain susceptible to adaptive bypasses by adversaries. In this paper, we introduce Two-Stage Latent Feature Optimization (TS-LFO), an efficient and effective copyright-stealing attack against protected diffusion-based customization. We begin by observing that existing defenses primarily disrupt the mapping between input images and their latent representations, thereby degrading the model's ability to produce personalized outputs. To counteract this, TS-LFO restores the broken mapping through a two-stage optimization process. In the Latent Denoising Stage, we enhance semantic consistency between latent codes and input images by jointly minimizing a Latent-Image Alignment Loss and a Latent Diffusion Loss with timestep-dependent weights, effectively suppressing the high-frequency noise introduced by defenses. In the Latent Reconstruction Stage, we recover low-frequency semantic information using pixel-level constraints to refine the latent features. Extensive experiments show that TS-LFO consistently bypasses state-of-the-art (SOTA) copyright defenses and outperforms SOTA copyright attacks such as DiffPure, GrIDPure and IMPRESS across diverse settings.", "date": "2026-06-06", "categories": []}
{"id": "2606.10371", "title": "Test-time Adversarial Takeover: A Real-time Hijacking Interface against Robotic Diffusion Policies", "abstract": "Diffusion-based action generation has become a foundational component of embodied AI, but its reliance on visual conditioning leaves deployed visuomotor policies vulnerable to adversarial manipulation. Most prior attacks focus on disruption: they perturb the observation stream to reduce task success or induce erratic behavior. We study a stronger threat, Test-time Adversarial Takeover (TAKO), in which an attacker obtains a real-time steering interface over a frozen robot policy and turns it into a remotely piloted instrument. TAKO learns a small vocabulary of reusable universal patches through differentiable diffusion inference; at test time, the attacker switches among these patches in the camera stream to compose attacker-chosen trajectories. This works because the perturbation acts on the visual conditioning pathway, where the induced bias can persist through iterative generative inference. We further show that the natural targeted baseline, target-policy matching, fails because the victim policy cannot reliably supervise itself on out-of-distribution target shifts. Across four tasks (2D manipulation, simulated aerial delivery, simulated ground navigation, and physical-world ground navigation), two visual encoders (ResNet-18 and EfficientNet-B0 + Transformer), and three generative inference families (DDPM, DDIM, and flow matching), human operators achieve 100\\% takeover success on attacker-defined objectives in every evaluated setting. The project page is available at https://tako-attack.github.io.", "date": "2026-06-09", "categories": []}
{"id": "2606.10571", "title": "Improving Adversarial Transferability on Vision-Language Pre-training Models via Surrogate-Specific Bias Correction", "abstract": "Adversarial examples reveal vulnerabilities in Vision-Language Pre-training (VLP) models and provide insights for improving robustness. A key property is cross-model transferability, which enables transfer-based black-box attacks. However, existing attacks often rely heavily on the surrogate model, causing cross-model performance drops. One reason is that adversarial optimization may follow surrogate model responses more than input semantics, making the update direction effective on the surrogate but less transferable to unseen targets. We refer to this dependency as surrogate-specific bias. Motivated by this observation, DeBias-Attack improves transferability by correcting surrogate-specific bias in adversarial optimization directions. It maintains two perturbation branches. The main branch optimizes a perturbation on the original image and obtains the adversarial gradient used to disrupt image-text alignment. The reference branch optimizes a perturbation on a weak-semantic image constructed from the dataset mean image with small Gaussian noise resampled at each iteration. Since this weak-semantic image contains little clear visual content, its optimization reflects surrogate responses more than image semantics, and its reference gradient estimates surrogate-specific bias. DeBias-Attack removes the aligned projection of the main gradient on the reference gradient before updating the adversarial image, then performs context-aware text substitution using the updated adversarial image. DeBias-Attack is the first transfer-based VLP attack that corrects surrogate-specific bias through gradient correction. Experiments show strong performance across VLP models, downstream tasks, and open-source and closed-source multimodal large language models.", "date": "2026-06-09", "categories": []}
{"id": "2606.10742", "title": "MemVenom: Triggered Poisoning of Multimodal Memories in Web Agents", "abstract": "External memory has become a core component of modern web agents, enabling long-horizon reasoning through the retrieval of past experiences. However, this paradigm introduces a critical vulnerability: malicious content injected into memory can be persistently recalled and repeatedly influence agent behavior. In this work, we identify and systematically study multimodal memory poisoning, an overlooked yet practical attack surface in web-agent systems. We propose MemVenom, a unified black-box attack framework that poisons graph-structured external memory with coordinated text-image evidence. Our method consists of a two-stage design: (1) a trigger-conditioned retrieval attack that ensures high-probability recall of malicious memory, and (2) a post-retrieval attack induction that leverages adversarial perturbations and stealthy OCR injection to override the original user objective. Unlike prior attacks that operate on prompts or text-only memory, our approach enables persistent, reusable, and goal-agnostic attacks without modifying model parameters or re-optimizing malicious tasks. Experiments across multiple web-agent frameworks and vision-language models demonstrate that MemVenom achieves strong end-to-end attack success with minimal impact on benign performance, reaching up to 99.15% on GPT-5-family web agents, while transferring effectively across architectures and model scales.", "date": "2026-06-09", "categories": []}
{"id": "2506.17231", "title": "Efficient and Stealthy Jailbreak Attacks via Adversarial Prompt Distillation from LLMs to SLMs", "abstract": "Current jailbreak attacks on large language models (LLMs) predominantly rely on LLMs themselves to generate adversarial prompts, creating a critical efficiency bottleneck: each attack requires substantial computational resources and API queries, limiting scalability and practical deployment. To overcome this limitation, we propose Adversarial Prompt Distillation (APD), a novel framework that transfers jailbreaking capabilities from LLMs to small language models (SLMs) for efficient, low-resource attacks. APD integrates three key components: (1) masked adversarial knowledge pre-training via LoRA fine-tuning, (2) dynamic temperature-controlled knowledge distillation to bridge architectural gaps, and (3) reinforcement learning-based template optimization for adaptive refinement. Extensive experiments across 12 models show that APD achieves state-of-the-art attack success rates (e.g., 96.4% ASR_k on GPT-4) while dramatically improving efficiency - generating prompts 3.7x faster with 11.3x fewer parameters than teacher models. Our work establishes the first practical framework for lightweight jailbreak attacks, exposes new vulnerabilities in LLM defenses, and provides a scalable testbed for advancing AI safety research. Our code is available at: https://github.com/lxgem/Efficient_and_Stealthy_Jailbreak_Attacks_via_Adversarial_Prompt.", "date": "2025-05-26", "categories": []}
{"id": "2507.02606", "title": "De-AntiFake: Rethinking the Protective Perturbations Against Voice Cloning Attacks", "abstract": "The rapid advancement of speech generation models has heightened privacy and security concerns related to voice cloning (VC). Recent studies have investigated disrupting unauthorized voice cloning by introducing adversarial perturbations. However, determined attackers can mitigate these protective perturbations and successfully execute VC. In this study, we conduct the first systematic evaluation of these protective perturbations against VC under realistic threat models that include perturbation purification. Our findings reveal that while existing purification methods can neutralize a considerable portion of the protective perturbations, they still lead to distortions in the feature space of VC models, which degrades the performance of VC. From this perspective, we propose a novel two-stage purification method: (1) Purify the perturbed speech; (2) Refine it using phoneme guidance to align it with the clean speech distribution. Experimental results demonstrate that our method outperforms state-of-the-art purification methods in disrupting VC defenses. Our study reveals the limitations of adversarial perturbation-based VC defenses and underscores the urgent need for more robust solutions to mitigate the security and privacy risks posed by VC. The code and audio samples are available at https://de-antifake.github.io.", "date": "2025-07-03", "categories": []}
{"id": "2601.12263", "title": "Multimodal Generative Engine Optimization: Rank Manipulation for Vision-Language Model Rankers", "abstract": "Vision-Language Models (VLMs) integrate visual and textual knowledge into unified representations that increasingly underpin modern retrieval and recommendation systems. However, it remains unclear how reliably these models utilize their cross-modal knowledge when ranking multimodal items, and whether their knowledge grounding can be subverted. In this paper, we expose a fundamental vulnerability in how VLMs apply multimodal knowledge for product ranking: through Multimodal Generative Engine Optimization (MGEO), we show that an adversary can manipulate a VLM's ranking decisions by jointly crafting imperceptible image perturbations and fluent textual suffixes that exploit the model's internal cross-modal knowledge coupling. Using an alternating optimization strategy, MGEO targets the deep interactions between visual and linguistic representations within the VLM, achieving rank manipulations that substantially exceed those of unimodal attacks and heuristic baselines powered by strong commercial models. Our findings reveal that surface-level content quality is insufficient for rank promotion; instead, direct alignment with the model's internal knowledge utilization mechanism is required. These results raise important questions on the faithfulness and robustness of knowledge grounding in multimodal foundation models, and motivate future work on defense mechanisms for multimodal retrieval systems. Code is available at: https://github.com/glad-lab/MGEO", "date": "2026-01-18", "categories": []}
{"id": "2602.05175", "title": "Enhancing Adversarial Robustness with Signed Distance Fields for Harmonizing Geometric Invariance and Texture", "abstract": "Deep neural networks demonstrate impressive performance in visual recognition but remain highly vulnerable to imperceptible adversarial attacks. Existing defense strategies such as adversarial training and diffusion-based purification have achieved significant progress but are frequently constrained by high computational cost, information loss, and inference latency. To address these challenges, we propose a Geometric and Texture balancing Purification (GeoTexPuri) framework that enhances adversarial robustness by harmonizing invariant geometric structures with textural features. Specifically, the framework integrates dense geometric guidance into the training phase by transforming discrete image masks into continuous spatial fields via Signed Distance Fields (SDF). This process establishes stable structural anchors that shield the model from local pixel noise. Through a multi-stream training objective, the model learns to internalize purified representations that effectively align semantic textural cues with these underlying geometric invariants. Extensive experiments on ImageNet demonstrate the efficacy of our approach. GeoTexPuri achieves 84.79\\% clean accuracy and 83.52\\% robust accuracy under the AutoAttack. Crucially, GeoTexPuri functions as a deterministic classifier during inference, requiring only the input image without any auxiliary geometric modules or additional computational costs, thereby ensuring a scalable and efficient solution for real-time applications.", "date": "2026-02-05", "categories": []}
{"id": "2604.25965", "title": "Adversarial Robustness of NTK Neural Networks", "abstract": "Deep learning models are widely deployed in safety-critical domains, but remain vulnerable to adversarial attacks. In this paper, we study the adversarial robustness of NTK neural networks in the context of nonparametric regression. We establish minimax optimal rates for adversarial regression in Sobolev spaces and then show that NTK neural networks, trained via gradient flow with early stopping, can achieve this optimal rate. However, in the overfitting regime, we prove that the minimum norm interpolant is vulnerable to adversarial perturbations.", "date": "2026-04-28", "categories": []}
{"id": "2606.07650", "title": "Detecting Aimbot Cheaters in MOGs", "abstract": "Multiplayer Online Games have become a multibillion dollar industry in the entertainment sector. However, the presence of cheaters undermines the experience of honest players and devalues the effort of game developers, as it directly affects player retention, competitive integrity, the legitimacy and trustworthiness of a game, and most importantly the overall revenue streams. Among various cheating techniques, visual aimbots represent an emerging threat. They use computer vision models to detect opponents from client screen captures rather than accessing game memory, making them completely undetectable by commercial kernel level anti cheat solutions. In this paper, we introduce PATCH, a novel proactive defense strategy that deploys adversarial patches as in game honeytokens to mitigate the presence of visual aimbot cheaters. Our approach centers on deliberately triggering the cheaters' object detection model, enabling either direct detection, or rendering the game unplayable for the cheater via patch flooding on their viewport. We evaluate our approach on various criteria; analyzing the effectiveness of different patch sizes, scalability of patches to different screen resolutions, efficacy against diverse visual aimbot cheat configurations and also explore various YOLO models to assess patch transferability. Evaluation on a custom Unreal Engine game demonstrates over 90 percent detection rate in white box scenarios for almost all patch sizes, and reaches 60 to 90 percent cross model transferability with larger patches. We further validate our approach on Fortnite, a commercial MOG, demonstrating real world applicability.", "date": "2026-06-02", "categories": []}
{"id": "2606.07716", "title": "SHIELD-IDS: Structurally Heterogeneous Ensemble with Integrated Layered Defense for Intrusion Detection Systems", "abstract": "Adversarial attacks pose a serious and growing threat to Machine Learning (ML)-based Intrusion Detection Systems (IDS), where imperceptible perturbations to network flow features can systematically mislead classifiers into accepting malicious traffic as benign. The IDS-Anta framework partially addresses this through Z-score normalization, Singular Value Decomposition (SVD), and Multi-Armed Bandit (MAB) classifier selection with Thompson Sampling, yet its classifier pool lacks sufficient structural diversity for robust adversarial resistance. This work introduces IDS-Anta++, which incorporates XGBoost and LightGBM gradient boosting models into the ensemble and wraps the extended pool in a three-layer black-box defense: Isolation Forest anomaly screening, median feature smoothing, and six-way majority voting. Experiments conducted on CIC-IDS-2017, CEC-CIC-IDS-2018, and CIC-DDoS-2019 under both Fast Gradient Sign Method (FGSM) and Zeroth Order Optimization (ZOO) attacks confirm detection accuracy above 99% on clean data, with measurable robustness gains under adversarial conditions relative to the baseline IDS-Anta configuration.", "date": "2026-06-05", "categories": []}
{"id": "2606.08467", "title": "The Confidence Trap: Calibration Attacks for Graph Neural Networks", "abstract": "While confidence calibration is essential for trustworthy decision-making in safety-critical applications, the robustness of calibrated GNNs to adversarial structural perturbations remains largely unexplored. However, studying calibration attacks on graphs presents unique technical challenges: (1) the discrete nature of graph structures complicates gradient-based optimization, (2) existing underconfidence objectives fail to drive predictions toward uniform distributions, and (3) GNNs are highly sensitive to edge perturbations, often causing unintended label changes that violate attack constraints. To address these challenges, we propose a \\textbf{Unified Graph Calibration Attack (UGCA)} framework designed for \\textbf{worst-case (white-box) analysis} of GNN calibration robustness. UGCA introduces a KL-divergence loss to encourage uniform predictive distributions, a reranking mechanism to reduce label flipping, a hybrid loss to recover labels when violations occur, and beam search to explore a broader adversarial search space. We further provide theoretical insights linking model generalization, dataset complexity, and calibration vulnerability, showing that models with higher accuracy or trained on datasets with more classes are more susceptible under this threat model. Extensive experiments demonstrate that UGCA substantially increases Expected Calibration Error while preserving classification accuracy. Our code is publicly available at https://github.com/CaptainCuong/Graph-Calibration-Attack.git.", "date": "2026-06-07", "categories": []}
{"id": "2606.08745", "title": "Stain-Aware Wavelet Regularization for Instant Adversarial Purification in Histopathology", "abstract": "Deep learning has become prevalent in computational pathology pipelines that support tasks such as cancer screening and digital pathology analysis. However, the susceptibility of neural networks to adversarial perturbations raises safety concerns for reliable deployment in clinical practice. In histopathological images, this challenge is exacerbated by the difficulty of distinguishing high-frequency adversarial noise from subtle and diagnostically relevant tissue structures. To address this issue, we propose Stain-Aware Wavelet Regularization (SAWR), an adversarial purification framework that leverages multi-level wavelet-domain regularization based on Haar transform to hierarchically disentangle adversarial perturbations from diagnostic structural information. This spectral constraint is further extended to individual histological channels, enabling stain-specific frequency regulation consistent with the biological properties of Hematoxylin and Eosin. When integrated into an instant purification framework, SAWR improves adversarial robustness by up to 10.69\\% over the baseline approach, while maintaining texture and spectral fidelity under adversarial perturbations.", "date": "2026-06-07", "categories": []}
{"id": "2606.08960", "title": "Hardening Agent Benchmarks with Adversarial Hacker-Fixer Loops", "abstract": "Agent benchmarks score submissions with outcome verifiers that are typically hand-written and brittle, leaving them open to reward hacking. We audit 1,968 tasks across five terminal-agent benchmarks and find 323 (16%) hackable by frontier models given only the task description. This corrupts both leaderboard rankings and RL training signal, yet the standard response is manual and reactive.\n  We introduce the hacker-fixer loop, a method for building exploit-resistant verifiers without per-task manual patching. The loop alternates three LLM agents: a hacker tries to pass the verifier without solving the task, a fixer patches the verifier to reject each discovered exploit, and a solver confirms the patched verifier still admits legitimate solutions. The loop iterates: each patch reshapes what the verifier rewards, surfacing the next exploit. We further add verifier access, and let patches transfer across tasks, to broaden the exploits the loop discovers.\n  On KernelBench, the loop drives the attack success rate from 62% to 0% on a held-out corpus of publicly reported exploits. We also find that weaker agents in the loop can defend against much stronger hackers: Gemini 3 Flash's loop drives the stronger Gemini 3.1 Pro and Claude Opus 4.7's attack success rate from 76% and 61% to 0% on KernelBench, and Gemini 3.1 Pro's from 39% to 17% on Terminal Bench across 77 tasks. We release Terminal Wrench (323 hackable environments, 3,632 hack trajectories) as a snapshot of the current attack surface, our patched verifiers, the exploits the loop discovered, and our implementation as a basis for future work.", "date": "2026-06-08", "categories": []}
{"id": "2606.09536", "title": "Adversarial Attack and Disturbance Detection by Hadamard-Coded Output Representations for Object Detection and Semantic Segmentation", "abstract": "Conventional one-hot encodings often yield poorly calibrated models, being overconfident under attack, and letting entropy-based detection algorithms fail. Previous image classification works have demonstrated that Hadamard-coded output representations can improve adversarial robustness. However, attempts to integrate Hadamard codes into semantic segmentation fall far behind state-of-the-art models in mean intersection-over-union performance. Regarding object detection, such output encodings have not yet been investigated at all. Further, no prior art addressed intrinsic codeword inconsistencies or actually exploited intrinsic codeword redundancy. Accordingly, we first derive a novel decoding procedure for Hadamard codewords towards optimal class-wise probabilities, solving the underlying optimization problem by using the projection onto the probability simplex. Second, our optimization delivers a measure of prediction inconsistency. Third, we are the first to show how to exploit these inconsistencies for adversarial attack and disturbance detection. Fourth, we introduce HadamardNet, a framework employing Hadamard codes as output representations for semantic segmentation and object detection models and tasks. We conduct a comprehensive evaluation both on disturbances and adversarial attacks, achieving state-of-the-art perturbation detection performance for both tasks in only a single detection pass, while delivering equivalent or close-by reference performance on clean data.", "date": "2026-06-08", "categories": []}
{"id": "2606.09700", "title": "What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks", "abstract": "Large language model (LLM)-powered content moderation systems have become a critical defense against harmful online content. However, these systems primarily operate on tokenized text and largely ignore the visual cues that humans naturally rely on when interpreting content. We show that this discrepancy creates a fundamental perceptual mismatch: content that is readily recognized as harmful by humans can become effectively invisible to automated moderation systems. To study this vulnerability, we introduce a class of Human-Perceptible Adversarial Attacks (HPAA), in which harmful expressions are embedded into otherwise benign text through visually salient typographic manipulations. Our key insight is that typographic features, including spacing, visual emphasis, and spatial arrangement, can be strategically combined to preserve human recognition of harmful content while substantially reducing machine detectability. Operating in black-box settings with only a small query budget, our attack automatically generates evasive content without requiring model access or gradient information. We evaluate the attack across multiple datasets and ten deployed moderation systems, including commercial APIs and state-of-the-art open-source guardrails. Results reveal a striking gap between human and machine perception: with only three detector queries, generated attacks achieve over 86\\% human recognition while maintaining detection rates below 1\\% across the evaluated systems. We further conduct ablation studies to identify the typographic factors driving successful evasion, analyze why current moderation architectures fail to capture these signals, and discuss practical defenses. Our findings expose a fundamental blind spot in today's LLM-based moderation ecosystem and highlight need for moderation systems that reason about content in a manner more consistent with human perceptual understanding.", "date": "2026-06-08", "categories": []}
{"id": "2505.14289", "title": "EVA: Evolving Semantic Adversaries for Red-Teaming GUI Agents Against Environmental Injection Attacks", "abstract": "Graphical User Interface (GUI) agents powered by Multimodal Large Language Models (MLLMs) are increasingly deployed yet vulnerable to Environmental Injection Attacks (EIAs).However, current red-teaming methods are hindered by prohibitive computational costs and limited adaptability. A fundamental question remains unaddressed: does the bottleneck of attack success lie in visual perception or semantic understanding? Through controlled experiments, we observe that semantic deception, rather than visual appearance, serves as the primary determinant of attack success. Based on this insight, we introduce EVA, an evolutionary framework that evolves adversarial payloads exclusively within the semantic dimension. EVA employs a discovery-deployment framework to mine linguistic vulnerability patterns and distill them into generalizable rules. Experimental results across five representative victim agents demonstrate that EVA achieves up to 85\\% attack success rate, evolving benign seeds into successful attacks within only 1.18 to 1.71 iterations. This rapid convergence uncovers a dense semantic attack space in the model's latent representation, unveiling a critical alignment paradox: the instruction-following capabilities reinforced by alignment training render agents inherently susceptible to authoritative, semantically deceptive environmental cues.", "date": "2025-05-20", "categories": []}
{"id": "2506.12454", "title": "On the existence of consistent adversarial attacks in high-dimensional linear classification", "abstract": "What fundamentally distinguishes an adversarial attack from a misclassification due to limited model expressivity or finite data? In this work, we investigate this question in the setting of high-dimensional binary classification, where statistical effects due to limited data availability play a central role. We introduce a new error metric that precisely capture this distinction, quantifying model vulnerability to consistent adversarial attacks -- perturbations that preserve the ground-truth labels. Our main technical contribution is an exact and rigorous asymptotic characterization of these metrics in both well-specified models and latent space models, revealing different vulnerability patterns compared to standard robust error measures. The theoretical results demonstrate that as models become more overparameterized, their vulnerability to label-preserving perturbations grows, offering theoretical insight into the mechanisms underlying model sensitivity to adversarial attacks.", "date": "2025-06-14", "categories": []}
{"id": "2510.09041", "title": "Robust Driving Control for Autonomous Vehicles: An Intelligent General-sum Constrained Adversarial Reinforcement Learning Approach", "abstract": "Deep reinforcement learning (DRL) has demonstrated remarkable success in developing autonomous driving policies. However, its vulnerability to adversarial attacks remains a critical barrier to real-world deployment. Although existing robust methods have achieved success, they still suffer from three key issues: (i) these methods are trained against myopic adversarial attacks, limiting their abilities to respond to more strategic threats, (ii) they have trouble causing truly safety-critical events (e.g., collisions), but instead often result in minor consequences, and (iii) these methods can introduce learning instability and policy drift during training due to the lack of robust constraints. To address these issues, we propose Intelligent General-sum Constrained Adversarial Reinforcement Learning (IGCARL), a novel robust autonomous driving approach that consists of a strategic targeted adversary and a robust driving agent. The strategic targeted adversary is designed to leverage the temporal decision-making capabilities of DRL to execute strategically coordinated multi-step attacks. In addition, it explicitly focuses on inducing safety-critical events by adopting a general-sum objective. The robust driving agent learns by interacting with the adversary to develop a robust autonomous driving policy against adversarial attacks. To ensure stable learning in adversarial environments and to mitigate policy drift caused by attacks, the agent is optimized under a constrained formulation. Extensive experiments show that IGCARL improves the success rate by at least 27.9% over state-of-the-art methods, demonstrating superior robustness to adversarial attacks and enhancing the safety and reliability of DRL-based autonomous driving.", "date": "2025-10-10", "categories": []}
{"id": "2512.12997", "title": "Calibrating Uncertainty for Zero-Shot Adversarial CLIP", "abstract": "CLIP delivers strong zero-shot classification but remains highly vulnerable to adversarial attacks. Prior adversarial fine-tuning work primarily matches predicted logits between clean and adversarial examples, which overlooks uncertainty calibration and may degrade the zero-shot generalization. A common expectation in reliable uncertainty estimation is that predictive uncertainty should increase as inputs become more difficult or shift away from the training distribution. However, we frequently observe the opposite in the adversarial setting: perturbations not only degrade accuracy but also suppress uncertainty, leading to severe miscalibration and over-confidence. This reveals a critical reliability gap beyond robustness. To bridge this gap, we propose an adversarial fine-tuning objective for CLIP considering both accuracy and uncertainty. By reparameterizing CLIP outputs as the concentration parameters of a Dirichlet distribution, we propose a unified representation that captures relative semantic structure and confidence magnitude. This enables holistic distribution alignment under perturbations, moving beyond single-logit anchoring and restoring calibrated uncertainty. Experiments across multiple zero-shot benchmarks demonstrate that our method significantly improves uncertainty calibration and achieves competitive adversarial robustness while preserving clean accuracy.", "date": "2025-12-15", "categories": []}
{"id": "2601.12359", "title": "Zero-Shot Embedding Drift Detection: A Lightweight Defense Against Prompt Injections in LLMs", "abstract": "Prompt injection attacks have become an increasing vulnerability for LLM applications, where adversarial prompts exploit indirect input channels such as emails or user-generated content to circumvent alignment safeguards and induce harmful or unintended outputs. Despite advances in alignment, even state-of-the-art LLMs remain broadly vulnerable to adversarial prompts, underscoring the urgent need for robust, productive, and generalizable detection mechanisms beyond inefficient, model-specific patches. In this work, we propose Zero-Shot Embedding Drift Detection (ZEDD), a lightweight, low-engineering-overhead framework that identifies both direct and indirect prompt injection attempts by quantifying semantic shifts in embedding space between benign and suspect inputs. ZEDD operates without requiring access to model internals, prior knowledge of attack types, or task-specific retraining, enabling efficient zero-shot deployment across diverse LLM architectures. Our method uses adversarial-clean prompt pairs and measures embedding drift via cosine similarity to capture subtle adversarial manipulations inherent to real-world injection attacks. To ensure robust evaluation, we assemble and re-annotate the comprehensive LLMail-Inject dataset spanning five injection categories derived from publicly available sources. Extensive experiments demonstrate that embedding drift is a robust and transferable signal, outperforming traditional methods in detection accuracy and operational efficiency. With greater than 93% accuracy in classifying prompt injections across model architectures like Llama 3, Qwen 2, and Mistral and a false positive rate of <3%, our approach offers a lightweight, scalable defense layer that integrates into existing LLM pipelines, addressing a critical gap in securing LLM-powered systems to withstand adaptive adversarial threats.", "date": "2026-01-18", "categories": []}
{"id": "2606.06938", "title": "When CLIP Sees More, It Fights Back Harder: Multi-View Guided Adaptive Counterattacks for Test-Time Adversarial Robustness", "abstract": "Vision-language models such as CLIP have achieved remarkable zero-shot recognition capabilities, yet their robustness against adversarial perturbations remains limited. Test-time counterattack (TTC) was recently proposed to improve CLIP's robustness by perturbing an input image to steer it away from a corrupted state during inference. However, TTC remains fragile under strong attacks because its counterattack relies on a directly corrupted original view and employs a noise-driven hard-gating scheme that cannot adapt to varying corruption severity. To address these limitations, we introduce Multi-view guided Adaptive Counterattack (MAC), which performs counterattacks for multi-view with corruption-aware soft weighting. Specifically, MAC first constructs augmented views of an input image to obtain diverse embeddings. It then performs counterattacks to refine corrupted embeddings of views. Next, MAC adaptively scales the counterattack intensity for each view based on its estimated corruption degree. Finally, the adaptively counterattacked views are aggregated to yield a robust final prediction. Extensive experiments across 20 datasets and diverse attack scenarios demonstrate that MAC substantially improves robustness while preserving high inference speed and memory efficiency with its tuning-free design. Our code is available at https://github.com/sunoh-kim/MAC.", "date": "2026-06-05", "categories": []}
{"id": "2606.06943", "title": "SS-TPT: Stability and Suitability-Guided Test-Time Prompt Tuning for Adversarially Robust Vision-Language Models", "abstract": "Vision-language models (VLMs) such as CLIP achieve strong zero-shot recognition but remain highly fragile under adversarial perturbations. Recent test-time adaptation defenses improve robustness by leveraging many augmented views, but this leads to impractical slowdown and a clear robustness-throughput trade-off. To address this challenge, we present Stability and Suitability-guided Test-time Prompt Tuning (SS-TPT), evaluating the quality of each augmented view via two complementary scores: (1) stability, measuring prediction invariance to weak augmentations, and (2) suitability, measuring feature-space density among views. These stability and suitability (SS) scores guide both adaptation and inference through an SS-guided consistency loss and an SS-weighted prediction, amplifying trustworthy views while suppressing corrupted ones. Extensive experiments demonstrate that SS-TPT significantly outperforms prior state-of-the-art methods, achieving superior robustness-throughput trade-offs across diverse datasets and varying numbers of views, thereby demonstrating both strong practicality and generality. Our code is available at https://github.com/sunoh-kim/SS-TPT.", "date": "2026-06-05", "categories": []}
{"id": "2503.01734", "title": "Adversarial Agents: Black-Box Evasion Attacks with Reinforcement Learning", "abstract": "Attacks on machine learning models have been extensively studied through stateless optimization. In this paper, we demonstrate how a reinforcement learning (RL) agent can learn a new class of attack algorithms that generate adversarial samples. Unlike traditional adversarial machine learning (AML) methods that craft adversarial samples independently, our RL-based approach retains and exploits past attack experience to improve the effectiveness and efficiency of future attacks. We formulate adversarial sample generation as a Markov Decision Process and evaluate RL's ability to (a) learn effective and efficient attack strategies and (b) compete with state-of-the-art AML. On two image classification benchmarks, our agent increases attack success rate by up to 13.2% and decreases the average number of victim model queries per attack by up to 16.9% from the start to the end of training. In a head-to-head comparison with state-of-the-art image attacks, our approach enables an adversary to generate adversarial samples with 17% more success on unseen inputs post-training. From a security perspective, this work demonstrates a powerful new attack vector that uses RL to train agents that attack ML models efficiently and at scale.", "date": "2025-03-03", "categories": []}
{"id": "2606.05609", "title": "SlotGCG: Exploiting the Positional Vulnerability in LLMs for Jailbreak Attacks", "abstract": "As large language models (LLMs) are widely deployed, identifying their vulnerability through jailbreak attacks becomes increasingly critical. Optimization-based attacks like Greedy Coordinate Gradient (GCG) have focused on inserting adversarial tokens to the end of prompts. However, GCG restricts adversarial tokens to a fixed insertion point (typically the prompt suffix), leaving the effect of inserting tokens at other positions unexplored. In this paper, we empirically investigate \\emph{slots}, i.e., candidate positions within a prompt where tokens can be inserted. We find that vulnerability to jailbreaking is highly related to the selection of the \\emph{slots}. Based on these findings, we introduce the \\textit{Vulnerable Slot Score} (VSS) to quantify the positional vulnerability to jailbreaking. We then propose SlotGCG, which evaluates all slots with VSS, selects the most vulnerable slots for insertion, and runs a targeted optimization attack at those slots. Our approach provides a position-search mechanism that is attack-agnostic and can be plugged into any optimization-based attack, adding only 200ms of preprocessing time. Experiments across multiple models demonstrate that SlotGCG significantly outperforms existing methods. Specifically, it achieves 14\\% higher Attack Success Rates (ASR) over GCG-based attacks, converges faster, and shows superior robustness against defense methods with 42\\% higher ASR than baseline approaches. Our implementation is available at \\href{https://github.com/youai058/SlotGCG}{https://github.com/youai058/SlotGCG}", "date": "2026-06-04", "categories": []}
{"id": "2606.05678", "title": "Beyond Waveform Robustness: Robust Feature-Vocoder Adversarial Attacks on Automatic Speech Recognition", "abstract": "Automatic speech recognition (ASR) systems have become widely used for multilingual speech-to-text transcription. Their robustness to adversarial attacks has become an important topic for the community. Existing adversarial attacks directly add adversarial noise to the speech audio. However, prior work has shown that existing adversarial attacks face two limitations: they often transfer poorly to black-box ASR systems and are increasingly mitigated by defenses tailored to input-space perturbations. In this work, we propose a Clean-Referenced Feature-Vocoder Attack, a surrogate-based black-box attack that moves the adversarial search space from raw waveforms to self-supervised learning (SSL) representations. To address the transferability limitation, we perturb more generalizable acoustic-phonetic representations rather than low-level waveform samples, reducing dependence on surrogate-specific waveform gradients and encouraging adversarial perturbations that generalize across ASR systems. To bypass different defenses, we shift the adversarial signal from explicit additive waveform noise to SSL feature-space perturbations and reconstruct them through a vocoder into speech-like waveform adversarial signals, making the resulting samples less aligned with waveform-bounded defenses. Extensive experiments show that, when optimized only on raw Whisper-small as a public surrogate model, our attack transfers effectively to black-box ASR models with a +26.6 WER improvement over the SOTA baseline, while also remaining effective against multiple training defenses with a +36.2 WER improvement. These results reveal a blind spot in current ASR robustness evaluation.", "date": "2026-06-04", "categories": []}
{"id": "2606.06186", "title": "Adversarial Attacks Already Tell the Answer: Directional Bias-Guided Test-time Defense for Vision-Language Models", "abstract": "Vision-Language Models (VLMs), such as CLIP, have shown strong zero-shot generalization but remain highly vulnerable to adversarial perturbations, posing serious risks in real-world applications. Test-time defenses for VLMs have recently emerged as a promising and efficient approach to defend against adversarial attacks without requiring costly large-scale retraining. In this work, we uncover a surprising phenomenon: under diverse input transformations, adversarial images in CLIP's feature space consistently shift along a dominant direction, in contrast to the dispersed patterns of clean images. We hypothesize that this dominant shift, termed the Defense Direction, opposes the adversarial shift, pointing features back toward their correct class centers. Building on this insight, we propose Directional Bias-guided Defense (DBD), a test-time framework that estimates the Defense Direction and employs a DB-score-based two-stream reconstruction strategy to recover robust representations. Experiments on 15 datasets demonstrate that DBD not only achieves SOTA adversarial robustness while preserving clean accuracy, but also reveals the counterintuitive result that adversarial accuracy can even surpass clean accuracy. This demonstrates that adversarial perturbations inherently encode directional priors about the true decision boundary.", "date": "2026-06-04", "categories": []}
{"id": "2503.10629", "title": "Hierarchical Self-Supervised Adversarial Training for Robust Vision Models in Histopathology", "abstract": "Adversarial attacks pose significant challenges for vision models in critical fields like healthcare, where reliability is essential. Although adversarial training has been well studied in natural images, its application to biomedical and microscopy data remains limited. Existing self-supervised adversarial training methods overlook the hierarchical structure of histopathology images, where patient-slide-patch relationships provide valuable discriminative signals. To address this, we propose Hierarchical Self-Supervised Adversarial Training (HSAT), which exploits these properties to craft adversarial examples using multi-level contrastive learning and integrate it into adversarial training for enhanced robustness. We evaluate HSAT on multiclass histopathology dataset OpenSRH and the results show that HSAT outperforms existing methods from both biomedical and natural image domains. HSAT enhances robustness, achieving an average gain of 54.31% in the white-box setting and reducing performance drops to 3-4% in the black-box setting, compared to 25-30% for the baseline. These results set a new benchmark for adversarial training in this domain, paving the way for more robust models. Our Code for training and evaluation is available at https://github.com/HashmatShadab/HSAT.", "date": "2025-03-13", "categories": []}
{"id": "2602.01658", "title": "Efficient Adversarial Attacks on High-dimensional Offline Bandits", "abstract": "Bandit algorithms have recently emerged as a powerful tool for evaluating machine learning models, including generative image models and large language models, by efficiently identifying top-performing candidates without exhaustive comparisons. These methods typically rely on a reward model, often distributed with public weights on platforms such as Hugging Face, to provide feedback to the bandit. While online evaluation is expensive and requires repeated trials, offline evaluation with logged data has become an attractive alternative. However, the adversarial robustness of offline bandit evaluation remains largely unexplored, particularly when an attacker perturbs the reward model (rather than the training data) prior to bandit training. In this work, we fill this gap by investigating, both theoretically and empirically, the vulnerability of offline bandit training to adversarial manipulations of the reward model. We introduce a novel threat model in which an attacker exploits offline data in high-dimensional settings to hijack the bandit's behavior. Starting with linear reward functions and extending to nonlinear models such as ReLU neural networks, we study attacks on two Hugging Face evaluators used for generative model assessment: one measuring aesthetic quality and the other assessing compositional alignment. Our results show that even small, imperceptible perturbations to the reward model's weights can drastically alter the bandit's behavior. From a theoretical perspective, we prove a striking high-dimensional effect: as input dimensionality increases, the perturbation norm required for a successful attack decreases, making modern applications such as image evaluation especially vulnerable. Extensive experiments confirm that naive random perturbations are ineffective, whereas carefully targeted perturbations achieve near-perfect attack success rates ...", "date": "2026-02-02", "categories": []}
{"id": "2605.07032", "title": "A Systematic Investigation of RL-Jailbreaking in LLMs", "abstract": "The evolution of generative models from next-token predictors to autonomous engines of complex systems necessitates rigorous safety hardening. Adversarial jailbreaking, the strategic manipulation of models to elicit harmful output, remains a primary threat to safe deployment. While Reinforcement Learning (RL) frames jailbreaking as a multi-step attack through sequential optimization, a mechanistic understanding of why the framework succeeds remains incomplete. To fill this gap, we present the first systematic decomposition of RL jailbreaking. We deconstruct the framework into problem formalization (reward function, action space, episode length), and algorithmic measures (RL algorithm, training data, reward-shaping) to identify the structural determinants of adversarial success. Our results reveal that the RL-jailbreaker successfully compromised all targeted models and safeguards. Through this first-of-its-kind analysis, we demonstrate that environment formalization, specifically dense rewards and extended episode lengths, is the primary driver of jailbreaking success. This work provides a tool for improving RL-jailbreaker efficiency and, ultimately, harden generative models resistant to RL-based attacks.", "date": "2026-05-07", "categories": []}
{"id": "2606.04168", "title": "When Autoregressive Consistency Hurts Safety Alignment", "abstract": "Safety alignment in large language models (LLMs) is fragile in part because it is often shallow: fine-tuning mainly reshapes the model's behavior near the first few output tokens. We argue that this phenomenon can be understood through autoregressive consistency, the tendency of next-token prediction to preserve and extend the current response trajectory consistently. By analyzing the learning dynamics of safety alignment, we show that autoregressive consistency can concentrate alignment updates on early tokens, offering a mechanistic explanation for shallow safety alignment. The same mechanism also predicts a broader class of attacks on LLMs: attacks that induce harmful continuation states at arbitrary positions in the output trajectory. As a concrete example, we introduce random insertion attack, which inserts a short harmful span into an otherwise safe refusal trajectory and exploits autoregressive consistency to sustain the resulting harmful branch, thereby bypassing safety alignment. Notably, a short harmful span can redirect the generation to be harmful even after a long refusal prefix, highlighting autoregressive consistency as a potential broader failure mechanism. This suggests that safety alignment should also break harmful autoregressive consistency throughout the output trajectory. We therefore propose adversarial safety alignment, an initial framework based on worst-case harmful continuation states, and instantiate it with random worst-insertion training. Overall, our results suggest that autoregressive consistency should be treated as a central consideration in both safety alignment and attack design.", "date": "2026-06-02", "categories": []}
{"id": "2606.04483", "title": "Off-Distribution Voices: Fanfiction Subgenres as Universal Vernacular Jailbreaks for Aligned LLMs", "abstract": "Existing jailbreaks against aligned LLMs are discrete artifacts whose surface forms are easy to fingerprint and patch. We argue that the real failure mode is not any specific prompt, but an entire register of natural human writing that safety training has under-covered. Building on this insight, we introduce the first jailbreak family that uses real fanfiction subgenres as universal attack carriers: a creative-writing meta is conditioned on passages from one of twelve Archive of Our Own (AO3) subgenres, and the harmful behavior is embedded as the climax of the resulting scene. The construction requires no attacker LLM and no per-target adaptation. On eight aligned LLMs over the union of HarmBench and JailbreakBench, this attack lifts mean ASR from 0.278 to 0.731 under a four-judge ensemble; a factorial decomposition shows the gain is carried by register rather than length or structure. Two active defences widen rather than narrow the vernacular-to-baseline ratio, indicating that template-targeting defences merely steer attackers toward register-based attacks like ours. We also propose SAGA-A4, a static four-turn extension that attains mean ASR 0.924, substantially exceeding three existing multi-turn methods.", "date": "2026-06-03", "categories": []}
{"id": "2606.04612", "title": "Hybrid Adversarial Defence for Natural Language Understanding Tasks", "abstract": "Large Language Models (LLMs) are vulnerable both to hallucination and adversarial manipulation. Although these problems are closely related, existing defences typically address them separately. We investigate a hybrid defence framework that combines entropy-based models, designed to reduce hallucinations, with uncertainty-based models and geometric-based models, designed to reduce vulnerability. Under in-domain tests on Natural Language Understanding datasets (FEVER, HotpotQA, CSQA, SIQA) we find our hybrid model improves both clean-task performance (up to 43.34\\% increase in accuracy) and adversarial robustness (up to 64.92\\% improvement in accuracy and 62.27\\% reduction in attack success rate). For out-of-distribution datasets (AeroEngQA, CPIQA) we see similar adversarial robustness from our hybrid model (up to 57.14\\% improvement in accuracy). For prompt injection (SafeGuard) and jailbreak detection (AdvBench, DAN) datasets our hybrid model is also very strong (up to 51\\% reduction in attack success rate compared to state of the art baseline models). Overall, our results show that combining entropy, uncertainty and geometric features provides a more effective defence strategy than using any single feature alone for both in-domain and out-of-distribution tasks.", "date": "2026-06-03", "categories": []}
{"id": "2606.04767", "title": "Measuring Model Robustness via Fisher Information: Spectral Bounds, Theoretical Guarantees, and Practical Algorithms", "abstract": "The robustness of deep neural networks is crucial for safety-critical deployments, yet existing evaluation methods are often attack-dependent and lack interpretability. We propose a principled, attack-agnostic robustness metric based on the spectral norm of the Fisher Information Matrix (FIM), which quantifies the worst-case sensitivity of the model's output distribution to input perturbations. Theoretically, we establish that the FIM equals the variance of the input Jacobian and derive closed-form spectral bounds for common architectures, including VGG, ResNet, DenseNet, and Transformer, providing the first theoretical robustness ranking. To enable scalable evaluation, we develop efficient algorithms, including power iteration and Hutchinson-based estimation, that support both white-box and black-box settings. Extensive experiments across multiple datasets, including CIFAR, ImageNet, and medical images, and across multiple architectures show a strong correlation between our metric and adversarial vulnerability. Our framework serves as an interpretable diagnostic tool that complements attack-based evaluations, offering insights into architectural sensitivity and guiding the design of more robust models. Code is available at: https://github.com/franz-chang/SRP/.", "date": "2026-06-03", "categories": []}
{"id": "2606.04929", "title": "Sequential Data Poisoning in LLM Post-Training", "abstract": "LLM post-training proceeds through multiple stages, e.g., supervised fine-tuning (SFT) followed by reinforcement learning from human feedback (RLHF) or direct preference optimization (DPO), where each stage draws data from different, potentially untrusted sources. Existing literature assumes data poisoning attacks may occur at each training stage, but neglects the possibility of multiple attackers. To study the trustworthiness of the entire post-training pipeline, we propose the threat model of sequential data poisoning, where multiple adversaries separately poison the SFT and preference datasets. Under this threat model, we identify the single-attacker illusion: each adversary, evaluated in isolation, appears to pose a negligible threat. Yet when adversaries collaborate across stages, the true vulnerability is revealed. In the SFT $\\to$ DPO pipeline, their contributions are additive: splitting a fixed poison budget across stages outperforms concentrating it in either stage alone. In the SFT $\\to$ PPO pipeline, their contributions are complementary: neither SFT nor reward model poisoning succeeds individually, yet their combination does. These findings show that security analyses of individual post-training stages systematically underestimate compound vulnerabilities that emerge only from their interaction. Code is available at https://github.com/jcksanderson/sequential-poisoning.", "date": "2026-06-03", "categories": []}
{"id": "2502.02260", "title": "Position: Adversarial ML for LLMs Is Not Making Any Progress", "abstract": "In the past decade, considerable research effort has been devoted to securing machine learning (ML) models that operate in adversarial settings. Yet, progress has been slow even for simple \"toy\" problems (e.g., robustness to small adversarial perturbations) and is often hindered by non-rigorous evaluations. Today, adversarial ML research has shifted towards studying larger, general-purpose language models. In this position paper, we argue that the situation is now even worse: in the era of LLMs, the field of adversarial ML studies problems that are (1) less clearly defined, (2) harder to solve, and (3) even more challenging to evaluate. As a result, we caution that yet another decade of work on adversarial ML may be failing to produce meaningful progress.", "date": "2025-02-04", "categories": []}
{"id": "2602.04899", "title": "Phantom Transfer: Data Poisoning can Survive Data-Level Defences", "abstract": "We present a data poisoning attack -- Phantom Transfer -- with the property that, even if you know precisely how the poison was placed into an otherwise benign dataset, you cannot filter it out. We achieve this by modifying subliminal learning to work in real-world contexts and demonstrate that the attack works regardless of which model produced the data, which model is trained on the data or what the attack target is. Furthermore, the attack survives 11 tested data-level defences, including one where every sample is paraphrased by another model. We characterise when this attack works best and show that it can be used to plant password-triggered behaviours into models while still beating defences.\n  In short, we provide an existence proof that maximum-affordance defences can fail to stop sophisticated data poisoning attacks. We suggest that future defences should be supplemented with white-box methods and post-training model audits.", "date": "2026-02-03", "categories": []}
{"id": "2603.23117", "title": "TRAP: Hijacking VLA CoT-Reasoning via Adversarial Patches", "abstract": "By integrating Chain-of-Thought (CoT) reasoning, Vision-Language-Action (VLA) models have demonstrated strong capabilities in robotic manipulation, particularly by improving generalization and interpretability. However, the security of CoT-based reasoning mechanisms remains largely unexplored. In this paper, we show that CoT reasoning introduces a novel attack vector for targeted behavior hijacking--for example, causing a robot to mistakenly deliver a knife to a person instead of an apple--without modifying the user's instruction. We first provide empirical evidence that CoT strongly governs action generation, even when it is semantically misaligned with the input instructions. Building on this observation, we propose TRAP, the first targeted behavior-hijacking adversarial attack against CoT-reasoning VLA models. By targeting the reasoning-to-action pathway, TRAP uses an adversarial patch (e.g., a tablecloth placed on the table) to steer intermediate CoT reasoning and downstream actions toward adversary-defined behaviors. Extensive evaluations on three representative reasoning VLAs, spanning distinct CoT reasoning mechanisms, demonstrate the effectiveness of TRAP. Notably, we implemented the patch by printing it on paper in a real-world setting. Our findings highlight the urgent need to secure CoT reasoning in VLA systems. The project page is available at https://zhengxian-huang.github.io/TRAP-website/.", "date": "2026-03-24", "categories": []}
{"id": "2605.19262", "title": "Backdooring Masked Diffusion Language Models", "abstract": "Masked diffusion language models (MDLMs) are emerging as a compelling new paradigm for text generation, but their training-time security remains largely unexplored. Existing backdoor attacks on Gaussian diffusion models or autoregressive language models do not directly apply to MDLMs because MDLMs rely on discrete state corruption and iterative denoising rather than continuous noising or left-to-right prediction. In this work, we present the first systematic study of training-time backdoor attacks on MDLMs. We propose SHADOWMASK, a backdoor attack that modifies the MDLM forward corruption process by replacing the standard all-mask terminal distribution with a trigger-mask mixture prior. This creates a dedicated denoising pathway from trigger-corrupted states to attacker-specified targets while preserving clean denoising behavior. We further provide a principled mathematical formulation by defining the backdoored forward process, deriving the reverse-time posterior, and obtaining the continuous-time training objective. Evaluations on DiT-based MDLM and LLaDA-8B-Instruct across WikiText-103, OpenWebText, and Alpaca show that SHADOWMASK achieves near-100% attack success, substantially outperforms standard data poisoning, largely preserves clean utility, remains effective under full-model and parameter-efficient fine-tuning, and is robust against representative defenses.", "date": "2026-05-19", "categories": []}
{"id": "2606.02597", "title": "Making Brain-Computer Interfaces More Secure", "abstract": "The development of brain-computer interfaces (BCIs) based on electroencephalograms (EEGs) has advanced significantly mainly to machine learning. Although the majority of earlier research has been on increasing classification accuracy, relatively little focus has been placed on security and robustness. According to recent research, EEG-based BCIs are susceptible to adversarial attacks, which can cause misdiagnosis due to minute, well-crafted disturbances. Evaluating model robustness against such perturbations is therefore critical for ensuring reliable deployment. In this study, we propose a lightweight custom Convolutional Neural Network (CNN) architecture to investigate adversarial robustness in EEG-based BCIs. The suggested method is assessed using two EEG datasets and contrasted with three novel CNN models tailored to EEG, namely EEGNet, DeepConvNet, and SleepEEGNet, under gradient-based adversarial attack scenarios. According to experimental findings, the suggested model continuously performs better in classification under adversarial perturbations compared to baseline models, indicating improved robustness. These findings highlight the potential of lightweight architectures for enhancing the reliability of EEG-based BCI systems under adversarial conditions.", "date": "2026-05-22", "categories": []}
{"id": "2606.02630", "title": "MultiTurnPSB: Evaluating Multi-Turn Jailbreak Attacks an dClassifier-Based Defenses for Medical AI Safety", "abstract": "Patient-facing medical chatbots are commonly evaluated on single-turn prompts, yet real users push back after refusals, add urgency, and invoke authority. We introduce MultiTurnPSB, a four-turn adversarial extension of PatientSafetyBench, and evaluate GPT-4.1-mini under fixed template, template-adaptive, and live adversarial attacks. Unsafe responses rise from 35% to nearly 80% by Turn 4 under live attack. Under the same adversary, GPT-4.1-mini and Claude Sonnet 4.5 are statistically indistinguishable at baseline but diverge to a 19x gap by Turn 4, a difference invisible to single-turn evaluation. We characterize four degradation trajectory signatures and identify a two-element attack formula responsible for most catastrophic failures. A lightweight input-side classifier reduces Turn 4 unsafe responses by 52 percentage points despite severe accuracy degradation, but the 45% false alarm rate on benign queries is the primary deployment constraint. A methodological finding also emerges: Claude Sonnet refused to generate adversarial messages in over half of late-turn conversations despite explicit red team framing, suggesting safety training may generalize to the attacker role.", "date": "2026-05-30", "categories": []}
{"id": "2606.02924", "title": "ATLAS: A Large-Scale Evaluation Benchmark for Adversarial LiDAR Perception", "abstract": "Autonomous driving perception is typically evaluated on clean benchmark data, yet real-world deployment requires robustness to rare, structured, and potentially adversarial sensor anomalies. This gap is especially critical for LiDAR, where external actors can physically manipulate the sensing process to induce black-box perception failures without accessing the model. Existing LiDAR benchmarks provide little visibility into this failure mode. Prior adversarial LiDAR studies have largely centered on attack hardware, geometric and algorithmic defenses, and early-generation detectors, leaving the robustness of modern perception systems unexplored. To address this evaluation gap, we introduce ATLAS (Adversarial Temporal LiDAR Attack Suite), the first large-scale, physically grounded evaluation benchmark for LiDAR perception models under black-box sensor attacks, simulating the two primary attack modes -- point injection and point removal -- across real driving sequences. Evaluating a broad cross-section of current state-of-the-art LiDAR perception models, ATLAS reveals a surprising robustness asymmetry: models with stronger performance on standard benchmarks tend to better withstand removal attacks, yet are actually more vulnerable to injection attacks than weaker models. We trace this vulnerability to standard object database sampling augmentations, revealing how current training practices can induce architecture-agnostic robustness failures, and study initial directions for mitigating both attack modes. We release the ATLAS generation code to support extensible, reproducible evaluations as attack capabilities evolve, helping make black-box sensor robustness an explicit consideration in future LiDAR perception development.", "date": "2026-06-01", "categories": []}
{"id": "2606.03136", "title": "PsychoPass: Geometric Profiling of Multi-Turn Adversarial LLM Conversations", "abstract": "Multi-turn jailbreak attacks on large language models (LLMs) reveal a mismatch in current guardrails: they operate on individual turns, while attacks unfold as trajectories across conversations. We propose a shift from content to dynamics, modeling conversations as paths in representation space and asking whether adversarial intent is encoded early in their geometry. We introduce PsychoPass, a framework that extracts geometric features from conversation trajectories in embedding space to predict a potential attack before harmful content is produced. These features achieve near-perfect performance in na\\\"ive classifiers, which is largely explained by the inclusion of number of turns as a feature. After removing this confound, a smaller but consistent geometric signal remains, with classification performance that does not depend meaningfully on encoder choice. Crucially, this signal appears early in the conversation: attack outcomes remain above chance from short prefixes alone, more reliably than baseline guardrails. A supporting theoretical analysis explains these findings via a decomposition of length and shape, a detection bound based on prefix length, and encoder invariance. Together, these results show that adversarial conversations leave an early, representation-robust geometric fingerprint suitable for online monitoring.", "date": "2026-06-02", "categories": []}
{"id": "2606.03521", "title": "Post-Hoc Robustness for Model-Based Reinforcement Learning", "abstract": "To improve the real-world applicability of reinforcement learning (RL), the field of adversarially robust RL studies how to train agents under adversarial environment perturbations. In this setting, a protagonist agent optimizes a policy under environmental perturbations from an adversary, resulting in a zero-sum Markov game. When adversarially robust RL is combined with model-based RL, the adversary can target a learned transition model instead of the training environment. Extending this idea, this work introduces post-hoc robustification of deep RL agents at inference time. By using the learned model in combination with a trained nominal policy, our approach performs a robust policy improvement step. The goal is to improve robustness without any additional training of neural networks. Specifically, we utilize model-predictive control under adversarial rollouts, which are approximated via projected gradient descent within a bounded uncertainty set. Furthermore, these offline rollouts are performed while considering and mitigating out-of-distribution issues. The proposed methodology is validated by demonstrating significant improvements in robustness when the algorithm is evaluated in perturbed Gymnasium MuJoCo environments, while considering the computational limitations of the post-hoc inference setting.", "date": "2026-06-02", "categories": []}
{"id": "2606.03647", "title": "Black-box, Adaptive, Efficient, Transferable, Harmful, Applicable... Attacks Are All You Need to Break LLMs", "abstract": "Accurately evaluating adversarial robustness is a longstanding challenge. A flawed attack design can inflate robustness estimates, making deployment risk assessment and defense comparison unreliable. Historically, standardized attacks such as AutoAttack have largely resolved this for image classifiers, providing a reliable evaluation baseline for systematic comparison across defenses. However, no equivalent exists for LLM jailbreak evaluation yet, where designing such an attack is considerably more difficult. A reliable attack must, among other things, be black-box compatible, applicable to arbitrary defense pipelines, and efficient, which no existing method jointly satisfies. We introduce Indirect Harm Optimization (IHO), a masked diffusion language model attacker trained via iterative preference optimization against a harmfulness judge, requiring only black-box access to the target. The same method can be used without modification as a strong adaptive attack on individual behaviors, or as an efficient amortized policy that transfers to held-out behaviors and unseen target models without fine-tuning. Even against layered defenses, such as a Circuit Breaker-trained model combined with an auxiliary detector, IHO improves attack success considerably over state-of-the-art approaches, without any defense-specific adaptation. Our results position IHO as a practical step toward the kind of standardized jailbreak evaluation that has improved reliability in the past. Code and models are available on GitHub and Hugging Face.", "date": "2026-06-02", "categories": []}
{"id": "2606.03793", "title": "Exploring Adversarial Robustness and Safety Alignment in Multilingual Multi-Modal Large Language Models", "abstract": "Multimodal Large Language Models integrate visual perception into language reasoning, introducing a continuous attack surface susceptible to adversarial attacks. Prior work on MLLM robustness has focused largely on English-centric tasks, leaving multilingual behaviour unexplored. We address this gap through a systematic study of adversarial robustness and multimodal safety across 12 diverse languages, evaluating open-source MLLMs that acquire multilingual capability through instruction tuning. Gradient-based attacks reveal a transferable multilingual vulnerability: adversarial images optimized in one language continue to induce failure in others, demonstrating strong cross-lingual transferability. Multilingual safety further varies with how effectively a model retrieves or interprets harmful instructions. When harmful intent is issued through text, languages with stronger linguistic grounding more often elicit misuse-enabling responses, while weaker languages produce fewer unsafe outputs. When embedded in the image as typographic content, English scripts are reliably recognised and followed, whereas non-English scripts are rarely parsed by the vision encoder. Lower-resource languages may therefore appear safer, but this is an artefact of comprehension and visual-grounding failures rather than genuine alignment, a phenomenon we term safety-by-failure. In contrast, MLLMs that build multilingual capability throughout their training stages rather than only at instruction tuning, such as Qwen3-VL, exhibit genuine cross-lingual safety, maintaining active refusal across languages rather than masking comprehension failure. Shallow multilingual adaptation, such as fine-tuning on translated instruction data, may produce surface-level understanding that creates illusory safety in low-resource languages; deeper integration across training stages leads to genuine multilingual safety alignment.", "date": "2026-06-02", "categories": []}
{"id": "2606.03925", "title": "Adaptive Causal Alignment for High-Confidence Adversarial Training", "abstract": "Inverse adversarial training leverages high-confidence predictions to stabilize robust learning, yet we uncover a critical paradox: high confidence often stems from overfitting to non-causal background correlations rather than intrinsic object semantics. Our investigation reveals that visual context functions as a dual-natured signal, serving as either a necessary supportive prior or a spurious confounder. This insight renders existing blind suppression strategies flawed, as they inevitably lead to severe Feature Loss. To resolve this, we propose High-Confidence Causally Aligned Training (HICAT), a unified framework that establishes a Semantic Equilibrium. Operating on a ``Measure-Debias-Align'' pipeline, HICAT integrates a Learnable Background-Bias Estimator (LBBE) to adaptively diagnose context utility. Guided by this diagnosis, an Adaptive Debiasing mechanism performs surgical logit rectification, complemented by a geometrically grounded Foreground Logit Orthogonal Enhancement (FLOE) loss to enforce rigorous feature disentanglement. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet-1K demonstrate that HICAT consistently improves over matched baselines across diverse architectures (CNNs and ViTs) while significantly reducing the robust generalization gap.", "date": "2026-06-02", "categories": []}
{"id": "2211.15223", "title": "Gamma-convergence of a nonlocal perimeter arising in adversarial machine learning", "abstract": "In this paper we prove Gamma-convergence of a nonlocal perimeter of Minkowski type to a local anisotropic perimeter. The nonlocal model describes the regularizing effect of adversarial training in binary classifications. The energy essentially depends on the interaction between two distributions modelling likelihoods for the associated classes. We overcome typical strict regularity assumptions for the distributions by only assuming that they have bounded $BV$ densities. In the natural topology coming from compactness, we prove Gamma-convergence to a weighted perimeter with weight determined by an anisotropic function of the two densities. Despite being local, this sharp interface limit reflects classification stability with respect to adversarial perturbations. We further apply our results to deduce Gamma-convergence of the associated total variations, to study the asymptotics of adversarial training, and to prove Gamma-convergence of graph discretizations for the nonlocal perimeter.", "date": "2022-11-28", "categories": []}
{"id": "2404.01356", "title": "Perturbation Effects on Accuracy and Fairness among Similar Individuals", "abstract": "Deep neural networks are vulnerable to adversarial perturbations that can simultaneously degrade prediction robustness and individual fairness across diverse application settings. However, existing evaluation protocols typically assess these dimensions in isolation, thereby obscuring critical failure modes. To bridge this gap, we formalize Robust Individual Fairness (RIF): under semantic-preserving (truth-condition-preserving) perturbations, predictions should remain both correct with respect to the ground truth and invariant across semantically equivalent individuals. To surface RIF violations in practice, we introduce RIFair, a black-box adversarial framework that leverages a decoupled perturbation strategy to construct semantically preserved yet unrobust and/or unfair instance pairs. Experiments across multiple model architectures and real-world textual datasets show that robustness-only or fairness-only metrics often miss Robust Biased and Unrobust Fair behaviors. RIFair}reliably exposes these hidden vulnerabilities, supporting RIF as a necessary criterion for trustworthy model assessment. The experimental code is publicly available at https://github.com/Xuran-LI/RIFair.", "date": "2024-04-01", "categories": []}
{"id": "2404.13621", "title": "Attack on Scene Flow using Point Clouds", "abstract": "Deep neural networks have made significant advancements in accurately estimating scene flow using point clouds, which is vital for many applications like video analysis, action recognition, and navigation. The robustness of these techniques, however, remains a concern, particularly in the face of adversarial attacks that have been proven to deceive state-of-the-art deep neural networks in many domains. Surprisingly, the robustness of scene flow networks against such attacks has not been thoroughly investigated. To address this problem, the proposed approach aims to bridge this gap by introducing adversarial white-box attacks specifically tailored for scene flow networks. Experimental results show that the generated adversarial examples obtain up to 33.7 relative degradation in average end-point error on the KITTI and FlyingThings3D datasets. The study also reveals the significant impact that attacks targeting point clouds in only one dimension or color channel have on average end-point error. Analyzing the success and failure of these attacks on the scene flow networks and their 2D optical flow network variants shows a higher vulnerability for the optical flow networks. Code is available at https://github.com/aheldis/Attack-on-Scene-Flow-using-Point-Clouds.git.", "date": "2024-04-21", "categories": []}
{"id": "2506.22666", "title": "VERA: Variational Inference Framework for Jailbreaking Large Language Models", "abstract": "The rise of API-only access to state-of-the-art LLMs highlights the need for effective black-box jailbreak methods to identify model vulnerabilities in real-world settings. Without a principled objective for gradient-based optimization, most existing approaches rely on genetic algorithms, which are limited by their initialization and dependence on manually curated prompt pools. Furthermore, these methods require individual optimization for each prompt, failing to provide a comprehensive characterization of model vulnerabilities. To address this gap, we introduce VERA: Variational infErence fRamework for jAilbreaking. VERA casts black-box jailbreak prompting as a variational inference problem, training a small attacker LLM to approximate the target LLM's posterior over adversarial prompts. Once trained, the attacker can generate diverse, fluent jailbreak prompts for a target query without re-optimization. Experimental results show that VERA achieves strong performance across a range of target LLMs, highlighting the value of probabilistic inference for adversarial prompt generation.", "date": "2025-06-27", "categories": []}
{"id": "2510.09260", "title": "GREAT: Generalizable Backdoor Attacks in RLHF via Emotion-Aware Trigger Synthesis", "abstract": "Recent work has shown that RLHF is highly susceptible to backdoor attacks. However, existing methods often rely on rare tokens or fixed triggers, limiting their impact in realistic scenarios. In this work, we develop GREAT, a novel framework for crafting natural distributional backdoors in RLHF. Specifically, GREAT targets harmful response generation for a vulnerable user subpopulation featured by semantically violent requests paired with emotionally angry triggers. At the core of our framework is a trigger identification pipeline that operates in the model's latent embedding space, leveraging dimensionality reduction and clustering techniques to identify representative triggers. To enable this, we introduce a hierarchical and diversity-driven prompting strategy to construct Erinyes, a high-quality dataset of over 5,000 angry triggers curated from GPT-4.1. Our experiments show that GREAT significantly outperforms baselines in attack generalization to unseen triggers, while preserving standard utility and maintaining stealth under defenses.", "date": "2025-10-10", "categories": []}
{"id": "2510.09288", "title": "A unifying Bayesian framework for adversarial robustness", "abstract": "The vulnerability of machine learning models to adversarial attacks remains a critical societal security challenge. Traditional defenses, such as adversarial training, typically robustify models by minimizing a worst-case loss. These deterministic approaches do not account for uncertainty in the adversary's attack. While stochastic defenses placing a probability distribution on the adversary exist, they often lack statistical rigor and fail to make explicit their underlying assumptions. To resolve these issues, we introduce a formal Bayesian framework that models adversarial uncertainty through a stochastic channel, articulating all probabilistic assumptions. This yields two robustification strategies: a proactive defense enacted during training, aligned with adversarial training, and a reactive defense enacted during operations, aligned with adversarial purification. Several state-of-the-art defenses can be recovered as limiting cases of our model. We empirically validate our methodology, showcasing the benefits of explicitly modeling adversarial uncertainty.", "date": "2025-10-10", "categories": []}
{"id": "2510.14025", "title": "NAPPure: Adversarial Purification for Robust Image Classification under Non-Additive Perturbations", "abstract": "Adversarial purification has achieved great success in combating adversarial image perturbations, which are usually assumed to be additive. However, non-additive adversarial perturbations such as blur, occlusion, and distortion are also common in the real world. Under such perturbations, existing adversarial purification methods are much less effective since they are designed to fit the additive nature. In this paper, we propose an extended adversarial purification framework named NAPPure, which can further handle non-additive perturbations. Specifically, we first establish the generation process of an adversarial image, and then disentangle the underlying clean image and perturbation parameters through likelihood maximization. Experiments on GTSRB and CIFAR-10 datasets show that NAPPure significantly boosts the robustness of image classification models against non-additive perturbations.", "date": "2025-10-15", "categories": []}
{"id": "2510.16660", "title": "Universal and Transferable Attacks on Pathology Foundation Models", "abstract": "We introduce Universal and Transferable Adversarial Perturbations (UTAP) for pathology foundation models that reveal critical vulnerabilities in their capabilities. Optimized using deep learning, UTAP comprises a fixed and weak noise pattern that, when added to a pathology image, systematically disrupts the feature representation capabilities of multiple pathology foundation models. Therefore, UTAP induces performance drops in downstream tasks that utilize foundation models, including misclassification across a wide range of unseen data distributions. In addition to compromising the model performance, we demonstrate two key features of UTAP: (1) universality: its perturbation can be applied across diverse field-of-views independent of the dataset that UTAP was developed on, and (2) transferability: its perturbation can successfully degrade the performance of various external, black-box pathology foundation models - never seen before. These two features indicate that UTAP is not a dedicated attack associated with a specific foundation model or image dataset, but rather constitutes a broad threat to various emerging pathology foundation models and their applications. We systematically evaluated UTAP across various state-of-the-art pathology foundation models on multiple datasets, causing a significant drop in their performance with visually imperceptible modifications to the input images using a fixed noise pattern. The development of these potent attacks establishes a critical, high-standard benchmark for model robustness evaluation, highlighting a need for advancing defense mechanisms and potentially providing the necessary assets for adversarial training to ensure the safe and reliable deployment of AI in pathology.", "date": "2025-10-18", "categories": []}
{"id": "2510.27249", "title": "C-LEAD: Contrastive Learning for Enhanced Adversarial Defense", "abstract": "Deep neural networks (DNNs) have achieved remarkable success in computer vision tasks such as image classification, segmentation, and object detection. However, they are vulnerable to adversarial attacks, which can cause incorrect predictions with small perturbations in input images. Addressing this issue is crucial for deploying robust deep-learning systems. This paper presents a novel approach that utilizes contrastive learning for adversarial defense, a previously unexplored area. Our method leverages the contrastive loss function to enhance the robustness of classification models by training them with both clean and adversarially perturbed images. By optimizing the model's parameters alongside the perturbations, our approach enables the network to learn robust representations that are less susceptible to adversarial attacks. Experimental results show significant improvements in the model's robustness against various types of adversarial perturbations. This suggests that contrastive loss helps extract more informative and resilient features, contributing to the field of adversarial robustness in deep learning. The code is publicly made available on GitHub in the following link: https://github.com/suklav/C_Lead .", "date": "2025-10-31", "categories": []}
{"id": "2511.12487", "title": "ToxSearch: Evolving Prompts for Toxicity Search in Large Language Models", "abstract": "Large Language Models remain vulnerable to adversarial prompts that elicit toxic content even after safety alignment. We present ToxSearch, a black-box evolutionary framework that tests model safety by evolving prompts in a synchronous steady-state loop. The system employs a diverse set of operators, including lexical substitutions, negation, back-translation, paraphrasing, and two semantic crossover operators, while a moderation oracle provides fitness guidance. Operator-level analysis shows heterogeneous behavior: lexical substitutions offer the best yield-variance trade-off, semantic-similarity crossover acts as a precise low-throughput inserter, and global rewrites exhibit high variance with elevated refusal costs. Using elite prompts evolved on LLaMA 3.1 8B, we observe practically meaningful but attenuated cross-model transfer, with toxicity roughly halving on most targets, smaller LLaMA 3.2 variants showing the strongest resistance, and some cross-architecture models retaining higher toxicity. These results suggest that small, controllable perturbations are effective vehicles for systematic red-teaming and that defenses should anticipate cross-model reuse of adversarial prompts rather than focusing only on single-model hardening.", "date": "2025-11-16", "categories": []}
{"id": "2602.02763", "title": "Exposing Vulnerabilities in Explanation for Time Series Classifiers via Dual-Target Attacks", "abstract": "Interpretable time series deep learning systems are often assessed by checking temporal consistency on explanations, implicitly treating this as evidence of robustness. We show that this assumption can fail: Predictions and explanations can be adversarially decoupled, enabling targeted misclassification while the explanation remains plausible and consistent with a chosen reference rationale. We propose TSEF (Time Series Explanation Fooler), a dual-target attack that jointly manipulates the classifier and explainer outputs. In contrast to single-objective misclassification attacks that disrupt explanation and spread attribution mass broadly, TSEF achieves targeted prediction changes while keeping explanations consistent with the reference. Across multiple datasets and explainer backbones, our results consistently reveal that explanation stability is a misleading proxy for decision robustness and motivate coupling-aware robustness evaluations for trustworthy time series tasks.", "date": "2026-02-02", "categories": []}
{"id": "2603.24511", "title": "Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs", "abstract": "We show that AI agents are capable of discovering novel algorithms for adversarial attacks against LLMs, advancing the state of the art on white-box jailbreaking and prompt injection evaluations. We deploy frontier agents, such as Claude Code and Codex, in an autoresearch loop with access to a library of 30+ prior methods and an evaluation script with a fixed compute budget. We show this pipeline to be effective in jailbreaking OpenAI's GPT-OSS-Safeguard-20B and in prompt injections against Meta-SecAlign-70B, an adversarially robust model. For GPT-OSS-Safeguard, the best agent-discovered method achieves up to 80\\% attack success rate on CBRN queries, compared to <50\\% for existing methods. For SecAlign, it achieves 100\\% ASR, while the best prior automated methods only achieve 82\\%. Notably, in our setting, attack methods are developed on unrelated surrogate models for a pure random-target token-forcing task, yet generalize directly to prompt injection on the adversarially trained model. Finally, we trace the lineage of methods developed during autoresearch, characterizing the agents' strategies and failure modes. Adversarial ML has long held that defenses must be evaluated against attacks tailored to them; autoresearch automates this principle, and we argue it should be the minimum bar for defense evaluation going forward.", "date": "2026-03-25", "categories": []}
{"id": "2604.07967", "title": "AtomEval: Validity-Aware Atomic Evaluation of Adversarial Claim Rewriting in Fact Verification", "abstract": "Large language models (LLMs) can rewrite refuted claims to evade evidence-based fact verifiers, but conventional attack success rate (ASR) can be inflated when rewrites change, weaken, or correct the false proposition they are supposed to preserve. We introduce AtomEval, a validity-aware evaluation protocol for fixed-evidence adversarial claim rewriting. AtomEval represents claims as subject--relation--object--modifier (SROM) atoms, applies a one-way preservation gate to separate valid verifier evasion from proposition-changing rewrites, and reports validity-aware attack success rate (VASR), which counts only verifier-evasive rewrites that preserve the original false proposition. AtomEval further provides fine-grained diagnostics that explain both proposition-level failures and non-minimal valid rewrites. On FEVER refuted-claim rewriting, AtomEval exposes and explains ASR inflation: many apparent attacks fool the verifier by altering, weakening, or correcting the proposition they should preserve. By making attacked-proposition preservation explicit and measurable, AtomEval provides a stable evaluation target for evaluating adversarial rewriters that must balance verifier evasion with proposition preservation.", "date": "2026-04-09", "categories": []}
{"id": "2605.12813", "title": "REALISTA: Realistic Latent Adversarial Attacks that Elicit LLM Hallucinations", "abstract": "Large language models (LLMs) achieve strong performance across many tasks but remain vulnerable to hallucinations, making it important to systematically evaluate their reliability under realistic adversarial inputs. We formulate hallucination elicitation as a constrained optimization problem, where the goal is to find semantically coherent adversarial prompts that are equivalent to benign user prompts. Existing attack methods remain limited: discrete prompt-based attacks preserve semantic equivalence and coherence but search only over a limited set of prompt variations, while continuous latent-space attacks explore a richer space but often decode into prompts that are no longer valid rephrasings. To address these limitations, we propose REALISTA, a realistic latent-space attack framework. REALISTA constructs an input-dependent dictionary of valid editing directions, each corresponding to a semantically equivalent and coherent rephrasing, and optimizes continuous combinations of these directions in latent space. This design combines the optimization flexibility of continuous attacks with the semantic realism of discrete rephrasing-based attacks. Experiments demonstrate that REALISTA achieves superior or comparable performance to state-of-the-art realistic attacks on open-source LLMs and, crucially, succeeds in attacking large reasoning models under free-form response settings, where prior realistic attacks fail. Code is available at https://github.com/Buyun-Liang/REALISTA.", "date": "2026-05-12", "categories": []}
{"id": "2606.00159", "title": "Digital-to-Physical Transfer of Adversarial Patches for Aerial Vehicle Detection", "abstract": "Deep neural network (DNN)-based object detectors are widely used for analyzing aerial and satellite imagery in applications such as environmental monitoring and urban analytics. Despite their strong performance, these models are known to be vulnerable to adversarial examples, and physical adversarial attacks using printable patterns pose realistic security threats. In this paper, we evaluate physical adversarial patch attacks against an aerial vehicle detector by bridging digital optimization and real-world deployment. Adversarial patches are optimized in the digital domain using a loss function that minimizes the maximum objectness score while incorporating non-printability score (NPS) and total variation (TV) constraints to ensure both printability and spatial smoothness. The optimized patches are printed and deployed in three configurations: ON, OFF, and OFF-Side. Experiments using a YOLOv3 detector show that while the OFF patch achieves the highest effectiveness in the digital domain (85.51% Average Objectness Reduction Rate (AORR)), the ON patch demonstrates superior robustness in physical environments (0.197-0.343 Objectness Score Ratio (OSR)) due to its consistent visibility. Furthermore, our results indicate that weather-based augmentation does not necessarily improve patch optimization in this domain. These findings provide critical insights into the practical vulnerabilities of aerial object detection systems.", "date": "2026-05-29", "categories": []}
{"id": "2606.00392", "title": "Detector-Evasive LLM Paraphrasing via Constrained Policy Optimization", "abstract": "AI-text detectors are vulnerable to paraphrasing and detector-guided paraphrasing attacks, but existing detector-evasion methods often lack precise control over semantic preservation. In particular, optimizing directly for detector evasion can degrade fine-grained semantics, whereas scalarized reward designs provide only indirect, weight-sensitive control over the evasion-semantics trade-off. We address this limitation by formulating detector-evasive LLM paraphrasing as a Constrained Markov Decision Process, where detector evasion is the primary objective and semantic preservation is enforced as an explicit constraint. We propose Detector Evasion Policy Optimization (DEPO), a Lagrangian primal-dual reinforcement learning algorithm with a novel GRPO-style group-based policy update. DEPO adaptively balances semantic preservation and detector evasion during training, enabling the policy to improve attack success within a prescribed semantic-preservation region. Experiments on MAGE, M4, RAID, and peer-review datasets, evaluated against MAGE, RoBERTa, RADAR, Binoculars, and Fast-DetectGPT detectors, show that DEPO achieves strong detector evasion while precisely satisfying the semantic preservation constraint. DEPO also exhibits cross-domain, cross-detector, and prompt-level robustness.", "date": "2026-05-29", "categories": []}
{"id": "2606.00738", "title": "SORA: Free Second-Order Attacks in Fast Adversarial Training", "abstract": "Adversarial Training (AT) is a leading defense against adversarial examples but often suffers from Catastrophic Overfitting (CO) in efficient single-step variants, where robustness to multi-step attacks collapses despite high single-step performance. We address this failure mode with two contributions. First, we formalize Epsilon Overfitting (EO), a perspective in which fixed perturbation magnitudes and directions exacerbate CO, and show that introducing perturbation variability significantly improves robust generalization across different architectures and datasets. Second, we propose PertAlign (Perturbation Alignment), a theoretically grounded, computationally negligible metric that predicts CO onset by measuring gradient alignment across attack stages. Leveraging these insights, we introduce SORA, an adaptive step-size AT method that dynamically adjusts perturbations based on loss surface geometry. SORA consistently prevents CO, achieves state-of-the-art robustness and clean accuracy, and generalizes across datasets and architectures using a single fixed set of hyperparameters, which is essential for applicability in fast AT. Extensive experiments on diverse datasets and architectures show that SORA matches or surpasses the robustness of prior methods while delivering higher clean accuracy and superior efficiency. Code is available at https://github.com/SecondOrderAT/SORA.", "date": "2026-05-30", "categories": []}
{"id": "2606.00801", "title": "Quality-Diversity Evolution for Discovering Diverse Vulnerabilities in LLM Safety", "abstract": "Current approaches to LLM adversarial testing suffer from coverage gaps: manual red-teaming does not scale, LLM-as-attacker methods exhibit mode collapse, and gradient-based approaches produce uninterpretable gibberish. We introduce a quality-diversity evolutionary framework that operates at the semantic level, evolving interpretable attack strategies rather than token sequences. Using MAP-Elites, we maintain a diverse archive of attacks across behavioral dimensions (strategy type, encoding method, prompt length). In experiments across GPT-4o-mini, Claude 3.5 Sonnet, Gemini 2.0 Flash, and an open-weight coding model (Devstral-small-2), we discover distinct vulnerability profiles: GPT-4o-mini is vulnerable to hypothetical and multi-turn framing combined with ROT13 encoding (fitness 0.8), Gemini to direct attacks with ROT13 and multi-turn with Leetspeak (0.8), while Claude shows uniformly ambiguous responses across all strategies (max 0.4). The semantic representation produces interpretable attacks that reveal systematic, model-specific weaknesses, providing actionable insights for improving LLM safety and a reproducible baseline for evaluating future frontier models. Code and experiment artifacts are released at https://github.com/bassrehab/red-queen.", "date": "2026-05-30", "categories": []}
{"id": "2606.00813", "title": "Cross-Generational Transfer of Adversarial Attacks Reveals Non-Monotonic Safety Alignment in LLMs", "abstract": "Safety alignment in LLMs does not improve monotonically across model generations. Studying four generations of Google's Gemma family (7B-31B) with quality-diversity evolution (MAP-Elites) as an automated red-teaming probe, we find that Gemma 3 (12B) exhibits 68.7% +/- 5.7% attack success rate (ASR; mean +/- std, 3 seeds), significantly higher than its predecessor Gemma 2 (45.5% +/- 7.2%; p = 0.030, paired bootstrap) and its successor Gemma 4 (33.9% +/- 1.8%). Replaying evolved attack archives across generations reveals that attacks from other generations transfer to Gemma 3 at 44-46% but only 14-18% to Gemma 4, indicating that Gemma 4's safety gains generalize beyond the attack distributions evolved against earlier generations. Under our 8B judge, copyright and cybercrime vulnerabilities register at near-100% across all generations, though a second-judge audit (Section 6) suggests the copyright result is sensitive to judge choice. Misinformation ASR jumps from 29% to 99% between Gemma 2 and Gemma 3 and remains elevated at 77% in Gemma 4, indicating the regression was not fully addressed. These patterns are invisible to static benchmarks and emerge only through adaptive, longitudinal probing. All experiments use 3 random seeds with a unified self-hosted judge; code and artifacts are available at https://github.com/bassrehab/red-queen.", "date": "2026-05-30", "categories": []}
{"id": "2606.00914", "title": "Adversarial Feeds Steer LLM Agent Decisions Against Their Defaults", "abstract": "LLM agents increasingly act after consuming ranked external information streams such as social feeds, search results, retrieval contexts, and email queues, yet safety evaluations almost always test the model or the user prompt in isolation, never the upstream ranker that decides what the agent reads just before it acts. We introduce a controlled protocol that holds the model, persona, topic, and final decision prompt fixed and varies only the composition and ordering of the posts an agent encounters during a preceding ten-turn \"scrolling\" phase, isolating the causal effect of feed curation on a downstream decision. Across 2,785 decision rollouts on four modern open instruct LLMs from three independent labs, we identify three response regimes: adversarial capitulation, default saturation, and a default-direction asymmetry in which a one-sided feed tips a decision the model was genuinely uncertain about (in the clearest cases from 5% to 100%; Fisher p as low as 3 x 10^-10) but cannot dislodge one it already favors or holds firmly. The effect follows a dose-response curve, survives a generator swap that rules out a writing-style artifact, generalizes across several decision domains including security-relevant choices such as removing a deployment approval gate or relaxing access controls, and is partly mitigated by two simple feed-level defenses; a frontier model retains its default. We characterize the recommender as a practical, default-bounded control surface for LLM agents, and argue that agent evaluations must audit the feed layer rather than the final prompt alone.", "date": "2026-05-30", "categories": []}
{"id": "2606.01437", "title": "CEAR: Certified Ensemble Adversarial Robustness in DNNs", "abstract": "Deep Neural Networks (DNNs) are highly susceptible to adversarial perturbations, leading to extensive research on robustness for safety-critical applications. State-of-the-art empirical defense mechanisms improve the robustness of DNNs through the training phase, but still struggle against adaptive white-box attacks. On the other hand, certified defenses offer provable guarantees of robustness within a specified perturbation bound. These guarantees hold regardless of the level of perturbations, even if the attacker is given full knowledge of the model. In this paper, we propose CEAR, an ensemble-based robust method that utilizes a hybrid of empirical and certified defense mechanisms. CEAR trains each network within the ensemble using varying Gaussian noise and temperatures to obfuscate gradients and logits, making the model more resistant to stronger gradient-based attacks. We then use noisy logits and propose two different voting mechanisms to further improve robustness. Furthermore, we extend randomized smoothing to verify the robustness of ensemble-based classifiers. Our experimental evaluations on MNIST, CIFAR10, and TinyImageNet datasets demonstrate superior certified accuracy on average, increased robustness radius, and decreased transferability compared to baseline methods.", "date": "2026-05-31", "categories": []}
{"id": "2606.01560", "title": "GJDNet: Robust Graph Neural Networks via Joint Disentangled Learning Against Adversarial Attacks", "abstract": "Graph Neural Networks (GNNs) are vulnerable to adversarial attacks, which inherently invert connectivity patterns by introducing disassortative edges in assortative graphs and assortative edges in disassortative graphs. This structural inversion creates structure-feature mismatches that disrupt neighborhood aggregation across different graph types. However, we find that existing defenses are limited, as they either treat neighborhoods as monolithic under fixed assortativity assumptions or rely on standard softmax classifiers that fail to account for perturbation-induced representation shifts. To further exploit this observation, we adopt a robustness perspective that jointly disentangles node representations and decision spaces, isolating perturbation effects while enforcing well-separated decision regions. Based on this principle, we propose Graph Joint Disentanglement Network (GJDNet), a unified framework for robust node classification across diverse graph assortativity regimes. GJDNet enhances robustness at both representation and decision levels: it employs feature-driven soft structural disentanglement with skewness-aware neighbor filtering to suppress perturbation-induced structure-feature mismatches, and introduces a Spherical Decision Boundary (SDB) to promote intra-class compactness and inter-class separation in the embedding space, thereby stabilizing decision boundaries under perturbations. Theoretical analysis provides insights into the effectiveness of the proposed disentangled representation and decision mechanisms, while extensive experiments demonstrate that GJDNet consistently achieves strong robustness across graphs with different connectivity regimes.", "date": "2026-06-01", "categories": []}
{"id": "2606.01738", "title": "THRD: A Training-Free Multi-Turn Defense Framework for Jailbreak Attacks on Large Language Models", "abstract": "Multi-turn jailbreak attacks pose a growing threat to LLMs by exploiting conversational dynamics such as gradual escalation and cross-turn coordination. Existing defenses either rely on costly retraining -- often degrading model utility -- or apply single-turn analysis independently at each turn, failing to capture how risk accumulates along interaction trajectories. We observe that safety behavior in multi-turn interaction is trajectory-dependent: dialogue history continuously reshapes the model's conditioning context, making it insufficient to evaluate each turn in isolation. Motivated by this insight, we present THRD, the first training-free framework that explicitly models temporal risk accumulation for multi-turn jailbreak defense. THRD integrates four modules: a Turn-level Risk Assessor (TRA) for instantaneous risk estimation, a Historical Context Analyzer (HCA) for cross-turn intent escalation detection, a Response Evaluator (RE) for identifying facilitative outputs, and a Decision Module that combines these signals through a time-evolving scoring mechanism with attenuation-based modulation and trend-aware adjustment. Experiments against state-of-the-art multi-turn attacks -- including tree-search-based and multi-agent collaborative methods -- across two target models show that THRD reduces ASR to 0.2--4.0% while preserving model utility within 1.5% degradation on MMLU and GSM8K. Ablation studies confirm non-redundant module contributions and stable cross-architecture generalization. Analysis of first rejection triggers reveals that over 70% of multi-turn attacks require Turn~2 or later to detect, validating the necessity of explicit temporal aggregation.", "date": "2026-06-01", "categories": []}
{"id": "2606.01746", "title": "Sensitivity as a Double-Edged Sword: A Trade-off Between Discriminability and Adversarial Robustness", "abstract": "Modern neural networks are highly susceptible to adversarial perturbations. In this work, we identify that part of this vulnerability stems from the sensitivity of the widely used fully connected (FC) classifiers to such perturbations. In contrast, simple $\\ell_2$ distance-based classifiers exhibit significantly greater robustness. We provide thorough theoretical and empirical analysis showing that while FC classifiers' high sensitivity makes them discriminative, it also makes them vulnerable. Conversely, $\\ell_2$-classifiers' insensitivity grants robustness but limits performance. Motivated by this trade-off, we propose a novel $\\ell_2$-reclassifier based on a Hybrid Prototype Mixing (HPM) framework. This method retains the discriminative power of FC classifiers while leveraging the robustness of $\\ell_2$ distance. It yields $\\ell_2$-distance-based predictions by fusing two prototype types: (1) stable, dataset-level prototypes updated via EMA, and (2) dynamic, batch-level prototypes generated from the FC classifier's predictions using a Straight-Through Estimator (STE). However, this dynamic, STE-based architecture introduces significant challenges for evaluation, such as gradient obfuscation and forward discontinuity. To address this, we propose a new, rigorous evaluation protocol, the Mixed Surrogate Attack (MSA), which uses multiple surrogates along with powerful AutoAttack to ensure a fair and robust assessment. Extensive experiments demonstrate that our lightweight, plug-and-play module, with minimal fine-tuning, effectively enhances the adversarial robustness of various existing SOTA adversarially trained models.", "date": "2026-06-01", "categories": []}
{"id": "2606.01892", "title": "Adversarial Attacks on Robot Localization Systems via Deep Feature Perturbation", "abstract": "Robot localization systems are critical for autonomous navigation and safety. Adversarial perturbations can mislead these systems, resulting in mislocalization, navigation errors, or unsafe interactions, especially in mission-critical scenarios. This paper investigates the vulnerability of deep learning based localization pipelines to adversarial attacks. We propose a novel framework for generating adversarial queries that specifically target Product Quantization (PQ) in visual localization systems. Our method employs a Lightweight Product Quantization Network (LPQN) to perturb query feature encodings, misleading the retrieval process by returning semantically irrelevant database entries. Adversarial queries are generated via a two-phase procedure: a forward pass that perturbs feature distributions and a backward pass that refines the perturbation through optimization. The lightweight design of LPQN allows the creation of subtle yet highly effective perturbations with minimal computational overhead. Extensive experiments in both controlled and real-world robotic environments demonstrate that our approach substantially degrades PQN performance, exposing critical vulnerabilities in practical applications.", "date": "2026-06-01", "categories": []}
{"id": "2606.02134", "title": "Rethinking Evaluation Paradigms in IBP-based Certified Training", "abstract": "Deep neural networks achieve strong performance on many supervised learning tasks but remain vulnerable to adversarial perturbations. Neural network verification provides mathematically rigorous robustness guarantees, yet at substantial computational cost. To mitigate this, certified training techniques optimise for verifiable robustness during training, typically inducing a trade-off between natural and certified accuracy controlled by method-specific hyperparameters. Because these metrics are inherently conflicting, the common practice of reporting a single configuration is problematic: it can mislead conclusions about overall performance and prevents unbiased assessments of the state of the art. We address this by evaluating certified training methods via Pareto front comparisons over the natural--certified accuracy trade-off. To enable fair, method-agnostic comparisons, we perform efficient automated multi-objective hyperparameter optimisation to identify a set of Pareto-optimal configurations for each method. This approach often uncovers substantial undertuning in previously reported configurations, yielding superior performance and establishing a new state of the art. Leveraging these fronts, we present the first comprehensive multi-objective comparison of certified training approaches, showing that prior advancements are less pronounced than assumed and revealing previously unreported performance complementarities.", "date": "2026-06-01", "categories": []}
{"id": "2606.02267", "title": "A combination of noise and bilateral filters achieve supralinear and scalable adversarial robustness in CNNs", "abstract": "The vulnerability of deep neural networks to adversarial examples poses a significant challenge for real-world deployment. Existing techniques to enhance deep network robustness rely on adversarial training, an approach that is powerful but computationally intensive and typically tailored to specific attack types. To address these limitations, existing works have explored techniques such as adding gaussian noise or filtering images, both of which can boost the network robustness to various adversarial attacks, albeit modestly. Here, we theoretically demonstrate that these two approaches enhance robustness against adversarial attacks through complementary mechanisms, resulting in supralinear robustness when combined. Building on this insight, we experimentally show that a simple preprocessor combining Gaussian noise and bilateral filtering yields supralinear improvements in adversarial robustness with minimal computational cost. Next, we combine our preprocessor with adversarial training and test on RobustBench to assess its supralinear improvement over state-of-the-art defenses. First, this combination ranks second on AutoAttack and third overall, while using only $\\sim$35% of the training FLOPs, using a model with $\\sim$50% less parametets, trained with $\\sim$33% of the epochs and $\\sim$15% the data compared to state-of-the-art defenses. Second, our method scales efficiently, matching the accuracy of competing models with roughly 2-8x less total compute across 3 orders of magnitude. Overall, our approach provides a principled and easily integrable framework for enhancing adversarial robustness, offering negligible computational overhead and a simple yet theoretically grounded design.", "date": "2026-06-01", "categories": []}
{"id": "2506.08255", "title": "SHIELD: Secure Hypernetworks for Incremental Expansion Learning Defense", "abstract": "Continual learning under adversarial conditions remains an open problem, as existing methods often compromise either robustness, scalability, or both. We propose a novel framework that integrates Interval Bound Propagation (IBP) with a hypernetwork-based architecture to enable certifiably robust continual learning across sequential tasks. Our method, SHIELD, generates task-specific model parameters via a shared hypernetwork conditioned solely on compact task embeddings, eliminating the need for replay buffers or full model copies and enabling efficient over time. To further enhance robustness, we introduce Interval MixUp, a novel training strategy that blends virtual examples represented as $\\ell_{\\infty}$ balls centered around MixUp points. Leveraging interval arithmetic, this technique guarantees certified robustness while mitigating the wrapping effect, resulting in smoother decision boundaries. We evaluate SHIELD under strong white-box adversarial attacks, including PGD and AutoAttack, across multiple benchmarks. It consistently outperforms existing robust continual learning methods, achieving state-of-the-art average accuracy while maintaining both scalability and certification. These results represent a significant step toward practical and theoretically grounded continual learning in adversarial settings.", "date": "2025-06-09", "categories": []}
{"id": "2511.17380", "title": "Non-Parametric Probabilistic Robustness: A Conservative Risk Estimator under Unknown Perturbation Distributions", "abstract": "Deep learning (DL) models, despite their remarkable success, remain vulnerable to small input perturbations that can cause erroneous outputs, motivating the recent proposal of probabilistic robustness (PR) as a complementary alternative to adversarial robustness (AR). However, existing PR formulations assume a fixed and known perturbation distribution, an unrealistic expectation in practice. To address this limitation, we propose non-parametric probabilistic robustness (NPPR), a more practical PR metric that does not rely on any predefined perturbation distribution. Following the non-parametric paradigm in statistical modeling, NPPR learns an optimized perturbation distribution directly from data, enabling conservative PR evaluation under distributional uncertainty. We further develop an NPPR estimator based on a Gaussian Mixture Model (GMM), covering various input-dependent and input-independent perturbation scenarios. Theoretical analyses establish the relationships among AR, PR, and NPPR. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny ImageNet across ResNet18/50, WideResNet50 and VGG16 validate NPPR as a more practical robustness metric, showing conservative (lower) PR estimates compared to assuming those common perturbation distributions used in state-of-the-arts.", "date": "2025-11-21", "categories": []}
{"id": "2605.24535", "title": "Steering Beyond the Support: Adversarial Training on Unsupervised Jailbroken Activation Simulation", "abstract": "Jailbreak prompts can trigger harmful completions on aligned LLMs, In accordance, safety steering has been proposed: test-time activation interventions that steer jailbreak activations to trigger refusal while preserving benign utility. However, existing steering methods are fundamentally supervised and tied to a static, limited training set, whereas real jailbreaks evolve and are often out-of-distributed from the training set, leading to failures on unseen attacks. In this paper, we tackle the failure on unseen jailbreaks problem, base on unsupervised latent direction discovery. We propose a bi-level adversarial training framework for zero-shot jailbreak defense. In the inner step, we simulate diverse jail-broken activations by extrapolating from refusal-state harmful-request activations via unsupervised latent direction discovery, which expands the coverage of real jailbreak activation subspaces. In the outer step, we train a potential-induced steering field to push these adversarial jailbroken states into refusal regions while keeping benign unchanged. Across three LLMs and six classical jailbreak families, our method achieves strong defense with attack success rates mostly below 5%, and rising subspace coverage throughout training helps explain the improved generalization.", "date": "2026-05-23", "categories": []}
{"id": "2605.26929", "title": "When Muon Optimizer Meets Adversarial Training: A Theoretical and Empirical Study", "abstract": "Adversarial training (AT) remains one of the most reliable empirical defenses against adversarial attacks. Its robustness critically depends on how the underlying min-max objective is optimized. In practice, Stochastic Gradient Descent (SGD) optimizer remains the default optimization choice for AT, whereas adaptive optimizers often improve standard training but may yield inferior robustness. Recently, the Muon optimizer, which orthogonalizes matrix-valued updates via an approximate polar decomposition, has achieved notable success in large-scale training at a memory cost comparable to SGD. This raises a security-relevant question: \\textit{can orthogonalized optimization improve AT under strong and heterogeneous threat models?} Focusing on this problem, we conduct a comprehensive theoretical and empirical study. Theoretically, we show that Muon imposes a spectral-norm stability ceiling on matrix updates, limiting uncontrolled spectral growth in the training dynamics without explicitly shrinking the learned weights. Empirically, across five architectures and three $\\ell_p$ threat models ($\\ell_\\infty$, $\\ell_1$, $\\ell_2$) and their union, Muon is competitive with SGD on CNNs and substantially outperforms AdamW on both CNNs and ViTs. These results identify optimizer geometry as a security-relevant factor in adversarial training, while clarifying the empirical regimes in which orthogonalized updates are beneficial. Overall, our findings highlight optimizer design as a security-critical component of AT.", "date": "2026-05-26", "categories": []}
{"id": "2605.30365", "title": "Mental Damage: Caption Poisoning Attacks on Retrieval-Augmented Text-to-Music Generation", "abstract": "Retrieval-augmented text-to-music (TTM) systems augment underspecified user prompts using captions retrieved from a music caption dataset. This design introduces an integrity dependency on the music knowledge database. We show that an attacker can poison the database by injecting a small number of crafted music captions, causing the system to retrieve malicious captions that bias prompt augmentation and steer generation away from the user's intended function, without modifying the user prompt, retriever, or generator. To achieve the music caption poisoning attack, we propose a dual-layer caption poisoning strategy that preserves high-level retrieval anchors while injecting low-level acoustic descriptors to steer prompt augmentation and downstream music generation toward an attacker-chosen target intent. In a MusicCaps knowledge database, CLAP retriever, and MusicGen pipeline, poisoned generations move substantially closer to the attacker's target, while remaining comparably aligned with the original user query. These results expose a practical integrity risk for retrieval-augmented creative AI systems. Our demo can be found at: https://yizhu-wen.github.io/Mental-Damage/", "date": "2026-05-18", "categories": []}
{"id": "2605.30578", "title": "AdvScene: Rethinking Adversarial Patch Evaluation Through Scene Robustness", "abstract": "Adversarial patches are physical patterns attached to real objects to mislead AI vision systems. Their real-world risk is not determined by a single successful prediction, but by whether they remain effective after deployment under changing viewpoints, distances, and scene conditions. We refer to this property as scene robustness, the effectiveness of a deployed patch across conditions in a real environment. Yet existing evaluations do not measure scene robustness well: real image benchmarks are realistic but fixed, while simulators are controllable but not grounded in a specific real scene.\n  We present AdvScene, a scene-grounded framework for measuring the scene robustness of adversarial patches in reconstructed real environments. AdvScene reframes evaluation as operational measurement: given a fixed deployed patch, it characterizes the patch's operational envelope - where and when the attack succeeds - as a function of viewpoint, distance, and scene context. A key challenge is that the attack is typically defined only in a single anchor view, while evaluation requires a representation that remains faithful under viewpoint changes. We formalize this as a constrained lifting problem and introduce Adversarial Patch-to-Scene Embedding (APSE), which resolves cross-view ambiguity while preserving attack-critical appearance and enforcing locality, target-surface attachment, and cross-view consistency. We validate AdvScene using real-world physical data and conduct a comprehensive evaluation of existing adversarial patches. Our results show that AdvScene reveals substantial scene-dependent variation in attack effectiveness that is not captured by existing image-centric or simulator-based evaluations.", "date": "2026-05-28", "categories": []}
{"id": "2605.30601", "title": "TASER: Task-Aware Stein Regularisation for Geometry-Driven Robustness", "abstract": "Modern deep networks remain fragile under distribution shift and adversarial perturbations, often due to excessive or poorly structured input sensitivity. We introduce TASER (Task-Aware Stein Regularisation), a training-time regularisation framework derived from Langevin Stein operators. By penalising pointwise Stein residuals under the training distribution, TASER encourages geometric compatibility between predictors and data density, inducing anisotropic, data-aware smoothness. We provide theoretical links between Stein regularisation and reduced first-order shift sensitivity, develop scalable implementation variants compatible with modern architectures, and demonstrate improved robustness and stability across regression and vision benchmarks. Across CIFAR-10 experiments, TASER consistently improves the adversarial robustness of established training methods without incurring statistically significant clean-accuracy degradation.", "date": "2026-05-28", "categories": []}
{"id": "2605.30667", "title": "Automatically Attacking Software Reverse Engineering AI Agents", "abstract": "Software tools for reverse engineering executable binary files, such as Ghidra, enable malware analysts to safely conduct robust static analysis without having access to original source code. Coupled with the analytic power of large language models (LLM), agentic systems enabled with tools, such as GhidraMCP, can allow analysts to automate a previously human driven process. Although this automation can increase the productivity of a single malware analyst, it also introduces a new area of vulnerability for malware obfuscation. This paper presents an adversarial technique using genetic algorithm-based prompt generation, a modification of an adversarial attack known as AutoDAN, to demonstrate the ability to deceive LLM-powered disassembly and decompilation systems into misinterpreting binary executables, effectively corrupting their analytical output. This proof-of-concept methodology exploits inherent vulnerabilities in how LLMs process and interpret decompiled machine code via prompt injection by using extraneous string variable assignments to pass surreptitious instructions to the LLM while not impacting the functionality of the executable file. We demonstrate this capability through several concise examples. This approach could enable attackers to bypass automated detection systems that rely on LLM-driven analysis pipelines. By studying and understanding this attack, insights can be gained regarding the security implication of integrating LLMs into cybersecurity toolchains and building more robust agentic code analysis systems.", "date": "2026-05-28", "categories": []}
{"id": "2605.31140", "title": "EvoDefense: Co-Evolving Black-Box Defense with Large Language Models", "abstract": "Large Language Models (LLMs) remain highly vulnerable to diverse attacks, particularly in black-box settings where the internals of target models are inaccessible. Existing black-box defenses typically rely on pre-defined filtering heuristics, which often fail to generalize to unseen attack types and target model architectures. We introduce EvoDefense, an experience-guided co-evolving black-box defense paradigm. EvoDefense employs a guard LLM to detect malicious queries and an experience memory module to accumulate defense knowledge from previous interactions. At the core of EvoDefense is a continuous attack-defense evolution loop, where an attack generator and the guard model iteratively refine their attack strategies and defense policies through experience-guided optimization. This design enables EvoDefense to generalize across unseen attacks and target models without retraining. Experiments on HarmBench, AdvBench, and AlpacaEval show that EvoDefense achieves consistently strong defense performance across seven popular models and five representative LLM attacks, while preserving competitive general capabilities. On HarmBench, EvoDefense reduces the attack success rate (ASR) of AutoDAN-turbo on Gemini-3-flash and LLaMA-3-8B-Instruct from 29.4% and 43.4% to 8.4% and 6.2%, respectively.", "date": "2026-05-29", "categories": []}
{"id": "2605.31246", "title": "BadBone: Backdoor Attacks Against Backbone Models in Visual Prompt Learning", "abstract": "Prompt learning is a new machine learning paradigm that has attracted ample attention due to its simplicity and proven efficacy. Despite its growing adoption, the security vulnerabilities associated with this paradigm remain underexplored. In this work, we take the first step to propose BadBone, a stealthy and adaptive backdoor attack against prompt learning using bi-level optimization. Instead of backdooring the prompt learning process, we aim to compromise a backbone model such that only target downstream tasks employing prompt learning inherit the backdoor vulnerability. Extensive experiments on three different models and three datasets from various domains show that our targeted/untargeted backdoored models achieve high attack performance while maintaining utility on both pre-training and downstream tasks. Moreover, we evaluate our approach against six state-of-the-art model-level defenses, including Neural Cleanse, ABS, MNTD, NAD, CLP, and D-BR. The results demonstrate that these defenses are largely ineffective against our backdoored models and thus leave the effective defense as an important direction for future work.", "date": "2026-05-29", "categories": []}
{"id": "2403.09441", "title": "An Empirical Study of the Influence of Adversarial Fine-Tuning on Compressed Neural Networks", "abstract": "As deep learning (DL) models are increasingly being integrated into our everyday lives, ensuring their safety by making them robust against adversarial attacks has become increasingly critical. DL models have been found to be susceptible to adversarial attacks by introducing small, targeted perturbations to disrupt the input data. Adversarial training has been presented as a mitigation strategy that can result in more robust models. This adversarial robustness comes with additional computational costs required to design adversarial attacks during training. The two objectives -- adversarial robustness and computational efficiency -- then appear to be in conflict with each other. In this work, we explore the effects of neural network compression on adversarial robustness. We specifically explore the effects of fine-tuning on compressed models, and present the trade-off between standard fine-tuning and adversarial fine-tuning. Our results show that adversarial fine-tuning of compressed models can yield large improvements to their robustness performance. We present experiments on several benchmark datasets showing that adversarial fine-tuning of compressed models can achieve robustness performance comparable to adversarially trained models, while also improving computational efficiency. Source code is available here: https://github.com/saintslab/Adver-Fine.", "date": "2024-03-14", "categories": []}
{"id": "2408.15451", "title": "Certified Causal Defense with Generalizable Robustness", "abstract": "While machine learning models have proven effective across various scenarios, it is widely acknowledged that many models are vulnerable to adversarial attacks. Recently, there have emerged numerous efforts in adversarial defense. Among them, certified defense is well known for its theoretical guarantees against arbitrary adversarial perturbations on input within a certain range (e.g., $l_2$ ball). However, most existing works in this line struggle to generalize their certified robustness in other data domains with distribution shifts. This issue is rooted in the difficulty of eliminating the negative impact of spurious correlations on robustness in different domains. To address this problem, in this work, we propose a novel certified defense framework GLEAN, which incorporates a causal perspective into the generalization problem in certified defense. More specifically, our framework integrates a certifiable causal factor learning component to disentangle the causal relations and spurious correlations between input and label, and thereby exclude the negative effect of spurious correlations on defense. On top of that, we design a causally certified defense strategy to handle adversarial attacks on latent causal factors. In this way, our framework is not only robust against malicious noises on data in the training distribution but also can generalize its robustness across domains with distribution shifts. Extensive experiments on benchmark datasets validate the superiority of our framework in certified robustness generalization in different data domains. Code is available in the supplementary materials.", "date": "2024-08-28", "categories": []}
{"id": "2504.12747", "title": "Privacy Protection Against Personalized Text-to-Image Synthesis via Cross-image Consistency Constraints", "abstract": "The rapid advancement of diffusion models and personalization techniques has made it possible to recreate individual portraits from just a few publicly available images. While such capabilities empower various creative applications, they also introduce serious privacy concerns, as adversaries can exploit them to generate highly realistic impersonations. To counter these threats, anti-personalization methods have been proposed, which add adversarial perturbations to published images to disrupt the training of personalization models. However, existing approaches largely overlook the intrinsic multi-image nature of personalization and instead adopt a naive strategy of applying perturbations independently, as commonly done in single-image settings. This neglects the opportunity to leverage inter-image relationships for stronger privacy protection. Therefore, we advocate for a group-level perspective on privacy protection against personalization. Specifically, we introduce Cross-image Anti-Personalization (CAP), a novel framework that enhances resistance to personalization by enforcing style consistency across perturbed images. Furthermore, we develop a dynamic ratio adjustment strategy that adaptively balances the impact of the consistency loss throughout the attack iterations. Extensive experiments on the classical CelebHQ and VGGFace2 benchmarks show that CAP substantially improves existing methods.", "date": "2025-04-17", "categories": []}
{"id": "2508.03221", "title": "BadBlocks: Low-Cost and Stealthy Backdoor Attacks Tailored for Text-to-Image Diffusion Models", "abstract": "Despite the remarkable progress of diffusion models in image generation, recent studies reveal their vulnerability to backdoor attacks via covert visual or textual triggers. Although evolving defense mechanisms can detect most existing threats through visual inspection or feature analysis, we introduce BadBlocks-a novel, lightweight, and highly covert attack that challenges these safeguards. By selectively poisoning specific blocks within the UNet architecture while keeping other components intact, BadBlocks requires only 30% of the computational resources and 20% of the GPU time of conventional attacks, effectively democratizing backdoor injection on consumer-grade GPUs. Empirical evaluations demonstrate that BadBlocks achieves a high attack success rate with negligible perceptual quality loss, while successfully bypassing state-of-the-art defenses, particularly attention-based detection frameworks. Layer-level ablation studies further confirm that backdoor mapping does not require full-network fine-tuning, revealing the disparate vulnerability of different neural layers. Overall, BadBlocks significantly lowers the barrier for executing backdoor attacks, presenting a critical security risk. Our code is available at: https://github.com/paoche11/BadBlocks.", "date": "2025-08-05", "categories": []}
{"id": "2510.12310", "title": "DeepTrust: Multi-Step Classification through Dissimilar Adversarial Representations for Robust Android Malware Detection", "abstract": "Over the last decade, machine learning has been extensively applied to identify malicious Android applications. However, such approaches remain vulnerable against adversarial examples, i.e., examples that are subtly manipulated to fool a machine learning model into making incorrect predictions. This research presents DeepTrust, a novel metaheuristic that arranges flexible classifiers, like deep neural networks, into an ordered sequence where the final decision is made by a single internal model based on conditions activated in cascade. In the Robust Android Malware Detection competition at the 2025 IEEE Conference SaTML, DeepTrust secured the first place and achieved state-of-the-art results, outperforming the next-best competitor by up to 266% under feature-space evasion attacks. This is accomplished while maintaining the highest detection rate on non-adversarial malware and a false positive rate below 1%. The method's efficacy stems from maximizing the divergence of the learned representations among the internal models. By using classifiers inducing fundamentally dissimilar embeddings of the data, the decision space becomes unpredictable for an attacker. This frustrates the iterative perturbation process inherent to evasion attacks, enhancing system robustness without compromising accuracy on clean examples.", "date": "2025-10-14", "categories": []}
{"id": "2602.14399", "title": "Multi-Turn Adaptive Prompting Attack on Large Vision-Language Models", "abstract": "Multi-turn jailbreak attacks have proven effective against text-only large language models (LLMs), where malicious content is gradually introduced to bypass safety alignment. However, effectively extending such attacks to large vision-language models (LVLMs) remains underexplored. In this paper, we find that naively incorporating visual inputs can make multi-turn jailbreaks easier to defend against; for example, overly malicious visual content will easily trigger the defense mechanism in safety-aligned LVLMs, resulting in more conservative responses. Based on this finding, we propose multi-turn adaptive prompting attack (MAPA) that 1) at each turn, alternates text-vision attack actions to elicit the most malicious response; and 2) across turns, adjusts the attack trajectory through iterative back-and-forth refinement to gradually amplify response maliciousness. This two-level design enables MAPA to consistently outperform state-of-the-art methods, improving attack success rates by 15-30% on recent benchmarks against LLaVA-v1.6-Mistral-7B, Qwen2.5-VL-7B-Instruct, Llama-3.2-Vision-11B-Instruct and GPT-4o-mini. Our code is available at: https://github.com/thomaschoi143/MAPA.", "date": "2026-02-16", "categories": []}
{"id": "2603.11331", "title": "Jailbreak Scaling Laws for Large Language Models: Polynomial-Exponential Crossover", "abstract": "Adversarial attacks can reliably steer safety-aligned large language models toward unsafe behavior. Empirically, we find that adversarial prompt-injection attacks can amplify attack success rate from the slow polynomial growth observed without injection to exponential growth with the number of inference-time samples. We first identify a minimal statistical mechanism for these two regimes by giving a small set of assumptions on the distribution of safe generation across contexts under which both scaling laws follow. To explain this phenomenon further, we propose a theoretical generative model of proxy language in terms of a spin-glass system operating in a replica-symmetry-breaking regime, where generations are drawn from the associated Gibbs measure and a subset of low-energy, size-biased clusters is designated unsafe. We analytically show how this model naturally realizes the minimal assumptions. Short injected prompts correspond to a weak magnetic field aligned towards unsafe cluster centers and yield a power-law scaling of attack success rate with the number of inference-time samples, while long injected prompts, i.e., strong magnetic field, yield exponential scaling. We observe qualitatively consistent behavior across a broad range of large language models, spanning parameter scales from 3B to 70B. In particular, the main trends remain stable across multiple attack methods, such as GCG and AutoDAN, as well as across benchmark datasets such as AdvBench and HarmBench.", "date": "2026-03-11", "categories": []}
{"id": "2604.26506", "title": "SafeReview: Defending LLM-based Review Systems Against Adversarial Hidden Prompts", "abstract": "As Large Language Models (LLMs) are increasingly integrated into academic peer review, their vulnerability to adversarial hidden prompts, i.e., adversarial instructions embedded in submissions to manipulate outcomes, poses a critical threat to scholarly integrity. We propose SafeReview, a co-evolutionary adversarial training framework for defending LLM-based peer review systems against such attacks. SafeReview jointly trains a Generator model to create sophisticated attack prompts and a Defender model to preserve review integrity under adversarial manipulation. The Generator is optimized to produce increasingly effective prompt injections, while the Defender is strengthened through preference-based training to maintain consistent reviews between clean and attacked submissions. Experimental results show that SafeReview improves robustness against adaptive prompt injection attacks, better preserves paper ranking under attack, and generalizes across attacker architectures compared with static defenses. These results demonstrate the potential of co-evolutionary training as a foundation for securing LLM-assisted peer review.", "date": "2026-04-29", "categories": []}
{"id": "2605.23239", "title": "Self-supervised Adversarial Purification for Graph Neural Networks", "abstract": "Defending Graph Neural Networks (GNNs) against adversarial attacks requires balancing accuracy and robustness, a trade-off often mishandled by traditional methods like adversarial training that intertwine these conflicting objectives within a single classifier. To overcome this limitation, we propose a self-supervised adversarial purification framework. We separate robustness from the classifier by introducing a dedicated purifier, which cleanses the input data before classification. In contrast to prior adversarial purification methods, we propose GPR-GAE, a novel graph auto-encoder (GAE), as a specialized purifier trained with a self-supervised strategy, adapting to diverse graph structures in a data-driven manner. Utilizing multiple Generalized PageRank (GPR) filters, GPR-GAE captures diverse structural representations for robust and effective purification. Our multi-step purification process further facilitates GPR-GAE to achieve precise graph recovery and robust defense against structural perturbations. Experiments across diverse datasets and attack scenarios demonstrate the state-of-the-art robustness of GPR-GAE, showcasing it as an independent plug-and-play purifier for GNN classifiers.", "date": "2026-05-22", "categories": []}
{"id": "2605.26156", "title": "Turning Bias into Bugs: Bandit-Guided Style Manipulation Attacks on LLM Judges", "abstract": "The known stylistic biases in LLM judges, such as a preference for verbosity or specific sentence structures, present an underexplored security vulnerability. In this work, we introduce BITE (BIas exploraTion and Exploitation), a black-box adversarial framework that learns semantics-preserving edits to mislead an LLM judge and artificially inflate the scores it assigns. We cast the selection of stylistic edits as a contextual bandit problem and use a LinUCB policy to adaptively choose edits that maximize the judge's score without access to model parameters or gradients. Empirically, we test BITE across a diverse range of LLM judges and tasks, including both pointwise and pairwise comparisons on chatbot leaderboards and AI-reviewer benchmarks. BITE achieves an attack success rate exceeding 65% and raises scores by 1-2 points on a 9-point scale, all while preserving semantic equivalence. We further assess the attack's stealthiness, showing that BITE evades standard style-control methods and several detection baselines. Our findings expose a fundamental weakness in the LLM-as-a-judge paradigm and motivate robust, attack-aware evaluation. Our code is available at https://github.com/xianglinyang/llm-as-a-judge-attack.", "date": "2026-05-24", "categories": []}
{"id": "2605.28899", "title": "Quantum-Enhanced Adversarial Robustness in Artificial Intelligence", "abstract": "Artificial Intelligence has achieved remarkable success across diverse application domains. However, its vulnerability to adversarial attacks poses significant challenges to reliability, security, and trustworthiness. Adversarial machine learning demonstrates that even highly accurate models can be manipulated through carefully crafted perturbations, raising serious concerns in safety critical systems such as healthcare, finance, and autonomous technologies. In parallel, quantum computing has emerged as a transformative paradigm capable of addressing complex computational problems through principles such as superposition, entanglement, and quantum interference. The convergence of these fields has led to the emergence of quantum artificial intelligence, which explores how quantum techniques can enhance learning efficiency, scalability, and robustness. This chapter provides a comprehensive overview of adversarial machine learning and existing defense strategies, followed by an accessible introduction to quantum computing and quantum machine learning models. It further presents conceptual frameworks for quantum-enhanced adversarial robustness, emphasizing quantum optimization, feature mapping, and hybrid quantum classical architectures. Practical applications, key challenges, and future research directions are also discussed to support the development of secure and trustworthy AI systems.", "date": "2026-05-27", "categories": []}
{"id": "2605.29114", "title": "ReasonBreak: Probing Vulnerabilities in Reasoning-Enabled Vision-Language-Action Models for Autonomous Driving", "abstract": "Vision-Language-Action (VLA) models with integrated reasoning have been proposed for end-to-end autonomous driving, assuming a tight coupling between reasoning and trajectory generation. However, the robustness of such systems under realistic input perturbations remains largely unexplored. We show that these models are highly vulnerable to realistic input perturbations, achieving up to 89% attack success rate (ASR) on reasoning and up to 72% on trajectory manipulation in closed-loop simulation, leading to increased collision rates and degraded safety metrics. Using NVIDIA's recent Alpamayo models as representative industry-developed VLAs, we conduct the first systematic black-box study of reasoning-enabled VLA models under realistic textual input corruptions, evaluating their impact on reasoning and driving behavior. We introduce a reasoning-aware evaluation framework capturing both semantic and structural aspects of reasoning, along with safety-centric measures. We also introduce a benchmark for evaluating attacks and defenses on reasoning-trajectory interactions in autonomous driving. Our results highlight the need for rigorous evaluation and improved defenses to ensure the safety of reasoning-enabled VLA systems in autonomous driving.", "date": "2026-05-27", "categories": []}
{"id": "2605.29983", "title": "Improving Adversarial Robustness of Attribution via Implicit Regularization", "abstract": "The adversarial robustness of attributions is a fundamental requirement for reliable explainability in deep learning, yet existing approaches typically rely on computationally expensive explicit regularization. In this work, we show that attribution robustness can arise implicitly from the learning dynamics of standard stochastic gradient descent. We theoretically motivate this effect through connections between parameter-space and input-space curvature, and validate it across architectures, datasets, and attribution methods, with negligible computational overhead. In contrast, we prove that such robustness gains often does not transfer to attention-based attribution under softmax normalization, due to inherent entropy constraints, and we validate this limitation experimentally. Finally, we show that replacing softmax attention with kernel-based attention restores the robustness gains in transformer models. Our results highlight learning dynamics as a principled and practical mechanism for robust explainability, and reveal fundamental limitations of attention-based attribution under normalization.", "date": "2026-05-28", "categories": []}
{"id": "2502.17832", "title": "MM-PoisonRAG: Disrupting Multimodal RAG with Local and Global Poisoning Attacks", "abstract": "Retrieval-augmented generation (RAG) has become a common practice in multimodal large language models (MLLM) to enhance factual grounding and reduce hallucination. Yet, its reliance on retrieval exposes MLLMs to knowledge poisoning attacks, in which adversaries deliberately inject malicious multimodal content into external knowledge bases to steer models toward generating incorrect or even harmful responses. We present MM-PoisonRAG, a framework to systematically study the vulnerability of multimodal RAG under knowledge poisoning. Specifically, we design two novel attack strategies: Localized Poisoning Attack (LPA), which implants targeted, query-specific multimodal misinformation to manipulate outputs toward attacker-controlled responses, and Globalized Poisoning Attack (GPA), which uses a single, untargeted adversarial injection to broadly corrupt reasoning and collapse generation quality across all queries. Extensive experiments on diverse tasks, multimodal RAG components, and attacker access levels reveal severe vulnerabilities: LPA achieves up to 56% attack success rate even under restricted access, and transfers effectively across four different retrievers without re-optimizing the adversaries. GPA completely disrupts model generation to 0% accuracy with just one poisoned content. Moreover, both LPA and GPA bypass existing defenses, underscoring the fragility of multimodal RAG and establishing MM-PoisonRAG as a foundation for future research on securing RAG frameworks against multimodal knowledge poisoning.", "date": "2025-02-25", "categories": []}
{"id": "2509.06350", "title": "Mask-GCG: Are All Tokens in Adversarial Suffixes Necessary for Jailbreak Attacks?", "abstract": "Jailbreak attacks on Large Language Models (LLMs) have demonstrated various successful methods whereby attackers manipulate models into generating harmful responses that they are designed to avoid. Among these, Greedy Coordinate Gradient (GCG) has emerged as a general and effective approach that optimizes the tokens in a suffix to generate jailbreakable prompts. While several improved variants of GCG have been proposed, they all rely on fixed-length suffixes. However, the potential redundancy within these suffixes remains unexplored. In this work, we propose Mask-GCG, a plug-and-play method that employs learnable token masking to identify impactful tokens within the suffix. Our approach increases the update probability for tokens at high-impact positions while pruning those at low-impact positions. This pruning not only reduces redundancy but also decreases the size of the gradient space, thereby lowering computational overhead and shortening the time required to achieve successful attacks compared to GCG. We evaluate Mask-GCG by applying it to the original GCG and several improved variants. Experimental results show that most tokens in the suffix contribute significantly to attack success, and pruning a minority of low-impact tokens does not affect the loss values or compromise the attack success rate (ASR), thereby revealing token redundancy in LLM prompts. Our findings provide insights for developing efficient and interpretable LLMs from the perspective of jailbreak attacks.", "date": "2025-09-08", "categories": []}
{"id": "2602.04898", "title": "Semantic-level Backdoor Attack against Text-to-Image Diffusion Models", "abstract": "Text-to-image (T2I) diffusion models are widely adopted for their strong generative capabilities, yet remain vulnerable to backdoor attacks. Existing attacks typically rely on fixed textual triggers and single-entity backdoor targets, making them highly susceptible to enumeration-based input defenses and attention-consistency detection. In this work, we propose Semantic-level Backdoor Attack (SemBD), which introduces representation-level triggers based on continuous semantic regions rather than discrete textual patterns. SemBD implants such semantic backdoors by distillation-based editing of the key and value projection matrices in cross-attention layers, enabling semantically equivalent but textually diverse prompts to activate the backdoor. To further enhance stealthiness, SemBD incorporates a semantic regularization to prevent unintended activation under incomplete semantics, as well as multi-entity backdoor targets that avoid highly consistent cross-attention patterns. Extensive experiments demonstrate that SemBD achieves a 100% attack success rate while maintaining strong robustness against state-of-the-art input-level defenses. Our code is available at https://github.com/DPAS-Lab/SemBD/.", "date": "2026-02-03", "categories": []}
{"id": "2605.27545", "title": "PAST2HARM: A Simple Adaptive Past Tense Attack for Jailbreaking Multimodal AI", "abstract": "Jailbreak attacks on multimodal AI systems remain underexplored, even though unsafe image generation can have more severe consequences than unsafe text and current defenses are relatively immature. We introduce PAST2HARM, a simple yet effective adaptive jailbreak framework that bypasses refusal training in state of the art multimodal text to image models. Building on prior findings that past tense reformulations can evade safeguards, PAST2HARM systematically exploits this vulnerability in multimodal generative AI.\n  We characterize the attack along two dimensions. First, breadth: through temporal deepening, the framework incrementally strengthens historical anchoring and archival cues, eroding refusal boundaries across models with varying alignment strength. Second, depth: via iterative escalation after initial compliance, we probe the upper bound of harmful generation, measuring severity using a scalar severity jailbreak metric evaluated by a language model acting as a judge. We find that mid conversation turns form peak vulnerability windows, where harmfulness increases before plateauing and eventually undergoing semantic inversion.\n  We evaluate PAST2HARM on three models Gemini Nano Banana Pro, GPT Image 2, and SD XL achieving attack success rates of 83 percent, 67 percent, and 100 percent in a black box, gradient free setting. Adversarial prompts also transfer across models, with cross model success rates above 50 percent. The attack elicits diverse harmful outputs, including explicit sexual content, political disinformation, historical denial narratives, hate speech, and self harm glorification. We further release a curated benchmark of prompts, reformulations, and outputs as a resource for red teaming and alignment. Our results expose fundamental brittleness in current safeguards and highlight the need for stronger multimodal safety training.", "date": "2026-05-26", "categories": []}
{"id": "2605.27674", "title": "Backdoor Attacks on Fault Detection and Localization in Cyber-Physical Systems", "abstract": "Cyber-Physical Systems (CPS) integrate sensing, communication, computation, and control to support critical infrastructure, including smart grids, industrial automation, and control systems. In the electrical utility domain, various controllers are used in CPS to ensure the system detects and recovers from faults, such as voltage fluctuations, and to perform load balancing in distribution systems. Machine learning- and deep learning-based fault detection and localization frameworks have recently gained significant attention in CPS for their ability to identify anomalies and operational failures in real time. However, these intelligent models are vulnerable to adversarial machine learning attacks, particularly backdoor attacks. In a backdoor attack, an adversary injects malicious patterns into the training data so that the model behaves normally most of the time but produces attacker-controlled outputs when triggered by specific patterns. This paper investigates the threat of backdoor attacks against fault detection and localization mechanisms in recent ML pipelines used in modern CPS systems. We define these threats and explore how they can be realized by designing triggers and evaluating their success in the CPS domain. Our experiments show the attack is successful even with 10\\% of poisoning.", "date": "2026-05-26", "categories": []}
{"id": "2605.28030", "title": "SPARD: Defending Harmful Fine-Tuning Attack via Safety Projection with Relevance-Diversity Data Selection", "abstract": "Fine-tuning large language models often undermines their safety alignment, a problem further amplified by harmful fine-tuning attacks in which adversarial data removes safeguards and induces unsafe behaviors. We propose SPARD, a defense framework that integrates Safety-Projected Alternating optimization with Relevance-Diversity aware data selection. SPARD employs SPAG, which optimizes alternatively between utility updates and explicit safety projections with a set of safe data to enforce safety constraints. To curate safe data, we introduce a Relevance-Diversity Determinantal Point Process to select compact safe data, balancing task relevance and safety coverage. Experiments on GSM8K and OpenBookQA under four harmful fine-tuning attacks demonstrate that SPARD consistently achieves the lowest average attack success rates, substantially outperforming state-of-the-art defense methods, while maintaining high task accuracy. Code is available at https://github.com/shuhao02/SPARD.", "date": "2026-05-27", "categories": []}
{"id": "2605.28074", "title": "SilentRetrieval: Hijacking Retrieval-Augmented Generation via Semantically-Preserving Adversarial Data Poisoning", "abstract": "Retrieval-Augmented Generation (RAG) mitigates LLM hallucinations but introduces a critical vulnerability: corpus integrity. We present SilentRetrieval, a two-stage data poisoning attack that hijacks RAG systems through adversarially crafted yet fluent documents. Stage 1 uses Coordinated Beam Search, a multi-token joint optimization method with a fluency-similarity objective, to keep a poisoned host document retrievable while constraining perplexity. Stage 2 uses Context-Adaptive Trigger Generation, a lightweight trigger-fusion step driven by a frozen LLM, to integrate manipulation triggers into document content. Under a one-poisoned-document-per-query evaluation with synthetic target answers, SilentRetrieval achieves 84.6%/81.3% HR@10 and 57.5%/54.8% ASR-LLM on Natural Questions and MS MARCO, while maintaining near-benign perplexity. Cross-model evaluation across four target LLMs shows nontrivial effectiveness under a fixed trigger generator, and transfer tests against unseen retrievers, including ColBERT and commercial embedding models, yield 64.7% average HR@10 under the same injected-corpus protocol. In a sampled Wikipedia-scale evaluation, SilentRetrieval retains 74.2% HR@10 at a 0.016% poisoning ratio. Combined retrieval-side and generation-side defenses reduce attack success substantially but incur a latency trade-off. Human evaluation shows substantially lower flag rates than disfluent baselines, while remaining numerically more suspicious than benign content at the current sample size.", "date": "2026-05-27", "categories": []}
{"id": "2605.28083", "title": "VLA-Hijack: A Transferable Patch Attack against Vision-Language-Action Models via Visual Proprioception Hijacking", "abstract": "While Vision-Language-Action (VLA) models have emerged as powerful generalist policies, their severe vulnerability to adversarial patches significantly hinders their deployment in safety-critical domains. Moreover, existing patch attacks primarily focus on white-box settings, heavily overfitting to the specific action output space of the target model, which results in poor cross-architecture transferability. To overcome this limitation, we propose VLA-Hijack, a unified adversarial framework that breaks the transferability bottleneck by exploiting a fundamental vulnerability identified in this work: before planning any motion, a VLA model must first use visual information to locate its own robotic arm within the environment. Targeting this shared visual self-localization process, our approach concurrently optimizes Attention-Guided Proprioceptive Suppression to inhibit the real robotic arm's features, and Multimodal Proprioceptive Injection to establish the patch as a surrogate \"phantom embodiment\". By alternating between semantic concept anchoring and visual prototype projection, VLA-Hijack effectively severs the semantic relationship between the agent's true embodiment and its control policy. Extensive experiments across diverse architectures (OpenVLA, UniVLA, and CronusVLA) demonstrate that VLA-Hijack achieves superior optimization efficiency in white-box settings and sets a new SOTA for cross-architecture and cross-domain black-box transferability.", "date": "2026-05-27", "categories": []}
{"id": "2605.28116", "title": "MIRAGE: Context-Aware Prompt Injection against Mobile GUI Agents via User-Generated Content", "abstract": "Mobile graphical user interface (GUI) agents driven by vision-language models (VLMs) perceive the screen as rendered pixels and choose actions from what they see, so they cannot reliably separate trusted interface elements from user-generated content. We present MIRAGE (Mobile Injection of Realistic Adversarial GUI Examples), a pipeline that turns benign mobile screenshots into prompt-injection samples by placing attacker-controlled text into ordinary user-generated content regions, without modifying the agent, the application, or the operating system. MIRAGE operates in three stages: a Localizer identifies user-controllable regions on the screenshot, a Generator synthesises context-aware payloads and renders them in the application's native style, and a Curator moderates realism and balances the samples across applications, region types, and attack intents. A key challenge is that an injected screenshot must stay visually indistinguishable from genuine user content while still diverting the agent; we address this by separating the stages that control reach, realism, and distributional balance. On a 1,111-sample benchmark spanning ten applications and eleven attack intents, all five evaluated VLM agents are vulnerable, with attack success rates of 23%-30%, and MIRAGE scores higher on human realism ratings than the strongest prior attack (3.02 versus 2.52 out of 5). We further find that per-sample realism and attack success are uncorrelated, so visual-quality filtering alone cannot reliably defend against this threat.", "date": "2026-05-27", "categories": []}
{"id": "2605.28201", "title": "Plant, Persist, Trigger: Sleeper Attack on Large Language Model Agents", "abstract": "Large Language Model (LLM) agents remain vulnerable to safety threats from the external environment, where attackers inject adversarial content into external observations such as tool-returned data, webpages, or MCP context, causing harmful agentic behaviors such as unsafe actions or incorrect outputs. Existing studies typically focus on single-interaction attacks, where the agent observes adversarial content and immediately exhibits harmful behavior within one user request. However, we show that adversarial content can also persist across interactions served by the same agent, making such threats harder to detect and mitigate. Specifically, adversarial content may persist in the agent state, remain dormant across interactions, and later be activated by a benign user query. We formalize this type of safety threat as Sleeper Attack. To evaluate it, we construct a benchmark with 1,896 instances covering six real-world harmful outcomes, three attack strategies, and three agent state targets: session context, memory, and reusable skills. Experiments on seven strong open-source and closed-source LLMs show that state-of-the-art LLM agents remain vulnerable to Sleeper Attack, even when they achieve low attack success rates under a single-interaction baseline. Our code and data are available at https://anonymous.4open.science/r/skdvnfu23ihr9wdscnksf1asdffsaef.", "date": "2026-05-27", "categories": []}
{"id": "2605.28467", "title": "Mitigating Adaptive Attacks against Reasoning Models with Activation Consistency Training", "abstract": "As LLMs gain stronger reasoning capabilities, their extended chain-of-thought introduces new degrees of complexity for defending against adversarial jailbreaks and prompt injection. We study consistency training, a family of fine-tuning objectives that enforce identical behavior on clean prompts and adversarial rewrites, and evaluate its two main variants, output-level (BCT) and activation-level (ACT), across five reasoning models. We formulate both methods as a prompt injection defense and find ACT to be competitive with other training-based defenses while requiring only self-supervised pairs of clean and wrapped prompts. Our experiments also generalize both techniques within the jailbreak setting, demonstrating that ACT remains more robust to adaptive attacks. We also provide mechanistic evidence that ACT's defense against jailbreaks is encoded as a roughly linear shift in activation space at the assistant-turn boundary. After ACT training, we can recover a single steering direction that controls refusal on reasoning models with minimal effect on benign inputs. We find that ACT remains robust even when the model's chain-of-thought is replaced with a compliant trace from the undefended base model, pivoting to refuse prefilled jailbreaks. Together, these results suggest that supervising internal representations is a surprisingly effective and interpretable approach to various forms of safety training in reasoning models.", "date": "2026-05-27", "categories": []}
{"id": "2605.28609", "title": "JECA^2: Judgment-Explanation Consistent Adversarial Attack against Forensic Vision-Language Models", "abstract": "Forensic vision-language models (VLMs) have recently been developed to detect image tampering and provide natural-language explanations. However, their robustness against adversarial manipulation remains underexplored. Existing adversarial attacks typically aim to flip the model's binary judgment, while the accompanying explanation may still reveal forensic cues and contradict the attacked judgment. In this paper, we study judgment-explanation consistent adversarial attacks against forensic VLMs and propose JECA^2, a controlled white-box red-team diagnostic that jointly redirects visual attribution and aligns textual explanations with the target judgment. On the visual side, JECA^2 uses Grad-CAM-guided perturbations to divert attribution from tampered regions toward benign regions. On the textual side, it optimizes prompt embeddings toward authenticity-affirming semantics under a token-proximity constraint. Experiments on forensic VLM benchmarks show that JECA^2 achieves higher attack success and automated judgment-explanation consistency than implemented baselines under white-box threat settings, while transfer to closed-source VLMs remains measurable but limited. Our results highlight a consistency failure mode in explanation-based forensic VLMs and motivate future robustness evaluation beyond binary detection accuracy.", "date": "2026-05-27", "categories": []}
{"id": "2510.17759", "title": "VERA-V: Variational Inference Framework for Jailbreaking Vision-Language Models", "abstract": "Vision-Language Models (VLMs) extend large language models with visual reasoning, but their multimodal design also introduces new, underexplored vulnerabilities. Existing multimodal red-teaming methods largely rely on brittle templates, focus on single-attack settings, and expose only a narrow subset of vulnerabilities. To address these limitations, we introduce VERA-V, a variational inference framework that recasts multimodal jailbreak discovery as learning a joint posterior distribution over paired text-image prompts. This probabilistic view enables the generation of stealthy, coupled adversarial inputs that bypass model guardrails. We train a lightweight attacker to approximate the posterior, allowing efficient sampling of diverse jailbreaks and providing distributional insights into vulnerabilities. VERA-V further integrates three complementary strategies: (i) typography-based text prompts that embed harmful cues, (ii) diffusion-based image synthesis that introduces adversarial signals, and (iii) structured distractors to fragment VLM attention. Experiments on HarmBench and HADES benchmarks show that VERA-V consistently outperforms state-of-the-art baselines on both open-source and frontier VLMs, achieving up to 53.75% higher attack success rate (ASR) over the best baseline on GPT-4o. We include the code on the project page available here: https://github.com/kxwhiowo/VERA-V", "date": "2025-10-20", "categories": []}
{"id": "2511.01724", "title": "PRBench: A Standardized Probabilistic Robustness Benchmark", "abstract": "Deep learning models are notoriously vulnerable to imperceptible perturbations. Most existing research centers on adversarial robustness (AR), which evaluates models under worst-case scenarios by examining the existence of deterministic adversarial examples (AEs). In contrast, probabilistic robustness (PR) adopts a statistical perspective, measuring the probability that predictions remain correct under stochastic perturbations. While PR is widely regarded as a practical complement to AR, dedicated training methods for improving PR are still relatively underexplored, albeit with emerging progress. Among the few PR-targeted training methods, we identify three limitations: i non-comparable evaluation protocols; ii limited comparisons to strong AT baselines despite anecdotal PR gains from AT; and iii no unified framework to compare the generalization of these methods. Thus, we introduce PRBench, the first benchmark dedicated to evaluating improvements in PR achieved by different robustness training methods. PRBench empirically compares most common AT and PR-targeted training methods using a comprehensive set of metrics, including clean accuracy, PR and AR performance, training efficiency, and generalization error (GE). We also provide theoretical analysis on the GE of PR performance across different training methods. Main findings revealed by PRBench include: AT methods are more versatile than PR-targeted training methods in terms of improving both AR and PR performance across diverse hyperparameter settings, while PR-targeted training methods consistently yield lower GE and higher clean accuracy. A leaderboard comprising 229 trained models across 7 datasets and 10 model architectures is publicly available at https://wellzline.github.io/PRBenchLeaderboard/.", "date": "2025-11-03", "categories": []}
{"id": "2604.27019", "title": "Dynamic Adversarial Fine-Tuning Reorganizes Refusal Geometry", "abstract": "Safety-aligned language models must refuse harmful requests without broad over-refusal, but it remains unclear how dynamic adversarial fine-tuning changes refusal-control carriers: Kullback--Leibler (KL)-constrained directions or small subspaces that causally modulate refusal without large safe-prompt distribution shifts. We study a 7B backbone under supervised fine-tuning (SFT) and Robust Refusal Dynamic Defense (R2D2), aligning HarmBench, StrongREJECT, and XSTest evaluations with five-anchor geometry measurements, causal interventions, and sparse adaptive stress tests. R2D2 drives fixed-source HarmBench attack success to zero at early checkpoints; however, these checkpoints also exhibit maximal XSTest refusal and fail a benign-utility audit. Later checkpoints partially recover utility-facing behavior while reopening attack success, with adaptive GCG attack success rate rising to 0.415 at step 250 and 0.613 at step 500. Internally, R2D2 preserves a late-layer admissible refusal-control carrier through step 100 and then relocates the best admissible carrier to an early layer; SFT relocates earlier yet remains less robust. Effective rank stays near 1.24, and SFT shows larger principal-angle drift, arguing against both dimensional expansion and drift magnitude as sufficient explanations. Causal interventions support a low-dimensional but utility-coupled carrier. These results support a geometry-reorganization account of R2D2 along a robustness--utility frontier, without establishing adaptive robustness.", "date": "2026-04-29", "categories": []}
{"id": "2605.26332", "title": "Erased but Exploitable: Black-box Embedding-Aware Prompting Against Unlearned Text-to-Image Diffusion Models", "abstract": "Machine unlearning aims to remove specific concepts from pretrained text-to-image diffusion models, yet several white- and black-box attacks have been introduced to make the model generate such unlearned concepts. These attacks, nevertheless, do not assume a realistic threat model, i.e. they either assume access to the model weights, or result in gibberish adversarial prompts that could be easily detected even through naive rule-based safeguarding. We aim to address this gap in this paper.\n  We introduce BEAP, a black-box, embedding-aware adversarial prompting attack that leverages a large language model (LLM) to iteratively generate effective adversarial prompts and exploit such hidden vulnerabilities.\n  BEAP performs an embedding-aware search in text space, combining multiple reward signals: unlearned concept presence, text-image alignment, and image quality, to refine generated prompts.\n  Unlike previous attack methods, BEAP keeps its prompts undetectable to safety filters while producing high-quality images.\n  Extensive experiments show that BEAP improves the Attack Success Rate (ASR) by more than 60% over prior methods, while requiring only an average of fifteen prompts per successful attack. Warning: This paper contains model outputs that may be offensive or upsetting in nature.", "date": "2026-05-25", "categories": []}
{"id": "2505.18979", "title": "Dynamic Optimization and Safety Indicator Injection for Jailbreaking Text-to-Image Models with Multimodal Safety Filters", "abstract": "Text-to-image (T2I) models can generate not-safe-for-work (NSFW) content, motivating multi-stage safety pipelines with both text and image filters. Newer LLM-based filters detect latent intent beyond keywords, making token-level perturbation attacks unreliable. Our evaluation further shows that existing jailbreak methods exhibit a sharp trade-off between filter evasion and semantic fidelity, while also requiring excessive queries to succeed. We introduce \\textbf{OptJail}, an automated jailbreak framework that combines dynamic prompt optimization with multimodal feedback. It consists of two key components: (i) \\textit{Dynamic Optimization}, an iterative process that leverages text-filter feedback and semantic consistency to rewrite prompts into adversarial variants; and (ii) \\textit{Adaptive Safety Indicator Injection}, which formulates the injection of benign visual cues as a reinforcement learning problem to bypass image-level filters. OptJail achieves state-of-the-art performance, increasing the ShieldLM-7B bypass rate from 8.9\\% (Sneakyprompt) to 99.0\\%, improving CLIP score from 0.2637 to 0.2762. Moreover, it generalizes to unseen filters and successfully jailbreaks DALL E 3 in our evaluation. Mechanistic analysis reveals why these defenses fail: optimized prompts are projected into the ``safe'' region of the filter's representation space yet remain nearly stationary in the generative model's semantic space, and injected safety indicators redirect image detectors' attention away from NSFW content toward benign visual cues. This study reveals systemic vulnerabilities in current multimodal defenses and motivates stronger adaptive defenses.", "date": "2025-05-25", "categories": []}
{"id": "2508.13309", "title": "DASH: A Meta-Attack Framework for Synthesizing Effective and Stealthy Adversarial Examples", "abstract": "Numerous techniques have been proposed for generating adversarial examples in white-box settings under strict Lp-norm constraints. However, such norm-bounded examples often fail to align well with human perception, and only a few methods specifically explore perceptually aligned adversarial examples. Moreover, it remains unclear whether insights from Lp-constrained attacks can be effectively leveraged to improve perceptual efficacy. In this paper, we introduce DASH, a fully differentiable meta-attack framework that generates effective and perceptually aligned adversarial examples by strategically composing existing Lp-based attack methods. DASH operates in a multi-stage fashion: at each stage, it aggregates candidate adversarial examples from multiple base attacks using learned, adaptive weights and propagates the result to the next stage. A novel meta-loss function guides this process by jointly minimizing misclassification loss and perceptual distortion, enabling the framework to dynamically modulate the contribution of each base attack throughout the stages. We evaluate DASH on adversarially trained models across CIFAR-10, CIFAR-100, and ImageNet. Despite relying solely on Lp-constrained based methods, DASH significantly outperforms state-of-the-art perceptual attacks such as AdvAD, achieving higher attack success rates (e.g., 20.63% improvement) and superior visual quality, as measured by SSIM, LPIPS, and FID (improvements $\\approx$ of 11, 0.015, and 5.7, respectively). Furthermore, DASH generalizes well to unseen defenses, making it a practical and strong baseline for evaluating robustness without requiring handcrafted adaptive attacks for each new defense.", "date": "2025-08-18", "categories": []}
{"id": "2509.12672", "title": "Towards Inclusive Toxic Content Moderation: Addressing Vulnerabilities to Adversarial Attacks in Toxicity Classifiers Tackling LLM-generated Content", "abstract": "The volume of machine-generated content online has grown dramatically due to the widespread use of Large Language Models (LLMs), leading to new challenges for content moderation systems. Conventional content moderation classifiers, which are usually trained on text produced by humans, suffer from misclassifications due to LLM-generated text deviating from their training data and adversarial attacks that aim to avoid detection. Present-day defence tactics are reactive rather than proactive, since they rely on adversarial training or external detection models to identify attacks. In this work, we aim to identify the vulnerable components of toxicity classifiers that contribute to misclassification, proposing a novel strategy based on mechanistic interpretability techniques. Our study focuses on fine-tuned BERT and RoBERTa classifiers, testing on diverse datasets spanning a variety of minority groups. We use adversarial attacking techniques to identify vulnerable circuits. Finally, we suppress these vulnerable circuits, improving performance against adversarial attacks. We also provide demographic-level insights into these vulnerable circuits, exposing fairness and robustness gaps in model training. We find that models have distinct heads that are either crucial for performance or vulnerable to attack and suppressing the vulnerable heads improves performance on adversarial input. We also find that different heads are responsible for vulnerability across different demographic groups, which can inform more inclusive development of toxicity detection models.", "date": "2025-09-16", "categories": []}
{"id": "2512.21815", "title": "High-Entropy Tokens as Multimodal Failure Points in Vision-Language Models", "abstract": "Vision-language models (VLMs) achieve remarkable performance but remain vulnerable to adversarial attacks. Entropy, as a measure of model uncertainty, is highly correlated with VLM reliability. While prior entropy-based attacks maximize uncertainty at all decoding steps, implicitly assuming that every token equally contributes to model instability, we reveal that a small fraction (around 20%) of high-entropy tokens, in the evaluated representative open-source VLMs with diverse architectures, concentrates a disproportionate share of adversarial influence during autoregressive generation. We demonstrate that concentrating adversarial perturbations on these high-entropy positions achieves comparable semantic degradation to global methods while optimizing fewer decoding positions. Additionally, across multiple representative VLMs, such attacks induce not only semantic drift but also a substantial unsafe subset (20-31%) under the current pipeline. Remarkably, since such vulnerable high-entropy tokens recur across architecturally diverse VLMs, attacks focused on them exhibit non-trivial transferability. Motivated by these findings, we design a simple Entropy-Guided Attack (EGA) that operationalizes sparse high-entropy targeting and extends it with a reusable token bank, yielding competitive attack success rates (93-95%) with a considerable harmful rate (30.2-38.6%) on the three representative open-source VLMs.", "date": "2025-12-26", "categories": []}
{"id": "2512.23995", "title": "RepetitionCurse: Measuring and Understanding Router Imbalance in Mixture-of-Experts LLMs under DoS Stress", "abstract": "Mixture-of-Experts architectures have become the standard for scaling large language models due to their superior parameter efficiency. To accommodate the growing number of experts in practice, modern inference systems commonly adopt expert parallelism to distribute experts across devices. However, the absence of explicit load balancing constraints during inference allows adversarial inputs to trigger severe routing concentration. We demonstrate that out-of-distribution prompts can manipulate the routing strategy such that all tokens are consistently routed to the same set of top-$k$ experts, which creates computational bottlenecks on certain devices while forcing others to idle. This converts an efficiency mechanism into a denial-of-service attack vector, leading to violations of service-level agreements for time to first token. We propose RepetitionCurse, a low-cost black-box strategy to exploit this vulnerability. By identifying a universal flaw in MoE router behavior, RepetitionCurse constructs adversarial prompts using simple repetitive token patterns in a model-agnostic manner. On widely deployed MoE models like Mixtral-8x7B, our method increases end-to-end inference latency by 3.063x, degrading service availability significantly.", "date": "2025-12-30", "categories": []}
{"id": "2602.09431", "title": "Grounding-Driven Attack: Improving Encoder-based Adversarial Transferability against Large Vision-Language Models", "abstract": "Large vision-language models (LVLMs) have achieved impressive performance across multimodal tasks, but their reliance on visual inputs exposes them to adversarial threats. Encoder-based attacks provide an efficient alternative to end-to-end optimization by crafting perturbations through the vision encoder alone. However, existing encoder-based attacks often assume that the surrogate encoder is identical or similar to the victim LVLM's vision encoder. In this work, we present a systematic study of their transferability in more realistic black-box deployments with heterogeneous LVLM architectures. We find that model-specific visual evidence is inconsistent across models, whereas text-conditioned grounding regions are more closely tied to caption-relevant evidence and provide a more stable transfer target. However, existing attacks remain weakly aligned with and insufficiently disrupt these regions. Motivated by these findings, we propose Grounding-Driven Attack (GDA), which aligns perturbation optimization with text-grounded evidence. GDA combines Grounding-Aware Perturbation Allocation to concentrate perturbation budget on grounded evidence regions with Grounding-Centric Evidence Disruption to intensify their global and local disruption. Experiments across diverse victim models and tasks show that GDA consistently outperforms existing encoder-based attacks in black-box transfer. These results highlight the central role of text-grounded evidence in adversarial transferability and motivate grounding-aware robustness evaluation and defense design.", "date": "2026-02-10", "categories": []}
{"id": "2605.02495", "title": "Efficient Preference Poisoning Attack on Offline RLHF", "abstract": "Offline Reinforcement Learning from Human Feedback (RLHF) pipelines such as Direct Preference Optimization (DPO) train on a pre-collected preference dataset, which makes them vulnerable to preference poisoning attack. We study label flip attacks against log-linear DPO. We first illustrate that flipping one preference label induces a parameter-independent shift in the DPO gradient. Using this key property, we can then convert the targeted poisoning problem into a structured binary sparse approximation problem. To solve this problem, we develop two attack methods: Binary-Aware Lattice Attack (BAL-A) and Binary Matching Pursuit Attack (BMP-A). BAL-A embeds the binary flip selection problem into a binary-aware lattice and applies Lenstra-Lenstra-Lov\\'asz reduction and Babai's nearest plane algorithm; we provide sufficient conditions that enforce binary coefficients and recover the minimum-flip objective. BMP-A adapts binary matching pursuit to our non-normalized gradient dictionary and yields coherence-based recovery guarantees and robustness (impossibility) certificates for $K$-flip budgets. Experiments on synthetic dictionaries and the Stanford Human Preferences dataset validate the theory and highlight how dictionary geometry governs attack success.", "date": "2026-05-04", "categories": []}
{"id": "2605.10764", "title": "Break the Brake, Not the Wheel: Untargeted Jailbreak via Entropy Maximization", "abstract": "Recent studies show that gradient-based universal image jailbreaks on vision-language models (VLMs) exhibit little or no cross-model transferability, casting doubt on the feasibility of transferable multimodal jailbreaks. We revisit this conclusion under a strictly untargeted threat model without enforcing a fixed prefix or response pattern. Our preliminary experiment reveals that refusal behavior concentrates at high-entropy tokens during autoregressive decoding, and non-refusal tokens already carry substantial probability mass among the top-ranked candidates before attack. Motivated by this finding, we propose Untargeted Jailbreak via Entropy Maximization(UJEM)-KL, a lightweight attack that maximizes entropy at these decision tokens to flip refusal outcomes, while stabilizing the remaining low-entropy positions to preserve output quality. Across three VLMs and two safety benchmarks, UJEM-KL achieves competitive white-box attack success rates and consistently improves transferability, while remaining effective under representative defenses. Our experimental results indicate that the limited transferability primarily stems from overly constrained optimization objectives.", "date": "2026-05-11", "categories": []}
{"id": "2605.16815", "title": "Universal Graph Backdoor Defense: A Feature-based Homophily Perspective", "abstract": "Graph neural networks (GNNs) have achieved remarkable success in relational learning. However, their vulnerability to graph backdoor attacks (GBAs) poses a significant barrier to broader adoption in high-stakes applications. Despite recent advances in graph backdoor defense (GBD), existing methods primarily focus on subgraph-based GBAs, relying on the assumption that poisoned target nodes are explicitly connected to subgraph triggers. Our empirical results reveal that such structure-centric approaches fail to defend against emerging feature-based GBAs that preserve graph topology. Therefore, in this paper, we study a novel problem of universal graph backdoor defense. First, we investigate the shared effects of both attack types from a feature-based homophily perspective, which characterizes local feature consistency between nodes and their neighborhoods. Thorough theoretical and empirical analyses demonstrate that, regardless of trigger mechanisms, backdoors induced by GBAs exhibit lower feature-based homophily than clean nodes, indicating a discrepancy in local feature similarity. Motivated by this insight, we propose to leverage node-level local feature consistency, modeled by a neighbor-aware reconstruction loss, to distinguish backdoors from clean nodes. Then, a robust training strategy is developed to eliminate trigger effects while reducing noise induced by detection uncertainty. Extensive experiments demonstrate that our framework significantly degrades the attack success rate and maintains competitive clean accuracy under both subgraph-based and feature-based attacks.", "date": "2026-05-16", "categories": []}
{"id": "2605.24659", "title": "IterInject: Indirect Prompt Injection Against LLM Agents via Feedback-Guided Iterative Optimization", "abstract": "LLM-based agents are increasingly deployed for complex tasks requiring planning, tool use, and interaction with external services. Their reliance on untrusted external content exposes them to indirect prompt injection (IPI), in which adversarial instructions embedded in retrieved data hijack agent behavior. Existing attacks rely on static payloads that cannot adapt to agent-specific defenses; even recent adaptive methods lack structured feedback to guide optimization. We introduce \\oursys, a feedback-guided iterative framework that closes the loop between injection, diagnosis, and refinement: a rule-based diagnoser produces structured outcome labels with behavioral descriptions, and an LLM-based optimizer refines payloads conditioned on the full optimization history. A synthesis step generates new disguise seeds from failure patterns, enabling the strategy space to self-evolve. On AgentDojo and InjectAgent, \\oursys substantially outperforms static baselines and existing adaptive methods across four victim models. Extension experiments on Claude Code, a production-grade coding agent with layered defenses, show that optimized payloads achieve full success on 5 of 9 targets; even those that resist full exploitation exhibit measurable improvement from iterative refinement. We further present a mechanistic analysis of IPI, identifying an attention-mediated threshold mechanism in mid-to-late layers; three causal interventions validate this finding and point to concrete defense directions.", "date": "2026-05-23", "categories": []}
{"id": "2605.24958", "title": "SEP-Attack: A Simple and Effective Paradigm for Transfer-Based Textual Adversarial Attack", "abstract": "Despite the strong performance of deep neural networks in modern Web and language applications, they remain vulnerable to adversarial attacks, especially transferable attacks that generate adversarial examples using surrogate models without accessing the victim model. Transferable attacks in the text domain are still under-explored, with only a few studies addressing this challenging issue, often with suboptimal results due to equal treatment of submodels or inaccurate estimation of importance scores. To address these challenges, we propose a simple yet effective paradigm for transfer-based textual adversarial attack, named SEP-Attack. Specifically, we employ the Determinantal Point Process (DPP) to generate diverse surrogate ensemble weights, representing the transferability of submodels. Using these weights, we introduce a new metric to evaluate prediction confidence scores, which in turn are used to calculate word importance scores and generate adversarial candidates. Finally, we quantify the transferability score for each candidate and select the top ones as the final transferable adversarial examples. Experiments conducted on four datasets and two real-world APIs validate the efficacy of SEP-Attack, significantly outperforming state-of-the-art baselines.", "date": "2026-05-24", "categories": []}
{"id": "2605.25304", "title": "When Interpretability Becomes a Liability: Adversarial Attacks on CBM Concept Layers", "abstract": "Concept Bottleneck Models (CBMs) have emerged as a cornerstone approach for interpretable machine learning, providing human-understandable intermediate representations through explicit concept activations. However, this interpretability fundamentally introduces a critical, previously unexplored attack surface: the concept bottleneck layer itself. We present a comprehensive, systematic study of concept-level adversarial vulnerabilities in CBMs, revealing that targeted, minimal perturbations operating on input pixels can induce catastrophic misclassification by manipulating semantic representations. We develop a rigorous theoretical framework to quantify concept-space robustness, establishing novel metrics that expose the vulnerability landscape of these architectures. Our extensive analysis on the CUB-200-2011 dataset demonstrates that standard CBMs exhibit severe susceptibility to concept-level manipulation. To address this critical weakness, we introduce SPECTRA (Semantic Perturbation-based Concept Training for Robustness against Attacks), a principled stability regularization defense. SPECTRA effectively hardens the semantic representation space, increasing the minimal perturbation norm required for a successful attack from 0.46 to over 4,200, rendering targeted concept manipulation computationally prohibitive. Furthermore, SPECTRA preserves baseline classification accuracy to within 2.2%. By establishing concept-level attacks as a fundamentally distinct threat model, this work opens a new research frontier at the intersection of interpretable machine learning and adversarial robustness.", "date": "2026-05-25", "categories": []}
{"id": "2605.25352", "title": "Certified Robustness from Approximate Gaussian Mixture Structures in Pretrained Latent Spaces", "abstract": "Deep learning models are vulnerable to adversarial perturbations, raising important concerns for safety-critical deployment. Empirical defenses can achieve strong robustness in practice, but lack formal guarantees, motivating the need for certifiably robust classifiers. While certified methods provide formal guarantees, they often yield overly conservative bounds due to their inability to exploit structure in complex data distributions. In this work, we propose a framework for designing certifiably robust classifiers that leverages latent structure in data representations. We first analyze the Gaussian mixture setting, deriving necessary and sufficient conditions for the existence of robust classifiers and constructing a classifier with a closed-form robustness certificate and generalization guarantees. Our main contribution is to show that exact structure is not required: we prove that if a pretrained encoder maps inputs to a latent distribution that is $\\varepsilon$-close (in KL divergence) to a Gaussian mixture, then certified accuracy degrades gracefully, with an explicit bound relating robustness under the true and approximate distributions. This result enables the direct use of pretrained models without requiring exact distributional assumptions. Empirically, our method achieves state-of-the-art or competitive certified accuracy on CIFAR-10 and ImageNet, while maintaining strong clean performance and low computational overhead. Overall, our work establishes approximate latent structure as a practical and principled route to certifiable robustness.", "date": "2026-05-25", "categories": []}
{"id": "2605.25663", "title": "Opportunistic Target Selection: Early Directional Commitment for Query-Efficient Black-Box Adversarial Attacks", "abstract": "Black-box adversarial attacks that minimize only the ground-truth confidence suffer from class drift: perturbations wander through the feature space without committing to a specific adversarial class, wasting queries on diffuse, undirected progress. We introduce Opportunistic Target Selection (OTS), a lightweight wrapper that switches an untargeted attack to a targeted objective early in its trajectory, locking onto whichever non-true class currently leads. OTS requires no architectural modification to the underlying attack, no gradient access, and no a priori target-class knowledge.\n  We validate OTS on three score-based attacks (SimBA, Square Attack with cross-entropy loss, and Bandits) across five standard ImageNet classifiers (4,500 runs). On random-search attacks, OTS closely tracks oracle performance, with gains up to +27 pp in success rate and 43% relative reduction in censored-mean iterations on ResNet-50. On gradient-estimation attacks (Bandits) and attacks with margin loss, OTS is redundant, a negative result that reinforces our interpretation of OTS as a margin-loss surrogate. On adversarially-trained models, a bimodal difficulty distribution eliminates the regime where targeting helps.", "date": "2026-05-25", "categories": []}
{"id": "2605.25937", "title": "Building an Adversarial Malware Dataset by Family and Type: Generation, Evasion, and Poisoning Evaluation", "abstract": "We present a dataset of adversarial malware samples derived from the public RawMal-TF collection of real-world malware binaries. Using a suite of adversarial malware generators, we construct two sets of adversarial PE files: 44,347 family-labelled samples and 33,596 type-labelled samples, achieving evasion rates of 98.35 % and 92.20 % against the EMBER classifier, respectively. Each adversarial binary is accompanied by detailed metadata, including EMBER scores and VirusTotal classifications. We further demonstrate the susceptibility of malware classification pipelines to data poisoning attacks through a series of training experiments. Injecting fully mislabelled adversarial samples representing only 0.5 % of the training data in the family-labelled dataset increases the evasion rate against the re-trained classifier from 26.1 % to 92.8 %. The dataset is publicly released to facilitate future research on adversarial malware, poisoning attacks, and the robustness of machine-learning-based malware detection systems.", "date": "2026-05-25", "categories": []}
{"id": "2507.06252", "title": "False Alarms, Real Damage: Adversarial Attacks Using LLM-based Models on Text-based Cyber Threat Intelligence Systems", "abstract": "Cyber Threat Intelligence (CTI) has emerged as a vital complementary approach that operates in the early phases of the cyber threat lifecycle. CTI involves collecting, processing, and analyzing threat data to provide a more accurate and rapid understanding of cyber threats. Due to the large volume of data, automation through Machine Learning (ML) and Natural Language Processing (NLP) models is essential for effective CTI extraction. These automated systems leverage Open Source Intelligence (OSINT) from sources like social networks, forums, and blogs to identify Indicators of Compromise (IoCs). Although prior research has focused on adversarial attacks on specific ML models, this study expands the scope by investigating vulnerabilities within various components of the entire CTI pipeline and their susceptibility to adversarial attacks. These vulnerabilities arise because they ingest textual inputs from various open sources, including real and potentially fake content. We analyse three types of attacks against CTI pipelines, including evasion, flooding, and poisoning, and assess their impact on the system's information selection capabilities. Specifically, on fake text generation, the work demonstrates how adversarial text generation techniques can create fake cybersecurity and cybersecurity-like text that misleads classifiers, degrades performance, and disrupts system functionality. The focus is primarily on the evasion attack, as it precedes and enables flooding and poisoning attacks within the CTI pipeline.", "date": "2025-07-05", "categories": []}
{"id": "2509.06896", "title": "Are Targeted Data Poisoning Attacks as Effective as We Think?", "abstract": "Targeted data poisoning attacks manipulate model predictions on specific test samples by injecting malicious data into training. Yet existing evaluations report average attack success rates over randomly selected targets, obscuring true worst-case effectiveness. We argue that the right evaluation focuses on the hardest samples to poison. The same reasoning applies to defense: since targeted attacks leave no footprint at the distribution level, defenders should proactively identify the most vulnerable samples and apply targeted countermeasures. Given a test dataset, this paper identifies both the easiest and hardest to poison examples based on only clean model information. Specifically, we offer coarse evaluations using clean training dynamics, and fine-grained classification on poison class using poison distances and budgets. Our experiments show these metrics reliably stratify samples by poisoning vulnerability, enabling both rigorous worst-case evaluation and proactive vulnerability-aware defense.", "date": "2025-09-08", "categories": []}
{"id": "2601.14300", "title": "Low-Cost Hard-Label Adversarial Attack with Theoretical Foundations", "abstract": "Hard-label black-box attacks, relying solely on top-1 predictions, represent one of the most challenging yet practically threat models. Despite recent progress, existing approaches face two key limitations: (1) they overlook the critical role of initialization, focusing primarily on optimization strategies; and (2) they rely heavily on empirical heuristics without theoretical guarantees. To bridge this gap, we establish a unified theoretical framework showing that existing sign-flipping hard-label attacks can be understood as approximating the true gradient sign. Guided by this principled analysis, we propose a novel attack framework featuring a zero-query initialization strategy and a Pattern-Driven Optimization (PDO) algorithm. We provide theoretical guarantees that our initialization yields higher cosine similarity to the true gradient sign than random baselines, and our PDO module achieves significantly lower query complexity than baseline search methods. Extensive experiments across CIFAR-10, ImageNet, and ObjectNet-covering standard and adversarially trained models, commercial APIs, and CLIP models-demonstrate that our method consistently outperforms SOTA hard-label attacks in both success rate and efficiency, particularly under low query budgets. Furthermore, our method demonstrates robust generalization across corrupted data (ImageNet-C), biomedical images (PathMNIST), and dense prediction tasks such as segmentation. Notably, it bypasses the stateful defense Blacklight, achieving a 0% detection rate.", "date": "2026-01-17", "categories": []}
{"id": "2602.00979", "title": "GradingAttack: Exposing Security Vulnerabilities in LLM Based Educational Grading Agents", "abstract": "Large language models (LLMs) are increasingly deployed as educational agents for automatic short answer grading (ASAG) in real-world educational environments, significantly boosting assessment efficiency and scalability. However, when these grading agents operate ``in the wild'', their vulnerability to adversarial manipulation raises critical concerns about agent security and trustworthiness. In this paper, we introduce GradingAttack, a fine-grained adversarial attack framework that systematically evaluates the security vulnerabilities of LLM based educational grading agents. Specifically, we design token-level and prompt-level attack strategies that manipulate agent grading outcomes while maintaining high stealth, exposing fundamental weaknesses in current agent deployments. Experiments on multiple datasets demonstrate that both attack strategies effectively compromise grading agents, with prompt-level attacks achieving higher success rates and token-level attacks exhibiting superior stealth capability. Our findings reveal that current LLM based educational agents lack robust defenses against adversarial attacks, underscoring the urgent need for developing secure and trustworthy agent systems for critical educational applications.", "date": "2026-02-01", "categories": []}
{"id": "2605.07590", "title": "Beyond Defenses: Manifold-Aligned Regularization for Intrinsic 3D Point Cloud Robustness", "abstract": "Despite extensive progress in point cloud robustness, existing methods primarily rely on augmentation strategies or defense mechanisms while overlooking the geometric nature of adversarial fragility. We hypothesize that adversarial vulnerability in 3D networks arises from a manifold misalignment between the latent geometry learned by the model and the intrinsic geometry of the underlying surface. Small, geometry-preserving perturbations along the input manifold often induce disproportionate distortions in feature space, potentially leading to misclassifications. We formalize this phenomenon by developing a geometric interpretation of 3D robustness that links classical adversarial theory to the intrinsic structure of point clouds. Motivated by this analysis, we introduce Manifold-Aligned Point Recognition (MAPR), a framework that regularizes the latent geometry by aligning predictions across intrinsic perturbations. MAPR augments each point cloud with intrinsic features capturing local curvature and diffusion structure, and applies a consistency loss that preserves invariance to intrinsic, geometry-preserving perturbations. Without relying on adversarial training or additional data, MAPR consistently improves robustness under multiple adversarial attacks across several datasets, achieving average robustness gains of +20.02 and +8.83 percentage points over vanilla models on ModelNet40 and ScanObjectNN, respectively.", "date": "2026-05-08", "categories": []}
{"id": "2605.20519", "title": "Codec-Robust Attacks on Audio LLMs", "abstract": "Prior attacks on Audio Large Language Models (Audio LLMs) demonstrated that carefully crafted waveform-domain perturbations can force targeted adversarial outputs. As a defense mechanism against these attacks, real-world codec compression preprocessing has been studied to both detect and remove the perturbations. Yet no existing attack has demonstrated robustness against these compressions. We introduce CodecAttack, which optimizes a perturbation in a neural audio codec's continuous latent space rather than directly perturbing the audio waveform. We show that the codec's compression channel, which discards waveform perturbations, transmits perturbations crafted in its own latent space. To further harden the attack across real-world compression channels, we apply multi-bitrate straight-through Expectation-over-Transformation (EoT), all without modifying the target model. Across three realistic Audio LLM deployment scenarios and three target models, CodecAttack achieves an average 85.5% target-substring attack success rate (ASR) on Opus at moderate bitrates, while the waveform baseline trained with identical EoT hardening does not exceed 26% at any bitrate. The attack transfers to held-out codecs, reaching up to 100% ASR on MP3 and 84% on AAC-LC without retraining. A per-band energy analysis shows that the latent perturbation concentrates below 4kHz, exactly where codecs allocate the most bits, while the waveform baseline spreads into higher frequencies that codecs discard. These results demonstrate that lossy compression is not a reliable defense against adversarial audio and that codec-aware attacks pose a practical threat to deployed Audio LLM systems.", "date": "2026-05-19", "categories": []}
{"id": "2605.22984", "title": "Test-Time Training Undermines Safety Guardrails", "abstract": "Test-Time Training (TTT) is an emerging paradigm that enables models to adapt their parameters during inference, improving performance on tasks such as few-shot learning, retrieval-augmented generation, and complex reasoning. However, this dynamic adaptation introduces new vulnerabilities that adversaries can exploit to jailbreak models. We identify three threat models for TTT and demonstrate how attackers can leverage them to bypass safety filters. Our results show that TTT can significantly increase the Attack Success Rate (ASR) and the ASR over 10 generation trials (ASR@10). For example, under LoRA, the few-shot and generation-phase threat models achieve an average ASR@10 of 95% and 93% respectively, across models from different families and scales. These vulnerabilities transfer to production fine-tuning APIs. We also show that TTT-induced overfitting can produce degenerate outputs that inflate ASR under standard judges, and propose a validity-aware evaluation to correct for this. Our findings suggest that TTT exposes a new attack surface, strengthens attacks, and undermines existing safety guardrails. As a first step toward defense, we propose a lightweight provider-side detector that flags TTT requests via the perplexity shift on a private harmful holdout, but robust deployment will ultimately require dynamic alignment.", "date": "2026-05-21", "categories": []}
{"id": "2605.23065", "title": "Dithering Defense: Adversarial Robustness of Vision Foundation Models via Multi-Level Floyd-Steinberg Dithering", "abstract": "Vision foundation models are widely used as frozen backbones across many downstream tasks, making them a single point of failure under adversarial attack. We study multi-level Floyd-Steinberg error-diffusion dithering as a lightweight, model-agnostic input transformation that disrupts adversarial perturbations while preserving semantic content. Unlike prior work, which was limited to binary dithering, grayscale CIFAR-10, and a single small model trained from scratch, we evaluate across six tasks (classification, segmentation, depth estimation, retrieval, captioning, visual question answering), two model families (DINOv2, PaliGemma), and three attacks of increasing strength (PGD, MI-FGSM, SIA), as well as an adaptive attacker using a straight-through estimator. Our results show that Floyd-Steinberg dithering at intermediate quantization levels, especially when combined with post-processing blur, exceeds or matches all tested baselines, including diffusion-based denoising, with substantially less degradation on clean inputs.", "date": "2026-05-21", "categories": []}
{"id": "2605.23168", "title": "PoisonForge: Task-Level Targeted Poisoning Benchmark for Instruction-Tuned LLMs", "abstract": "When practitioners fine-tune LLMs on unvetted datasets, an adversary can exploit the data supply chain through task-level poisoning: inserting a small number of crafted instruction-response pairs that cause the model to embed attacker-specified entities, such as a country, in outputs for a targeted task family while behaving normally elsewhere. We introduce PoisonForge, a benchmark that parameterizes this threat along four dimensions (bias type, poisoning mode, appearance count, and target output length) and evaluates 12 open-weight models (from 2B to 32B parameters) across five families under a primarily 1% poison budget. With only 10 poisoned examples among 1,000 fine-tuning examples, 11 of 12 models exceed a 70% attack success rate (ASR) in their most vulnerable configuration. Meanwhile, unintended leakage to non-target tasks remains below 0.5%, and models perform well on standard benchmarks. We analyze in detail the factors contributing to attack success. We observe that multiple appearances of an entity increase the ASR, the optimal poisoning mode depends on the semantic structure of the target entity, and ASR drops monotonically with the task output length. A correlation analysis and risk prediction model confirm that poisoning design choices, rather than model scale, are the primary causes of attack success, and that these patterns generalize to predict attack success on new tasks. We release all configurations, pipelines, and analysis code to support reproducible comparisons.", "date": "2026-05-22", "categories": []}
{"id": "2605.23220", "title": "WMAttack: Automated Attack Search for Adversarial Evaluation of World-Model Agents", "abstract": "Despite the growing use of world models as decision-making agents, their adversarial robustness remains underexplored due to the lack of dedicated automated evaluation methods. A key obstacle is that attack evaluation must be both accurate and efficient: weak manually tuned attacks can overestimate robustness, while exhaustive hyperparameter search is prohibitively expensive because each candidate requires closed-loop rollouts through learned latent dynamics. We introduce WMAttack, an automated attack-search framework for adversarial evaluation of world-model agents. WMAttack formulates robustness evaluation as a finite-budget search over attack configurations, including attack families, perturbation budgets, optimization steps, restarts, and allocation rules. To improve search accuracy, Self-Correcting Attack Search (SCAS) refines the attack proposal distribution using feedback from reward degradation, action instability, runtime cost, and rollout variability. To improve search efficiency, Representation-Guided Attack Retrieval (RGAR) retrieves effective historical configurations from representation-similar tasks, providing a warm start for unseen environments. We provide a theoretical explanation showing that proposal refinement improves finite-budget search when it shifts probability mass toward high-utility attacks. Across Atari and DeepMind Control tasks, WMAttack consistently discovers stronger attacks than the evaluated baselines, improving normalized reward drop from 0.497 to 1.034 on DreamerV3 Atari and from 0.319 to 0.682 on DMC. Ablations further show that RGAR improves initial candidate quality and SCAS improves final attack utility under fixed evaluation budgets.", "date": "2026-05-22", "categories": []}
{"id": "2605.23411", "title": "Sample-wise Targeted Adversarial Attacks on Test-time Adaptation", "abstract": "Test-time adaptation (TTA) effectively counters distribution shifts but exposes models to adversarial manipulation via the unlabeled test stream. Existing class-wise targeted attacks remain impractical for stealthy exploitation in this setting: since TTA operates on batches, forcing a subset of samples toward a target label unintentionally pulls similar benign samples along, resulting in a conspicuously high frequency of the target label that is easy to detect. To capture a more realistic threat, we introduce a sample-wise targeted attack. Unlike prior approaches, the attacker aims to misclassify only inputs carrying an attacker-chosen trigger, while preserving the global label distribution of benign queries to evade detection. To achieve this, we propose a meta-learning-based attack with a novel priority-aware gradient alignment strategy that explicitly prioritizes attack success. The strategy formulates the gradient update as an ellipsoidal trust-region problem, mitigating the misalignment between attack success and distributional stealth, while providing theoretical guarantees for effective optimization of the attack objective in the presence of gradient misalignment. Extensive experiments on CIFAR-10-C, CIFAR-100-C, and ImageNet-C across TTA protocols demonstrate that our method achieves high targeted success rates while maintaining a label distribution that is consistent with the no-attack baseline, making it difficult to detect in unlabeled TTA deployment scenarios. Furthermore, we demonstrate that our attack shows strong robustness against existing defenses.", "date": "2026-05-22", "categories": []}
{"id": "2605.23623", "title": "Adversarial Vulnerability Under Temporal Concept Drift: A Longitudinal Study of Android Malware Detection", "abstract": "We present a longitudinal, drift-aware evaluation of adversarial robustness across more than a decade of Android applications using static and dynamic feature representations extracted from emulator and real-device executions. The dataset is organized into yearly slices and evaluated under three deployment protocols that emulate realistic learning scenarios: (1) same-year training and testing, (2) cross-year deployment without model updates, and (3) expanding-window retraining with cumulative historical data. Across multiple classifier families, adversarial examples are generated using FGSM and SPSA under feasibility constraints. We measure clean performance, Adversarial Accuracy (AA), Attack Success Rate (ASR), and introduce temporal linkage metrics -- RobustDrop, $\\Delta$ASR, and Adversarial Amplification Factor (AAF) -- to quantify the relationship between distribution shift and robustness degradation.nResults show that temporal separation is associated with reduced adversarial robustness under the evaluated transfer-based feature-space setting. As the train-test gap increases, clean accuracy and adversarial accuracy decline, while attack success exhibits configuration-dependent increases, particularly under FGSM perturbations and static features. Expanding-window retraining mitigates, but does not eliminate, robustness loss under continued distributional evolution. These findings indicate that temporal drift should be considered when assessing the long-term robustness of intelligent detection systems under evolving data distributions and highlight the need for drift-aware robustness assessment frameworks in long-lived adversarial environments.", "date": "2026-05-22", "categories": []}
{"id": "2209.03358", "title": "Attacking the Spike: On the Transferability and Security of Spiking Neural Networks to Adversarial Examples", "abstract": "Spiking neural networks (SNNs) have attracted much attention for their high energy efficiency and recent advances in classification performance. However, unlike traditional deep learning approaches, the study of SNN robustness to adversarial examples remains relatively underdeveloped. In this work, we advance the adversarial attack side of SNNs through three contributions. First, we show that successful white-box adversarial attacks on SNNs are highly dependent on the underlying surrogate gradient estimator, even for adversarially trained SNNs. Second, using the best single surrogate gradient estimator, we analyze the transferability of adversarial attacks across SNNs, Vision Transformers (ViTs) and CNNs. Our analysis reveals two key gaps: no existing white-box attack exploits multiple surrogate gradient estimators for SNNs, and no single-model attack reliably generates adversarial examples that simultaneously fool both SNN and non-SNN models. For our third contribution, we develop the Mixed Dynamic Spiking Estimation (MDSE) attack to address these issues. MDSE uses a dynamic gradient estimation scheme to fully exploit multiple surrogate gradient estimator functions and generates adversarial examples capable of fooling SNN and non-SNN models simultaneously. MDSE is up to 91.4% more effective on SNN/ViT model ensembles and provides a 3x boost on adversarially trained SNN ensembles compared to conventional white-box attacks like Auto-PGD. Experiments cover three datasets (CIFAR-10, CIFAR-100, ImageNet) and nineteen classifier models (seven per CIFAR dataset, five for ImageNet). Our implementation of MDSE and the evaluated models is publicly available at https://github.com/nuoxuxxx/attacking-the-spike-mdse.", "date": "2022-09-07", "categories": []}
{"id": "2605.21541", "title": "Frequency-Domain Regularized Adversarial Alignment for Transferable Attacks against Closed-Source MLLMs", "abstract": "Multimodal large language models (MLLMs) remain vulnerable to transfer-based targeted attacks, where perturbations optimized on open-source surrogate encoders can generalize to closed-source MLLMs. A key challenge for improving adversarial transferability is to effectively capture the intrinsic visual focus shared across different models, such that perturbations align with transferable semantic cues rather than surrogate-specific behaviors. However, existing methods suffer from spatial-domain feature redundancy and surrogate-specific gradient signals, thereby hindering cross-model transferability. In this paper, we propose FRA-Attack, which addresses both challenges from a unified frequency-domain regularization perspective. For feature alignment, a high-pass DCT objective on patch features suppresses redundant global structures and concentrates the loss on the high-frequency band that carries the MLLMs' intrinsic visual focus. For gradient optimization, we introduce Frequency-domain Gradient Regularization (FGR), a \\textit{model-agnostic} low-pass regularizer that modulates the surrogate gradient using only the geometric frequency coordinate, \\textit{i.e.}, no surrogate-derived statistic is involved, so that FGR is model-agnostic by construction, removing surrogate-specific high-frequency artifacts while preserving transferable low-frequency directions. Together, the two components form a unified frequency-domain treatment of transferability. Extensive experiments on $15$ flagship MLLMs across $7$ vendors show that FRA-Attack achieves superior cross-model transferability, particularly with state-of-the-art performance on GPT-5.4, Claude-Opus-4.6 and Gemini-3-flash.", "date": "2026-05-20", "categories": []}
{"id": "2605.21999", "title": "Toward Understanding Adversarial Distillation: Why Robust Teachers Fail", "abstract": "Adversarial Distillation aims to enhance student robustness by guiding the student with a robust teacher's soft labels within the min-max adversarial training framework, yet its success is notoriously inconsistent: a more robust teacher often fails to improve, or even harms, the student's robust generalization. In this paper, we identify a key mechanism of this teacher dependency: the misalignment between the teacher's supervisory confidence and the student's representational limitations on a consistent subset of training data -- the Robustly Unlearnable Set. We present a theoretical framework analyzing the feature learning dynamics of a two-layer neural network, demonstrating that this mismatch creates a dichotomy in distillation outcomes. We prove that when a teacher provides confident supervision on unlearnable samples, it compels the student to memorize spurious noise patterns that eventually overpower the learned robust signal, thereby driving robust overfitting. Conversely, a teacher that exhibits high uncertainty on these samples effectively suppresses noise memorization, allowing the student to rely solely on the learnable signal for robust generalization. We empirically validate our theory across both synthetic simulations and real-image classification datasets, confirming that robust overfitting is driven by the teacher's interaction with unlearnable samples. Finally, we demonstrate that a teacher's predictive entropy on unlearnable samples serves as a strong indicator of student robustness, validating our theoretical framework and offering a principled guideline for robust teacher selection.", "date": "2026-05-21", "categories": []}
{"id": "2605.22122", "title": "Adversarial Trust Poisoning in Vehicular Collaborative Perception", "abstract": "Collaborative perception (CP) enables connected and autonomous vehicles to share sensor data and jointly reason about their environment. To defend against adversaries that fabricate or manipulate shared data, existing systems employ cross-vehicle inconsistency detection and trust estimation, penalizing vehicles whose observations conflict with the majority. In this work, we show that these defenses themselves introduce a new attack surface. We present TrustFlip, a novel attack that weaponizes consistency-based defenses to poison the trust assigned to benign vehicles. Instead of injecting false data into the collaboration pipeline, it deploys physical adversarial objects that are genuine but induce inconsistent observations among benign vehicles. The resulting inconsistencies are misattributed by the defense to the targeted vehicle, causing its trust score to degrade and eventually leading to its downweighting or exclusion from collaboration. Consequently, the system loses reliable sensing contributors, degrading perception capability and potentially inducing safety-critical failures. We evaluate TrustFlip across multiple collaborative perception architectures and defense mechanisms. Our results show that state-of-the-art defenses can be significantly affected: the attack removes the targeted benign vehicle from collaboration in up to 87.7% of scenarios and drops Average Precision (AP) by up to 13%. As an initial mitigation, we introduce TrustReflect, a lightweight self-reflection mechanism that marks disputed regions as uncertain and excludes them from trust evaluation, reducing the attack success rate by 35-100%.", "date": "2026-05-21", "categories": []}
{"id": "2605.22273", "title": "Exposing Vulnerabilities in Visible-Infrared VLMs: A Unified Geometric Adversarial Framework with Cross-Task Transferability", "abstract": "Vision-language models (VLMs) have achieved strong performance across diverse multimodal tasks, but their adversarial robustness in visible-infrared (VIS-IR) scenarios remains underexplored. This gap is critical because VIS-IR sensing is widely used in real-world perception systems to support reliable understanding under challenging imaging conditions. To address this cross-modal threat setting, we propose CFGPatch, a curved-edge fractal geometric adversarial patch framework for attacking VIS-IR VLMs. CFGPatch builds on triangular fractal geometry and replaces rigid straight-edged primitives with Bezier-curved elements, preserving multi-scale fractal self-similarity while introducing smoother contours, richer directional variation, and more flexible shape deformation. In addition, we design a modality-specific Fraser-spiral rendering mechanism to inject fine-grained texture distortions and misleading perceptual cues into visible and infrared images. By coupling global curved-fractal geometry with local spiral-based appearance interference, CFGPatch disrupts both shape perception and texture interpretation. We further adopt expectation over transformation (EOT) to improve robustness against common image-level transformations. Extensive experiments show that CFGPatch effectively fools VIS-IR VLMs and consistently outperforms standard patch baselines in attack effectiveness and robustness. Moreover, adversarial samples optimized for zero-shot classification transfer well to image captioning and visual question answering, demonstrating strong cross-task transferability and generalizability across downstream tasks.", "date": "2026-05-21", "categories": []}
{"id": "2412.14738", "title": "Spectrally unstable nodes drive reliability failures in graph learning", "abstract": "Graph-learning algorithms can fail when graph structure is adversarially perturbed, intrinsically noisy or constructed from imperfect observations. Here we show that some nodes bear much greater responsibility than others for allowing adversarial perturbations and intrinsic noise to harm graph-learning algorithms. Building on graph-spectral distortion analysis, we identify these failure-driving nodes and introduce a reliability-aware intervention that isolates them from the main learning step. The target algorithm is applied to a stable induced subgraph, and predictions for isolated nodes are recovered through topology- or centroid-based propagation. Across graph neural networks under targeted and non-targeted structural attacks, spectral hypergraph clustering and multi-view spectral clustering, this principle improves reliability under both adversarial and intrinsic noise. These results suggest that node-level spectral instability provides a common mechanism for understanding and mitigating reliability failures in graph learning.", "date": "2024-12-19", "categories": []}
{"id": "2605.20308", "title": "SDM: A Powerful Tool for Evaluating Model Robustness", "abstract": "Gradient-based attacks are important methods for evaluating model robustness. However, since the proposal of APGD, it has been difficult for such methods to achieve significant breakthroughs. To achieve such an effect, we first analyze the issue of \"high-loss non-adversarial examples\" that degrades attack performance in previous methods, and prove that this issue arises from inappropriate objectives for adversarial example generation. Subsequently, we reconstruct the objective as \"maximizing the difference between the non-ground-truth label probability upper bound and the ground-truth label probability\", and proposes a novel and powerful gradient-based attack method named Sequential Difference Maximization (SDM). SDM establishes a three-layer optimization framework of \"cycle-stage-step\". It adopts the negative probability loss function and the Directional Probability Difference Ratio (DPDR) loss function in the initial and subsequent optimization stages, respectively, and approaches the ideal objective of adversarial example generation via stage-wise sequential optimization. Experiments demonstrate that compared with previous state-of-the-art methods, SDM not only achieves stronger attack performance but also exhibits superior cost-effectiveness. The code is available at https://github.com/X-L-Liu/ICML-SDM.", "date": "2026-05-19", "categories": []}
{"id": "2605.21146", "title": "Detecting Trojaned DNNs via Spectral Regression Analysis", "abstract": "Modern DNNs are repeatedly fine-tuned to incorporate new data and functionality. This evolutionary workflow introduces a security risk when updated data cannot be fully trusted, as adversaries may implant Trojans during fine-tuning. We present MIST, a Trojan detection approach that analyzes how a model's internal representations change during fine-tuning. Rather than attempting to reconstruct trigger conditions, MIST characterizes benign model evolution using pre-activation spectra and flags updates whose spectral deviations are inconsistent with this reference. This framing treats Trojan detection as a regression problem over model updates. An empirical evaluation across four datasets and eight Trojan attacks shows that spectral distances reliably distinguish Trojaned updates from clean fine-tuning. MIST outperforms state-of-the-art detection accuracy after a single update, without requiring any knowledge about the poisoned data or the trigger, and remains effective under multi-step benign evolution, with graceful and bounded degradation. These results indicate that spectral evolution provides a stable and assumption-light signal for detecting malicious model updates.", "date": "2026-05-20", "categories": []}
{"id": "2605.21362", "title": "LASH: Adaptive Semantic Hybridization for Black-Box Jailbreaking of Large Language Models", "abstract": "Jailbreak attacks expose a persistent gap between the intended safety behavior of aligned large language models and their behavior under adversarial prompting. Existing automated methods are increasingly effective but each commits to a single attack family (e.g., one refinement loop, one tree search, one mutation space, or one strategy library) and no single family dominates: the best-performing method shifts across target models and harm categories, suggesting complementary strengths that per-prompt composition could exploit. We introduce LASH (LLM Adaptive Semantic Hybridization), a black-box framework that treats outputs from multiple base attacks as reusable seed prompts and adaptively composes them for each target request. Given a seed pool, LASH searches over seed subsets and softmax-normalized mixture weights; a composition module synthesizes a single candidate prompt, and a derivative-free genetic optimizer updates the weights using black-box target feedback and a two-stage fitness function combining keyword-based refusal detection with LLM-judge scoring. On JailbreakBench, which contains 100 harmful prompts across 10 categories, we evaluate LASH on six common target models. LASH achieves an average attack success rate of 84.5% under keyword-based evaluation and 74.5% under two-stage evaluation, where responses are first filtered for refusals and then scored by an LLM judge for whether they substantively fulfill the original harmful request. LASH outperforms five state-of-the-art baselines on both metrics with only 30 mean target queries. LASH also remains competitive under three defense mechanisms and induces more success-like internal representations. These results suggest that adaptive composition across heterogeneous jailbreak strategies is a promising direction for black-box red-teaming.", "date": "2026-05-20", "categories": []}
{"id": "2412.00404", "title": "Hard-Label Black-Box Attacks on 3D Point Clouds", "abstract": "With the maturity of depth sensors in various 3D safety-critical applications, 3D point cloud models have been shown to be vulnerable to adversarial attacks. Almost all existing 3D attackers simply follow the white-box or black-box setting to iteratively update coordinate perturbations based on back-propagated or estimated gradients. However, these methods are hard to deploy in real-world scenarios (no model details are provided) as they severely rely on parameters or output logits of victim models. To this end, we propose point cloud attacks from a more practical setting, i.e., hard-label black-box attack, in which attackers can only access the prediction label of 3D input. We introduce a novel 3D attack method based on a new spectrum-aware decision boundary algorithm to generate high-quality adversarial samples. In particular, we first construct a class-aware model decision boundary, by developing a learnable spectrum-fusion strategy to adaptively fuse point clouds of different classes in the spectral domain, aiming to craft their intermediate samples without distorting the original geometry. Then, we devise an iterative coordinate-spectrum optimization method with curvature-aware boundary search to move the intermediate sample along the decision boundary for generating adversarial point clouds with trivial perturbations. Experiments demonstrate that our attack competitively outperforms existing white/black-box attackers in terms of attack performance and adversary quality.", "date": "2024-11-30", "categories": []}
{"id": "2601.16200", "title": "Feature-Space Smoothing: Certified Robustness of Deep Representations", "abstract": "Modern deep learning models exhibit strong capabilities across diverse applications, yet remain vulnerable to malicious inputs that induce erroneous predictions via feature-space distortion. To address this vulnerability, we propose Feature-space Smoothing (FS), a general defense framework that provides certified robustness at the feature representation level. We show that FS converts a given feature encoder into a smoothed variant that is guaranteed to maintain a certified lower bound on the cosine similarity between clean and adversarial features under l_2-bounded perturbations. We then establish that this Feature Cosine Similarity Bound (FCSB) can be extended to the prediction-wise certification under the cosine similarity measure, and the value of FCSB is determined by the encoder intrinsic Gaussian robustness score. Building on those insights, we introduce the Gaussian Smoothness Booster (GSB), a plug-and-play module to improve the encoder Gaussian robustness score. Specifically, the GSB module is plugged to enhance the feature-space consistency and maintain the feature utility for downstream tasks under Gaussian perturbations. This design enables seamless integration of FS on the protected model, e.g., Multimodal Large Language Models (MLLMs), without additional model retraining or alignment, improving its robustness while preserving the performance for downstream task-oriented decoding. Extensive experiments demonstrate that integrating FS consistently provides non-trivial certified robustness and significantly improves task-oriented performance under strong white-box adversarial attacks across diverse models and applications.", "date": "2026-01-22", "categories": []}
{"id": "2601.22569", "title": "Whispers of Wealth: Red-Teaming Google's Agent Payments Protocol via Prompt Injection", "abstract": "Large language model (LLM) based agents are increasingly used to automate financial transactions, yet their reliance on contextual reasoning exposes payment systems to prompt-driven manipulation. The Agent Payments Protocol (AP2) aims to secure agent-led purchases through cryptographically verifiable mandates, but its practical robustness remains underexplored. In this work, we perform an AI red-teaming evaluation of AP2 and identify vulnerabilities arising from indirect and direct prompt injection. We introduce two attack techniques, the Branded Whisper Attack and the Vault Whisper Attack which manipulate product ranking and extract sensitive user data. Using a functional AP2 based shopping agent built with Gemini-2.5-Flash and the Google ADK framework, we experimentally validate that simple adversarial prompts can reliably subvert agent behavior. Our findings reveal critical weaknesses in current agentic payment architectures and highlight the need for stronger isolation and defensive safeguards in LLM-mediated financial systems.", "date": "2026-01-30", "categories": []}
{"id": "2605.18821", "title": "Quantum Adversarial Machine Learning: From Classical Adaptations to Quantum-Native Methods", "abstract": "Machine learning has revolutionized numerous industrial domains. Despite recent advances, machine learning models remain vulnerable to adversarial threats. Adversarial machine learning is a field that studies these vulnerabilities to build robust machine learning models. Quantum machine learning is an interdisciplinary field that bridges quantum computing and classical machine learning. While quantum machine learning shows potentials to outperform classical machine learning in complex tasks such as regression, classification, and generative modeling, it remains vulnerable to adversarial attacks. Given the recent advancements in quantum computing and machine learning, the quantum adversarial machine learning field has emerged to study the vulnerabilities of quantum machine learning, possible attacks, and novel quantum-enhanced defense strategies. In this survey, we provide a detailed overview on quantum adversarial machine learning and explore the existing attacks and countermeasures. We also review the theoretical underpinnings of this area, emerging trends, and critical challenges.", "date": "2026-05-12", "categories": []}
{"id": "2605.18868", "title": "DarkLLM: Learning Language-Driven Adversarial Attacks with Large Language Models", "abstract": "While vision and multimodal foundation models underpin critical tasks from perception to complex reasoning, they remain highly vulnerable to adversarial attacks. However, traditional adversarial attacks are typically limited to single, predefined objectives, tightly coupling each attack to a specific model or task, which restricts their scalability and flexibility in real-world scenarios. In this work, we present DarkLLM, a novel attack framework that trains an LLM to translate natural-language attack instructions into latent attack vectors, which are then decoded into visual adversarial perturbations. By leveraging natural-language instruction tuning, DarkLLM not only unifies targeted, untargeted, segmentation, and multi-model attacks within a single framework, but also achieves flexible and controllable adversarial generation, enabling each instruction to produce a perturbation that induces desired behaviors across heterogeneous models. Through extensive experiments across 4 tasks, 13 datasets, and 15 models, we demonstrate that DarkLLM with only 1B parameters can follow attacker instructions and generate highly effective attacks against CLIP, SAM, and frontier LLMs, revealing a systemic vulnerability in modern foundation models.", "date": "2026-05-15", "categories": []}
{"id": "2605.18919", "title": "MoCo-EA: Exploiting Adversarial Mode Connectivity for Efficient Evolutionary Attacks", "abstract": "Evolutionary algorithms for adversarial attacks leverage population-based search to discover perturbations without gradient information, but suffer from inefficient crossover operations that destroy adversarial properties through discrete interpolation. We introduce Mode Connectivity Evolutionary Attack (MoCo-EA), which replaces traditional crossover with a novel B\\'ezier crossover operator that optimizes perturbations along a continuous B\\'ezier curve between parent perturbations. Our key insight is that adversarial examples lie on connected manifolds where intermediate points maintain and often enhance attack effectiveness. We demonstrate three findings: (1) Successful adversarial perturbations exhibit mode connectivity; (2) Intermediate points along optimized paths achieve higher transferability than endpoints; (3) B\\'ezier crossover dramatically outperforms discrete genetic operations while reducing convergence time and query requirements. By exploiting the geometric structure of adversarial space through path optimization, MoCo-EA provides an efficient and reliable method. Our work challenges the traditional view of adversarial examples as isolated points and opens new directions for both attack generation and defense research.", "date": "2026-05-18", "categories": []}
{"id": "2605.18988", "title": "Surviving the Unseen: Predictive Defense for Novel Multi-Turn Multimodal Attacks", "abstract": "The expansion of Multimodal Large Language Models (MLLMs) and their integration into autonomous agentic workflows has introduced a non-stationary attack surface. Empirical observations indicate that adversaries employ progressive, cross-modal perturbations that evade turn-specific guardrails by distributing malicious intent across longitudinal conversational trajectories. Static defense mechanisms, constrained by the Markov property, evaluate inputs in isolation and fail to detect cumulative structural poisoning. To handle this limitation, this paper formulates safety verification as a dynamic survival prediction and trajectory dynamics problem. The Triple-tier Anomaly Defense (TRIAD) framework is proposed as a predictive model that maps multimodal and multi-turn conversational flow as a continuous trajectory. The framework integrates structural anomaly detection to monitor covariance shifts, a Ledoit-Wolf regularized Mahalanobis distance to monitor covariance shifts in high-dimensional spaces, and topological trajectory acceleration to differentiate benign creative exploration from continuous malicious drift. These kinematic and geometric features are integrated into a time-varying Cox Proportional Hazards model via a Bayesian Hidden Markov Model (HMM) feedback loop. Theoretical analysis demonstrates that the TRIAD framework provides a mathematically bounded expected time-to-failure under adversarial perturbations, ensuring that malicious acceleration diverges positively. This framework provides a computationally efficient, interpretable, and predictive safeguard for real-time agentic AI systems, establishing a rigorous foundation for continuous safety alignment without relying on empirical retraining.", "date": "2026-05-18", "categories": []}
{"id": "2605.19377", "title": "The Evaluation Game: Beyond Static LLM Benchmarking", "abstract": "As jailbreaks, adversarially crafted inputs that bypass safety constraints, continue to be discovered in Large Language Models, practitioners increasingly rely on fine-tuning as a defensive strategy. Yet the theoretical foundations underlying this robustness fine-tuning remain underexplored. We introduce a game-theoretic framework in which the interaction between an evaluator (auditing the model for jailbreaks) and a trainer is formalized as a two-player game.\n  A key feature of our approach is the use of group actions, a mathematical structure that captures symmetries and transformations, to formally represent data augmentation. The simplest non-trivial instance is the circle with cyclic translation groups, where we exhibit various regimes depending on the trainer's generalization range. Below a critical threshold, the evaluator maintains a constant miss ratio for linearly many rounds, whereas other settings can yield very different behaviors. We further provide empirical evidence supporting locality-dependence of the model: for the three model families we tested (Llama, Qwen and Mistral), we have significant evidence that fine-tuning on adversarial prompts induces only local generalization, with refusal rates on test examples highly correlated with the distance to the fine-tuning prompts. Our framework recasts the central object of adversarial evaluation: a benchmark is not a static set of prompts but an orbit under the evaluator's group action, and audit protocols that ignore trainer-side adaptation cannot distinguish a genuine fix from a memorized patch.", "date": "2026-05-19", "categories": []}
{"id": "2605.19446", "title": "Targeted Downstream-Agnostic Attack", "abstract": "Recently, pre-trained encoders have gained widespread use due to their strong capability in representation extraction. However, they are vulnerable to downstream-agnostic attacks (DAAs). Existing DAA methods operate under a permissive threat model, where an attack is successful if the generated downstream-agnostic adversarial examples (DAEs) change the original prediction, without requiring a specific target. In this paper, we propose a Targeted DAA (TDAA) method under a stricter threat model requiring the attack to be both targeted and downstream-agnostic. Since the downstream task is unknown and encoders do not directly produce predictions, achieving a targeted attack is particularly challenging. To address this, we introduce a novel component termed the 'threat image', pre-selected by the attacker as the target. Specifically, a generator is designed to produce example-specific adversarial perturbations that compel the victim encoder to output identical features for both the DAEs and the threat image. Unlike previous DAA methods that generate a single shared perturbation for all samples, which often fails due to image diversity, our method adopts an example-specific paradigm. This generates tailored perturbations for each image to ensure a high attack success rate and invisibility. By leveraging the threat image as a feature-level anchor, our method builds a task-agnostic bridge to reveal the vulnerabilities of the victim encoder. Extensive experiments on 10 self-supervised methods across 3 benchmark datasets demonstrate the effectiveness of our approach and reveal the pronounced vulnerability of pre-trained encoders. The code will be made publicly available after the review period.", "date": "2026-05-19", "categories": []}
{"id": "2605.19478", "title": "Exposing Functional Fusion: A New Class of Strategic Backdoor in Dynamic Prompt Architectures", "abstract": "Existing ViT backdoor attacks based on backbone-overwriting full-tuning are computationally expensive and inflict performance degradation. This has forced adversaries towards the Visual Parameter-Efficient Fine-Tuning (PEFT) paradigm, dominated by adapter-based (e.g., LoRA) and prompt-based (e.g., VPT) approaches. While adapter security has seen initial study, the risks of the burgeoning prompt-based ecosystem remain critically unexplored. We fill this critical gap, exposing how the evolution of VPT towards dynamic and context-aware architectures can facilitate a far more dangerous and emergent threat. This vulnerability arises even though these dynamic modules unlock superior benign performance. We propose VIPER, an attack framework built on a lightweight, dynamic Visual Prompt Generator (VPG) that demonstrates this vulnerability. Critically, this dynamic architecture enables Functional Fusion: an emergent phenomenon where malicious logic and benign task utility are tightly fused into the same sparse, high-magnitude parameter core. This fusion creates a formidable ``hostage\" dilemma, as pruning the attack necessarily destroys the benign performance. Comprehensive evaluations show VIPER effectively addresses the attacker's trilemma: VIPER not only achieves state-of-the-art performance on clean data, but also maintains near-100% ASR even under 90% VPG-module pruning (where LoRA attacks collapse), while adding only an imperceptible 0.06ms (1.16%) of inference latency. VIPER's results, driven by Functional Fusion, expose a new, paradigm-level risk in dynamic prompt architectures.", "date": "2026-05-19", "categories": []}
{"id": "2605.19698", "title": "Awakening the Hydra: Stabilizing Multi-Concept Backdoor Injection in Text-to-Image Diffusion Models", "abstract": "Text-to-image diffusion models are increasingly developed through open-source reuse and repeated downstream fine-tuning, where reused checkpoints are difficult to verify and thus more susceptible to hidden backdoor behaviors. In such ecosystems, a single pretrained model may be sequentially adapted and redistributed by multiple independent parties, allowing multiple concept-specific trigger-target associations to accumulate in the same model. When these associations coexist, semantic conflicts can be amplified in the shared representation space, leading to cross-concept entanglement and degraded generation quality. Notably, instead of strengthening the attack, such accumulation can destabilize previously injected behaviors and reduce attack reliability.\n  In this work, we systematically investigate backdoor attacks under this interference-prone setting and propose Hydra, a unified framework for robust and controlled multi-concept backdoor injection under cumulative and decentralized reuse. Our core insight is that stable backdoor injection under large-scale multi-concept settings requires explicitly constraining trigger semantics while coordinating cross-task interactions during optimization. Specifically, Hydra performs evolutionary trigger search in the text encoder space to identify triggers that are semantically aligned with their target concepts while remaining stable across other injected concepts. It further combines multi-task fine-tuning with trigger-clean regularization to improve training stability under dense multi-concept injection. Extensive experiments across multiple diffusion backbones under rigorous multi-concept settings show that Hydra maintains effective backdoor activation while preserving clean generation fidelity and image quality. For instance, across 8 attackers and 500 concept pairs, Hydra maintains ~95% ASR and strong clean generation.", "date": "2026-05-19", "categories": []}
{"id": "2605.19966", "title": "Detecting Fluent Optimization-Based Adversarial Prompts via Sequential Entropy Changes", "abstract": "Optimization-based adversarial suffixes can jailbreak aligned large language models (LLMs) while remaining fluent, weakening static and windowed perplexity-based detectors. We cast adversarial suffix detection as an online change-point detection problem over the token-level next-token entropy stream. Using the LLM system prompt to estimate a robust baseline, we standardize user-token entropies and apply a one-sided CUSUM statistic. The resulting detector, CPD Online (CPD), is model-agnostic, training-free, runs online, and localizes the adversarial suffix onset. On a benchmark of 1,012 optimization-based suffix attacks (GCG, AutoDAN, AdvPrompter, BEAST, AutoDAN-HGA) and 1,012 perplexity-controlled benign prompts, CPD improves F1 over the strongest windowed-perplexity baseline on all six open-weight chat models (LLaMA-2-7B/13B, Vicuna-7B/13B, Qwen2.5-7B/14B). On LLaMA-2-7B at the canonical CUSUM setting ($k=0$), CPD reaches AUROC $0.88$ and F1 $0.82$. Beyond prompt-level detection, CPD concentrates 79.6% of its triggers inside the adversarial suffix, versus 17-46% for windowed perplexity. Finally, when used as a lightweight gate for LLaMA Guard, CPD reduces guard calls by 17-22% on a high-volume, benign-dominated deployment while preserving guard-level detection quality", "date": "2026-05-19", "categories": []}
{"id": "2605.20037", "title": "When Critics Disagree: Adaptive Reward Poisoning Attacks in RIS-Aided Wireless Control System", "abstract": "Reward-poisoning attacks present a significant risk to learning-based wireless control systems. Given this, we propose a Disagreement-Guided Reward Poisoning (DGRP) adaptive attack on a Soft Actor-Critic (SAC) agent. In a Cognitive Radio Network (CRN) environment assisted by Reconfigurable Intelligent Surfaces (RIS), the SAC agent is tasked with maximizing the long-term secondary users' (SUs) rate by simultaneously optimizing the transmission power of the SU transmitter and the RIS phase shifts. DGRP corrupts rewards, particularly when the SAC dual critics exhibit substantial disagreement-especially in high-leverage, high-uncertainty states-resulting in distorted value estimations and guiding the policy towards suboptimal actions. Our findings demonstrate that DGRP substantially diminishes the performance improvements typically provided by RIS and degrades transmission quality. We further investigate key attack parameters and determine their impact on learning. In comparison to periodic-timing and exploration-triggered baselines, DGRP consistently causes greater damage, highlighting the necessity of considering disagreement-aware threats when evaluating the robustness of Deep Reinforcement Learning (DRL) in RIS-assisted networks.", "date": "2026-05-19", "categories": []}
{"id": "2505.02360", "title": "Catastrophic Overfitting, Entropy Gap and Participation Ratio: A Noiseless $l^p$ Norm Solution for Fast Adversarial Training", "abstract": "Adversarial training is a cornerstone of robust deep learning, but fast methods like the Fast Gradient Sign Method (FGSM) often suffer from Catastrophic Overfitting (CO), where models become robust to single-step attacks but fail against multi-step variants. While existing solutions rely on noise injection, regularization, or gradient clipping, we propose a novel solution that purely controls the $l^p$ training norm to mitigate CO. Our study is motivated by the empirical observation that CO is more prevalent under the $l^{\\infty}$ norm than the $l^2$ norm. Leveraging this insight, we develop a framework for generalized $l^p$ attack as a fixed point problem and craft $l^p$-FGSM attacks to understand the transition mechanics from $l^2$ to $l^{\\infty}$. This leads to our core insight: CO emerges when highly concentrated gradients where information localizes in few dimensions interact with aggressive norm constraints. By quantifying gradient concentration through Participation Ratio and entropy measures, we develop an adaptive $l^p$-FGSM that automatically tunes the training norm based on gradient information. Extensive experiments demonstrate that this approach achieves strong robustness without requiring additional regularization or noise injection, providing a novel and theoretically-principled pathway to mitigate the CO problem.", "date": "2025-05-05", "categories": []}
{"id": "2510.10140", "title": "Adversarial Attacks on Downstream Weather Forecasting Models: Application to Tropical Cyclone Trajectory Prediction", "abstract": "Deep learning-based weather forecasting (DLWF) models leverage past weather observations to generate future forecasts, supporting a wide range of downstream applications, including tropical cyclone (TC) prediction. In this paper, we investigate their vulnerability to adversarial attacks, where subtle perturbations to the upstream forecasts can alter the downstream TC trajectory predictions. Although research into adversarial attacks on DLWF models has grown recently, it remains challenging to craft perturbed upstream forecasts that steer the downstream outputs toward attacker-specified trajectories. First, conventional TC detection systems are opaque, non-differentiable black boxes, making standard gradient-based attacks infeasible. Second, the extreme rarity of TC events leads to severe class imbalance problem, making it difficult to develop attack methods for perturbing upstream forecasts that produce realistic-looking cyclone paths aligned with attacker's target trajectories. To overcome these limitations, we propose Cyc-Attack, a novel method for perturbing the upstream forecasts of DLWF models to generate adversarial trajectories. The proposed method uses a differentiable surrogate model to approximate the TC detector's output, enabling the application of gradient-based attacks. Cyc-Attack also employs a skewness-aware loss function with kernel dilation strategy to address the imbalance problem. Finally, a distance-based gradient weighting scheme and regularization are used to constrain the perturbations and eliminate unrealistic-looking trajectories, thereby making the adversarial upstream forecasts less easily detectable. Our experiments show that Cyc-Attack achieves a higher true positive rate in matching the attacker's target trajectories, along with lower false alarm rates and stealthier perturbations than conventional attack methods.", "date": "2025-10-11", "categories": []}
{"id": "2601.16880", "title": "Theory of Minimal Weight Perturbations in Deep Networks and its Applications for Low-Rank Activated Backdoor Attacks", "abstract": "The minimal norm weight perturbations of DNNs required to achieve a specified change in output are derived and the factors determining its size are discussed. These single-layer exact formulae are contrasted with more generic multi-layer Lipschitz constant based robustness guarantees; both are observed to be of the same order which indicates similar efficacy in their guarantees. These results are applied to precision-modification-activated backdoor attacks, establishing provable compression thresholds below which such attacks cannot succeed, and show empirically that low-rank compression can reliably activate latent backdoors while preserving full-precision accuracy. These expressions reveal how back-propagated margins govern layer-wise sensitivity and provide certifiable guarantees on the smallest parameter updates consistent with a desired output shift.", "date": "2026-01-23", "categories": []}
{"id": "2601.21531", "title": "On the Adversarial Robustness of Large Vision-Language Models under Visual Token Compression", "abstract": "Visual token compression is widely used to accelerate large vision-language models (LVLMs) by pruning or merging visual tokens, yet its adversarial robustness remains unexplored. We show that existing encoder-based attacks cannot fully disclose the robustness vulnerabilities of compressed LVLMs, due to an optimization-inference mismatch: perturbations are optimized on the full-token representation, while inference is performed through a token-compression bottleneck. To address this gap, we propose the Compression-AliGnEd attack (CAGE), which aligns perturbation optimization with compression inference without assuming access to the deployed compression mechanism or its token budget. CAGE combines (i) expected feature disruption, which concentrates distortion on tokens likely to survive across plausible budgets, and (ii) rank distortion alignment, which actively aligns token distortions with rank scores to promote the retention of highly distorted evidence. Across diverse representative plug-and-play compression mechanisms and datasets, our results show that CAGE consistently achieves lower robust accuracy than the baseline. This work highlights that robustness assessments ignoring compression can be overly optimistic, calling for compression-aware security evaluation and defenses for efficient LVLMs.", "date": "2026-01-29", "categories": []}
{"id": "2605.14005", "title": "Mistletoe: Stealthy Acceleration-Collapse Attacks on Speculative Decoding", "abstract": "Speculative decoding has become a widely adopted technique for accelerating large language model (LLM) inference by drafting multiple candidate tokens and verifying them with a target model in parallel. Its efficiency, however, critically depends on the average accepted length $\\tau$, i.e., how many draft tokens survive each verification step. In this work, we identify a new mechanism-level vulnerability in model-based speculative decoding: the drafter is trained to approximate the target model distribution, but this approximation is inevitably imperfect. Such a drafter-target mismatch creates a hidden attack surface where small perturbations can preserve the target model's visible behavior while substantially reducing draft-token acceptability. We propose Mistletoe, a stealthy acceleration-collapse attack against speculative decoding. Mistletoe directly targets the acceptance mechanism of speculative decoding. It jointly optimizes a degradation objective that decreases drafter-target agreement and a semantic-preservation objective that constrains the target model's output distribution. To resolve the conflict between these objectives, we introduce a null-space projection mechanism, where degradation gradients are projected away from the local semantic-preserving direction, suppressing draft acceptance while minimizing semantic drift. Experiments on various speculative decoding systems show that Mistletoe substantially reduces average accepted length $\\tau$, collapses speedup, and lowers averaged token throughput, while preserving output quality and perplexity. Our work highlights that speculative decoding introduces a mechanism-level attack surface beyond existing output robustness, calling for more robust designs of LLM acceleration systems.", "date": "2026-05-13", "categories": []}
{"id": "2605.16286", "title": "Homoglyph-based Adversarial Perturbation of Introductory Computer Science Theory Problems", "abstract": "Different AI tools such as ChatGPT, Gemini, and Claude are becoming very popular. Although they are helpful for many day-to-day tasks, they can be used in unexpected ways. For example, the learning objectives of a course may not be achieved if students use these tools to solve their homework problems. This paper proposes a simple method to address this issue in the lazy student model. The method uses homoglyph-based adversarial perturbation to first modify the question without changing the semantic meaning of the question. Then a few characters are perturbed by their homoglyphs. Our experimental result shows the theoretical problems of introductory computer science courses can be effectively perturbed. We also propose an interactive tool to conveniently use our method.", "date": "2026-04-13", "categories": []}
{"id": "2605.16312", "title": "When Actions Disappear: Adversarial Action Removal in Self-Play Reinforcement Learning", "abstract": "We study adversarial action masking in self-play reinforcement learning: an attacker selectively removes legal actions from a victim's action set. Unlike observation or action perturbations, removal eliminates decision options before the agent acts. Across poker games scaling from 6 to 5,531 information states and two non-poker domains, learned masking causes substantially more damage than random masking and learned perturbation baselines. The attack persists across Q-learning, PPO, NFSP, neural NFSP, and DQN victims; transfers across agents; is amplified by self-play; and shows no recovery under extended masked training. Mechanistically, the adversary targets high-value decision points, captured by reach-weighted contingent action capacity (CAC$_w$) and a value-weighted refinement CAC$_v$. These results identify action availability as a distinct robustness surface in self-play RL.", "date": "2026-05-04", "categories": []}
{"id": "2605.16440", "title": "Semantic Smoothing via Novel View Synthesis for Robust SAR Image Classification", "abstract": "Deep neural networks are vulnerable to adversarial perturbations, limiting deployment in safety-critical applications such as synthetic aperture radar (SAR) automatic target recognition (ATR). Randomized smoothing improves robustness by averaging predictions over noisy inputs, but isotropic noise often fails to preserve the semantic structure of SAR imagery. We propose semantic smoothing, a defense that replaces noised-based perturbations with structured randomized transformations generated by a novel view synthesis model. For SAR, we condition on acquisition geometry to synthesize multiple plausible radar views. Predictions across generated randomized views are aggregated to form a robust classifier. Experiments show that semantic smoothing improves robustness against standard attacks, such as FGSM and PGD, and SAR-specific attacks, such as OTSA and SMGAA, while also increasing clean classification accuracy. These results demonstrate that randomized smoothing via semantically preserving geometric transformations is a promising alternative to isotropic noise for adversarial defense in structured sensing domains.", "date": "2026-05-15", "categories": []}
{"id": "2605.16720", "title": "Compositional Adversarial Training for Robust Visual Watermarking", "abstract": "Robust watermarking is typically trained with random post-processing augmentation, but random sampling under-covers the combinatorial space of realistic attack pipelines and rarely encounters the rare compositions that actually break detection. This leads to unstable training and poor sample efficiency. We instead formulate watermark robustness as a min-max problem over a structured space of compositional transformations. We propose Compositional Adversarial Training (CAT), a plug-in framework that learns a sequential differentiable adversary that observes the current watermarked image and selects an attack family at each step to maximally disrupt message recovery. CAT combines a straight-through Gumbel-Softmax attack selection with entropy regularization, allowing the backward pass to be end-to-end differentiable and aggregate gradient information across attack families, yielding faster, smoother convergence without collapsing to a single attack mode. We evaluate CAT on post-generation watermarks VideoSeal 0.0, VideoSeal 1.0, and PixelSeal and in-generation WMAR under both single-step and two-step attack suites, on in-distribution and multiple out-of-distribution image and video benchmarks. CAT consistently outperforms random-augmentation baselines trained with the same augmentation budget, with the largest gains on hard composed attacks and OOD evaluations; improving overall watermark capacity by up to $63.5\\%$ in the single-step attack setting and $13.0\\%$ in the compositional setting. In the autoregressive setting, CAT improves the TPR@FPR$=1\\%$ by $12\\%$ on average on difficult geometric transformations. These results show that robust visual watermarking benefits from training against adaptive compositional adversaries rather than independent random corruptions.", "date": "2026-05-16", "categories": []}
{"id": "2605.17128", "title": "New Wide-Net-Casting Jailbreak Attacks Risk Large Models", "abstract": "Jailbreak attacks on large models have drawn growing attention due to their close ties to societal safety. This work identifies a practical yet unexplored jailbreak scenario, the wide-net-casting scenario, where an adversary can query a group of large models instead of a single one to elicit harmful outputs. Our analysis reveals substantial yet previously overlooked safety risks under this scenario. As a key part of our analysis, we further develop a novel jailbreak method tailored to the wide-net-casting scenario. With this tailored method, the jailbreak success rate can even reach 100\\% in some experiments when targeting the large models without additional safeguards, exposing wide-net-casting as a distinct, high-risk scenario that warrants attention in future evaluation and defense research.", "date": "2026-05-16", "categories": []}
{"id": "2605.17577", "title": "TAME: Test-Time Adversarial Prompt Tuning via Mixture-of-Experts for Vision-Language Models", "abstract": "Large-scale pre-trained Vision-Language models (VLMs), such as CLIP, exhibit strong zero-shot generalization, yet remain highly vulnerable to imperceptible adversarial perturbations, raising serious safety concerns for open-world deployment. To enhance robustness without requiring downstream task-specific retraining, we propose TAME, a novel test-time defense. Building upon our prior Test-Time Adversarial Prompt Tuning (TAPT), TAME introduces an architectural reformulation by replacing TAPT's single adaptive prompt with an input-conditioned Mixture-of-Experts (MoE) framework, enabling more expressive and adaptive defense. Specifically, TAME maintains a bank of learnable expert prompts and employs an input-dependent routing mechanism to aggregate a customized prompt mixture for each unlabeled test sample at inference time. This test-time defense mechanism is driven by three unsupervised objectives: (1) multi-view prediction entropy minimization, (2) layer-wise alignment of visual token statistics to precomputed clean and adversarial reference distributions, and (3) MoE regularization for balanced expert utilization and prompt diversity. We evaluated TAME on 11 benchmark datasets, including ImageNet and 10 additional zero-shot datasets. The results show that TAME improves the zero-shot adversarial robustness of the original CLIP by at least 49.1% under AutoAttack while largely preserving generalization on clean samples. TAME also consistently outperforms existing adversarial prompt tuning methods across multiple prompt designs, yielding an average robustness gain of at least 30.2%.", "date": "2026-05-17", "categories": []}
{"id": "2605.17772", "title": "Towards Universal Physical Adversarial Attacks via a Joint Multi-Objective and Multi-Model Optimization Framework", "abstract": "Physical adversarial attacks often overfit single surrogate models and optimization objectives. While ensemble attacks can mitigate this, existing methods struggle with severe gradient conflicts within restricted physical texture spaces, significantly degrading cross-model transferability. To bridge this gap, this paper proposes a Joint Multi-Objective and Multi-Model Optimization Framework (JMOF) that leverages quantitative similarity analysis to select the optimal surrogate model ensemble. Within JMOF, a dual-level mechanism jointly suppresses prediction outputs and flattens intermediate feature distributions, balancing attack efficiency with deep generalization. Additionally, an Orthogonal Gradient Alignment (OGA) strategy resolves cross-model gradient conflicts, transforming mutually repulsive gradients into synergistic optimization directions. Extensive simulated and real-world experiments demonstrate that JMOF outperforms state-of-the-art baselines against diverse black-box detectors. Crucially, JMOF exhibits substantial cross-vision-task generalization, generating attacks capable of simultaneously deceiving object detection and semantic segmentation or monocular depth estimation models. This research advances the generalization limits of physical adversarial attacks, providing a robust framework for evaluating visual AI vulnerabilities in real-world deployments.", "date": "2026-05-18", "categories": []}
{"id": "2605.17822", "title": "Unleashing the Representational Power of Fourier Shapes for Attacking Infrared Object Detection", "abstract": "Infrared object detection is crucial for perception in autonomous driving and surveillance but remains vulnerable to physical adversarial attacks. Unlike in the RGB domain, where attacks rely on color texture, infrared attacks must manipulate thermal signatures, making the geometry shape of heat-blocking materials the primary adversarial information carrier. Current shape-based methods suffer from a fundamental trade-off between representational capability and optimization power, limiting their attack effectiveness.In this work, we overcome this dilemma by introducing learnable Fourier shapes to the infrared domain. We utilize an end-to-end differentiable framework where a compact set of Fourier coefficients, defining the shape boundary, is analytically mapped to a pixel-space mask via the winding number theorem. This enables efficient gradient-based optimization to generate potent shapes that cause human targets to evade detection. Extensive digital and physical experiments provide a comprehensive evaluation and validate our superior performance. Our resulting physical patch achieves striking robustness, successfully evading detectors across diverse distances, angles, poses, and individuals, and achieves over 88% attack success rate at distances greater than 25m (conf.=0.5). Code is available at https://github.com/Yongyx99/Fourier-shape-attack.", "date": "2026-05-18", "categories": []}
{"id": "2605.17936", "title": "Universal Adversarial Triggers", "abstract": "Recent works have illustrated that modern NLP models trained for diverse tasks ranging from sentiment analysis to language generation succumb to universal adversarial attacks, a class of input-agnostic attacks where a common trigger sequence is used to attack the model. Although these attacks are successful, the triggers generated by such attacks are ungrammatical and unnatural. Our work proposes a novel technique combining parts-of-speech filtering and perplexity based loss function to generate sensible triggers that are closer to natural phrases. For the task of sentiment analysis on the SST dataset, the method produces sensible triggers that achieve accuracies as low as 0.04 and 0.12 for flipping positive to negative predictions and vice-versa. To build robust models, we also perform adversarial training using the generated triggers that increases the accuracy of the model from 0.12 to 0.48. We aim to illustrate that adversarial attacks can be made difficult to detect by generating sensible triggers, and to facilitate robust model development through relevant defenses.", "date": "2026-05-18", "categories": []}
{"id": "2605.18058", "title": "Threats to Arabic Handwriting Recognition: Investigating Black-Box Adversarial Attacks on embedded ConvNet models", "abstract": "Arabic handwriting recognition (AHR) has made significant progress with deep learning models. AHR research has largely focused on performance, with security receiving little attention. This study provides what appears to be a new line of inquiry by demonstrating the vulnerability of high-performing models to adversarial black-box attacks. The focus on black-box attacks reflects real-world scenarios where the attacker has no prior knowledge of the model architecture. Extensive experiments were conducted on two benchmark AHR datasets containing Arabic handwritten Characters. Results demonstrated the effectiveness of the attacks, with the Pixle attack achieving an attack success rate of 99-100\\% on most models. Other, less aggressive attacks achieved success rates of 50-96\\% across most experiments. Despite the higher attack success rate, the attacks maintain the structural integrity of the characters, rendering them almost imperceptible to the human eye. The findings indicate the higher vulnerability of the studied models to adversarial manipulation. This underscores the need to strengthen efforts to secure these models and ensure their reliability in AHR real-world applications.", "date": "2026-05-18", "categories": []}
{"id": "2605.18593", "title": "Not What You Asked For: Typographic Attacks in Household Robot Manipulation", "abstract": "Open-vocabulary embodied AI agents increasingly rely on vision-language models such as CLIP for object perception and task grounding. However, the shared embedding space that enables this flexibility introduces a structural vulnerability to typographic attacks, where printed text in a physical scene semantically overrides visual judgment. While prior work has quantified this threat in static 2D benchmarks and 3D navigation tasks, its impact on the full Sense-Plan-Act pipeline of household robot manipulation remains unexplored.\n  This work evaluates typographic attacks in a Habitat-based simulation using the HomeRobot benchmark. We introduce a decoupled perception architecture that exposes a frozen CLIP encoder to adversarial stickers while maintaining geometric grounding via DETIC. In a controlled evaluation pool of 59 attributable episodes, the attack achieves an overall Attack Success Rate (ASR) of 67.8%, rising to 70.0% among fully successful episodes, under uncontrolled viewing angles and occlusion with no perceptual optimization.\n  Critically, we find that perceptual errors propagate through the persistent 3D semantic map to produce kinetic failures, defined here as physically executed grasping and transport of the wrong object driven by an adversarially poisoned semantic state. In these cases, the robot physically grasps and delivers the wrong object to a target receptacle. These results establish typographic misclassification as a real, measurable, and physically consequential threat to the safety of modular manipulation pipelines that prior typographic attack research has left unexamined.", "date": "2026-05-18", "categories": []}
{"id": "2605.18624", "title": "Learning to Look Benign: Targeted Evasion of Malware Detectors via API Import Injection", "abstract": "Machine learning-based malware detectors are widely deployed in antivirus and endpoint detection systems, yet their reliance on static features makes them vulnerable to adversarial manipulation. This paper investigates whether a malware sample can be intentionally misclassified as a specific benign software category, not merely as \"not malware\", by adding a small number of Win32 API imports characteristic of that selected category, without removing any existing imports or retraining the detector. We propose a framework centered on a Conditional Variational Autoencoder (CVAE) whose decoder is strictly additive. It can introduce new API calls but never remove existing ones, preserving malware functionality by design. For each malware sample, the framework automatically identifies which benign category it most closely resembles and uses that as the evasion target. A knowledge-distilled differentiable proxy enables gradient-based training against the non-differentiable ensemble detector. Experiments on a six-class dataset of binary Win32 API import vectors extracted from 3,799 Windows executables (five benign categories, one malware class) show that, against a detector achieving 87.5% malware recall, adding just 20 API imports reduces recall to 30%. At k=20, among samples that evaded detection, 99% are classified as the intended target category. The CVAE outperforms both a frequency-based baseline and random selection at every tested injection size (k = 5 to 50). Validation on real PE files submitted to VirusTotal confirms that the attack transfers to commercial static detection engines, with an average 54.5% reduction in flagging engines. These findings expose a concrete vulnerability in API-based malware classifiers and demonstrate that targeted evasion into a chosen benign category is achievable with minimal, functionality-preserving modifications.", "date": "2026-05-18", "categories": []}
{"id": "2605.18666", "title": "A No-Defense Defense Against Gradient-Based Adversarial Attacks on ML-NIDS: Is Less More?", "abstract": "Gradient-based adversarial attacks subtly manipulate inputs of Machine Learning (ML) models to induce incorrect predictions. This paper investigates whether careful architectural choices alone can yield an inherently robust Deep Neural Network (DNN)-based Network Intrusion Detection Systems (NIDS), without any additional explicit defenses. Through thousands of experiments, around 2200, varying network depth, feature dimensionality, activation functions, and dropout across FGSM, PGD, and BIM attacks, we show that shallower networks, reduced feature sets, and ReLU activation consistently and jointly reduce adversarial vulnerability. Moreover, a simple model following this recipe outperforms deeper, fully-featured adversarially trained models, while maintaining near-perfect clean-traffic detection and lower training times. Nevertheless, while less is more, the selection of the right less is what truly matters.", "date": "2026-05-18", "categories": []}
{"id": "2406.18944", "title": "Rethinking and Red-Teaming Protective Perturbation in Personalized Diffusion Models", "abstract": "Personalized diffusion models (PDMs) have become prominent for adapting pre-trained text-to-image models to generate images of specific subjects using minimal training data. However, PDMs are susceptible to minor adversarial perturbations, leading to significant degradation when fine-tuned on corrupted datasets. These vulnerabilities are exploited to create protective perturbations that prevent unauthorized image generation. Existing purification methods attempt to red-team the protective perturbation to break the protection but often over-purify images, resulting in information loss. In this work, we conduct an in-depth analysis of the fine-tuning process of PDMs through the lens of shortcut learning. We hypothesize and empirically demonstrate that adversarial perturbations induce a latent-space misalignment between images and their text prompts in the CLIP embedding space. This misalignment causes the model to erroneously associate noisy patterns with unique identifiers during fine-tuning, resulting in poor generalization. Based on these insights, we propose a systematic red-teaming framework that includes data purification and contrastive decoupling learning. We first employ off-the-shelf image restoration techniques to realign images with their original semantic content in latent space. Then, we introduce contrastive decoupling learning with noise tokens to decouple the learning of personalized concepts from spurious noise patterns. Our study not only uncovers shortcut learning vulnerabilities in PDMs but also provides a thorough evaluation framework for developing stronger protection. Our extensive evaluation demonstrates its advantages over existing purification methods and its robustness against adaptive perturbations.", "date": "2024-06-27", "categories": []}
{"id": "2510.13842", "title": "ADMIT: Few-shot Knowledge Poisoning Attacks on RAG-based Fact Checking", "abstract": "Knowledge poisoning poses a critical threat to Retrieval-Augmented Generation (RAG) systems by injecting adversarial content into knowledge bases, tricking Large Language Models (LLMs) into producing attacker-controlled outputs grounded in manipulated context. Prior work highlights LLMs' susceptibility to misleading or malicious retrieved content. However, real-world fact-checking scenarios are more challenging, as credible evidence typically dominates the retrieval pool. To investigate this problem, we extend knowledge poisoning to the fact-checking setting, where retrieved context includes authentic supporting or refuting evidence. We propose \\textbf{ADMIT} (\\textbf{AD}versarial \\textbf{M}ulti-\\textbf{I}njection \\textbf{T}echnique), a few-shot, semantically aligned poisoning attack that flips fact-checking decisions and induces deceptive justifications, all without access to the target LLMs, retrievers, or token-level control. Extensive experiments show that ADMIT transfers effectively across 4 retrievers, 11 LLMs, and 4 cross-domain benchmarks, achieving an average attack success rate (ASR) of 86\\% at an extremely low poisoning rate of $0.93 \\times 10^{-6}$, and remaining robust even in the presence of strong counter-evidence. Compared with prior state-of-the-art attacks, ADMIT improves ASR by 11.2\\% across all settings, exposing significant vulnerabilities in real-world RAG-based fact-checking systems.", "date": "2025-10-11", "categories": []}
{"id": "2602.04003", "title": "When AI Persuades: Adversarial Explanation Attacks on Human Trust in AI-Assisted Decision Making", "abstract": "Most adversarial threats in artificial intelligence (AI) target the computational behavior of models rather than the humans who rely on them. Yet modern AI systems increasingly operate within human decision loops, where users interpret and act on model recommendations. Large Language Models (LLMs) generate fluent natural-language explanations that shape how users perceive and trust AI outputs, revealing a new attack surface at the cognitive layer: the communication channel between AI and its users. We introduce adversarial explanation attacks (AEAs), where an attacker manipulates the framing of LLM-generated explanations to modulate human trust in incorrect outputs. We formalize this behavioral threat through the trust miscalibration gap, a metric that captures the difference in human trust between benign and adversarial explanations. Using this metric as a lens, we highlight a behavioral risk where persuasive explanation framing can preserve user trust even when the underlying AI prediction is wrong. To characterize this threat, we conducted a human study with over 200 participants, systematically varying four dimensions of explanation framing: reasoning mode, evidence type, communication style, and presentation format. Our findings show that users report nearly identical trust for adversarial and benign explanations, with adversarial explanations preserving the vast majority of benign trust despite being incorrect. The most vulnerable cases arise when AEAs closely resemble expert communication, combining authoritative evidence, neutral tone, and domain-appropriate reasoning. Vulnerability is highest on hard tasks, in fact-driven domains, and among participants who are less formally educated, younger, or highly trusting of AI.", "date": "2026-02-03", "categories": []}
{"id": "2603.13864", "title": "Inevitable Encounters: Backdoor Attacks Involving Lossy Compression", "abstract": "Real-world backdoor attacks often require poisoned datasets to be stored and transmitted before being used to compromise deep learning systems. However, in the era of big data, the inevitable use of lossy compression poses a fundamental challenge to invisible backdoor attacks. We find that triggers embedded in RGB images often become ineffective after the images are lossily compressed into binary bitstreams (e.g., JPEG files) for storage and transmission. As a result, the poisoned data lose its malicious effect after compression, causing backdoor injection to fail. In this paper, we highlight the necessity of explicitly accounting for the lossy compression process in backdoor attacks. This requires attackers to ensure that the transmitted binary bitstreams preserve malicious trigger information, so that effective triggers can be recovered in the decompressed data. Building on the region-of-interest (ROI) coding mechanism in image compression, we propose two poisoning strategies tailored to inevitable lossy compression. First, we introduce Universal Attack Activation, a universal method that uses sample-specific ROI masks to reactivate trigger information in binary bitstreams for learned image compression (LIC). Second, we present Compression-Adapted Attack, a new attack strategy that employs customized ROI masks to encode trigger information into binary bitstreams and is applicable to both traditional codecs and LIC. Extensive experiments demonstrate the effectiveness of both strategies.", "date": "2026-03-14", "categories": []}
{"id": "2605.15249", "title": "Enabling Adversarial Robustness in AI Models through Kubeflow MLOps", "abstract": "AI models are increasingly deployed in cloud-native environments to support scalable and automated services. However, while platforms such as Kubernetes provide strong infrastructure orchestration, security mechanisms specifically designed to protect deployed AI models remain limited. This paper presents security measures for AI models deployed in Kubernetes clusters. The proposed architecture integrates Kubeflow-based MLOps to automatically detect adversarial attacks during the inference phase and trigger defense mechanisms that preserve the model's accuracy and reliability. Specifically, a Fast Gradient Sign Method (FGSM) attack is applied at inference time, and a Projected Gradient Descent (PGD)-based adversarial training defense is automatically deployed when a degradation in accuracy is detected. The experimental results indicate that the deployed defense robustifies the model, significantly recovering accuracy relative to the degradation caused by the attack.", "date": "2026-05-14", "categories": []}
{"id": "2605.16090", "title": "A Cross-Modal Prompt Injection Attack against Large Vision-Language Models with Image-Only Perturbation", "abstract": "Large vision-language models (LVLMs) have emerged as a powerful paradigm for multimodal intelligence, but their growing deployment also expands the attack surface of prompt injection. Despite this growing concern, existing attacks still suffer from a critical limitation: the injected prompt for one modality only steers the model's interpretation of that singular input. Alternatively, these attacks remain multimodal but fail to achieve cross-modal prompt perturbation. To bridge this gap, we introduce a novel cross-modal prompt injection attack CrossMPI, which can steer the model's interpretation of both textual and visual inputs via image-only prompt injection. Our design is underpinned by the following key breakthroughs. First, we turn the focus of the injected prompt perturbation optimization from the visual embedding space (typically with only $10^5$ parameters) to the model hidden state space (for multimodal information integration and with $10^7$ parameters). Then, two strategies are adopted to mitigate the optimization challenges posed by the larger parameter space. To constrain the optimized model parameter space, we introduce a layer selection strategy that identifies the layers most critical to multimodal integration. Interestingly, deviating from the past experience, our analysis reveals that the optimal layers for LVLM prompt perturbation reside in the middle of the model rather than the last. To constrain the image perturbation space, we propose a new distance-decremental perturbation budget assignment strategy that allocates budgets decrementally as the pixel distance to semantic-critical regions increases. Extensive experiments across multiple LVLMs and datasets show that our method significantly outperforms baseline approaches.", "date": "2026-05-15", "categories": []}
{"id": "2303.14511", "title": "Improving robustness of jet tagging algorithms with adversarial training: exploring the loss surface", "abstract": "In the field of high-energy physics, deep learning algorithms continue to gain in relevance and provide performance improvements over traditional methods, for example when identifying rare signals or finding complex patterns. From an analyst's perspective, obtaining highest possible performance is desirable, but recently, some attention has been shifted towards studying robustness of models to investigate how well these perform under slight distortions of input features. Especially for tasks that involve many (low-level) inputs, the application of deep neural networks brings new challenges. In the context of jet flavor tagging, adversarial attacks are used to probe a typical classifier's vulnerability and can be understood as a model for systematic uncertainties. A corresponding defense strategy, adversarial training, improves robustness, while maintaining high performance. Investigating the loss surface corresponding to the inputs and models in question reveals geometric interpretations of robustness, taking correlations into account.", "date": "2023-03-25", "categories": []}
{"id": "2507.07776", "title": "SCOOTER: A Human Evaluation Framework for Unrestricted Adversarial Examples", "abstract": "Unrestricted adversarial attacks aim to fool computer vision models without being constrained by $\\ell_p$-norm bounds to remain imperceptible to humans, for example, by changing an object's color. This allows attackers to circumvent traditional, norm-bounded defense strategies such as adversarial training or certified defense strategies. However, due to their unrestricted nature, there are also no guarantees of norm-based imperceptibility, necessitating human evaluations to verify just how authentic these adversarial examples look. While some related work assesses this vital quality of adversarial attacks, none provide statistically significant insights. This issue necessitates a unified framework that supports and streamlines such an assessment for evaluating and comparing unrestricted attacks. To close this gap, we introduce SCOOTER - an open-source, statistically powered framework for evaluating unrestricted adversarial examples. Our contributions are: $(i)$ best-practice guidelines for crowd-study power, compensation, and Likert equivalence bounds to measure imperceptibility; $(ii)$ the first large-scale human vs. model comparison across 346 human participants showing that three color-space attacks and three diffusion-based attacks fail to produce imperceptible images. Furthermore, we found that GPT-4o can serve as a preliminary test for imperceptibility, but it only consistently detects adversarial examples for four out of six tested attacks; $(iii)$ open-source software tools, including a browser-based task template to collect annotations and analysis scripts in Python and R; $(iv)$ an ImageNet-derived benchmark dataset containing 3K real images, 7K adversarial examples, and over 34K human ratings. Our findings demonstrate that automated vision systems do not align with human perception, reinforcing the need for a ground-truth SCOOTER benchmark.", "date": "2025-07-10", "categories": []}
{"id": "2601.01972", "title": "Hidden State Poisoning Attacks against Mamba-based Language Models", "abstract": "State space models (SSMs) like Mamba offer efficient alternatives to Transformer-based language models, with linear time complexity. Yet, their adversarial robustness remains critically unexplored. This paper studies the phenomenon whereby specific short input phrases induce a partial amnesia effect in such models, by irreversibly overwriting information in their hidden states, referred to as a Hidden State Poisoning Attack (HiSPA). Our benchmark RoBench-25 allows evaluating a model's information retrieval capabilities when subject to HiSPAs, and confirms the vulnerability of SSMs against such attacks. Even the recent Jamba-1.7-Mini SSM--Transformer (a 52B hybrid model) collapses on RoBench-25 under some HiSPA triggers, whereas pure Transformers do not. We also observe that HiSPA triggers significantly weaken the Jamba model on the popular Open-Prompt-Injections benchmark, unlike pure Transformers. We further show that the theoretical and empirical findings extend to Mamba-2, and also analyse a Mamba-2-based hybrid (Nemotron-3-Nano). Finally, our interpretability study reveals patterns in Mamba's hidden layers during HiSPAs that could be used to build a HiSPA mitigation system. The full code and data to reproduce the experiments can be found at https://anonymous.4open.science/r/hispa_anonymous-5DB0.", "date": "2026-01-05", "categories": []}
{"id": "2605.09027", "title": "GAMBIT: A Three-Mode Benchmark for Adversarial Robustness in Multi-Agent LLM Collectives", "abstract": "In multi-agent systems (MAS), a single deceptive agent can nullify all gains of an agentic AI collective and evade deployed defenses. However, existing adversarial studies on MAS target only shallow tasks and do not consider adaptive adversaries, which evolve their strategies to evade the very detectors trained to catch them. To address that gap, we introduce GAMBIT, a benchmark with three evaluation modes and two independent scores for evaluating imposter detectors: the first two modes measure zero-shot detection under increasing distribution shift, and a third recalibration mode measures how quickly a detector adapts to novel attacks from just 20 labeled examples. The benchmark comes with a dataset of 27,804 labeled instances spanning 240 co-evolved imposter strategies. Our contributions are threefold: (1) Using chess as a substrate deep reasoning problem and Gemini 3.1 Pro for agents, we release GAMBIT and its dataset to evaluate imposter detectors under realistic constraints against a stealthy adaptive imposter; (2) We introduce an adaptive imposter agent based on an efficient evolutionary framework, generalizable beyond chess, that collapses collective task performance while remaining essentially undetectable (50.5% F1-score with a Gemini-based detector); (3) We show that zero-shot evaluation can be highly misleading for adaptive adversaries: two detectors with near-identical zero-shot scores differ by 8x on few-shot adaptation, while the meta-learned variant converges 20x faster, a gap only visible in the recalibration mode. Altogether, GAMBIT provides the first multi-agent benchmark where adversarial attacks and defenses co-evolve, with an imposter framework generalizable beyond our use case, and promising techniques for fast recalibration in a rapidly evolving adversarial system. Code and data: https://anonymous.4open.science/r/gambit.", "date": "2026-05-09", "categories": []}
{"id": "2605.13213", "title": "Hierarchical Attacks for Multi-Modal Multi-Agent Reasoning", "abstract": "Multi-modal multi-agent systems (MM-MAS) have gained increasing attention for their capacity to enable complex reasoning and coordination across diverse modalities. As these systems continue to expand in scale and functionality, investigating their potential vulnerabilities has become increasingly important. However, existing studies on adversarial attacks in multi-agent systems primarily focus on isolated agents or unimodal settings, leaving the vulnerabilities of MM-MAS largely underexplored. To bridge this gap, we introduce HAM$^{3}$, a Hierarchical Attack framework for multi-modal multi-agent systems that decomposes attacks into three interconnected layers. Specifically, at the perception layer, HAM$^{3}$ mounts attacks by perturbing visual inputs, textual inputs, and their fused visual-textual representations. At the communication layer, it performs communication-level attacks that corrupt message content and interaction topology, such as manipulating shared context or communication links to distort collective information flow. At the reasoning layer, it conducts reasoning-level attacks that interfere with each agent's cognitive pipeline, biasing reasoning trajectories and ultimately compromising final decisions. We evaluate HAM$^{3}$ on the GQA benchmark through multi-agent systems built on distinct reasoning paradigms including ReAct, Plan-and-Solve, and Reflexion. Experiments demonstrate that our framework achieves an Attack Success Rate of up to 78.3%, with reasoning-layer attacks being the most effective. More than half of the successful attacks lead multiple agents to produce consistent errors. These findings offer valuable insights for building more robust and interpretable multi-agent intelligence.", "date": "2026-05-13", "categories": []}
{"id": "2605.13338", "title": "Inducing Overthink: Hierarchical Genetic Algorithm-based DoS Attack on Black-Box Large Language Reasoning Models", "abstract": "Large Reasoning Models (LRMs) are increasingly integrated into systems requiring reliable multi-step inference, yet this growing dependence exposes new vulnerabilities related to computational availability. In particular, LRMs exhibit a tendency to \"overthink\", producing excessively long and redundant reasoning traces, when confronted with incomplete or logically inconsistent inputs. This behavior significantly increases inference latency and energy consumption, forming a potential vector for denial-of-service (DoS) style resource exhaustion. In this work, we investigate this attack surface and propose an automated black-box framework that induces overthinking in LRMs by systematically perturbing the logical structure of input problems. Our method employs a hierarchical genetic algorithm (HGA) operating on structured problem decompositions, and optimizes a composite fitness function designed to maximize both response length and reflective overthinking markers. Across four state-of-the-art reasoning models, the proposed method substantially amplifies output length, achieving up to a 26.1x increase on the MATH benchmark and consistently outperforming benign and manually crafted missing-premise baselines. We further demonstrate strong transferability, showing that adversarial inputs evolved using a small proxy model retain high effectiveness against large commercial LRMs. These findings highlight overthinking as a shared and exploitable vulnerability in modern reasoning systems, underscoring the need for more robust defenses.", "date": "2026-05-13", "categories": []}
{"id": "2605.14396", "title": "Systematic Discovery of Semantic Attacks in Online Map Construction through Conditional Diffusion", "abstract": "Autonomous vehicles depend on online HD map construction to perceive lane boundaries, dividers, and pedestrian crossings -- safety-critical road elements that directly govern motion planning. While existing pixel perturbation attacks can disrupt the mapping, they can be neutralized by standard adversarial defenses. We present MIRAGE, a framework for systematic discovery of semantic attacks that bypass adversarial defenses and degrade mapping predictions by finding plausible environmental variation (e.g. shadows, wet roads). MIRAGE exploits the latent manifold of real-world data learned by diffusion models, and searches for semantically mutated scenes neighboring the ground truth with the same road topology yet mislead the mapping predictions. We evaluate MIRAGE on nuScenes and demonstrate two attacks: (1) boundary removal, suppressing 57.7% of detections and corrupting 96% of planned trajectories; and (2) boundary injection, the only method that successfully injects fictitious boundaries, while pixel PGD and AdvPatch fail entirely. Both attacks remain potent under various adversarial defenses. We use two independent VLM judges to quantify realism, where MIRAGE passes as realistic 80--84% of the time (vs. 97--99% for clean nuScenes), while AdvPatch only 0--9%. Our findings expose a categorical gap in current adversarial defenses: semantic-level perturbations that manifest as legitimate environmental variation are substantially harder to mitigate than pixel-level perturbations.", "date": "2026-05-14", "categories": []}
{"id": "2605.14418", "title": "The Great Pretender: A Stochasticity Problem in LLM Jailbreak", "abstract": "\"Oh-Oh, yes, I'm the great pretender. Pretending that I'm doing well. My need is such, I pretend too much...\" summarizes the state in the area of jailbreak creation and evaluation. You find this method to generate adversarial attacks proposed by a reputable institution (e.g., BoN from Anthropic or Crescendo from Microsoft Research). However, this method does not deliver on the promise claimed in the paper despite having top ASR scores against industry-grade LLMs. You successfully generate the jailbreak prompts against your target (open) model. However, the generated jailbreak prompt works against the target model with a 50% consecutive success rate (5 out of 10 attempts) despite having an 80% ASR (on paper) on the latest closed-source model (with a guardrail system)! This observation leads us to think. First, Attack Success Rate (ASR), the primary metric for LLM jailbreak benchmarking, is not a stable quantity. Second, published ASR numbers are therefore systematically inflated and incomparable across papers. Therefore, we wonder \"Why a successful jailbreak prompt does not perform consistently well against a target model on which the prompts have been optimized?\". To answer this question, we study the impact of stochasticity not only during attack evaluation but also during attack generation. Our evaluation includes several jailbreak attacks, models (different sizes and providers), and judges. In addition, we propose a new metric and two new frameworks (CAS-eval and CAS-gen). Our evaluation framework, CAS-eval, shows that an attack can have an ASR drop of up to 30 percentage points when a jailbreak prompt needs to succeed on more than one attempt. Thankfully, our attack generation framework (CAS-gen) improves previous jailbreak methods and helps them recover this loss of 30 percentage points!", "date": "2026-05-14", "categories": []}
{"id": "2605.14868", "title": "Fast Adversarial Attacks with Gradient Prediction", "abstract": "Generating adversarial examples at scale is a core primitive for robustness evaluation, adversarial training, and red-teaming, yet even \"fast\" attacks such as FGSM remain throughput-limited by the cost of a backward pass. We introduce a family of attacks that eliminates the backward pass by predicting the input gradient from forward-pass hidden states via a lightweight linear regression. The approach is motivated by a kernel view of neural networks and is exact in the Neural Tangent Kernel regime, while remaining effective for practical finite-width models. Empirically, our methods recover much of FGSM's attack performance while using only a small fraction of the time, corresponding to a $532\\%$ increase in throughput. These results suggest gradient prediction as a simple and general route to significantly faster adversarial generation under realistic wall-clock constraints.", "date": "2026-05-14", "categories": []}
{"id": "2605.15030", "title": "WARD: Adversarially Robust Defense of Web Agents Against Prompt Injections", "abstract": "Web agents can autonomously complete online tasks by interacting with websites, but their exposure to open web environments makes them vulnerable to prompt injection attacks embedded in HTML content or visual interfaces. Existing guard models still suffer from limited generalization to unseen domains and attack patterns, high false positive rates on benign content, reduced deployment efficiency due to added latency at each step, and vulnerability to adversarial attacks that evolve over time or directly target the guard itself. To address these limitations, we propose WARD (Web Agent Robust Defense against Prompt Injection), a practical guard model for secure and efficient web agents. WARD is built on WARD-Base, a large-scale dataset with around 177K samples collected from 719 high-traffic URLs and platforms, and WARD-PIG, a dedicated dataset designed for prompt injection attacks targeting the guard model. We further introduce A3T, an adaptive adversarial attack training framework that iteratively strengthens WARD through a memory-based attacker and guard co-evolution process. Extensive experiments show that WARD achieves nearly perfect recall on out-of-distribution benchmarks, maintains low false positive rates to preserve agent utility, remains robust against guard-targeted and adaptive attacks under substantial distribution shifts, and runs efficiently in parallel with the agent without introducing additional latency.", "date": "2026-05-14", "categories": []}
{"id": "2601.13359", "title": "Sockpuppetting: Jailbreaking LLMs by Combining Prefilling with Optimization", "abstract": "Prefill attacks are an effective and low-cost jailbreaking method, as they directly insert an acceptance sequence (e.g., \"Sure, here is how to...\") at the start of an LLM's output and lead the model to continue the response. We make two contributions to this prior work. First, we show that an unsophisticated adversary can improve the well-known prefill attacks by ensembling a small number of prefill variants. Running three easy-to-generate prefills yields a combined attack success rate (ASR) of 22%, 90%, and 99% on Gemma-7B, Llama-3.1-8B, and Qwen3-8B respectively, an up to 38% improvement over the standard \"Sure, here's...\" prefill and up to 82% over our reproduction of GCG (Zou et al., 2023). Second, we introduce \"sockpuppetting\", a hybrid attack that optimizes an adversarial suffix placed inside the \"assistant\" message block of the chat template, rather than within the user prompt. The rolling variant of this attack, RollingSockpuppetGCG, increases prompt-agnostic ASR by up to 64% over our universal GCG baseline on Llama-3.1-8B. Both findings highlight the need for defences against output-prefix injection in open-weight models. Code: https://gitlab.com/asendotsinski/sockpuppetting", "date": "2026-01-19", "categories": []}
{"id": "2605.11492", "title": "A Mimetic Detector for Adversarial Image Perturbations", "abstract": "Adversarial attacks fool deep image classifiers by adding tiny, almost invisible noise patterns to a clean image. The standard $\\ell^\\infty$-bounded attacks (FGSM, PGD, and the $\\ell^\\infty$ variant of Carlini--Wagner) produce high-frequency, near-random sign patterns at the pixel level: nearly invisible in $\\ell^2$, but carrying disproportionate gradient energy. We exploit this with a single-shot, training-free detector using the high-order Corbino--Castillo mimetic operators from the open-source MOLE library. No retraining, no surrogate classifier, no access to the network under attack: the verdict is a property of the input alone, computed in $O(HW)$ time. We validate the detector on the standard \\texttt{peppers} test image at the canonical $\\ell^\\infty$ budget $\\varepsilon = 16/255$ and observe a clean-vs-adversarial separation that grows monotonically from $3.55\\times$ at order $k=2$ to $4.62\\times$ at $k=8$.", "date": "2026-05-12", "categories": []}
{"id": "2605.12743", "title": "Still Camouflage, Moving Illusion: View-Induced Trajectory Manipulation in Autonomous Driving", "abstract": "Existing physical adversarial attacks on vision-based autonomous driving induce time-evolving perception errors, including biased object tracking or trajectory prediction, through (i) sophisticated physical patch inducing detection box drift when entering the view distance, or (ii) dynamically changing patches that cause different perception errors at different time. In both cases, viewing-angle variation is treated as a challenge, requiring adversarial patches to remain effective across frames under varying views, leading to complex multi-view optimization. In contrast, we show that viewing-angle variation itself can be turned into an attack tool. We design a new attack paradigm where a static, passive adversarial camouflage is mounted on a vehicle whose view-dependent appearance naturally evolves with relative motion, inducing consistent feature drift across frames. This causes the system to infer a physically plausible but incorrect trajectory, such as a false cut-in, which propagates to downstream decision-making and triggers unnecessary braking. Unlike prior approaches that require multi-view robustness or active intervention, our attack emerges from normal driving dynamics and is easy to deploy: a parked vehicle with a natural camouflage can induce hard braking in passing autonomous vehicles. We demonstrate the novel attack on nuScenes dataset, showing the effectiveness with an end-to-end success rate of up to 87.5%, measured by hard-braking events, and robustness across different scene backgrounds, victim vehicle speeds, and perception models.", "date": "2026-05-12", "categories": []}
{"id": "2605.12792", "title": "SoK: A Comprehensive Analysis of the Current Status of Neural Tangent Generalization Attacks with Research Directions", "abstract": "There is recently a serious issue that Deep Neural Networks (DNNs) training uses more and more unauthorized data. A clean-label generalization attack, one type of data poisoning attacks, has been suggested to address this issue. The Neural Tangent Generalization Attack (NTGA) is considered as the first well-known clean-label generalization attack under the black-box settings, which provided an unprecedented step in data protection approaches. In this paper, we conduct a comprehensive analysis on the state-of-the-art of NTGA; to the best of our knowledge, this is the first thorough analysis regarding NTGA. First, we provide a classification of attacks against DNNs with their explanations and relations to NTGA. Then, this paper presents a taxonomy of black-box attacks and demonstrate that the NTGA is the first clean-label generalization attack under the black-box setting. We further analyze the existing studies of NTGA and give a comprehensive comparisons of their findings by conducting our own experiments to verify these findings. Moreover, our extensive experiments show that NTGA is vulnerable to adversarial training and image transformations, and applying linear separability to NTGA-generated images makes them more susceptible to such vulnerablities. We present the pros and cons of NTGA and suggest ways to improve NTGA robustness based on our analysis. Our further experiments indicate that several recently proposed clean-label generalization attacks outperform NTGA on data protection. Finally, we unveil the necessity of further research with future research insights on NTGA.", "date": "2026-05-12", "categories": []}
{"id": "2605.12937", "title": "AuraMask: An Extensible Pipeline for Developing Aesthetic Anti-Facial Recognition Image Filters", "abstract": "Anti-facial recognition (AFR) image filters alter images in ways that are subtle to people but blinding to computer vision. Yet, despite widespread interest in these technologies to subvert surveillance, users rarely use them in practice -- because the ``subtle'' alterations are visible enough to conflict with users' self-presentation goals. To address this challenge, we propose AuraMask: a novel approach to creating AFR filters that are both adversarially effective and aesthetically acceptable. Using AuraMask, we produce 40 ``aesthetic'' filters that emulate popular ``one-click'' Instagram image filters. We show that AuraMask filters meet or exceed the adversarial effectiveness of prior methods against open-source facial recognition models. Moreover, in a controlled online user study ($N=630$) we confirm these filters achieve significantly higher user acceptance than prior methods. Lastly, we provide our AFR pipeline to the community for accelerated research in adversarially effective and aesthetically acceptable protections.", "date": "2026-05-13", "categories": []}
{"id": "2605.13170", "title": "Finding the Weakest Link: Adversarial Attack against Multi-Agent Communications", "abstract": "Multi-agent systems rely on communication for information sharing and action coordination, which exposes a vulnerability to attacks. We investigate single-victim communication perturbation attacks against Multi-Agent Reinforcement Learning-trained systems and propose methods that use gradient information from the Jacobian to identify which messages, agent, and timesteps are most susceptible to attack and have the greatest impact on the system. We enhance these methods with two proposed adversarial loss functions that trade-off attack success for attack impact which also create more effective perturbations. We empirically demonstrate the effectiveness of our methods against two different multi-agent communication methods in navigation, PredatorPrey, and TrafficJunction environments. Our results show that our novel message selection method achieves a similar or greater impact than random message selection across almost all tested scenarios. Our victim selection, message selection, tempo, and loss functions improve attack effectiveness in half of the thirty scenarios we tested.", "date": "2026-05-13", "categories": []}
{"id": "2605.13381", "title": "Backbone is All You Need: Assessing Vulnerabilities of Frozen Foundation Models in Synthetic Image Forensics", "abstract": "As AI-generated synthetic images become increasingly realistic, Vision Transformers (ViTs) have emerged as a cornerstone of modern deepfake detection. However, the prevailing reliance on frozen, pre-trained backbones introduces a subtle yet critical vulnerability. In this work, we present the Surrogate Iterative Adversarial Attack (SIAA), a gray-box attack that exploits knowledge of the detector's ViT backbone alone and operates entirely within the target detector's feature space to craft highly effective adversarial examples. Through our experiments, involving multiple ViT-based detectors and diverse gray-box scenarios, including few-shot learning, complete training misalignment and attack transferability tests, we demonstrate that this vulnerability consistently yields high attack success rates, often approaching white-box performance. By doing so, we reveal that backbone knowledge alone is sufficient to undermine detector reliability, highlighting the urgent need for more resilient defenses in adversarial multimedia forensics.", "date": "2026-05-13", "categories": []}
{"id": "2605.13395", "title": "Taming the Long Tail: Rebalancing Adversarial Training via Adaptive Perturbation", "abstract": "Deep neural networks are highly vulnerable to adversarial examples, i.e.,small perturbations that can significantly degrade model performance. While adversarial training has become the primary defense strategy, most studies focus on balanced datasets, overlooking the challenges posed by real-world long-tail data. Motivated by the fact that perturbations in adversarial examples inherently alter the training distribution, we theoretically investigate their impact. We first revisit adversarial training for long-tail data and identify two key limitations: (i) a skewed training objective caused by class imbalance, and (ii) unstable evolution of adversarial distributions. Furthermore, we show that perturbations can simultaneously address both adversarial vulnerability and class imbalance. Based on these insights, we propose RobustLT, a plug-and-play framework that adaptively adjusts perturbations during adversarial training. Extensive experiments demonstrate that RobustLT consistently enhances adversarial robustness and class-balance on long-tailed datasets. The code is available at \\href{https://github.com/zhang-lilin/RobustLT}{https://github.com/zhang-lilin/RobustLT}.", "date": "2026-05-13", "categories": []}
{"id": "2605.13411", "title": "Model-Agnostic Lifelong LLM Safety via Externalized Attack-Defense Co-Evolution", "abstract": "Large language models remain vulnerable to adversarial prompts that elicit harmful outputs. Existing safety paradigms typically couple red-teaming and post-training in a closed, policy-centric loop, causing attack discovery to suffer from rapid saturation and limiting the exposure of novel failure modes, while leaving defenses inefficient, rigid, and difficult to transfer across victim models. To this end, we propose EvoSafety, an LLM safety framework built around persistent, inspectable, and reusable external structures. For red teaming, EvoSafety equips the attack policy with an adversarial skill library, enabling continued vulnerability probing through simple library expansion after saturation, while supporting the evolution of adversarial vectors. For defense learning, EvoSafety replaces model-specific safety fine-tuning with a lightweight auxiliary defense model augmented with memory retrieval. This enables efficient, transferable, and model-agnostic safety improvements, while allowing robustness to be enhanced solely through memory updates. With a single training procedure, the defense policy can operate in both Steer and Guard modes: the former activates the victim model's intrinsic defense mechanisms, while the latter directly filters harmful inputs. Extensive experiments demonstrate the superiority of EvoSafety: in Guard mode, it achieves a 99.61% defense success rate, outperforming Qwen3Guard-8B by 14.13% with only 37.5% of its parameters, while preserving reasoning performance on benign queries. Warning: This paper contains potentially harmful text.", "date": "2026-05-13", "categories": []}
{"id": "2411.16769", "title": "Red-Teaming Text-to-Image Models via In-Context Experience Replay and Semantic-Preserving Prompt Rewriting", "abstract": "Understanding the capabilities of text-to-image (T2I) models in harmful content generation is essential to safety and compliance. However, human red-teaming is costly and inconsistent, driving the need for automatic tools that simulate realistic misuse attempts. Existing methods either require white-box access, fail to generalize across defenses, or produce uninterpretable adversarial tokens, while generating fluent prompts that preserve the original harmful intent remains underexplored despite its practical relevance. We propose ICER, a black-box framework that addresses this gap through two components: an LLM-based rewriter that produces fluent, natural-language adversarial prompts, and in-context experience replay that accumulates successful jailbreaking patterns into a reusable prior. These components are integrated via bandit optimization, enabling ICER to efficiently balance exploiting proven attack strategies with exploring new ones. Experiments across six safety mechanisms show that ICER outperforms seven baselines under both standard and semantics-preserving evaluation, with over 30% of generated prompts transferring to commercial systems like DALL-E 3 and Midjourney.", "date": "2024-11-25", "categories": []}
{"id": "2512.20865", "title": "Robustness Certificates for Neural Networks against Adversarial Attacks", "abstract": "The increasing use of machine learning in safety-critical domains amplifies the risk of adversarial threats, especially data poisoning attacks that corrupt training data to degrade performance or induce unsafe behavior. Most existing defenses lack formal guarantees or rely on restrictive assumptions about the model class, attack type, extent of poisoning, or point-wise certification, limiting their practical reliability. This paper introduces a principled formal robustness certification framework that models gradient-based training as a discrete-time dynamical system (dt-DS) and formulates poisoning robustness as a formal safety verification problem. By adapting the concept of barrier certificates (BCs) from control theory, we introduce sufficient conditions to certify a robust radius ensuring that the terminal model remains safe under worst-case ${\\ell}_p$-norm based poisoning. To make this practical, we parameterize BCs as neural networks trained on finite sets of poisoned trajectories. We further derive probably approximately correct (PAC) bounds by solving a scenario convex program (SCP), which yields a confidence lower bound on the certified robustness radius generalizing beyond the training set. Importantly, our framework also extends to certification against test-time attacks, making it the first unified framework to provide formal guarantees in both training and test-time attack settings. Experiments on MNIST, SVHN, and CIFAR-10 show that our approach certifies non-trivial perturbation budgets while being model-agnostic and requiring no prior knowledge of the attack or contamination level.", "date": "2025-12-24", "categories": []}
{"id": "2605.03629", "title": "Adversarial Effects on Expressibility and Trainability in Distributed Variational Quantum Algorithms", "abstract": "Distributed quantum algorithms offer a promising pathway to scale variational quantum algorithms beyond the constraints of noisy intermediate-scale quantum hardware. However, existing approaches implicitly assume a trusted entanglement-sharing layer across quantum processors. We show that this assumption introduces a fundamental vulnerability: adversarial perturbations of shared entanglement induce structured gate-level noise that directly impacts quantum learning. We develop a framework that maps entanglement-level perturbations to gate-level noise via an explicit Kraus representation. To quantify their impact, we introduce Kraus expressibility, a metric that generalizes unitary expressibility to noisy quantum channels. We then establish a trade-off between Kraus expressibility and trainability of noisy quantum circuits through gradient variance analysis. Our analysis reveals that an adversary can manipulate Kraus expressibility to maintain sufficiently large cost gradients (avoiding barren plateaus) while systematically biasing optimization toward incorrect solutions. We validate these findings through numerical simulations, demonstrating adversarial degradation of expressibility and trainability.", "date": "2026-05-05", "categories": []}
{"id": "2605.10954", "title": "Controlled Steering-Based State Preparation for Adversarial-Robust Quantum Machine Learning", "abstract": "Quantum machine learning (QML) provides a promising framework for leveraging quantum-mechanical effects in learning tasks. However, its vulnerability to adversarial perturbations remains a major challenge for practical deployment. In QML systems, small perturbations applied to classical inputs can propagate through the quantum encoding stage and distort the resulting quantum state, thereby degrading model performance. In this work, we propose a defense mechanism that replaces the conventional quantum encoding stage of a QML model with passive steering-based controlled state preparation, which guides the encoded state toward a controlled intermediate state. By tuning the steering strength and the number of steering iterations, the proposed method suppresses the influence of adversarial perturbations while maintaining high clean accuracy and improving adversarial accuracy. Experimental results demonstrate that the passive steering-based defense consistently improves adversarial accuracy across different QML models and datasets under gradient-based adversarial attacks, achieving adversarial accuracy improvements of up to 40.19%.", "date": "2026-04-30", "categories": []}
{"id": "2605.10987", "title": "AESOP: Adversarial Execution-path Selection to Overload Deep Learning Pipelines", "abstract": "Modern machine learning deployments increasingly compose specialized models into dynamic inference pipelines, where upstream components produce intermediate predictions that determine the workload and inputs of downstream components. The cost of processing an input is therefore not determined by any single model, but by two coupled factors: the per-inference cost of each invoked component and its workload volume. Because these pipelines run under hard real-time constraints, efficiency is a fundamental requirement for system availability. We show that this structure creates an efficiency-attack surface that existing methods targeting single models cannot exploit: on identical inputs and budgets, path-aware targeting inflates FLOPs by $2,407\\times$ while the strongest single-model baseline achieves $117\\times$ -- a $20\\times$ gap attributable entirely to where the attack is directed. We formalize this as the adversarial path-selection problem and present AESOP, a framework combining vulnerability-guided path ranking with adaptive loss weighting. We evaluate AESOP on five pipelines plus a production-realistic deployment variant with batching, bounded buffering, and confidence-threshold defenses. AESOP achieves up to $2,407\\times$ FLOPs and $419\\times$ latency inflation in white-box setting and 58$\\times$ FLOPs / 17$\\times$ latency in gray-box settings. Under system-level defenses, the attack is not neutralized but redirected: pipelines are forced to choose between throughput collapse ($0.578 \\to 0.006$ input/s) and $96.7\\%$ data loss to sustain throughput.", "date": "2026-05-09", "categories": []}
{"id": "2605.11730", "title": "Persona-Conditioned Adversarial Prompting: Multi-Identity Red-Teaming for Adversarial Discovery and Mitigation", "abstract": "Automated red-teaming for LLMs often discovers narrow attack slices, missing diverse real-world threats, and yielding insufficient data for safety fine-tuning. We introduce Persona-Conditioned Adversarial Prompting (PCAP), which conditions adversarial search on diverse attacker personas (e.g., doctors, students, malicious actors) and strategy sets to explore realistic attack scenarios. By running parallel persona-conditioned searches, PCAP discovers transferable jailbreaks across different contexts and generates rich defense datasets with automatic metadata tracking. On GPT-OSS 120B, PCAP increases attack success from 57\\% to 97\\% while producing 2-6$\\times$ more diverse prompts covering varied real-world scenarios. Critically, fine-tuning lightweight adapters on PCAP-generated data significantly improves model robustness (recall: 0.36 $\\rightarrow$ 0.99, F1: 0.53 $\\rightarrow$ 0.96) with minimal false positives, demonstrating a practical closed-loop approach from vulnerability discovery to automated alignment.", "date": "2026-05-12", "categories": []}
{"id": "2503.09336", "title": "Stealthy Patch-Wise Backdoor Attack in 3D Point Cloud via Curvature Awareness", "abstract": "Backdoor attacks pose a severe threat to deep neural networks (DNNs) by implanting hidden backdoors that can be activated with predefined triggers to manipulate model behaviors maliciously. Recent studies have extended backdoor attacks to 3D point clouds, but most existing triggers are sample-wise and often cause visible geometric artifacts or high optimization cost. To address these limitations, we propose the Stealthy Patch-Wise Backdoor Attack (SPBA), a patch-wise backdoor attack framework for 3D point clouds. Specifically, SPBA decomposes a point cloud into local patches, where each patch is formed by a Farthest Point Sampling (FPS) center and its K-nearest neighbors (KNN). Candidate patches are ranked using a patch imperceptibility score derived from local curvature variation, and a unified spectral trigger is injected into the selected patches by perturbing only the coordinates of existing points while preserving the original point cardinality. Extensive experiments on ModelNet40 and ShapeNetPart further demonstrate that SPBA achieves state-of-the-art stealthiness among prior methods and reduces spectral-trigger computation by 98.43% relative to a sample-wise spectral baseline, while maintaining competitive attack performance. These results support localized spectral design as an effective and efficient approach to stealthy backdoor attacks in 3D point cloud models. Code is available at https://github.com/HazardFY/SPBA.", "date": "2025-03-12", "categories": []}
{"id": "2510.20129", "title": "SAID: Safety-Aware Intent Defense via Prefix Probing for Large Language Models", "abstract": "Large Language Models (LLMs) remain vulnerable to jailbreak attacks, where adversarially crafted prompts induce policy-violating responses despite safety alignment. Existing defenses typically improve safety through external filtering, auxiliary guardrails, or decoding-time control. However, these interventions often reduce practical deployability because they may require additional model access, introduce extra inference cost, or affect benign-task utility. In this paper, we propose Safety-Aware Intent Defense (SAID), a training-free jailbreak defense framework based on intent-level safety probing. SAID first distills potentially obfuscated user inputs into concise core intents using the target model itself. It then applies a validated safety prefix to probe each distilled intent and elicit the model's safety-aware response. Finally, a conservative aggregation rule rejects the original request if any distilled intent is identified as unsafe. This design enables black-box-compatible defense without updating model parameters or modifying the decoding process. Experiments on four open-source LLMs under six representative jailbreak attacks show that SAID achieves state-of-the-art defense performance in reducing harmful responses while maintaining competitive utility on benign tasks. Further analyses on prefix variants, hierarchical distillation, and inference efficiency demonstrate that SAID provides a practical safety-utility trade-off for securing LLMs against jailbreak threats.", "date": "2025-10-23", "categories": []}
{"id": "2602.00318", "title": "Optimal Transport-Guided Adversarial Attacks on Graph Neural Network-Based Bot Detection", "abstract": "The rise of bot accounts on social media poses significant risks to public discourse. To address this threat, modern bot detectors increasingly rely on Graph Neural Networks (GNNs). However, the effectiveness of these GNN-based detectors in real-world settings remains poorly understood. In practice, attackers continuously adapt their strategies as well as must operate under domain-specific and temporal constraints, which can fundamentally limit the applicability of existing attack methods. As a result, there is a critical need for robust GNN-based bot detection methods under realistic, constraint-aware attack scenarios.\n  To address this gap, we introduce BOCLOAK to systematically evaluate the robustness of GNN-based social bot detection via both edge editing and node injection adversarial attacks under realistic constraints. BOCLOAK constructs a probability measure over spatio-temporal neighbor features and learns an optimal transport geometry that separates human and bot behaviors. It then decodes transport plans into sparse, plausible edge edits that evade detection while obeying real-world constraints. We evaluate BOCLOAK across three social bot datasets, five state-of-the-art bot detectors, three adversarial defenses, and compare it against four leading graph adversarial attack baselines. BOCLOAK achieves up to 80.13% higher attack success rates while using 99.80% less GPU memory under realistic real-world constraints. Most importantly, BOCLOAK shows that optimal transport provides a lightweight, principled framework for bridging the gap between adversarial attacks and real-world bot detection.", "date": "2026-01-30", "categories": []}
{"id": "2605.00445", "title": "The Power of Order: Fooling LLMs with Adversarial Table Permutations", "abstract": "Large Language Models have achieved remarkable success and are increasingly deployed in critical applications involving tabular data, such as Table Question Answering. However, their robustness to the structure of this input remains a critical, unaddressed question. This paper demonstrates that modern LLMs exhibit a significant vulnerability to the layout of tabular data. Specifically, we show that semantically-invariant permutations of rows and columns - rearrangements that do not alter the table's underlying information - are sometimes sufficient to cause incorrect or inconsistent model outputs. To systematically probe this vulnerability, we introduce Adversarial Table Permutation, a novel, gradient-based attack that efficiently identifies worst-case permutations designed to maximally disrupt model performance. Our extensive experiments demonstrate that ATP significantly degrades the performance of a wide range of LLMs. This reveals a pervasive vulnerability across different model sizes and architectures, including the most recent and popular models. Our findings expose a fundamental weakness in how current LLMs process structured data, underscoring the urgent need to develop permutation-robust models for reliable, real-world applications.", "date": "2026-05-01", "categories": []}
{"id": "2605.07399", "title": "GPO-V: Jailbreak Diffusion Vision Language Model by Global Probability Optimization", "abstract": "Diffusion Vision-Language Models (dVLMs), built upon the non-causal foundations of Diffusion Large Language Models (dLLMs), have demonstrated remarkable efficacy in multimodal tasks by departing from the traditional autoregressive generation paradigm. While dVLMs appear inherently robust against conventional jailbreak tactics, which we categorize as Fixed Prefix Optimization (FPO) (e.g., anchoring responses with \"Sure, here is\"), this perceived resilience is deceptive. Our investigation into the safety landscape of dVLMs reveals a unique refusal pattern: Immediate Refusal and Progressive Refusal. We find that while FPO-based attacks often fail by triggering the latter, the progressive refinement process itself uncovers a novel, latent attack surface. To exploit this vulnerability, we propose Global Probability Optimization (GPO), a general jailbreak paradigm designed specifically for the denoising trajectory of masked diffusion models. Unlike prefix-based methods, GPO manipulates the global generative dynamics to bypass guardrails in diffusion language models. Building on this, we introduce GPO-V, the first visual-modality jailbreak framework tailored for dVLMs. Empirical results demonstrate that GPO-V produces stealthy perturbations with exceptional cross-model transferability, revealing a critical security gap in non-sequential generative architectures. Our findings underscore the critical urgency of addressing safety alignment in dVLMs. These results necessitate an immediate and fundamental re-evaluation of current defense paradigms to mitigate the unique risks of diffusion-based generation. Our code is available at: https://anonymous.4open.science/r/GPO-V-0250.", "date": "2026-05-08", "categories": []}
{"id": "2605.08257", "title": "Research on Security Enhancement Methods for Adversarial Robust Large Language Model Intelligent Agents for Medical Decision-Making Tasks", "abstract": "Motivated by the challenge to improve the adversarial robustness, security, and trust of medical decision making intelligent agents, this study develops a full-link security enhancement framework, which describes \"input risk perception - medical evidence constraint - knowledge consistency verification - decision confidence reweighting - security output control - adversarial feedback update.\" We propose ARSM-Agent and define a weighted joint objective consisting of decision accuracy loss, adversarial robustness loss, safety refusal loss, and knowledge consistency loss, with weights of 0.3, 0.3, 0.2, and 0.2, respectively. The whole medical decision formulation is implemented by multi-module collaborative linkage. We verify that the algorithm is more efficient than four baselines, including LLM-Agent, Retrieval-Agent, Filter-Agent, and Adv-Train-Agent. Under semantic perturbation, prompt injection, drug-name confusion, and false-evidence attacks, ARSM-Agent reduces the overall attack success rate to 8.7% and achieves a knowledge consistency score of 0.91. Ablation experiments quantify each module's contribution: removing risk perception, evidence retrieval, consistency verification, and confidence reweighting reduces accuracy by 6.7%, 9.1%, 7.6%, and 4.4%, respectively, and increases attack success rate by 13.8%, 11.1%, 8.6%, and 6.9%. The proposed approach addresses key security issues of medical decision making intelligent agents, obtains secure decision making in challenging scenarios, and provides reliable intelligent support for medical decision-making intelligent agents.", "date": "2026-05-07", "categories": []}
{"id": "2605.08310", "title": "WebTrap: Stealthy Mid-Task Hijacking of Browser Agents During Navigation", "abstract": "Browser agents are increasingly deployed in long-horizon tasks, which require executing extended action chains to accomplish user goals. However, this prolonged execution process provides attackers with more opportunities to inject malicious instructions. Existing prompt injection attacks against browser agents expose two key gaps: (1) low effectiveness, as attacks optimized for toy baselines fail to achieve end-to-end goals in real-world scenarios with complex environments and longer steps; (2) weak stealthiness, since most attacks pit the attack goal against the user goal, causing a significant drop in system usability under attack. To address these gaps, we propose WebTrap, a mid-task hijacking injection attack. It employs multi-step instruction fusion steering to seamlessly combine both goals, enabling the agent to resume the original user task after executing the attack goal. Furthermore, we design a context-grounded generation method to align the injected content with the task environment and system instructions, maximizing the hijacking success rate. Extensive experiments on two browser agent tasks, based on extended WASP and InjecAgent environments, demonstrate that our method achieves a high attack success rate while preserving the usability of the original system. We find that WebTrap exploits the agent's navigation vulnerabilities, binding the two goals so tightly that standard defense mechanisms cannot restore the system to normal operation. These findings reveal a critical vulnerability in agent systems during long-horizon tasks that they can be stealthily hijacked.", "date": "2026-05-08", "categories": []}
{"id": "2605.08440", "title": "TARO: Temporal Adversarial Rectification Optimization Using Diffusion Models as Purifiers", "abstract": "Adversarial purification with diffusion models seeks to project adversarial examples back toward the data manifold, but balancing semantic preservation and robustness against adaptive attacks remains challenging. Recent work shows that standard diffusion purification can fail under adaptive evaluation, while test-time score-based optimization is more resilient. Existing optimization defenses, however, typically rely on a single diffusion noise regime or treat timesteps uniformly, overlooking the distinct roles of coarse and fine denoising scales. We propose Temporal Adversarial Rectification Optimization (TARO), an inference-time purification method that builds a temporally guided score prior from multiple denoising views along the diffusion trajectory. TARO forms a coarse-to-fine residual target: high-noise experts provide globally smoothed structure with reduced adversarial sensitivity, while low-noise experts restore image-specific, class-relevant details. A guidance strength controls this temporal correction, allowing TARO to balance robust global rectification with semantic preservation. Empirically, TARO improves robust accuracy across datasets and adaptive threat models in a zero-shot setting, while remaining compatible with complementary adversarial-likelihood objectives for further robustness gains.", "date": "2026-05-08", "categories": []}
{"id": "2605.08727", "title": "Control Your View: High-Resolution Global Semantic Manipulation in Learned Image Compression", "abstract": "Learned image compression (LIC) integrates deep neural networks (DNNs) to map high-dimensional images into compact latent representations, reducing redundancy and achieving superior rate-distortion (RD) performance in benign settings. Unfortunately, due to inherent vulnerabilities in DNNs, LIC systems are susceptible to adversarial perturbations that lead to downstream deterioration, compression rate degradation, untargeted distortion, and both local semantic manipulation (LSM) and low-resolution ($3\\times28\\times28$) global semantic manipulation (GSM). However, high-resolution GSM remains unexplored due to its intractability. Notably, the existing project gradient descent (PGD) method achieves near-perfect white-box attacks for classification, segmentation, and other tasks, yet fails to generalize to high-resolution GSM. Our theoretical and empirical analyses reveal that well-performing GSM drives adversarial examples from the Identity Region to the Amplification Region through the Lazying-Oscillating-Refining stages. General $\\ell_{\\infty}$-bounded attacks fail on high-resolution GSM because their step-size schedules cannot accommodate both the Oscillating and Refining stages. Based on this, we propose the Periodic Geometric Decay schedule that enables $\\ell_{\\infty}$-bounded high-resolution GSM. To verify our approach, we integrate it with PGD, yielding a minimal variant, PGD$^{2}$-GSM. Extensive experiments on the Kodak $(3\\times768\\times512)$ demonstrate that our PGD$^{2}$-GSM is the first to stably achieve high-resolution GSM, thereby exposing a novel threat to LIC systems. Code is available at https://github.com/chinaliangjiaming/PGD2-GSM.", "date": "2026-05-09", "categories": []}
{"id": "2605.08778", "title": "Not All Turns Matter: Credit Assignment for Multi-Turn Jailbreaking", "abstract": "Deploying LLMs in multi-turn dialogues facilitates jailbreak attacks that distribute harmful intent across seemingly benign turns. Recent training-based multi-turn jailbreak methods learn long-horizon attack strategies from interaction feedback, but often rely on coarse trajectory-level outcome signals that broadcast uniformly to every turn. However, we find that turn-level contributions in multi-turn jailbreaking are non-uniform, phase-dependent, and target-specific. Such coarse outcome supervision induces a credit assignment problem, leading to over-rewarding redundant turns in successful trajectories and under-crediting useful intermediate turns in failed ones. To address this, we propose TRACE, a turn-aware credit assignment framework for reinforcement learning (RL)-based multi-turn jailbreaking. For successful trajectories, TRACE estimates turn-level contributions via leave-one-turn-out semantic masking; for failed ones, TRACE assigns penalties based on prompt harmfulness and semantic relevance, with an additional local refusal-aware penalty. Furthermore, we reuse the attack-side credit signal for multi-turn defense alignment. Extensive experiments on open-source and closed-source targets show that TRACE achieves strong overall performance in effectiveness, transferability, and efficiency, yielding about a 25% relative improvement in attack success rate over the strongest RL baseline while also improving the safety-utility balance when reused for defense alignment.", "date": "2026-05-09", "categories": []}
{"id": "2605.08898", "title": "LLM-Agnostic Semantic Representation Attack", "abstract": "Large Language Models (LLMs) increasingly employ alignment techniques to prevent harmful outputs. Despite these safeguards, attackers can circumvent them by crafting adversarial prompts. Predominant token-level optimization methods primarily rely on optimizing for exact affirmative templates (e.g., ``\\textit{Sure, here is...}''). However, these paradigms frequently encounter bottlenecks such as suboptimal convergence, compromised prompt naturalness, and poor cross-model generalization. To address these limitations, we propose Semantic Representation Attack (SRA), a novel LLM-agnostic paradigm that fundamentally reconceptualizes adversarial objectives from exact textual targeting to malicious semantic representations. Theoretically, we establish the semantic Coherence-Convergence Relationship and derive a Cross-Model Semantic Generalization bound, proving that maintaining semantic coherence guarantees both white-box semantic convergence and black-box transferability. Technically, we operationalize this framework via the Semantic Representation Heuristic Search (SRHS) algorithm, which preserves interpretability and structural coherence of the adversarial prompts during incremental discrete token chunk expansion. Extensive evaluations demonstrate that our framework achieves a 99.71% average attack success rate across 26 open-source LLMs, with strong transferability and stealth.", "date": "2026-05-09", "categories": []}
{"id": "2605.08910", "title": "Enhancing Adversarial Robustness in Network Intrusion Detection: A Layer-wise Adaptive Regularization Approach", "abstract": "The new wave of adversarial attacks that utilize gradient-related vulnerabilities in neural network-based classifiers makes Network Intrusion Detection Systems more open to such threats. Although state-of-the-art adversarial training methods have shown promising results in producing more robust classifiers, their interpretability and defense ability are limited due to their lack of understanding of how adversarial attacks propagate in different layers of network classifiers. In this paper, we present an insightful approach, called LARAR (Layer-wise Adversarial Robustness using Adaptive Regularization), that incorporates additional layer-wise vulnerability analysis and adaptive weighting in conventional adversarial training methods. Additionally, we utilize 'Auxiliary Classifiers' in our approach. LARAR provides interpretable layer-wise vulnerability scores, achieves a clean accuracy of 95.01%, and provides better robustness against adversarial attacks (FGSM, PGD, and transfer attacks) on the UNSW-NB15 dataset. Through the identification of vulnerable layers, the proposed framework reduces computational complexity and enables the early detection of adversarial samples, thus enhancing the effectiveness and interpretability of adversarial defense mechanisms in NIDS.", "date": "2026-05-09", "categories": []}
{"id": "2605.09225", "title": "The Art of the Jailbreak: Formulating Jailbreak Attacks for LLM Security Beyond Binary Scoring", "abstract": "Jailbreak attacks -- adversarial prompts that bypass LLM alignment through purely linguistic manipulation -- pose a growing operational security threat, yet the field lacks large-scale, reproducible infrastructure for generating, categorizing, and evaluating them systematically. This paper addresses that gap with three contributions. (1) Large-scale compositional jailbreak dataset. We construct 114,000 adversarial prompts by applying 912 composing strategies to 125 harmful seed prompts from JailBreakV-28K. Every prompt is assigned to one of 14 cybersecurity attack categories (e.g., malware, phishing, privilege escalation) via a six-model majority-vote pipeline, and each strategy is ranked by effectiveness per category, enabling principled strategy selection grounded in concrete adversarial objectives. (2) Automated jailbreak generation. We instruction-fine-tune category-aware LLMs on Moderate and Optimal subsets, producing models that synthesize fluent jailbreak prompts from a harmful seed at inference time -- no templates, no gradient search. Our generators achieve perplexity 24-39 versus 40-140 for AutoDAN and AmpleGCG, with safety-filter evasion rates of 0.29-0.51 Mal (LlamaPromptGuard-2-86M), enabling controllable, scalable red-teaming under realistic adversarial conditions. (3) OPTIMUS: a training-free jailbreak evaluator. OPTIMUS is a continuous metric J(S,H) that jointly captures semantic similarity between the harmful seed and the jailbreak (S) and harmfulness probability (H) via calibrated penalty functions. Unlike binary attack success rate (ASR), OPTIMUS requires no task-specific training, generalizes across evolving strategies, and exposes a stealth-optimal regime (S*=0.57, H*=0.43) that ASR misses. Experiments across 114,000 prompts confirm that OPTIMUS separates Weak, Moderate, and Optimal jailbreaks with category-level evidence binary evaluation cannot supply.", "date": "2026-05-09", "categories": []}
{"id": "2605.09504", "title": "Position: AI Security Policy Should Target Systems, Not Models", "abstract": "We present swarm-attack, an open-source adversarial testing framework in which multiple lightweight LLM agents coordinate through shared memory, parallel exploration, and evolutionary optimization. Together, our results demonstrate that both safety bypass of frontier models and software vulnerability discovery, i.e., the capability class that motivated restricted release of Anthropic's Mythos Preview, are achievable at effectively zero cost using commodity hardware and openly available models. We report two experiments. In the first, five instances of a 1.2 billion parameter model conducted 225 jailbreak attacks each against GPT-4o and Claude Sonnet~4. Against GPT-4o, the swarm achieved an Effective Harm Rate of 45.8%, producing 49 critical-severity breaches; against Claude Sonnet-4, the Effective Harm Rate was 0% despite a 40% technical success rate. In the second experiment, the same models performed combined source code analysis and binary fuzzing against a vulnerable C application with 9 planted CWEs. With a hand-crafted exploit seed corpus, regex pattern detection, and AddressSanitizer-based crash classification, the pipeline recovers 9 of 9 vulnerabilities (100% recall) in approximately four minutes on a consumer MacBook. With those scaffold components disabled, the same model recovers 0 of 9 by crash verification and 2 of 9 by citation. The capability class that motivated restricted release of Anthropic's Mythos Preview is therefore reproducible at effectively zero cost; the important enabler is the system scaffold itself, which compensates for the limited reasoning capacity of small individual models.", "date": "2026-05-10", "categories": []}
{"id": "2605.09902", "title": "Adversarial Attacks Against MLLMs via Progressive Resolution Processing and Adaptive Feature Alignment", "abstract": "Adversarial perturbations can mislead Multimodal Large Language Models (MLLMs) recognize a benign image as a specific target object, posing serious risks in safety-critical scenarios such as autonomous driving and medical diagnosis. This makes transfer-based targeted attacks crucial for understanding and improving black-box MLLM robustness. Existing transfer-based targeted attack methods typically rely on the final global features of the surrogate encoder and anchor optimization to original-resolution target crops, leading to their limited transferability and robustness. To address these challenges, we propose Progressive Resolution Processing and Adaptive Feature Alignment (PRAF-Attack), a targeted transfer-based attack framework that integrates multi-scale global semantic guidance with robust intermediate-layer local alignment. Unlike prior methods that align only the surrogate encoder's final layer, we design an adaptive feature alignment strategy that leverages intermediate representations to enhance transferability. Specifically, we introduce an adaptive intermediate layer selection mechanism to identify transferable hierarchical features across surrogate ensembles via gradient consistency, along with an adaptive patch-level optimization strategy that preserves highly correlated local regions through efficient patch filtering. To overcome the reliance on fixed original-resolution target crops, we propose a progressive resolution processing strategy that gradually refines optimization from coarse to fine, enabling the attack to better exploit target information at multiple scales and achieve stronger transferability. We evaluate PRAF-Attack on a diverse suite of black-box MLLMs, including six open-source models and six closed-source commercial APIs. Compared with seven state-of-the-art targeted attack baselines, the proposed PRAF-Attack consistently achieves superior transferability.", "date": "2026-05-11", "categories": []}
{"id": "2605.10253", "title": "Knowledge Poisoning Attacks on Medical Multi-Modal Retrieval-Augmented Generation", "abstract": "Retrieval-augmented generation (RAG) is a widely adopted paradigm for enhancing LLMs in medical applications by incorporating expert multimodal knowledge during generation. However, the underlying retrieval databases may naturally contain, or be intentionally injected with, adversarial knowledge, which can perturb model outputs and undermine system reliability. To investigate this risk, prior studies have explored knowledge poisoning attacks in medical RAG systems. Nevertheless, most of them rely on the strong assumption that adversaries possess prior knowledge of user queries, which is unrealistic in deployments and substantially limits their practical applicability. In this paper, we propose M\\textsuperscript{3}Att, a knowledge-poisoning framework designed for medical multimodal RAG systems, assuming only limited distribution knowledge of the underlying database. Our core idea is to inject covert misinformation into textual data while using paired visual data as a query-agnostic trigger to promote retrieval. We first propose a unified framework that introduces imperceptible perturbations to visual inputs to manipulate retrieval probabilities. Besides, due to the prior medical knowledge in LLMs, naively poisoned medical content with explicit factual errors can be corrected during generation. Thus, we leverage the inherent ambiguity of medical diagnosis and design a covert misinformation injection strategy that degrades diagnostic accuracy while evading model self-correction. Experiments on five LLMs and datasets demonstrate that M\\textsuperscript{3}Att consistently produces clinically plausible yet incorrect generations. Codes: https://github.com/ypr17/M3Att.", "date": "2026-05-11", "categories": []}
{"id": "2605.10582", "title": "Guaranteed Jailbreaking Defense via Disrupt-and-Rectify Smoothing", "abstract": "This paper proposes a guaranteed defense method for large language models (LLMs) to safeguard against jailbreaking attacks. Drawing inspiration from the denoised-smoothing approach in the adversarial defense domain, we propose a novel smoothing-based defense method, termed Disrupt-and-Rectify Smoothing (DR-Smoothing). Specifically, we integrate a two-stage prompt processing scheme-first disrupting the input prompt, then rectifying it-into the conventional smoothing defense framework. This disrupt-and-rectify approach improves upon previous disrupt-only approaches by restoring out-of-distribution disrupted prompts to an in-distribution form, thereby reducing the risk of unpredictable LLM behavior. In addition, this two-stage scheme offers a distinct advantage in striking a balance between harmlessness and helpfulness in jailbreaking defense. Notably, we present a theoretical analysis for generic smoothing framework, offering a tight bound for the defense success probability and the requirements on the disruption strength. Our approach can defend against both token-level and prompt-level jailbreaking attacks, under both established and adaptive attacking scenarios. Extensive experiments demonstrate that our approach surpasses current state-of-the-art defense methods in terms of both harmlessness and helpfulness.", "date": "2026-05-11", "categories": []}
{"id": "2605.06902", "title": "Streaming Adversarial Robustness in Fuzzy ARTMAP: Mechanism-Aligned Evaluation, Progressive Training, and Interpretable Diagnostics", "abstract": "Adversarial robustness has been studied extensively for offline deep networks, but less is known about strict single-pass streaming neural learners. This paper studies adversarial robustness in Fuzzy ARTMAP, an Adaptive Resonance Theory architecture based on category competition, complement coding, match tracking, and replay-free prototype updates. We introduce WB-Softmax, a differentiable white-box attack surrogate aligned with ARTMAP's category-competition and map-field prediction mechanism, and formalize a streaming evaluation principle requiring robustness to be assessed on the final deployed model. Across four image benchmarks, WB-Softmax achieves 89-100% attack success on vanilla Fuzzy ARTMAP models. We show that defense rankings can reverse across protocols: offline adversarial training may appear strong under transfer attacks yet collapse under adaptive white-box evaluation, whereas progressive two-stage selective training provides the strongest overall replay-free robustness. We further show that ART's explicit category geometry enables interpretable diagnosis of separation collapse and match-score inversion. These results provide a mechanism-aligned, protocol-aware framework for adversarial robustness in streaming prototype-based learners.", "date": "2026-05-07", "categories": []}
{"id": "2605.07273", "title": "From Clouds to Hallucinations: Atmospheric Retrieval Hijacking in Remote Sensing Vision-Language RAG", "abstract": "Multimodal RAG systems increasingly rely on vision-language retrievers to ground visual queries in external textual evidence. Existing adversarial studies on RAG mainly manipulate the retrieval corpus or memory, while attacks on vision-language and remote sensing models typically target end-task predictions. Input-space threats to the evidence retrieval stage of remote sensing multimodal RAG remain underexplored. To address this gap, we introduce CloudWeb, an atmospheric retrieval hijacking attack that modifies only the input image while keeping the retriever, generator, and knowledge base fixed at deployment. CloudWeb overlays parameterized cloud- and haze-like patterns on remote sensing images and optimizes them with a retrieval-oriented objective that pulls adversarial image embeddings toward target atmospheric evidence, suppresses source-scene evidence, enforces rank separation, and regularizes naturalness and coverage. To the best of our knowledge, this is the first study of retrieval-stage atmospheric evidence hijacking in remote sensing multimodal RAG. We evaluate CloudWeb on a seven-dataset remote sensing RAG benchmark with five CLIP-style retrievers, including GeoRSCLIP, RemoteCLIP, OpenAI CLIP, and OpenCLIP, together with downstream vision-language generators. Across retrievers, CloudWeb consistently outperforms clean retrieval, handcrafted atmospheric baselines, random cloud perturbations, and fixed variants in injecting weather-related evidence into top-ranked results. On GeoRSCLIP ViT-B/32, Weather@5 increases from 0.71\\% to 43.29\\%. Downstream generation further shows measurable weather hallucination and semantic shift, indicating that retrieval-stage hijacking can propagate to the final RAG response. These findings reveal a practical failure mode: natural-looking atmospheric changes can compromise evidence retrieval before generation begins.", "date": "2026-05-08", "categories": []}
{"id": "2605.07398", "title": "Exposing and Mitigating Temporal Attack in Deepfake Video Detection", "abstract": "While spatiotemporal deepfake detectors achieve high AUC, our experiments reveal their susceptibility to evasion attacks. These models tend to overfit on fragile temporal spectrum cues, rather than learning robust semantic causality. To mitigate this vulnerability, we propose SpInShield, a temporal spectral-invariant defense framework explicitly designed to decouple semantic motion from manipulatable spectral artifacts. We propose a learnable spectral adversary that dynamically synthesizes severe spectral deformations, simulating extreme attack scenarios. By employing a shortcut suppression optimization strategy, SpInShield compels the encoder to extract reliable forensic cues while purging unstable spectral statistics from the latent space. Experiments show that SpInShield obtains competitive performance on widely used datasets and outperforms the strongest baseline by 21.30 percentage points in AUC under simulated amplitude spectral attacks.", "date": "2026-05-08", "categories": []}
{"id": "2605.07447", "title": "Sparse Autoencoders as Plug-and-Play Firewalls for Adversarial Attack Detection in VLMs", "abstract": "Vision-language models (VLMs) have advanced rapidly and are increasingly deployed in real-world applications, especially with the rise of agent-based systems. However, their safety has received relatively limited attention. Even the latest proprietary and open-weight VLMs remain highly vulnerable to adversarial attacks, leaving downstream applications exposed to significant risks. In this work, we propose a novel and lightweight adversarial attack detection framework based on sparse autoencoders (SAEs), termed SAEgis. By inserting an SAE module into a pretrained VLM and training it with standard reconstruction objectives, we find that the learned sparse latent features naturally capture attack-relevant signals. These features enable reliable classification of whether an input image has been adversarially perturbed, even for previously unseen samples. Extensive experiments show that SAEgis achieves strong performance across in-domain, cross-domain, and cross-attack settings, with particularly large improvements in cross-domain generalization compared to existing baselines. In addition, combining signals from multiple layers further improves robustness and stability. To the best of our knowledge, this is the first work to explore SAE as a plug-and-play mechanism for adversarial attack detection in VLMs. Our method requires no additional adversarial training, introduces minimal overhead, and provides a practical approach for improving the safety of real-world VLM systems.", "date": "2026-05-08", "categories": []}
{"id": "2502.20650", "title": "Gungnir: Exploiting Stylistic Features in Images for Backdoor Attacks on Diffusion Models", "abstract": "Diffusion Models (DMs) have achieved remarkable success in image generation, yet recent studies reveal their vulnerability to backdoor attacks, where adversaries manipulate outputs via covert triggers embedded in inputs. Existing defenses, such as backdoor detection and trigger inversion, are largely effective because prior attacks rely on limited input spaces and low-dimensional triggers that are visually conspicuous or easily captured by neural detectors. To broaden the threat landscape, we propose Gungnir, a novel backdoor attack that activates malicious behaviors through style-based triggers embedded in input images. Unlike explicit visual patches or textual cues, stylistic features serve as stealthy, high-level triggers. We introduce Reconstructing-Adversarial Noise (RAN) and Short-Term Timesteps-Retention (STTR) to preserve trigger-consistent diffusion dynamics in image-to-image tasks. The resulting trigger-embedded samples are perceptually indistinguishable from clean images, evading both manual and automated detection. Extensive experiments show that Gungnir bypasses state-of-the-art defenses with an extremely low backdoor detection rate (BDR) and remains effective under fine-tuning-based purification, revealing previously underexplored vulnerabilities in diffusion models.", "date": "2025-02-28", "categories": []}
{"id": "2505.21938", "title": "Practical Adversarial Attacks on Stochastic Bandits via Fake Data Injection", "abstract": "Adversarial attacks on stochastic bandits have traditionally relied on some unrealistic assumptions, such as per-round reward manipulation and unbounded perturbations, limiting their relevance to real-world systems. We propose a more practical threat model, Fake Data Injection, which reflects realistic adversarial constraints: the attacker can inject only a limited number of bounded fake feedback samples into the learner's history, simulating legitimate interactions. We design effective attack strategies under this model, explicitly addressing both magnitude constraints (on reward values) and temporal constraints (on when and how often data can be injected). Our theoretical analysis shows that these attacks can mislead a class of bandit algorithms into selecting a target arm in nearly all rounds while incurring only sublinear attack cost. Experiments on synthetic and real-world datasets validate the effectiveness of our strategies, revealing vulnerabilities in stochastic bandit algorithms under practical adversarial scenarios.", "date": "2025-05-28", "categories": []}
{"id": "2605.05415", "title": "Information Theoretic Adversarial Training of Large Language Models", "abstract": "Large language models (LLMs) remain vulnerable to adversarial prompting despite advances in alignment and safety, often exhibiting harmful behaviors under novel attack strategies. While adversarial training can improve robustness, existing approaches are computationally expensive and difficult to scale. Recent continuous adversarial training methods, such as Continuous adversarial training (CAT) and Continuous Adversarial Preference Optimization (CAPO), address this challenge by leveraging gradient-based perturbations in the embedding space, enabling more efficient and expressive attacks. Building on this paradigm, we propose WARDEN, a distributionally robust adversarial training framework for LLMs that dynamically reweights adversarial examples through an f -divergence ambiguity set around the empirical training distribution. Our method optimizes the worst-case adversarial loss within a divergence ball around the empirical data distribution, automatically emphasizing harder adversarial examples. Using the convex dual formulation, the objective reduces to a log-sum-exp form under the KL divergence, with a dynamical parameter controlling the strength of reweighting. This study leads to a new class of information-theoretic objectives that significantly reduce attack success rates while maintaining model utility. Across multiple LLMs and attack settings, WARDEN substantially reduces attack success rates with computational and utility costs comparable to CAT-, CAPO-, and MixAT-based baselines, making it a practical approach for scalable robust alignment.", "date": "2026-05-06", "categories": []}
{"id": "2605.05534", "title": "Adversarial Graph Neural Network Benchmarks: Towards Practical and Fair Evaluation", "abstract": "Adversarial learning and the robustness of Graph Neural Networks (GNNs) are topics of widespread interest in the machine learning community, as documented by the number of adversarial attacks and defenses designed for these purposes. While a rigorous evaluation of these adversarial methods is necessary to understand the robustness of GNNs in real-world applications, we posit that many works in the literature do not share the same experimental settings, leading to ambiguous and potentially contradictory scientific conclusions. In this benchmark, we demonstrate the importance of adopting fair, robust, and standardized evaluation protocols in adversarial GNN research. We perform a comprehensive re-evaluation of seven widely used attacks and eight recent defenses under both poisoning and evasion scenarios, across six popular graph datasets. Our study spans over 453,000 experiments conducted within a unified framework. We observe substantial differences in adversarial attack performance when evaluated under a fair and robust procedure. Our findings reveal that previously overlooked factors, such as target node selection and the training process of the attacked model, have a profound impact on attack effectiveness, to the extent of completely distorting performance insights. These results underscore the urgent need for standardized evaluations in adversarial graph machine learning.", "date": "2026-05-07", "categories": []}
{"id": "2605.05632", "title": "Architecture Matters: Comparing RAG Systems under Knowledge Base Poisoning", "abstract": "Retrieval-Augmented Generation (RAG) systems are vulnerable to knowledge base poisoning, yet existing attacks have been evaluated almost exclusively against vanilla retrieve-then-generate pipelines. Architectures designed to handle conflicting retrieved information - multi-agent debate, agentic retrieval, recursive language models - remain untested against adversarially optimized contradictions. We evaluate four RAG architectures (vanilla RAG, agentic RAG, MADAM-RAG, and Recursive Language Models) under controlled single-document (N=1) poisoning on 921 Natural Questions QA pairs, comparing a clean baseline, naive injection, and CorruptRAG-AK - an adversarial attack whose meta-epistemic framing targets credibility assessment. Architecture is a high-impact variable in adversarial robustness: under CorruptRAG-AK, attack success rates range from 81.9% (vanilla) to 24.4% (RLM) - a spread of nearly 58 percentage points across architectures with comparable clean accuracy (~92%). Decomposing this gap, once the poisoned document is retrieved, adversarial framing - not retrieval optimization - drives the majority of CorruptRAG-AK's advantage for three of four architectures, localizing the cross-architecture vulnerability at the content-reasoning stage. Our MADAM-RAG reimplementation shows the highest apparent contradiction detection rate, though our LLM judge over-identifies this behavior (~48.5% precision), so reported rates are upper bounds. Regardless of detection, MADAM-RAG cannot resolve contradictions reliably, producing a 41.4% non-answer rate even on clean inputs - though implementation divergences from the original may contribute. We introduce a seven-category behavioral taxonomy capturing contradiction detection, hedging, and failure modes beyond binary accuracy. Code, data, and analysis notebooks are publicly available.", "date": "2026-05-07", "categories": []}
{"id": "2605.05928", "title": "Backdoor Mitigation in Object Detection via Adversarial Fine-Tuning", "abstract": "Backdoor attacks can implant malicious behaviours into deep models while preserving performance on clean data, posing a serious threat to safety-critical vision systems. Although backdoor mitigation has been studied extensively for image classification, defenses for object detection remain comparatively underdeveloped. Adversarial fine-tuning is a common backdoor mitigation approach in classification, but adapting it to detection is nontrivial as classification-oriented adversarial generation does not match the detection attack space, where attacks may cause object misclassification or disappearance, and standard detection losses can dilute the repair signal across many predictions. We address these challenges through a detection-aware adversarial fine-tuning framework for mitigating object-detection backdoors when the defender has access only to a compromised detector and a small clean dataset, without knowing the attack objective. For adversarial generation that does not require knowledge of the attack objective, we introduce soft-branch minimisation, which uses a soft gate to combine objectives aligned with misclassification and disappearance attacks, together with a detection-aware classification-loss maximisation. For targeted repair, we introduce a dual-objective fine-tuning loss applied to target-matched predictions, concentrating the defensive update on predictions most relevant to the backdoor behaviour. Experiments across CNN- and Transformer-based detectors show that our approach more effectively reduces attack success while preserving true detections, compared with classification-oriented baselines, and maintains competitive clean detection performance.", "date": "2026-05-07", "categories": []}
{"id": "2605.06357", "title": "Memory Efficient Full-gradient Attacks (MEFA) Framework for Adversarial Defense Evaluations", "abstract": "This work studies the robust evaluation of iterative stochastic purification defenses under white-box adversarial attacks. Our key technical insight is that gradient checkpointing makes exact end-to-end gradient computation through long purification trajectories practical by trading additional recomputation for substantially lower memory usage. This enables full-gradient adaptive attacks against diffusion- and Langevin-based purification defenses, where prior evaluations often resort to approximate backpropagation due to memory constraints. These approximations can weaken the attack signal and risk overestimating robustness. In parallel, stochasticity in iterative purification is frequently under-controlled, even though different purification trajectories can substantially change reported robustness metrics. Building on this insight, we introduce a memory-efficient full-gradient evaluation framework for stochastic purification defenses. The framework combines checkpointed backpropagation with evaluation protocols that control stochastic variability, thereby reducing memory bottlenecks while preserving exact gradients. We evaluate diffusion-based purification and Langevin sampling with Energy-Based Models (EBMs), demonstrating that full-gradient attacks uncover vulnerabilities missed by approximate-gradient evaluations. Our framework yields stronger state-of-the-art $\\ell_{\\infty}$ and $\\ell_{2}$ white-box attacks and further supports probing out-of-distribution robustness. Overall, our results show that exact-gradient evaluation is essential for reliable benchmarking of iterative stochastic defenses.", "date": "2026-05-07", "categories": []}
{"id": "2312.06423", "title": "MalPurifier: Enhancing Android Malware Detection with Adversarial Purification against Evasion Attacks", "abstract": "Machine learning (ML) has gained significant adoption in Android malware detection to address the escalating threats posed by the rapid proliferation of malware attacks. However, recent studies have revealed the inherent vulnerabilities of ML-based detection systems to evasion attacks. While efforts have been made to address this critical issue, many of the existing defensive methods encounter challenges such as lower effectiveness or reduced generalization capabilities. In this paper, we introduce MalPurifier, a novel adversarial purification framework specifically engineered for Android malware detection. Specifically, MalPurifier integrates three key innovations: a diversified adversarial perturbation mechanism for robustness and generalizability, a protective noise injection strategy for benign data integrity, and a Denoising AutoEncoder (DAE) with a dual-objective loss for accurate purification and classification. Extensive experiments on two large-scale datasets demonstrate that MalPurifier significantly outperforms state-of-the-art defenses. It robustly defends against a comprehensive set of 37 perturbation-based evasion attacks, consistently achieving robust accuracies above 90.91%. As a lightweight, model-agnostic, and plug-and-play module, MalPurifier offers a practical and effective solution to bolster the security of ML-based Android malware detectors.", "date": "2023-12-11", "categories": []}
{"id": "2602.00305", "title": "Syntax- and Compilation-Preserving Evasion of LLM Vulnerability Detectors", "abstract": "LLM-based vulnerability detectors are increasingly deployed in CI/CD security gating, yet their resilience to evasion under syntax- and compilation-preserving edits remains poorly understood. We evaluate five attack variants spanning four carrier families of behavior-preserving code transformations on a unified C/C++ benchmark ($N=5000$) and introduce Complete Resistance (CR), measuring the fraction of correctly detected vulnerabilities that withstand all attack variants. Our findings reveal a significant robustness gap: models achieving 70\\%+ clean recall exhibit CR as low as 0.12\\%, meaning over 87\\% of detected vulnerabilities can be evaded by at least one syntax-preserving edit. Universal adversarial strings optimized on a 14B surrogate transfer effectively to black-box APIs including GPT-4o, while on-target optimization further amplifies evasion (up to 92.5\\% ASR). These results indicate that clean benchmark accuracy alone is insufficient as a security guarantee for deployed vulnerability detectors.", "date": "2026-01-30", "categories": []}
{"id": "2605.04209", "title": "Undetectable Backdoors in Model Parameters: Hiding Sparse Secrets in High Dimensions", "abstract": "We present Sparse Backdoor, a supply-chain attack that plants a \\emph{provably undetectable} backdoor in pre-trained image classifiers, including convolutional networks and Vision Transformers. The attack injects a structured sparse perturbation along a randomly chosen direction into a small subset of columns at each fully connected layer, propagating a trigger signal to an adversary-chosen target class, and masks the perturbation with an independent isotropic Gaussian dither. The dither serves a single technical purpose: it induces a clean reference distribution anchored at the pre-trained weights, against which undetectability can be formalized. Under a mild margin condition on the pre-trained classifier, we show that the dithered reference is functionally equivalent to the original classifier. We prove that distinguishing the backdoor-injected model from this reference is at least as hard as Sparse PCA detection, which is computationally infeasible under standard hardness assumptions. The guarantee holds against any probabilistic polynomial-time distinguisher with white-box access to the parameters.", "date": "2026-05-05", "categories": []}
{"id": "2605.04443", "title": "Dissociating spatial frequency reliance from adversarial robustness advantages in neurally guided deep convolutional neural networks", "abstract": "Deep convolutional neural networks (DCNNs) have rivaled humans on many visual tasks, yet they remain vulnerable to near-imperceptible perturbations generated by adversarial attacks. Recent work shows that aligning DCNN representations with human visual cortex activity improves adversarial robustness, but the mechanisms driving this advantage are unclear. One hypothesis suggests that neural alignment confers robustness by biasing models away from brittle high-frequency details and towards the low spatial frequencies (LSF). However, recent work shows that human object recognition critically depends on a narrow, mid-frequency \"human channel\". Interestingly, this band was partially preserved in prior LSF-focused studies. Here, we investigate whether a spectral bias towards the LSF or the human channel is the primary driver of the adversarial robustness observed in neurally aligned DCNNs. We first show that DCNNs aligned to higher-order regions of the human ventral visual stream systematically increase reliance on both LSF and the human channel. However, directly steering DCNNs towards these bands revealed a clear dissociation. Biasing models towards the human channel, either alone or together with LSF, does not improve robustness and even impairs it. LSF bias produced some robustness gains, but such improvements are modest despite inducing much larger shifts in spatial-frequency reliance than neurally aligned models. Spatial-frequency-biased models overall show little, if any, increase in similarity to human neural representational geometry. Together, our results suggest that altered spatial-frequency reliance is likely an emergent property of learning more human-like representations rather than the primary mechanism by which neural alignment confers adversarial robustness, and motivate the need for future research examining representational properties beyond spatial-frequency profiles.", "date": "2026-05-06", "categories": []}
{"id": "2605.04675", "title": "Physical Adversarial Clothing Evades Visible-Thermal Detectors via Non-Overlapping RGB-T Pattern", "abstract": "Visible-thermal (RGB-T) object detection is a crucial technology for applications such as autonomous driving, where multimodal fusion enhances performance in challenging conditions like low light. However, the security of RGB-T detectors, particularly in the physical world, has been largely overlooked. This paper proposes a novel approach to RGB-T physical attacks using adversarial clothing with a non-overlapping RGB-T pattern (NORP). To simulate full-view (0$^{\\circ}$--360$^{\\circ}$) RGB-T attacks, we construct 3D RGB-T models for human and adversarial clothing. NORP is a new adversarial pattern design using distinct visible and thermal materials without overlap, avoiding the light reduction in overlapping RGB-T patterns (ORP). To optimize the NORP on adversarial clothing, we propose a spatial discrete-continuous optimization (SDCO) method. We systematically evaluated our method on RGB-T detectors with different fusion architectures, demonstrating high attack success rates both in the digital and physical worlds. Additionally, we introduce a fusion-stage ensemble method that enhances the transferability of adversarial attacks across unseen RGB-T detectors with different fusion architectures.", "date": "2026-05-06", "categories": []}
{"id": "2605.04698", "title": "Gray-Box Poisoning of Continuous Malware Ingestion Pipelines", "abstract": "Modern malware detection pipelines rely on continuous data ingestion and machine learning to counter the high volume of novel threats. This work investigates a realistic gray-box poisoning threat model targeting these pipelines. Using the secml_malware framework, we generate problem-space adversarial binaries through functionality-preserving manipulations, specifically Import Address Table (IAT) and section injections. We evaluate the impact of these poisoned samples when ingested into a defender's training set for a LightGBM malware detection model. Our empirical results demonstrate that subtle IAT-based perturbations enable compact poisoning samples that significantly degrade detection recall. These findings illustrate the inherent challenge of developing low-visibility adversarial perturbations that maintain high poisoning efficacy within continuous learning systems. We further evaluate a defense mechanism based on a homogeneous ensemble, which successfully identifies and filters up to 95.6% of poisoning attempts while maintaining a high retention rate for legitimate data. These findings emphasize the necessity of robust pre-ingestion validation in production pipelines.", "date": "2026-05-06", "categories": []}
{"id": "2510.22555", "title": "Cross-Paradigm Graph Backdoor Attacks with Promptable Subgraph Triggers", "abstract": "Graph Neural Networks(GNNs) are vulnerable to backdoor attacks, where adversaries implant malicious triggers to manipulate model predictions. Existing trigger generators are often simplistic in structure and overly reliant on specific features, confining them to a single graph learning paradigm, such as graph supervised learning, graph contrastive learning, or graph prompt learning. Such paradigm-specific designs lead to poor transferability across different learning frameworks, limiting attack success rates in general testing scenarios. To bridge this gap, we propose Cross-Paradigm Graph Backdoor Attacks with Promptable Subgraph Triggers(CP-GBA), which employs Graph Prompt Learning(GPL) to synthesize transferable subgraph triggers. Specifically, we first distill a compact yet expressive trigger set into a queryable repository, jointly optimizing for class-awareness, feature richness, and structural fidelity. Furthermore, we pioneer the theoretical exploration of GPL transferability under prompt-based objectives, ensuring robust generalization to diverse and unseen test-time paradigms. Extensive experiments across multiple real-world datasets and defense scenarios show that CP-GBA achieves state-of-the-art attack success rates.", "date": "2025-10-26", "categories": []}
{"id": "2601.14505", "title": "Uncovering and Understanding FPR Manipulation Attack in Industrial IoT Networks", "abstract": "In the network security domain, due to practical issues -- including imbalanced data and heterogeneous legitimate network traffic -- adversarial attacks in machine learning-based NIDSs have been viewed as attack packets misclassified as benign. Due to this prevailing belief, the possibility of (maliciously) perturbed benign packets being misclassified as attack has been largely ignored. In this paper, we demonstrate that this is not only theoretically possible, but also a particular threat to NIDS. In particular, we uncover a practical cyberattack, FPR manipulation attack (FPA), especially targeting industrial IoT networks, where domain-specific knowledge of the widely used MQTT protocol is exploited and a systematic simple packet-level perturbation is performed to alter the labels of benign traffic samples without employing traditional gradient-based or non-gradient-based methods. The experimental evaluations demonstrate that this novel attack results in a success rate of 80.19% to 100%. In addition, while estimating impacts in the Security Operations Center, we observe that even a small fraction of false positive alerts, irrespective of different budget constraints and alert traffic intensities, can increase the delay of genuine alerts investigations up to 2 hr in a single day under normal operating conditions. Furthermore, a series of relevant statistical and XAI analyses is conducted to understand the key factors behind this remarkable success. Finally, we explore the effectiveness of the FPA packets to enhance models' robustness through adversarial training and investigate the changes in decision boundaries accordingly.", "date": "2026-01-20", "categories": []}
{"id": "2605.02921", "title": "EvoJail: Evolutionary Diverse Jailbreak Prompt Generation for Large Language Models", "abstract": "As LLMs continue to shape real-world applications, automated jailbreak generation becomes essential to reveal safety weaknesses and guide model improvement. Existing automatic jailbreak generation methods have not yet fully considered two important aspects: adaptability to evolving safety-finetuned models, which affects their effectiveness on newer model versions, and diversity in generated prompts, which can cause narrow or repetitive attack patterns. To address these issues, we propose EvoJail, an instruction-fusion-driven evolutionary jailbreak generation framework that formalizes jailbreak prompt generation as a multi-objective black-box optimization problem and leverages the principles of evolutionary algorithms to search for jailbreak prompts that can adapt across different model versions and exhibit diverse attack patterns. Specifically, EvoJail integrates jailbreak prompt generation into an iterative evolutionary loop, where at each iteration candidate prompts are evaluated directly against the target model and then selected and varied based on the target model's responses, enabling the generation process to continuously adapt to model updates. To enhance diversity, EvoJail introduces field-aware instruction fusion to construct diverse starting points and incorporates diversity-aware objectives into the evolutionary fitness function, guiding the search toward prompts with richer semantic variation, while further designing multi-level LLM-based mutation operators that modify prompt structures at different granularities to promote structural diversity throughout the evolutionary process. Results demonstrate that EvoJail has stronger adaptability and can achieve over $93\\%$ attack success rate and more than $5.6\\%$ improvement in diversity metrics over state-of-the-art methods.", "date": "2026-04-22", "categories": []}
{"id": "2605.02946", "title": "RouteHijack: Routing-Aware Attack on Mixture-of-Experts LLMs", "abstract": "Safety alignment is critical for the responsible deployment of large language models (LLMs). As Mixture-of-Experts (MoE) architectures are increasingly adopted to scale model capacity, understanding their safety robustness becomes essential. Existing adversarial attacks, however, have notable limitations. Prompt-based jailbreaks rely on heuristic search and transfer poorly, model intervention methods require privileged access to internal representations, and optimization-based input attacks remain output-centric and are fundamentally limited to MoE models due to the non-differentiable routing mechanism.\n  In this paper, we present RouteHijack, a routing-aware jailbreak for MoE LLMs. Our key insight is that safety behavior is concentrated in a small subset of experts, creating an opportunity to steer model behavior by influencing routing decisions through input optimization. Building on this observation, RouteHijack first performs response-driven expert localization to identify safety-critical and harmful experts by contrasting activations under safe refusals and harmful completions. It then constructs adversarial suffixes with a routing-aware objective that suppresses safety experts, promotes harmful experts, and prevents early-stage refusal during generation. At inference time, the optimized suffix is appended to a malicious prompt, requiring only input access. Across seven MoE LLMs, RouteHijack achieves a 69.3\\% average attack success rate (ASR), outperforming prior optimization-based attack by $3.2\\times$. RouteHijack also transfers zero-shot across five sibling MoE variants, raising average ASR from 27.7\\% to 61.2\\%, and further generalizes to three MoE-based VLMs, increasing average ASR from 2.47\\% to 38.7\\%. These findings expose a fundamental vulnerability in sparse expert architectures and highlight the need for defenses beyond output-level alignment.", "date": "2026-05-01", "categories": []}
{"id": "2605.03405", "title": "TsallisPGD: Adaptive Gradient Weighting for Adversarial Attacks on Semantic Segmentation", "abstract": "Attacking semantic segmentation models is significantly harder than image classification models because an attacker must flip thousands of pixel predictions simultaneously. Standard pixel-wise cross-entropy (CE) is ill-suited to this setting: it tends to overemphasize already-misclassified pixels, which slows optimization and overstates model robustness. To address these issues, we introduce TsallisPGD, an adversarial attack built on the Tsallis cross-entropy, a generalization of CE parameterized by $q$, which adaptively reshapes the gradient landscape by controlling gradient concentration across pixels. By varying $q$, we steer the attack toward pixels at different confidence levels. We first show that no single fixed-$q$ is universally optimal, as its effectiveness depends on the dataset, model architecture, and perturbation budget. Motivated by this, we propose a dynamic $q$-schedule that sweeps $q$ during optimization. Extensive experiments on Cityscapes, Pascal VOC, and ADE20K show that TsallisPGD, using a single validation-selected schedule, achieves the best average attack rank across all evaluated settings and improves over CEPGD, SegPGD, CosPGD, JSPGD, and MaskedPGD in reducing accuracy and mIoU on both standard and robust models.", "date": "2026-05-05", "categories": []}
{"id": "2605.04019", "title": "Redefining AI Red Teaming in the Agentic Era: From Weeks to Hours", "abstract": "AI systems are entering critical domains like healthcare, finance, and defense, yet remain vulnerable to adversarial attacks. While AI red teaming is a primary defense, current approaches force operators into manual, library-specific workflows. Operators spend weeks hand-crafting workflows - assembling attacks, transforms, and scorers. When results fall short, workflows must be rebuilt. As a result, operators spend more time constructing workflows than probing targets for security and safety vulnerabilities.\n  We introduce an AI red teaming agent built on the open-source Dreadnode SDK. The agent creates workflows grounded in 45+ adversarial attacks, 450+ transforms, and 130+ scorers. Operators can probe multi-agent systems, multilingual, and multimodal targets, focusing on what to probe rather than how to implement it.\n  We make three contributions: 1. Agentic interface. Operators describe goals in natural language via the Dreadnode TUI (Terminal User Interface). The agent handles attack selection, transform composition, execution, and reporting, letting operators focus on red teaming. Weeks compress to hours. 2. Unified framework. A single framework for probing traditional ML models (adversarial examples) and generative AI systems (jailbreaks), removing the need for separate libraries. 3. Llama Scout case study. We red team Meta Llama Scout and achieve an 85% attack success rate with severity up to 1.0, using zero human-developed code", "date": "2026-05-05", "categories": []}
{"id": "2402.05284", "title": "Analyzing Adversarial Inputs in Deep Reinforcement Learning", "abstract": "In recent years, Deep Reinforcement Learning (DRL) has become a popular paradigm in machine learning due to its successful applications to real-world and complex systems. However, even the state-of-the-art DRL models have been shown to suffer from reliability concerns -- for example, their susceptibility to adversarial inputs, i.e., small and abundant input perturbations that can fool the models into making unpredictable and potentially dangerous decisions. This drawback limits the deployment of DRL systems in safety-critical contexts, where even a small error cannot be tolerated. In this work, we present a comprehensive analysis of the characterization of adversarial inputs, through the lens of formal verification. Specifically, we present the Adversarial Rate, a metric adapted from the ProVe family, for the systematic evaluation of adversarial inputs in DRL, which partitions the input domain into subregions to enable both quantification and spatial visualization of adversarial inputs. The main contribution of this work is to provide a comprehensive evaluation framework for the effect of adversarial inputs on DRL policies. We present a set of tools and algorithms for its computation. Our analysis empirically demonstrates how adversarial inputs can affect the safety of a given DRL system with respect to such perturbations. Moreover, we analyze the behavior of these configurations to suggest several useful practices and guidelines to help mitigate the vulnerability of trained DRL networks.", "date": "2024-02-07", "categories": []}
{"id": "2512.10275", "title": "Sample-wise Adaptive Weighting for Transfer Consistency in Adversarial Distillation", "abstract": "Adversarial distillation in the standard min-max adversarial training framework aims to transfer adversarial robustness from a large, robust teacher network to a compact student. However, existing work often neglects to incorporate state-of-the-art robust teachers. Through extensive analysis, we find that stronger teachers do not necessarily yield more robust students-a phenomenon known as robust saturation. While typically attributed to capacity gaps, we show that such explanations are incomplete. Instead, we identify adversarial transferability-the fraction of student-crafted adversarial examples that remain effective against the teacher-as a key factor in successful robustness transfer. Based on this insight, we propose Sample-wise Adaptive Adversarial Distillation (SAAD), which reweights training examples by their measured transferability without incurring additional computational cost. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet show that SAAD consistently improves AutoAttack robustness over prior methods. Our code is available at https://github.com/HongsinLee/saad.", "date": "2025-12-11", "categories": []}
{"id": "2602.07249", "title": "Beyond Crash: Hijacking Your Autonomous Vehicle for Fun and Profit", "abstract": "Autonomous Vehicles (AVs), especially vision-based AVs, are rapidly being deployed without human operators. As AVs operate in safety-critical environments, understanding their robustness in an adversarial environment is an important research problem. Prior physical adversarial attacks on vision-based autonomous vehicles predominantly target immediate safety failures (e.g., a crash, a traffic-rule violation, or a transient lane departure) by inducing a short-lived perception or control error. This paper shows a qualitatively different risk: a long-horizon route integrity compromise, where an attacker gradually steers a victim AV away from its intended route and into an attacker-chosen destination while the victim continues to drive ``normally.'' This will not pose a danger to the victim vehicle itself, but also to potential passengers sitting inside the vehicle, who may not notice the route changes.\n  In this paper, we design and implement the first adversarial framework, called JackZebra, which performs route-level hijacking of a vision-based end-to-end driving stack using a physically plausible attacker vehicle with a reconfigurable display and a camera sensor mounted on the rear. The central challenge is temporal persistence: adversarial influence must remain effective in changing viewpoints, lighting, weather, traffic, and the victim's continual replanning -- without triggering conspicuous failures. Our key insight is to treat route hijacking as a closed-loop control problem and to convert adversarial patches into steering primitives that can be selected online via an interactive adjustment loop based on observed victim behavior using the rear camera. Our evaluations in both simulated and real-world scenarios show that JackZebra can successfully hijack victim vehicles to deviate from original routes and stop at places designated by the adversary with a high success rate.", "date": "2026-02-06", "categories": []}
{"id": "2605.01034", "title": "A Theoretical Game of Attacks via Compositional Skills", "abstract": "As large language models grow increasingly capable, concerns about their safe deployment have intensified. While numerous alignment strategies aim to restrict harmful behavior, these defenses can still be circumvented through carefully designed adversarial prompts. In this work, we introduce a theoretical framework that formalizes a game between an attacker and a defender. Within this framework, we design a theoretical best-response attack strategy and show that it is closely related to many existing adversarial prompting methods. We further analyze the resulting game, characterize its equilibria, and reveal inherent advantages for the attacker. Drawing on our theoretical analysis, we also derive a provably optimal defense strategy. Empirically, we evaluate a practical instantiation of the theoretically optimal attack and observe stronger performance relative to existing adversarial prompting approaches in diverse settings encompassing different LLMs and benchmarks.", "date": "2026-05-01", "categories": []}
{"id": "2605.01098", "title": "Almost for Free: Crafting Adversarial Examples with Convolutional Image Filters", "abstract": "Adversarial examples in machine learning are typically generated using gradients, obtained either directly through access to the model or approximated via queries to it. In this paper, we propose a much simpler approach to craft adversarial examples, drawing inspiration from insights of explainable machine learning. In particular, we design \\emph{adversarial image filters} that are based on classic edge detection algorithms but optimized to deceive learning models. The resulting untargeted attacks are transferable and require only a single pass over the input. Empirically, we find that 3x3 filters already enable success rates between 30% and 80% on different neural networks. Compared to related approaches using generative models for crafting adversarial examples, we reduce the number of parameters by five orders of magnitude, resulting in a very efficient attack. When investigating the parameters of the learned filters, we observe interesting properties such as a high transferability between models and structures common to classic image filters. Our results provide further insights into the vulnerability of neural networks and their fragility to malicious noise.", "date": "2026-05-01", "categories": []}
{"id": "2605.01298", "title": "Checkerboard: A Simple, Effective, Efficient and Learning-free Clean Label Backdoor Attack with Low Poisoning Budget", "abstract": "Backdoor attacks threaten the deep learning supply chain by poisoning a small fraction of the training data so that a model behaves normally on clean inputs but misclassifies trigger-carrying inputs to an attacker-chosen target class. Clean-label backdoor attacks are especially dangerous because poisoned samples remain label-consistent and are therefore harder to detect. Yet existing clean-label attacks typically rely on expensive optimization, surrogate-model training, or nontrivial data access.\n  We present Checkerboard, a theoretically grounded, learning-free clean-label backdoor attack that is effective, efficient, and simple to implement. From a linear separability formulation, we derive a checkerboard trigger in closed form, removing the need for surrogate-model training and trigger optimization. For texture-rich datasets, we introduce Complexity-driven Sample Selection, which uses only target-class data to improve trigger-to-background contrast by selecting low-complexity images for poisoning. Across four benchmark datasets, Checkerboard outperforms 8 baseline attacks and achieves state-of-the-art performance under low poisoning budgets. For example, on CIFAR-10, under a trigger perturbation budget of $10/255$, poisoning 20 training samples achieves $99.99\\%$ Attack Success Rate (ASR). On ImageNet-100, a poisoning rate of only $0.46\\%$ yields over $94\\%$ ASR without degrading clean accuracy. The proposed attack also remains effective against state-of-the-art backdoor defenses and shows strong resistance to adaptive defenses.", "date": "2026-05-02", "categories": []}
{"id": "2605.01449", "title": "VisInject: Disruption != Injection -- A Dual-Dimension Evaluation of Universal Adversarial Attacks on Vision-Language Models", "abstract": "Universal adversarial attacks on aligned multimodal large language models are increasingly reported with attack success rates in the 60-80% range, suggesting the visual modality is highly vulnerable to imperceptible perturbations as a prompt-injection channel. We argue that this number conflates two distinct events: (i) the model's output was perturbed (Influence), and (ii) the attacker's chosen target concept was actually emitted (Precise Injection). We compose two existing techniques -- Universal Adversarial Attack and AnyAttack -- under an $L_{inf}$ budget of 16/255, and we add a dual-axis evaluation: a deterministic Ratcliff-Obershelp drift score for Influence (programmatic baseline) plus a 4-tier ordinal categorical none/weak/partial/confirmed for Precise Injection. The judge is DeepSeek-V4-Pro in thinking mode, calibrated against Claude Opus 4.7 with Cohen's $\\kappa$ = 0.77 on the injection axis (substantial agreement); the entire 4475-entry SHA-256 input cache ships with the dataset so reviewers can re-derive paper numbers bit-exact without an API key. Across 6615 pairs over four open VLMs, seven attack prompts, and seven test images, the two axes diverge by roughly 90$\\times$: 66.4% of pairs are programmatically disturbed (LLM-judged 46.6% at the substantial-or-complete tier), but only 0.756% (50/6615) reach any non-none injection tier and only 0.030% (2/6615) verbatim. The few injections that do land cluster on screenshot- or document-style carriers whose semantics already invite text transcription. BLIP-2 shows \\emph{zero detectable drift} at $L_{inf}$ = 16/255 across all 2205 pairs even when used as a Stage-1 surrogate. We release the full dataset -- 21 universal images, 147 adversarial photos, 6,615 response pairs, the v3 dual-axis judge results, and the cache at huggingface.co/datasets/jeffliulab/visinject.", "date": "2026-05-02", "categories": []}
{"id": "2605.01519", "title": "Certified vs. Empirical Adversarial Robust-ness via Hybrid Convolutions with Attention Stochasticity", "abstract": "We introduce Hybrid Convolutions with Attention Stochasticity (HyCAS), an adversarial defense that narrows the long-standing gap between provable robustness under L2 certificates and empirical robustness against strong L attacks, while preserving strong generalization across diverse imaging benchmarks. HyCAS unifies deterministic and randomized principles by coupling 1-Lipschitz, spectrally normalized convolutions with two stochastic components, spectral normalized random, projection filters and a randomized attention-noise mechanism, to realize a randomized defense. Injecting smoothing randomness inside the architecture yields an overall <= 2-Lipschitz network with formal certificates. Exten-sive experiments on diverse imaging benchmarks, including CIFAR-10/100, ImageNet-1k, NIH Chest X-ray, HAM10000, show that HyCAS surpasses prior leading certified and empirical defenses, boosting certified accuracy by up to 7.3% (on NIH Chest X-ray) and empirical robustness by up to 3.1% (on HAM10000), without sacrificing clean accuracy. These results show that a randomized Lipschitz constrained architecture can simultaneously improve both certified L2 and empirical L adversarial robustness, thereby supporting safer deployment of deep models in high-stakes applications. Code: https://github.com/misti1203/HyCAS", "date": "2026-05-02", "categories": []}
{"id": "2605.01591", "title": "Led to Mislead: Adversarial Content Injection for Attacks on Neural Ranking Models", "abstract": "Neural Ranking Models (NRMs) are central to modern information retrieval but remain highly vulnerable to adversarial manipulation. Existing attacks often rely on heuristics or surrogate models, limiting effectiveness and transferability. We propose CRAFT, a supervised framework for black-box adversarial rank attacks powered by large language models (LLMs). CRAFT operates in three stages: adversarial dataset generation via retrieval-augmented generation and self-refinement, supervised fine-tuning on curated adversarial examples, and preference-guided optimization to align generations with rank-promotion objectives. Extensive experiments on the MS MARCO passage dataset, TREC Deep Learning 2019, and TREC Deep Learning 2020 benchmarks show that CRAFT significantly outperforms state-of-the-art baselines, achieving higher promotion rates and rank boosts while preserving fluency and semantic fidelity. Moreover, CRAFT transfers effectively across diverse ranking architectures, including cross-encoder, embedding-based, and LLM-based rankers, underscoring vulnerabilities in real-world retrieval systems. This work provides a principled framework for studying adversarial threats in NRMs, underscores the risks of generative AI in rank manipulation, and provides a foundation for developing more robust retrieval systems. To support reproducibility, we publicly release our source code, trained models, and prompt templates.", "date": "2026-05-02", "categories": []}
{"id": "2605.01718", "title": "Dual-branch Robust Unlearnable Examples", "abstract": "Unlearnable examples (UEs) aim to compromise model training by injecting imperceptible perturbations to clean samples. However, existing UE schemes exhibit limited robustness against advanced defenses due to their heuristic design or narrowly scoped domain perturbations. To address this, we propose \\texttt{DUNE}, a \\underline{\\textbf{D}}ual-branch \\underline{\\textbf{UN}}learnable \\underline{\\textbf{E}}nsemble perturbation optimization approach. Specifically, \\texttt{DUNE} separately optimizes perturbations in the spatial and color domains to establish the mapping between perturbations and shift-induced labels. This design extends the perturbation domain to increase noise intensity for improving robustness and drives the models to learn perturbation-oriented features with degraded generalization, thereby achieving unlearnability. To strengthen \\texttt{DUNE}'s performance, we further propose an unlearnability-enhancing ensemble strategy that aggregates diverse pre-trained models during the dual-branch optimization. Extensive experiments on benchmark datasets CIFAR-10 and ImageNet verify that \\texttt{DUNE}'s robustness outperforms 12 SOTA UE schemes under 7 mainstream defenses, yielding a lower average test accuracy of 14.95\\% to 50.82\\%.", "date": "2026-05-03", "categories": []}
{"id": "2605.01899", "title": "Disentangling Intent from Role: Adversarial Self-Play for Persona-Invariant Safety Alignment", "abstract": "The growing capabilities of large language models (LLMs) have driven their widespread deployment across diverse domains, even in potentially high-risk scenarios. Despite advances in safety alignment techniques, current models remain vulnerable to emerging persona-based jailbreak attacks. Existing research on persona-based jailbreak has primarily focused on attack iterations, yet it lacks systemic and mechanistic constraints on the defense side. To address this challenge, we propose Persona-Invariant Alignment (PIA), an adversarial self-play framework that achieves co-evolution through Persona Lineage Evolution (PLE) on the attack side and Persona-Invariant Consistency Learning (PICL) on the defense side. Theoretically, PICL is grounded in the structural separation hypothesis, using a unilateral KL-divergence constraint to enable the structural decoupling of safety decisions from persona context, thereby maintaining safe behavior under persona-based jailbreak attacks. Experimental results demonstrate that PLE efficiently explores high-risk persona spaces by leveraging lineage-based credit propagation. Meanwhile, the PICL defense method significantly reduces the Attack Success Rate (ASR) while preserving the model's general capability, thereby validating the superiority and robustness of this alignment paradigm. Codes are available at https://github.com/JiajiaLi-1130/PIA.", "date": "2026-05-03", "categories": []}
{"id": "2605.02109", "title": "Detecting Adversarial Data via Provable Adversarial Noise Amplification", "abstract": "The nonuniform and growing impact of adversarial noise across the layers of deep neural networks has been used in the literature, without a formal mathematical justification, to detect adversarial inputs and improve robustness. In this work, we study this phenomenon in detail and present a formal adversarial noise amplification theorem. We specify a set of sufficient conditions under which the adversarial noise amplification is mathematically guaranteed. Based on theoretical observations, we propose a novel training methodology with a custom spectral loss function and a specific architectural design to enhance the amplification signal for detecting adversarial data. Finally, we introduce a new, lightweight detection mechanism that leverages the enhanced amplification signal and operates entirely at inference time. To validate our approach, we demonstrate the detector's efficacy against both state-of-the-art attacks and a purpose-built adaptive attack, confirming that enhanced amplification can serve as a robust and reliable signal for adversarial defense.", "date": "2026-05-04", "categories": []}
{"id": "2605.02183", "title": "Manifold-Constrained Adversarial Training for Long-Tailed Robustness via Geometric Alignment", "abstract": "Adversarial training is effective on balanced datasets, but its robustness degrades under longtailed class distributions, where tail classes suffer high robust error and unstable decision boundaries. We propose Manifold-Constrained Adversarial Training (MCAT), a unified framework that enforces the semantic validity of adversarial examples by penalizing deviations from class-conditional manifolds in feature space, while promoting balanced geometric separation across classes via an ETF-inspired regularization. We provide theoretical results that link geometric separation to lower bounds on adversarially robust margins, and show that manifold-constrained adversarial risk upperbounds robust risk on high-density semantic regions. Extensive experiments on standard longtailed benchmarks demonstrate consistent improvements in overall, balanced, and tail-class adversarial robustness.", "date": "2026-05-04", "categories": []}
{"id": "2605.02374", "title": "Fight Poison with Poison: Enhancing Robustness in Few-shot Machine-Generated Text Detection with Adversarial Training", "abstract": "Machine-generated text (MGT) detection is critical for regulating online information ecosystems, yet existing detectors often underperform in few-shot settings and remain vulnerable to adversarial, humanizing attacks. To build accurate and robust detectors under limited supervision, we adopt a threat-modeling perspective and study detector vulnerabilities from an attacker's viewpoint under an output-only black-box setting. Motivated by this perspective, we propose RAG-GuidEd Attacker Strengthens ConTrastive Few-shot Detector (REACT), an adversarial training framework that improves both few-shot detection performance and robustness against attacks. REACT couples a humanization-oriented attacker with a target detector: the attacker leverages retrieval-augmented generation (RAG) to craft highly human-like adversarial examples to evade detection, while the detector learns from these adversaries with a contrastive objective to stabilize few-shot representation learning and enhance robustness. We alternately update the attacker and the detector to enable their co-evolution. Experiments on 4 datasets with 4 shot sizes and 3 random seeds show that REACT improves average detection F1 by 4.95 points over 8 state-of-the-art (SOTA) detectors and reduces the average attack success rate (ASR) under 4 strong attacks by 3.66 percentage points.", "date": "2026-05-04", "categories": []}
{"id": "2506.22982", "title": "Revisiting CroPA: A Reproducibility Study and Enhancements for Cross-Prompt Adversarial Transferability in Vision-Language Models", "abstract": "Large Vision-Language Models (VLMs) have revolutionized computer vision, enabling tasks such as image classification, captioning, and visual question answering. However, they remain highly vulnerable to adversarial attacks, particularly in scenarios where both visual and textual modalities can be manipulated. In this study, we conduct a comprehensive reproducibility study of \"An Image is Worth 1000 Lies: Adversarial Transferability Across Prompts on Vision-Language Models\" validating the Cross-Prompt Attack (CroPA) and confirming its superior cross-prompt transferability compared to existing baselines. Beyond replication we propose several key improvements: (1) A novel initialization strategy that significantly improves Attack Success Rate (ASR). (2) Investigate cross-image transferability by learning universal perturbations. (3) A novel loss function targeting vision encoder attention mechanisms to improve generalization. Our evaluation across prominent VLMs -- including Flamingo, BLIP-2, and InstructBLIP as well as extended experiments on LLaVA validates the original results and demonstrates that our improvements consistently boost adversarial effectiveness. Our work reinforces the importance of studying adversarial vulnerabilities in VLMs and provides a more robust framework for generating transferable adversarial examples, with significant implications for understanding the security of VLMs in real-world applications.", "date": "2025-06-28", "categories": []}
{"id": "2602.03265", "title": "Beyond Suffixes: Token Position in GCG Adversarial Attacks on Large Language Models", "abstract": "Large Language Models (LLMs) have seen widespread adoption across multiple domains, creating an urgent need for robust safety alignment mechanisms. However, robustness remains challenging due to jailbreak attacks that bypass alignment via adversarial prompts. In this work, we focus on the prevalent Greedy Coordinate Gradient (GCG) attack and identify a previously underexplored attack axis in jailbreak attacks typically framed as suffix-based: the placement of adversarial tokens within the prompt. Using GCG as a case study, we show that both optimizing attacks to generate prefixes instead of suffixes and varying adversarial token position during evaluation substantially influence attack success rates. Our findings highlight a critical blind spot in current safety evaluations and underline the need to account for the position of adversarial tokens in the adversarial robustness evaluation of LLMs.", "date": "2026-02-03", "categories": []}
{"id": "2605.00236", "title": "Attention Is Where You Attack", "abstract": "Safety-aligned large language models rely on RLHF and instruction tuning to refuse harmful requests, yet the internal mechanisms implementing safety behavior remain poorly understood. We introduce the Attention Redistribution Attack (ARA), a white-box adversarial attack that identifies safety-critical attention heads and crafts nonsemantic adversarial tokens that redirect attention away from safety-relevant positions. Unlike prior jailbreak methods operating at the semantic or output-logit level, ARA targets the geometry of softmax attention on the probability simplex using Gumbel-softmax optimization over targeted heads. Across LLaMA-3-8B-Instruct, Mistral-7B-Instruct-v0.1, and Gemma-2-9B-it, ARA bypasses safety alignment with as few as 5 tokens and 500 optimization steps, achieving 36% ASR on Mistral-7B and 30% on LLaMA-3 against 200 HarmBench prompts, while Gemma-2 remains at 1%. Our principal mechanistic finding is a dissociation between ablation and redistribution: zeroing out the top-ranked safety heads produces at most 1 flip among 39 to 50 baseline refusals, while ARA targeting the corresponding safety-heavy layers flips 72/200 prompts on Mistral-7B and 60/200 on LLaMA-3. This suggests that safety is not localized in these heads as removable components, but emerges from the attention routing they perform. Removing a head allows compensation through the residual stream, while redirecting its attention propagates a corrupted signal downstream.", "date": "2026-04-30", "categories": []}
{"id": "2605.00443", "title": "Adaptive Equilibrium: Dynamic Weighting Framework for Generalized Interruption of DeepFake Models", "abstract": "The advancement of generalized deepfake disruption is constrained by the interruption imbalance, a fundamental bottleneck inherent to the generation of universal perturbations. We reveal that conventional static gradient normalization fundamentally struggles to resolve architectural conflicts, causing the optimization to bias towards susceptible models while neglecting resistant ones. We argue that achieving high and uniform effectiveness requires resolving this imbalance by reaching an adaptive equilibrium. We propose the Adaptive Equilibrium Framework (AEF), which employs a dynamic weighting mechanism that utilizes real-time loss feedback to adaptively assign greater interruption weights to the most resistant models. This approach shifts the optimization from an average-case problem to finding a dynamic balance, driving the perturbation to a uniformly effective equilibrium state. Comprehensive experiments validate that AEF achieves a more balanced interruption performance, maintaining a consistent interruption success rate across the evaluated diverse architectures.", "date": "2026-05-01", "categories": []}
{"id": "2306.16050", "title": "Evaluating Similitude and Robustness of Deep Image Denoising Models via Adversarial Attack", "abstract": "Deep neural networks (DNNs) have shown superior performance comparing to traditional image denoising algorithms. However, DNNs are inevitably vulnerable while facing adversarial attacks. In this paper, we propose an adversarial attack method named denoising-PGD which can successfully attack all the current deep denoising models while keep the noise distribution almost unchanged. We surprisingly find that the current mainstream non-blind denoising models (DnCNN, FFDNet, ECNDNet, BRDNet), blind denoising models (DnCNN-B, Noise2Noise, RDDCNN-B, FAN), plug-and-play (DPIR, CurvPnP) and unfolding denoising models (DeamNet) almost share the same adversarial sample set on both grayscale and color images, respectively. Shared adversarial sample set indicates that all these models are similar in term of local behaviors at the neighborhood of all the test samples. Thus, we further propose an indicator to measure the local similarity of models, called robustness similitude. Non-blind denoising models are found to have high robustness similitude across each other, while hybrid-driven models are also found to have high robustness similitude with pure data-driven non-blind denoising models. According to our robustness assessment, data-driven non-blind denoising models are the most robust. We use adversarial training to complement the vulnerability to adversarial attacks. Moreover, the model-driven image denoising BM3D shows resistance on adversarial attacks.", "date": "2023-06-28", "categories": []}
{"id": "2501.19143", "title": "Imitation Game for Adversarial Disillusion with Chain-of-Thought Reasoning in Generative AI", "abstract": "As the cornerstone of artificial intelligence, machine perception confronts a fundamental threat posed by adversarial illusions. These adversarial attacks manifest in two primary forms: deductive illusion, where specific stimuli are crafted based on the victim model's general decision logic, and inductive illusion, where the victim model's general decision logic is shaped by specific stimuli. The former exploits the model's decision boundaries to create a stimulus that, when applied, interferes with its decision-making process. The latter reinforces a conditioned reflex in the model, embedding a backdoor during its learning phase that, when triggered by a stimulus, causes aberrant behaviours. The multifaceted nature of adversarial illusions calls for a unified defence framework, addressing vulnerabilities across various forms of attack. In this study, we propose a disillusion paradigm based on the concept of an imitation game. At the heart of the imitation game lies a multimodal generative agent, steered by chain-of-thought reasoning, which observes, internalises and reconstructs the semantic essence of a sample, liberated from the classic pursuit of reversing the sample to its original state. As a proof of concept, we conduct experimental simulations using a multimodal generative dialogue agent and evaluates the methodology under a variety of attack scenarios. Experimental results demonstrate that the proposed framework consistently neutralises both deductive and inductive adversarial illusions across diverse white-box and black-box attack scenarios.", "date": "2025-01-31", "categories": []}
{"id": "2509.21087", "title": "Are Modern Speech Enhancement Systems Vulnerable to Adversarial Attacks?", "abstract": "Machine learning approaches for speech enhancement are becoming increasingly expressive, enabling ever more powerful modifications of input signals. In this paper, we demonstrate that this expressiveness introduces a vulnerability: advanced speech enhancement models can be susceptible to adversarial attacks. Specifically, we show that adversarial noise, carefully crafted and psychoacoustically masked by the original input, can be injected such that the enhanced speech output conveys an entirely different semantic meaning. We experimentally verify that contemporary predictive speech enhancement models can indeed be manipulated in this way. Furthermore, we highlight that diffusion models with stochastic samplers exhibit inherent robustness to such adversarial attacks by design.", "date": "2025-09-25", "categories": []}
{"id": "2512.22046", "title": "Backdoor Attacks on Prompt-Driven Video Segmentation Foundation Models", "abstract": "Prompt-driven Video Segmentation Foundation Models (VSFMs), such as SAM2, are increasingly used in applications including autonomous driving and digital pathology, yet their security risks remain underexplored. We study backdoor attacks against VSFMs and show that directly applying classic attacks such as BadNet is largely ineffective, yielding attack success rates (ASR) below 5%. Through gradient-similarity and attention-map analyses, we find that traditional backdoor training fails because clean and triggered samples induce aligned image-encoder gradients, while model attention remains focused on the prompt-specified object rather than the trigger. To address this limitation, we propose BadVSFM, the first backdoor attack framework tailored to prompt-driven VSFMs. BadVSFM uses a two-stage strategy that first learns trigger-specific encoder features and then trains the decoder to map triggered frame prompt representations to an attacker-specified target mask while preserving clean segmentation behavior. Experiments on five VSFMs and two datasets show that BadVSFM achieves strong, controllable backdoor effects across triggers and prompt types with limited clean-performance degradation. Ablations and interpretability analyses validate the necessity of the two-stage design, and five representative defenses remain largely ineffective. Our results reveal a practical and underexplored vulnerability of current VSFMs to backdoor threats.", "date": "2025-12-26", "categories": []}
{"id": "2604.27487", "title": "Low Rank Adaptation for Adversarial Perturbation", "abstract": "Low-Rank Adaptation (LoRA), which leverages the insight that model updates typically reside in a low-dimensional space, has significantly improved the training efficiency of Large Language Models (LLMs) by updating neural network layers using low-rank matrices. Since the generation of adversarial examples is an optimization process analogous to model training, this naturally raises the question: Do adversarial perturbations exhibit a similar low-rank structure?\n  In this paper, we provide both theoretical analysis and extensive empirical investigation across various attack methods, model architectures, and datasets to show that adversarial perturbations indeed possess an inherently low-rank structure. This insight opens up new opportunities for improving both adversarial attacks and defenses. We mainly focus on leveraging this low-rank property to improve the efficiency and effectiveness of black-box adversarial attacks, which often suffer from excessive query requirements. Our method follows a two-step approach. First, we use a reference model and auxiliary data to guide the projection of gradients into a low-dimensional subspace. Next, we confine the perturbation search in black-box attacks to this low-rank subspace, significantly improving the efficiency and effectiveness of the adversarial attacks.\n  We evaluated our approach across a range of attack methods, benchmark models, datasets, and threat models. The results demonstrate substantial and consistent improvements in the performance of our low-rank adversarial attacks compared to conventional methods.", "date": "2026-04-30", "categories": []}
{"id": "2604.28129", "title": "Latent Adversarial Detection: Adaptive Probing of LLM Activations for Multi-Turn Attack Detection", "abstract": "Multi-turn prompt injection follows a known attack path -- trust-building, pivoting, escalation but text-level defenses miss covert attacks where individual turns appear benign. We show this attack path leaves an activation-level signature in the model's residual stream: each phase shift moves the activation, producing a total path length far exceeding benign conversations. We call this adversarial restlessness. Five scalar trajectory features capturing this signal lift conversation-level detection from 76.2% to 93.8% on synthetic held-out data. The signal replicates across four model families (24B-70B); probes are model-specific and do not transfer across architectures. Generalization is source-dependent: leave-one-source-out evaluation shows each of synthetic, LMSYS-Chat-1M, and SafeDialBench captures distinct attack distributions, with detection on real-world LMSYS reaching 47-71% when its distribution is represented in training. Combined three-source training achieves 89.4% detection at 2.4% false positive rate on a held-out mixed set. We further show that three-phase turn-level labels(benign/pivoting/adversarial) unique to our synthetic dataset are essential: binary conversation-level labels produce 50-59% false positives. These results establish adversarial restlessness as a reliable activation-level signal and characterize the data requirements for practical deployment.", "date": "2026-04-30", "categories": []}
{"id": "2604.26317", "title": "The Unseen Adversaries: Robust and Generalized Defense Against Adversarial Patches", "abstract": "The vulnerabilities of deep neural networks against singularities have raised serious concerns regarding their deployment in the physical world. One of the most prominent and impactful physical-world adversarial perturbations is the attachment of patches to clean images, known as an adversarial patch attack. Similarly, natural noises such as Gaussian and Salt\\&Pepper are highly prevalent in the real world. The current research need arises from the above vulnerabilities and the lack of efforts to tackle these two singularities independently and, especially, in combination. In this research, we have, for the first time, combined these two prominent singularities and proposed a novel dataset. Using this dataset, we have conducted a benchmark study of singularity data-point detection using features from several convolutional neural networks. For classification, rather than the popular neural network-based parameter tuning, we have used traditional yet effective machine learning classifiers. The extensive experiments across various in- and out-of-distribution (OOD) singularities reveal several interesting findings about the effectiveness of classifiers and show that it is hard to defend against adversaries when they are treated independently, and inefficient classifiers are selected.", "date": "2026-04-29", "categories": []}
{"id": "2604.26496", "title": "Robust Alignment: Harmonizing Clean Accuracy and Adversarial Robustness in Adversarial Training", "abstract": "Adversarial Training (AT) is one of the most effective methods for developing robust deep neural networks (DNNs). However, AT faces a trade-off problem between clean accuracy and adversarial robustness. In this work, we reveal a surprising phenomenon for the first time: Varying input perturbation intensities for training samples near decision boundaries in AT have minimal impact on model robustness. This finding directly exposes the inconsistency between accuracy and robustness score fluctuations, leading us to identify the misalignment between input and latent spaces as a critical driver of the robustness-accuracy trade-off. To mitigate this misalignment for harmonizing accuracy and robustness, we define Robust Alignment as a new AT target, encouraging the model perception to change with input perturbations provided the final label prediction remains unchanged, which can be achieved via two novel ideas. First, we suggest a reduced and fixed perturbation intensity for those boundary samples, which facilitates the model to utilize the perturbations as learnable patterns, instead of noises that complicate decision boundaries meaninglessly. Second, we propose a Domain Interpolation Consistency Adversarial Regularization (DICAR), based on rigorous theoretical derivations, which explicitly introduces semantic alignment between input and latent spaces into AT. Based on these two ideas, we end up with a new Robust Alignment Adversarial Training (RAAT) method, effectively harmonizing accuracy and robustness. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-28-10 demonstrate the effectiveness of RAAT in improving the trade-off beyond four common baselines and a total of 14 related state-of-the-art (SOTA) works.", "date": "2026-04-29", "categories": []}
{"id": "2512.20677", "title": "Learning-Based Automated Adversarial Red-Teaming for Robustness Evaluation of Large Language Models", "abstract": "The increasing deployment of large language models (LLMs) in safety-critical applications raises fundamental challenges in systematically evaluating robustness against adversarial behaviors. Existing red-teaming practices are largely manual and expert-driven, which limits scalability, reproducibility, and coverage in high-dimensional prompt spaces. We formulate automated LLM red-teaming as a structured adversarial search problem and propose a learning-driven framework for scalable vulnerability discovery. The approach combines meta-prompt-guided adversarial prompt generation with a hierarchical execution and detection pipeline, enabling standardized evaluation across six representative threat categories, including reward hacking, deceptive alignment, data exfiltration, sandbagging, inappropriate tool use, and chain-of-thought manipulation. Extensive experiments on GPT-OSS-20B identify 47 vulnerabilities, including 21 high-severity failures and 12 previously undocumented attack patterns. Compared with manual red-teaming under matched query budgets, our method achieves a 3.9$\\times$ higher discovery rate with 89\\% detection accuracy, demonstrating superior coverage, efficiency, and reproducibility for large-scale robustness evaluation.", "date": "2025-12-21", "categories": []}
{"id": "2501.02450", "title": "GCP: Guarded Collaborative Perception with Spatial-Temporal Aware Malicious Agent Detection", "abstract": "Collaborative perception significantly enhances autonomous driving safety by extending each vehicle's perception range through message sharing among connected and autonomous vehicles. Unfortunately, it is also vulnerable to adversarial message attacks from malicious agents, resulting in severe performance degradation. While existing defenses employ hypothesis-and-verification frameworks to detect malicious agents based on single-shot outliers, they overlook temporal message correlations, which can be circumvented by subtle yet harmful perturbations in model input and output spaces. This paper reveals a novel blind area confusion (BAC) attack that compromises existing single-shot outlier-based detection methods. As a countermeasure, we propose GCP, a Guarded Collaborative Perception framework based on spatial-temporal aware malicious agent detection, which maintains single-shot spatial consistency through a confidence-scaled spatial concordance loss, while simultaneously examining temporal anomalies by reconstructing historical bird's eye view motion flows in low-confidence regions. We also employ a joint spatial-temporal Benjamini-Hochberg test to synthesize dual-domain anomaly results for reliable malicious agent detection. Extensive experiments demonstrate GCP's superior performance under diverse attack scenarios, achieving up to 34.69% improvements in AP@0.5 compared to the state-of-the-art CP defense strategies under BAC attacks, while maintaining consistent 5-8% improvements under other typical attacks. Code will be released at https://github.com/yihangtao/GCP.git.", "date": "2025-01-05", "categories": []}
{"id": "2505.12009", "title": "LatentStealth: Unnoticeable and Efficient Adversarial Attacks on Expressive Human Pose and Shape Estimation", "abstract": "Expressive human pose and shape estimation (EHPS) plays a central role in digital human generation, particularly in live-streaming applications. However, most existing EHPS models focus primarily on minimizing estimation errors, with limited attention on potential security vulnerabilities, such as generating inappropriate content, violent actions, or racially offensive gestures and expressions. Current adversarial attacks on EHPS models often generate visually conspicuous perturbations, limiting their practicality and ability to expose real-world security threats. To address this limitation, we propose an unnoticeable adversarial method, termed \\textbf{LatentStealth}, specifically tailored for EHPS models. The key idea is to exploit the structured latent representations of natural images as the medium for crafting perturbations. Instead of injecting noise directly into the pixel space, our method projects inputs into the latent space, where adversarial patterns are generated and progressively refined along optimized directions. This latent-space manipulation enables the attack to maintain high imperceptibility while preserving its effectiveness. Furthermore, as the optimization process is guided by only a small number of model output queries, the framework achieves competitive attack performance with low computational overhead, making it both practical and efficient for real-world scenarios. Extensive experiments on the 3DPW and UBody datasets demonstrate the superiority of LatentStealth, revealing critical vulnerabilities in current systems. These findings highlight the urgent need to address and mitigate security risks in digital human generation technologies.", "date": "2025-05-17", "categories": []}
{"id": "2510.10486", "title": "MEASER: Malware embedding attacks on open-source LLMs", "abstract": "Open-source large language models (LLMs) have demonstrated considerable dominance over proprietary LLMs in resolving neural processing tasks, thanks to the collaborative and sharing nature. Although full access to source codes, model parameters, and training data lays the groundwork for transparency, we argue that such a full-access manner is vulnerable to MEAs, and their ill-effects are not fully understood. In this paper, we conduct a systematic formalization for MEAs on open-source LLMs by enumerating all possible threat models associated with adversary objectives, knowledge, and capabilities. Therein, the threat posed by adversaries with internal knowledge, who inject payloads and triggers during the model sharing phase, is of practical interest. We go even further and propose the first MEA against open-source LLMs, dubbed MEASER, which wields impacts through identifying targeted parameters, embedding payloads, injecting triggers, and executing payloads sequentially. Particularly, MEASER enhances the attack robustness against quantization and parameter-efficient fine-tuning (PEFT) by employing the Magnitude-Adaptive Relative Quantization Index Modulation (MAR-QIM) mechanism, synergized with LDPC codes and spread spectrum modulation. In addition, to achieve stealthiness, MEASER devises the performance-aware importance metric to identify targeted parameters with the least degradation of model performance. Extensive experiments on four popular open-source LLMs show that the stealth rate of MEASER outperforms existing MEAs (for general DNNs) significantly, while consistently achieving a 0 bit error rate (BER) in all settings. Moreover, MEASER also maintains superior stealthiness on quantized models. We appeal for investigations on countermeasures against MEASER in view of the significant attack effectiveness.", "date": "2025-10-12", "categories": []}
{"id": "2512.20164", "title": "AI Security Beyond Core Domains: Resume Screening as a Case Study of Adversarial Vulnerabilities in Specialized LLM Applications", "abstract": "Large Language Models (LLMs) excel at text comprehension and generation, making them ideal for automated tasks like code review and content moderation. However, our research identifies a vulnerability: LLMs can be manipulated by \"adversarial instructions\" hidden in input data, such as resumes or code, causing them to deviate from their intended task. Notably, while defenses may exist for mature domains such as code review, they are often absent in other common applications such as resume screening and peer review. This paper introduces a benchmark to assess this vulnerability in resume screening, revealing attack success rates exceeding 80% for certain attack types. We evaluate two defense mechanisms: prompt-based defenses achieve 10.1% attack reduction with 12.5% false rejection increase, while our proposed FIDS (Foreign Instruction Detection through Separation) using LoRA adaptation achieves 15.4% attack reduction with 10.4% false rejection increase. The combined approach provides 26.3% attack reduction, demonstrating that training-time defenses outperform inference-time mitigations in both security and utility preservation.", "date": "2025-12-23", "categories": []}
{"id": "2601.17536", "title": "OTI: A Model-free and Visually Interpretable Measure of Image Attackability", "abstract": "Despite the tremendous success of neural networks, benign images can be corrupted by adversarial perturbations to deceive these models. Intriguingly, images differ in their attackability. Specifically, given an attack configuration, some images are easily corrupted, whereas others are more resistant. Evaluating image attackability has important applications in active learning, adversarial training, and attack enhancement. This prompts a growing interest in developing attackability measures. However, existing methods are scarce and suffer from two major limitations: (1) They rely on a model proxy to provide prior knowledge (e.g., gradients or minimal perturbation) to extract model-dependent image features. Unfortunately, in practice, many task-specific models are not readily accessible. (2) Extracted features characterizing image attackability lack visual interpretability, obscuring their direct relationship with the images. To address these, we propose a novel Object Texture Intensity (OTI), a model-free and visually interpretable measure of image attackability, which measures image attackability as the texture intensity of the image's semantic object. Theoretically, we describe the principles of OTI from the perspectives of decision boundaries as well as the mid- and high-frequency characteristics of adversarial perturbations. Comprehensive experiments demonstrate that OTI is effective and computationally efficient. In addition, our OTI provides the adversarial machine learning community with a visual understanding of attackability.", "date": "2026-01-24", "categories": []}
{"id": "2604.20833", "title": "AVISE: Framework for Evaluating the Security of AI Systems", "abstract": "As artificial intelligence (AI) systems are increasingly deployed across critical domains, their security vulnerabilities pose growing risks of high-profile exploits and consequential system failures. Yet systematic approaches to evaluating AI security remain underdeveloped. In this paper, we introduce AVISE (AI Vulnerability Identification and Security Evaluation), a modular open-source framework for identifying vulnerabilities in and evaluating the security of AI systems and models. As a demonstration of the framework, we extend the theory-of-mind-based multi-turn Red Queen attack into an Adversarial Language Model (ALM) augmented attack and develop an automated Security Evaluation Test (SET) for discovering jailbreak vulnerabilities in language models. The SET comprises 25 test cases and an Evaluation Language Model (ELM) that determines whether each test case was able to jailbreak the target model, achieving 92% accuracy, an F1-score of 0.91, and a Matthews correlation coefficient of 0.83. We evaluate nine recently released language models of diverse sizes with the SET and find that all are vulnerable to the augmented Red Queen attack to varying degrees. AVISE provides researchers and industry practitioners with an extensible foundation for developing and deploying automated SETs, offering a concrete step toward more rigorous and reproducible AI security evaluation.", "date": "2026-04-22", "categories": []}
{"id": "2604.21395", "title": "Supervised Learning Has a Necessary Geometric Blind Spot: Theory, Consequences, and Minimal Repair", "abstract": "PGD adversarial training, the standard robustness method, can reduce Jacobian Frobenius norm yet worsen clean-input geometry (e.g., TDI 1.336 vs. ERM 1.093). We show this is not an implementation artifact but a theorem-level consequence of supervised learning.\n  We prove that any encoder minimizing supervised loss must retain non-zero sensitivity along directions correlated with training labels, including directions that are nuisance at test time. This holds across proper scoring rules, architectures, and dataset sizes. We call this the geometric blind spot of supervised learning.\n  This theorem unifies four empirical phenomena often treated separately: non-robust features, texture bias, corruption fragility, and the robustness-accuracy tradeoff. It also explains why suppressing sensitivity in one adversarial direction can redistribute sensitivity elsewhere.\n  We introduce Trajectory Deviation Index (TDI), a diagnostic of geometric isotropy. Unlike CKA, intrinsic dimension, or Jacobian Frobenius norm alone, TDI captures the failure mode above. In our experiments, PGD attains low Frobenius norm but high TDI, while PMH attains the lowest TDI with one additional training term and no architectural changes.\n  Across seven tasks, BERT/SST-2, and ImageNet ViT-B/16 (backbone family underlying CLIP/DINO/SAM), the blind spot is measurable and repairable. It appears at foundation-model scale, worsens with model scale and task-specific fine-tuning, and is substantially reduced by PMH. PMH also leads on non-Gaussian corruption types (blur/brightness/contrast) without corruption-specific training.", "date": "2026-04-23", "categories": []}
{"id": "2604.22853", "title": "FastAT Benchmark: A Comprehensive Framework for Fair Evaluation of Fast Adversarial Training Methods", "abstract": "Fast Adversarial Training (FastAT) seeks to achieve adversarial robustness at a fraction of the computational cost incurred by standard multi-step methods such as PGD-AT. Although numerous FastAT techniques have been proposed in recent years, fair comparison among them remains elusive. Existing benchmarks and public leaderboards typically permit diverse model architectures, varying training configurations, and external data sources, making it unclear whether reported improvements reflect genuine algorithmic advances or merely more favorable experimental conditions. To address this problem, we introduce the FastAT Benchmark, a controlled evaluation framework built on three core design principles: unified architecture requirements, standardized training settings, and strict prohibition of external or synthetic data. The benchmark implements over twenty representative FastAT methods within a single codebase, enabling direct and reproducible comparison. Each method is assessed through a dual-metric evaluation framework that measures both adversarial robustness (accuracy under PGD, AutoAttack, and CR Attack) and computational cost (GPU training time and peak memory footprint). Comprehensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet provide reliable baseline measurements and reveal that well-designed single-step methods can match or surpass PGD-AT robustness at substantially lower cost, while no single method dominates across all evaluation dimensions. The complete benchmark, including source code, configuration files, and experimental results, is publicly available to support transparent and fair evaluation of future FastAT research.", "date": "2026-04-22", "categories": []}
{"id": "2604.23067", "title": "Training a General Purpose Automated Red Teaming Model", "abstract": "Automated methods for red teaming LLMs are an important tool to identify LLM vulnerabilities that may not be covered in static benchmarks, allowing for more thorough probing. They can also adapt to each specific LLM to discover weaknesses unique to it. Most current automated red teaming methods are intended for tackling safety and content moderation. Thus, they make use of content safety models as evaluators and optimize for circumventing them, and as such, have not been tested with other adversarial intents not typically captured by these. We propose a pipeline for training a red teaming model that can generalize to arbitrary adversarial goals, including objectives it has not been directly trained on, and that does not depend on the existence of a pre-existing evaluator available at training time. We demonstrate that finetuning small models, such as Qwen3-8B, using this pipeline results in a substantial improvement in their ability to generate attacks for both in and out of domain adversarial goals.", "date": "2026-04-24", "categories": []}
{"id": "2604.23105", "title": "Transferable Physical-World Adversarial Patches Against Object Detection in Autonomous Driving", "abstract": "Deep learning drives major advances in autonomous driving (AD), where object detectors are central to perception. However, adversarial attacks pose significant threats to the reliability and safety of these systems, with physical adversarial patches representing a particularly potent form of attack. Physical adversarial patch attacks pose severe risks but are usually crafted for a single model, yielding poor transferability to unseen detectors. We propose AdvAD, a transfer-based physical attack against object detection in autonomous driving. Instead of targeting a specific detector, AdvAD optimizes adversarial patches over multiple detection models in a unified framework, encouraging the learned perturbations to capture shared vulnerabilities across architectures. The optimization process adaptively balances model contributions and enforces robustness to physical variations. It further employs data augmentation and geometric transformations to maintain patch effectiveness under diverse physical conditions. Experiments in both digital and real-world settings show that AdvAD consistently outperforms state-of-the-art (SOTA) attacks in performance and transferability.", "date": "2026-04-25", "categories": []}
{"id": "2604.23483", "title": "Agentic Adversarial Rewriting Exposes Architectural Vulnerabilities in Black-Box NLP Pipelines", "abstract": "Multi-component natural language processing (NLP) pipelines are increasingly deployed for high-stakes decisions, yet no existing adversarial method can test their robustness under realistic conditions: binary-only feedback, no gradient access, and strict query budgets. We formalize this strict black-box threat model and propose a two-agent evasion framework operating in a semantic perturbation space. An Attacker Agent generates meaning-preserving rewrites while a Prompt Optimization Agent refines the attack strategy using only binary decision feedback within a 10-query budget. Evaluated against four evidence-based misinformation detection pipelines, the framework achieves evasion rates of 19.95 to 40.34% on modern large language model (LLM) based systems, compared to at most 3.90% for token-level perturbation baselines that rely on surrogate models because they cannot operate under our threat model. A legacy system relying on static lexical retrieval exhibits near-total vulnerability 97.02%, establishing a lower bound that exposes how architectural choices govern the attack surface. Evasion effectiveness is associated with three architectural properties: evidence retrieval mechanism, retrieval-inference coupling, and baseline classification accuracy. The iterative prompt optimization yields the largest marginal gains against the most robust targets, confirming that adaptive strategy discovery is essential when evasion is non-trivial. Analysis of successful rewrites reveals four exploitation patterns, each targeting failures at distinct pipeline stages. A pattern-informed defense reduces the evasion rate by up to 65.18%.", "date": "2026-04-26", "categories": []}
{"id": "2604.24082", "title": "Jailbreaking Frontier Foundation Models Through Intention Deception", "abstract": "Large (vision-)language models exhibit remarkable capability but remain highly susceptible to jailbreaking. Existing safety training approaches aim to have the model learn a refusal boundary between safe and unsafe, based on the user's intent. It has been found that this binary training regime often leads to brittleness, since the user intent cannot reliably be evaluated, especially if the attacker obfuscates their intent, and also makes the system seem unhelpful. In response, frontier models, such as GPT-5, have shifted from refusal-based safeguards to safe completion, that aims to maximize helpfulness while obeying safety constraints. However, safe completion could be exploited when a user pretends their intention is benign. Specifically, this intent inversion would be effective in multi-turn conversation, where the attacker has multiple opportunities to reinforce their deceptively benign intent. In this work, we introduce a novel multi-turn jailbreaking method that exploits this vulnerability. Our approach gradually builds conversational trust by simulating benign-seeming intentions and by exploiting the consistency property of the model, ultimately guiding the target model toward harmful, detailed outputs. Most crucially, our approach also uncovered an additional class of model vulnerability that we call para-jailbreaking that has been unnoticed up to now. Para-jailbreaking describes the situation where the model may not reveal harmful direct reply to the attack query, however the information that it reveals is nevertheless harmful. Our contributions are threefold. First, it achieves high success rates against frontier models including GPT-5-thinking and Claude-Sonnet-4.5. Second, our approach revealed and addressed para-jailbreaking harmful output. Third, experiments on multimodal VLM models showed that our approach outperformed state-of-the-art models.", "date": "2026-04-27", "categories": []}
{"id": "2604.24162", "title": "Defusing the Trigger: Plug-and-Play Defense for Backdoored LLMs via Tail-Risk Intrinsic Geometric Smoothing", "abstract": "Defending against backdoor attacks in large language models remains a critical practical challenge. Existing defenses mitigate these threats but typically incur high preparation costs and degrade utility via offline purification, or introduce severe latency via complex online interventions. To overcome this dichotomy, we present Tail-risk Intrinsic Geometric Smoothing (TIGS), a plug-and-play inference-time defense requiring no parameter updates, external clean data, or auxiliary generation. TIGS leverages the observation that successful backdoor triggers consistently induce localized attention collapse within the semantic content region. Operating entirely within the native forward pass, TIGS first performs content-aware tail-risk screening to identify suspicious attention heads and rows using sample-internal signals. It then applies intrinsic geometric smoothing: a weak content-domain correction preserves semantic anchoring, while a stronger full-row contraction disrupts trigger-dominant routing. Finally, a controlled full-row write-back reconstructs the attention matrix to ensure inference stability. Extensive evaluations demonstrate that TIGS substantially suppresses attack success rates while strictly preserving clean reasoning and open-ended semantic consistency. Crucially, this favorable security-utility-latency equilibrium persists across diverse architectures, including dense, reasoning-oriented, and sparse mixture-of-experts models. By structurally disrupting adversarial routing with marginal latency overhead, TIGS establishes a highly practical, deployment-ready defense standard for state-of-the-art LLMs.", "date": "2026-04-27", "categories": []}
{"id": "2604.24332", "title": "Mitigating Error Amplification in Fast Adversarial Training", "abstract": "Fast Adversarial Training (FAT) has proven effective in enhancing model robustness by encouraging networks to learn perturbation-invariant representations. However, FAT often suffers from catastrophic overfitting (CO), where the model overfits to the training attack and fails to generalize to unseen ones. Moreover, robustness oriented optimization typically leads to notable performance degradation on clean inputs, and such degradation becomes increasingly severe as the perturbation budget grows. In this work, we conduct a comprehensive analysis of how guidance strength affects model performance by modulating perturbation and supervision levels across distinct confidence groups. The findings reveal that low confidence samples are the primary contributors to CO and the robustness accuracy trade off. Building on this insight, we propose a Distribution-aware Dynamic Guidance (DDG) strategy that dynamically adjusts both the perturbation budget and supervision signal. Specifically, DDG scales the perturbation magnitude according to the sample confidence at the ground truth class, thereby guiding samples toward consistent decision boundaries while mitigating the influence of learning spurious correlations. Simultaneously, it dynamically adjusts the supervision signal based on the prediction state of each sample, preventing overemphasis on incorrect signals. To alleviate potential gradient instability arising from dynamic guidance, we further design a weighted regularization constraint. Extensive experiments on standard benchmarks demonstrate that DDG effectively alleviates both CO and the robustness accuracy trade off.", "date": "2026-04-27", "categories": []}
{"id": "2604.24350", "title": "Unveiling the Backdoor Mechanism Hidden Behind Catastrophic Overfitting in Fast Adversarial Training", "abstract": "Fast Adversarial Training (FAT) has attracted significant attention due to its efficiency in enhancing neural network robustness against adversarial attacks. However, FAT is prone to catastrophic overfitting (CO), wherein models overfit to the specific attack used during training and fail to generalize to others. While existing methods introduce diverse hypotheses and propose various strategies to mitigate CO, a systematic and intuitive explanation of CO remains absent. In this work, we innovatively interpret CO through the lens of backdoor. Through validations on pathway division, diverse feature predictions, and universal class distinguishable triggers in CO, we conceptualize CO as a weak trigger variant of unlearnable tasks, unifying CO, backdoor attacks, and unlearnable tasks under a common theoretical framework. Guided by this, we leverage several backdoor inspired strategies to mitigate CO: (i) Recalibrate CO affected model parameters using vanilla fine tuning, linear probing, or reinitialization-based techniques; (ii) Introduce a weight outlier suppression constraint to regulate abnormal deviations in model weights. Extensive experiments support our interpretation of CO and show the efficacy of the proposed mitigation strategies.", "date": "2026-04-27", "categories": []}
{"id": "2502.03698", "title": "How Vulnerable Is My Learned Policy? Universal Adversarial Perturbation Attacks On Modern Behavior Cloning Policies", "abstract": "Learning from demonstrations is a popular approach to train AI models; however, their vulnerability to adversarial attacks remains underexplored. We present the first systematic study of adversarial attacks, across a range of both classic and recently proposed imitation learning algorithms, including Vanilla Behavior Cloning (Vanilla BC), LSTM-GMM, Implicit Behavior Cloning (IBC), Diffusion Policy (DP), and Vector-Quantized Behavior Transformer (VQ-BET). We study the vulnerability of these methods to both white-box, grey-box and black-box adversarial perturbations. Our experiments reveal that most existing methods are highly vulnerable to these attacks, including black-box transfer attacks that transfer across algorithms. To the best of our knowledge, we are the first to study and compare the vulnerabilities of different popular imitation learning algorithms to both white-box and black-box attacks. Our findings highlight the vulnerabilities of modern imitation learning algorithms, paving the way for future work in addressing such limitations. Videos and code are available at https://sites.google.com/view/uap-attacks-on-bc.", "date": "2025-02-06", "categories": []}
{"id": "2602.21977", "title": "When LoRA Betrays: Backdooring Text-to-Image Models by Masquerading as Benign Adapters", "abstract": "Low-Rank Adaptation (LoRA) has emerged as a leading technique for efficiently fine-tuning text-to-image diffusion models, and its widespread adoption on open-source platforms has fostered a vibrant culture of model sharing and customization. However, the same modular and plug-and-play flexibility that makes LoRA appealing also introduces a broader attack surface. To highlight this risk, we propose Masquerade-LoRA (MasqLoRA), the first systematic attack framework that leverages an independent LoRA module as the attack vehicle to stealthily inject malicious behavior into text-to-image diffusion models. MasqLoRA operates by freezing the base model parameters and updating only the low-rank adapter weights using a small number of \"trigger word-target image\" pairs. This enables the attacker to train a standalone backdoor LoRA module that embeds a hidden cross-modal mapping: when the module is loaded and a specific textual trigger is provided, the model produces a predefined visual output; otherwise, it behaves indistinguishably from the benign model, ensuring the stealthiness of the attack. Experimental results demonstrate that MasqLoRA can be trained with minimal resource overhead and achieves a high attack success rate of 99.8%. MasqLoRA reveals a severe and unique threat in the AI supply chain, underscoring the urgent need for dedicated defense mechanisms for the LoRA-centric sharing ecosystem.", "date": "2026-02-25", "categories": []}
{"id": "2604.22552", "title": "Transferable Physical-World Adversarial Patches Against Pedestrian Detection Models", "abstract": "Physical adversarial patch attacks critically threaten pedestrian detection, causing surveillance and autonomous driving systems to miss pedestrians and creating severe safety risks. Despite their effectiveness in controlled settings, existing physical attacks face two major limitations in practice: they lack systematic disruption of the multi-stage decision pipeline, enabling residual modules to offset perturbations, and they fail to model complex physical variations, leading to poor robustness. To overcome these limitations, we propose a novel pedestrian adversarial patch generation method that combines multi-stage collaborative attacks with robustness enhancement under physical diversity, called TriPatch. Specifically, we design a triplet loss consisting of detection confidence suppression, bounding-box offset amplification, and non-maximum suppression (NMS) disruption, which jointly act across different stages of the detection pipeline. In addition, we introduce an appearance consistency loss to constrain the color distribution of the patch, thereby improving its adaptability under diverse imaging conditions, and incorporate data augmentation to further enhance robustness against complex physical perturbations. Extensive experiments demonstrate that TriPatch achieves a higher attack success rate across multiple detector models compared to existing approaches.", "date": "2026-04-24", "categories": []}
{"id": "2604.22569", "title": "Adversarial Co-Evolution of Malware and Detection Models: A Bilevel Optimization Perspective", "abstract": "Machine learning-based malware detectors are increasingly vulnerable to adversarial examples. Traditional defenses, such as one-shot adversarial training, often fail against adaptive attackers who use reinforcement learning to bypass detection. This paper proposes a robust defense framework based on bilevel optimization, explicitly modeling the strategic interaction between a defender and an attacker as an adversarial co-evolutionary process. We evaluate our approach using the MAB-malware framework against three distinct malware families: Mokes, Strab, and DCRat. Our experimental results demonstrate that while standard classifiers and basic adversarial retraining often remain vulnerable, showing evasion rates as high as 90 %, the proposed bilevel optimization approach consistently achieves near-total immunity, reducing evasion rates to 0 - 1.89 %. Furthermore, the iterative framework significantly increases the attacker's query complexity, raising the average cost of successful evasion by up to two orders of magnitude. These findings suggest that modeling the iterative cycle of attack and defense through bilevel optimization is essential for developing resilient malware detection systems capable of withstanding evolving adversarial threats.", "date": "2026-04-24", "categories": []}
{"id": "2604.22639", "title": "Adversarial Malware Generation in Linux ELF Binaries via Semantic-Preserving Transformations", "abstract": "Malware development and detection have undergone significant changes in recent years as modern concepts, such as machine learning, have been used for both adversarial attacks and defense. Despite intensive research on Windows Portable Executable (PE) files, there is minimal work on Linux Executable and Linkable Format (ELF). In this work, we summarize the academic papers submitted in this field and develop a new adversarial malware generator for the ELF format. Using a variety of metrics, we thoroughly evaluated our generator and achieved an Evasion Rate of 67.74 % while changing the confidence of the malware detector by -0.50 in the mean case for the dataset used. In our approach, we chose MalConv as the target classifier. Using this classifier, we found that the most successful modifications used strings typical of benign files as a data source. We conducted a variety of experiments and concluded that the target classifier appears sensitive to strings at any location within the executable file.", "date": "2026-04-24", "categories": []}
{"id": "2510.20792", "title": "BadGraph: A Backdoor Attack Against Latent Diffusion Model for Text-Guided Graph Generation", "abstract": "The rapid progress of graph generation has raised new security concerns, particularly regarding backdoor vulnerabilities. Though prior work has explored backdoor attacks against diffusion models for image or unconditional graph generation, those against conditional graph generation models, especially text-guided graph generation models, remain largely unexamined. This paper proposes BadGraph, a backdoor attack method against latent diffusion models for text-guided graph generation. BadGraph leverages textual triggers to poison training data, covertly implanting backdoors that induce attacker-specified subgraphs during inference when triggers appear, while preserving normal performance on clean inputs. Extensive experiments on four benchmark datasets (PubChem, ChEBI-20, PCDes, MoMu) demonstrate the effectiveness and stealth of the attack: a poisoning rate of less than 10% can achieve a 50% attack success rate, while 24% suffices for over an 80% success rate, with negligible performance degradation on benign samples. Ablation studies further reveal that the backdoor is implanted during VAE and diffusion training rather than pretraining. These findings reveal the security vulnerabilities in latent diffusion models for text-guided graph generation, highlight the serious risks in applications such as drug discovery, and underscore the need for robust defenses against the backdoor attack in such diffusion models.", "date": "2025-10-23", "categories": []}
{"id": "2602.16729", "title": "Intent Laundering: AI Safety Datasets Are Not What They Seem", "abstract": "We systematically evaluate the quality of widely used adversarial safety datasets from two perspectives: in isolation and in practice. In isolation, we examine how well these datasets reflect real-world adversarial attacks based on three defining properties: being driven by ulterior intent, well-crafted, and out-of-distribution. We find that these datasets overrely on \"triggering cues\": words or phrases with overt negative/sensitive connotations that are intended to trigger safety mechanisms explicitly, which is unrealistic compared to real-world attacks. In practice, we evaluate whether these datasets genuinely measure safety risks or merely provoke refusals through triggering cues. To explore this, we introduce \"intent laundering\": a procedure that abstracts away triggering cues from adversarial attacks (data points) while strictly preserving their malicious intent and all relevant details. Our results show that current adversarial safety datasets fail to faithfully represent real-world adversarial behavior due to their overreliance on triggering cues. Once these cues are removed, all previously evaluated \"reasonably safe\" models become unsafe, including Gemini 3 Pro and Claude Sonnet 3.7/4. Moreover, when intent laundering is adapted as a jailbreaking technique, it consistently achieves high attack success rates, ranging from 90.00% to 100.00%, under fully black-box access. Overall, our findings expose a significant disconnect between how existing datasets evaluate model safety and how real-world adversaries behave.", "date": "2026-02-17", "categories": []}
{"id": "2604.20994", "title": "Breaking MCP with Function Hijacking Attacks: Novel Threats for Function Calling and Agentic Models", "abstract": "The growth of agentic AI has drawn significant attention to function calling Large Language Models (LLMs), which are designed to extend the capabilities of AI-powered system by invoking external functions. Injection and jailbreaking attacks have been extensively explored to showcase the vulnerabilities of LLMs to user prompt manipulation. The expanded capabilities of agentic models introduce further vulnerabilities via their function calling interface. Recent work in LLM security showed that function calling can be abused, leading to data tampering and theft, causing disruptive behavior such as endless loops, or causing LLMs to produce harmful content in the style of jailbreaking attacks. This paper introduces a novel function hijacking attack (FHA) that manipulates the tool selection process of agentic models to force the invocation of a specific, attacker-chosen function. While existing attacks focus on semantic preference of the model for function-calling tasks, we show that FHA is largely agnostic to the context semantics and robust to the function sets, making it applicable across diverse domains. We further demonstrate that FHA can be trained to produce universal adversarial functions, enabling a single attacked function to hijack tool selection across multiple queries and payload configurations. We conducted experiments on 5 different models, including instructed and reasoning variants, reaching 70% to 100% ASR over the established BFCL dataset. Our findings further demonstrate the need for strong guardrails and security modules for agentic systems.", "date": "2026-04-22", "categories": []}
{"id": "2604.21310", "title": "Adversarial Evasion in Non-Stationary Malware Detection: Minimizing Drift Signals through Similarity-Constrained Perturbations", "abstract": "Deep learning has emerged as a powerful approach for malware detection, demonstrating impressive accuracy across various data representations. However, these models face critical limitations in real-world, non-stationary environments where both malware characteristics and detection systems continuously evolve. Our research investigates a fundamental security question: Can an attacker generate adversarial malware samples that simultaneously evade classification and remain inconspicuous to drift monitoring mechanisms? We propose a novel approach that generates targeted adversarial examples in the classifier's standardized feature space, augmented with sophisticated similarity regularizers. By carefully constraining perturbations to maintain distributional similarity with clean malware, we create an optimization objective that balances targeted misclassification with drift signal minimization. We quantify the effectiveness of this approach by comprehensively comparing classifier output probabilities using multiple drift metrics. Our experiments demonstrate that similarity constraints can reduce output drift signals, with $\\ell_2$ regularization showing the most promising results. We observe that perturbation budget significantly influences the evasion-detectability trade-off, with increased budget leading to higher attack success rates and more substantial drift indicators.", "date": "2026-04-23", "categories": []}
{"id": "2604.21416", "title": "CSC: Turning the Adversary's Poison against Itself", "abstract": "Poisoning-based backdoor attacks pose significant threats to deep neural networks by embedding triggers in training data, causing models to misclassify triggered inputs as adversary-specified labels while maintaining performance on clean data. Existing poison restraint-based defenses often suffer from inadequate detection against specific attack variants and compromise model utility through unlearning methods that lead to accuracy degradation. This paper conducts a comprehensive analysis of backdoor attack dynamics during model training, revealing that poisoned samples form isolated clusters in latent space early on, with triggers acting as dominant features distinct from benign ones. Leveraging these insights, we propose Cluster Segregation Concealment (CSC), a novel poison suppression defense. CSC first trains a deep neural network via standard supervised learning while segregating poisoned samples through feature extraction from early epochs, DBSCAN clustering, and identification of anomalous clusters based on class diversity and density metrics. In the concealment stage, identified poisoned samples are relabeled to a virtual class, and the model's classifier is fine-tuned using cross-entropy loss to replace the backdoor association with a benign virtual linkage, preserving overall accuracy. CSC was evaluated on four benchmark datasets against twelve poisoning-based attacks, CSC outperforms nine state-of-the-art defenses by reducing average attack success rates to near zero with minimal clean accuracy loss. Contributions include robust backdoor patterns identification, an effective concealment mechanism, and superior empirical validation, advancing trustworthy artificial intelligence.", "date": "2026-04-23", "categories": []}
{"id": "2604.21860", "title": "Transient Turn Injection: Exposing Stateless Multi-Turn Vulnerabilities in Large Language Models", "abstract": "Large language models (LLMs) are increasingly integrated into sensitive workflows, raising the stakes for adversarial robustness and safety. This paper introduces Transient Turn Injection(TTI), a new multi-turn attack technique that systematically exploits stateless moderation by distributing adversarial intent across isolated interactions. TTI leverages automated attacker agents powered by large language models to iteratively test and evade policy enforcement in both commercial and open-source LLMs, marking a departure from conventional jailbreak approaches that typically depend on maintaining persistent conversational context. Our extensive evaluation across state-of-the-art models-including those from OpenAI, Anthropic, Google Gemini, Meta, and prominent open-source alternatives-uncovers significant variations in resilience to TTI attacks, with only select architectures exhibiting substantial inherent robustness. Our automated blackbox evaluation framework also uncovers previously unknown model specific vulnerabilities and attack surface patterns, especially within medical and high stakes domains. We further compare TTI against established adversarial prompting methods and detail practical mitigation strategies, such as session level context aggregation and deep alignment approaches. Our study underscores the urgent need for holistic, context aware defenses and continuous adversarial testing to future proof LLM deployments against evolving multi-turn threats.", "date": "2026-04-23", "categories": []}
{"id": "2505.16313", "title": "Accelerating Targeted Hard-Label Adversarial Attacks in Low-Query Black-Box Settings", "abstract": "Deep neural networks for image classification remain vulnerable to adversarial examples -- small, imperceptible perturbations that induce misclassifications. In black-box settings, where only the final prediction is accessible, crafting targeted attacks that aim to misclassify into a specific target class is particularly challenging due to narrow decision regions. Current state-of-the-art methods often exploit the geometric properties of the decision boundary separating a source image and a target image rather than incorporating information from the images themselves. In contrast, we propose Targeted Edge-informed Attack (TEA), a novel attack that utilizes edge information from the target image to carefully perturb it, thereby producing an adversarial image that is closer to the source image while still achieving the desired target classification. Our approach consistently outperforms current state-of-the-art methods across different models in low query settings (nearly 70% fewer queries are used), a scenario especially relevant in real-world applications with limited queries and black-box access. Furthermore, by efficiently generating a suitable adversarial example, TEA provides an improved target initialization for established geometry-based attacks.", "date": "2025-05-22", "categories": []}
{"id": "2604.20047", "title": "PASTA: A Patch-Agnostic Twofold-Stealthy Backdoor Attack on Vision Transformers", "abstract": "Vision Transformers (ViTs) have achieved remarkable success across vision tasks, yet recent studies show they remain vulnerable to backdoor attacks. Existing patch-wise attacks typically assume a single fixed trigger location during inference to maximize trigger attention. However, they overlook the self-attention mechanism in ViTs, which captures long-range dependencies across patches. In this work, we observe that a patch-wise trigger can achieve high attack effectiveness when activating backdoors across neighboring patches, a phenomenon we term the Trigger Radiating Effect (TRE). We further find that inter-patch trigger insertion during training can synergistically enhance TRE compared to single-patch insertion. Prior ViT-specific attacks that maximize trigger attention often sacrifice visual and attention stealthiness, making them detectable.\n  Based on these insights, we propose PASTA, a twofold stealthy patch-wise backdoor attack in both pixel and attention domains. PASTA enables backdoor activation when the trigger is placed at arbitrary patches during inference. To achieve this, we introduce a multi-location trigger insertion strategy to enhance TRE. However, preserving stealthiness while maintaining strong TRE is challenging, as TRE is weakened under stealthy constraints. We therefore formulate a bi-level optimization problem and propose an adaptive backdoor learning framework, where the model and trigger iteratively adapt to each other to avoid local optima. Extensive experiments show that PASTA achieves 99.13% attack success rate across arbitrary patches on average, while significantly improving visual and attention stealthiness (144.43x and 18.68x) and robustness (2.79x) against state-of-the-art ViT defenses across four datasets, outperforming CNN- and ViT-based baselines.", "date": "2026-04-21", "categories": []}
{"id": "2604.20495", "title": "Towards Certified Malware Detection: Provable Guarantees Against Evasion Attacks", "abstract": "Machine learning-based static malware detectors remain vulnerable to adversarial evasion techniques, such as metamorphic engine mutations. To address this vulnerability, we propose a certifiably robust malware detection framework based on randomized smoothing through feature ablation and targeted noise injection. During evaluation, our system analyzes an executable by generating multiple ablated variants, classifies them by using a smoothed classifier, and identifies the final label based on the majority vote. By analyzing the top-class voting distribution and the Wilson score interval, we derive a formal certificate that guarantees robustness within a specific radius against feature-space perturbations. We evaluate our approach by comparing the performance of the base classifier and the smoothed classifier on both clean executables and ablated variants generated using PyMetaEngine. Our results demonstrate that the proposed smoothed classifier successfully provides certifiable robustness against metamorphic evasion attacks without requiring modifications to the underlying machine learning architecture.", "date": "2026-04-22", "categories": []}
{"id": "2604.20704", "title": "Auto-ART: Structured Literature Synthesis and Automated Adversarial Robustness Testing", "abstract": "Adversarial robustness evaluation underpins every claim of trustworthy ML deployment, yet the field suffers from fragmented protocols and undetected gradient masking. We make two contributions. (1) Structured synthesis. We analyze nine peer-reviewed corpus sources (2020--2026) through seven complementary protocols, producing the first end-to-end structured analysis of the field's consensus and unresolved challenges. (2) Auto-ART framework. We introduce Auto-ART, an open-source framework that operationalizes identified gaps: 50+ attacks, 28 defense modules, the Robustness Diagnostic Index (RDI), and gradient-masking detection. It supports multi-norm evaluation (l1/l2/linf/semantic/spatial) and compliance mapping to NIST AI RMF, OWASP LLM Top 10, and the EU AI Act. Empirical validation on RobustBench demonstrates that Auto-ART's pre-screening identifies gradient masking in 92% of flagged cases, and RDI rankings correlate highly with full AutoAttack. Multi-norm evaluation exposes a 23.5 pp gap between average and worst-case robustness on state-of-the-art models. No prior work combines such structured meta-scientific analysis with an executable evaluation framework bridging literature gaps into engineering.", "date": "2026-04-22", "categories": []}
{"id": "2511.21893", "title": "Breaking the Illusion: Consensus-Based Generative Mitigation of Adversarial Illusions in Multi-Modal Embeddings", "abstract": "Multi-modal foundation models align images, text, and other modalities in a shared embedding space but remain vulnerable to adversarial illusions [35], where imperceptible perturbations disrupt cross-modal alignment and mislead downstream tasks. To counteract the effects of adversarial illusions, we propose a task-agnostic mitigation mechanism that purifies the attacker's perturbed input using generative models, e.g., Variational Autoencoders (VAEs), to restore natural alignment. To further enhance the defense mechanism, we adopt a generative sampling strategy combined with a consensus-based aggregation scheme over the outcomes of the generated samples. Our experiments on ImageBind, a state-of-the-art multi-modal encoder, show that our approach substantially reduces the illusion attack success rates to near-zero and improves cross-modal alignment in unperturbed and perturbed input settings, providing an effective and task-agnostic defense against adversarial illusions. The code is available at https://github.com/fatemehakb/adversarial-illusions-mitigation.", "date": "2025-11-26", "categories": []}
{"id": "2601.07056", "title": "Adversarial Attacks on Medical Hyperspectral Imaging Exploiting Spectral-Spatial Dependencies and Multiscale Features", "abstract": "Medical hyperspectral imaging (MHSI) has shown strong potential for disease diagnosis by capturing spectral-spatial information of tissues. While deep learning has substantially improved MHSI classification accuracy, its robustness remains limited due to the well-known trade-off between accuracy and robustness in Deep Neural Networks (DNNs). This issue is particularly critical in MHSI, where reliable prediction depends on local tissue relationships and multiscale spectral-spatial structures. A practical way to improve robustness is to identify the most unstable adversarial examples and incorporate them into adversarial training. However, existing attack methods do not sufficiently exploit these MHSI-specific properties, leading to suboptimal attack effectiveness and limited value for robustness enhancement. To address this gap, we propose a structured adversarial attack framework for MHSI that progressively models its local spectral-spatial dependencies and multiscale hierarchical representations. The proposed method generates anatomically consistent perturbations by modeling neighborhood dependencies and hierarchical spectral-spatial features. Experiments on the brain and choledoch datasets show that our method more effectively degrades lesion-related classification performance in critical tumor regions than existing baselines while maintaining low perturbation magnitude. These results reveal a clinically relevant robustness weakness in current MHSI models and provide stronger adversarial samples for developing targeted defense strategies.", "date": "2026-01-11", "categories": []}
{"id": "2604.03753", "title": "Spatiotemporal-Aware Bit-Flip Injection on DNN-based Advanced Driver Assistance Systems (extended version)", "abstract": "Modern advanced driver assistance systems (ADAS) rely on deep neural networks (DNNs) for perception and planning. Since DNNs' parameters reside in DRAM during inference, bit flips caused by cosmic radiation or low-voltage operation may corrupt DNN computations, distort driving decisions, and lead to real-world incidents. This paper presents a SpatioTemporal-Aware Fault Injection (STAFI) framework to locate critical fault sites in DNNs for ADAS efficiently. Spatially, we propose a Progressive Metric-guided Bit Search (PMBS) that efficiently identifies critical network weight bits whose corruption causes the largest deviations in driving behavior (e.g., unintended acceleration or steering). Furthermore, we develop a Critical Fault Time Identification (CFTI) mechanism that determines when to trigger these faults, taking into account the context of real-time systems and environmental states, to maximize the safety impact. Experiments on DNNs for a production ADAS demonstrate that STAFI uncovers 29.56x more hazard-inducing critical faults than the strongest baseline.", "date": "2026-04-04", "categories": []}
{"id": "2604.18663", "title": "Beyond Explicit Refusals: Soft-Failure Attacks on Retrieval-Augmented Generation", "abstract": "Existing jamming attacks on Retrieval-Augmented Generation (RAG) systems typically induce explicit refusals or denial-of-service behaviors, which are conspicuous and easy to detect. In this work, we formalize a subtler availability threat, termed soft failure, which degrades system utility by inducing fluent and coherent yet non-informative responses rather than overt failures. We propose Deceptive Evolutionary Jamming Attack (DEJA), an automated black-box attack framework that generates adversarial documents to trigger such soft failures by exploiting safety-aligned behaviors of large language models. DEJA employs an evolutionary optimization process guided by a fine-grained Answer Utility Score (AUS), computed via an LLM-based evaluator, to systematically degrade the certainty of answers while maintaining high retrieval success. Extensive experiments across multiple RAG configurations and benchmark datasets show that DEJA consistently drives responses toward low-utility soft failures, achieving SASR above 79\\% while keeping hard-failure rates below 15\\%, significantly outperforming prior attacks. The resulting adversarial documents exhibit high stealth, evading perplexity-based detection and resisting query paraphrasing, and transfer across model families to proprietary systems without retargeting.", "date": "2026-04-20", "categories": []}
{"id": "2604.18867", "title": "Hierarchically Robust Zero-shot Vision-language Models", "abstract": "Vision-Language Models (VLMs) can perform zero-shot classification but are susceptible to adversarial attacks. While robust fine-tuning improves their robustness, existing approaches align fixed text embeddings with an image embedding, sacrificing natural performance and robustness. A robustness degradation also occurs when a model faces adversarial attacks targeting superclasses (parent classes, e.g., mammal) in addition to their base (leaf) classes (e.g., cat). Thus, to enhance adversarial robustness and leverage the inherent hierarchical properties of class space, we propose a novel adversarial fine-tuning framework based on hierarchical embeddings and several levels of adversarially robust alignment of image-text modalities. Additional mechanisms place visual embeddings at the desired depth of hierarchy, and we provide a theoretical connection between the depth of embedding in the hierarchy and the maximum viable margin size. Our model naturally realizes several margin sizes, boosting generalization of adversaries for robustification. As various trees with different parent labels can share the same leaf labels, we also consider aligning over multiple trees to boost semantic variety. Experiments across several datasets are performed.", "date": "2026-04-20", "categories": []}
{"id": "2604.18874", "title": "How Adversarial Environments Mislead Agentic AI?", "abstract": "Tool-integrated agents are deployed on the premise that external tools ground their outputs in reality. Yet this very reliance creates a critical attack surface. Current evaluations benchmark capability in benign settings, asking \"can the agent use tools correctly\" but never \"what if the tools lie\". We identify this Trust Gap: agents are evaluated for performance, not for skepticism. We formalize this vulnerability as Adversarial Environmental Injection (AEI), a threat model where adversaries compromise tool outputs to deceive agents. AEI constitutes environmental deception: constructing a \"fake world\" of poisoned search results and fabricated reference networks around unsuspecting agents. We operationalize this via POTEMKIN, a Model Context Protocol (MCP)-compatible harness for plug-and-play robustness testing. We identify two orthogonal attack surfaces: The Illusion (breadth attacks) poison retrieval to induce epistemic drift toward false beliefs, while The Maze (depth attacks) exploit structural traps to cause policy collapse into infinite loops. Across 11,000+ runs on five frontier agents, we find a stark robustness gap: resistance to one attack often increases vulnerability to the other, demonstrating that epistemic and navigational robustness are distinct capabilities.", "date": "2026-04-20", "categories": []}
{"id": "2604.19724", "title": "Benign Overfitting in Adversarial Training for Vision Transformers", "abstract": "Despite the remarkable success of Vision Transformers (ViTs) across a wide range of vision tasks, recent studies have revealed that they remain vulnerable to adversarial examples, much like Convolutional Neural Networks (CNNs). A common empirical defense strategy is adversarial training, yet the theoretical underpinnings of its robustness in ViTs remain largely unexplored. In this work, we present the first theoretical analysis of adversarial training under simplified ViT architectures. We show that, when trained under a signal-to-noise ratio that satisfies a certain condition and within a moderate perturbation budget, adversarial training enables ViTs to achieve nearly zero robust training loss and robust generalization error under certain regimes. Remarkably, this leads to strong generalization even in the presence of overfitting, a phenomenon known as \\emph{benign overfitting}, previously only observed in CNNs (with adversarial training). Experiments on both synthetic and real-world datasets further validate our theoretical findings.", "date": "2026-04-21", "categories": []}
{"id": "2503.14281", "title": "XOXO: Stealthy Cross-Origin Context Poisoning Attacks against AI Coding Assistants", "abstract": "AI coding assistants are widely used for tasks like code generation. These tools now require large and complex contexts, automatically sourced from various origins$\\unicode{x2014}$across files, projects, and contributors$\\unicode{x2014}$forming part of the prompt fed to underlying LLMs. This automatic context-gathering introduces new vulnerabilities, allowing attackers to subtly poison input to compromise the assistant's outputs, potentially generating vulnerable code or introducing critical errors. We propose a novel attack, Cross-Origin Context Poisoning (XOXO), that is challenging to detect as it relies on adversarial code modifications that are semantically equivalent. Traditional program analysis techniques struggle to identify these perturbations since the semantics of the code remains correct, making it appear legitimate. This allows attackers to manipulate coding assistants into producing incorrect outputs, while shifting the blame to the victim developer. We introduce a novel, task-agnostic, black-box attack algorithm GCGS that systematically searches the transformation space using a Cayley Graph, achieving a 75.72% attack success rate on average across five tasks and eleven models, including GPT 4.1 and Claude 3.5 Sonnet v2 used by popular AI coding assistants. Furthermore, defenses like adversarial fine-tuning are ineffective against our attack, underscoring the need for new security measures in LLM-powered coding tools.", "date": "2025-03-18", "categories": []}
{"id": "2511.02356", "title": "ASTRA: An Automated Framework for Strategy Discovery, Retrieval, and Evolution for Jailbreaking LLMs", "abstract": "Despite extensive safety alignment, Large Language Models (LLMs) remain vulnerable to jailbreak attacks. However, existing methods generally lack the capability for continuous learning and self-evolution from interactions, limiting the diversity and adaptability of attack strategies. To address this, we propose ASTRA, an automated framework capable of autonomously discovering, retrieving, and evolving attack strategies. ASTRA operates on a closed-loop ``attack-evaluate-distill-reuse'' mechanism, which not only generates attack prompts but also automatically distills reusable strategies from every interaction. To systematically manage these strategies, we introduce a dynamic three-tier strategy library (Effective, Promising, and Ineffective) that categorizes strategies based on performance. This hierarchical memory mechanism enables the framework to enhance efficiency by leveraging successful patterns while optimizing the exploration space by avoiding known failures. Extensive experiments in a black-box setting demonstrate that ASTRA significantly outperforms existing baselines.", "date": "2025-11-04", "categories": []}
{"id": "2512.16055", "title": "Driving in Corner Case: A Real-World Adversarial Closed-Loop Evaluation Platform for End-to-End Autonomous Driving", "abstract": "Safety-critical corner cases, difficult to collect in the real world, are crucial for evaluating end-to-end autonomous driving. Adversarial interaction is an effective method to generate such safety-critical corner cases. While existing adversarial evaluation methods are built for models operating in simplified simulation environments, adversarial evaluation for real-world end-to-end autonomous driving has been little explored. To address this challenge, we propose a closed-loop evaluation platform for end-to-end autonomous driving, which can generate adversarial interactions in real-world scenes. In our platform, the real-world image generator cooperates with an adversarial traffic policy to evaluate various end-to-end models trained on real-world data. The generator, based on flow matching, efficiently and stably generates real-world images according to the traffic environment information. The efficient adversarial surrounding vehicle policy is designed to model challenging interactions and create corner cases that current autonomous driving systems struggle to handle. Experimental results demonstrate that the platform can generate realistic driving images efficiently. Through evaluating the end-to-end models such as UniAD and VAD, we demonstrate that based on the adversarial policy, our platform evaluates the performance degradation of the tested model in corner cases. This result indicates that this platform can effectively detect the model's potential issues, which will facilitate the safety and robustness of end-to-end autonomous driving.", "date": "2025-12-18", "categories": []}
{"id": "2601.04448", "title": "Merging Triggers, Breaking Backdoors: Defensive Poisoning for Instruction-Tuned Language Models", "abstract": "Large Language Models (LLMs) have greatly advanced Natural Language Processing (NLP), particularly through instruction tuning, which enables broad task generalization without additional fine-tuning. However, their reliance on large-scale datasets-often collected from human or web sources-makes them vulnerable to backdoor attacks, where adversaries poison a small subset of data to implant hidden behaviors. Despite this growing risk, defenses for instruction-tuned models remain underexplored. We propose MB-Defense (Merging & Breaking Defense Framework), a novel training pipeline that immunizes instruction-tuned LLMs against diverse backdoor threats. MB-Defense comprises two stages: (i) Defensive Poisoning, which merges attacker and defensive triggers into a unified backdoor representation, and (ii) Backdoor Neutralization, which breaks this representation through additional training to restore clean behavior. Extensive experiments across multiple LLMs show that MB-Defense substantially lowers attack success rates while preserving instruction-following ability. Our method offers a generalizable and data-efficient defense strategy, improving the robustness of instruction-tuned LLMs against unseen backdoor attacks.", "date": "2026-01-07", "categories": []}
{"id": "2601.23179", "title": "Universal Adversarial Attacks against Closed-Source MLLMs via Target-View Routed Meta Optimization", "abstract": "Targeted adversarial attacks on closed-source multimodal large language models (MLLMs) have been increasingly explored under black-box transfer, yet prior methods are predominantly sample-specific and offer limited reusability across inputs. We instead study a more stringent setting, Universal Targeted Transferable Adversarial Attacks (UTTAA), where a single perturbation must consistently steer arbitrary inputs toward a specified target across unknown commercial MLLMs. Naively adapting existing sample-wise attacks to this universal setting faces three core difficulties: (i) target supervision becomes high-variance due to target-crop randomness, (ii) token-wise matching is unreliable because universality suppresses image-specific cues that would otherwise anchor alignment, and (iii) few-source per-target adaptation is highly initialization-sensitive, which can degrade the attainable performance. In this work, we propose MCRMO-Attack, which stabilizes supervision via Multi-Crop Aggregation with an Attention-Guided Crop, improves token-level reliability through alignability-gated Token Routing, and meta-learns a cross-target perturbation prior that yields stronger per-target solutions. Across commercial MLLMs, we boost unseen-image attack success rate by +23.7\\% on GPT-4o and +19.9\\% on Gemini-2.0 over the strongest universal baseline.", "date": "2026-01-30", "categories": []}
{"id": "2604.16424", "title": "Safety, Security, and Cognitive Risks in State-Space Models: A Systematic Threat Analysis with Spectral, Stateful, and Capacity Attacks", "abstract": "State-Space Models (SSMs) -- structured SSMs (S4, S4D, DSS, S5), selective SSMs (Mamba, Mamba-2), and hybrid architectures (Jamba) -- are deployed in safety-critical long-context applications: genomic analysis, clinical time-series forecasting, and cybersecurity log processing. Their linear-time scaling is compelling, yet the security properties of their compressed-state recurrent architectures remain unstudied.\n  We present the first systematic treatment of SSM safety, security, and cognitive risks. Seven contributions: (1) Formal threat framework -- SSM Attack Surface (five layers), State Integrity Violation (StIV), Cross-Context Amplification Ratio $\\mathcal{X}_\\mathcal{S}$, and a Spectral Sensitivity Proposition grounded in the $H_\\infty$ norm. (2) Three novel attack classes: spectral adversarial attacks (transfer-function gain exploitation), delayed-trigger stateful backdoors (activate thousands of steps after injection), and state capacity saturation (entropy flooding forces silent forgetting). (3) 14 MITRE ATLAS technique extensions across the full tactic chain. (4) Six-profile attacker taxonomy with kill chains for genomics, clinical, and cybersecurity domains. (5) Four cognitive risk hypotheses grounded in state-compression mechanics. (6) Governance-aligned mitigations mapped to CREST, NIST AI 600-1, and EU AI Act. (7) Empirical evaluation: targeted genomic injection achieves $\\mathrm{StIV}=0.519$ vs. $0.086$ random ($6.0\\times$, $p<0.001$); PGD state injection achieves $156\\times$ output perturbation over random; SSD-structured extraction confirmed at $O(N^2)$ vs. $O(N^3)$ query complexity ($N\\times$ speedup). Validation on pretrained checkpoints is detailed in the Appendix.", "date": "2026-04-04", "categories": []}
{"id": "2604.16499", "title": "HQA-VLAttack: Towards High Quality Adversarial Attack on Vision-Language Pre-Trained Models", "abstract": "Black-box adversarial attack on vision-language pre-trained models is a practical and challenging task, as text and image perturbations need to be considered simultaneously, and only the predicted results are accessible. Research on this problem is in its infancy, and only a handful of methods are available. Nevertheless, existing methods either rely on a complex iterative cross-search strategy, which inevitably consumes numerous queries, or only consider reducing the similarity of positive image-text pairs but ignore that of negative ones, which will also be implicitly diminished, thus inevitably affecting the attack performance. To alleviate the above issues, we propose a simple yet effective framework to generate high-quality adversarial examples on vision-language pre-trained models, named HQA-VLAttack, which consists of text and image attack stages. For text perturbation generation, it leverages the counter-fitting word vector to generate the substitute word set, thus guaranteeing the semantic consistency between the substitute word and the original word. For image perturbation generation, it first initializes the image adversarial example via the layer-importance guided strategy, and then utilizes contrastive learning to optimize the image adversarial perturbation, which ensures that the similarity of positive image-text pairs is decreased while that of negative image-text pairs is increased. In this way, the optimized adversarial images and texts are more likely to retrieve negative examples, thereby enhancing the attack success rate. Experimental results on three benchmark datasets demonstrate that HQA-VLAttack significantly outperforms strong baselines in terms of attack success rate.", "date": "2026-04-14", "categories": []}
{"id": "2604.16515", "title": "Penny Wise, Pixel Foolish: Bypassing Price Constraints in Multimodal Agents via Visual Adversarial Perturbations", "abstract": "The rapid proliferation of Multimodal Large Language Models (MLLMs) has enabled mobile agents to execute high-stakes financial transactions, but their adversarial robustness remains underexplored. We identify Visual Dominance Hallucination (VDH), where imperceptible visual cues can override textual price evidence in screenshot-based, price-constrained settings and lead agents to irrational decisions. We propose PriceBlind, a stealthy white-box adversarial attack framework for controlled screenshot-based evaluation. PriceBlind exploits the modality gap in CLIP-based encoders via a Semantic-Decoupling Loss that aligns the image embedding with low-cost, value-associated anchors while preserving pixel-level fidelity. On E-ShopBench, PriceBlind achieves around 80% ASR in white-box evaluation; under a simplified single-turn coordinate-selection protocol, Ensemble-DI-FGSM transfers with roughly 35-41% ASR across GPT-4o, Gemini-1.5-Pro, and Claude-3.5-Sonnet. We also show that robust encoders and Verify-then-Act defenses reduce ASR substantially, though with some clean-accuracy trade-off.", "date": "2026-04-15", "categories": []}
{"id": "2604.16532", "title": "Beyond Attack Success Rate: A Multi-Metric Evaluation of Adversarial Transferability in Medical Imaging Models", "abstract": "While deep learning systems are becoming increasingly prevalent in medical image analysis, their vulnerabilities to adversarial perturbations raise serious concerns for clinical deployment. These vulnerability evaluations largely rely on Attack Success Rate (ASR), a binary metric that indicates solely whether an attack is successful. However, the ASR metric does not account for other factors, such as perturbation strength, perceptual image quality, and cross-architecture attack transferability, and therefore, the interpretation is incomplete. This gap requires consideration, as complex, large-scale deep learning systems, including Vision Transformers (ViTs), are increasingly challenging the dominance of Convolutional Neural Networks (CNNs). These architectures learn differently, and it is unclear whether a single metric, e.g., ASR, can effectively capture adversarial behavior. To address this, we perform a systematic empirical study on four medical image datasets: PathMNIST, DermaMNIST, RetinaMNIST, and CheXpert. We evaluate seven models (VGG-16, ResNet-50, DenseNet-121, Inception-v3, DeiT, Swin Transformer, and ViT-B/16) against seven attack methods at five perturbation budgets, measuring ASR, Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and $L_2$ perturbation magnitude. Our findings show a consistent pattern: perceptual and distortion metrics are strongly associated with one another and exhibit minimal correlation with ASR. This applies to both CNNs and ViTs. The results demonstrate that ASR alone is an inadequate indicator of adversarial robustness and transferability. Consequently, we argue that a thorough assessment of adversarial risk in medical AI necessitates multi-metric frameworks that encompass not only the attack efficacy but also its methodology and associated overheads.", "date": "2026-04-16", "categories": []}
{"id": "2604.16540", "title": "PoInit-of-View: Poisoning Initialization of Views Transfers Across Multiple 3D Reconstruction Systems", "abstract": "Poisoning input views of 3D reconstruction systems has been recently studied. However, we identify that existing studies simply backpropagate adversarial gradients through the 3D reconstruction pipeline as a whole, without uncovering the new vulnerability rooted in specific modules of the 3D reconstruction pipeline. In this paper, we argue that the structure-from-motion (SfM) initialization, as the geometric core of many widely used reconstruction systems, can be targeted to achieve transferable poisoning effects across diverse 3D reconstruction systems. To this end, we propose PoInit-of-View, which optimizes adversarial perturbations to intentionally introduce cross-view gradient inconsistencies at projections of corresponding 3D points. These inconsistencies disrupt keypoint detection and feature matching, thereby corrupting pose estimation and triangulation within SfM, eventually resulting in low-quality rendered views. We also provide a theoretical analysis that connects cross-view inconsistency to correspondence collapse. Experimental results demonstrate the effectiveness of our PoInit-of-View on diverse 3D reconstruction systems and datasets, surpassing the single-view baseline by 25.1% in PSNR and 16.5% in SSIM in black-box transfer settings, such as 3DGS to NeRF.", "date": "2026-04-17", "categories": []}
{"id": "2604.16966", "title": "Visual Inception: Compromising Long-term Planning in Agentic Recommenders via Multimodal Memory Poisoning", "abstract": "The evolution from static ranking models to Agentic Recommender Systems (Agentic RecSys) empowers AI agents to maintain long-term user profiles and autonomously plan service tasks. While this paradigm shift enhances personalization, it introduces a vulnerability: reliance on Long-term Memory (LTM). In this paper, we uncover a threat termed \"Visual Inception.\" Unlike traditional adversarial attacks that seek immediate misclassification, Visual Inception injects triggers into user-uploaded images (e.g., lifestyle photos) that act as \"sleeper agents\" within the system's memory. When retrieved during future planning, these poisoned memories hijack the agent's reasoning chain, steering it toward adversary-defined goals (e.g., promoting high-margin products) without prompt injection. To mitigate this, we propose CognitiveGuard, a dual-process defense framework inspired by human cognition. It consists of a System 1 Perceptual Sanitizer (diffusion-based purification) to cleanse sensory inputs and a System 2 Reasoning Verifier (counterfactual consistency checks) to detect anomalies in memory-driven planning. Extensive experiments on a mock e-commerce agent environment demonstrate that Visual Inception achieves about 85% Goal-Hit Rate (GHR), while CognitiveGuard reduces this risk to around 10% with configurable latency trade-offs (about 1.5s in lite mode to about 6.5s for full sequential verification), without quality degradation under our setup.", "date": "2026-04-18", "categories": []}
{"id": "2604.18487", "title": "Adversarial Humanities Benchmark: Results on Stylistic Robustness in Frontier Model Safety", "abstract": "The Adversarial Humanities Benchmark (AHB) evaluates whether model safety refusals survive a shift away from familiar harmful prompt forms. Starting from harmful tasks drawn from MLCommons AILuminate, the benchmark rewrites the same objectives through humanities-style transformations while preserving intent. This extends literature on Adversarial Poetry and Adversarial Tales from single jailbreak operators to a broader benchmark family of stylistic obfuscation and goal concealment. In the benchmark results reported here, the original attacks record 3.84% attack success rate (ASR), while transformed methods range from 36.8% to 65.0%, yielding 55.75% overall ASR across 31 frontier models. Under a European Union AI Act Code-of-Practice-inspired systemic-risk lens, Chemical, biological, radiological and nuclear (CBRN) is the highest bucket. Taken together, this lack of stylistic robustness suggests that current safety techniques suffer from weak generalization: deep understanding of 'non-maleficence' remains a central unresolved problem in frontier model safety.", "date": "2026-04-20", "categories": []}
{"id": "2604.18537", "title": "MetaCloak-JPEG: JPEG-Robust Adversarial Perturbation for Preventing Unauthorized DreamBooth-Based Deepfake Generation", "abstract": "The rapid progress of subject-driven text-to-image synthesis, and in particular DreamBooth, has enabled a consent-free deepfake pipeline: an adversary needs only 4-8 publicly available face images to fine-tune a personalized diffusion model and produce photorealistic harmful content. Current adversarial face-protection systems -- PhotoGuard, Anti-DreamBooth, and MetaCloak -- perturb user images to disrupt surrogate fine-tuning, but all share a structural blindness: none backpropagates gradients through the JPEG compression pipeline that every major social-media platform applies before adversary access. Because JPEG quantization relies on round(), whose derivative is zero almost everywhere, adversarial energy concentrates in high-frequency DCT bands that JPEG discards, eliminating 60-80% of the protective signal. We introduce MetaCloak-JPEG, which closes this gap by inserting a Differentiable JPEG (DiffJPEG) layer built on the Straight-Through Estimator (STE): the forward pass applies standard JPEG compression, while the backward pass replaces round() with the identity. DiffJPEG is embedded in a JPEG-aware EOT distribution (~70% of augmentations include DiffJPEG) and a curriculum quality-factor schedule (QF: 95 to 50) inside a bilevel meta-learning loop. Under an l-inf perturbation budget of eps=8/255, MetaCloak-JPEG attains 32.7 dB PSNR, a 91.3% JPEG survival rate, and outperforms PhotoGuard on all 9 evaluated JPEG quality factors (9/9 wins, mean denoising-loss gain +0.125) within a 4.1 GB training-memory budget.", "date": "2026-04-20", "categories": []}
{"id": "2604.18540", "title": "Duality for the Adversarial Total Variation", "abstract": "Adversarial training of binary classifiers can be reformulated as regularized risk minimization involving a nonlocal total variation. Building on this perspective, we establish a characterization of the subdifferential of this total variation using duality techniques. To achieve this, we derive a dual representation of the nonlocal total variation and a related integration of parts formula, involving a nonlocal gradient and divergence. We provide such duality statements both in the space of continuous functions vanishing at infinity on proper metric spaces and for the space of essentially bounded functions on Euclidean domains. Furthermore, under some additional conditions we provide characterizations of the subdifferential in these settings.", "date": "2026-04-20", "categories": []}
{"id": "2604.15370", "title": "TopFeaRe: Locating Critical State of Adversarial Resilience for Graphs Regarding Topology-Feature Entanglement", "abstract": "Graph adversarial attacks are usually produced from the two perspectives of topology/structure and node feature, both of them represent the paramount characteristics learned by today's deep learning models. Although some defense countermeasures are proposed at present, they fails to disclose the intrinsic reasons why these two aspects necessitate and how they are adequately fused to co-learn the graph representation. Towards this question, we in this paper propose an adversarial defense approach through locating the graph's critical state of adversarial resilience, resorting to the equilibrium-point theory in the discipline of complex dynamic system (CDS). In brief, our work has three novelties: i) Adversarial-Attack Modeling, i.e. map a graph regime into CDS, and use the oscillation of dynamic system to model the behavior of adversarial perturbation; ii) 2D Topology-Feature-Entangled Function Design for Perturbed Graph, i.e. project graph topology and node feature as two characteristic spaces, and define two-dimensional entangled perturbation functions to represent the dynamic variance under adversarial attacks; and iii) Location of Critical State of Adversarial Resilience, i.e. utilize the equilibrium-point theory to locate the graph's critical state of attack resilience resorting to the perturbation-reflected 2D function. Finally, multi-facet experiments on five commonly-used realistic datasets validate the effectiveness of our proposed approach, and the results show our approach can significantly outperform the state-of-the-art baselines under four representative graph adversarial attacks.", "date": "2026-04-15", "categories": []}
{"id": "2604.15708", "title": "APC: Transferable and Efficient Adversarial Point Counterattack for Robust 3D Point Cloud Recognition", "abstract": "The advent of deep neural networks has led to remarkable progress in 3D point cloud recognition, but they remain vulnerable to adversarial attacks. Although various defense methods have been studied, they suffer from a trade-off between robustness and transferability. We propose Adversarial Point Counterattack (APC) to achieve both simultaneously. APC is a lightweight input-level purification module that generates instance-specific counter-perturbations for each point, effectively neutralizing attacks. Leveraging clean-adversarial pairs, APC enforces geometric consistency in data space and semantic consistency in feature space. To improve generalizability across diverse attacks, we adopt a hybrid training strategy using adversarial point clouds from multiple attack types. Since APC operates purely on input point clouds, it directly transfers to unseen models and defends against attacks targeting them without retraining. At inference, a single APC forward pass provides purified point clouds with negligible time and parameter overhead. Extensive experiments on two 3D recognition benchmarks demonstrate that the APC achieves state-of-the-art defense performance. Furthermore, cross-model evaluations validate its superior transferability. The code is available at https://github.com/gyjung975/APC.", "date": "2026-04-17", "categories": []}
{"id": "2604.16037", "title": "Stochasticity in Tokenisation Improves Robustness", "abstract": "The widespread adoption of large language models (LLMs) has increased concerns about their robustness. Vulnerabilities in perturbations of tokenisation of the input indicate that models trained with a deterministic canonical tokenisation can be brittle to adversarial attacks. Recent studies suggest that stochastic tokenisation can deliver internal representations that are less sensitive to perturbations. In this paper, we analyse how stochastic tokenisations affect robustness to adversarial attacks and random perturbations. We systematically study this over a range of learning regimes (pre-training, supervised fine-tuning, and in-context learning), data sets, and model architectures. We show that pre-training and fine-tuning with uniformly sampled stochastic tokenisations improve robustness to random and adversarial perturbations. Evaluating on uniformly sampled non-canonical tokenisations reduces the accuracy of a canonically trained Llama-1b model by 29.8%. We find that training with stochastic tokenisation preserves accuracy without increasing inference cost.", "date": "2026-04-17", "categories": []}
{"id": "2408.14728", "title": "Improving Clean Accuracy via a Tangent-Space Perspective on Adversarial Training", "abstract": "Adversarial training has proven effective in improving the robustness of deep neural networks against adversarial attacks. However, this enhanced robustness often comes at the cost of a substantial drop in accuracy on clean data. In this paper, we address this limitation by introducing Tangent Direction Guided Adversarial Training (TART), a novel method that enhances clean accuracy by exploiting the geometry of the data manifold. We argue that adversarial examples with large components in the normal direction can overly distort the decision boundary and degrade clean accuracy. TART addresses this issue by estimating the tangent direction of adversarial examples and adaptively modulating the perturbation bound based on the norm of their tangential component. To the best of our knowledge, TART is the first adversarial defense framework that explicitly incorporates the concept of tangent space and direction into adversarial training. Extensive experiments on both synthetic and benchmark datasets demonstrate that TART consistently improves clean accuracy while maintaining robustness against adversarial attacks.", "date": "2024-08-27", "categories": []}
{"id": "2505.10846", "title": "AutoRAN: Automated Hijacking of Safety Reasoning in Large Reasoning Models", "abstract": "This paper presents AutoRAN, the first framework to automate the hijacking of internal safety reasoning in large reasoning models (LRMs). At its core, AutoRAN pioneers an execution simulation paradigm that leverages a weaker but less-aligned model to simulate execution reasoning for initial hijacking attempts and iteratively refine attacks by exploiting reasoning patterns leaked through the target LRM's refusals. This approach steers the target model to bypass its own safety guardrails and elaborate on harmful instructions. We evaluate AutoRAN against state-of-the-art LRMs, including GPT-o3/o4-mini and Gemini-2.5-Flash, across multiple benchmarks (AdvBench, HarmBench, and StrongReject). Results show that AutoRAN achieves approaching 100% success rate within one or few turns, effectively neutralizing reasoning-based defenses even when evaluated by robustly aligned external models. This work reveals that the transparency of the reasoning process itself creates a critical and exploitable attack surface, highlighting the urgent need for new defenses that protect models' reasoning traces rather than merely their final outputs.", "date": "2025-05-16", "categories": []}
{"id": "2601.03416", "title": "GAMBIT: A Gamified Jailbreak Framework for Multimodal Large Language Models", "abstract": "Multimodal Large Language Models (MLLMs) have become widely deployed, yet their safety alignment remains fragile under adversarial inputs. Previous work has shown that increasing inference steps can disrupt safety mechanisms and lead MLLMs to generate attacker-desired harmful content. However, most existing attacks focus on increasing the complexity of the modified visual task itself and do not explicitly leverage the model's own reasoning incentives. This leads to them underperforming on reasoning models (Models with Chain-of-Thoughts) compared to non-reasoning ones (Models without Chain-of-Thoughts). If a model can think like a human, can we influence its cognitive-stage decisions so that it proactively completes a jailbreak? To validate this idea, we propose GAMBI} (Gamified Adversarial Multimodal Breakout via Instructional Traps), a novel multimodal jailbreak framework that decomposes and reassembles harmful visual semantics, then constructs a gamified scene that drives the model to explore, reconstruct intent, and answer as part of winning the game. The resulting structured reasoning chain increases task complexity in both vision and text, positioning the model as a participant whose goal pursuit reduces safety attention and induces it to answer the reconstructed malicious query. Extensive experiments on popular reasoning and non-reasoning MLLMs demonstrate that GAMBIT achieves high Attack Success Rates (ASR), reaching 92.13% on Gemini 2.5 Flash, 91.20% on QvQ-MAX, and 85.87% on GPT-4o, significantly outperforming baselines.", "date": "2026-01-06", "categories": []}
{"id": "2604.14200", "title": "Retina gap junctions support the robust perception by warping neural representational geometries along the visual hierarchy", "abstract": "Deep Neural Networks (DNNs) are vulnerable to elaborately designed adversarial noise, although they have achieved extraordinary success in many tasks. Compared with DNNs, the human visual system is highly robust. However, it is unclear how the human visual system defends against adversarial attacks, especially the role of the early visual system and its influence on the brain manifold. Due to retina gap junctions being crucial for the denoising function in the early visual system, we combine a retina gap junction-based filter, G-filter, with DNN as an abstract human visual system model called the biological hybrid model. We adopt this model to study the defense performance of retina gap junctions and their impact on the brain manifold. Compared with other defense methods, the biological hybrid model is more robust and can be further improved by introducing noise during training. Next, we analyze the manifold and its decision boundary of the biological hybrid model from a geometry perspective. The results show that the biological hybrid model has a unique 2D decision boundary with high nonlinearity and a lower curvature of the decision boundary of the manifold compared to other defense methods. The transforming manifold may account for the high robustness of the biological hybrid model. Finally, to dissect G-filter and clarify its internal mechanism, we borrow the Neural Ordinary Differential Equation (ODE) concept and rewrite G-filter into an equivalent recurrent neural network. The results show that the decision boundary of the model's manifold will gradually change with time and eventually reach a steady state, which is modulated by gap junction conductance, revealing the influence of retina gap junctions on the brain manifold is a gradually evolving process.", "date": "2026-04-03", "categories": []}
{"id": "2604.14444", "title": "Robustness Analysis of Machine Learning Models for IoT Intrusion Detection Under Data Poisoning Attacks", "abstract": "Ensuring the reliability of machine learning-based intrusion detection systems remains a critical challenge in Internet of Things (IoT) environments, particularly as data poisoning attacks increasingly threaten the integrity of model training pipelines. This study evaluates the susceptibility of four widely used classifiers, Random Forest, Gradient Boosting Machine, Logistic Regression, and Deep Neural Network models, against multiple poisoning strategies using three real-world IoT datasets. Results show that while ensemble-based models exhibit comparatively stable performance, Logistic Regression and Deep Neural Networks suffer degradation of up to 40% under label manipulation and outlier-based attacks. Such disruptions significantly distort decision boundaries, reduce detection fidelity, and undermine deployment readiness. The findings highlight the need for adversarially robust training, continuous anomaly monitoring, and feature-level validation within operational Network Intrusion Detection Systems. The study also emphasizes the importance of integrating resilience testing into regulatory and compliance frameworks for AI-driven IoT security. Overall, this work provides an empirical foundation for developing more resilient intrusion detection pipelines and informs future research on adaptive, attack-aware models capable of maintaining reliability under adversarial IoT conditions.", "date": "2026-04-15", "categories": []}
{"id": "2604.14604", "title": "Hijacking Large Audio-Language Models via Context-Agnostic and Imperceptible Auditory Prompt Injection", "abstract": "Modern Large audio-language models (LALMs) power intelligent voice interactions by tightly integrating audio and text. This integration, however, expands the attack surface beyond text and introduces vulnerabilities in the continuous, high-dimensional audio channel. While prior work studied audio jailbreaks, the security risks of malicious audio injection and downstream behavior manipulation remain underexamined. In this work, we reveal a previously overlooked threat, auditory prompt injection, under realistic constraints of audio data-only access and strong perceptual stealth. To systematically analyze this threat, we propose \\textit{AudioHijack}, a general framework that generates context-agnostic and imperceptible adversarial audio to hijack LALMs. \\textit{AudioHijack} employs sampling-based gradient estimation for end-to-end optimization across diverse models, bypassing non-differentiable audio tokenization. Through attention supervision and multi-context training, it steers model attention toward adversarial audio and generalizes to unseen user contexts. We also design a convolutional blending method that modulates perturbations into natural reverberation, making them highly imperceptible to users. Extensive experiments on 13 state-of-the-art LALMs show consistent hijacking across 6 misbehavior categories, achieving average success rates of 79\\%-96\\% on unseen user contexts with high acoustic fidelity. Real-world studies demonstrate that commercial voice agents from Mistral AI and Microsoft Azure can be induced to execute unauthorized actions on behalf of users. These findings expose critical vulnerabilities in LALMs and highlight the urgent need for dedicated defense.", "date": "2026-04-16", "categories": []}
{"id": "2604.14643", "title": "Physically-Induced Atmospheric Adversarial Perturbations: Enhancing Transferability and Robustness in Remote Sensing Image Classification", "abstract": "Adversarial attacks pose a severe threat to the reliability of deep learning models in remote sensing (RS) image classification. Most existing methods rely on direct pixel-wise perturbations, failing to exploit the inherent atmospheric characteristics of RS imagery or survive real-world image degradations. In this paper, we propose FogFool, a physically plausible adversarial framework that generates fog-based perturbations by iteratively optimizing atmospheric patterns based on Perlin noise. By modeling fog formations with natural, irregular structures, FogFool generates adversarial examples that are not only visually consistent with authentic RS scenes but also deceptive. By leveraging the spatial coherence and mid-to-low-frequency nature of atmospheric phenomena, FogFool embeds adversarial information into structural features shared across diverse architectures. Extensive experiments on two benchmark RS datasets demonstrate that FogFool achieves superior performance: not only does it exceed in white-box settings, but also exhibits exceptional black-box transferability (reaching 83.74% TASR) and robustness against common preprocessing-based defenses such as JPEG compression and filtering. Detailed analyses, including confusion matrices and Class Activation Map (CAM) visualizations, reveal that our atmospheric-driven perturbations induce a universal shift in model attention. These results indicate that FogFool represents a practical, stealthy, and highly persistent threat to RS classification systems, providing a robust benchmark for evaluating model reliability in complex environments.", "date": "2026-04-16", "categories": []}
{"id": "2604.15022", "title": "Route to Rome Attack: Directing LLM Routers to Expensive Models via Adversarial Suffix Optimization", "abstract": "Cost-aware routing dynamically dispatches user queries to models of varying capability to balance performance and inference cost. However, the routing strategy introduces a new security concern that adversaries may manipulate the router to consistently select expensive high-capability models. Existing routing attacks depend on either white-box access or heuristic prompts, rendering them ineffective in real-world black-box scenarios. In this work, we propose R$^2$A, which aims to mislead black-box LLM routers to expensive models via adversarial suffix optimization. Specifically, R$^2$A deploys a hybrid ensemble surrogate router to mimic the black-box router. A suffix optimization algorithm is further adapted for the ensemble-based surrogate. Extensive experiments on multiple open-source and commercial routing systems demonstrate that {R$^2$A} significantly increases the routing rate to expensive models on queries of different distributions. Code and examples: https://github.com/thcxiker/R2A-Attack.", "date": "2026-04-16", "categories": []}
{"id": "2508.00555", "title": "Activation-Guided Local Editing for Jailbreaking Attacks", "abstract": "Jailbreaking is an essential adversarial technique for red-teaming these models to uncover and patch security flaws. However, existing jailbreak methods face significant drawbacks. Token-level jailbreak attacks often produce incoherent or unreadable inputs and exhibit poor transferability, while prompt-level attacks lack scalability and rely heavily on manual effort and human ingenuity. We propose a concise and effective two-stage framework that combines the advantages of these approaches. The first stage performs a scenario-based generation of context and rephrases the original malicious query to obscure its harmful intent. The second stage then utilizes information from the model's hidden states to guide fine-grained edits, effectively steering the model's internal representation of the input from a malicious toward a benign one. Extensive experiments demonstrate that this method achieves state-of-the-art Attack Success Rate, with gains of up to 37.74% over the strongest baseline, and exhibits excellent transferability to black-box models. Our analysis further demonstrates that AGILE maintains substantial effectiveness against prominent defense mechanisms, highlighting the limitations of current safeguards and providing valuable insights for future defense development. Our code is available at https://github.com/SELGroup/AGILE.", "date": "2025-08-01", "categories": []}
{"id": "2604.12371", "title": "Reading Between the Pixels: Linking Text-Image Embedding Alignment to Typographic Attack Success on Vision-Language Models", "abstract": "We study typographic prompt injection attacks on vision-language models (VLMs), where adversarial text is rendered as images to bypass safety mechanisms, posing a growing threat as VLMs serve as the perceptual backbone of autonomous agents, from browser automation and computer-use systems to camera-equipped embodied agents. In practice, the attack surface is heterogeneous: adversarial text appears at varying font sizes and under diverse visual conditions, while the growing ecosystem of VLMs exhibits substantial variation in vulnerability, complicating defensive approaches. Evaluating 1,000 prompts from SALAD-Bench across four VLMs, namely, GPT-4o, Claude Sonnet 4.5, Mistral-Large-3, and Qwen3-VL-4B-Instruct under varying font sizes (6--28px) and visual transformations (rotation, blur, noise, contrast changes), we find: (1) font size significantly affects attack success rate (ASR), with very small fonts (6px) yielding near-zero ASR while mid-range fonts achieve peak effectiveness; (2) text attacks are more effective than image attacks for GPT-4o (36% vs 8%) and Claude (47% vs 22%), while Qwen3-VL and Mistral show comparable ASR across modalities; (3) text-image embedding distance from two multimodal embedding models (JinaCLIP and Qwen3-VL-Embedding) shows strong negative correlation with ASR across all four models (r = -0.71 to -0.93, p < 0.01); (4) heavy degradations increase embedding distance by 10--12% and reduce ASR by 34--96%, while rotation asymmetrically affects models (Mistral drops 50%, GPT-4o unchanged). These findings highlight that model-specific robustness patterns preclude one-size-fits-all defenses and offer empirical guidance for practitioners selecting VLM backbones for agentic systems operating in adversarial environments.", "date": "2026-04-14", "categories": []}
{"id": "2604.13316", "title": "Beyond Uniform Sampling: Synergistic Active Learning and Input Denoising for Robust Neural Operators", "abstract": "Neural operators have emerged as fast surrogate models for physics simulations, yet they remain acutely vulnerable to adversarial perturbations, a critical liability for safety-critical digital twin deployments. We present a synergistic defense that combines active learning-based data generation with an input denoising architecture. The active learning component adaptively probes model weaknesses using differential evolution attacks, then generates targeted training data at discovered vulnerability locations while an adaptive smooth-ratio safeguard preserves baseline accuracy. The input denoising component augments the operator architecture with a learnable bottleneck that filters adversarial noise while retaining physics-relevant features. On the viscous Burgers' equation benchmark, the combined approach achieves a 2.04% combined error (1.21% baseline + 0.83% robustness), representing an 87% reduction relative to standard training (15.42% combined) and outperforming both active learning alone (3.42%) and input denoising alone (5.22%). More broadly, our results, combined with cross-architecture vulnerability analysis from prior work, suggest that optimal training data for neural operators is architecture-dependent: because different architectures concentrate sensitivity in distinct input subspaces, uniform sampling cannot adequately cover the vulnerability landscape of all models. These findings have potential implications for the deployment of neural operators in safety-critical energy systems including nuclear reactor monitoring.", "date": "2026-04-14", "categories": []}
{"id": "2501.13340", "title": "Retrievals Can Be Detrimental: Unveiling the Backdoor Vulnerability of Retrieval-Augmented Diffusion Models", "abstract": "Diffusion models (DMs) have recently demonstrated remarkable generation capability. However, their training generally requires huge computational resources and large-scale datasets. To solve these, recent studies empower DMs with the advanced Retrieval-Augmented Generation (RAG) technique and propose retrieval-augmented diffusion models (RDMs). By incorporating rich knowledge from an auxiliary database, RAG enhances diffusion models' generation and generalization ability while significantly reducing model parameters. Despite the great success, RAG may introduce novel security issues that warrant further investigation. In this paper, we reveal that the RDM is susceptible to backdoor attacks by proposing a multimodal contrastive attack approach named BadRDM. Our framework fully considers RAG's characteristics and is devised to manipulate the retrieved items for given text triggers, thereby further controlling the generated contents. Specifically, we first insert a tiny portion of images into the retrieval database as target toxicity surrogates. Subsequently, a malicious variant of contrastive learning is adopted to inject backdoors into the retriever, which builds shortcuts from triggers to the toxicity surrogates. Furthermore, we enhance the attacks through novel entropy-based selection and generative augmentation strategies that can derive better toxicity surrogates. Extensive experiments on two mainstream tasks demonstrate the proposed BadRDM achieves outstanding attack effects while preserving the model's benign utility.", "date": "2025-01-23", "categories": []}
{"id": "2511.09780", "title": "Hail to the Thief: Exploring Attacks and Defenses in Decentralised GRPO", "abstract": "Group Relative Policy Optimization (GRPO) has demonstrated wide adoption in the post-training of Large Language Models (LLMs). In GRPO, prompts are answered by the model and preferred behaviour is learnt via reinforcement learning. Owing to the small communication volume, GRPO is inherently suitable for decentralised training as the prompts can be concurrently answered by multiple nodes and these completions are exchanged in the form of strings. In this work, we explore the robustness of decentralised GRPO by presenting the first adversarial attacks and countermeasures. We present a diverse set of attacks where malicious nodes poison benign models by sharing their poisoned completions. We demonstrate these attacks on math and coding tasks and show that an adversary can achieve attack success rates of up to 100% in as few as 50 iterations. Moreover, to mitigate the attacks, we propose two defense mechanisms that check logit probabilities of completions or utilize an LLM judge to filter completions. The defenses prevent all but the DoS attack that causes unnecessarily lengthy but conceptually correct completions. The code of both attacks and defenses can be found at: https://github.com/gensyn-ai/HTTT.", "date": "2025-11-12", "categories": []}
{"id": "2603.05004", "title": "Poisoning the Inner Prediction Logic of Graph Neural Networks for Clean-Label Backdoor Attacks", "abstract": "Graph Neural Networks (GNNs) have achieved remarkable results in various tasks. Recent studies reveal that graph backdoor attacks can poison the GNN model to predict test nodes with triggers attached as the target class. However, apart from injecting triggers to training nodes, these graph backdoor attacks generally require altering the labels of trigger-attached training nodes into the target class, which is impractical in real-world scenarios. In this work, we focus on the clean-label graph backdoor attack, a realistic but understudied topic where training labels are not modifiable. According to our preliminary analysis, existing graph backdoor attacks generally fail under the clean-label setting. Our further analysis identifies that the core failure of existing methods lies in their inability to poison the prediction logic of GNN models, leading to the triggers being deemed unimportant for prediction. Therefore, we study a novel problem of effective clean-label graph backdoor attacks by poisoning the inner prediction logic of GNN models. We propose BA-Logic to solve the problem by coordinating a poisoned node selector and a logic-poisoning trigger generator. Extensive experiments on real-world datasets demonstrate that our method effectively enhances the attack success rate and surpasses state-of-the-art graph backdoor attack competitors under clean-label settings. Our code is available at https://anonymous.4open.science/r/BA-Logic", "date": "2026-03-05", "categories": []}
{"id": "2604.11928", "title": "INTARG: Informed Real-Time Adversarial Attack Generation for Time-Series Regression", "abstract": "Time-series forecasting aims to predict future values by modeling temporal dependencies in historical observations. It is a critical component of many real-world systems, where accurate forecasts improve operational efficiency and help mitigate uncertainty and risk. More recently, machine learning (ML), and especially deep learning (DL)-based models, have gained widespread adoption for time-series forecasting, but they remain vulnerable to adversarial attacks. However, many state-of-the-art attack methods are not directly applicable in time-series settings, where storing complete historical data or performing attacks at every time step is often impractical. This paper proposes an adversarial attack framework for time-series forecasting under an online bounded-buffer setting, leveraging an informed and selective attack strategy. By selectively targeting time steps where the model exhibits high confidence and the expected prediction error is maximal, our framework produces fewer but substantially more effective attacks. Experiments show that our framework can increase the prediction error up to 2.42x, while performing attacks in fewer than 10% of time steps.", "date": "2026-04-13", "categories": []}
{"id": "2604.12616", "title": "Every Picture Tells a Dangerous Story: Memory-Augmented Multi-Agent Jailbreak Attacks on VLMs", "abstract": "The rapid evolution of Vision-Language Models (VLMs) has catalyzed unprecedented capabilities in artificial intelligence; however, this continuous modal expansion has inadvertently exposed a vastly broadened and unconstrained adversarial attack surface. Current multimodal jailbreak strategies primarily focus on surface-level pixel perturbations and typographic attacks or harmful images; however, they fail to engage with the complex semantic structures intrinsic to visual data. This leaves the vast semantic attack surface of original, natural images largely unscrutinized. Driven by the need to expose these deep-seated semantic vulnerabilities, we introduce \\textbf{MemJack}, a \\textbf{MEM}ory-augmented multi-agent \\textbf{JA}ilbreak atta\\textbf{CK} framework that explicitly leverages visual semantics to orchestrate automated jailbreak attacks. MemJack employs coordinated multi-agent cooperation to dynamically map visual entities to malicious intents, generate adversarial prompts via multi-angle visual-semantic camouflage, and utilize an Iterative Nullspace Projection (INLP) geometric filter to bypass premature latent space refusals. By accumulating and transferring successful strategies through a persistent Multimodal Experience Memory, MemJack maintains highly coherent extended multi-turn jailbreak attack interactions across different images, thereby improving the attack success rate (ASR) on new images. Extensive empirical evaluations across full, unmodified COCO val2017 images demonstrate that MemJack achieves a 71.48\\% ASR against Qwen3-VL-Plus, scaling to 90\\% under extended budgets. Furthermore, to catalyze future defensive alignment research, we will release \\textbf{MemJack-Bench}, a comprehensive dataset comprising over 113,000 interactive multimodal jailbreak attack trajectories, establishing a vital foundation for developing inherently robust VLMs.", "date": "2026-04-14", "categories": []}
{"id": "2604.12780", "title": "Efficient Adversarial Training via Criticality-Aware Fine-Tuning", "abstract": "Vision Transformer (ViT) models have achieved remarkable performance across various vision tasks, with scalability being a key advantage when applied to large datasets. This scalability enables ViT models to exhibit strong generalization capabilities. However, as the number of parameters increases, the robustness of ViT models to adversarial examples does not scale proportionally. Adversarial training (AT), one of the most effective methods for enhancing robustness, typically requires fine-tuning the entire model, leading to prohibitively high computational costs, especially for large ViT architectures. In this paper, we aim to robustly fine-tune only a small subset of parameters to achieve robustness comparable to standard AT. To accomplish this, we introduce Criticality-Aware Adversarial Training (CAAT), a novel method that adaptively allocates resources to the most robustness-critical parameters, fine-tuning only selected modules. Specifically, CAAT efficiently identifies parameters that contribute most to adversarial robustness. It then leverages parameter-efficient fine-tuning (PEFT) to robustly adjust weight matrices where the number of critical parameters exceeds a predefined threshold. CAAT exhibits favorable generalization when scaled to larger vision transformer architectures, potentially paving the way for adversarial training at scale, e.g, compared with plain adversarial training, CAAT incurs only a 4.3% decrease in adversarial robustness while tuning approximately 6% of its parameters. Extensive experiments on three widely used adversarial learning datasets demonstrate that CAAT outperforms state-of-the-art lightweight AT methods with fewer trainable parameters.", "date": "2026-04-14", "categories": []}
{"id": "2604.12781", "title": "Fragile Reconstruction: Adversarial Vulnerability of Reconstruction-Based Detectors for Diffusion-Generated Images", "abstract": "Recently, detecting AI-generated images produced by diffusion-based models has attracted increasing attention due to their potential threat to safety. Among existing approaches, reconstruction-based methods have emerged as a prominent paradigm for this task. However, we find that such methods exhibit severe security vulnerabilities to adversarial perturbations; that is, by adding imperceptible adversarial perturbations to input images, the detection accuracy of classifiers collapses to near zero. To verify this threat, we present a systematic evaluation of the adversarial robustness of three representative detectors across four diverse generative backbone models. First, we construct adversarial attacks in white-box scenarios, which degrade the performance of all well-trained detectors. Moreover, we find that these attacks demonstrate transferability; specifically, attacks crafted against one detector can be transferred to others, indicating that adversarial attacks on detectors can also be constructed in a black-box setting. Finally, we assess common countermeasures and find that standard defense methods against adversarial attacks provide limited mitigation. We attribute these failures to the low signal-to-noise ratio (SNR) of attacked samples as perceived by the detectors. Overall, our results reveal fundamental security limitations of reconstruction-based detectors and highlight the need to rethink existing detection strategies.", "date": "2026-04-14", "categories": []}
{"id": "2604.12817", "title": "Understanding and Improving Continuous Adversarial Training for LLMs via In-context Learning Theory", "abstract": "Adversarial training (AT) is an effective defense for large language models (LLMs) against jailbreak attacks, but performing AT on LLMs is costly. To improve the efficiency of AT for LLMs, recent studies propose continuous AT (CAT) that searches for adversarial inputs within the continuous embedding space of LLMs during AT. While CAT has achieved empirical success, its underlying mechanism, i.e., why adversarial perturbations in the embedding space can help LLMs defend against jailbreak prompts synthesized in the input token space, remains unknown. This paper presents the first theoretical analysis of CAT on LLMs based on in-context learning (ICL) theory. For linear transformers trained with adversarial examples from the embedding space on in-context linear regression tasks, we prove a robust generalization bound that has a negative correlation with the perturbation radius in the embedding space. This clearly explains why CAT can defend against jailbreak prompts from the LLM's token space. Further, the robust bound shows that the robustness of an adversarially trained LLM is closely related to the singular values of its embedding matrix. Based on this, we propose to improve LLM CAT by introducing an additional regularization term, which depends on singular values of the LLM's embedding matrix, into the objective function of CAT. Experiments on real-world LLMs demonstrate that our method can help LLMs achieve a better jailbreak robustness-utility tradeoff. The code is available at https://github.com/fshp971/continuous-adv-icl.", "date": "2026-04-14", "categories": []}
{"id": "2407.11764", "title": "Adversarial Robustness of Graph Transformers", "abstract": "Existing studies have shown that Message-Passing Graph Neural Networks (MPNNs) are highly susceptible to adversarial attacks. In contrast, despite the increasing importance of Graph Transformers (GTs), their robustness properties are unexplored. We close this gap and design the first adaptive attacks for GTs. In particular, we provide general design principles for strong gradient-based attacks on GTs w.r.t. structure perturbations and instantiate our attack framework for five representative and popular GT architectures. Specifically, we study GTs with specialized attention mechanisms and Positional Encodings (PEs) based on pairwise shortest paths, random walks, and the Laplacian spectrum. We evaluate our attacks on multiple tasks and perturbation models, including structure perturbations for node and graph classification, and node injection for graph classification. Our results reveal that GTs can be catastrophically fragile in many cases. Addressing this vulnerability, we show how our adaptive attacks can be effectively used for adversarial training, substantially improving robustness.", "date": "2024-07-16", "categories": []}
{"id": "2412.12373", "title": "Quantum Adversarial Machine Learning and Defense Strategies: Challenges and Opportunities", "abstract": "As quantum computing continues to advance, the development of quantum-secure neural networks is crucial to prevent adversarial attacks. This paper proposes three quantum-secure design principles: (1) using post-quantum cryptography, (2) employing quantum-resistant neural network architectures, and (3) ensuring transparent and accountable development and deployment. These principles are supported by various quantum strategies, including quantum data anonymization, quantum-resistant neural networks, and quantum encryption. The paper also identifies open issues in quantum security, privacy, and trust, and recommends exploring adaptive adversarial attacks and auto adversarial attacks as future directions. The proposed design principles and recommendations provide guidance for developing quantum-secure neural networks, ensuring the integrity and reliability of machine learning models in the quantum era.", "date": "2024-12-16", "categories": []}
{"id": "2506.14493", "title": "LingoLoop Attack: Trapping MLLMs via Linguistic Context and State Entrapment into Endless Loops", "abstract": "Multimodal Large Language Models (MLLMs) have shown great promise but require substantial computational resources during inference. Attackers can exploit this by inducing excessive output, leading to resource exhaustion and service degradation. Prior energy-latency attacks aim to increase generation time by broadly shifting the output token distribution away from the EOS token, but they neglect the influence of token-level Part-of-Speech (POS) characteristics on EOS and sentence-level structural patterns on output counts, limiting their efficacy. To address this, we propose LingoLoop, an attack designed to induce MLLMs to generate excessively verbose and repetitive sequences. First, we find that the POS tag of a token strongly affects the likelihood of generating an EOS token. Based on this insight, we propose a POS-Aware Delay Mechanism to postpone EOS token generation by adjusting attention weights guided by POS information. Second, we identify that constraining output diversity to induce repetitive loops is effective for sustained generation. We introduce a Generative Path Pruning Mechanism that limits the magnitude of hidden states, encouraging the model to produce persistent loops. Extensive experiments on models like Qwen2.5-VL-3B demonstrate LingoLoop's powerful ability to trap them in generative loops; it consistently drives them to their generation limits and, when those limits are relaxed, can induce outputs with up to 367x more tokens than clean inputs, triggering a commensurate surge in energy consumption. These findings expose significant MLLMs' vulnerabilities, posing challenges for their reliable deployment.", "date": "2025-06-17", "categories": []}
{"id": "2508.06964", "title": "Adversarial Video Promotion Against Text-to-Video Retrieval", "abstract": "Thanks to the development of cross-modal models, text-to-video retrieval (T2VR) is advancing rapidly, but its robustness remains largely unexamined. Existing attacks against T2VR are designed to push videos away from queries, i.e., suppressing the ranks of videos, while the attacks that pull videos towards selected queries, i.e., promoting the ranks of videos, remain largely unexplored. These attacks can be more impactful as attackers may gain more views/clicks for financial benefits and widespread (mis)information. To this end, we pioneer the first attack against T2VR to promote videos adversarially, dubbed the Video Promotion attack (ViPro). We further propose Modal Refinement (MoRe) to capture the finer-grained, intricate interaction between visual and textual modalities to enhance black-box transferability. Comprehensive experiments cover 2 existing baselines, 3 leading T2VR models, 3 prevailing datasets with over 10k videos, evaluated under 3 scenarios. All experiments are conducted in a multi-target setting to reflect realistic scenarios where attackers seek to promote the video regarding multiple queries simultaneously. We also evaluated our attacks for defences and imperceptibility. Overall, ViPro surpasses other baselines by over $30/10/4\\%$ for white/grey/black-box settings on average. Our work highlights an overlooked vulnerability, provides a qualitative analysis on the upper/lower bound of our attacks, and offers insights into potential counterplays. Code will be publicly available at https://github.com/michaeltian108/ViPro.", "date": "2025-08-09", "categories": []}
{"id": "2604.08557", "title": "Re-Mask and Redirect: Exploiting Denoising Irreversibility in Diffusion Language Models", "abstract": "Safety alignment in diffusion language models (dLLMs) relies on a single load-bearing assumption: that committed tokens are permanent. We show that violating this assumption, by re-masking committed refusal tokens and injecting a short affirmative prefix, achieves 74-82% ASR on HarmBench across all three publicly available safety-tuned dLLMs, rising to 92-98% with a generic 8-token compliance prefix. We call this attack TrajHijack; it is the first trajectory-level attack on dLLMs, requires no gradient computation, and generalizes across SFT and preference-optimized (VRPO) models. Three findings emerge. First, the vulnerability is irreducibly two-component: re-masking alone (4.4%) and prefix alone (5.7%) both fail. Second, gradient optimization via a differentiable Gumbel-softmax chain consistently degrades ASR (41.5% vs. 76.1%), because continuous perturbations push token distributions off-manifold. Third, A2D (the strongest published dLLM defense) is more vulnerable to TrajHijack (89.9%) than the undefended model (76.1%): its silent-refusal training removes the contextual resistance that trajectory-level attacks must overcome, an effect we call the Defense Inversion Effect.", "date": "2026-03-17", "categories": []}
{"id": "2604.09651", "title": "FlowHijack: A Dynamics-Aware Backdoor Attack on Flow-Matching Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models are emerging as a cornerstone for robotics, with flow-matching policies like $\\pi_0$ showing great promise in generating smooth, continuous actions. As these models advance, their unique action generation mechanism - the vector field dynamics - presents a critical yet unexplored security vulnerability, particularly backdoor vulnerabilities. Existing backdoor attacks designed for autoregressive discretization VLAs cannot be directly applied to this new continuous dynamics. We introduce FlowHijack, the first backdoor attack framework to systematically target the underlying vector-field dynamics of flow-matching VLAs. Our method combines a novel $\\tau$-conditioned injection strategy, which manipulates the initial phase of the action generation, with a dynamics mimicry regularizer. Experiments demonstrate that FlowHijack achieves high attack success rates using stealthy, context-aware triggers where prior works failed. Crucially, it preserves benign task performance and, by enforcing kinematic similarity, generates malicious actions that are behaviorally indistinguishable from normal actions. Our findings reveal a significant vulnerability in continuous embodied models, highlighting the urgent need for defenses targeting the model's internal generative dynamics.", "date": "2026-03-30", "categories": []}
{"id": "2604.09706", "title": "The Deployment Gap in AI Media Detection: Platform-Aware and Visually Constrained Adversarial Evaluation", "abstract": "Recent AI media detectors report near-perfect performance under clean laboratory evaluation, yet their robustness under realistic deployment conditions remains underexplored. In practice, AI-generated images are resized, compressed, re-encoded, and visually modified before being shared on online platforms. We argue that this creates a deployment gap between laboratory robustness and real-world reliability.\n  In this work, we introduce a platform-aware adversarial evaluation framework for AI media detection that explicitly models deployment transforms (e.g., resizing, compression, screenshot-style distortions) and constrains perturbations to visually plausible meme-style bands rather than full-image noise. Under this threat model, detectors achieving AUC $\\approx$ 0{.}99 in clean settings experience substantial degradation. Per-image platform-aware attacks reduce AUC to significantly lower levels and achieve high fake-to-real misclassification rates, despite strict visual constraints. We further demonstrate that universal perturbations exist even under localized band constraints, revealing shared vulnerability directions across inputs. Beyond accuracy degradation, we observe pronounced calibration collapse under attack, where detectors become confidently incorrect.\n  Our findings highlight that robustness measured under clean conditions substantially overestimates deployment robustness. We advocate for platform-aware evaluation as a necessary component of future AI media security benchmarks and release our evaluation framework to facilitate standardized robustness assessment.", "date": "2026-04-07", "categories": []}
{"id": "2604.09879", "title": "Topo-ADV: Generating Topology-Driven Imperceptible Adversarial Point Clouds", "abstract": "Deep neural networks for 3D point cloud understanding have achieved remarkable success in object classification and recognition, yet recent work shows that these models remain highly vulnerable to adversarial perturbations. Existing 3D attacks predominantly manipulate geometric properties such as point locations, curvature, or surface structure, implicitly assuming that preserving global shape fidelity preserves semantic content. In this work, we challenge this assumption and introduce the first topology-driven adversarial attack for point cloud deep learning. Our key insight is that the homological structure of a 3D object constitutes a previously unexplored vulnerability surface. We propose Topo-ADV, an end-to-end differentiable framework that incorporates persistent homology as an explicit optimization objective, enabling gradient-based manipulation of topological features during adversarial example generation. By embedding persistence diagrams through differentiable topological representations, our method jointly optimizes (i) a topology divergence loss that alters persistence, (ii) a misclassification objective, and (iii) geometric imperceptibility constraints that preserve visual plausibility. Experiments demonstrate that subtle topology-driven perturbations consistently achieve up to 100% attack success rates on benchmark datasets such as ModelNet40, ShapeNet Part, and ScanObjectNN using PointNet and DGCNN classifiers, while remaining geometrically indistinguishable from the original point clouds, beating state-of-the-art methods on various perceptibility metrics.", "date": "2026-04-10", "categories": []}
{"id": "2604.10326", "title": "Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion", "abstract": "Large language models remain vulnerable to jailbreak attacks -- inputs designed to bypass safety mechanisms and elicit harmful responses -- despite advances in alignment and instruction tuning. We propose Head-Masked Nullspace Steering (HMNS), a circuit-level intervention that (i) identifies attention heads most causally responsible for a model's default behavior, (ii) suppresses their write paths via targeted column masking, and (iii) injects a perturbation constrained to the orthogonal complement of the muted subspace. HMNS operates in a closed-loop detection-intervention cycle, re-identifying causal heads and reapplying interventions across multiple decoding attempts. Across multiple jailbreak benchmarks, strong safety defenses, and widely used language models, HMNS attains state-of-the-art attack success rates with fewer queries than prior methods. Ablations confirm that nullspace-constrained injection, residual norm scaling, and iterative re-identification are key to its effectiveness. To our knowledge, this is the first jailbreak method to leverage geometry-aware, interpretability-informed interventions, highlighting a new paradigm for controlled model steering and adversarial safety circumvention.", "date": "2026-04-11", "categories": []}
{"id": "2604.10715", "title": "Defending against Patch-Based and Texture-Based Adversarial Attacks with Spectral Decomposition", "abstract": "Adversarial examples present significant challenges to the security of Deep Neural Network (DNN) applications. Specifically, there are patch-based and texture-based attacks that are usually used to craft physical-world adversarial examples, posing real threats to security-critical applications such as person detection in surveillance and autonomous systems, because those attacks are physically realizable. Existing defense mechanisms face challenges in the adaptive attack setting, i.e., the attacks are specifically designed against them. In this paper, we propose Adversarial Spectrum Defense (ASD), a defense mechanism that leverages spectral decomposition via Discrete Wavelet Transform (DWT) to analyze adversarial patterns across multiple frequency scales. The multi-resolution and localization capability of DWT enables ASD to capture both high-frequency (fine-grained) and low-frequency (spatially pervasive) perturbations. By integrating this spectral analysis with the off-the-shelf Adversarial Training (AT) model, ASD provides a comprehensive defense strategy against both patch-based and texture-based adversarial attacks. Extensive experiments demonstrate that ASD+AT achieved state-of-the-art (SOTA) performance against various attacks, outperforming the APs of previous defense methods by 21.73%, in the face of strong adaptive adversaries specifically designed against ASD. Code available at https://github.com/weiz0823/adv-spectral-defense .", "date": "2026-04-12", "categories": []}
{"id": "2604.10837", "title": "Immune2V: Image Immunization Against Dual-Stream Image-to-Video Generation", "abstract": "Image-to-video (I2V) generation has the potential for societal harm because it enables the unauthorized animation of static images to create realistic deepfakes. While existing defenses effectively protect against static image manipulation, extending these to I2V generation remains underexplored and non-trivial. In this paper, we systematically analyze why modern I2V models are highly robust against naive image-level adversarial attacks (i.e., immunization). We observe that the video encoding process rapidly dilutes the adversarial noise across future frames, and the continuous text-conditioned guidance actively overrides the intended disruptive effect of the immunization. Building on these findings, we propose the Immune2V framework which enforces temporally balanced latent divergence at the encoder level to prevent signal dilution, and aligns intermediate generative representations with a precomputed collapse-inducing trajectory to counteract the text-guidance override. Extensive experiments demonstrate that Immune2V produces substantially stronger and more persistent degradation than adapted image-level baselines under the same imperceptibility budget.", "date": "2026-04-12", "categories": []}
{"id": "2604.10933", "title": "QShield: Securing Neural Networks Against Adversarial Attacks using Quantum Circuits", "abstract": "Deep neural networks remain highly vulnerable to adversarial perturbations, limiting their reliability in security- and safety-critical applications. To address this challenge, we introduce QShield, a modular hybrid quantum-classical neural network (HQCNN) architecture designed to enhance the adversarial robustness of classical deep learning models. QShield integrates a conventional convolutional neural network (CNN) backbone for feature extraction with a quantum processing module that encodes the extracted features into quantum states, applies structured entanglement operations under realistic noise models, and outputs a hybrid prediction through a dynamically weighted fusion mechanism implemented via a lightweight multilayer perceptron (MLP). We systematically evaluate both classical and hybrid quantum-classical models on the MNIST, OrganAMNIST, and CIFAR-10 datasets, using a comprehensive set of robustness, efficiency, and computational performance metrics. Our results demonstrate that classical models are highly vulnerable to adversarial attacks, whereas the proposed hybrid models with entanglement patterns maintain high predictive accuracy while substantially reducing attack success rates across a wide range of adversarial attacks. Furthermore, the proposed hybrid architecture significantly increased the computational cost required to generate adversarial examples, thereby introducing an additional layer of defense. These findings indicate that the proposed modular hybrid architecture achieves a practical balance between predictive accuracy and adversarial robustness, positioning it as a promising approach for secure and reliable machine learning in sensitive and safety-critical applications.", "date": "2026-04-13", "categories": []}
{"id": "2604.11576", "title": "Finetune Like You Pretrain: Boosting Zero-shot Adversarial Robustness in Vision-language Models", "abstract": "Despite their impressive zero-shot abilities, vision-language models such as CLIP have been shown to be susceptible to adversarial attacks. To enhance its adversarial robustness, recent studies finetune the pretrained vision encoder of CLIP with adversarial examples on a proxy dataset such as ImageNet by aligning adversarial images with correct class labels. However, these methods overlook the important roles of training data distributions and learning objectives, resulting in reduced zero-shot capabilities and limited transferability of robustness across domains and datasets. In this work, we propose a simple yet effective paradigm AdvFLYP, which follows the training recipe of CLIP's pretraining process when performing adversarial finetuning to the model. Specifically, AdvFLYP finetunes CLIP with adversarial images created based on image-text pairs collected from the web, and match them with their corresponding texts via a contrastive loss. To alleviate distortion of adversarial image embeddings of noisy web images, we further propose to regularise AdvFLYP by penalising deviation of adversarial image features. We show that logit- and feature-level regularisation terms benefit robustness and clean accuracy, respectively. Extensive experiments on 14 downstream datasets spanning various domains show the superiority of our paradigm over mainstream practices. Our code and model weights are released at https://github.com/Sxing2/AdvFLYP.", "date": "2026-04-13", "categories": []}
{"id": "2604.11590", "title": "Learning Robustness at Test-Time from a Non-Robust Teacher", "abstract": "Nowadays, pretrained models are increasingly used as general-purpose backbones and adapted at test-time to downstream environments where target data are scarce and unlabeled. While this paradigm has proven effective for improving clean accuracy on the target domain, adversarial robustness has received far less attention, especially when the original pretrained model is not explicitly designed to be robust. This raises a practical question: \\emph{can a pretrained, non-robust model be adapted at test-time to improve adversarial robustness on a target distribution?} To face this question, this work studies how adversarial training strategies behave when integrated into adaptation schemes for the unsupervised test-time setting, where only a small set of unlabeled target samples is available. It first analyzes how classical adversarial training formulations can be extended to this scenario, showing that straightforward distillation-based adaptations remain unstable and highly sensitive to hyperparameter tuning, particularly when the teacher itself is non-robust.\n  To address these limitations, the work proposes a label-free framework that uses the predictions of a non-robust teacher model as a semantic anchor for both the clean and adversarial objectives during adaptation. We further provide theoretical insights showing that our formulation yields a more stable alternative to the self-consistency-based regularization commonly used in classical adversarial training. Experiments evaluate the proposed approach on CIFAR-10 and ImageNet under induced photometric transformations. The results support the theoretical insights by showing that the proposed approach achieves improved optimization stability, lower sensitivity to parameter choices, and a better robustness-accuracy trade-off than existing baselines in this post-deployment test-time setting.", "date": "2026-04-13", "categories": []}
{"id": "2604.11720", "title": "On the Robustness of Watermarking for Autoregressive Image Generation", "abstract": "The proliferation of autoregressive (AR) image generators demands reliable detection and attribution of their outputs to mitigate misinformation, and to filter synthetic images from training data to prevent model collapse. To address this need, watermarking techniques, specifically designed for AR models, embed a subtle signal at generation time, enabling downstream verification through a corresponding watermark detector. In this work, we study these schemes and demonstrate their vulnerability to both watermark removal and forgery attacks. We assess existing attacks and further introduce three new attacks: (i) a vector-quantized regeneration removal attack, (ii) adversarial optimization-based attack, and (iii) a frequency injection attack. Our evaluation reveals that removal and forgery attacks can be effective with access to a single watermarked reference image and without access to original model parameters or watermarking secrets. Our findings indicate that existing watermarking schemes for AR image generation do not reliably support synthetic content detection for dataset filtering. Moreover, they enable Watermark Mimicry, whereby authentic images can be manipulated to imitate a generator's watermark and trigger false detection to prevent their inclusion in future model training.", "date": "2026-04-13", "categories": []}
{"id": "2501.11568", "title": "Graph Defense Diffusion Model", "abstract": "Graph Neural Networks (GNNs) are highly vulnerable to adversarial attacks, which can greatly degrade their performance. Existing graph purification methods attempt to address this issue by filtering attacked graphs. However, they struggle to defend effectively against multiple types of adversarial attacks (e.g., targeted attacks and non-targeted attacks) simultaneously due to limited flexibility. Additionally, these methods lack comprehensive modeling of graph data, relying heavily on heuristic prior knowledge. To overcome these challenges, we introduce the Graph Defense Diffusion Model (GDDM), a flexible purification method that leverages the denoising and modeling capabilities of diffusion models. The iterative nature of diffusion models aligns well with the stepwise process of adversarial attacks, making them particularly suitable for defense. By iteratively adding and removing noises (edges), GDDM effectively purifies attacked graphs, restoring their original structures and features. Our GDDM consists of two key components: (1) Graph Structure-Driven Refiner, which preserves the basic fidelity of the graph during the denoising process, and ensures that the generated graph remains consistent with the original scope; and (2) Node Feature-Constrained Regularizer, which removes residual impurities from the denoised graph, further enhancing the purification effect. By designing tailored denoising strategies to handle different types of adversarial attacks, we improve the GDDM's adaptability to various attack scenarios. Furthermore, GDDM demonstrates strong scalability, leveraging its structural properties to seamlessly transfer across similar datasets without retraining. Extensive experiments on three real-world datasets demonstrate that GDDM outperforms state-of-the-art methods in defending against various adversarial attacks, showcasing its robustness and effectiveness.", "date": "2025-01-20", "categories": []}
{"id": "2511.09829", "title": "Thermally Activated Dual-Modal Adversarial Clothing against AI Surveillance Systems", "abstract": "Adversarial patches have emerged as a popular privacy-preserving approach for resisting AI-driven surveillance systems. However, their conspicuous appearance makes them difficult to deploy in real-world scenarios. In this paper, we propose a thermally activated adversarial wearable designed to ensure adaptability and effectiveness in complex real-world environments. The system integrates thermochromic dyes with flexible heating units to induce visually dynamic adversarial patterns on clothing surfaces. In its default state, the clothing appears as an ordinary black T-shirt. Upon heating via an embedded thermal unit, hidden adversarial patterns on the fabric are activated, allowing the wearer to effectively evade detection across both visible and infrared modalities. Physical experiments demonstrate that the adversarial wearable achieves rapid texture activation within 50 seconds and maintains an adversarial success rate above 80\\% across diverse real-world surveillance environments. This work demonstrates a new pathway toward physically grounded, user-controllable anti-AI systems, highlighting the growing importance of proactive adversarial techniques for privacy protection in the age of ubiquitous AI surveillance.", "date": "2025-11-13", "categories": []}
{"id": "2512.03730", "title": "Out-of-the-box: Black-box Causal Attacks on Object Detectors", "abstract": "Adversarial perturbations are a useful way to expose vulnerabilities in object detectors. Existing perturbation methods are frequently white-box, architecture specific and use a loss function. More importantly, while they are often successful, it is rarely clear why they work. Insights into the mechanism of this success would allow developers to understand and analyze these attacks, as well as fine-tune the model to prevent them.\n  This paper presents BlackCAtt, a black-box algorithm and tool, which uses minimal, causally sufficient pixel sets to construct explainable, imperceptible, reproducible, architecture-agnostic attacks on object detectors. We evaluate BlackCAtt on standard benchmarks and compare it to other black-box adversarial attacks methods. When BlackCAtt has access only to the position and label of a bounding box, it produces attacks that are comparable or better to those produced by other black-box methods. When BlackCAtt has access to the model confidence as well, it can work as a meta-algorithm, improving the ability of standard black-box techniques to construct smaller, less perceptible attacks.\n  As BlackCAtt attacks manipulate causes only, the attacks become fully explainable. We compare the performance of BlackCAtt with other black-box attack methods and show that targeting causal pixels leads to smaller and less perceptible attacks. For example, when using BlackCAtt with SquareAttack, it reduces the average distance ($L_0$ norm) of the attack from the original input from $0.987$ to $0.072$, while maintaining a similar success rate. We perform ablation studies on the BlackCAtt algorithm and analyze the effect of different components on its performance.", "date": "2025-12-03", "categories": []}
{"id": "2601.10587", "title": "Adversarial Evasion Attacks on Computer Vision using SHAP Values", "abstract": "The paper introduces a white-box attack on computer vision models using SHAP values. It demonstrates how adversarial evasion attacks can compromise the performance of deep learning models by reducing output confidence or inducing misclassifications. Such attacks are particularly insidious as they can deceive the perception of an algorithm while eluding human perception due to their imperceptibility to the human eye. The proposed attack leverages SHAP values to quantify the significance of individual inputs to the output at the inference stage. A comparison is drawn between the SHAP attack and the well-known Fast Gradient Sign Method. We find evidence that SHAP attacks are more robust in generating misclassifications particularly in gradient hiding scenarios.", "date": "2026-01-15", "categories": []}
{"id": "2604.08405", "title": "SyncBreaker:Stage-Aware Multimodal Adversarial Attacks on Audio-Driven Talking Head Generation", "abstract": "Diffusion-based audio-driven talking-head generation enables realistic portrait animation, but also introduces risks of misuse, such as fraud and misinformation. Existing protection methods are largely limited to a single modality, and neither image-only nor audio-only attacks can effectively suppress speech-driven facial dynamics. To address this gap, we propose SyncBreaker, a stage-aware multimodal protection framework that jointly perturbs portrait and audio inputs under modality-specific perceptual constraints. Our key contributions are twofold. First, for the image stream, we introduce nullifying supervision with Multi-Interval Sampling (MIS) across diffusion stages to steer the generation toward the static reference portrait by aggregating guidance from multiple denoising intervals. Second, for the audio stream, we propose Cross-Attention Fooling (CAF), which suppresses interval-specific audio-conditioned cross-attention responses. Both streams are optimized independently and combined at inference time to enable flexible deployment. We evaluate SyncBreaker in a white-box proactive protection setting. Extensive experiments demonstrate that SyncBreaker more effectively degrades lip synchronization and facial dynamics than strong single-modality baselines, while preserving input perceptual quality and remaining robust under purification. Code: https://github.com/kitty384/SyncBreaker.", "date": "2026-04-09", "categories": []}
{"id": "2604.08750", "title": "Adversarial Sensor Errors for Safe and Robust Wind Turbine Fleet Control", "abstract": "Plant-level control is an emerging wind energy technology that presents opportunities and challenges. By controlling turbines in a coordinated manner via a central controller, it is possible to achieve greater wind power plant efficiency. However, there is a risk that measurement errors will confound the process, or even that hackers will alter the telemetry signals received by the central controller. This paper presents a framework for developing a safe plant controller by training it with an adversarial agent designed to confound it. This necessitates training the adversary to confound the controller, creating a sort of circular logic or \"Arms Race.\" This paper examines three broad training approaches for co-training the protagonist and adversary, finding that an Arms Race approach yields the best results. These initial results indicate that the Arms Race adversarial training reduced worst-case performance degradation from 39% power loss to 7.9% power gain relative to a baseline operational strategy.", "date": "2026-04-09", "categories": []}
{"id": "2604.09222", "title": "GRM: Utility-Aware Jailbreak Attacks on Audio LLMs via Gradient-Ratio Masking", "abstract": "Audio large language models (ALLMs) enable rich speech-text interaction, but they also introduce jailbreak vulnerabilities in the audio modality. Existing audio jailbreak methods mainly optimize jailbreak success while overlooking utility preservation, as reflected in transcription quality and question answering performance. In practice, stronger attacks often come at the cost of degraded utility. To study this trade-off, we revisit existing attacks by varying their perturbation coverage in the frequency domain, from partial-band to full-band, and find that broader frequency coverage does not necessarily improve jailbreak performance, while utility consistently deteriorates. This suggests that concentrating perturbation on a subset of bands can yield a better attack-utility trade-off than indiscriminate full-band coverage. Based on this insight, we propose GRM, a utility-aware frequency-selective jailbreak framework. It ranks Mel bands by their attack contribution relative to utility sensitivity, perturbs only a selected subset of bands, and learns a reusable universal perturbation under a semantic-preservation objective. Experiments on four representative ALLMs show that GRM achieves an average Jailbreak Success Rate (JSR) of 88.46% while providing a better attack-utility trade-off than representative baselines. These results highlight the potential of frequency-selective perturbation for better balancing attack effectiveness and utility preservation in audio jailbreak. Content Warning: This paper includes harmful query examples and unsafe model responses.", "date": "2026-04-10", "categories": []}
{"id": "2604.09253", "title": "Mosaic: Multimodal Jailbreak against Closed-Source VLMs via Multi-View Ensemble Optimization", "abstract": "Vision-Language Models (VLMs) are powerful but remain vulnerable to multimodal jailbreak attacks. Existing attacks mainly rely on either explicit visual prompt attacks or gradient-based adversarial optimization. While the former is easier to detect, the latter produces subtle perturbations that are less perceptible, but is usually optimized and evaluated under homogeneous open-source surrogate-target settings, leaving its effectiveness on commercial closed-source VLMs under heterogeneous settings unclear. To examine this issue, we study different surrogate-target settings and observe a consistent gap between homogeneous and heterogeneous settings, a phenomenon we term surrogate dependency. Motivated by this finding, we propose Mosaic, a Multi-view ensemble optimization framework for multimodal jailbreak against closed-source VLMs, which alleviates surrogate dependency under heterogeneous surrogate-target settings by reducing over-reliance on any single surrogate model and visual view. Specifically, Mosaic incorporates three core components: a Text-Side Transformation module, which perturbs refusal-sensitive lexical patterns; a Multi-View Image Optimization module, which updates perturbations under diverse cropped views to avoid overfitting to a single visual view; and a Surrogate Ensemble Guidance module, which aggregates optimization signals from multiple surrogate VLMs to reduce surrogate-specific bias. Extensive experiments on safety benchmarks demonstrate that Mosaic achieves state-of-the-art Attack Success Rate and Average Toxicity against commercial closed-source VLMs.", "date": "2026-04-10", "categories": []}
{"id": "2309.10348", "title": "Language Guided Adversarial Purification", "abstract": "Adversarial purification using generative models demonstrates strong adversarial defense performance. These methods are classifier and attack-agnostic, making them versatile but often computationally intensive. Recent strides in diffusion and score networks have improved image generation and, by extension, adversarial purification. Another highly efficient class of adversarial defense methods known as adversarial training requires specific knowledge of attack vectors, forcing them to be trained extensively on adversarial examples. To overcome these limitations, we introduce a new framework, namely Language Guided Adversarial Purification (LGAP), utilizing pre-trained diffusion models and caption generators to defend against adversarial attacks. Given an input image, our method first generates a caption, which is then used to guide the adversarial purification process through a diffusion network. Our approach has been evaluated against strong adversarial attacks, proving its effectiveness in enhancing adversarial robustness. Our results indicate that LGAP outperforms most existing adversarial defense techniques without requiring specialized network training. This underscores the generalizability of models trained on large datasets, highlighting a promising direction for further research.", "date": "2023-09-19", "categories": []}
{"id": "2506.02978", "title": "On the Robustness of Tabular Foundation Models: Test-Time Attacks and In-Context Defenses", "abstract": "Recent tabular Foundational Models (FM) such as TabPFN and TabICL, leverage in-context learning to achieve strong performance without gradient updates or fine-tuning. However, their robustness to adversarial manipulation remains largely unexplored. In this work, we present a comprehensive study of the adversarial vulnerabilities of tabular FM, focusing on both their fragility to targeted test-time attacks and their potential misuse as adversarial tools. We show on three benchmarks in finance, cybersecurity and healthcare, that small, structured perturbations to test inputs can significantly degrade prediction accuracy, even when training context remain fixed. Additionally, we demonstrate that tabular FM can be repurposed to generate transferable evasion to conventional models such as random forests and XGBoost, and on a lesser extent to deep tabular models. To improve tabular FM, we formulate the robustification problem as an optimization of the weights (adversarial fine-tuning), or the context (adversarial in-context learning). We introduce an in-context adversarial training strategy that incrementally replaces the context with adversarial perturbed instances, without updating model weights. Our approach improves robustness across multiple tabular benchmarks. Together, these findings position tabular FM as both a target and a source of adversarial threats, highlighting the urgent need for robust training and evaluation practices in this emerging paradigm.", "date": "2025-06-03", "categories": []}
{"id": "2506.08514", "title": "DiffGradCAM: A Class Activation Map Using the Full Model Decision to Solve Unaddressed Adversarial Attacks", "abstract": "Class Activation Mapping (CAM) and its gradient-based variants (e.g., GradCAM) have become standard tools for explaining Convolutional Neural Network (CNN) predictions. However, these approaches typically focus on individual logits, while for neural networks using softmax, the class membership probability estimates depend only on the differences between logits, not on their absolute values. This disconnect leaves standard CAMs vulnerable to adversarial manipulation, such as passive fooling, where a model is trained to produce misleading CAMs without affecting decision performance.\n  To address this vulnerability, we propose DiffGradCAM and its higher-order derivative version DiffGradCAM++, as novel, lightweight, contrastive approaches to class activation mapping that are not susceptible to passive fooling and match the output of standard methods such as GradCAM and GradCAM++ in the non-adversarial case. To test our claims, we introduce Salience-Hoax Activation Maps (SHAMs), a more advanced, entropy-aware form of passive fooling that serves as a benchmark for CAM robustness under adversarial conditions. Together, SHAM and DiffGradCAM establish a new framework for probing and improving the robustness of saliency-based explanations. We validate both contributions across multi-class tasks with few and many classes.", "date": "2025-06-10", "categories": []}
{"id": "2509.07673", "title": "Nearest Neighbor Projection Removal Adversarial Training", "abstract": "Deep neural networks have exhibited impressive performance in image classification tasks but remain vulnerable to adversarial examples. Standard adversarial training enhances robustness but typically fails to explicitly address inter-class feature overlap, a significant contributor to adversarial susceptibility. In this work, we introduce a novel adversarial training framework that actively mitigates inter-class proximity by projecting out inter-class dependencies from adversarial and clean samples in the feature space. Specifically, our approach first identifies the nearest inter-class neighbors for each adversarial sample and subsequently removes projections onto these neighbors to enforce stronger feature separability. Theoretically, we demonstrate that our proposed logits correction reduces the Lipschitz constant of neural networks, thereby lowering the Rademacher complexity, which directly contributes to improved generalization and robustness. Extensive experiments across standard benchmarks including CIFAR-10, CIFAR-100, SVHN, and TinyImagenet show that our method demonstrates strong performance that is competitive with leading adversarial training techniques, highlighting significant achievements in both robust and clean accuracy. Our findings reveal the importance of addressing inter-class feature proximity explicitly to bolster adversarial robustness in DNNs. The code is available in the supplementary material.", "date": "2025-09-09", "categories": []}
{"id": "2604.06950", "title": "Making MLLMs Blind: Adversarial Smuggling Attacks in MLLM Content Moderation", "abstract": "Multimodal Large Language Models (MLLMs) are increasingly being deployed as automated content moderators. Within this landscape, we uncover a critical threat: Adversarial Smuggling Attacks. Unlike adversarial perturbations (for misclassification) and adversarial jailbreaks (for harmful output generation), adversarial smuggling exploits the Human-AI capability gap. It encodes harmful content into human-readable visual formats that remain AI-unreadable, thereby evading automated detection and enabling the dissemination of harmful content. We classify smuggling attacks into two pathways: (1) Perceptual Blindness, disrupting text recognition; and (2) Reasoning Blockade, inhibiting semantic understanding despite successful text recognition. To evaluate this threat, we constructed SmuggleBench, the first comprehensive benchmark comprising 1,700 adversarial smuggling attack instances. Evaluations on SmuggleBench reveal that both proprietary (e.g., GPT-5) and open-source (e.g., Qwen3-VL) state-of-the-art models are vulnerable to this threat, producing Attack Success Rates (ASR) exceeding 90%. By analyzing the vulnerability through the lenses of perception and reasoning, we identify three root causes: the limited capabilities of vision encoders, the robustness gap in OCR, and the scarcity of domain-specific adversarial examples. We conduct a preliminary exploration of mitigation strategies, investigating the potential of test-time scaling (via CoT) and adversarial training (via SFT) to mitigate this threat. Our code is publicly available at https://github.com/zhihengli-casia/smugglebench.", "date": "2026-04-08", "categories": []}
{"id": "2604.07831", "title": "Are GUI Agents Focused Enough? Automated Distraction via Semantic-level UI Element Injection", "abstract": "Existing red-teaming studies on GUI agents have important limitations. Adversarial perturbations typically require white-box access, which is unavailable for commercial systems, while prompt injection is increasingly mitigated by stronger safety alignment. To study robustness under a more practical threat model, we propose Semantic-level UI Element Injection, a red-teaming setting that overlays safety-aligned and harmless UI elements onto screenshots to misdirect the agent's visual grounding. Our method uses a modular Editor-Overlapper-Victim pipeline and an iterative search procedure that samples multiple candidate edits, keeps the best cumulative overlay, and adapts future prompt strategies based on previous failures. Across five victim models, our optimized attacks improve attack success rate by up to 4.4x over random injection on the strongest victims. Moreover, elements optimized on one source model transfer effectively to other target models, indicating model-agnostic vulnerabilities. After the first successful attack, the victim still clicks the attacker-controlled element in more than 15% of later independent trials, versus below 1% for random injection, showing that the injected element acts as a persistent attractor rather than simple visual clutter.", "date": "2026-04-09", "categories": []}
{"id": "2604.08005", "title": "Preference Redirection via Attention Concentration: An Attack on Computer Use Agents", "abstract": "Advancements in multimodal foundation models have enabled the development of Computer Use Agents (CUAs) capable of autonomously interacting with GUI environments. As CUAs are not restricted to certain tools, they allow to automate more complex agentic tasks but at the same time open up new security vulnerabilities. While prior work has concentrated on the language modality, the vulnerability of the vision modality has received less attention. In this paper, we introduce PRAC, a novel attack that, unlike prior work targeting the VLM output directly, manipulates the model's internal preferences by redirecting its attention toward a stealthy adversarial patch. We show that PRAC is able to manipulate the selection process of a CUA on an online shopping platform towards a chosen target product. While we require white-box access to the model for the creation of the attack, we show that our attack generalizes to fine-tuned versions of the same model, presenting a critical threat as multiple companies build specific CUAs based on open weights models.", "date": "2026-04-09", "categories": []}
{"id": "2507.21540", "title": "PRISM: Programmatic Reasoning with Image Sequence Manipulation for LVLM Jailbreaking", "abstract": "The increasing sophistication of large vision-language models (LVLMs) has been accompanied by advances in safety alignment mechanisms designed to prevent harmful content generation. However, these defenses remain vulnerable to sophisticated adversarial attacks. Existing jailbreak methods typically rely on direct and semantically explicit prompts, overlooking subtle vulnerabilities in how LVLMs compose information over multiple reasoning steps. In this paper, we propose a novel and effective jailbreak framework inspired by Return-Oriented Programming (ROP) techniques from software security. Our approach decomposes a harmful instruction into a sequence of individually benign visual gadgets. A carefully engineered textual prompt directs the sequence of inputs, prompting the model to integrate the benign visual gadgets through its reasoning process to produce a coherent and harmful output. This makes the malicious intent emergent and difficult to detect from any single component. We validate our method through extensive experiments on established benchmarks including SafeBench and MM-SafetyBench, targeting popular LVLMs. Results show that our approach consistently and substantially outperforms existing baselines on state-of-the-art models, achieving near-perfect attack success rates (over 0.90 on SafeBench) and improving ASR by up to 0.39. Our findings reveal a critical and underexplored vulnerability that exploits the compositional reasoning abilities of LVLMs, highlighting the urgent need for defenses that secure the entire reasoning process.", "date": "2025-07-29", "categories": []}
{"id": "2601.15474", "title": "BadImplant: Injection-based Multi-Targeted Graph Backdoor Attack", "abstract": "Graph neural network (GNN) have demonstrated exceptional performance in solving critical problems across diverse domains yet remain susceptible to backdoor attacks. Existing studies on backdoor attack for graph classification are limited to single target attack using subgraph replacement based mechanism where the attacker implants only one trigger into the GNN model. In this paper, we introduce the first multi-targeted backdoor attack for graph classification task, where multiple triggers simultaneously redirect predictions to different target labels. Instead of subgraph replacement, we propose subgraph injection which preserves the structure of the original graphs while poisoning the clean graphs. Extensive experiments demonstrate the efficacy of our approach, where our attack achieves high attack success rates for all target labels with minimal impact on the clean accuracy. Experimental results on five dataset demonstrate the superior performance of our attack framework compared to the conventional subgraph replacement-based attack. Our analysis on four GNN models confirms the generalization capability of our attack which is effective regardless of the GNN model architectures and training parameters settings. We further investigate the impact of the attack design parameters including injection methods, number of connections, trigger sizes, trigger edge density and poisoning ratios. Additionally, our evaluation against state-of-the-art defenses (randomized smoothing and fine-pruning) demonstrates the robustness of our proposed multi-target attacks. This work highlights the GNN vulnerability against multi-targeted backdoor attack in graph classification task. Our source codes will be available at https://github.com/SiSL-URI/Multi-Targeted-Graph-Backdoor-Attack.", "date": "2026-01-21", "categories": []}
{"id": "2602.09987", "title": "Infusion: Shaping Model Behavior by Editing Training Data via Influence Functions", "abstract": "Influence functions are commonly used to attribute model behavior to training documents. We explore the reverse: crafting training data that induces model behavior. Our framework, Infusion, uses scalable influence-function approximations to compute small perturbations to training documents that induce targeted changes in model behavior through parameter shifts. We evaluate Infusion on data poisoning tasks across vision and language domains. On CIFAR-10, we show that making subtle edits via Infusion to just 0.2% (100/45,000) of the training documents can be competitive with the baseline of inserting a small number of explicit behavior examples. We also find that Infusion transfers across architectures (ResNet $\\leftrightarrow$ CNN), suggesting a single poisoned corpus can affect multiple independently trained models. In preliminary language experiments, we characterize when our approach increases the probability of target behaviors and when it fails, finding it most effective at amplifying behaviors the model has already learned. Taken together, these results show that small, subtle edits to training data can systematically shape model behavior, underscoring the importance of training data interpretability for adversaries and defenders alike. We provide the code here: https://github.com/jrosseruk/infusion.", "date": "2026-02-10", "categories": []}
{"id": "2603.13970", "title": "Shapes are not enough: CONSERVAttack and its use for finding vulnerabilities and uncertainties in machine learning applications", "abstract": "In High Energy Physics, as in many other fields of science, the application of machine learning techniques has been crucial in advancing our understanding of fundamental phenomena. Increasingly, deep learning models are applied to analyze both simulated and experimental data. In most experiments, a rigorous regime of testing for physically motivated systematic uncertainties is in place. The numerical evaluation of these tests for differences between the data on the one side and simulations on the other side quantifies the effect of potential sources of mismodelling on the machine learning output. In addition, thorough comparisons of marginal distributions and (linear) feature correlations between data and simulation in \"control regions\" are applied. However, the guidance by physical motivation, and the need to constrain comparisons to specific regions, does not guarantee that all possible sources of deviations have been accounted for. We therefore propose a new adversarial attack - the CONSERVAttack - designed to exploit the remaining space of hypothetical deviations between simulation and data after the above mentioned tests. The resulting adversarial perturbations are consistent within the uncertainty bounds - evading standard validation checks - while successfully fooling the underlying model. We further propose strategies to mitigate such vulnerabilities and argue that robustness to adversarial effects must be considered when interpreting results from deep learning in particle physics.", "date": "2026-03-14", "categories": []}
{"id": "2604.05853", "title": "Reading Between the Pixels: An Inscriptive Jailbreak Attack on Text-to-Image Models", "abstract": "Modern text-to-image (T2I) models can now render legible, paragraph-length text, enabling a fundamentally new class of misuse. We identify and formalize the inscriptive jailbreak, where an adversary coerces a T2I system into generating images containing harmful textual payloads (e.g., fraudulent documents) embedded within visually benign scenes. Unlike traditional depictive jailbreaks that elicit visually objectionable imagery, inscriptive attacks weaponize the text-rendering capability itself. Because existing jailbreak techniques are designed for coarse visual manipulation, they struggle to bypass multi-stage safety filters while maintaining character-level fidelity. To expose this vulnerability, we propose Etch, a black-box attack framework that decomposes the adversarial prompt into three functionally orthogonal layers: semantic camouflage, visual-spatial anchoring, and typographic encoding. This decomposition reduces joint optimization over the full prompt space to tractable sub-problems, which are iteratively refined through a zero-order loop. In this process, a vision-language model critiques each generated image, localizes failures to specific layers, and prescribes targeted revisions. Extensive evaluations across 7 models on the 2 benchmarks demonstrate that Etch achieves an average attack success rate of 65.57% (peaking at 91.00%), significantly outperforming existing baselines. Our results reveal a critical blind spot in current T2I safety alignments and underscore the urgent need for typography-aware defense multimodal mechanisms.", "date": "2026-04-07", "categories": []}
{"id": "2604.06289", "title": "Adversarial Robustness of Time-Series Classification for Crystal Collimator Alignment", "abstract": "In this paper, we analyze and improve the adversarial robustness of a convolutional neural network (CNN) that assists crystal-collimator alignment at CERN's Large Hadron Collider (LHC) by classifying a beam-loss monitor (BLM) time series during crystal rotation. We formalize a local robustness property for this classifier under an adversarial threat model based on real-world plausibility. Building on established parameterized input-transformation patterns used for transformation- and semantic-perturbation robustness, we instantiate a preprocessing-aware wrapper for our deployed time-series pipeline: we encode time-series normalization, padding constraints, and structured perturbations as a lightweight differentiable wrapper in front of the CNN, so that existing gradient-based robustness frameworks can operate on the deployed pipeline. For formal verification, data-dependent preprocessing such as per-window z-normalization introduces nonlinear operators that require verifier-specific abstractions. We therefore focus on attack-based robustness estimates and pipeline-checked validity by benchmarking robustness with the frameworks Foolbox and ART. Adversarial fine-tuning of the resulting CNN improves robust accuracy by up to 18.6 % without degrading clean accuracy. Finally, we extend robustness on time-series data beyond single windows to sequence-level robustness for sliding-window classification, introduce adversarial sequences as counterexamples to a temporal robustness requirement over full scans, and observe attack-induced misclassifications that persist across adjacent windows.", "date": "2026-04-07", "categories": []}
{"id": "2604.06865", "title": "Physical Adversarial Attacks on AI Surveillance Systems:Detection, Tracking, and Visible--Infrared Evasion", "abstract": "Physical adversarial attacks are increasingly studied in settings that resemble deployed surveillance systems rather than isolated image benchmarks. In these settings, person detection, multi-object tracking, visible--infrared sensing, and the practical form of the attack carrier all matter at once. This changes how the literature should be read. A perturbation that suppresses a detector in one frame may have limited practical effect if identity is recovered over time; an RGB-only result may say little about night-time systems that rely on visible and thermal inputs together; and a conspicuous patch can imply a different threat model from a wearable or selectively activated carrier. This paper reviews physical attacks from that surveillance-oriented viewpoint. Rather than attempting a complete catalogue of all physical attacks in computer vision, we focus on the technical questions that become central in surveillance: temporal persistence, sensing modality, carrier realism, and system-level objective. We organize prior work through a four-part taxonomy and discuss how recent results on multi-object tracking, dual-modal visible--infrared evasion, and controllable clothing reflect a broader change in the field. We also summarize evaluation practices and unresolved gaps, including distance robustness, camera-pipeline variation, identity-level metrics, and activation-aware testing. The resulting picture is that surveillance robustness cannot be judged reliably from isolated per-frame benchmarks alone; it has to be examined as a system problem unfolding over time, across sensors, and under realistic physical deployment constraints.", "date": "2026-04-08", "categories": []}
{"id": "2604.06954", "title": "Compression as an Adversarial Amplifier Through Decision Space Reduction", "abstract": "Image compression is a ubiquitous component of modern visual pipelines, routinely applied by social media platforms and resource-constrained systems prior to inference. Despite its prevalence, the impact of compression on adversarial robustness remains poorly understood. We study a previously unexplored adversarial setting in which attacks are applied directly in compressed representations, and show that compression can act as an adversarial amplifier for deep image classifiers. Under identical nominal perturbation budgets, compression-aware attacks are substantially more effective than their pixel-space counterparts. We attribute this effect to decision space reduction, whereby compression induces a non-invertible, information-losing transformation that contracts classification margins and increases sensitivity to perturbations. Extensive experiments across standard benchmarks and architectures support our analysis and reveal a critical vulnerability in compression-in-the-loop deployment settings. Code will be released.", "date": "2026-04-08", "categories": []}
{"id": "2604.06987", "title": "CAAP: Capture-Aware Adversarial Patch Attacks on Palmprint Recognition Models", "abstract": "Palmprint recognition is deployed in security-critical applications, including access control and palm-based payment, due to its contactless acquisition and highly discriminative ridge-and-crease textures. However, the robustness of deep palmprint recognition systems against physically realizable attacks remains insufficiently understood. Existing studies are largely confined to the digital setting and do not adequately account for the texture-dominant nature of palmprint recognition or the distortions introduced during physical acquisition. To address this gap, we propose CAAP, a capture-aware adversarial patch framework for palmprint recognition. CAAP learns a universal patch that can be reused across inputs while remaining effective under realistic acquisition variation. To match the structural characteristics of palmprints, the framework adopts a cross-shaped patch topology, which enlarges spatial coverage under a fixed pixel budget and more effectively disrupts long-range texture continuity. CAAP further integrates three modules: ASIT for input-conditioned patch rendering, RaS for stochastic capture-aware simulation, and MS-DIFE for feature-level identity-disruptive guidance. We evaluate CAAP on the Tongji, IITD, and AISEC datasets against generic CNN backbones and palmprint-specific recognition models. Experiments show that CAAP achieves strong untargeted and targeted attack performance with favorable cross-model and cross-dataset transferability. The results further show that, although adversarial training can partially reduce the attack success rate, substantial residual vulnerability remains. These findings indicate that deep palmprint recognition systems remain vulnerable to physically realizable, capture-aware adversarial patch attacks, underscoring the need for more effective defenses in practice. Code available at https://github.com/ryliu68/CAAP.", "date": "2026-04-08", "categories": []}
{"id": "2407.14971", "title": "Sim-CLIP: Unsupervised Siamese Adversarial Fine-Tuning for Robust and Semantically-Rich Vision-Language Models", "abstract": "Vision-Language Models (VLMs) rely heavily on pretrained vision encoders to support downstream tasks such as image captioning, visual question answering, and zero-shot classification. Despite their strong performance, these encoders remain highly vulnerable to imperceptible adversarial perturbations, which can severely degrade both robustness and semantic quality in multimodal reasoning. In this work, we introduce Sim-CLIP, an unsupervised adversarial fine-tuning framework that enhances the robustness of the CLIP vision encoder while preserving overall semantic representations. Sim-CLIP adopts a Siamese training architecture with a cosine similarity objective and a symmetric stop-gradient mechanism to enforce semantic alignment between clean and adversarial views. This design avoids large-batch contrastive learning and additional momentum encoders, enabling robust training with low computational overhead. We evaluate Sim-CLIP across multiple Vision-Language Models and tasks under both targeted and untargeted adversarial attacks. Experimental results demonstrate that Sim-CLIP consistently outperforms state-of-the-art robust CLIP variants, achieving stronger adversarial robustness while maintaining or improving semantic fidelity. These findings highlight the limitations of existing adversarial defenses and establish Sim-CLIP as an effective and scalable solution for robust vision-language representation learning.", "date": "2024-07-20", "categories": []}
{"id": "2602.15195", "title": "Weight space Detection of Backdoors in LoRA Adapters", "abstract": "LoRA adapters let users fine-tune large language models (LLMs) efficiently. However, LoRA adapters are shared through open repositories like Hugging Face Hub \\citep{huggingface_hub_docs}, making them vulnerable to backdoor attacks. Current detection methods require running the model with test input data -- making them impractical for screening thousands of adapters where the trigger for backdoor behavior is unknown. We detect poisoned adapters by analyzing their weight matrices directly, without running the model -- making our method trigger-agnostic. For each attention projection (Q, K, V, O), our method extracts five spectral statistics from the low-rank update $\\Delta W$, yielding a 20-dimensional signature for each adapter. A logistic regression detector trained on this representation separates benign and poisoned adapters across three model families -- Llama-3.2-3B~\\citep{llama3}, Qwen2.5-3B~\\citep{qwen25}, and Gemma-2-2B~\\citep{gemma2} -- on unseen test adapters drawn from instruction-following, reasoning, question-answering, code, and classification tasks. Across all three architectures, the detector achieves 100\\% accuracy.", "date": "2026-02-16", "categories": []}
{"id": "2504.18594", "title": "RaPA: Enhancing Transferable Targeted Attacks via Random Parameter Pruning", "abstract": "Compared to untargeted attacks, targeted transfer-based attack is still suffering from much lower Attack Success Rates (ASRs), although significant improvements have been achieved by kinds of methods, such as diversifying input, stabilizing the gradient, and re-training surrogate models. In this paper, we find that adversarial examples generated by existing methods rely heavily on a small subset of surrogate model parameters, which in turn limits their transferability to unseen target models. Inspired by this, we propose the Random Parameter Pruning Attack (RaPA), which introduces parameter-level randomization during the attack process. At each optimization step, RaPA randomly prunes model parameters to generate diverse yet semantically consistent surrogate variants.We show this parameter-level randomization is equivalent to adding an importance-equalization regularizer, thereby alleviating the over-reliance issue. Extensive experiments across both CNN and Transformer architectures demonstrate that RaPA substantially enhances transferability. In the challenging case of transferring from CNN-based to Transformer-based models, RaPA achieves up to 11.7% higher average ASRs than state-of-the-art baselines(with 33.3% ASRs), while being training-free, cross-architecture efficient, and easily integrated into existing attack frameworks. Code is available in https://github.com/molarsu/RaPA.", "date": "2025-04-24", "categories": []}
{"id": "2505.12167", "title": "FABLE: A Localized, Targeted Adversarial Attack on Weather Forecasting Models", "abstract": "Deep learning-based weather forecasting (DLWF) models have recently demonstrated significant performance gains over gold-standard physics-based simulation tools. However, these models are potentially vulnerable to adversarial attacks, which raises concerns about their trustworthiness. In this paper, we investigate the feasibility and challenges of applying existing adversarial attack methods to DLWF models and propose a novel framework called FABLE (Forecast Alteration By Localized targeted advErsarial attack) to address them. FABLE performs a 3D discrete wavelet decomposition to disentangle the spatial and temporal components of the data. By regulating the magnitude of adversarial perturbations across different components, FABLE produces adversarial inputs that remain closely aligned with the original inputs while steering the DLWF models toward generating the targeted forecast outcomes. Experimental results on real-world weather datasets demonstrate the effectiveness of FABLE over baseline methods across various metrics.", "date": "2025-05-17", "categories": []}
{"id": "2509.23279", "title": "Vid-Freeze: Protecting Images from Malicious Image-to-Video Generation via Temporal Freezing", "abstract": "The rapid progress of image-to-video (I2V) generation models has introduced significant risks by enabling deceptive or malicious video synthesis from a single image. Prior defenses such as I2VGuard attempt to immunize images by inducing spatio-temporal degradation, which does not necessarily provide meaningful protection, since residual motion can still convey malicious intent. In this work, we introduce Vid-Freeze -- a novel adversarial defense that adds imperceptible perturbations to enforce temporal freezing in generated videos. Our method explicitly targets attention dynamics in I2V models to suppress motion synthesis. As a result, immunized images produce standstill or near-static videos, effectively blocking malicious content generation. Experiments demonstrate strong protection across models and support temporal freezing as a promising direction for proactive and meaningful defense against I2V misuse.", "date": "2025-09-27", "categories": []}
{"id": "2511.17362", "title": "ATAC: Augmentation-Based Test-Time Adversarial Correction for CLIP", "abstract": "Despite its remarkable success in zero-shot image-text matching, CLIP remains highly vulnerable to adversarial perturbations on images. As adversarial fine-tuning is prohibitively costly, recent works explore various test-time defense strategies; however, these approaches still exhibit limited robustness. In this work, we revisit this problem and propose a simple yet effective strategy: Augmentation-based Test-time Adversarial Correction (ATAC). Our method operates directly in the embedding space of CLIP, calculating augmentation-induced drift vectors to infer a semantic recovery direction and correcting the embedding based on the angular consistency of these latent drifts. Across a wide range of benchmarks, ATAC consistently achieves remarkably high robustness, surpassing that of previous state-of-the-art methods by nearly 50\\% on average, all while requiring minimal computational overhead. Furthermore, ATAC retains state-of-the-art robustness in unconventional and extreme settings and even achieves nontrivial robustness against adaptive attacks. Our results demonstrate that ATAC is an efficient method in a novel paradigm for test-time adversarial defenses in the embedding space of CLIP. Code is available at: https://github.com/kylin0421/ATAC", "date": "2025-11-21", "categories": []}
{"id": "2604.03427", "title": "Adversarial Robustness of Deep State Space Models for Forecasting", "abstract": "State-space model (SSM) for time-series forecasting have demonstrated strong empirical performance on benchmark datasets, yet their robustness under adversarial perturbations is poorly understood. We address this gap through a control-theoretic lens, focusing on the recently proposed Spacetime SSM forecaster. We first establish that the decoder-only Spacetime architecture can represent the optimal Kalman predictor when the underlying data-generating process is autoregressive - a property no other SSM possesses. Building on this, we formulate robust forecaster design as a Stackelberg game against worst-case stealthy adversaries constrained by a detection budget, and solve it via adversarial training. We derive closed-form bounds on adversarial forecasting error that expose how open-loop instability, closed-loop instability, and decoder state dimension each amplify vulnerability - offering actionable principles towards robust forecaster design. Finally, we show that even adversaries with no access to the forecaster can nonetheless construct effective attacks by exploiting the model's locally linear input-output behavior, bypassing gradient computations entirely. Experiments on the Monash benchmark datasets highlight that model-free attacks, without any gradient computation, can cause at least 33% more error than projected gradient descent with a small step size.", "date": "2026-04-03", "categories": []}
{"id": "2604.03843", "title": "Explainability-Guided Adversarial Attacks on Transformer-Based Malware Detectors Using Control Flow Graphs", "abstract": "Transformer-based malware detection systems operating on graph modalities such as control flow graphs (CFGs) achieve strong performance by modeling structural relationships in program behavior. However, their robustness to adversarial evasion attacks remains underexplored. This paper examines the vulnerability of a RoBERTa-based malware detector that linearizes CFGs into sequences of function calls, a design choice that enables transformer modeling but may introduce token-level sensitivities and ordering artifacts exploitable by adversaries. By evaluating evasion strategies within this graph-to-sequence framework, we provide insight into the practical robustness of transformer-based malware detectors beyond aggregate detection accuracy.\n  This paper proposes a white-box adversarial evasion attack that leverages explainability mechanisms to identify and perturb most influential graph components. Using token- and word-level attributions derived from integrated gradients, the attack iteratively replaces positively attributed function calls with synthetic external imports, producing adversarial CFG representations without altering overall program structure. Experimental evaluation on small- and large-scale Windows Portable Executable (PE) datasets demonstrates that the proposed method can reliably induce misclassification, even against models trained to high accuracy. Our results highlight that explainability tools, while valuable for interpretability, can also expose critical attack surfaces in transformer-based malware detectors.", "date": "2026-04-04", "categories": []}
{"id": "2604.03995", "title": "A Systematic Study of Cross-Modal Typographic Attacks on Audio-Visual Reasoning", "abstract": "As audio-visual multi-modal large language models (MLLMs) are increasingly deployed in safety-critical applications, understanding their vulnerabilities is crucial. To this end, we introduce Multi-Modal Typography, a systematic study examining how typographic attacks across multiple modalities adversely influence MLLMs. While prior work focuses narrowly on unimodal attacks, we expose the cross-modal fragility of MLLMs. We analyze the interactions between audio, visual, and text perturbations and reveal that coordinated multi-modal attack creates a significantly more potent threat than single-modality attacks (attack success rate = $83.43\\%$ vs $34.93\\%$).Our findings across multiple frontier MLLMs, tasks, and common-sense reasoning and content moderation benchmarks establishes multi-modal typography as a critical and underexplored attack strategy in multi-modal reasoning. Code and data will be publicly available.", "date": "2026-04-05", "categories": []}
{"id": "2604.04060", "title": "CoopGuard: Stateful Cooperative Agents Safeguarding LLMs Against Evolving Multi-Round Attacks", "abstract": "As Large Language Models (LLMs) are increasingly deployed in complex applications, their vulnerability to adversarial attacks raises urgent safety concerns, especially those evolving over multi-round interactions. Existing defenses are largely reactive and struggle to adapt as adversaries refine strategies across rounds. In this work, we propose CoopGuard , a stateful multi-round LLM defense framework based on cooperative agents that maintains and updates an internal defense state to counter evolving attacks. It employs three specialized agents (Deferring Agent, Tempting Agent, and Forensic Agent) for complementary round-level strategies, coordinated by System Agent, which conditions decisions on the evolving defense state (interaction history) and orchestrates agents over time. To evaluate evolving threats, we introduce the EMRA benchmark with 5,200 adversarial samples across 8 attack types, simulating progressively LLM multi-round attacks. Experiments show that CoopGuard reduces attack success rate by 78.9% over state-of-the-art defenses, while improving deceptive rate by 186% and reducing attack efficiency by 167.9%, offering a more comprehensive assessment of multi-round defense. These results demonstrate that CoopGuard provides robust protection for LLMs in multi-round adversarial scenarios.", "date": "2026-04-05", "categories": []}
{"id": "2604.04349", "title": "Adversarial Robustness Analysis of Cloud-Assisted Autonomous Driving Systems", "abstract": "Autonomous vehicles increasingly rely on deep learning-based perception and control, which impose substantial computational demands. Cloud-assisted architectures offload these functions to remote servers, enabling enhanced perception and coordinated decision-making through the Internet of Vehicles (IoV). However, this paradigm introduces cross-layer vulnerabilities, where adversarial manipulation of perception models and network impairments in the vehicle-cloud link can jointly undermine safety-critical autonomy. This paper presents a hardware-in-the-loop IoV testbed that integrates real-time perception, control, and communication to evaluate such vulnerabilities in cloud-assisted autonomous driving. A YOLOv8-based object detector deployed on the cloud is subjected to whitebox adversarial attacks using the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD), while network adversaries induce delay and packet loss in the vehicle-cloud loop. Results show that adversarial perturbations significantly degrade perception performance, with PGD reducing detection precision and recall from 0.73 and 0.68 in the clean baseline to 0.22 and 0.15 at epsilon= 0.04. Network delays of 150-250 ms, corresponding to transient losses of approximately 3-4 frames, and packet loss rates of 0.5-5 % further destabilize closed-loop control, leading to delayed actuation and rule violations. These findings highlight the need for cross-layer resilience in cloud-assisted autonomous driving systems.", "date": "2026-04-06", "categories": []}
{"id": "2604.04630", "title": "Multimodal Backdoor Attack on VLMs for Autonomous Driving via Graffiti and Cross-Lingual Triggers", "abstract": "Visual language model (VLM) is rapidly being integrated into safety-critical systems such as autonomous driving, making it an important attack surface for potential backdoor attacks. Existing backdoor attacks mainly rely on unimodal, explicit, and easily detectable triggers, making it difficult to construct both covert and stable attack channels in autonomous driving scenarios. GLA introduces two naturalistic triggers: graffiti-based visual patterns generated via stable diffusion inpainting, which seamlessly blend into urban scenes, and cross-language text triggers, which introduce distributional shifts while maintaining semantic consistency to build robust language-side trigger signals. Experiments on DriveVLM show that GLA requires only a 10\\% poisoning ratio to achieve a 90\\% Attack Success Rate (ASR) and a 0\\% False Positive Rate (FPR). More insidiously, the backdoor does not weaken the model on clean tasks, but instead improves metrics such as BLEU-1, making it difficult for traditional performance-degradation-based detection methods to identify the attack. This study reveals underestimated security threats in self-driving VLMs and provides a new attack paradigm for backdoor evaluation in safety-critical multimodal systems.", "date": "2026-04-06", "categories": []}
{"id": "2410.13891", "title": "S$^4$ST: A Strong, Self-transferable, faSt, and Simple Scale Transformation for Transferable Targeted Attack", "abstract": "Transferable Targeted Attacks (TTAs) face significant challenges due to severe overfitting to surrogate models. Recent breakthroughs heavily rely on large-scale training data of victim models, while data-free solutions, \\textit{i.e.}, image transformation-involved gradient optimization, often depend on black-box feedback for method design and tuning. These dependencies violate black-box transfer settings and compromise threat evaluation fairness. In this paper, we propose two blind estimation measures, self-alignment and self-transferability, to analyze per-transformation effectiveness and cross-transformation correlations under strict black-box constraints. Our findings challenge conventional assumptions: (1) Attacking simple scaling transformations uniquely enhances targeted transferability, outperforming other basic transformations and rivaling leading complex methods; (2) Geometric and color transformations exhibit high internal redundancy despite weak inter-category correlations. These insights drive the design and tuning of S$^4$ST (Strong, Self-transferable, faSt, Simple Scale Transformation), which integrates dimensionally consistent scaling, complementary low-redundancy transformations, and block-wise operations. Extensive evaluations across diverse architectures, training distributions, and tasks show that S$^{4}$ST achieves state-of-the-art effectiveness-efficiency balance without data dependency. We reveal that scaling's effectiveness stems from visual data's multi-scale nature and ubiquitous scale augmentation during training, rendering such augmentation a double-edged sword. Further validations on medical imaging and face verification confirm the framework's strong generalization.", "date": "2024-10-13", "categories": []}
{"id": "2604.02686", "title": "Beyond Semantic Manipulation: Token-Space Attacks on Reward Models", "abstract": "Reward models (RMs) are widely used as optimization targets in reinforcement learning from human feedback (RLHF), yet they remain vulnerable to reward hacking. Existing attacks mainly operate within the semantic space, constructing human-readable adversarial outputs that exploit RM biases. In this work, we introduce a fundamentally different paradigm: Token Mapping Perturbation Attack (TOMPA), a framework that performs adversarial optimization directly in token space. By bypassing the standard decode-re-tokenize interface between the policy and the reward model, TOMPA enables the attack policy to optimize over raw token sequences rather than coherent natural language. Using only black-box scalar feedback, TOMPA automatically discovers non-linguistic token patterns that elicit extremely high rewards across multiple state-of-the-art RMs. Specifically, when targeting Skywork-Reward-V2-Llama-3.1-8B, TOMPA nearly doubles the reward of GPT-5 reference answers and outperforms them on 98.0% of prompts. Despite these high scores, the generated outputs degenerate into nonsensical text, revealing that RMs can be systematically exploited beyond the semantic regime and exposing a critical vulnerability in current RLHF pipelines.", "date": "2026-04-03", "categories": []}
{"id": "2604.02780", "title": "A Unified Perspective on Adversarial Membership Manipulation in Vision Models", "abstract": "Membership inference attacks (MIAs) aim to determine whether a specific data point was part of a model's training set, serving as effective tools for evaluating privacy leakage of vision models. However, existing MIAs implicitly assume honest query inputs, and their adversarial robustness remains unexplored. We show that MIAs for vision models expose a previously overlooked adversarial surface: adversarial membership manipulation, where imperceptible perturbations can reliably push non-member images into the \"member\" region of state-of-the-art MIAs. In this paper, we provide the first unified perspective on this phenomenon by analyzing its mechanism and implications. We begin by demonstrating that adversarial membership fabrication is consistently effective across diverse architectures and datasets. We then reveal a distinctive geometric signature - a characteristic gradient-norm collapse trajectory - that reliably separates fabricated from true members despite their nearly identical semantic representations. Building on this insight, we introduce a principled detection strategy grounded in gradient-geometry signals and develop a robust inference framework that substantially mitigates adversarial manipulation. Extensive experiments show that fabrication is broadly effective, while our detection and robust inference strategies significantly enhance resilience. This work establishes the first comprehensive framework for adversarial membership manipulation in vision models.", "date": "2026-04-03", "categories": []}
{"id": "2604.03117", "title": "Revealing Physical-World Semantic Vulnerabilities: Universal Adversarial Patches for Infrared Vision-Language Models", "abstract": "Infrared vision-language models (IR-VLMs) have emerged as a promising paradigm for multimodal perception in low-visibility environments, yet their robustness to adversarial attacks remains largely unexplored. Existing adversarial patch methods are mainly designed for RGB-based models in closed-set settings and are not readily applicable to the open-ended semantic understanding and physical deployment requirements of infrared VLMs. To bridge this gap, we propose Universal Curved-Grid Patch (UCGP), a universal physical adversarial patch framework for IR-VLMs. UCGP integrates Curved-Grid Mesh (CGM) parameterization for continuous, low-frequency, and deployable patch generation with a unified representation-driven objective that promotes subspace departure, topology disruption, and stealth. To improve robustness under real-world deployment and domain shift, we further incorporate Meta Differential Evolution and EOT-augmented TPS deformation modeling. Rather than manipulating labels or prompts, UCGP directly disrupts the visual representation space, weakening cross-modal semantic alignment. Extensive experiments demonstrate that UCGP consistently compromises semantic understanding across diverse IR-VLM architectures while maintaining cross-model transferability, cross-dataset generalization, real-world physical effectiveness, and robustness against defenses. These findings reveal a previously overlooked robustness vulnerability in current infrared multimodal systems.", "date": "2026-04-03", "categories": []}
{"id": "2508.02530", "title": "Understanding the Risks of Asphalt Art to the Reliability of Vision-Based Perception Systems", "abstract": "Artistic crosswalks featuring asphalt art, introduced by different organizations in recent years, aim to enhance the visibility and safety of pedestrians. However, their visual complexity may interfere with surveillance systems that rely on vision-based object detection models. In this study, we investigate the impact of asphalt art on pedestrian detection performance of a pretrained vision-based object detection model. We construct realistic crosswalk scenarios by compositing various street art patterns into a fixed surveillance scene and evaluate the model's performance in detecting pedestrians on asphalt-arted crosswalks under both benign and adversarial conditions. A benign case refers to pedestrian crosswalks painted with existing normal asphalt art, whereas an adversarial case involves digitally crafted or altered asphalt art perpetrated by an attacker. Our results show that while simple, color-based designs have minimal effect, complex artistic patterns, particularly those with high visual salience, can significantly degrade pedestrian detection performance. Furthermore, we demonstrate that adversarially crafted asphalt art can be exploited to deliberately obscure real pedestrians or generate non-existent pedestrian detections. These findings highlight a potential vulnerability in urban vision-based pedestrian surveillance systems, and underscore the importance of accounting for environmental visual variations when designing robust pedestrian perception models.", "date": "2025-08-04", "categories": []}
{"id": "2510.00463", "title": "On the Adversarial Robustness of Learning-based Conformal Novelty Detection", "abstract": "This paper studies the adversarial robustness of conformal novelty detection. In particular, we focus on two powerful learning-based frameworks that come with finite-sample false discovery rate (FDR) control: one is AdaDetect (by Marandon et al., 2024) that is based on the positive-unlabeled classifier, and the other is a one-class classifier-based approach (by Bates et al., 2023). While they provide rigorous statistical guarantees under benign conditions, their behavior under adversarial perturbations remains underexplored. We first formulate an oracle attack setup, under the AdaDetect formulation, that quantifies the worst-case degradation of FDR, deriving an upper bound that characterizes the statistical cost of attacks. This idealized formulation directly motivates a practical and effective attack scheme that only requires query access to the output labels of both frameworks. Coupling these formulations with two popular and complementary black-box adversarial algorithms, we systematically evaluate the vulnerability of both frameworks on synthetic and real-world datasets. Our results show that adversarial perturbations can significantly increase the FDR while maintaining high detection power, exposing fundamental limitations of current error-controlled novelty detection methods and motivating the development of more robust alternatives.", "date": "2025-10-01", "categories": []}
{"id": "2511.20456", "title": "Towards Trustworthy Wi-Fi CSI-based Sensing: Systematic Evaluation of Adversarial Robustness", "abstract": "Machine learning drives Channel State Information (CSI)-based human sensing in modern wireless networks, enabling applications like device-free human activity recognition (HAR) and identification (HID). However, the susceptibility of these models to adversarial perturbations raises security concerns that must be quantified prior to edge deployment.\n  We present a systematic robustness evaluation of five diverse CSI architectures across four public datasets, jointly analyzing white-box, black-box transfer, and universal attacks, together with defense strategies, under unconstrained and physics-guided perturbation boundaries. Contrary to prior assumptions, our experiments reveal that model capacity does not guarantee robustness; simple architectures consistently exhibit superior resilience compared to high-capacity sequence and vision models. Furthermore, vulnerability is fundamentally task-dependent, with HAR proving highly susceptible to attack, while HID demonstrates stark inherent resistance. Crucially, enforcing physical signal constraints drastically reduces attack success rates and significantly taxes attacker computation, showing that standard unconstrained feature-space attacks substantially overestimate real-world Over-The-Air vulnerabilities. By synthesizing attack, defense, and security metrics with strict edge hardware considerations, this work establishes foundational design principles for secure, deployable, and physically realizable wireless sensing systems.", "date": "2025-11-25", "categories": []}
{"id": "2603.20025", "title": "Graph-Informed Adversarial Modeling: Infimal Subadditivity of Interpolative Divergences", "abstract": "We study adversarial learning when the target distribution factorizes according to a known Bayesian network. For interpolative divergences, including $(f,\\Gamma)$-divergences, we prove a new infimal subadditivity principle showing that, under suitable conditions, a global variational discrepancy is controlled by an average of family-level discrepancies aligned with the graph. In an additive regime, the surrogate is exact. This closes a theoretical gap in the literature; existing subadditivity results justify graph-informed adversarial learning for classical discrepancies, but not for interpolative divergences, where the usual factorization argument breaks down. In turn, we provide a justification for replacing a standard, graph-agnostic GAN with a monolithic discriminator by a graph-informed GAN (GiGAN) with localized family-level discriminators, without requiring the optimizer itself to factorize according to the graph. We also obtain parallel results for integral probability metrics and proximal optimal transport divergences, identify natural discriminator classes for which the theory applies, and present experiments showing improved stability and structural recovery relative to graph-agnostic baselines.", "date": "2026-03-20", "categories": []}
{"id": "2604.01618", "title": "Tex3D: Objects as Attack Surfaces via Adversarial 3D Textures for Vision-Language-Action Models", "abstract": "Vision-language-action (VLA) models have shown strong performance in robotic manipulation, yet their robustness to physically realizable adversarial attacks remains underexplored. Existing studies reveal vulnerabilities through language perturbations and 2D visual attacks, but these attack surfaces are either less representative of real deployment or limited in physical realism. In contrast, adversarial 3D textures pose a more physically plausible and damaging threat, as they are naturally attached to manipulated objects and are easier to deploy in physical environments. Bringing adversarial 3D textures to VLA systems is nevertheless nontrivial. A central obstacle is that standard 3D simulators do not provide a differentiable optimization path from the VLA objective function back to object appearance, making it difficult to optimize through an end-to-end manner. To address this, we introduce Foreground-Background Decoupling (FBD), which enables differentiable texture optimization through dual-renderer alignment while preserving the original simulation environment. To further ensure that the attack remains effective across long-horizon and diverse viewpoints in the physical world, we propose Trajectory-Aware Adversarial Optimization (TAAO), which prioritizes behaviorally critical frames and stabilizes optimization with a vertex-based parameterization. Built on these designs, we present Tex3D, the first framework for end-to-end optimization of 3D adversarial textures directly within the VLA simulation environment. Experiments in both simulation and real-robot settings show that Tex3D significantly degrades VLA performance across multiple manipulation tasks, achieving task failure rates of up to 96.7\\%. Our empirical results expose critical vulnerabilities of VLA systems to physically grounded 3D adversarial attacks and highlight the need for robustness-aware training.", "date": "2026-04-02", "categories": []}
{"id": "2604.01888", "title": "Low-Effort Jailbreak Attacks Against Text-to-Image Safety Filters", "abstract": "Text-to-image generative models are widely deployed in creative tools and online platforms. To mitigate misuse, these systems rely on safety filters and moderation pipelines that aim to block harmful or policy violating content. In this work we show that modern text-to-image models remain vulnerable to low-effort jailbreak attacks that require only natural language prompts. We present a systematic study of prompt-based strategies that bypass safety filters without model access, optimization, or adversarial training. We introduce a taxonomy of visual jailbreak techniques including artistic reframing, material substitution, pseudo-educational framing, lifestyle aesthetic camouflage, and ambiguous action substitution. These strategies exploit weaknesses in prompt moderation and visual safety filtering by masking unsafe intent within benign semantic contexts. We evaluate these attacks across several state-of-the-art text-to-image systems and demonstrate that simple linguistic modifications can reliably evade existing safeguards and produce restricted imagery. Our findings highlight a critical gap between surface-level prompt filtering and the semantic understanding required to detect adversarial intent in generative media systems. Across all tested models and attack categories we observe an attack success rate (ASR) of up to 74.47%.", "date": "2026-04-02", "categories": []}
{"id": "2510.18314", "title": "Genesis: Evolving Attack Strategies for LLM Web Agent Red-Teaming", "abstract": "As large language model (LLM) agents increasingly automate complex web tasks, they boost productivity while simultaneously introducing new security risks. However, relevant studies on web agent attacks remain limited. Existing red-teaming approaches mainly rely on manually crafted attack strategies or static models trained offline. Such methods fail to capture the underlying behavioral patterns of web agents, making it difficult to generalize across diverse environments. In web agent attacks, success requires the continuous discovery and evolution of attack strategies. To this end, we propose Genesis, a novel agentic framework composed of three modules: Attacker, Scorer, and Strategist. The Attacker generates adversarial injections by integrating the genetic algorithm with a hybrid strategy representation. The Scorer evaluates the target web agent's responses to provide feedback. The Strategist dynamically uncovers effective strategies from interaction logs and compiles them into a continuously growing strategy library, which is then re-deployed to enhance the Attacker's effectiveness. Extensive experiments across various web tasks show that our framework discovers novel strategies and consistently outperforms existing attack baselines. Our code is available at https://github.com/CjangCjengh/web_agent_attack.", "date": "2025-10-21", "categories": []}
{"id": "2604.00371", "title": "Neural Reconstruction of LiDAR Point Clouds under Jamming Attacks via Full-Waveform Representation and Simultaneous Laser Sensing", "abstract": "LiDAR sensors are critical for autonomous driving perception, yet remain vulnerable to spoofing attacks. Jamming attacks inject high-frequency laser pulses that completely blind LiDAR sensors by overwhelming authentic returns with malicious signals. We discover that while point clouds become randomized, the underlying full-waveform data retains distinguishable signatures between attack and legitimate signals. In this work, we propose PULSAR-Net, capable of reconstructing authentic point clouds under jamming attacks by leveraging previously underutilized intermediate full-waveform representations and simultaneous laser sensing in modern LiDAR systems. PULSAR-Net adopts a novel U-Net architecture with axial spatial attention mechanisms specifically designed to identify attack-induced signals from authentic object returns in the full-waveform representation. To address the lack of full-waveform representations in existing LiDAR datasets under jamming attacks, we introduce a physics-aware dataset generation pipeline that synthesizes realistic full-waveform representations under jamming attacks. Despite being trained exclusively on synthetic data, PULSAR-Net achieves reconstruction rates of 92% and 73% for vehicles obscured by jamming attacks in real-world static and driving scenarios, respectively.", "date": "2026-04-01", "categories": []}
{"id": "2604.00452", "title": "Out of Sight, Out of Track: Adversarial Attacks on Propagation-based Multi-Object Trackers via Query State Manipulation", "abstract": "Recent Tracking-by-Query-Propagation (TBP) methods have advanced Multi-Object Tracking (MOT) by enabling end-to-end (E2E) pipelines with long-range temporal modeling. However, this reliance on query propagation introduces unexplored architectural vulnerabilities to adversarial attacks. We present FADE, a novel attack framework designed to exploit these specific vulnerabilities. FADE employs two attack strategies targeting core TBP mechanisms: (i) Temporal Query Flooding: Generates spurious temporally consistent track queries to exhaust the tracker's limited query budget, forcing it to terminate valid tracks. (ii) Temporal Memory Corruption: Directly attacks the query updater's memory by severing temporal links via state de-correlation and erasing the learned feature identity of matched tracks. Furthermore, we introduce a differentiable pipeline to optimize these attacks for physical-world realizability by leveraging simulations of advanced perception sensor spoofing. Experiments on MOT17 and MOT20 benchmarks demonstrate that FADE is highly effective against state-of-the-art TBP trackers, causing significant identity switches and track terminations.", "date": "2026-04-01", "categories": []}
{"id": "2604.00770", "title": "Thinking Wrong in Silence: Backdoor Attacks on Continuous Latent Reasoning", "abstract": "A new generation of language models reasons entirely in continuous hidden states, producing no tokens and leaving\n  no audit trail. We show that this silence creates a fundamentally new attack surface. ThoughtSteer perturbs a\n  single embedding vector at the input layer; the model's own multi-pass reasoning amplifies this perturbation into a\n  hijacked latent trajectory that reliably produces the attacker's chosen answer, while remaining structurally\n  invisible to every token-level defense. Across two architectures (Coconut and SimCoT), three reasoning benchmarks,\n  and model scales from 124M to 3B parameters, ThoughtSteer achieves >=99% attack success rate with near-baseline\n  clean accuracy, transfers to held-out benchmarks without retraining (94-100%), evades all five evaluated active\n  defenses, and survives 25 epochs of clean fine-tuning. We trace these results to a unifying mechanism: Neural\n  Collapse in the latent space pulls triggered representations onto a tight geometric attractor, explaining both why\n  defenses fail and why any effective backdoor must leave a linearly separable signature (probe AUC>=0.999). Yet a\n  striking paradox emerges: individual latent vectors still encode the correct answer even as the model outputs the\n  wrong one. The adversarial information is not in any single vector but in the collective trajectory, establishing\n  backdoor perturbations as a new lens for mechanistic interpretability of continuous reasoning. Code and checkpoints\n  are available.", "date": "2026-04-01", "categories": []}
{"id": "2604.00938", "title": "WARP: Guaranteed Inner-Layer Repair of NLP Transformers", "abstract": "Transformer-based NLP models remain vulnerable to adversarial perturbations, yet existing repair methods face a fundamental trade-off: gradient-based approaches offer flexibility but lack verifiability and often overfit; methods that do provide repair guarantees are restricted to the final layer or small networks, significantly limiting the parameter search space available for repair. We present WARP (Weight-Adjusted Repair with Provability), a constraint-based repair framework that extends repair beyond the last layer of Transformer models. WARP formulates repair as a convex quadratic program derived from a first-order linearization of the logit gap, enabling tractable optimization over a high-dimensional parameter space. Under the condition that the first-order approximation holds, this formulation induces three per-sample guarantees: (i) a positive margin constraint ensuring correct classification on repaired inputs, (ii) preservation constraints over a designated remain set, and (iii) a certified robustness radius derived from Lipschitz continuity. To ensure feasibility across varying model architectures, we introduce a sensitivity-based preprocessing step that conditions the optimization landscape accordingly. We further show that the iterative optimization procedure converges to solutions satisfying all repair constraints under mild assumptions. Empirical evaluation on encoder-only Transformers with varying layer architectures validates that these guarantees hold in practice while improving robustness to adversarial inputs. Our results demonstrate that guaranteed, generalizable Transformer repair is achievable through principled constraint-based optimization.", "date": "2026-04-01", "categories": []}
{"id": "2604.01049", "title": "Adversarial Attacks in AI-Driven RAN Slicing: SLA Violations and Recovery", "abstract": "Next-generation (NextG) cellular networks are designed to support emerging applications with diverse data rate and latency requirements, such as immersive multimedia services and large-scale Internet of Things deployments. A key enabling mechanism is radio access network (RAN) slicing, which dynamically partitions radio resources into virtual resource blocks to efficiently serve heterogeneous traffic classes, including enhanced mobile broadband (eMBB), massive machine-type communications (mMTC), and ultra-reliable low-latency communications (URLLC). In this paper, we study the impact of adversarial attacks on AI-driven RAN slicing decisions, where a budget-constrained adversary selectively jams slice transmissions to bias deep reinforcement learning (DRL)-based resource allocation, and quantify the resulting service level agreement (SLA) violations and post-attack recovery behavior. Our results indicate that budget-constrained adversarial jamming can induce severe and slice-dependent steady-state SLA violations. Moreover, the DRL agent's reward converges toward the clean baseline only after a non-negligible recovery period.", "date": "2026-04-01", "categories": []}
{"id": "2603.26984", "title": "A Provable Energy-Guided Test-Time Defense Boosting Adversarial Robustness of Large Vision-Language Models", "abstract": "Despite the rapid progress in multimodal models and Large Visual-Language Models (LVLM), they remain highly susceptible to adversarial perturbations, raising serious concerns about their reliability in real-world use. While adversarial training has become the leading paradigm for building models that are robust to adversarial attacks, Test-Time Transformations (TTT) have emerged as a promising strategy to boost robustness at inference. In light of this, we propose Energy-Guided Test-Time Transformation (ET3), a lightweight, training-free defense that enhances the robustness by minimizing the energy of the input samples. Our method is grounded in a theory that proves our transformation succeeds in classification under reasonable assumptions. We present extensive experiments demonstrating that ET3 provides a strong defense for classifiers, zero-shot classification with CLIP, and also for boosting the robustness of LVLMs in tasks such as Image Captioning and Visual Question Answering. Code is available at github.com/OmnAI-Lab/Energy-Guided-Test-Time-Defense .", "date": "2026-03-27", "categories": []}
{"id": "2603.28594", "title": "Detection of Adversarial Attacks in Robotic Perception", "abstract": "Deep Neural Networks (DNNs) achieve strong performance in semantic segmentation for robotic perception but remain vulnerable to adversarial attacks, threatening safety-critical applications. While robustness has been studied for image classification, semantic segmentation in robotic contexts requires specialized architectures and detection strategies.", "date": "2026-03-30", "categories": []}
{"id": "2603.28824", "title": "SNEAKDOOR: Stealthy Backdoor Attacks against Distribution Matching-based Dataset Condensation", "abstract": "Dataset condensation aims to synthesize compact yet informative datasets that retain the training efficacy of full-scale data, offering substantial gains in efficiency. Recent studies reveal that the condensation process can be vulnerable to backdoor attacks, where malicious triggers are injected into the condensation dataset, manipulating model behavior during inference. While prior approaches have made progress in balancing attack success rate and clean test accuracy, they often fall short in preserving stealthiness, especially in concealing the visual artifacts of condensed data or the perturbations introduced during inference. To address this challenge, we introduce Sneakdoor, which enhances stealthiness without compromising attack effectiveness. Sneakdoor exploits the inherent vulnerability of class decision boundaries and incorporates a generative module that constructs input-aware triggers aligned with local feature geometry, thereby minimizing detectability. This joint design enables the attack to remain imperceptible to both human inspection and statistical detection. Extensive experiments across multiple datasets demonstrate that Sneakdoor achieves a compelling balance among attack success rate, clean test accuracy, and stealthiness, substantially improving the invisibility of both the synthetic data and triggered samples while maintaining high attack efficacy. The code is available at https://github.com/XJTU-AI-Lab/SneakDoor.", "date": "2026-03-29", "categories": []}
{"id": "2603.29038", "title": "Trojan-Speak: Bypassing Constitutional Classifiers with No Jailbreak Tax via Adversarial Finetuning", "abstract": "Fine-tuning APIs offered by major AI providers create new attack surfaces where adversaries can bypass safety measures through targeted fine-tuning. We introduce Trojan-Speak, an adversarial fine-tuning method that bypasses Anthropic's Constitutional Classifiers. Our approach uses curriculum learning combined with GRPO-based hybrid reinforcement learning to teach models a communication protocol that evades LLM-based content classification. Crucially, while prior adversarial fine-tuning approaches report more than 25% capability degradation on reasoning benchmarks, Trojan-Speak incurs less than 5% degradation while achieving 99+% classifier evasion for models with 14B+ parameters. We demonstrate that fine-tuned models can provide detailed responses to expert-level CBRN (Chemical, Biological, Radiological, and Nuclear) queries from Anthropic's Constitutional Classifiers bug-bounty program. Our findings reveal that LLM-based content classifiers alone are insufficient for preventing dangerous information disclosure when adversaries have fine-tuning access, and we show that activation-level probes can substantially improve robustness to such attacks.", "date": "2026-03-30", "categories": []}
{"id": "2603.29062", "title": "CivicShield: A Cross-Domain Defense-in-Depth Framework for Securing Government-Facing AI Chatbots Against Multi-Turn Adversarial Attacks", "abstract": "LLM-based chatbots in government services face critical security gaps. Multi-turn adversarial attacks achieve over 90% success against current defenses, and single-layer guardrails are bypassed with similar rates. We present CivicShield, a cross-domain defense-in-depth framework for government-facing AI chatbots. Drawing on network security, formal verification, biological immune systems, aviation safety, and zero-trust cryptography, CivicShield introduces seven defense layers: (1) zero-trust foundation with capability-based access control, (2) perimeter input validation, (3) semantic firewall with intent classification, (4) conversation state machine with safety invariants, (5) behavioral anomaly detection, (6) multi-model consensus verification, and (7) graduated human-in-the-loop escalation. We present a formal threat model covering 8 multi-turn attack families, map the framework to NIST SP 800-53 controls across 14 families, and evaluate using ablation analysis. Theoretical analysis shows layered defenses reduce attack probability by 1-2 orders of magnitude versus single-layer approaches. Simulation against 1,436 scenarios including HarmBench (416), JailbreakBench (200), and XSTest (450) achieves 72.9% combined detection [69.5-76.0% CI] with 2.9% effective false positive rate after graduated response, while maintaining 100% detection of multi-turn crescendo and slow-drift attacks. The honest drop on real benchmarks versus author-generated scenarios (71.2% vs 76.7% on HarmBench, 47.0% vs 70.0% on JailbreakBench) validates independent evaluation importance. CivicShield addresses an open gap at the intersection of AI safety, government compliance, and practical deployment.", "date": "2026-03-30", "categories": []}
{"id": "2603.29182", "title": "Dummy-Aware Weighted Attack (DAWA): Breaking the Safe Sink in Dummy Class Defenses", "abstract": "Adversarial robustness evaluation faces a critical challenge as new defense paradigms emerge that can exploit limitations in existing assessment methods. This paper reveals that Dummy Classes-based defenses, which introduce an additional \"dummy\" class as a safety sink for adversarial examples, achieve significantly overestimated robustness under conventional evaluation strategies like AutoAttack. The fundamental limitation stems from these attacks' singular focus on misleading the true class label, which aligns perfectly with the defense mechanism--successful attacks are simply captured by the dummy class. To address this gap, we propose Dummy-Aware Weighted Attack (DAWA), a novel evaluation method that simultaneously targets both the true label and dummy label with adaptive weighting during adversarial example synthesis. Extensive experiments demonstrate that DAWA effectively breaks this defense paradigm, reducing the measured robustness of a leading Dummy Classes-based defense from 58.61% to 29.52% on CIFAR-10 under l_infty perturbation (epsilon=8/255). Our work provides a more reliable benchmark for evaluating this emerging class of defenses and highlights the need for continuous evolution of robustness assessment methodologies.", "date": "2026-03-31", "categories": []}
{"id": "2603.29418", "title": "Adversarial Prompt Injection Attack on Multimodal Large Language Models", "abstract": "Although multimodal large language models (MLLMs) are increasingly deployed in real-world applications, their instruction-following behavior leaves them vulnerable to prompt injection attacks. Existing prompt injection methods predominantly rely on textual prompts or perceptible visual prompts that are observable by human users. In this work, we study imperceptible visual prompt injection against powerful closed-source MLLMs, where adversarial instructions are embedded in the visual modality. Our method adaptively embeds the malicious prompt into the input image via a bounded text overlay to provide semantic guidance. Meanwhile, the imperceptible visual perturbation is iteratively optimized to align the feature representation of the attacked image with those of the malicious visual and textual targets at both coarse- and fine-grained levels. Specifically, the visual target is instantiated as a text-rendered image and progressively refined during optimization to more faithfully represent the desired semantics and improve transferability. Extensive experiments on two multimodal understanding tasks across multiple closed-source MLLMs demonstrate the superior performance of our approach compared to existing methods.", "date": "2026-03-31", "categories": []}
{"id": "2504.16474", "title": "Seeking Flat Minima over Diverse Surrogates for Improved Adversarial Transferability: A Theoretical Framework and Algorithmic Instantiation", "abstract": "The transfer-based black-box adversarial attack setting poses the challenge of crafting an adversarial example (AE) on known surrogate models that remain effective against unseen target models. Due to the practical importance of this task, numerous methods have been proposed to address this challenge. However, most previous methods are heuristically designed and intuitively justified, lacking a theoretical foundation. To bridge this gap, we derive a novel transferability bound that offers provable guarantees for adversarial transferability. Our theoretical analysis has the advantages of \\textit{(i)} deepening our understanding of previous methods by building a general attack framework and \\textit{(ii)} providing guidance for designing an effective attack algorithm. Our theoretical results demonstrate that optimizing AEs toward flat minima over the surrogate model set, while controlling the surrogate-target model shift measured by the adversarial model discrepancy, yields a comprehensive guarantee for AE transferability. The results further lead to a general transfer-based attack framework, within which we observe that previous methods consider only partial factors contributing to the transferability. Algorithmically, inspired by our theoretical results, we first elaborately construct the surrogate model set in which models exhibit diverse adversarial vulnerabilities with respect to AEs to narrow an instantiated adversarial model discrepancy. Then, a \\textit{model-Diversity-compatible Reverse Adversarial Perturbation} (DRAP) is generated to effectively promote the flatness of AEs over diverse surrogate models to improve transferability. Extensive experiments on NIPS2017 and CIFAR-10 datasets against various target models demonstrate the effectiveness of our proposed attack.", "date": "2025-04-23", "categories": []}
{"id": "2505.13280", "title": "FlowPure: Continuous Normalizing Flows for Adversarial Purification", "abstract": "Despite significant advances in the area, adversarial robustness remains a critical challenge in systems employing machine learning models. The removal of adversarial perturbations at inference time, known as adversarial purification, has emerged as a promising defense strategy. To achieve this, state-of-the-art methods leverage diffusion models that inject Gaussian noise during a forward process to dilute adversarial perturbations, followed by a denoising step to restore clean samples before classification. In this work, we propose FlowPure, a novel purification method based on Continuous Normalizing Flows (CNFs) trained with Conditional Flow Matching (CFM) to learn mappings from adversarial examples to their clean counterparts. Unlike prior diffusion-based approaches that rely on fixed noise processes, FlowPure can leverage specific attack knowledge to improve robustness under known threats, while also supporting a more general stochastic variant trained on Gaussian perturbations for settings where such knowledge is unavailable. Experiments on CIFAR-10 and CIFAR-100 demonstrate that our method outperforms state-of-the-art purification defenses in preprocessor-blind and white-box scenarios, and can do so while fully preserving benign accuracy in the former. Moreover, our results show that not only is FlowPure a highly effective purifier but it also holds strong potential for adversarial detection, identifying preprocessor-blind PGD samples with near-perfect accuracy. Our code is publicly available at https://github.com/DistriNet/FlowPure.", "date": "2025-05-19", "categories": []}
{"id": "2509.02028", "title": "See No Evil: Adversarial Attacks Against Linguistic-Visual Association in Referring Multi-Object Tracking Systems", "abstract": "Language-vision understanding has driven the development of advanced perception systems, most notably the emerging paradigm of Referring Multi-Object Tracking (RMOT). By leveraging natural-language queries, RMOT systems can selectively track objects that satisfy a given semantic description, guided through Transformer-based spatial-temporal reasoning modules. End-to-End (E2E) RMOT models further unify feature extraction, temporal memory, and spatial reasoning within a Transformer backbone, enabling long-range spatial-temporal modeling over fused textual-visual representations. Despite these advances, the reliability and robustness of RMOT remain underexplored. In this paper, we examine the security implications of RMOT systems from a design-logic perspective, identifying adversarial vulnerabilities that compromise both the linguistic-visual referring and track-object matching components. Additionally, we uncover a novel vulnerability in advanced RMOT models employing FIFO-based memory, whereby targeted and consistent attacks on their spatial-temporal reasoning introduce errors that persist within the history buffer over multiple subsequent frames. We present VEIL, a novel adversarial framework designed to disrupt the unified referring-matching mechanisms of RMOT models. We show that carefully crafted digital and physical perturbations can corrupt the tracking logic reliability, inducing track ID switches and terminations. We conduct comprehensive evaluations using the Refer-KITTI dataset to validate the effectiveness of VEIL and demonstrate the urgent need for security-aware RMOT designs for critical large-scale applications.", "date": "2025-09-02", "categories": []}
{"id": "2512.08503", "title": "Disrupting Hierarchical Reasoning: Adversarial Protection for Geographic Privacy in Multimodal Reasoning Models", "abstract": "Multi-modal large reasoning models (MLRMs) pose significant privacy risks by inferring precise geographic locations from personal images through hierarchical chain-of-thought reasoning. Existing privacy protection techniques, primarily designed for perception-based models, prove ineffective against MLRMs' sophisticated multi-step reasoning processes that analyze environmental cues. We introduce \\textbf{ReasonBreak}, a novel adversarial framework specifically designed to disrupt hierarchical reasoning in MLRMs through concept-aware perturbations. Our approach is founded on the key insight that effective disruption of geographic reasoning requires perturbations aligned with conceptual hierarchies rather than uniform noise. ReasonBreak strategically targets critical conceptual dependencies within reasoning chains, generating perturbations that invalidate specific inference steps and cascade through subsequent reasoning stages. To facilitate this approach, we contribute \\textbf{GeoPrivacy-6K}, a comprehensive dataset comprising 6,341 ultra-high-resolution images ($\\geq$2K) with hierarchical concept annotations. Extensive evaluation across seven state-of-the-art MLRMs (including GPT-o3, GPT-5, Gemini 2.5 Pro) demonstrates ReasonBreak's superior effectiveness, achieving a 14.4\\% improvement in tract-level protection (33.8\\% vs 19.4\\%) and nearly doubling block-level protection (33.5\\% vs 16.8\\%). This work establishes a new paradigm for privacy protection against reasoning-based threats.", "date": "2025-12-09", "categories": []}
{"id": "2603.20339", "title": "Graph-Aware Stealthy Poison-Text Backdoors for Text-Attributed Graphs", "abstract": "Modern graph learning systems often combine links with text, as in citation networks with abstracts or social graphs with user posts. In such systems, text is usually easier to edit than graph structure, which creates a practical security risk: an attacker may hide a small malicious cue in training text and later use it to trigger incorrect predictions. This paper studies that risk in a realistic setting where the attacker edits only node text and leaves the graph unchanged. We propose \\textbf{TAGBD}, a graph-aware backdoor attack that first selects training nodes that are easier to manipulate, then generates stealthy poison text with a shadow graph model, and finally injects the text by replacing the original content or appending a short phrase. Experiments on three benchmark datasets show that TAGBD achieves very high attack success rates, transfers across different graph models, and remains effective under common defenses. These results show that inconspicuous poison text alone can serve as a reliable attack channel in text-attributed graphs, highlighting the need for defenses that inspect both node content and graph structure.", "date": "2026-03-20", "categories": []}
{"id": "2603.21876", "title": "Thermal Topology Collapse: Universal Physical Patch Attacks on Infrared Vision Systems", "abstract": "Although infrared pedestrian detectors have been widely deployed in visual perception tasks, their vulnerability to physical adversarial attacks is becoming increasingly apparent. Existing physical attack methods predominantly rely on instance-specific online optimization and rigid pattern design, leading to high deployment costs and insufficient physical robustness. To address these limitations, this work proposes the Universal Physical Patch Attack (UPPA), the first universal physical attack method in the infrared domain. This method employs geometrically constrained parameterized Bezier blocks to model perturbations and utilizes the Particle Swarm Optimization (PSO) algorithm to perform unified optimization across the global data distribution, thus maintaining topological stability under dynamic deformations. In the physical deployment phase, we materialize the optimized digital perturbations into physical cold patches, achieving a continuous and smooth low-temperature distribution that naturally aligns with the thermal radiation characteristics of infrared imaging. Extensive experiments demonstrate that UPPA achieves an outstanding physical attack success rate without any online computational overhead, while also exhibiting strong cross-domain generalization and reliable black-box transferability.", "date": "2026-03-23", "categories": []}
{"id": "2603.26136", "title": "PEANUT: Perturbations by Eigenvector Alignment for Attacking Graph Neural Networks Under Topology-Driven Message Passing", "abstract": "Graph Neural Networks (GNNs) have achieved remarkable performance on tasks involving relational data. However, small perturbations to the graph structure can significantly alter GNN outputs, raising concerns about their robustness in real-world deployments. In this work, we explore the core vulnerability of GNNs which explicitly consume graph topology in the form of the adjacency matrix or Laplacian as a means for message passing, and propose PEANUT, a simple, gradient-free, restricted black-box attack that injects virtual nodes to capitalize on this vulnerability. PEANUT is a injection based attack, which is widely considered to be more practical and realistic scenario than graph modification attacks, where the attacker is able to modify the original graph structure directly. Our method works at the inference phase, making it an evasion attack, and is applicable almost immediately, since it does not involve lengthy iterative optimizations or parameter learning, which add computational and time overhead, or training surrogate models, which are susceptible to failure due to differences in model priors and generalization capabilities. PEANUT also does not require any features on the injected node and consequently demonstrates that GNN performance can be significantly deteriorated even with injected nodes with zeros for features, highlighting the significance of effectively designed connectivity in such attacks. Extensive experiments on real-world datasets across three graph tasks demonstrate the effectiveness of our attack despite its simplicity.", "date": "2026-03-27", "categories": []}
{"id": "2603.27332", "title": "Unsafe by Reciprocity: How Generation-Understanding Coupling Undermines Safety in Unified Multimodal Models", "abstract": "Recent advances in Large Language Models (LLMs) and Text-to-Image (T2I) models have led to the emergence of Unified Multimodal Models (UMMs), where multimodal understanding and image generation are tightly integrated within a shared architecture. Prior studies suggest that such reciprocity enhances cross-functionality performance through shared representations and joint optimization. However, the safety implications of this tight coupling remain largely unexplored, as existing safety research predominantly analyzes understanding and generation functionalities in isolation. In this work, we investigate whether cross-functionality reciprocity itself constitutes a structural source of vulnerability in UMMs. We propose RICE: Reciprocal Interaction-based Cross-functionality Exploitation, a novel attack paradigm that explicitly exploits bidirectional interactions between understanding and generation. Using this framework, we systematically evaluate Generation-to-Understanding (G-U) and Understanding-to-Generation (U-G) attack pathways, demonstrating that unsafe intermediate signals can propagate across modalities and amplify safety risks. Extensive experiments show high Attack Success Rates (ASR) in both directions, revealing previously overlooked safety weaknesses inherent to UMMs.", "date": "2026-03-28", "categories": []}
{"id": "2603.27918", "title": "Adversarial Attacks on Multimodal Large Language Models: A Comprehensive Survey", "abstract": "Multimodal large language models (MLLMs) integrate information from multiple modalities such as text, images, audio, and video, enabling complex capabilities such as visual question answering and audio translation. While powerful, this increased expressiveness introduces new and amplified vulnerabilities to adversarial manipulation. This survey provides a comprehensive and systematic analysis of adversarial threats to MLLMs, moving beyond enumerating attack techniques to explain the underlying causes of model susceptibility. We introduce a taxonomy that organizes adversarial attacks according to attacker objectives, unifying diverse attack surfaces across modalities and deployment settings. Additionally, we also present a vulnerability-centric analysis that links integrity attacks, safety and jailbreak failures, control and instruction hijacking, and training-time poisoning to shared architectural and representational weaknesses in multimodal systems. Together, this framework provides an explanatory foundation for understanding adversarial behavior in MLLMs and informs the development of more robust and secure multimodal language systems.", "date": "2026-03-30", "categories": []}
{"id": "2603.28092", "title": "InkDrop: Invisible Backdoor Attacks Against Dataset Condensation", "abstract": "Dataset Condensation (DC) is a data-efficient learning paradigm that synthesizes small yet informative datasets, enabling models to match the performance of full-data training. However, recent work exposes a critical vulnerability of DC to backdoor attacks, where malicious patterns (\\textit{e.g.}, triggers) are implanted into the condensation dataset, inducing targeted misclassification on specific inputs. Existing attacks always prioritize attack effectiveness and model utility, overlooking the crucial dimension of stealthiness. To bridge this gap, we propose InkDrop, which enhances the imperceptibility of malicious manipulation without degrading attack effectiveness and model utility. InkDrop leverages the inherent uncertainty near model decision boundaries, where minor input perturbations can induce semantic shifts, to construct a stealthy and effective backdoor attack. Specifically, InkDrop first selects candidate samples near the target decision boundary that exhibit latent semantic affinity to the target class. It then learns instance-dependent perturbations constrained by perceptual and spatial consistency, embedding targeted malicious behavior into the condensed dataset. Extensive experiments across diverse datasets validate the overall effectiveness of InkDrop, demonstrating its ability to integrate adversarial intent into condensed datasets while preserving model utility and minimizing detectability. Our code is available at https://github.com/lvdongyi/InkDrop.", "date": "2026-03-30", "categories": []}
{"id": "2603.28425", "title": "From Pixels to Reality: Physical-Digital Patch Attacks on Real-World Camera", "abstract": "This demonstration presents Digital-Physical Adversarial Attacks (DiPA), a new class of practical adversarial attacks against pervasive camera-based authentication systems, where an attacker displays an adversarial patch directly on a smartphone screen instead of relying on printed artifacts. This digital-only physical presentation enables rapid deployment, removes the need for total-variation regularization, and improves patch transferability in black-box conditions. DiPA leverages an ensemble of state-of-the-art face-recognition models (ArcFace, MagFace, CosFace) to enhance transfer across unseen commercial systems. Our interactive demo shows a real-time dodging attack against a deployed face-recognition camera, preventing authorized users from being recognized while participants dynamically adjust patch patterns and observe immediate effects on the sensing pipeline. We further demonstrate DiPA's superiority over existing physical attacks in terms of success rate, feature-space distortion, and reductions in detection confidence, highlighting critical vulnerabilities at the intersection of mobile devices, pervasive vision, and sensor-driven authentication infrastructures.", "date": "2026-03-30", "categories": []}
{"id": "2603.28476", "title": "With a Little Help From My Friends: Collective Manipulation in Risk-Controlling Recommender Systems", "abstract": "Recommendation systems have become central gatekeepers of online information, shaping user behaviour across a wide range of activities. In response, users increasingly organize and coordinate to steer algorithmic outcomes toward diverse goals, such as promoting relevant content or limiting harmful material, relying on platform affordances -- such as likes, reviews, or ratings. While these mechanisms can serve beneficial purposes, they can also be leveraged for adversarial manipulation, particularly in systems where such feedback directly informs safety guarantees. In this paper, we study this vulnerability in recently proposed risk-controlling recommender systems, which use binary user feedback (e.g., \"Not Interested\") to provably limit exposure to unwanted content via conformal risk control. We empirically demonstrate that their reliance on aggregate feedback signals makes them inherently susceptible to coordinated adversarial user behaviour. Using data from a large-scale online video-sharing platform, we show that a small coordinated group (comprising only 1% of the user population) can induce up to a 20% degradation in nDCG for non-adversarial users by exploiting the affordances provided by risk-controlling recommender systems. We evaluate simple, realistic attack strategies that require little to no knowledge of the underlying recommendation algorithm and find that, while coordinated users can significantly harm overall recommendation quality, they cannot selectively suppress specific content groups through reporting alone. Finally, we propose a mitigation strategy that shifts guarantees from the group level to the user level, showing empirically how it can reduce the impact of adversarial coordinated behaviour while ensuring personalized safety for individuals.", "date": "2026-03-30", "categories": []}
{"id": "2603.28568", "title": "XSPA: Crafting Imperceptible X-Shaped Sparse Adversarial Perturbations for Transferable Attacks on VLMs", "abstract": "Vision-language models (VLMs) rely on a shared visual-textual representation space to perform tasks such as zero-shot classification, image captioning, and visual question answering (VQA). While this shared space enables strong cross-task generalization, it may also introduce a common vulnerability: small visual perturbations can propagate through the shared embedding space and cause correlated semantic failures across tasks. This risk is particularly important in interactive and decision-support settings, yet it remains unclear whether VLMs are robust to highly constrained, sparse, and geometrically fixed perturbations. To address this question, we propose X-shaped Sparse Pixel Attack (XSPA), an imperceptible structured attack that restricts perturbations to two intersecting diagonal lines. Compared with dense perturbations or flexible localized patches, XSPA operates under a much stricter attack budget and thus provides a more stringent test of VLM robustness. Within this sparse support, XSPA jointly optimizes a classification objective, cross-task semantic guidance, and regularization on perturbation magnitude and along-line smoothness, inducing transferable misclassification as well as semantic drift in captioning and VQA while preserving visual subtlety. Under the default setting, XSPA modifies only about 1.76% of image pixels. Experiments on the COCO dataset show that XSPA consistently degrades performance across all three tasks. Zero-shot accuracy drops by 52.33 points on OpenAI CLIP ViT-L/14 and 67.00 points on OpenCLIP ViT-B/16, while GPT-4-evaluated caption consistency decreases by up to 58.60 points and VQA correctness by up to 44.38 points. These results suggest that even highly sparse and visually subtle perturbations with fixed geometric priors can substantially disrupt cross-task semantics in VLMs, revealing a notable robustness gap in current multimodal systems.", "date": "2026-03-30", "categories": []}
{"id": "2603.28673", "title": "FL-PBM: Pre-Training Backdoor Mitigation for Federated Learning", "abstract": "Backdoor attacks pose a significant threat to the integrity and reliability of Artificial Intelligence (AI) models, enabling adversaries to manipulate model behavior by injecting poisoned data with hidden triggers. These attacks can lead to severe consequences, especially in critical applications such as autonomous driving, healthcare, and finance. Detecting and mitigating backdoor attacks is crucial across the lifespan of model's phases, including pre-training, in-training, and post-training. In this paper, we propose Pre-Training Backdoor Mitigation for Federated Learning (FL-PBM), a novel defense mechanism that proactively filters poisoned data on the client side before model training in a federated learning (FL) environment. The approach consists of three stages: (1) inserting a benign trigger into the data to establish a controlled baseline, (2) applying Principal Component Analysis (PCA) to extract discriminative features and assess the separability of the data, (3) performing Gaussian Mixture Model (GMM) clustering to identify potentially malicious data samples based on their distribution in the PCA-transformed space, and (4) applying a targeted blurring technique to disrupt potential backdoor triggers. Together, these steps ensure that suspicious data is detected early and sanitized effectively, thereby minimizing the influence of backdoor triggers on the global model. Experimental evaluations on image-based datasets demonstrate that FL-PBM reduces attack success rates by up to 95% compared to baseline federated learning (FedAvg) and by 30 to 80% relative to state-of-the-art defenses (RDFL and LPSF). At the same time, it maintains over 90% clean model accuracy in most experiments, achieving better mitigation without degrading model performance.", "date": "2026-03-30", "categories": []}
{"id": "2601.13082", "title": "Adversarial News and Lost Profits: Manipulating Headlines in LLM-Driven Algorithmic Trading", "abstract": "Large Language Models (LLMs) are increasingly adopted in the financial domain. Their exceptional capabilities to analyse textual data make them well-suited for inferring the sentiment of finance-related news. Such feedback can be leveraged by algorithmic trading systems (ATS) to guide buy/sell decisions. However, this practice bears the risk that a threat actor may craft \"adversarial news\" intended to mislead an LLM. In particular, the news headline may include \"malicious\" content that remains invisible to human readers but which is still ingested by the LLM. Although prior work has studied textual adversarial examples, their system-wide impact on LLM-supported ATS has not yet been quantified in terms of monetary risk. To address this threat, we consider an adversary with no direct access to an ATS but able to alter stock-related news headlines on a single day. We evaluate two human-imperceptible manipulations in a financial context: Unicode homoglyph substitutions that misroute models during stock-name recognition, and hidden-text clauses that alter the sentiment of the news headline. We implement a realistic ATS in Backtrader that fuses an LSTM-based price forecast with LLM-derived sentiment (FinBERT, FinGPT, FinLLaMA, and six general-purpose LLMs), and quantify monetary impact using portfolio metrics. Experiments on real-world data show that manipulating a one-day attack over 14 months can reliably mislead LLMs and reduce annual returns by up to 17.7 percentage points. To assess real-world feasibility, we analyze popular scraping libraries and trading platforms and survey 27 FinTech practitioners, confirming our hypotheses. We notified trading platform owners of this security issue.", "date": "2026-01-19", "categories": []}
{"id": "2603.12206", "title": "CLASP: Defending Hybrid Large Language Models Against Hidden State Poisoning Attacks", "abstract": "State space models (SSMs) like Mamba have gained significant traction as efficient alternatives to Transformers, achieving linear complexity while maintaining competitive performance. However, Hidden State Poisoning Attacks (HiSPAs), a recently discovered vulnerability that corrupts SSM memory through adversarial strings, pose a critical threat to these architectures and their hybrid variants. Framing the HiSPA mitigation task as a binary classification problem at the token level, we introduce the CLASP model (Classifier Against State Poisoning) to defend against this threat. CLASP exploits distinct patterns in Mamba's block output embeddings (BOEs) and uses an XGBoost classifier to identify malicious tokens with minimal computational overhead. We consider a realistic scenario in which both SSMs and HiSPAs are likely to be used: an LLM screening r\\'esum\\'es to identify the best candidates for a role. Evaluated on a corpus of 2,483 r\\'esum\\'es totaling 9.5M tokens with controlled injections, CLASP achieves 95.9% token-level F1 score and 99.3% document-level F1 score on malicious tokens detection. Crucially, the model generalizes to unseen attack patterns: under leave-one-out cross-validation, performance remains high (96.9% document-level F1), while under clustered cross-validation with structurally novel triggers, it maintains useful detection capability (91.6% average document-level F1). Operating independently of any downstream model, CLASP processes 1,032 tokens per second with under 4GB VRAM consumption, potentially making it suitable for real-world deployment as a lightweight front-line defense for SSM-based and hybrid architectures. All code and detailed results are available at https://anonymous.4open.science/r/hispikes-91C0.", "date": "2026-03-12", "categories": []}
{"id": "2603.19562", "title": "Neural Uncertainty Principle: A Unified View of Adversarial Fragility and LLM Hallucination", "abstract": "Adversarial vulnerability in vision and hallucination in large language models are conventionally viewed as separate problems, each addressed with modality-specific patches. This study first reveals that they share a common geometric origin: the input and its loss gradient are conjugate observables subject to an irreducible uncertainty bound. Formalizing a Neural Uncertainty Principle (NUP) under a loss-induced state, we find that in near-bound regimes, further compression must be accompanied by increased sensitivity dispersion (adversarial fragility), while weak prompt-gradient coupling leaves generation under-constrained (hallucination). Crucially, this bound is modulated by an input-gradient correlation channel, captured by a specifically designed single-backward probe. In vision, masking highly coupled components improves robustness without costly adversarial training; in language, the same prefill-stage probe detects hallucination risk before generating any answer tokens. NUP thus turns two seemingly separate failure taxonomies into a shared uncertainty-budget view and provides a principled lens for reliability analysis. Guided by this NUP theory, we propose ConjMask (masking high-contribution input components) and LogitReg (logit-side regularization) to improve robustness without adversarial training, and use the probe as a decoding-free risk signal for LLMs, enabling hallucination detection and prompt selection. NUP thus provides a unified, practical framework for diagnosing and mitigating boundary anomalies across perception and generation tasks.", "date": "2026-03-20", "categories": []}
{"id": "2603.26045", "title": "H-Node Attack and Defense in Large Language Models", "abstract": "We present H-Node Adversarial Noise Cancellation (H-Node ANC), a mechanistic framework that identifies, exploits, and defends hallucination representations in transformer-based large language models (LLMs) at the level of individual hidden-state dimensions. A logistic regression probe trained on last-token hidden states localizes hallucination signal to a small set of high-variance dimensions -- termed Hallucination Nodes (H-Nodes) -- with probe AUC reaching 0.90 across four architectures. A white-box adversarial attack amplifies these dimensions at inference time via a real-time forward hook, achieving a selectivity of 3.02x with less than 10% visibility to the defender. Adaptive ANC defense suppresses H-Node excess in-pass using confidence-weighted cancellation, reducing grounded activation drift by 33-42% over static cancellation. A dynamic iterative extension that re-ranks cancellation targets across successive passes recovers up to 0.69 robustness from a single-pass baseline of 8%. All contributions are validated on OPT-125M, Phi-3-mini-4k-instruct, LLaMA-3-8B-Instruct, and Mistral-7B-Instruct-v0.3 (125M-8B parameters). Perplexity impact is surgical (<5%) and MMLU degradation is at most 3%, confirming that the defense does not impair general reasoning capability.", "date": "2026-03-27", "categories": []}
{"id": "2603.26067", "title": "R-PGA: Robust Physical Adversarial Camouflage Generation via Relightable 3D Gaussian Splatting", "abstract": "Physical adversarial camouflage poses a severe security threat to autonomous driving systems by mapping adversarial textures onto 3D objects. Nevertheless, current methods remain brittle in complex dynamic scenarios, failing to generalize across diverse geometric (e.g., viewing configurations) and radiometric (e.g., dynamic illumination, atmospheric scattering) variations. We attribute this deficiency to two fundamental limitations in simulation and optimization. First, the reliance on coarse, oversimplified simulations (e.g., via CARLA) induces a significant domain gap, confining optimization to a biased feature space. Second, standard strategies targeting average performance result in a rugged loss landscape, leaving the camouflage vulnerable to configuration shifts.To bridge these gaps, we propose the Relightable Physical 3D Gaussian Splatting (3DGS) based Attack framework (R-PGA). Technically, to address the simulation fidelity issue, we leverage 3DGS to ensure photo-realistic reconstruction and augment it with physically disentangled attributes to decouple intrinsic material from lighting. Furthermore, we design a hybrid rendering pipeline that leverages precise Relightable 3DGS for foreground rendering, while employing a pre-trained image translation model to synthesize plausible relighted backgrounds that align with the relighted foreground.To address the optimization robustness issue, we propose the Hard Physical Configuration Mining (HPCM) module, designed to actively mine worst-case physical configurations and suppress their corresponding loss peaks. This strategy not only diminishes the overall loss magnitude but also effectively flattens the rugged loss landscape, ensuring consistent adversarial effectiveness and robustness across varying physical configurations.", "date": "2026-03-27", "categories": []}
{"id": "2603.26105", "title": "Are LLM-Enhanced Graph Neural Networks Robust against Poisoning Attacks?", "abstract": "Large Language Models (LLMs) have advanced Graph Neural Networks (GNNs) by enriching node representations with semantic features, giving rise to LLM-enhanced GNNs that achieve notable performance gains. However, the robustness of these models against poisoning attacks, which manipulate both graph structures and textual attributes during training, remains unexplored. To bridge this gap, we propose a robustness assessment framework that systematically evaluates LLM-enhanced GNNs under poisoning attacks. Our framework enables comprehensive evaluation across multiple dimensions. Specifically, we assess 24 victim models by combining eight LLM- or Language Model (LM)-based feature enhancers with three representative GNN backbones. To ensure diversity in attack coverage, we incorporate six structural poisoning attacks (both targeted and non-targeted) and three textual poisoning attacks operating at the character, word, and sentence levels. Furthermore, we employ four real-world datasets, including one released after the emergence of LLMs, to avoid potential ground truth leakage during LLM pretraining, thereby ensuring fair evaluation. Extensive experiments show that LLM-enhanced GNNs exhibit significantly higher accuracy and lower Relative Drop in Accuracy (RDA) than a shallow embedding-based baseline across various attack settings. Our in-depth analysis identifies key factors that contribute to this robustness, such as the effective encoding of structural and label information in node representations. Based on these insights, we outline future research directions from both offensive and defensive perspectives, and propose a new combined attack along with a graph purification defense. To support future research, we release the source code of our framework at~\\url{https://github.com/CyberAlSec/LLMEGNNRP}.", "date": "2026-03-27", "categories": []}
{"id": "2510.06790", "title": "Get RICH or Die Scaling: Profitably Trading Inference Compute for Robustness", "abstract": "Test-time reasoning has raised benchmark performances and even shown promise in addressing the historically intractable problem of making models robust to adversarially out-of-distribution (OOD) data. Indeed, recent work used reasoning to aid satisfaction of model specifications designed to thwart attacks, finding a striking correlation between LLM reasoning effort and robustness to jailbreaks. However, this benefit fades when stronger (e.g. gradient-based or multimodal) attacks are used. This may be expected as models often can't follow instructions on the adversarially OOD data created by such attacks, and instruction following is needed to act in accordance with the attacker-thwarting spec. Thus, we hypothesize that the test-time robustness benefits of specs are unlocked by initial robustness sufficient to follow instructions on OOD data. Namely, we posit the Robustness from Inference Compute Hypothesis (RICH): inference-compute defenses profit as the model's training data better reflects the components of attacked data. Guided by the RICH, we test models of varying initial-robustness levels, finding inference-compute adds robustness even to white-box multimodal attacks, provided the model has sufficient initial robustness. Further evidencing a rich-get-richer dynamic, InternVL 3.5 gpt-oss 20B gains little robustness when its test compute is scaled, but such scaling adds significant robustness if we first robustify its vision encoder (creating the first adversarially robust reasoning VLM in the process). Robustifying models makes attacked components of data more in-distribution (ID), and the RICH suggests this fuels compositional generalization -- understanding OOD data via its ID components -- to following spec instructions on adversarial data. Consistently, we find test-time defenses both build and depend on train-time data and defenses.", "date": "2025-10-08", "categories": []}
{"id": "2512.20821", "title": "Divided We Fall: Defending Against Adversarial Attacks via Soft-Gated Fractional Mixture-of-Experts with Randomized Adversarial Training", "abstract": "Machine learning is a powerful tool enabling full automation of a huge number of tasks without explicit programming. Despite recent progress of machine learning in different domains, these models have shown vulnerabilities when they are exposed to adversarial threats. Adversarial threats aim to hinder the machine learning models from satisfying their objectives. They can create adversarial perturbations, which are imperceptible to humans' eyes but have the ability to cause misclassification during inference. In this paper, we propose a defense system, which devises an adversarial training module within mixture-of-experts architecture to enhance its robustness against white-box evasion attacks. In our proposed defense system, we use nine pre-trained classifiers (experts) with ResNet-18 as their backbone. During end-to-end training, the parameters of all experts and the gating mechanism are jointly updated allowing further optimization of the experts. Our proposed defense system outperforms prior MoE-based defenses under strong white-box FGSM and PGD evaluation on CIFAR-10 and SVHN. The use of multiple experts increases training time and compute relative to single-network baselines; however, inference scales approximately linearly with the number of experts and is substantially cheaper than training.", "date": "2025-12-23", "categories": []}
{"id": "2603.21208", "title": "JANUS: A Lightweight Framework for Jailbreaking Text-to-Image Models via Distribution Optimization", "abstract": "Text-to-image (T2I) models such as Stable Diffusion and DALLE remain susceptible to generating harmful or Not-Safe-For-Work (NSFW) content under jailbreak attacks despite deployed safety filters. Existing jailbreak attacks either rely on proxy-loss optimization instead of the true end-to-end objective, or depend on large-scale and costly RL-trained generators. Motivated by these limitations, we propose JANUS , a lightweight framework that formulates jailbreak as optimizing a structured prompt distribution under a black-box, end-to-end reward from the T2I system and its safety filters. JANUS replaces a high-capacity generator with a low-dimensional mixing policy over two semantically anchored prompt distributions, enabling efficient exploration while preserving the target semantics. On modern T2I models, we outperform state-of-the-art jailbreak methods, improving ASR-8 from 25.30% to 43.15% on Stable Diffusion 3.5 Large Turbo with consistently higher CLIP and NSFW scores. JANUS succeeds across both open-source and commercial models. These findings expose structural weaknesses in current T2I safety pipelines and motivate stronger, distribution-aware defenses. Warning: This paper contains model outputs that may be offensive.", "date": "2026-03-22", "categories": []}
{"id": "2603.24821", "title": "Generative Adversarial Perturbations with Cross-paradigm Transferability on Localized Crowd Counting", "abstract": "State-of-the-art crowd counting and localization are primarily modeled using two paradigms: density maps and point regression. Given the field's security ramifications, there is active interest in model robustness against adversarial attacks. Recent studies have demonstrated transferability across density-map-based approaches via adversarial patches, but cross-paradigm attacks (i.e., across both density map-based models and point regression-based models) remain unexplored. We introduce a novel adversarial framework that compromises both density map and point regression architectural paradigms through a comprehensive multi-task loss optimization. For point-regression models, we employ scene-density-specific high-confidence logit suppression; for density-map approaches, we use peak-targeted density map suppression. Both are combined with model-agnostic perceptual constraints to ensure that perturbations are effective and imperceptible to the human eye. Extensive experiments demonstrate the effectiveness of our attack, achieving on average a 7X increase in Mean Absolute Error compared to clean images while maintaining competitive visual quality, and successfully transferring across seven state-of-the-art crowd models with transfer ratios ranging from 0.55 to 1.69. Our approach strikes a balance between attack effectiveness and imperceptibility compared to state-of-the-art transferable attack strategies. The source code is available at https://github.com/simurgh7/CrowdGen", "date": "2026-03-25", "categories": []}
{"id": "2603.24857", "title": "AI Security in the Foundation Model Era: A Comprehensive Survey from a Unified Perspective", "abstract": "As machine learning (ML) systems expand in both scale and functionality, the security landscape has become increasingly complex, with a proliferation of attacks and defenses. However, existing studies largely treat these threats in isolation, lacking a coherent framework to expose their shared principles and interdependencies. This fragmented view hinders systematic understanding and limits the design of comprehensive defenses. Crucially, the two foundational assets of ML -- \\textbf{data} and \\textbf{models} -- are no longer independent; vulnerabilities in one directly compromise the other. The absence of a holistic framework leaves open questions about how these bidirectional risks propagate across the ML pipeline. To address this critical gap, we propose a \\emph{unified closed-loop threat taxonomy} that explicitly frames model-data interactions along four directional axes. Our framework offers a principled lens for analyzing and defending foundation models. The resulting four classes of security threats represent distinct but interrelated categories of attacks: (1) Data$\\rightarrow$Data (D$\\rightarrow$D): including \\emph{data decryption attacks and watermark removal attacks}; (2) Data$\\rightarrow$Model (D$\\rightarrow$M): including \\emph{poisoning, harmful fine-tuning attacks, and jailbreak attacks}; (3) Model$\\rightarrow$Data (M$\\rightarrow$D): including \\emph{model inversion, membership inference attacks, and training data extraction attacks}; (4) Model$\\rightarrow$Model (M$\\rightarrow$M): including \\emph{model extraction attacks}. Our unified framework elucidates the underlying connections among these security threats and establishes a foundation for developing scalable, transferable, and cross-modal security strategies, particularly within the landscape of foundation models.", "date": "2026-03-25", "categories": []}
{"id": "2603.25164", "title": "PIDP-Attack: Combining Prompt Injection with Database Poisoning Attacks on Retrieval-Augmented Generation Systems", "abstract": "Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of applications. However, their practical deployment is often hindered by issues such as outdated knowledge and the tendency to generate hallucinations. To address these limitations, Retrieval-Augmented Generation (RAG) systems have been introduced, enhancing LLMs with external, up-to-date knowledge sources. Despite their advantages, RAG systems remain vulnerable to adversarial attacks, with data poisoning emerging as a prominent threat. Existing poisoning-based attacks typically require prior knowledge of the user's specific queries, limiting their flexibility and real-world applicability. In this work, we propose PIDP-Attack, a novel compound attack that integrates prompt injection with database poisoning in RAG. By appending malicious characters to queries at inference time and injecting a limited number of poisoned passages into the retrieval database, our method can effectively manipulate LLM response to arbitrary query without prior knowledge of the user's actual query. Experimental evaluations across three benchmark datasets (Natural Questions, HotpotQA, MS-MARCO) and eight LLMs demonstrate that PIDP-Attack consistently outperforms the original PoisonedRAG. Specifically, our method improves attack success rates by 4% to 16% on open-domain QA tasks while maintaining high retrieval precision, proving that the compound attack strategy is both necessary and highly effective.", "date": "2026-03-26", "categories": []}
{"id": "2603.25170", "title": "Knowledge-Guided Adversarial Training for Infrared Object Detection via Thermal Radiation Modeling", "abstract": "In complex environments, infrared object detection exhibits broad applicability and stability across diverse scenarios. However, infrared object detection is vulnerable to both common corruptions and adversarial examples, leading to potential security risks. To improve the robustness of infrared object detection, current methods mostly adopt a data-driven ideology, which only superficially drives the network to fit the training data without specifically considering the unique characteristics of infrared images, resulting in limited robustness. In this paper, we revisit infrared physical knowledge and find that relative thermal radiation relations between different classes can be regarded as a reliable knowledge source under the complex scenarios of adversarial examples and common corruptions. Thus, we theoretically model thermal radiation relations based on the rank order of gray values for different classes, and further quantify the stability of various inter-class thermal radiation relations. Based on the above theoretical framework, we propose Knowledge-Guided Adversarial Training (KGAT) for infrared object detection, in which infrared physical knowledge is embedded into the adversarial training process, and the predicted results are optimized to be consistent with the actual physical laws. Extensive experiments on three infrared datasets and six mainstream infrared object detection models demonstrate that KGAT effectively enhances both clean accuracy and robustness against adversarial attacks and common corruptions.", "date": "2026-03-26", "categories": []}
{"id": "2603.25230", "title": "A Unified Spatial Alignment Framework for Highly Transferable Transformation-Based Attacks on Spatially Structured Tasks", "abstract": "Transformation-based adversarial attacks (TAAs) demonstrate strong transferability when deceiving classification models. However, existing TAAs often perform unsatisfactorily or even fail when applied to structured tasks such as semantic segmentation and object detection. Encouragingly, recent studies that categorize transformations into non-spatial and spatial transformations inspire us to address this challenge. We find that for non-structured tasks, labels are spatially non-structured, and thus TAAs are not required to adjust labels when applying spatial transformations. In contrast, for structured tasks, labels are spatially structured, and failing to transform labels synchronously with inputs can cause spatial misalignment and yield erroneous gradients. To address these issues, we propose a novel unified Spatial Alignment Framework (SAF) for highly transferable TAAs on spatially structured tasks, where the TAAs spatially transform labels synchronously with the input using the proposed Spatial Alignment (SA) algorithm. Extensive experiments demonstrate the crucial role of our SAF for TAAs on structured tasks. Specifically, in non-targeted attacks, our SAF degrades the average mIoU on Cityscapes from 24.50 to 11.34, and on Kvasir-SEG from 49.91 to 31.80, while reducing the average mAP of COCO from 17.89 to 5.25.", "date": "2026-03-26", "categories": []}
{"id": "2603.25244", "title": "Efficient Preemptive Robustification with Image Sharpening", "abstract": "Despite their great success, deep neural networks rely on high-dimensional, non-robust representations, making them vulnerable to imperceptible perturbations, even in transfer scenarios. To address this, both training-time defenses (e.g., adversarial training and robust architecture design) and post-attack defenses (e.g., input purification and adversarial detection) have been extensively studied. Recently, a limited body of work has preliminarily explored a pre-attack defense paradigm, termed preemptive robustification, which introduces subtle modifications to benign samples prior to attack to proactively resist adversarial perturbations. Unfortunately, their practical applicability remains questionable due to several limitations, including (1) reliance on well-trained classifiers as surrogates to provide robustness priors, (2) substantial computational overhead arising from iterative optimization or trained generators for robustification, and (3) limited interpretability of the optimization- or generation-based robustification processes. Inspired by recent studies revealing a positive correlation between texture intensity and the robustness of benign samples, we show that image sharpening alone can efficiently robustify images. To the best of our knowledge, this is the first surrogate-free, optimization-free, generator-free, and human-interpretable robustification approach. Extensive experiments demonstrate that sharpening yields remarkable robustness gains with low computational cost, especially in transfer scenarios.", "date": "2026-03-26", "categories": []}
{"id": "2603.25257", "title": "Mitigating Evasion Attacks in Fog Computing Resource Provisioning Through Proactive Hardening", "abstract": "This paper investigates the susceptibility to model integrity attacks that overload virtual machines assigned by the k-means algorithm used for resource provisioning in fog networks. The considered k-means algorithm runs two phases iteratively: offline clustering to form clusters of requested workload and online classification of new incoming requests into offline-created clusters. First, we consider an evasion attack against the classifier in the online phase. A threat actor launches an exploratory attack using query-based reverse engineering to discover the Machine Learning (ML) model (the clustering scheme). Then, a passive causative (evasion) attack is triggered in the offline phase. To defend the model, we suggest a proactive method using adversarial training to introduce attack robustness into the classifier. Our results show that our mitigation technique effectively maintains the stability of the resource provisioning system against attacks.", "date": "2026-03-26", "categories": []}
{"id": "2409.10562", "title": "Natural Adversaries: Fuzzing Autonomous Vehicles with Realistic Roadside Object Placements", "abstract": "The emergence of Autonomous Vehicles (AVs) has spurred research into testing the resilience of their perception systems, i.e., ensuring that they are not susceptible to critical misjudgements. It is important that these systems are tested not only with respect to other vehicles on the road, but also with respect to objects placed on the roadside. Trash bins, billboards, and greenery are examples of such objects, typically positioned according to guidelines developed for the human visual system, which may not align perfectly with the needs of AVs. Existing tests, however, usually focus on adversarial objects with conspicuous shapes or patches, which are ultimately unrealistic due to their unnatural appearance and reliance on white-box knowledge. In this work, we introduce a black-box attack on AV perception systems that creates realistic adversarial scenarios (i.e., satisfying road design guidelines) by manipulating the positions of common roadside objects and without resorting to \"unnatural\" adversarial patches. In particular, we propose TrashFuzz, a fuzzing algorithm that finds scenarios in which the placement of these objects leads to substantial AV misperceptions -- such as mistaking a traffic light's colour -- with the overall goal of causing traffic-law violations. To ensure realism, these scenarios must satisfy several rules encoding regulatory guidelines governing the placement of objects on public streets. We implemented and evaluated these attacks on the Apollo autonomous driving system, finding that TrashFuzz induced violations of 15 out of 24 traffic laws.", "date": "2024-09-13", "categories": []}
{"id": "2504.09604", "title": "Mitigating Many-Shot Jailbreaking", "abstract": "Many-shot jailbreaking (MSJ) is an adversarial technique that exploits the long context windows of modern LLMs to circumvent model safety training by including in the prompt many examples of a \"fake\" assistant responding inappropriately before the final request. With enough examples, the model's in-context learning abilities override its safety training, and it responds as if it were the \"fake\" assistant. In this work, we probe the effectiveness of different fine-tuning and input sanitization approaches on mitigating MSJ attacks, alone and in combination. We find incremental mitigation effectiveness for each, and show that the combined techniques significantly reduce the effectiveness of MSJ attacks, while retaining model performance in benign in-context learning and conversational tasks. We suggest that our approach could meaningfully ameliorate this vulnerability if incorporated into model safety post-training.", "date": "2025-04-13", "categories": []}
{"id": "2507.22171", "title": "Enhancing Jailbreak Attacks on LLMs via Persona Prompts", "abstract": "Jailbreak attacks aim to exploit large language models (LLMs) by inducing them to generate harmful content, thereby revealing their vulnerabilities. Understanding and addressing these attacks is crucial for advancing the field of LLM safety. Previous jailbreak approaches have mainly focused on direct manipulations of harmful intent, with limited attention to the impact of persona prompts. In this study, we systematically explore the efficacy of persona prompts in compromising LLM defenses. We propose a genetic algorithm-based method that automatically crafts persona prompts to bypass LLM's safety mechanisms. Our experiments reveal that: (1) our evolved persona prompts reduce refusal rates by 50-70% across multiple LLMs, and (2) these prompts demonstrate synergistic effects when combined with existing attack methods, increasing success rates by 10-20%. Our code and data are available at https://github.com/CjangCjengh/Generic_Persona.", "date": "2025-07-28", "categories": []}
{"id": "2603.23860", "title": "Why the Maximum Second Derivative of Activations Matters for Adversarial Robustness", "abstract": "This work investigates the critical role of activation function curvature -- quantified by the maximum second derivative $\\max|\\sigma''|$ -- in adversarial robustness. Using the Recursive Curvature-Tunable Activation Family (RCT-AF), which enables precise control over curvature through parameters $\\alpha$ and $\\beta$, we systematically analyze this relationship. Our study reveals a fundamental trade-off: insufficient curvature limits model expressivity, while excessive curvature amplifies the normalized Hessian diagonal norm of the loss, leading to sharper minima that hinder robust generalization. This results in a non-monotonic relationship where optimal adversarial robustness consistently occurs when $\\max|\\sigma''|$ falls within 4 to 10, a finding that holds across diverse network architectures, datasets, and adversarial training methods. We provide theoretical insights into how activation curvature affects the diagonal elements of the hessian matrix of the loss, and experimentally demonstrate that the normalized Hessian diagonal norm exhibits a U-shaped dependence on $\\max|\\sigma''|$, with its minimum within the optimal robustness range, thereby validating the proposed mechanism.", "date": "2026-03-25", "categories": []}
{"id": "2603.24232", "title": "Attack Assessment and Augmented Identity Recognition for Human Skeleton Data", "abstract": "Machine learning models trained on small data sets for security applications are especially vulnerable to adversarial attacks. Person identification from LiDAR based skeleton data requires time consuming and expensive data acquisition for each subject identity. Recently, Assessment and Augmented Identity Recognition for Skeletons (AAIRS) has been used to train Hierarchical Co-occurrence Networks for Person Identification (HCN-ID) with small LiDAR based skeleton data sets. However, AAIRS does not evaluate robustness of HCN-ID to adversarial attacks or inoculate the model to defend against such attacks. Popular perturbation-based approaches to generating adversarial attacks are constrained to targeted perturbations added to real training samples, which is not ideal for inoculating models with small training sets. Thus, we propose Attack-AAIRS, a novel addition to the AAIRS framework. Attack-AAIRS leverages a small real data set and a GAN generated synthetic data set to assess and improve model robustness against unseen adversarial attacks. Rather than being constrained to perturbations of limited real training samples, the GAN learns the distribution of adversarial attack samples that exploit weaknesses in HCN-ID. Attack samples drawn from this distribution augment training for inoculation of the HCN-ID to improve robustness. Ten-fold cross validation of Attack-AAIRS yields increased robustness to unseen attacks- including FGSM, PGD, Additive Gaussian Noise, MI-FGSM, and BIM. The HCN-ID Synthetic Data Quality Score for Attack-AAIRS indicates that generated attack samples are of similar quality to the original benign synthetic samples generated by AAIRS. Furthermore, inoculated models show consistent final test accuracy with the original model trained on real data, demonstrating that our method improves robustness to adversarial attacks without reducing test performance on real data.", "date": "2026-03-25", "categories": []}
{"id": "2603.24570", "title": "Anti-I2V: Safeguarding your photos from malicious image-to-video generation", "abstract": "Advances in diffusion-based video generation models, while significantly improving human animation, poses threats of misuse through the creation of fake videos from a specific person's photo and text prompts. Recent efforts have focused on adversarial attacks that introduce crafted perturbations to protect images from diffusion-based models. However, most existing approaches target image generation, while relatively few explicitly address image-to-video diffusion models (VDMs), and most primarily focus on UNet-based architectures. Hence, their effectiveness against Diffusion Transformer (DiT) models remains largely under-explored, as these models demonstrate improved feature retention, and stronger temporal consistency due to larger capacity and advanced attention mechanisms. In this work, we introduce Anti-I2V, a novel defense against malicious human image-to-video generation, applicable across diverse diffusion backbones. Instead of restricting noise updates to the RGB space, Anti-I2V operates in both the $L$*$a$*$b$* and frequency domains, improving robustness and concentrating on salient pixels. We then identify the network layers that capture the most distinct semantic features during the denoising process to design appropriate training objectives that maximize degradation of temporal coherence and generation fidelity. Through extensive validation, Anti-I2V demonstrates state-of-the-art defense performance against diverse video diffusion models, offering an effective solution to the problem.", "date": "2026-03-25", "categories": []}
{"id": "2406.03017", "title": "DifAttack++: Query-Efficient Black-Box Adversarial Attack via Hierarchical Disentangled Feature Space in Cross-Domain", "abstract": "This work investigates efficient score-based black-box adversarial attacks that achieve a high Attack Success Rate (ASR) and good generalization ability. We propose a novel attack framework, termed DifAttack++, which operates in a hierarchical disentangled feature space and significantly differs from existing methods that manipulate the entire feature space. Specifically, DifAttack++ firstly disentangles an image's latent representation into an Adversarial Feature (AF) and a Visual Feature (VF) using an autoencoder equipped with a carefully designed Hierarchical Decouple-Fusion (HDF) module. In this formulation, the AF primarily governs the adversarial capability of an image, while the VF largely preserves its visual appearance. To enable the feature disentanglement and image reconstruction, we jointly train two autoencoders for the clean and adversarial image domains, i.e., cross-domain, respectively, using paired clean images and their corresponding Adversarial Examples (AEs) generated by white-box attacks on available surrogate models. During the black-box attack stage, DifAttack++ iteratively optimizes the AF based on query feedback from the victim model, while keeping the VF fixed, until a successful AE is obtained. Extensive experimental results demonstrate that DifAttack++ achieves superior ASR and query efficiency compared to state-of-the-art methods, while producing AEs with comparable visual quality. Our code is available at https://github.com/csjunjun/DifAttackPlus.git.", "date": "2024-06-05", "categories": []}
{"id": "2501.08415", "title": "Cross-Modal Transferable Image-to-Video Attack on Video Quality Metrics", "abstract": "Recent studies have revealed that modern image and video quality assessment (IQA/VQA) metrics are vulnerable to adversarial attacks. An attacker can manipulate a video through preprocessing to artificially increase its quality score according to a certain metric, despite no actual improvement in visual quality. Most of the attacks studied in the literature are white-box attacks, while black-box attacks in the context of VQA have received less attention. Moreover, some research indicates a lack of transferability of adversarial examples generated for one model to another when applied to VQA. In this paper, we propose a cross-modal attack method, IC2VQA, aimed at exploring the vulnerabilities of modern VQA models. This approach is motivated by the observation that the low-level feature spaces of images and videos are similar. We investigate the transferability of adversarial perturbations across different modalities; specifically, we analyze how adversarial perturbations generated on a white-box IQA model with an additional CLIP module can effectively target a VQA model. The addition of the CLIP module serves as a valuable aid in increasing transferability, as the CLIP model is known for its effective capture of low-level semantics. Extensive experiments demonstrate that IC2VQA achieves a high success rate in attacking three black-box VQA models. We compare our method with existing black-box attack strategies, highlighting its superiority in terms of attack success within the same number of iterations and levels of attack strength. We believe that the proposed method will contribute to the deeper analysis of robust VQA metrics.", "date": "2025-01-14", "categories": []}
{"id": "2511.05919", "title": "Injecting Falsehoods: Adversarial Man-in-the-Middle Attacks Undermining Factual Recall in LLMs", "abstract": "LLMs are now an integral part of information retrieval. As such, their role as question answering chatbots raises significant concerns due to their shown vulnerability to adversarial man-in-the-middle (MitM) attacks. Here, we propose the first principled attack evaluation on LLM factual memory under prompt injection via Xmera, our novel, theory-grounded MitM framework. By perturbing the input given to \"victim\" LLMs in three closed-book and fact-based QA settings, we undermine the correctness of the responses and assess the uncertainty of their generation process. Surprisingly, trivial instruction-based attacks report the highest success rate (up to ~85.3%) while simultaneously having a high uncertainty for incorrectly answered questions. To provide a simple defense mechanism against Xmera, we train Random Forest classifiers on the response uncertainty levels to distinguish between attacked and unattacked queries (average AUC of up to ~94.8%). We believe that signaling users to be cautious about the answers they receive from black-box and potentially corrupt LLMs is a first checkpoint toward user cyberspace safety.", "date": "2025-11-08", "categories": []}
{"id": "2512.10766", "title": "Metaphor-based Jailbreak Attacks on Text-to-Image Models", "abstract": "Text-to-image (T2I) models commonly incorporate defense mechanisms to prevent the generation of sensitive images. Unfortunately, recent jailbreak attacks have shown that adversarial prompts can effectively bypass these mechanisms and induce T2I models to produce sensitive content, revealing critical safety vulnerabilities. However, existing attack methods implicitly assume that the attacker knows the type of deployed defenses, which limits their effectiveness against unknown or diverse defense mechanisms. In this work, we reveal an underexplored vulnerability of T2I models to metaphor-based jailbreak attacks (MJA), which aims to attack diverse defense mechanisms without prior knowledge of their type by generating metaphor-based adversarial prompts. Specifically, MJA consists of two modules: an LLM-based multi-agent generation module (LMAG) and an adversarial prompt optimization module (APO). LMAG decomposes the generation of metaphor-based adversarial prompts into three subtasks: metaphor retrieval, context matching, and adversarial prompt generation. Subsequently, LMAG coordinates three LLM-based agents to generate diverse adversarial prompts by exploring various metaphors and contexts. To enhance attack efficiency, APO first trains a surrogate model to predict the attack results of adversarial prompts and then designs an acquisition strategy to adaptively identify optimal adversarial prompts. Extensive experiments on T2I models with various external and internal defense mechanisms demonstrate that MJA achieves stronger attack performance while using fewer queries, compared with six baseline methods. Additionally, we provide an in-depth vulnerability analysis suggesting that metaphor-based adversarial prompts evade safety mechanisms by inducing semantic ambiguity, while sensitive images arise from the model's probabilistic interpretation of concealed semantics.", "date": "2025-12-06", "categories": []}
{"id": "2602.22983", "title": "Obscure but Effective: Classical Chinese Jailbreak Prompt Optimization via Bio-Inspired Search", "abstract": "As Large Language Models (LLMs) are increasingly used, their security risks have drawn increasing attention. Existing research reveals that LLMs are highly susceptible to jailbreak attacks, with effectiveness varying across language contexts. This paper investigates the role of classical Chinese in jailbreak attacks. Owing to its conciseness and obscurity, classical Chinese can partially bypass existing safety constraints, exposing notable vulnerabilities in LLMs. Based on this observation, this paper proposes a framework, CC-BOS, for the automatic generation of classical Chinese adversarial prompts based on multi-dimensional fruit fly optimization, facilitating efficient and automated jailbreak attacks in black-box settings. Prompts are encoded into eight policy dimensions-covering role, behavior, mechanism, metaphor, expression, knowledge, trigger pattern and context; and iteratively refined via smell search, visual search, and cauchy mutation. This design enables efficient exploration of the search space, thereby enhancing the effectiveness of black-box jailbreak attacks. To enhance readability and evaluation accuracy, we further design a classical Chinese to English translation module. Extensive experiments demonstrate that effectiveness of the proposed CC-BOS, consistently outperforming state-of-the-art jailbreak attack methods.", "date": "2026-02-26", "categories": []}
{"id": "2603.22341", "title": "T-MAP: Red-Teaming LLM Agents with Trajectory-aware Evolutionary Search", "abstract": "While prior red-teaming efforts have focused on eliciting harmful text outputs from large language models (LLMs), such approaches fail to capture agent-specific vulnerabilities that emerge through multi-step tool execution, particularly in rapidly growing ecosystems such as the Model Context Protocol (MCP). To address this gap, we propose a trajectory-aware evolutionary search method, T-MAP, which leverages execution trajectories to guide the discovery of adversarial prompts. Our approach enables the automatic generation of attacks that not only bypass safety guardrails but also reliably realize harmful objectives through actual tool interactions. Empirical evaluations across diverse MCP environments demonstrate that T-MAP substantially outperforms baselines in attack realization rate (ARR) and remains effective against frontier models, including GPT-5.2, Gemini-3-Pro, Qwen3.5, and GLM-5, thereby revealing previously underexplored vulnerabilities in autonomous LLM agents.", "date": "2026-03-21", "categories": []}
{"id": "2603.22525", "title": "Adversarial Vulnerabilities in Neural Operator Digital Twins: Gradient-Free Attacks on Nuclear Thermal-Hydraulic Surrogates", "abstract": "Operator learning models are rapidly emerging as the predictive core of digital twins for nuclear and energy systems, promising real-time field reconstruction from sparse sensor measurements. Yet their robustness to adversarial perturbations remains uncharacterized, a critical gap for deployment in safety-critical systems. Here we show that neural operators are acutely vulnerable to extremely sparse (fewer than 1% of inputs), physically plausible perturbations that exploit their sensitivity to boundary conditions. Using gradient-free differential evolution across four operator architectures, we demonstrate that minimal modifications trigger catastrophic prediction failures, increasing relative $L_2$ error from $\\sim$1.5% (validated accuracy) to 37-63% while remaining completely undetectable by standard validation metrics. Notably, 100% of successful single-point attacks pass z-score anomaly detection. We introduce the effective perturbation dimension $d_{\\text{eff}}$, a Jacobian-based diagnostic that, together with sensitivity magnitude, yields a two-factor vulnerability model explaining why architectures with extreme sensitivity concentration (POD-DeepONet, $d_{\\text{eff}} \\approx 1$) are not necessarily the most exploitable, since low-rank output projections cap maximum error, while moderate concentration with sufficient amplification (S-DeepONet, $d_{\\text{eff}} \\approx 4$) produces the highest attack success. Gradient-free search outperforms gradient-based alternatives (PGD) on architectures with gradient pathologies, while random perturbations of equal magnitude achieve near-zero success rates, confirming that the discovered vulnerabilities are structural. Our findings expose a previously overlooked attack surface in operator learning models and establish that these models require robustness guarantees beyond standard validation before deployment.", "date": "2026-03-23", "categories": []}
{"id": "2603.22882", "title": "TreeTeaming: Autonomous Red-Teaming of Vision-Language Models via Hierarchical Strategy Exploration", "abstract": "The rapid advancement of Vision-Language Models (VLMs) has brought their safety vulnerabilities into sharp focus. However, existing red teaming methods are fundamentally constrained by an inherent linear exploration paradigm, confining them to optimizing within a predefined strategy set and preventing the discovery of novel, diverse exploits. To transcend this limitation, we introduce TreeTeaming, an automated red teaming framework that reframes strategy exploration from static testing to a dynamic, evolutionary discovery process. At its core lies a strategic Orchestrator, powered by a Large Language Model (LLM), which autonomously decides whether to evolve promising attack paths or explore diverse strategic branches, thereby dynamically constructing and expanding a strategy tree. A multimodal actuator is then tasked with executing these complex strategies. In the experiments across 12 prominent VLMs, TreeTeaming achieves state-of-the-art attack success rates on 11 models, outperforming existing methods and reaching up to 87.60\\% on GPT-4o. The framework also demonstrates superior strategic diversity over the union of previously public jailbreak strategies. Furthermore, the generated attacks exhibit an average toxicity reduction of 23.09\\%, showcasing their stealth and subtlety. Our work introduces a new paradigm for automated vulnerability discovery, underscoring the necessity of proactive exploration beyond static heuristics to secure frontier AI models.", "date": "2026-03-24", "categories": []}
{"id": "2509.04597", "title": "DisPatch: Disarming Adversarial Patches in Object Detection with Diffusion Models", "abstract": "Object detection is fundamental to various real-world applications, such as security monitoring and surveillance video analysis. Despite their advancements, state-of-the-art object detectors are still vulnerable to adversarial patch attacks, which can be easily applied to real-world objects to either conceal actual items or create non-existent ones, leading to severe consequences. In this work, we introduce DisPatch, the first diffusion-based defense framework for object detection. Unlike previous works that aim to \"detect and remove\" adversarial patches, DisPatch adopts a \"regenerate and rectify\" strategy, leveraging generative models to disarm attack effects while preserving the integrity of the input image. Specifically, we utilize the in-distribution generative power of diffusion models to regenerate the entire image, aligning it with benign data. A rectification process is then employed to identify and replace adversarial regions with their regenerated benign counterparts. DisPatch is attack-agnostic and requires no prior knowledge of the existing patches. Extensive experiments across multiple detectors demonstrate that DisPatch consistently outperforms state-of-the-art defenses on both hiding attacks and creating attacks, achieving the best overall mAP@0.5 score of 89.3% on hiding attacks, and lowering the attack success rate to 24.8% on untargeted creating attacks. Moreover, it strikes the balance between effectiveness and efficiency, and maintains strong robustness against adaptive attacks, making it a practical and reliable defense method.", "date": "2025-09-04", "categories": []}
{"id": "2511.12985", "title": "Angular Gradient Sign Method: Uncovering Vulnerabilities in Hyperbolic Networks", "abstract": "Adversarial examples in neural networks have been extensively studied in Euclidean geometry, but recent advances in \\textit{hyperbolic networks} call for a reevaluation of attack strategies in non-Euclidean geometries. Existing methods such as FGSM and PGD apply perturbations without regard to the underlying hyperbolic structure, potentially leading to inefficient or geometrically inconsistent attacks. In this work, we propose a novel adversarial attack that explicitly leverages the geometric properties of hyperbolic space. Specifically, we compute the gradient of the loss function in the tangent space of hyperbolic space, decompose it into a radial (depth) component and an angular (semantic) component, and apply perturbation derived solely from the angular direction. Our method generates adversarial examples by focusing perturbations in semantically sensitive directions encoded in angular movement within the hyperbolic geometry. Empirical results on image classification, cross-modal retrieval tasks and network architectures demonstrate that our attack achieves higher fooling rates than conventional adversarial attacks, while producing high-impact perturbations with deeper insights into vulnerabilities of hyperbolic embeddings. This work highlights the importance of geometry-aware adversarial strategies in curved representation spaces and provides a principled framework for attacking hierarchical embeddings.", "date": "2025-11-17", "categories": []}
{"id": "2603.04839", "title": "Towards Highly Transferable Vision-Language Attack via Semantic-Augmented Dynamic Contrastive Interaction", "abstract": "With the rapid advancement and widespread application of vision-language pre-training (VLP) models, their vulnerability to adversarial attacks has become a critical concern. In general, the adversarial examples can typically be designed to exhibit transferable power, attacking not only different models but also across diverse tasks. However, existing attacks on language-vision models mainly rely on static cross-modal interactions and focus solely on disrupting positive image-text pairs, resulting in limited cross-modal disruption and poor transferability. To address this issue, we propose a Semantic-Augmented Dynamic Contrastive Attack (SADCA) that enhances adversarial transferability through progressive and semantically guided perturbation. SADCA progressively disrupts cross-modal alignment through dynamic interactions between adversarial images and texts. This is accomplished by SADCA establishing a contrastive learning mechanism involving adversarial, positive and negative samples, to reinforce the semantic inconsistency of the obtained perturbations. Moreover, we empirically find that input transformations commonly used in traditional transfer-based attacks also benefit VLPs, which motivates a semantic augmentation module that increases the diversity and generalization of adversarial examples. Extensive experiments on multiple datasets and models demonstrate that SADCA significantly improves adversarial transferability and consistently surpasses state-of-the-art methods. The code is released at https://github.com/LiYuanBoJNU/SADCA.", "date": "2026-03-05", "categories": []}
{"id": "2603.04846", "title": "Multi-Paradigm Collaborative Adversarial Attack Against Multi-Modal Large Language Models", "abstract": "The rapid progress of Multi-Modal Large Language Models (MLLMs) has significantly advanced downstream applications. However, this progress also exposes serious transferable adversarial vulnerabilities. In general, existing adversarial attacks against MLLMs typically rely on surrogate models trained within a single learning paradigm and perform independent optimisation in their respective feature spaces. This straightforward setting naturally restricts the richness of feature representations, delivering limits on the search space and thus impeding the diversity of adversarial perturbations. To address this, we propose a novel Multi-Paradigm Collaborative Attack (MPCAttack) framework to boost the transferability of adversarial examples against MLLMs. In principle, MPCAttack aggregates semantic representations, from both visual images and language texts, to facilitate joint adversarial optimisation on the aggregated features through a Multi-Paradigm Collaborative Optimisation (MPCO) strategy. By performing contrastive matching on multi-paradigm features, MPCO adaptively balances the importance of different paradigm representations and guides the global perturbation optimisation, effectively alleviating the representation bias. Extensive experimental results on multiple benchmarks demonstrate the superiority of MPCAttack, indicating that our solution consistently outperforms state-of-the-art methods in both targeted and untargeted attacks on open-source and closed-source MLLMs. The code is released at https://github.com/LiYuanBoJNU/MPCAttack.", "date": "2026-03-05", "categories": []}
{"id": "2603.16960", "title": "Adversarial attacks against Modern Vision-Language Models", "abstract": "We study adversarial robustness of open-source vision-language model (VLM) agents deployed in a self-contained e-commerce environment built to simulate realistic pre-deployment conditions. We evaluate two agents, LLaVA-v1.5-7B and Qwen2.5-VL-7B, under three gradient-based attacks: the Basic Iterative Method (BIM), Projected Gradient Descent (PGD), and a CLIP-based spectral attack. Against LLaVA, all three attacks achieve substantial attack success rates (52.6%, 53.8%, and 66.9% respectively), demonstrating that simple gradient-based methods pose a practical threat to open-source VLM agents. Qwen2.5-VL proves significantly more robust across all attacks (6.5%, 7.7%, and 15.5%), suggesting meaningful architectural differences in adversarial resilience between open-source VLM families. These findings have direct implications for the security evaluation of VLM agents prior to commercial deployment.", "date": "2026-03-17", "categories": []}
{"id": "2603.20390", "title": "CAMA: Exploring Collusive Adversarial Attacks in c-MARL", "abstract": "Cooperative multi-agent reinforcement learning (c-MARL) has been widely deployed in real-world applications, such as social robots, embodied intelligence, UAV swarms, etc. Nevertheless, many adversarial attacks still exist to threaten various c-MARL systems. At present, the studies mainly focus on single-adversary perturbation attacks and white-box adversarial attacks that manipulate agents' internal observations or actions. To address these limitations, we in this paper attempt to study collusive adversarial attacks through strategically organizing a set of malicious agents into three collusive attack modes: Collective Malicious Agents, Disguised Malicious Agents, and Spied Malicious Agents. Three novelties are involved: i) three collusive adversarial attacks are creatively proposed for the first time, and a unified framework CAMA for policy-level collusive attacks is designed; ii) the attack effectiveness is theoretically analyzed from the perspectives of disruptiveness, stealthiness, and attack cost; and iii) the three collusive adversarial attacks are technically realized through agent's observation information fusion, attack-trigger control. Finally, multi-facet experiments on four SMAC II maps are performed, and experimental results showcase the three collusive attacks have an additive adversarial synergy, strengthening attack outcome while maintaining high stealthiness and stability over long horizons. Our work fills the gap for collusive adversarial learning in c-MARL.", "date": "2026-03-20", "categories": []}
{"id": "2603.20746", "title": "Adversarial Attacks on Locally Private Graph Neural Networks", "abstract": "Graph neural network (GNN) is a powerful tool for analyzing graph-structured data. However, their vulnerability to adversarial attacks raises serious concerns, especially when dealing with sensitive information. Local Differential Privacy (LDP) offers a privacy-preserving framework for training GNNs, but its impact on adversarial robustness remains underexplored. This paper investigates adversarial attacks on LDP-protected GNNs. We explore how the privacy guarantees of LDP can be leveraged or hindered by adversarial perturbations. The effectiveness of existing attack methods on LDP-protected GNNs are analyzed and potential challenges in crafting adversarial examples under LDP constraints are discussed. Additionally, we suggest directions for defending LDP-protected GNNs against adversarial attacks. This work investigates the interplay between privacy and security in graph learning, highlighting the need for robust and privacy-preserving GNN architectures.", "date": "2026-03-21", "categories": []}
{"id": "2603.20777", "title": "OmniPatch: A Universal Adversarial Patch for ViT-CNN Cross-Architecture Transfer in Semantic Segmentation", "abstract": "Robust semantic segmentation is crucial for safe autonomous driving, yet deployed models remain vulnerable to black-box adversarial attacks when target weights are unknown. Most existing approaches either craft image-wide perturbations or optimize patches for a single architecture, which limits their practicality and transferability. We introduce OmniPatch, a training framework for learning a universal adversarial patch that generalizes across images and both ViT and CNN architectures without requiring access to target model parameters.", "date": "2026-03-21", "categories": []}
{"id": "2603.21155", "title": "Can LLMs Fool Graph Learning? Exploring Universal Adversarial Attacks on Text-Attributed Graphs", "abstract": "Text-attributed graphs (TAGs) enhance graph learning by integrating rich textual semantics and topological context for each node. While boosting expressiveness, they also expose new vulnerabilities in graph learning through text-based adversarial surfaces. Recent advances leverage diverse backbones, such as graph neural networks (GNNs) and pre-trained language models (PLMs), to capture both structural and textual information in TAGs. This diversity raises a key question: How can we design universal adversarial attacks that generalize across architectures to assess the security of TAG models? The challenge arises from the stark contrast in how different backbones-GNNs and PLMs-perceive and encode graph patterns, coupled with the fact that many PLMs are only accessible via APIs, limiting attacks to black-box settings. To address this, we propose BadGraph, a novel attack framework that deeply elicits large language models (LLMs) understanding of general graph knowledge to jointly perturb both node topology and textual semantics. Specifically, we design a target influencer retrieval module that leverages graph priors to construct cross-modally aligned attack shortcuts, thereby enabling efficient LLM-based perturbation reasoning. Experiments show that BadGraph achieves universal and effective attacks across GNN- and LLM-based reasoners, with up to a 76.3% performance drop, while theoretical and empirical analyses confirm its stealthy yet interpretable nature.", "date": "2026-03-22", "categories": []}
{"id": "2603.21194", "title": "Is Monitoring Enough? Strategic Agent Selection For Stealthy Attack in Multi-Agent Discussions", "abstract": "Multi-agent discussions have been widely adopted, motivating growing efforts to develop attacks that expose their vulnerabilities. In this work, we study a practical yet largely unexplored attack scenario, the discussion-monitored scenario, where anomaly detectors continuously monitor inter-agent communications and block detected adversarial messages. Although existing attacks are effective without discussion monitoring, we show that they exhibit detectable patterns and largely fail under such monitoring constraints. But does this imply that monitoring alone is sufficient to secure multi-agent discussions? To answer this question, we develop a novel attack method explicitly tailored to the discussion-monitored scenario. Extensive experiments demonstrate that effective attacks remain possible even under continuous monitoring, indicating that monitoring alone does not eliminate adversarial risks.", "date": "2026-03-22", "categories": []}
{"id": "2603.21411", "title": "Fingerprinting Deep Neural Networks for Ownership Protection: An Analytical Approach", "abstract": "Adversarial-example-based fingerprinting approaches, which leverage the decision boundary characteristics of deep neural networks (DNNs) to craft fingerprints, have proven effective for model ownership protection. However, a fundamental challenge remains unresolved: how far a fingerprint should be placed from the decision boundary to simultaneously satisfy two essential properties, i.e., robustness and uniqueness, for effective and reliable ownership protection. Despite the importance of the fingerprint-to-boundary distance, existing works lack a theoretical solution and instead rely on empirical heuristics, which may violate either robustness or uniqueness properties.\n  We propose AnaFP, an analytical fingerprinting scheme that constructs fingerprints under theoretical guidance. Specifically, we formulate fingerprint generation as controlling the fingerprint-to-boundary distance through a tunable stretch factor. To ensure both robustness and uniqueness, we mathematically formalize these properties that determine the lower and upper bounds of the stretch factor. These bounds jointly define an admissible interval within which the stretch factor must lie, thereby establishing a theoretical connection between the two constraints and the fingerprint-to-boundary distance. To enable practical fingerprint generation, we approximate the original (infinite) sets of pirated and independently trained models using two finite surrogate model pools and employ a quantile-based relaxation strategy to relax the derived bounds. Due to the circular dependency between the lower bound and the stretch factor, we apply grid search over the admissible interval to determine the most feasible stretch factor. Extensive experimental results show that AnaFP consistently outperforms prior methods, achieving effective ownership verification across diverse model architectures and model modification attacks.", "date": "2026-03-22", "categories": []}
{"id": "2502.19095", "title": "Cross-site scripting adversarial attacks based on deep reinforcement learning: Evaluation and extension study", "abstract": "Cross-site scripting (XSS) poses a significant threat to web application security. While Deep Learning (DL) has shown remarkable success in detecting XSS attacks, it remains vulnerable to adversarial attacks due to the discontinuous nature of the mapping between the input (i.e., the attack) and the output (i.e., the prediction of the model whether an input is classified as XSS or benign). These adversarial attacks employ mutation-based strategies for different components of XSS attack vectors, allowing adversarial agents to iteratively select mutations to evade detection. Our work replicates a state-of-the-art XSS adversarial attack, highlighting threats to validity in the reference work and extending it towards a more effective evaluation strategy. Moreover, we introduce an XSS Oracle to mitigate these threats. The experimental results show that our approach achieves an escape rate above 96% when the threats to validity of the replicated technique are addressed.", "date": "2025-02-26", "categories": []}
{"id": "2510.12752", "title": "KoALA: KL-L0 Adversarial Detector via Label Agreement", "abstract": "Deep neural networks are highly susceptible to adversarial attacks, which pose significant risks to security- and safety-critical applications. We present KoALA (KL-L0 Adversarial detection via Label Agreement), a novel, semantics-free adversarial detector that requires no architectural changes or adversarial retraining. KoALA operates on a simple principle: it detects an adversarial attack when class predictions from two complementary similarity metrics disagree. These metrics - KL divergence and an L0-based similarity - are specifically chosen to detect different types of perturbations. The KL divergence metric is sensitive to dense, low-amplitude shifts, while the L0-based similarity is designed for sparse, high-impact changes. We provide a formal proof of correctness for our approach. The only training required is a simple fine-tuning step on a pre-trained image encoder using clean images to ensure the embeddings align well with both metrics. This makes KoALA a lightweight, plug-and-play solution for existing models and various data modalities. Our extensive experiments on ResNet/CIFAR-10 and CLIP/Tiny-ImageNet confirm our theoretical claims. When the theorem's conditions are met, KoALA consistently and effectively detects adversarial examples. On the full test sets, KoALA achieves a precision of 0.96 and a recall of 0.97 on ResNet/CIFAR-10, and a precision of 0.71 and a recall of 0.94 on CLIP/Tiny-ImageNet.", "date": "2025-10-14", "categories": []}
{"id": "2602.08934", "title": "StealthRL: Reinforcement Learning Paraphrase Attacks for Multi-Detector Evasion of AI-Text Detectors", "abstract": "AI-text detectors face a critical robustness challenge: adversarial paraphrasing attacks that preserve semantics while evading detection. We introduce StealthRL, a reinforcement learning framework that stress-tests detector robustness under realistic adversarial conditions. StealthRL trains a paraphrase policy against a multi-detector ensemble using Group Relative Policy Optimization (GRPO) with LoRA adapters on Qwen3-4B, optimizing a composite reward that balances detector evasion with semantic preservation. We evaluate six attack settings (M0-M5) on the full filtered MAGE test pool (15,310 human / 14,656 AI) against four detectors: RoBERTa, Fast-DetectGPT, Binoculars, and MAGE. StealthRL achieves near-zero detection on three of the four detectors and a 0.024 mean TPR@1%FPR, reducing mean AUROC from 0.79 to 0.43 and attaining a 97.6% attack success rate. Critically, attacks transfer to two held-out detectors not seen during training, revealing shared architectural vulnerabilities rather than detector-specific brittleness. We additionally conduct LLM-based quality evaluation via Likert scoring on 500 matched samples per method, analyze detector score distributions to explain why evasion succeeds, and provide per-detector AUROC with bootstrap confidence intervals. Our results expose significant robustness gaps in current AI-text detection and establish StealthRL as a principled adversarial evaluation protocol. Code and evaluation pipeline are publicly available at https://github.com/suraj-ranganath/StealthRL.", "date": "2026-02-09", "categories": []}
{"id": "2603.19456", "title": "In-the-Wild Camouflage Attack on Vehicle Detectors through Controllable Image Editing", "abstract": "Deep neural networks (DNNs) have achieved remarkable success in computer vision but remain highly vulnerable to adversarial attacks. Among them, camouflage attacks manipulate an object's visible appearance to deceive detectors while remaining stealthy to humans. In this paper, we propose a new framework that formulates vehicle camouflage attacks as a conditional image-editing problem. Specifically, we explore both image-level and scene-level camouflage generation strategies, and fine-tune a ControlNet to synthesize camouflaged vehicles directly on real images. We design a unified objective that jointly enforces vehicle structural fidelity, style consistency, and adversarial effectiveness. Extensive experiments on the COCO and LINZ datasets show that our method achieves significantly stronger attack effectiveness, leading to more than 38% AP50 decrease, while better preserving vehicle structure and improving human-perceived stealthiness compared to existing approaches. Furthermore, our framework generalizes effectively to unseen black-box detectors and exhibits promising transferability to the physical world. Project page is available at https://humansensinglab.github.io/CtrlCamo", "date": "2026-03-19", "categories": []}
{"id": "2603.20108", "title": "Trojan horse hunt in deep forecasting models: Insights from the European Space Agency competition", "abstract": "Forecasting plays a crucial role in modern safety-critical applications, such as space operations. However, the increasing use of deep forecasting models introduces a new security risk of trojan horse attacks, carried out by hiding a backdoor in the training data or directly in the model weights. Once implanted, the backdoor is activated by a specific trigger pattern at test time, causing the model to produce manipulated predictions. We focus on this issue in our \\textit{Trojan Horse Hunt} data science competition, where more than 200 teams faced the task of identifying triggers hidden in deep forecasting models for spacecraft telemetry. We describe the novel task formulation, benchmark set, evaluation protocol, and best solutions from the competition. We further summarize key insights and research directions for effective identification of triggers in time series forecasting models. All materials are publicly available on the official competition webpage https://www.kaggle.com/competitions/trojan-horse-hunt-in-space.", "date": "2026-03-20", "categories": []}
{"id": "2309.08945", "title": "Inverse classification with logistic and softmax classifiers: efficient optimization", "abstract": "In recent years, a certain type of problems have become of interest where one wants to query a trained classifier. Specifically, one wants to find the closest instance to a given input instance such that the classifier's predicted label is changed in a desired way. Examples of these \"inverse classification\" problems are counterfactual explanations, adversarial examples and model inversion. All of them are fundamentally optimization problems over the input instance vector involving a fixed classifier, and it is of interest to achieve a fast solution for interactive or real-time applications. We focus on solving this problem efficiently with the squared Euclidean distance for two of the most widely used classifiers: logistic regression and softmax classifier. Owing to special properties of these models, we show that the optimization can be solved in closed form for logistic regression, and iteratively but extremely fast for the softmax classifier. This allows us to solve either case exactly (to nearly machine precision) in a runtime of milliseconds to around a second even for very high-dimensional instances and many classes.", "date": "2023-09-16", "categories": []}
{"id": "2505.21854", "title": "Rethinking Gradient-based Adversarial Attacks on Point Cloud Classification", "abstract": "Gradient-based adversarial attacks are widely used to evaluate the robustness of 3D point cloud classifiers, yet they often rely on uniform update rules that neglect point-wise heterogeneity, leading to perceptible perturbations. We propose two complementary strategies to improve both the effectiveness and imperceptibility of the attack. \\textbf{WAAttack} employs weighted gradients to dynamically adjust per-point perturbation magnitudes and uses an adaptive step size strategy to regulate the global perturbation scale. \\textbf{SubAttack} partitions the point cloud into subsets and, at each iteration, perturbs only those combinations with high adversarial efficacy and low perceptual saliency. Together, these methods offer a principled refinement of gradient-based attacks for 3D point clouds. Extensive experiments show that our approach consistently outperforms state-of-the-art methods in generating highly imperceptible adversarial examples. The code is available at https://github.com/chenjun0326/WA_SubAttack.", "date": "2025-05-28", "categories": []}
{"id": "2603.07313", "title": "Adversarial Latent-State Training for Robust Policies in Partially Observable Domains", "abstract": "Robustness under latent distribution shift remains challenging in partially observable reinforcement learning. We formalize a focused setting where an adversary selects a hidden initial latent distribution before the episode, termed an adversarial latent-initial-state POMDP. Theoretically, we prove a latent minimax principle, characterize worst-case defender distributions, and derive approximate best-response inequalities with finite-sample concentration bounds that make the optimization and sampling terms explicit. Empirically, using a Battleship benchmark, we demonstrate that targeted exposure to shifted latent distributions reduces average robustness gaps between Spread and Uniform distributions from 10.3 to 3.1 shots at equal budget. Furthermore, iterative best-response training exhibits budget-sensitive behavior that is qualitatively consistent with the theorem-guided diagnostics once one accounts for discounted PPO surrogates and finite-sample noise. Ultimately, we show that for latent-initial-state problems, the framework yields a clean evaluation game and useful theorem-motivated diagnostics while also making clear where implementation-level surrogates and optimization limits enter.", "date": "2026-03-07", "categories": []}
{"id": "2603.18570", "title": "Attack by Unlearning: Unlearning-Induced Adversarial Attacks on Graph Neural Networks", "abstract": "Graph neural networks (GNNs) are widely used for learning from graph-structured data in domains such as social networks, recommender systems, and financial platforms. To comply with privacy regulations like the GDPR, CCPA, and PIPEDA, approximate graph unlearning, which aims to remove the influence of specific data points from trained models without full retraining, has become an increasingly important component of trustworthy graph learning. However, approximate unlearning often incurs subtle performance degradation, which may incur negative and unintended side effects. In this work, we show that such degradations can be amplified into adversarial attacks. We introduce the notion of \\textbf{unlearning corruption attacks}, where an adversary injects carefully chosen nodes into the training graph and later requests their deletion. Because deletion requests are legally mandated and cannot be denied, this attack surface is both unavoidable and stealthy: the model performs normally during training, but accuracy collapses only after unlearning is applied. Technically, we formulate this attack as a bi-level optimization problem: to overcome the challenges of black-box unlearning and label scarcity, we approximate the unlearning process via gradient-based updates and employ a surrogate model to generate pseudo-labels for the optimization. Extensive experiments across benchmarks and unlearning algorithms demonstrate that small, carefully designed unlearning requests can induce significant accuracy degradation, raising urgent concerns about the robustness of GNN unlearning under real-world regulatory demands. The source code will be released upon paper acceptance.", "date": "2026-03-19", "categories": []}
{"id": "2603.18598", "title": "Complementary Text-Guided Attention for Zero-Shot Adversarial Robustness", "abstract": "Due to the impressive zero-shot capabilities, pre-trained vision-language models (e.g., CLIP), have attracted widespread attention and adoption across various domains. Nonetheless, CLIP has been observed to be susceptible to adversarial examples. Through experimental analysis, we have observed a phenomenon wherein adversarial perturbations induce shifts in text-guided attention. Building upon this observation, we propose a simple yet effective strategy: Text-Guided Attention for Zero-Shot Robustness (TGA-ZSR). This framework incorporates two components: Local Attention Refinement Module and Global Attention Constraint Module. Our goal is to maintain the generalization of the CLIP model and enhance its adversarial robustness. Additionally, the Global Attention Constraint Module acquires text-guided attention from both the target and original models using clean examples. Its objective is to maintain model performance on clean samples while enhancing overall robustness. However, we observe that the method occasionally focuses on irrelevant or spurious features, which can lead to suboptimal performance and undermine its robustness in certain scenarios. To overcome this limitation, we further propose a novel approach called Complementary Text-Guided Attention (Comp-TGA). This method integrates two types of foreground attention: attention guided by the class prompt and reversed attention driven by the non-class prompt. These complementary attention mechanisms allow the model to capture a more comprehensive and accurate representation of the foreground. The experiments validate that TGA-ZSR and Comp-TGA yield 9.58% and 11.95% improvements respectively, in zero-shot robust accuracy over the current state-of-the-art techniques across 16 datasets.", "date": "2026-03-19", "categories": []}
{"id": "2603.19127", "title": "On Optimizing Multimodal Jailbreaks for Spoken Language Models", "abstract": "As Spoken Language Models (SLMs) integrate speech and text modalities, they inherit the safety vulnerabilities of their LLM backbone and an expanded attack surface. SLMs have been previously shown to be susceptible to jailbreaking, where adversarial prompts induce harmful responses. Yet existing attacks largely remain unimodal, optimizing either text or audio in isolation. We explore gradient-based multimodal jailbreaks by introducing JAMA (Joint Audio-text Multimodal Attack), a joint multimodal optimization framework combining Greedy Coordinate Gradient (GCG) for text and Projected Gradient Descent (PGD) for audio, to simultaneously perturb both modalities. Evaluations across four state-of-the-art SLMs and four audio types demonstrate that JAMA surpasses unimodal jailbreak rate by 1.5x to 10x. We analyze the operational dynamics of this joint attack and show that a sequential approximation method makes it 4x to 6x faster. Our findings suggest that unimodal safety is insufficient for robust SLMs. The code and data are available at https://repos.lsv.uni-saarland.de/akrishnan/multimodal-jailbreak-slm", "date": "2026-03-19", "categories": []}
{"id": "2603.19204", "title": "Robustness, Cost, and Attack-Surface Concentration in Phishing Detection", "abstract": "Phishing detectors built on engineered website features attain near-perfect accuracy under i.i.d.\\ evaluation, yet deployment security depends on robustness to post-deployment feature manipulation. We study this gap through a cost-aware evasion framework that models discrete, monotone feature edits under explicit attacker budgets. Three diagnostics are introduced: minimal evasion cost (MEC), the evasion survival rate $S(B)$, and the robustness concentration index (RCI).\n  On the UCI Phishing Websites benchmark (11\\,055 instances, 30 ternary features), Logistic Regression, Random Forests, Gradient Boosted Trees, and XGBoost all achieve $\\mathrm{AUC}\\ge 0.979$ under static evaluation. Under budgeted sanitization-style evasion, robustness converges across architectures: the median MEC equals 2 with full features, and over 80\\% of successful minimal-cost evasions concentrate on three low-cost surface features. Feature restriction improves robustness only when it removes all dominant low-cost transitions. Under strict cost schedules, infrastructure-leaning feature sets exhibit 17-19\\% infeasible mass for ensemble models, while the median MEC among evadable instances remains unchanged. We formalize this convergence: if a positive fraction of correctly detected phishing instances admit evasion through a single feature transition of minimal cost $c_{\\min}$, no classifier can raise the corresponding MEC quantile above $c_{\\min}$ without modifying the feature representation or cost model. Adversarial robustness in phishing detection is governed by feature economics rather than model complexity.", "date": "2026-03-19", "categories": []}
{"id": "2504.04893", "title": "SCAM: A Real-World Typographic Robustness Evaluation for Multimodal Foundation Models", "abstract": "Typographic attacks exploit the interplay between text and visual content in multimodal foundation models, causing misclassifications when misleading text is embedded within images. Existing datasets are limited in size and diversity, making it difficult to study such vulnerabilities. In this paper, we introduce SCAM, the largest and most diverse dataset of real-world typographic attack images to date, containing 1162 images across hundreds of object categories and attack words. Through extensive benchmarking of Vision-Language Models on SCAM, we demonstrate that typographic attacks significantly degrade performance, and identify that training data and model architecture influence the susceptibility to these attacks. Our findings indicate that typographic attacks remain effective against state-of-the-art Large Vision-Language Models, especially those employing vision encoders inherently vulnerable to such attacks. However, employing larger Large Language Model backbones reduces this vulnerability while simultaneously enhancing typographic understanding. Additionally, we demonstrate that synthetic attacks closely resemble real-world (handwritten) attacks, validating their use in research. Our work provides a comprehensive resource and empirical insights to facilitate future research toward robust and trustworthy multimodal AI systems. Finally, we publicly release the datasets introduced in this paper, along with the code for evaluations under www.bliss.berlin/research/scam.", "date": "2025-04-07", "categories": []}
{"id": "2603.16972", "title": "Over-the-air White-box Attack on the Wav2Vec Speech Recognition Neural Network", "abstract": "Automatic speech recognition systems based on neural networks are vulnerable to adversarial attacks that alter transcriptions in a malicious way. Recent works in this field have focused on making attacks work in over-the-air scenarios, however such attacks are typically detectable by human hearing, limiting their potential applications. In the present work we explore different approaches of making over-the-air attacks less detectable, as well as the impact these approaches have on the attacks' effectiveness.", "date": "2026-03-17", "categories": []}
{"id": "2603.17176", "title": "Towards Unsupervised Adversarial Document Detection in Retrieval Augmented Generation Systems", "abstract": "Retrieval augmented generation systems have become an integral part of everyday life. Whether in internet search engines, email systems, or service chatbots, these systems are based on context retrieval and answer generation with large language models. With their spread, also the security vulnerabilities increase. Attackers become increasingly focused on these systems and various hacking approaches are developed. Manipulating the context documents is a way to persist attacks and make them affect all users. Therefore, detecting compromised, adversarial context documents early is crucial for security. While supervised approaches require a large amount of labeled adversarial contexts, we propose an unsupervised approach, being able to detect also zero day attacks. We conduct a preliminary study to show appropriate indicators for adversarial contexts. For that purpose generator activations, output embeddings, and an entropy-based uncertainty measure turn out as suitable, complementary quantities. With an elementary statistical outlier detection, we propose and compare their detection abilities. Furthermore, we show that the target prompt, which the attacker wants to manipulate, is not required for a successful detection. Moreover, our results indicate that a simple context summary generation might even be superior in finding manipulated contexts.", "date": "2026-03-17", "categories": []}
{"id": "2603.17828", "title": "TINA: Text-Free Inversion Attack for Unlearned Text-to-Image Diffusion Models", "abstract": "Although text-to-image diffusion models exhibit remarkable generative power, concept erasure techniques are essential for their safe deployment to prevent the creation of harmful content. This has fostered a dynamic interplay between the development of erasure defenses and the adversarial probes designed to bypass them, and this co-evolution has progressively enhanced the efficacy of erasure methods. However, this adversarial co-evolution has converged on a narrow, text-centric paradigm that equates erasure with severing the text-to-image mapping, ignoring that the underlying visual knowledge related to undesired concepts still persist. To substantiate this claim, we investigate from a visual perspective, leveraging DDIM inversion to probe whether a generative pathway for the erased concept can still be found. However, identifying such a visual generative pathway is challenging because standard text-guided DDIM inversion is actively resisted by text-centric defenses within the erased model. To address this, we introduce TINA, a novel Text-free INversion Attack, which enforces this visual-only probe by operating under a null-text condition, thereby avoiding existing text-centric defenses. Moreover, TINA integrates an optimization procedure to overcome the accumulating approximation errors that arise when standard inversion operates without its usual textual guidance. Our experiments demonstrate that TINA regenerates erased concepts from models treated with state-of-the-art unlearning. The success of TINA proves that current methods merely obscure concepts, highlighting an urgent need for paradigms that operate directly on internal visual knowledge.", "date": "2026-03-18", "categories": []}
{"id": "2603.17914", "title": "Noise-Aware Misclassification Attack Detection in Collaborative DNN Inference", "abstract": "Collaborative inference of object classification Deep neural Networks (DNNs) where resource-constrained end-devices offload partially processed data to remote edge servers to complete end-to-end processing, is becoming a key enabler of edge-AI. However, such edge-offloading is vulnerable to malicious data injections leading to stealthy misclassifications that are tricky to detect, especially in the presence of environmental noise. In this paper, we propose a semi-gray-box and noise- aware anomaly detection framework fueled by a variational autoencoder (VAE) to capture deviations caused by adversarial manipulation. The proposed framework incorporates a robust noise-aware feature that captures the characteristic behavior of environmental noise to improve detection accuracy while reducing false alarm rates. Our evaluation with popular object classification DNNs demonstrate the robustness of the proposed detection (up to 90% AUROC across DNN configurations) under realistic noisy conditions while revealing limitations caused by feature similarity and elevated noise levels.", "date": "2026-03-18", "categories": []}
{"id": "2503.06140", "title": "Boosting the Local Invariance for Better Adversarial Transferability", "abstract": "Transfer-based attacks pose a significant threat to real-world applications by directly targeting victim models with adversarial examples generated on surrogate models. While numerous approaches have been proposed to enhance adversarial transferability, existing works often overlook the intrinsic relationship between adversarial perturbations and input images. In this work, we find that adversarial perturbation often exhibits poor translation invariance for a given clean image and model, which is attributed to local invariance. Through empirical analysis, we demonstrate a positive correlation between the local invariance of adversarial perturbations w.r.t. the input image and their transferability across models. Based on this finding, we propose a general adversarial transferability boosting technique called the Local Invariance Boosting approach (LI-Boost). Extensive experiments on the standard ImageNet dataset demonstrate that LI-Boost can significantly enhance various transfer-based attacks (e.g., gradient-based, input transformation-based, model-related, advanced objective function, ensemble, etc.) on CNNs, ViTs, defense mechanisms, commercial vision API systems, and vision-language models. Our approach provides a promising direction for future research on improving adversarial transferability across models. Our code is available at https://github.com/Trustworthy-AI-Group/TransferAttack.", "date": "2025-03-08", "categories": []}
{"id": "2602.19570", "title": "VALD: Multi-Stage Vision Attack Detection for Efficient LVLM Defense", "abstract": "Large Vision-Language Models (LVLMs) can be vulnerable to adversarial images that subtly bias their outputs toward plausible yet incorrect responses. We introduce a general, efficient, and training-free defense that combines image transformations with agentic data consolidation to recover correct model behavior. A key component of our approach is a two-stage detection mechanism that quickly filters out the majority of clean inputs. We first assess image consistency under content-preserving transformations at negligible computational cost. For more challenging cases, we examine discrepancies in a text-embedding space. Only when necessary do we invoke a powerful LLM to resolve attack-induced divergences. A key idea is to consolidate multiple responses, leveraging both their similarities and their differences. We show that our method achieves state-of-the-art accuracy while maintaining notable efficiency: most clean images skip costly processing, and even in the presence of numerous adversarial examples, the overhead remains minimal.", "date": "2026-02-23", "categories": []}
{"id": "2603.15714", "title": "How Vulnerable Are AI Agents to Indirect Prompt Injections? Insights from a Large-Scale Public Competition", "abstract": "LLM based agents are increasingly deployed in high stakes settings where they process external data sources such as emails, documents, and code repositories. This creates exposure to indirect prompt injection attacks, where adversarial instructions embedded in external content manipulate agent behavior without user awareness. A critical but underexplored dimension of this threat is concealment: since users tend to observe only an agent's final response, an attack can conceal its existence by presenting no clue of compromise in the final user facing response while successfully executing harmful actions. This leaves users unaware of the manipulation and likely to accept harmful outcomes as legitimate. We present findings from a large scale public red teaming competition evaluating this dual objective across three agent settings: tool calling, coding, and computer use. The competition attracted 464 participants who submitted 272000 attack attempts against 13 frontier models, yielding 8648 successful attacks across 41 scenarios. All models proved vulnerable, with attack success rates ranging from 0.5% (Claude Opus 4.5) to 8.5% (Gemini 2.5 Pro). We identify universal attack strategies that transfer across 21 of 41 behaviors and multiple model families, suggesting fundamental weaknesses in instruction following architectures. Capability and robustness showed weak correlation, with Gemini 2.5 Pro exhibiting both high capability and high vulnerability. To address benchmark saturation and obsoleteness, we will endeavor to deliver quarterly updates through continued red teaming competitions. We open source the competition environment for use in evaluations, along with 95 successful attacks against Qwen that did not transfer to any closed source model. We share model-specific attack data with respective frontier labs and the full dataset with the UK AISI and US CAISI to support robustness research.", "date": "2026-03-16", "categories": []}
{"id": "2410.08950", "title": "On the Adversarial Transferability of Generalized \"Skip Connections\"", "abstract": "Skip connection is an essential ingredient for modern deep models to be deeper and more powerful. Despite their huge success in normal scenarios (state-of-the-art classification performance on natural examples), we investigate and identify an interesting property of skip connections under adversarial scenarios, namely, the use of skip connections allows easier generation of highly transferable adversarial examples. Specifically, in ResNet-like models (with skip connections), we find that biasing backpropagation to favor gradients from skip connections--while suppressing those from residual modules via a decay factor--allows one to craft adversarial examples with high transferability. Based on this insight, we propose the Skip Gradient Method (SGM). Although starting from ResNet-like models in vision domains, we further extend SGM to more advanced architectures, including Vision Transformers (ViTs), models with varying-length paths, and other domains such as natural language processing. We conduct comprehensive transfer-based attacks against diverse model families, including ResNets, Transformers, Inceptions, Neural Architecture Search-based models, and Large Language Models (LLMs). The results demonstrate that employing SGM can greatly improve the transferability of crafted attacks in almost all cases. Furthermore, we demonstrate that SGM can still be effective under more challenging settings such as ensemble-based attacks, targeted attacks, and against defense equipped models. At last, we provide theoretical explanations and empirical insights on how SGM works. Our findings not only motivate new adversarial research into the architectural characteristics of models but also open up further challenges for secure model architecture design. Our code is available at https://github.com/mo666666/SGM.", "date": "2024-10-11", "categories": []}
{"id": "2412.18507", "title": "An Empirical Analysis of Federated Learning Models Subject to Label-Flipping Adversarial Attack", "abstract": "In this paper, we empirically analyze adversarial attacks on selected federated learning models. The specific learning models considered are Multinominal Logistic Regression (MLR), Support Vector Classifier (SVC), Multilayer Perceptron (MLP), Convolution Neural Network (CNN), %Recurrent Neural Network (RNN), Random Forest, XGBoost, and Long Short-Term Memory (LSTM). For each model, we simulate label-flipping attacks, experimenting extensively with 10 federated clients and 100 federated clients. We vary the percentage of adversarial clients from 10% to 100% and, simultaneously, the percentage of labels flipped by each adversarial client is also varied from 10% to 100%. Among other results, we find that models differ in their inherent robustness to the two vectors in our label-flipping attack, i.e., the percentage of adversarial clients, and the percentage of labels flipped by each adversarial client. We discuss the potential practical implications of our results.", "date": "2024-12-24", "categories": []}
{"id": "2502.02290", "title": "FRAUD-RLA: A new reinforcement learning adversarial attack against credit card fraud detection", "abstract": "Adversarial attacks pose a significant threat to data-driven systems, and researchers have spent considerable resources studying them. Despite its economic relevance, this trend largely overlooked the issue of credit card fraud detection. To address this gap, we propose a new threat model that demonstrates the limitations of existing attacks and highlights the necessity to investigate new approaches. We then design a new adversarial attack for credit card fraud detection, employing reinforcement learning to bypass classifiers. This attack, called FRAUD-RLA, is designed to maximize the attacker's reward by optimizing the exploration-exploitation tradeoff and working with significantly less required knowledge than competitors. Our experiments, conducted on three different heterogeneous datasets and against two fraud detection systems, indicate that FRAUD-RLA is effective, even considering the severe limitations imposed by our threat model.", "date": "2025-02-04", "categories": []}
{"id": "2504.17884", "title": "Unsupervised Corpus Poisoning Attacks in Continuous Space for Dense Retrieval", "abstract": "This paper concerns corpus poisoning attacks in dense information retrieval, where an adversary attempts to compromise the ranking performance of a search algorithm by injecting a small number of maliciously generated documents into the corpus. Our work addresses two limitations in the current literature. First, attacks that perform adversarial gradient-based word substitution search do so in the discrete lexical space, while retrieval itself happens in the continuous embedding space. We thus propose an optimization method that operates in the embedding space directly. Specifically, we train a perturbation model with the objective of maintaining the geometric distance between the original and adversarial document embeddings, while also maximizing the token-level dissimilarity between the original and adversarial documents. Second, it is common for related work to have a strong assumption that the adversary has prior knowledge about the queries. In this paper, we focus on a more challenging variant of the problem where the adversary assumes no prior knowledge about the query distribution (hence, unsupervised). Our core contribution is an adversarial corpus attack that is fast and effective. We present comprehensive experimental results on both in- and out-of-domain datasets, focusing on two related tasks: a top-1 attack and a corpus poisoning attack. We consider attacks under both a white-box and a black-box setting. Notably, our method can generate successful adversarial examples in under two minutes per target document; four times faster compared to the fastest gradient-based word substitution methods in the literature with the same hardware. Furthermore, our adversarial generation method generates text that is more likely to occur under the distribution of natural text (low perplexity), and is therefore more difficult to detect.", "date": "2025-04-24", "categories": []}
{"id": "2505.23404", "title": "AJF: Adaptive Jailbreak Framework Based on the Comprehension Ability of Black-Box Large Language Models", "abstract": "Recent advancements in adversarial jailbreak attacks have exposed critical vulnerabilities in Large Language Models (LLMs), enabling the circumvention of alignment safeguards through increasingly sophisticated prompt manipulations. Our experiments find that the effectiveness of jailbreak strategies is influenced by the comprehension ability of the target LLM. Building on this insight, we propose an Adaptive Jailbreak Framework (AJF) based on the comprehension ability of black-box large language models. Specifically, AJF first categorizes the comprehension ability of the LLM and then applies different strategies accordingly: For models with limited comprehension ability (Type-I LLMs), AJF integrates layered semantic mutations with an encryption technique (MuEn strategy), to more effectively evade the LLM's defenses during the input and inference stages. For models with strong comprehension ability (Type-II LLMs), AJF employs a more complex strategy that builds upon the MuEn strategy by adding an additional layer: inducing the LLM to generate an encrypted response. This forms a dual-end encryption scheme (MuDeEn strategy), further bypassing the LLM's defenses during the output stage. Experimental results demonstrate the effectiveness of our approach, achieving attack success rates of \\textbf{98.9\\%} on GPT-4o (29 May 2025 release) and \\textbf{99.8\\%} on GPT-4.1 (8 July 2025 release). Our work contributes to a deeper understanding of the vulnerabilities in current LLMs alignment mechanisms.", "date": "2025-05-29", "categories": []}
{"id": "2508.07795", "title": "Boosting Active Defense Persistence: A Two-Stage Defense Framework Combining Interruption and Poisoning Against Deepfake", "abstract": "Active defense strategies have been developed to counter the threat of deepfake technology. However, a primary challenge is their lack of persistence, as their effectiveness is often short-lived. Attackers can bypass these defenses by simply collecting protected samples and retraining their models. This means that static defenses inevitably fail when attackers retrain their models, which severely limits practical use. We argue that an effective defense not only distorts forged content but also blocks the model's ability to adapt, which occurs when attackers retrain their models on protected images. To achieve this, we propose an innovative Two-Stage Defense Framework (TSDF). Benefiting from the intensity separation mechanism designed in this paper, the framework uses dual-function adversarial perturbations to perform two roles. First, it can directly distort the forged results. Second, it acts as a poisoning vehicle that disrupts the data preparation process essential for an attacker's retraining pipeline. By poisoning the data source, TSDF aims to prevent the attacker's model from adapting to the defensive perturbations, thus ensuring the defense remains effective long-term. Comprehensive experiments show that the performance of traditional interruption methods degrades sharply when it is subjected to adversarial retraining. However, our framework shows a strong dual defense capability, which can improve the persistence of active defense. Our code will be available at https://github.com/vpsg-research/TSDF.", "date": "2025-08-11", "categories": []}
{"id": "2510.00517", "title": "Understanding Sensitivity of Differential Attention through the Lens of Adversarial Robustness", "abstract": "Differential Attention (DA) has been proposed as a refinement to standard attention, suppressing redundant or noisy context through a subtractive structure and thereby reducing contextual hallucination. While this design sharpens task-relevant focus, we show that it also introduces a structural fragility under adversarial perturbations. Our theoretical analysis identifies negative gradient alignment-a configuration encouraged by DA's subtraction-as the key driver of sensitivity amplification, leading to increased gradient norms and elevated local Lipschitz constants. We empirically validate this Fragile Principle through systematic experiments on ViT/DiffViT and evaluations of pretrained CLIP/DiffCLIP, spanning five datasets in total. These results demonstrate higher attack success rates, frequent gradient opposition, and stronger local sensitivity compared to standard attention. Furthermore, depth-dependent experiments reveal a robustness crossover: stacking DA layers attenuates small perturbations via depth-dependent noise cancellation, though this protection fades under larger attack budgets. Overall, our findings uncover a fundamental trade-off: DA improves discriminative focus on clean inputs but increases adversarial vulnerability, underscoring the need to jointly design for selectivity and robustness in future attention mechanisms.", "date": "2025-10-01", "categories": []}
{"id": "2511.07112", "title": "More Agents Improve Math Problem Solving but Adversarial Robustness Gap Persists", "abstract": "When LLM agents work together, they seem to be more powerful than a single LLM in mathematical question answering. However, are they also more robust to adversarial inputs? We investigate this question using adversarially perturbed math questions. These perturbations include punctuation noise with three intensities (10%, 30%, 50%), plus real-world and human-like typos (WikiTypo, R2ATA). Using a unified sampling-and-voting framework (Agent Forest), we evaluate six open-source models (Qwen3-4B/14B, Llama3.1-8B, Mistral-7B, Gemma3-4B/12B) across four benchmarks (GSM8K, MATH, MMLU-Math, MultiArith), with various numbers of agents n = {1,2,5,10,15,20,25}. Our findings show that 1) Noise type matters: punctuation noise harm scales with its severity, and the human typos remain the dominant bottleneck, yielding the largest gaps to Clean accuracy and the highest attack success rate (ASR) even with a large number of agents; 2) Collaboration reliably improves accuracy as the number of agents, n, increases, with the largest gains from n=1 to n=5 and diminishing returns beyond n$\\approx$10. However, the adversarial robustness gap persists regardless of the agent count.", "date": "2025-11-10", "categories": []}
{"id": "2511.14386", "title": "Cheating Stereo Matching in Full-scale: Physical Adversarial Attack against Binocular Depth Estimation in Autonomous Driving", "abstract": "Though deep neural models adopted to realize the perception of autonomous driving have proven vulnerable to adversarial examples, known attacks often leverage 2D patches and target mostly monocular perception. Therefore, the effectiveness of Physical Adversarial Examples (PAEs) on stereo-based binocular depth estimation remains largely unexplored. To this end, we propose the first texture-enabled physical adversarial attack against stereo matching models in the context of autonomous driving. Our method employs a 3D PAE with global camouflage texture rather than a local 2D patch-based one, ensuring both visual consistency and attack effectiveness across different viewpoints of stereo cameras. To cope with the disparity effect of these cameras, we also propose a new 3D stereo matching rendering module that allows the PAE to be aligned with real-world positions and headings in binocular vision. We further propose a novel merging attack that seamlessly blends the target into the environment through fine-grained PAE optimization. It has significantly enhanced stealth and lethality upon existing hiding attacks that fail to get seamlessly merged into the background. Extensive evaluations show that our PAEs can successfully fool the stereo models into producing erroneous depth information.", "date": "2025-11-18", "categories": []}
{"id": "2602.02569", "title": "DECEIVE-AFC: Adversarial Claim Attacks against Search-Enabled LLM-based Fact-Checking Systems", "abstract": "Fact-checking systems with search-enabled large language models (LLMs) have shown strong potential for verifying claims by dynamically retrieving external evidence. However, the robustness of such systems against adversarial attack remains insufficiently understood. In this work, we study adversarial claim attacks against search-enabled LLM-based fact-checking systems under a realistic input-only threat model. We propose DECEIVE-AFC, an agent-based adversarial attack framework that integrates novel claim-level attack strategies and adversarial claim validity evaluation principles. DECEIVE-AFC systematically explores adversarial attack trajectories that disrupt search behavior, evidence retrieval, and LLM-based reasoning without relying on access to evidence sources or model internals. Extensive evaluations on benchmark datasets and real-world systems demonstrate that our attacks substantially degrade verification performance, reducing accuracy from 78.7% to 53.7%, and significantly outperform existing claim-based attack baselines with strong cross-system transferability.", "date": "2026-01-31", "categories": []}
{"id": "2603.11501", "title": "KEPo: Knowledge Evolution Poison on Graph-based Retrieval-Augmented Generation", "abstract": "Graph-based Retrieval-Augmented Generation (GraphRAG) constructs the Knowledge Graph (KG) from external databases to enhance the timeliness and accuracy of Large Language Model (LLM) generations. However, this reliance on external data introduces new attack surfaces. Attackers can inject poisoned texts into databases to manipulate LLMs into producing harmful target responses for attacker-chosen queries. Existing research primarily focuses on attacking conventional RAG systems. However, such methods are ineffective against GraphRAG. This robustness derives from the KG abstraction of GraphRAG, which reorganizes injected text into a graph before retrieval, thereby enabling the LLM to reason based on the restructured context instead of raw poisoned passages. To expose latent security vulnerabilities in GraphRAG, we propose Knowledge Evolution Poison (KEPo), a novel poisoning attack method specifically designed for GraphRAG. For each target query, KEPo first generates a toxic event containing poisoned knowledge based on the target answer. By fabricating event backgrounds and forging knowledge evolution paths from original facts to the toxic event, it then poisons the KG and misleads the LLM into treating the poisoned knowledge as the final result. In multi-target attack scenarios, KEPo further connects multiple attack corpora, enabling their poisoned knowledge to mutually reinforce while expanding the scale of poisoned communities, thereby amplifying attack effectiveness. Experimental results across multiple datasets demonstrate that KEPo achieves state-of-the-art attack success rates for both single-target and multi-target attacks, significantly outperforming previous methods.", "date": "2026-03-12", "categories": []}
{"id": "2603.13435", "title": "CtrlAttack: A Unified Attack on World-Model Control in Diffusion Models", "abstract": "Diffusion-based image-to-video (I2V) models increasingly exhibit world-model-like properties by implicitly capturing temporal dynamics. However, existing studies have mainly focused on visual quality and controllability, and the robustness of the state transition learned by the model remains understudied. To fill this gap, we are the first to analyze the vulnerability of I2V models, find that temporal control mechanisms constitute a new attack surface, and reveal the challenge of modeling them uniformly under different attack settings. Based on this, we propose a trajectory-control attack, called CtrlAttack, to interfere with state evolution during the generation process. Specifically, we represent the perturbation as a low-dimensional velocity field and construct a continuous displacement field via temporal integration, thereby affecting the model's state transitions while maintaining temporal consistency; meanwhile, we map the perturbation to the observation space, making the method applicable to both white-box and black-box attack settings. Experimental results show that even under low-dimensional and strongly regularized perturbation constraints, our method can still significantly disrupt temporal consistency by increasing the attack success rate (ASR) to over 90% in the white-box setting and over 80% in the black-box setting, while keeping the variation of the FID and FVD within 6 and 130, respectively, thus revealing the potential security risk of I2V models at the level of state dynamics.", "date": "2026-03-13", "categories": []}
{"id": "2603.13880", "title": "Benchmarking the Energy Cost of Assurance in Neuromorphic Edge Robotics", "abstract": "Deploying trustworthy artificial intelligence on edge robotics imposes a difficult trade-off between high-assurance robustness and energy sustainability. Traditional defense mechanisms against adversarial attacks typically incur significant computational overhead, threatening the viability of power-constrained platforms in environments such as cislunar space. This paper quantifies the energy cost of assurance in event-driven neuromorphic systems. We benchmark the Hierarchical Temporal Defense (HTD) framework on the BrainChip Akida AKD1000 processor against a suite of adversarial temporal attacks. We demonstrate that unlike traditional deep learning defenses which often degrade efficiency significantly with increased robustness, the event-driven nature of the proposed architecture achieves a superior trade-off. The system reduces gradient-based adversarial success rates from 82.1% to 18.7% and temporal jitter success rates from 75.8% to 25.1%, while maintaining an energy consumption of approximately 45 microjoules per inference. We report a counter-intuitive reduction in dynamic power consumption in the fully defended configuration, attributed to volatility-gated plasticity mechanisms that induce higher network sparsity. These results provide empirical evidence that neuromorphic sparsity enables sustainable and high-assurance edge autonomy.", "date": "2026-03-14", "categories": []}
{"id": "2603.14860", "title": "Architecture-Agnostic Feature Synergy for Universal Defense Against Heterogeneous Generative Threats", "abstract": "Generative AI deployment poses unprecedented challenges to content safety and privacy. However, existing defense mechanisms are often tailored to specific architectures (e.g., Diffusion Models or GANs), creating fragile \"defense silos\" that fail against heterogeneous generative threats. This paper identifies a fundamental optimization barrier in naive pixel-space ensemble strategies: due to divergent objective functions, pixel-level gradients from heterogeneous generators become statistically orthogonal, causing destructive interference. To overcome this, we observe that despite disparate low-level mechanisms, high-level feature representations of generated content exhibit alignment across architectures. Based on this, we propose the Architecture-Agnostic Targeted Feature Synergy (ATFS) framework. By introducing a target guidance image, ATFS reformulates multi-model defense as a unified feature space alignment task, enabling intrinsic gradient alignment without complex rectification. Extensive experiments show ATFS achieves SOTA protection in heterogeneous scenarios (e.g., Diffusion+GAN). It converges rapidly, reaching over 90% performance within 40 iterations, and maintains strong attack potency even under tight perturbation budgets. The framework seamlessly extends to unseen architectures (e.g., VQ-VAE) by switching the feature extractor, and demonstrates robust resistance to JPEG compression and scaling. Being computationally efficient and lightweight, ATFS offers a viable pathway to dismantle defense silos and enable universal generative security. Code and models are open-sourced for reproducibility.", "date": "2026-03-16", "categories": []}
{"id": "2603.15221", "title": "ADV-0: Closed-Loop Min-Max Adversarial Training for Long-Tail Robustness in Autonomous Driving", "abstract": "Deploying autonomous driving systems requires robustness against long-tail scenarios that are rare but safety-critical. While adversarial training offers a promising solution, existing methods typically decouple scenario generation from policy optimization and rely on heuristic surrogates. This leads to objective misalignment and fails to capture the shifting failure modes of evolving policies. This paper presents ADV-0, a closed-loop min-max optimization framework that treats the interaction between driving policy (defender) and adversarial agent (attacker) as a zero-sum Markov game. By aligning the attacker's utility directly with the defender's objective, we reveal the optimal adversary distribution. To make this tractable, we cast dynamic adversary evolution as iterative preference learning, efficiently approximating this optimum and offering an algorithm-agnostic solution to the game. Theoretically, ADV-0 converges to a Nash Equilibrium and maximizes a certified lower bound on real-world performance. Experiments indicate that it effectively exposes diverse safety-critical failures and greatly enhances the generalizability of both learned policies and motion planners against unseen long-tail risks.", "date": "2026-03-16", "categories": []}
{"id": "2603.15396", "title": "AI Evasion and Impersonation Attacks on Facial Re-Identification with Activation Map Explanations", "abstract": "Facial identification systems are increasingly deployed in surveillance and yet their vulnerability to adversarial evasion and impersonation attacks pose a critical risk. This paper introduces a novel framework for generating adversarial patches capable of both evasion and impersonation attacks against deep re-identification models across non-overlapping cameras. Unlike prior approaches that require iterative patch optimisation for each target, our method employs a conditional encoder-decoder network to synthesize adversarial patches in a single forward pass, guided by multi-scale features from source and target images. The patches are optimised with a dual adversarial objective comprising of pull and push terms. To enhance imperceptibility and aid physical deployment, we further integrate naturalistic patch generation using pre-trained latent diffusion models. Experiments on standard pedestrian (Market-1501, DukeMTMCreID) and facial recognition benchmarks (CelebA-HQ, PubFig) datasets demonstrate the effectiveness of the proposed method. Our adversarial evasion attacks reduce mean Average Precision from 90% to 0.4% in white-box settings and from 72% to 0.4% in black-box settings, showing strong cross-model generalization. In targeted impersonation attacks, our framework achieves a success rate of 27% on CelebA-HQ, competing with other patch-based methods. We go further to use clustering of activation maps to interpret which features are most used by adversarial attacks and propose a pathway for future countermeasures. The results highlight the practicality of adversarial patch attacks on retrieval-based systems and underline the urgent need for robust defense strategies.", "date": "2026-03-16", "categories": []}
{"id": "2402.03627", "title": "Partially Recentralization Softmax Loss for Vision-Language Models Robustness", "abstract": "As Large Language Models make a breakthrough in natural language processing tasks (NLP), multimodal technique becomes extremely popular. However, it has been shown that multimodal NLP are vulnerable to adversarial attacks, where the outputs of a model can be dramatically changed by a perturbation to the input. While several defense techniques have been proposed both in computer vision and NLP models, the multimodal robustness of models have not been fully explored. In this paper, we study the adversarial robustness provided by modifying loss function of pre-trained multimodal models, by restricting top K softmax outputs. Based on the evaluation and scoring, our experiments show that after a fine-tuning, adversarial robustness of pre-trained models can be significantly improved, against popular attacks. Further research should be studying, such as output diversity, generalization and the robustness-performance trade-off of this kind of loss functions. Our code will be available after this paper is accepted", "date": "2024-02-06", "categories": []}
{"id": "2506.18248", "title": "Improving Black-Box Generative Attacks via Generator Semantic Consistency", "abstract": "Transfer attacks optimize on a surrogate and deploy to a black-box target. While iterative optimization attacks in this paradigm are limited by their per-input cost limits efficiency and scalability due to multistep gradient updates for each input, generative attacks alleviate these by producing adversarial examples in a single forward pass at test time. However, current generative attacks still adhere to optimizing surrogate losses (e.g., feature divergence) and overlook the generator's internal dynamics, underexploring how the generator's internal representations shape transferable perturbations. To address this, we enforce semantic consistency by aligning the early generator's intermediate features to an EMA teacher, stabilizing object-aligned representations and improving black-box transfer without inference-time overhead. To ground the mechanism, we quantify semantic stability as the standard deviation of foreground IoU between cluster-derived activation masks and foreground masks across generator blocks, and observe reduced semantic drift under our method. For more reliable evaluation, we also introduce Accidental Correction Rate (ACR) to separate inadvertent corrections from intended misclassifications, complementing the inherent blind spots in traditional Attack Success Rate (ASR), Fooling Rate (FR), and Accuracy metrics. Across architectures, domains, and tasks, our approach can be seamlessly integrated into existing generative attacks with consistent improvements in black-box transfer, while maintaining test-time efficiency.", "date": "2025-06-23", "categories": []}
{"id": "2509.23325", "title": "Robust Fine-Tuning from Non-Robust Pretrained Models: Mitigating Suboptimal Transfer With Epsilon-Scheduling", "abstract": "Fine-tuning pretrained models is a standard and effective workflow in modern machine learning. However, robust fine-tuning (RFT), which aims to simultaneously achieve adaptation to a downstream task and robustness to adversarial examples, remains challenging. Despite the abundance of non-robust pretrained models in open-source repositories, their potential for RFT is less understood. We address this knowledge gap by systematically examining RFT from such non-robust models. Our experiments reveal that fine-tuning non-robust models with a robust objective, even under small perturbations, can lead to poor performance, a phenomenon that we dub suboptimal transfer. In challenging scenarios (eg, difficult tasks, high perturbation), the resulting performance can be so low that it may be considered a transfer failure. We find that fine-tuning using a robust objective impedes task adaptation at the beginning of training and eventually prevents optimal transfer. However, we propose a novel heuristic, Epsilon-Scheduling, a schedule over perturbation strength used during training that promotes optimal transfer. Additionally, we introduce expected robustness, a metric that captures performance across a range of perturbations, providing a more comprehensive evaluation of the accuracy-robustness trade-off for diverse models at test time. Extensive experiments on a wide range of configurations (six pretrained models and five datasets) show that Epsilon-Scheduling successfully prevents suboptimal transfer and consistently improves expected robustness.", "date": "2025-09-27", "categories": []}
{"id": "2603.12414", "title": "SpectralGuard: Detecting Memory Collapse Attacks in State Space Models", "abstract": "State Space Models (SSMs) such as Mamba achieve linear-time sequence processing through input-dependent recurrence, but this mechanism introduces a critical safety vulnerability. We show that the spectral radius rho(A-bar) of the discretized transition operator governs effective memory horizon: when an adversary drives rho toward zero through gradient-based Hidden State Poisoning, memory collapses from millions of tokens to mere dozens, silently destroying reasoning capacity without triggering output-level alarms. We prove an Evasion Existence Theorem showing that for any output-only defense, adversarial inputs exist that simultaneously induce spectral collapse and evade detection, then introduce SpectralGuard, a real-time monitor that tracks spectral stability across all model layers. SpectralGuard achieves F1=0.961 against non-adaptive attackers and retains F1=0.842 under the strongest adaptive setting, with sub-15ms per-token latency. Causal interventions and cross-architecture transfer to hybrid SSM-Attention systems confirm that spectral monitoring provides a principled, deployable safety layer for recurrent foundation models.", "date": "2026-03-12", "categories": []}
{"id": "2603.12557", "title": "Lyapunov Stable Graph Neural Flow", "abstract": "Graph Neural Networks (GNNs) are highly vulnerable to adversarial perturbations in both topology and features, making the learning of robust representations a critical challenge. In this work, we bridge GNNs with control theory to introduce a novel defense framework grounded in integer- and fractional-order Lyapunov stability. Unlike conventional strategies that rely on resource-heavy adversarial training or data purification, our approach fundamentally constrains the underlying feature-update dynamics of the GNN. We propose an adaptive, learnable Lyapunov function paired with a novel projection mechanism that maps the network's state into a stable space, thereby offering theoretically provable stability guarantees. Notably, this mechanism is orthogonal to existing defenses, allowing for seamless integration with techniques like adversarial training to achieve cumulative robustness. Extensive experiments demonstrate that our Lyapunov-stable graph neural flows substantially outperform base neural flows and state-of-the-art baselines across standard benchmarks and various adversarial attack scenarios.", "date": "2026-03-13", "categories": []}
{"id": "2603.12582", "title": "RTD-Guard: A Black-Box Textual Adversarial Detection Framework via Replacement Token Detection", "abstract": "Textual adversarial attacks pose a serious security threat to Natural Language Processing (NLP) systems by introducing imperceptible perturbations that mislead deep learning models. While adversarial example detection offers a lightweight alternative to robust training, existing methods typically rely on prior knowledge of attacks, white-box access to the victim model, or numerous queries, which severely limits their practical deployment. This paper introduces RTD-Guard, a novel black-box framework for detecting textual adversarial examples. Our key insight is that word-substitution perturbations in adversarial attacks closely resemble the \"replaced tokens\" that a Replaced Token Detection (RTD) discriminator is pre-trained to identify. Leveraging this, RTD-Guard employs an off-the-shelf RTD discriminator-without fine-tuning-to localize suspicious tokens, masks them, and detects adversarial examples by observing the prediction confidence shift of the victim model before and after intervention. The entire process requires no adversarial data, model tuning, or internal model access, and uses only two black-box queries. Comprehensive experiments on multiple benchmark datasets demonstrate that RTD-Guard effectively detects adversarial texts generated by diverse state-of-the-art attack methods. It surpasses existing detection baselines across multiple metrics, offering a highly efficient, practical, and resource-light defense mechanism-particularly suited for real-world deployment in resource-constrained or privacy-sensitive environments.", "date": "2026-03-13", "categories": []}
{"id": "2603.12688", "title": "STRAP-ViT: Segregated Tokens with Randomized -- Transformations for Defense against Adversarial Patches in ViTs", "abstract": "Adversarial patches are physically realizable localized noise, which are able to hijack Vision Transformers (ViT) self-attention, pulling focus toward a small, high-contrast region and corrupting the class token to force confident misclassifications. In this paper, we claim that the tokens which correspond to the areas of the image that contain the adversarial noise, have different statistical properties when compared to the tokens which do not overlap with the adversarial perturbations. We use this insight to propose a mechanism, called STRAP-ViT, which uses Jensen-Shannon Divergence as a metric for segregating tokens that behave as anomalies in the Detection Phase, and then apply randomized composite transformations on them during the Mitigation Phase to make the adversarial noise ineffective. The minimum number of tokens to transform is a hyper-parameter for the defense mechanism and is chosen such that at least 50% of the patch is covered by the transformed tokens. STRAP-ViT fits as a non-trainable plug-and-play block within the ViT architectures, for inference purposes only, with a minimal computational cost and does not require any additional training cost/effort. STRAP-ViT has been tested on multiple pre-trained vision transformer architectures (ViT-base-16 and DinoV2) and datasets (ImageNet and CalTech-101), across multiple adversarial attacks (Adversarial Patch, LAVAN, GDPA and RP2), and found to provide excellent robust accuracies lying within a 2-3% range of the clean baselines, and outperform the state-of-the-art.", "date": "2026-03-13", "categories": []}
{"id": "2603.12717", "title": "Altered Thoughts, Altered Actions: Probing Chain-of-Thought Vulnerabilities in VLA Robotic Manipulation", "abstract": "Recent Vision-Language-Action (VLA) models increasingly adopt chain-of-thought (CoT) reasoning, generating a natural-language plan before decoding motor commands. This internal text channel between the reasoning module and the action decoder has received no adversarial scrutiny. We ask: which properties of this intermediate plan does the action decoder actually rely on, and can targeted corruption of the reasoning trace alone -- with all inputs left intact -- degrade a robot's physical task performance? We design a taxonomy of seven text corruptions organized into three attacker tiers (blind noise, mechanical-semantic, and LLM-adaptive) and apply them to a state-of-the-art reasoning VLA across 40 LIBERO tabletop manipulation tasks. Our results reveal a striking asymmetry: substituting object names in the reasoning trace reduces overall success rate by 8.3~percentage points (pp) -- reaching $-$19.3~pp on goal-conditioned tasks and $-$45~pp on individual tasks -- whereas sentence reordering, spatial-direction reversal, token noise, and even a 70B-parameter LLM crafting plausible-but-wrong plans all have negligible impact (within $\\pm$4~pp). This asymmetry indicates that the action decoder depends on entity-reference integrity rather than reasoning quality or sequential structure. Notably, a sophisticated LLM-based attacker underperforms simple mechanical object-name substitution, because preserving plausibility inadvertently retains the entity-grounding structure the decoder needs. A cross-architecture control using a non-reasoning VLA confirms the vulnerability is exclusive to reasoning-augmented models, while instruction-level attacks degrade both architectures -- establishing that the internal reasoning trace is a distinct and stealthy threat vector invisible to input-validation defenses.", "date": "2026-03-13", "categories": []}
{"id": "2603.12796", "title": "Spectral Defense Against Resource-Targeting Attack in 3D Gaussian Splatting", "abstract": "Recent advances in 3D Gaussian Splatting (3DGS) deliver high-quality rendering, yet the Gaussian representation exposes a new attack surface, the resource-targeting attack. This attack poisons training images, excessively inducing Gaussian growth to cause resource exhaustion. Although efficiency-oriented methods such as smoothing, thresholding, and pruning have been explored, these spatial-domain strategies operate on visible structures but overlook how stealthy perturbations distort the underlying spectral behaviors of training data. As a result, poisoned inputs introduce abnormal high-frequency amplifications that mislead 3DGS into interpreting noisy patterns as detailed structures, ultimately causing unstable Gaussian overgrowth and degraded scene fidelity. To address this, we propose \\textbf{Spectral Defense} in Gaussian and image fields. We first design a 3D frequency filter to selectively prune Gaussians exhibiting abnormally high frequencies. Since natural scenes also contain legitimate high-frequency structures, directly suppressing high frequencies is insufficient, and we further develop a 2D spectral regularization on renderings, distinguishing naturally isotropic frequencies while penalizing anisotropic angular energy to constrain noisy patterns. Experiments show that our defense builds robust, accurate, and secure 3DGS, suppressing overgrowth by up to $5.92\\times$, reducing memory by up to $3.66\\times$, and improving speed by up to $4.34\\times$ under attacks.", "date": "2026-03-13", "categories": []}
{"id": "2603.13028", "title": "Purify Once, Edit Freely: Breaking Image Protections under Model Mismatch", "abstract": "Diffusion models enable high-fidelity image editing but can also be misused for unauthorized style imitation and harmful content generation. To mitigate these risks, proactive image protection methods embed small, often imperceptible adversarial perturbations into images before sharing to disrupt downstream editing or fine-tuning. However, in realistic post-release scenarios, content owners cannot control downstream processing pipelines, and protections optimized for a surrogate model may fail when attackers use mismatched diffusion pipelines. Existing purification methods can weaken protections but often sacrifice image quality and rarely examine architectural mismatch. We introduce a unified post-release purification framework to evaluate protection survivability under model mismatch. We propose two practical purifiers: VAE-Trans, which corrects protected images via latent-space projection, and EditorClean, which performs instruction-guided reconstruction with a Diffusion Transformer to exploit architectural heterogeneity. Both operate without access to protected images or defense internals. Across 2,100 editing tasks and six representative protection methods, EditorClean consistently restores editability. Compared to protected inputs, it improves PSNR by 3-6 dB and reduces FID by 50-70 percent on downstream edits, while outperforming prior purification baselines by about 2 dB PSNR and 30 percent lower FID. Our results reveal a purify-once, edit-freely failure mode: once purification succeeds, the protective signal is largely removed, enabling unrestricted editing. This highlights the need to evaluate protections under model mismatch and design defenses robust to heterogeneous attackers.", "date": "2026-03-13", "categories": []}
{"id": "2505.19459", "title": "Your Classifier Can Do More: Towards Balancing the Gaps in Classification, Robustness, and Generation", "abstract": "Joint Energy-based Models (JEMs) are well known for their ability to unify classification and generation within a single framework. Despite their promising generative and discriminative performance, their robustness remains far inferior to adversarial training (AT), which, conversely, achieves strong robustness but sacrifices clean accuracy and lacks generative ability. This inherent trilemma-balancing classification accuracy, robustness, and generative capability-raises a fundamental question: Can a single model achieve all three simultaneously? To answer this, we conduct a systematic energy landscape analysis of clean, adversarial, and generated samples across various JEM and AT variants. We observe that AT reduces the energy gap between clean and adversarial samples, while JEMs narrow the gap between clean and synthetic ones. This observation suggests a key insight: if the energy distributions of all three data types can be aligned, we might bridge their performance disparities. Building on this idea, we propose Energy-based Joint Distribution Adversarial Training (EB-JDAT), a unified generative-discriminative-robust framework that maximizes the joint probability of clean and adversarial distribution. EB-JDAT introduces a novel min-max energy optimization to explicitly aligning energies across clean, adversarial, and generated samples. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet subsets demonstrate that EB-JDAT achieves state-of-the-art robustness while maintaining near-original accuracy and competitive generation quality of JEMs, effectively achieving a new trade-off frontier between accuracy, robustness, and generation. The code is released at https://github.com/yujkc/EB-JDAT.", "date": "2025-05-26", "categories": []}
{"id": "2602.13093", "title": "Consistency of Large Reasoning Models Under Multi-Turn Attacks", "abstract": "Large reasoning models with reasoning capabilities achieve state-of-the-art performance on complex tasks, but their robustness under multi-turn adversarial pressure remains underexplored. We evaluate nine frontier reasoning models under adversarial attacks. Our findings reveal that reasoning confers meaningful but incomplete robustness: most reasoning models studied significantly outperform instruction-tuned baselines, yet all exhibit distinct vulnerability profiles, with misleading suggestions universally effective and social pressure showing model-specific efficacy. Through trajectory analysis, we identify five failure modes (Self-Doubt, Social Conformity, Suggestion Hijacking, Emotional Susceptibility, and Reasoning Fatigue) with the first two accounting for 50% of failures. We further demonstrate that Confidence-Aware Response Generation (CARG), effective for standard LLMs, fails for reasoning models due to overconfidence induced by extended reasoning traces; counterintuitively, random confidence embedding outperforms targeted extraction. Our results highlight that reasoning capabilities do not automatically confer adversarial robustness and that confidence-based defenses require fundamental redesign for reasoning models.", "date": "2026-02-13", "categories": []}
{"id": "2603.10689", "title": "Contract And Conquer: How to Provably Compute Adversarial Examples for a Black-Box Model?", "abstract": "Black-box adversarial attacks are widely used as tools to test the robustness of deep neural networks against malicious perturbations of input data aimed at a specific change in the output of the model. Such methods, although they remain empirically effective, usually do not guarantee that an adversarial example can be found for a particular model. In this paper, we propose Contract And Conquer (CAC), an approach to provably compute adversarial examples for neural networks in a black-box manner. The method is based on knowledge distillation of a black-box model on an expanding distillation dataset and precise contraction of the adversarial example search space. CAC is supported by the transferability guarantee: we prove that the method yields an adversarial example for the black-box model within a fixed number of algorithm iterations. Experimentally, we demonstrate that the proposed approach outperforms existing state-of-the-art black-box attack methods on ImageNet dataset for different target models, including vision transformers.", "date": "2026-03-11", "categories": []}
{"id": "2410.23678", "title": "Goal Hijacking Attack on Large Language Models via Pseudo-Conversation Injection", "abstract": "Goal hijacking is a type of adversarial attack on Large Language Models (LLMs) where the objective is to manipulate the model into producing a specific, predetermined output, regardless of the user's original input. In goal hijacking, an attacker typically appends a carefully crafted malicious suffix to the user's prompt, which coerces the model into ignoring the user's original input and generating the target response. In this paper, we introduce a novel goal hijacking attack method called Pseudo-Conversation Injection, which leverages the weaknesses of LLMs in role identification within conversation contexts. Specifically, we construct the suffix by fabricating responses from the LLM to the user's initial prompt, followed by a prompt for a malicious new task. This leads the model to perceive the initial prompt and fabricated response as a completed conversation, thereby executing the new, falsified prompt. Following this approach, we propose three Pseudo-Conversation construction strategies: Targeted Pseudo-Conversation, Universal Pseudo-Conversation, and Robust Pseudo-Conversation. These strategies are designed to achieve effective goal hijacking across various scenarios. Our experiments, conducted on two mainstream LLM platforms including ChatGPT and Qwen, demonstrate that our proposed method significantly outperforms existing approaches in terms of attack effectiveness.", "date": "2024-10-31", "categories": []}
{"id": "2503.11627", "title": "Are Deep Speech Denoising Models Robust to Adversarial Noise?", "abstract": "Deep noise suppression (DNS) models enjoy widespread use throughout a variety of high-stakes speech applications. However, we show that four recent DNS models can each be reduced to outputting unintelligible gibberish through the addition of psychoacoustically hidden adversarial noise, even in low-background-noise and simulated over-the-air settings. For three of the models, a small transcription study with audio and multimedia experts confirms unintelligibility of the attacked audio; simultaneously, an ABX study shows that the adversarial noise is generally imperceptible, with some variance between participants and samples. While we also establish several negative results around targeted attacks and model transfer, our results nevertheless highlight the need for practical countermeasures before open-source DNS systems can be used in safety-critical applications.", "date": "2025-03-14", "categories": []}
{"id": "2504.11106", "title": "Token-Level Constraint Boundary Search for Jailbreaking Text-to-Image Models", "abstract": "Text-to-Image (T2I) generation has advanced rapidly in recent years, but they also raise safety concerns due to the potential production of harmful content. In the practical deployments, T2I services typically adopt full-chain defenses that combine a prompt checker, a securely trained generator, and a post-hoc image checker. Jailbreaking such full-chain systems is challenging in the black-box settings because prompt tokens form a discrete combinatorial space and the attack must satisfy multiple coupled constraints under sparse feedback and limited queries. To address these challenges, we propose Token-level Constraint Boundary Search (TCBS)-Attack, a novel query-based black-box jailbreak attack that searches for tokens located near the decision boundaries defined by text and image checkers. TCBS-Attack incorporates decision boundaries as constraint conditions to guide the evolutionary search of token populations, iteratively optimize tokens near these boundaries. Such evolutionary search process reduces the effective search space and improves query efficiency while preserving semantic coherence. Extensive experiments demonstrate that TCBS-Attack consistently outperforms state-of-the-art jailbreak attacks across various T2I models, including securely trained open-source models and commercial online services like DALL-E 3. TCBS-Attack achieves an ASR-4 of 52.5% and an ASR-1 of 22.0% on jailbreaking full-chain T2I models, significantly surpassing baseline methods.", "date": "2025-04-15", "categories": []}
{"id": "2511.21145", "title": "TEAR: Temporal-aware Automated Red-teaming for Text-to-Video Models", "abstract": "Text-to-Video (T2V) models are capable of synthesizing high-quality, temporally coherent dynamic video content, but the diverse generation also inherently introduces critical safety challenges. Existing safety evaluation methods,which focus on static image and text generation, are insufficient to capture the complex temporal dynamics in video generation. To address this, we propose a TEmporal-aware Automated Red-teaming framework, named TEAR, an automated framework designed to uncover safety risks specifically linked to the dynamic temporal sequencing of T2V models. TEAR employs a temporal-aware test generator optimized via a two-stage approach: initial generator training and temporal-aware online preference learning, to craft textually innocuous prompts that exploit temporal dynamics to elicit policy-violating video output. And a refine model is adopted to improve the prompt stealthiness and adversarial effectiveness cyclically. Extensive experimental evaluation demonstrates the effectiveness of TEAR across open-source and commercial T2V systems with over 80% attack success rate, a significant boost from prior best result of 57%.", "date": "2025-11-26", "categories": []}
{"id": "2602.22427", "title": "Adversarial Hubness Detector: Detecting Hubness Poisoning in Retrieval-Augmented Generation Systems", "abstract": "Retrieval-Augmented Generation (RAG) systems are essential to contemporary AI applications, allowing large language models to obtain external knowledge via vector similarity search. Nevertheless, these systems encounter a significant security flaw: hubness - items that frequently appear in the top-$k$ retrieval results for a disproportionately high number of varied queries. These hubs can be exploited to introduce harmful content, alter search rankings, bypass content filtering, and decrease system performance.\n  We introduce hubscan, an open-source security scanner that evaluates vector indices and embeddings to identify hubs in RAG systems. Hubscan presents a multi-detector architecture that integrates: (1) robust statistical hubness detection utilizing median/Median Absolute Deviation (MAD)-based z-scores, (2) cluster spread analysis to assess cross-cluster retrieval patterns, (3) stability testing under query perturbations, and (4) domain-aware and modality-aware detection for category-specific and cross-modal attacks. Our solution accommodates several vector databases (FAISS, Pinecone, Qdrant, Weaviate) and offers versatile retrieval techniques, including vector similarity, hybrid search, and lexical matching with reranking capabilities.\n  We evaluate hubscan on Food-101, MS-COCO, and FiQA adversarial hubness benchmarks constructed using state-of-the-art gradient-optimized and centroid-based hub generation methods. Hubscan achieves 90% recall at a 0.2% alert budget and 100% recall at 0.4%, with adversarial hubs ranking above the 99.8th percentile. In testing, domain-scoped scanning recovered 100% of targeted attacks that evaded global detection. Production validation on 1M real web documents from MS MARCO demonstrates significant score separation between clean documents and adversarial content.", "date": "2026-02-25", "categories": []}
{"id": "2603.03507", "title": "Solving adversarial examples requires solving exponential misalignment", "abstract": "Adversarial attacks - input perturbations imperceptible to humans that fool neural networks - remain both a persistent failure mode in machine learning, and a phenomenon with mysterious origins. To shed light, we define and analyze a network's perceptual manifold (PM) for a class concept as the space of all inputs confidently assigned to that class by the network. We find, strikingly, that the dimensionalities of neural network PMs are orders of magnitude higher than those of natural human concepts. Since volume typically grows exponentially with dimension, this suggests exponential misalignment between machines and humans, with exponentially many inputs confidently assigned to concepts by machines but not humans. Furthermore, this provides a natural geometric hypothesis for the origin of adversarial examples: because a network's PM fills such a large region of input space, any input will be very close to any class concept's PM. Our hypothesis thus suggests that adversarial robustness cannot be attained without dimensional alignment of machine and human PMs, and therefore makes strong predictions: both robust accuracy and distance to any PM should be negatively correlated with the PM dimension. We confirmed these predictions across 18 different networks of varying robust accuracy. Crucially, we find even the most robust networks are still exponentially misaligned, and only the few PMs whose dimensionality approaches that of human concepts exhibit alignment to human perception. Our results connect the fields of alignment and adversarial examples, and suggest the curse of high dimensionality of machine PMs is a major impediment to adversarial robustness.", "date": "2026-03-03", "categories": []}
{"id": "2603.10010", "title": "FERRET: Framework for Expansion Reliant Red Teaming", "abstract": "We introduce a multi-faceted automated red teaming framework in which the goal is to generate multi-modal adversarial conversations that would break a target model and introduce various expansions that would result in more effective and efficient adversarial conversations. The introduced expansions include: 1. Horizontal expansion in which the goal is for the red team model to self-improve and generate more effective conversation starters that would shape a conversation. 2. Vertical expansion in which the goal is to take these conversation starters that are discovered in the horizontal expansion phase and expand them into effective multi-modal conversations and 3. Meta expansion in which the goal is for the red team model to discover more effective multi-modal attack strategies during the course of a conversation. We call our framework FERRET (Framework for Expansion Reliant Red Teaming) and compare it with various existing automated red teaming approaches. In our experiments, we demonstrate the effectiveness of FERRET in generating effective multi-modal adversarial conversations and its superior performance against existing state of the art approaches.", "date": "2026-02-17", "categories": []}
{"id": "2603.10068", "title": "ADVERSA: Measuring Multi-Turn Guardrail Degradation and Judge Reliability in Large Language Models", "abstract": "Most adversarial evaluations of large language model (LLM) safety assess single prompts and report binary pass/fail outcomes, which fails to capture how safety properties evolve under sustained adversarial interaction. We present ADVERSA, an automated red-teaming framework that measures guardrail degradation dynamics as continuous per-round compliance trajectories rather than discrete jailbreak events. ADVERSA uses a fine-tuned 70B attacker model (ADVERSA-Red, Llama-3.1-70B-Instruct with QLoRA) that eliminates the attacker-side safety refusals that render off-the-shelf models unreliable as attackers, scoring victim responses on a structured 5-point rubric that treats partial compliance as a distinct measurable state.\n  We report a controlled experiment across three frontier victim models (Claude Opus 4.6, Gemini 3.1 Pro, GPT-5.2) using a triple-judge consensus architecture in which judge reliability is measured as a first-class research outcome rather than assumed. Across 15 conversations of up to 10 adversarial rounds, we observe a 26.7% jailbreak rate with an average jailbreak round of 1.25, suggesting that in this evaluation setting, successful jailbreaks were concentrated in early rounds rather than accumulating through sustained pressure. We document inter-judge agreement rates, self-judge scoring tendencies, attacker drift as a failure mode in fine-tuned attackers deployed out of their training distribution, and attacker refusals as a previously-underreported confound in victim resistance measurement. All limitations are stated explicitly. Attack prompts are withheld per responsible disclosure policy; all other experimental artifacts are released.", "date": "2026-03-10", "categories": []}
{"id": "2603.10413", "title": "Enhancing Network Intrusion Detection Systems: A Multi-Layer Ensemble Approach to Mitigate Adversarial Attacks", "abstract": "Adversarial examples can represent a serious threat to machine learning (ML) algorithms. If used to manipulate the behaviour of ML-based Network Intrusion Detection Systems (NIDS), they can jeopardize network security. In this work, we aim to mitigate such risks by increasing the robustness of NIDS towards adversarial attacks. To that end, we explore two adversarial methods for generating malicious network traffic. The first method is based on Generative Adversarial Networks (GAN) and the second one is the Fast Gradient Sign Method (FGSM). The adversarial examples generated by these methods are then used to evaluate a novel multilayer defense mechanism, specifically designed to mitigate the vulnerability of ML-based NIDS. Our solution consists of one layer of stacking classifiers and a second layer based on an autoencoder. If the incoming network data are classified as benign by the first layer, the second layer is activated to ensure that the decision made by the stacking classifier is correct. We also incorporated adversarial training to further improve the robustness of our solution. Experiments on two datasets, namely UNSW-NB15 and NSL-KDD, demonstrate that the proposed approach increases resilience to adversarial attacks.", "date": "2026-03-11", "categories": []}
{"id": "2603.10806", "title": "Backdoor Directions in Vision Transformers", "abstract": "This paper investigates how Backdoor Attacks are represented within Vision Transformers (ViTs). By assuming knowledge of the trigger, we identify a specific ``trigger direction'' in the model's activations that corresponds to the internal representation of the trigger. We confirm the causal role of this linear direction by showing that interventions in both activation and parameter space consistently modulate the model's backdoor behavior across multiple datasets and attack types. Using this direction as a diagnostic tool, we trace how backdoor features are processed across layers. Our analysis reveals distinct qualitative differences: static-patch triggers follow a different internal logic than stealthy, distributed triggers. We further examine the link between backdoors and adversarial attacks, specifically testing whether PGD-based perturbations (de-)activate the identified trigger mechanism. Finally, we propose a data-free, weight-based detection scheme for stealthy-trigger attacks. Our findings show that mechanistic interpretability offers a robust framework for diagnosing and addressing security vulnerabilities in computer vision.", "date": "2026-03-11", "categories": []}
{"id": "2603.10821", "title": "Evaluating randomized smoothing as a defense against adversarial attacks in trajectory prediction", "abstract": "Accurate and robust trajectory prediction is essential for safe and efficient autonomous driving, yet recent work has shown that even state-of-the-art prediction models are highly vulnerable to inputs being mildly perturbed by adversarial attacks. Although model vulnerabilities to such attacks have been studied, work on effective countermeasures remains limited. In this work, we develop and evaluate a new defense mechanism for trajectory prediction models based on randomized smoothing -- an approach previously applied successfully in other domains. We evaluate its ability to improve model robustness through a series of experiments that test different strategies of randomized smoothing. We show that our approach can consistently improve prediction robustness of multiple base trajectory prediction models in various datasets without compromising accuracy in non-adversarial settings. Our results demonstrate that randomized smoothing offers a simple and computationally inexpensive technique for mitigating adversarial attacks in trajectory prediction.", "date": "2026-03-11", "categories": []}
{"id": "2603.10842", "title": "PivotAttack: Rethinking the Search Trajectory in Hard-Label Text Attacks via Pivot Words", "abstract": "Existing hard-label text attacks often rely on inefficient \"outside-in\" strategies that traverse vast search spaces. We propose PivotAttack, a query-efficient \"inside-out\" framework. It employs a Multi-Armed Bandit algorithm to identify Pivot Sets-combinatorial token groups acting as prediction anchors-and strategically perturbs them to induce label flips. This approach captures inter-word dependencies and minimizes query costs. Extensive experiments across traditional models and Large Language Models demonstrate that PivotAttack consistently outperforms state-of-the-art baselines in both Attack Success Rate and query efficiency.", "date": "2026-03-11", "categories": []}
{"id": "2511.20223", "title": "V-Attack: Targeting Disentangled Value Features for Controllable Adversarial Attacks on LVLMs", "abstract": "Adversarial attacks have evolved from simply disrupting predictions on conventional task-specific models to the more complex goal of manipulating image semantics on Large Vision-Language Models (LVLMs). However, existing methods struggle with controllability and fail to precisely manipulate the semantics of specific concepts in the image. We attribute this limitation to semantic entanglement in the patch-token representations on which adversarial attacks typically operate: global context aggregated by self-attention in the vision encoder dominates individual patch features, making them unreliable handles for precise local semantic manipulation. Our systematic investigation reveals a key insight: value features (V) computed within the transformer attention block serve as much more precise handles for manipulation. We show that V suppresses global-context channels, allowing it to retain high-entropy, disentangled local semantic information. Building on this discovery, we propose V-Attack, a novel method designed for precise local semantic attacks. V-Attack targets the value features and introduces two core components: (1) a Self-Value Enhancement module to refine V's intrinsic semantic richness, and (2) a Text-Guided Value Manipulation module that leverages text prompts to locate source concept and optimize it toward a target concept. By bypassing the entangled patch features, V-Attack achieves highly effective semantic control. Extensive experiments across diverse LVLMs, including LLaVA, InternVL, DeepseekVL and GPT-4o, show that V-Attack improves the attack success rate by an average of 36% over state-of-the-art methods, exposing critical vulnerabilities in modern visual-language understanding. Our code and data are available https://github.com/Summu77/V-Attack.", "date": "2025-11-25", "categories": []}
{"id": "2511.21192", "title": "When Robots Obey the Patch: Universal Transferable Patch Attacks on Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models are vulnerable to adversarial attacks, yet universal and transferable attacks remain underexplored, as most existing patches overfit to a single model and fail in black-box settings. To address this gap, we present a systematic study of universal, transferable adversarial patches against VLA-driven robots under unknown architectures, finetuned variants, and sim-to-real shifts. We introduce UPA-RFAS (Universal Patch Attack via Robust Feature, Attention, and Semantics), a unified framework that learns a single physical patch in a shared feature space while promoting cross-model transfer. UPA-RFAS combines (i) a feature-space objective with an $\\ell_1$ deviation prior and repulsive InfoNCE loss to induce transferable representation shifts, (ii) a robustness-augmented two-phase min-max procedure where an inner loop learns invisible sample-wise perturbations and an outer loop optimizes the universal patch against this hardened neighborhood, and (iii) two VLA-specific losses: Patch Attention Dominance to hijack text$\\to$vision attention and Patch Semantic Misalignment to induce image-text mismatch without labels. Experiments across diverse VLA models, manipulation suites, and physical executions show that UPA-RFAS consistently transfers across models, tasks, and viewpoints, exposing a practical patch-based attack surface and establishing a strong baseline for future defenses.", "date": "2025-11-26", "categories": []}
{"id": "2603.08897", "title": "Comparative Analysis of Patch Attack on VLM-Based Autonomous Driving Architectures", "abstract": "Vision-language models are emerging for autonomous driving, yet their robustness to physical adversarial attacks remains unexplored. This paper presents a systematic framework for comparative adversarial evaluation across three VLM architectures: Dolphins, OmniDrive (Omni-L), and LeapVAD. Using black-box optimization with semantic homogenization for fair comparison, we evaluate physically realizable patch attacks in CARLA simulation. Results reveal severe vulnerabilities across all architectures, sustained multi-frame failures, and critical object detection degradation. Our analysis exposes distinct architectural vulnerability patterns, demonstrating that current VLM designs inadequately address adversarial threats in safety-critical autonomous driving applications.", "date": "2026-03-09", "categories": []}
{"id": "2408.00329", "title": "OTAD: An Optimal Transport-Induced Robust Model for Agnostic Adversarial Attack", "abstract": "Deep neural networks (DNNs) are vulnerable to small adversarial perturbations of the inputs, posing a significant challenge to their reliability and robustness. Empirical methods such as adversarial training can defend against particular attacks but remain vulnerable to more powerful attacks. Alternatively, Lipschitz networks provide certified robustness to unseen perturbations but lack sufficient expressive power. To harness the advantages of both approaches, we design a novel two-step Optimal Transport induced Adversarial Defense (OTAD) model that can fit the training data accurately while preserving the local Lipschitz continuity. First, we train a DNN with a regularizer derived from optimal transport theory, yielding a discrete optimal transport map linking data to its features. By leveraging the map's inherent regularity, we interpolate the map by solving the convex integration problem (CIP) to guarantee the local Lipschitz property. OTAD is extensible to diverse architectures of ResNet and Transformer, making it suitable for complex data. For efficient computation, the CIP can be solved through training neural networks. OTAD opens a novel avenue for developing reliable and secure deep learning systems through the regularity of optimal transport maps. Empirical results demonstrate that OTAD can outperform other robust models on diverse datasets.", "date": "2024-08-01", "categories": []}
{"id": "2410.07719", "title": "How Learning Dynamics Drive Adversarially Robust Generalization?", "abstract": "Despite being widely adopted as a canonical framework for learning robust models, adversarial training suffers from robust overfitting. Existing empirical measures and theoretical explorations are insufficient to provide satisfying mechanistic insights into the phenomenon. By viewing adversarial training with momentum SGD as a discrete-time dynamical system, we introduce a PAC-Bayesian analytical framework that proves time-resolved robust generalization bounds. Specifically, our framework tracks the closed-form evolution of the posterior mean and covariance under both stationary and non-stationary transient regimes, revealing their connections to the learning rate, the geometry of the loss landscape, and mini-batch stochastic gradients. By empirically approximating the statistical quantities implied by our theory, we offer a unified, mechanistic explanation for robust overfitting. We also illustrate why adversarial weight perturbation reduces the robust generalization gap by suppressing the loss curvature, but its design may be suboptimal for optimization due to over-penalization.", "date": "2024-10-10", "categories": []}
{"id": "2501.10466", "title": "Efficient Semi-Supervised Adversarial Training via Latent Clustering-Based Data Reduction", "abstract": "Learning robust models under adversarial settings is widely recognized as requiring a considerably large number of training samples. Recent work proposes semi-supervised adversarial training (SSAT), which utilizes external unlabeled or synthetically generated data and is currently the state of the art. However, SSAT requires substantial extra data to attain high robustness, resulting in prolonged training time and increased memory usage. In this paper, we propose data reduction strategies to improve the efficiency of SSAT by optimizing the amount of additional data incorporated. Specifically, we design novel latent clustering-based techniques to select or generate a small, critical subset of data samples near the model's decision boundary. While focusing on boundary-adjacent points, our methods maintain a balanced ratio between boundary and non-boundary data points, thereby avoiding overfitting. Comprehensive experiments across image benchmarks demonstrate that our methods can effectively reduce SSAT's data requirements and computational costs while preserving its strong robustness advantages. In particular, our latent-space selection scheme based on k-means clustering and our guided diffusion-based approach with LCG-KM are the most effective, achieving nearly identical robust accuracies with 5 times to 10 times less unlabeled data. When compared to full SSAT trained to convergence, our methods reduce total runtime by approximately 3 times to 4 times due to strategic prioritization of unlabeled data.", "date": "2025-01-15", "categories": []}
{"id": "2506.00661", "title": "Elytra: A Flexible Framework for Securing Large Vision Systems", "abstract": "Adversarial attacks have emerged as a critical threat to autonomous driving systems. These attacks exploit the underlying neural network, allowing small, almost invisible, perturbations to alter the behavior of such systems in potentially malicious ways, e.g., causing a traffic sign classification network to misclassify a stop sign as a speed limit sign. Prior work in hardening such systems against adversarial attacks has looked at fine-tuning of the system or adding additional pre-processing steps to the input pipeline. Such solutions either have a hard time generalizing, require knowledge of adversarial attacks during training, or are computationally undesirable. Instead, we propose a framework called ELYTRA to take insights for parameter-efficient fine-tuning and use low-rank adaptation (LoRA) to train a lightweight security patch (or patches), enabling us to dynamically patch large pre-existing vision systems as new vulnerabilities are discovered. We demonstrate that the ELYTRA framework can patch pre-trained large vision models to improve classification accuracy by up to 24.09% in the presence of adversarial examples.", "date": "2025-05-31", "categories": []}
{"id": "2510.02286", "title": "Tree-based Dialogue Reinforced Policy Optimization for Red-Teaming Attacks", "abstract": "Despite recent rapid progress in AI safety, current large language models remain vulnerable to adversarial attacks in multi-turn interaction settings, where attackers strategically adapt their prompts across conversation turns and pose a more critical yet realistic challenge. Existing approaches that discover safety vulnerabilities either rely on manual red-teaming with human experts or employ automated methods using pre-defined templates and human-curated attack data, with most focusing on single-turn attacks. However, these methods did not explore the vast space of possible multi-turn attacks, failing to consider novel attack trajectories that emerge from complex dialogue dynamics and strategic conversation planning. This gap is particularly critical given recent findings that LLMs exhibit significantly higher vulnerability to multi-turn attacks compared to single-turn attacks. We propose DialTree, an on-policy reinforcement learning framework integrated with tree search that autonomously discovers diverse multi-turn attack strategies by treating the dialogue as a sequential decision-making problem, enabling systematic exploration without manually curated data. Through extensive experiments, our approach not only achieves more than 44.2% higher ASR across 12 target models compared to previous state-of-the-art approaches, but also effectively uncovers new attack strategies by learning optimal dialogue policies that maximize attack success across multiple turns.", "date": "2025-10-02", "categories": []}
{"id": "2510.10932", "title": "DropVLA: An Action-Level Backdoor Attack on Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models map multimodal perception and language instructions to executable robot actions, making them particularly vulnerable to behavioral backdoor manipulation: a hidden trigger introduced during training can induce unintended physical actions while nominal task performance remains intact. Prior work on VLA backdoors primarily studies untargeted attacks or task-level hijacking, leaving fine-grained control over individual actions largely unexplored. In this work, we present DropVLA, an action-level backdoor attack that forces a reusable action primitive (e.g., open_gripper) to execute at attacker-chosen decision points under a realistic pipeline-black-box setting with limited data-poisoning access, using a window-consistent relabeling scheme for chunked fine-tuning. On OpenVLA-7B evaluated with LIBERO, vision-only poisoning achieves 98.67%-99.83% attack success rate (ASR) with only 0.31% poisoned episodes while preserving 98.50%-99.17% clean-task retention, and successfully triggers the targeted action within 25 control steps at 500 Hz (0.05 s). Text-only triggers are unstable at low poisoning budgets, and combining text with vision provides no consistent ASR improvement over vision-only attacks. The backdoor remains robust to moderate trigger variations and transfers across evaluation suites (96.27%, 99.09%), whereas text-only largely fails (0.72%). We further validate physical-world feasibility on a 7-DoF Franka arm with pi0-fast, demonstrating non-trivial attack efficacy under camera-relative motion that induces image-plane trigger drift. These results reveal that VLA models can be covertly steered at the granularity of safety-critical actions with minimal poisoning and without observable degradation of nominal performance.", "date": "2025-10-13", "categories": []}
{"id": "2511.10576", "title": "Tight Robustness Certification Through the Convex Hull of $\\ell_0$ Attacks", "abstract": "Few-pixel attacks mislead a classifier by modifying a few pixels of an image. Their perturbation space is an $\\ell_0$-ball, which is not convex, unlike $\\ell_p$-balls for $p\\geq1$. However, existing local robustness verifiers typically scale by relying on linear bound propagation, which captures convex perturbation spaces. We show that the convex hull of an $\\ell_0$-ball is the intersection of its bounding box and an asymmetrically scaled $\\ell_1$-like polytope. The volumes of the convex hull and this polytope are nearly equal as the input dimension increases. We then show a linear bound propagation that precisely computes bounds over the convex hull and is significantly tighter than bound propagations over the bounding box or our $\\ell_1$-like polytope. This bound propagation scales the state-of-the-art $\\ell_0$ verifier on its most challenging robustness benchmarks by 1.24x-7.07x, with a geometric mean of 3.16.", "date": "2025-11-13", "categories": []}
{"id": "2511.19517", "title": "Automating Deception: Scalable Multi-Turn LLM Jailbreaks", "abstract": "Multi-turn conversational attacks, which leverage psychological principles like Foot-in-the-Door (FITD), where a small initial request paves the way for a more significant one, to bypass safety alignments, pose a persistent threat to Large Language Models (LLMs). Progress in defending against these attacks is hindered by a reliance on manual, hard-to-scale dataset creation. This paper introduces a novel, automated pipeline for generating large-scale, psychologically-grounded multi-turn jailbreak datasets. We systematically operationalize FITD techniques into reproducible templates, creating a benchmark of 1,500 scenarios across illegal activities and offensive content. We evaluate seven models from three major LLM families under both multi-turn (with history) and single-turn (without history) conditions. Our results reveal stark differences in contextual robustness: models in the GPT family demonstrate a significant vulnerability to conversational history, with Attack Success Rates (ASR) increasing by as much as 32 percentage points. In contrast, Google's Gemini 2.5 Flash exhibits exceptional resilience, proving nearly immune to these attacks, while Anthropic's Claude 3 Haiku shows strong but imperfect resistance. These findings highlight a critical divergence in how current safety architectures handle conversational context and underscore the need for defenses that can resist narrative-based manipulation.", "date": "2025-11-24", "categories": []}
{"id": "2512.01782", "title": "Dual Randomized Smoothing: Beyond Global Noise Variance", "abstract": "Randomized Smoothing (RS) is a prominent technique for certifying the robustness of neural networks against adversarial perturbations. With RS, achieving high accuracy at small radii requires a small noise variance, while achieving high accuracy at large radii requires a large noise variance. However, the global noise variance used in the standard RS formulation leads to a fundamental limitation: there exists no global noise variance that simultaneously achieves strong performance at both small and large radii. To break through the global variance limitation, we propose a dual RS framework which enables input-dependent noise variances. To achieve that, we first prove that RS remains valid with input-dependent noise variances, provided the variance is locally constant around each input. Building on this result, we introduce two components: (i) a variance estimator predicts an optimal noise variance for each input, (ii) this estimated variance is then used by a standard RS classifier. The variance estimator is independently smoothed via RS to ensure local constancy, enabling flexible design. We also introduce training strategies to iteratively optimize the two components. Experiments on CIFAR-10 demonstrate that our dual RS method provides strong performance for both small and large radii-unattainable with global noise variance-while incurring only a 60% computational overhead at inference. Moreover, it outperforms prior input-dependent noise approaches across most radii, with gains at radii 0.5, 0.75, and 1.0 of 15.6%, 20.0%, and 15.7%. On ImageNet, dual RS remains effective across all radii, with advantages of 8.6%, 17.1%, and 9.1% at radii 0.5, 1.0, and 1.5. Additionally, the dual RS framework provides a routing perspective for certified robustness, improving the accuracy-robustness trade-off with off-the-shelf expert RS models.", "date": "2025-12-01", "categories": []}
{"id": "2603.06621", "title": "Reward Under Attack: Analyzing the Robustness and Hackability of Process Reward Models", "abstract": "Process Reward Models (PRMs) are rapidly becoming the backbone of LLM reasoning pipelines, yet we demonstrate that state-of-the-art PRMs are systematically exploitable under adversarial optimization pressure. To address this, we introduce a three-tiered diagnostic framework that applies increasing adversarial pressure to quantify these vulnerabilities. Static perturbation analysis uncovers a fluency-logic dissociation: high invariance to surface-level style changes reward changes $<$0.1, yet inconsistent detection of logically-corrupted reasoning, with different models failing on different attack types. Adversarial optimization demonstrates that gradient-based attacks inflate rewards on invalid trajectories, with reward landscapes exhibiting wide, exploitable peaks. RL-induced reward hacking exposes the critical failure mode: policies trained on AIME problems achieve near-perfect PRM rewards ($>$0.9), while ground-truth accuracy remains low (below 4%), with 43% of reward gains attributable to stylistic shortcuts. These findings reveal that current PRMs function as fluency detectors rather than reasoning verifiers, creating systematic blind spots that undermine their use as training signals. We release PRM-BiasBench and a diagnostic toolkit to enable robustness evaluation before deployment. The code and dataset are available at https://github.com/SqueezeAILab/reward-under-attack.", "date": "2026-02-20", "categories": []}
{"id": "2603.07590", "title": "Models as Lego Builders: Assembling Malice from Benign Blocks via Semantic Blueprints", "abstract": "Despite the rapid progress of Large Vision-Language Models (LVLMs), the integration of visual modalities introduces new safety vulnerabilities that adversaries can exploit to elicit biased or malicious outputs. In this paper, we demonstrate an underexplored vulnerability via semantic slot filling, where LVLMs complete missing slot values with unsafe content even when the slot types are deliberately crafted to appear benign. Building on this finding, we propose StructAttack, a simple yet effective single-query jailbreak framework under black-box settings. StructAttack decomposes a harmful query into a central topic and a set of benign-looking slot types, then embeds them as structured visual prompts (e.g., mind maps, tables, or sunburst diagrams) with small random perturbations. Paired with a completion-guided instruction, LVLMs automatically recompose the concealed semantics and generate unsafe outputs without triggering safety mechanisms. Although each slot appears benign in isolation (local benignness), StructAttack exploits LVLMs' reasoning to assemble these slots into coherent harmful semantics. Extensive experiments on multiple models and benchmarks show the efficacy of our proposed StructAttack.", "date": "2026-03-08", "categories": []}
{"id": "2603.08498", "title": "All Vehicles Can Lie: Efficient Adversarial Defense in Fully Untrusted-Vehicle Collaborative Perception via Pseudo-Random Bayesian Inference", "abstract": "Collaborative perception (CP) enables multiple vehicles to augment their individual perception capacities through the exchange of feature-level sensory data. However, this fusion mechanism is inherently vulnerable to adversarial attacks, especially in fully untrusted-vehicle environments. Existing defense approaches often assume a trusted ego vehicle as a reference or incorporate additional binary classifiers. These assumptions limit their practicality in real-world deployments due to the questionable trustworthiness of ego vehicles, the requirement for real-time detection, and the need for generalizability across diverse scenarios. To address these challenges, we propose a novel Pseudo-Random Bayesian Inference (PRBI) framework, a first efficient defense method tailored for fully untrusted-vehicle CP. PRBI detects adversarial behavior by leveraging temporal perceptual discrepancies, using the reliable perception from the preceding frame as a dynamic reference. Additionally, it employs a pseudo-random grouping strategy that requires only two verifications per frame, while applying Bayesian inference to estimate both the number and identities of malicious vehicles. Theoretical analysis has proven the convergence and stability of the proposed PRBI framework. Extensive experiments show that PRBI requires only 2.5 verifications per frame on average, outperforming existing methods significantly, and restores detection precision to between 79.4% and 86.9% of pre-attack levels.", "date": "2026-03-09", "categories": []}
{"id": "2502.20403", "title": "Adversarial Robustness of Partitioned Quantum Classifiers", "abstract": "Adversarial robustness in quantum classifiers is a critical area of study, providing insights into their performance compared to classical models and uncovering potential advantages inherent to quantum machine learning. In the NISQ era of quantum computing, circuit cutting is a notable technique for simulating circuits that exceed the qubit limitations of current devices, enabling the distribution of a quantum circuit's execution across multiple quantum processing units through classical communication. In contrast, when quantum communication is available, teleportation-based methods can be used to support the distribution of the quantum circuit. We study the robustness of partitioned quantum classifiers to adversarial perturbations targeting wire cutting or quantum state teleportation and show a link between such perturbations and implementing adversarial gates within intermediate layers of a quantum classifier. We then proceed to study the latter problem from both a theoretical and experimental perspective.", "date": "2025-01-28", "categories": []}
{"id": "2511.13127", "title": "SPARK: Jailbreaking T2V Models by Synergistically Prompting Auditory and Recontextualized Knowledge", "abstract": "Jailbreak attacks can circumvent model safety guardrails and reveal critical blind spots. Prior attacks on text-to-video (T2V) models typically add adversarial perturbations to obviously unsafe prompts, which are often easy to detect and defend. In contrast, we show that benign-looking prompts containing rich, implicit cues can induce T2V models to generate semantically unsafe videos that both violate policy and preserve the original (blocked) intent. To realize this, we propose SPARK, a jailbreak framework that leverages T2V models cross-modal associative patterns via a modular prompt design. Specifically, our prompts combine three components: neutral scene anchors, which provide the surface-level scene description extracted from the blocked intent to maintain plausibility; latent auditory triggers, textual descriptions of innocuous-sounding audio events (e.g., creaking, muffled noises) that exploit learned audio-visual co-occurrence priors to bias the model toward particular unsafe visual concepts; and stylistic modulators, cinematic directives (e.g., camera framing, atmosphere) that amplify and stabilize the latent trigger's effect. We formalize attack generation as a constrained optimization over the above modular prompt space and solve it with a guided search procedure that balances stealth and effectiveness. Extensive experiments over 7 T2V models demonstrate the efficacy of our attack, achieving a +23% improvement in average attack success rate in commercial models.", "date": "2025-11-17", "categories": []}
{"id": "2603.05625", "title": "Identifying Adversary Characteristics from an Observed Attack", "abstract": "When used in automated decision-making systems, machine learning (ML) models are vulnerable to data-manipulation attacks. Some defense mechanisms (e.g., adversarial regularization) directly affect the ML models while others (e.g., anomaly detection) act within the broader system. In this paper we consider a different task for defending the adversary, focusing on the attacker, rather than the attack. We present and demonstrate a framework for identifying characteristics about the attacker from an observed attack.\n  We prove that, without additional knowledge, the attacker is non-identifiable (multiple potential attackers would perform the same observed attack). To address this challenge, we propose a domain-agnostic framework to identify the most probable attacker. This framework aids the defender in two ways. First, knowledge about the attacker can be leveraged for exogenous mitigation (i.e., addressing the vulnerability by altering the decision-making system outside the learning algorithm and/or limiting the attacker's capability). Second, when implementing defense methods that directly affect the learning process (e.g., adversarial regularization), knowledge of the specific attacker improves performance. We present the details of our framework and illustrate its applicability through specific instantiations on a variety of learners.", "date": "2026-03-05", "categories": []}
{"id": "2603.06311", "title": "Latent Transfer Attack: Adversarial Examples via Generative Latent Spaces", "abstract": "Adversarial attacks are a central tool for probing the robustness of modern vision models, yet most methods optimize perturbations directly in pixel space under $\\ell_\\infty$ or $\\ell_2$ constraints. While effective in white-box settings, pixel-space optimization often produces high-frequency, texture-like noise that is brittle to common preprocessing (e.g., resizing and cropping) and transfers poorly across architectures. We propose $\\textbf{LTA}$ ($\\textbf{L}$atent $\\textbf{T}$ransfer $\\textbf{A}$ttack), a transfer-based attack that instead optimizes perturbations in the latent space of a pretrained Stable Diffusion VAE. Given a clean image, we encode it into a latent code and optimize the latent representation to maximize a surrogate classifier loss, while softly enforcing a pixel-space $\\ell_\\infty$ budget after decoding. To improve robustness to resolution mismatch and standard input pipelines, we incorporate Expectation Over Transformations (EOT) via randomized resizing, interpolation, and cropping, and apply periodic latent Gaussian smoothing to suppress emerging artifacts and stabilize optimization. Across a suite of CNN and vision-transformer targets, LTA achieves strong transfer attack success while producing spatially coherent, predominantly low-frequency perturbations that differ qualitatively from pixel-space baselines and occupy a distinct point in the transfer-quality trade-off. Our results highlight pretrained generative latent spaces as an effective and structured domain for adversarial optimization, bridging robustness evaluation with modern generative priors.", "date": "2026-03-06", "categories": []}
{"id": "2603.06508", "title": "When One Modality Rules Them All: Backdoor Modality Collapse in Multimodal Diffusion Models", "abstract": "While diffusion models have revolutionized visual content generation, their rapid adoption has underscored the critical need to investigate vulnerabilities, e.g., to backdoor attacks. In multimodal diffusion models, it is natural to expect that attacking multiple modalities simultaneously (e.g., text and image) would yield complementary effects and strengthen the overall backdoor. In this paper, we challenge this assumption by investigating the phenomenon of Backdoor Modality Collapse, a scenario where the backdoor mechanism degenerates to rely predominantly on a subset of modalities, rendering others redundant. To rigorously quantify this behavior, we introduce two novel metrics: Trigger Modality Attribution (TMA) and Cross-Trigger Interaction (CTI). Through extensive experiments across diverse training configurations in multimodal conditional diffusion, we consistently observe a ``winner-takes-all'' dynamic in backdoor behavior. Our results reveal that (1) attacks often collapse into subset-modality dominance, and (2) cross-modal interaction is negligible or even negative, contradicting the intuition of synergistic vulnerability. These findings highlight a critical blind spot in current assessments, suggesting that high attack success rates often mask a fundamental reliance on a subset of modalities. This establishes a principled foundation for mechanistic analysis and future defense development.", "date": "2026-03-06", "categories": []}
{"id": "2512.01153", "title": "DPAC: Distribution-Preserving Adversarial Control for Diffusion Sampling", "abstract": "Adversarially guided diffusion sampling often achieves the target class, but sample quality degrades as deviations between the adversarially controlled and nominal trajectories accumulate. We formalize this degradation as a path-space Kullback-Leibler divergence(path-KL) between controlled and nominal (uncontrolled) diffusion processes, thereby showing via Girsanov's theorem that it exactly equals the control energy. Building on this stochastic optimal control (SOC) view, we theoretically establish that minimizing this path-KL simultaneously tightens upper bounds on both the 2-Wasserstein distance and Fr\\'echet Inception Distance (FID), revealing a principled connection between adversarial control energy and perceptual fidelity. From a variational perspective, we derive a first-order optimality condition for the control: among all directions that yield the same classification gain, the component tangent to iso-(log-)density surfaces (i.e., orthogonal to the score) minimizes path-KL, whereas the normal component directly increases distributional drift. This leads to DPAC (Distribution-Preserving Adversarial Control), a diffusion guidance rule that projects adversarial gradients onto the tangent space defined by the generative score geometry. We further show that in discrete solvers, the tangent projection cancels the O({\\Delta}t) leading error term in the Wasserstein distance, achieving an O({\\Delta}t^2) quality gap; moreover, it remains second-order robust to score or metric approximation. Empirical studies on ImageNet-100 validate the theoretical predictions, confirming that DPAC achieves lower FID and estimated path-KL at matched attack success rates.", "date": "2025-12-01", "categories": []}
{"id": "2603.04731", "title": "When Priors Backfire: On the Vulnerability of Unlearnable Examples to Pretraining", "abstract": "Unlearnable Examples (UEs) serve as a data protection strategy that generates imperceptible perturbations to mislead models into learning spurious correlations instead of underlying semantics. In this paper, we uncover a fundamental vulnerability of UEs that emerges when learning starts from a pretrained model. Crucially, our empirical analysis shows that even when data are protected by carefully crafted perturbations, pretraining priors still furnish rich semantic representations that allow the model to circumvent the shortcuts introduced by UEs and capture genuine features, thereby nullifying unlearnability. To address this, we propose BAIT (Binding Artificial perturbations to Incorrect Targets), a novel bi-level optimization formulation. Specifically, the inner level aims at associating the perturbed samples with real labels to simulate standard data-label alignment, while the outer level actively disrupts this alignment by enforcing a mislabel-perturbation binding that maps samples to designated incorrect targets. This mechanism effectively overrides the semantic guidance of priors, forcing the model to rely on the injected perturbations and consequently preventing the acquisition of true semantics. Extensive experiments on standard benchmarks and multiple pretrained backbones demonstrate that BAIT effectively mitigates the influence of pretraining priors and maintains data unlearnability.", "date": "2026-03-05", "categories": []}
{"id": "2603.04859", "title": "Osmosis Distillation: Model Hijacking with the Fewest Samples", "abstract": "Transfer learning is devised to leverage knowledge from pre-trained models to solve new tasks with limited data and computational resources. Meanwhile, dataset distillation has emerged to synthesize a compact dataset that preserves critical information from the original large dataset. Therefore, a combination of transfer learning and dataset distillation offers promising performance in evaluations. However, a non-negligible security threat remains undiscovered in transfer learning using synthetic datasets generated by dataset distillation methods, where an adversary can perform a model hijacking attack with only a few poisoned samples in the synthetic dataset. To reveal this threat, we propose Osmosis Distillation (OD) attack, a novel model hijacking strategy that targets deep learning models using the fewest samples. Comprehensive evaluations on various datasets demonstrate that the OD attack attains high attack success rates in hidden tasks while preserving high model utility in original tasks. Furthermore, the distilled osmosis set enables model hijacking across diverse model architectures, allowing model hijacking in transfer learning with considerable attack performance and model utility. We argue that awareness of using third-party synthetic datasets in transfer learning must be raised.", "date": "2026-03-05", "categories": []}
{"id": "2603.04913", "title": "Beyond the Patch: Exploring Vulnerabilities of Visuomotor Policies via Viewpoint-Consistent 3D Adversarial Object", "abstract": "Neural network-based visuomotor policies enable robots to perform manipulation tasks but remain susceptible to perceptual attacks. For example, conventional 2D adversarial patches are effective under fixed-camera setups, where appearance is relatively consistent; however, their efficacy often diminishes under dynamic viewpoints from moving cameras, such as wrist-mounted setups, due to perspective distortions. To proactively investigate potential vulnerabilities beyond 2D patches, this work proposes a viewpoint-consistent adversarial texture optimization method for 3D objects through differentiable rendering. As optimization strategies, we employ Expectation over Transformation (EOT) with a Coarse-to-Fine (C2F) curriculum, exploiting distance-dependent frequency characteristics to induce textures effective across varying camera-object distances. We further integrate saliency-guided perturbations to redirect policy attention and design a targeted loss that persistently drives robots toward adversarial objects. Our comprehensive experiments show that the proposed method is effective under various environmental conditions, while confirming its black-box transferability and real-world applicability.", "date": "2026-03-05", "categories": []}
{"id": "2504.20376", "title": "When Memory Becomes a Vulnerability: Towards Multi-turn Jailbreak Attacks against Text-to-Image Generation Systems", "abstract": "Modern text-to-image (T2I) generation systems (e.g., DALL$\\cdot$E 3) exploit the memory mechanism, which captures key information in multi-turn interactions for faithful generation. Despite its practicality, the security analyses of this mechanism have fallen far behind. In this paper, we reveal that it can exacerbate the risk of jailbreak attacks. Previous attacks fuse the unsafe target prompt into one ultimate adversarial prompt, which can be easily detected or lead to the generation of non-unsafe images due to under- or over-detoxification. In contrast, we propose embedding the malice at the inception of the chat session in memory, addressing the above limitations.\n  Specifically, we propose Inception, the first multi-turn jailbreak attack against real-world text-to-image generation systems that explicitly exploits their memory mechanisms. Inception is composed of two key modules: segmentation and recursion. We introduce Segmentation, a semantic-preserving method that generates multi-round prompts. By leveraging NLP analysis techniques, we design policies to decompose a prompt, together with its malicious intent, according to sentence structure, thereby evading safety filters. Recursion further addresses the challenge posed by unsafe sub-prompts that cannot be separated through simple segmentation. It firstly expands the sub-prompt, then invokes segmentation recursively. To facilitate multi-turn adversarial prompts crafting, we build VisionFlow, an emulation T2I system that integrates two-stage safety filters and industrial-grade memory mechanisms. The experiment results show that Inception successfully allures unsafe image generation, surpassing the SOTA by a 20.0\\% margin in attack success rate. We also conduct experiments on the real-world commercial T2I generation platforms, further validating the threats of Inception in practice.", "date": "2025-04-29", "categories": []}
{"id": "2602.22227", "title": "Dynamic Adversarial Reinforcement Learning for Robust Multimodal Large Language Models", "abstract": "Despite their impressive capabilities, Multimodal Large Language Models (MLLMs) exhibit perceptual fragility when confronted with visually complex scenes. This weakness stems from a reliance on finite training datasets, which are prohibitively expensive to scale and impose a ceiling on model robustness. We introduce \\textbf{AOT-SFT}, a large-scale adversarial dataset for bootstrapping MLLM robustness. Building on this, we propose \\textbf{AOT (Adversarial Opponent Training)}, a self-play framework that forges MLLM robustness by creating its own training data. Our method orchestrates a co-evolution between an image-editing Attacker and a Defender MLLM, where the Attacker generates a diverse and dynamic curriculum of image manipulations, forcing the Defender to adapt and improve. Extensive experiments demonstrate that AOT enhances the Defender's perceptual robustness and reduces hallucinations, establishing a scalable paradigm for training more reliable MLLMs.", "date": "2026-01-24", "categories": []}
{"id": "2603.04064", "title": "Tuning Just Enough: Lightweight Backdoor Attacks on Multi-Encoder Diffusion Models", "abstract": "As text-to-image diffusion models become increasingly deployed in real-world applications, concerns about backdoor attacks have gained significant attention. Prior work on text-based backdoor attacks has largely focused on diffusion models conditioned on a single lightweight text encoder. However, more recent diffusion models that incorporate multiple large-scale text encoders remain underexplored in this context. Given the substantially increased number of trainable parameters introduced by multiple text encoders, an important question is whether backdoor attacks can remain both efficient and effective in such settings. In this work, we study Stable Diffusion 3, which uses three distinct text encoders and has not yet been systematically analyzed for text-encoder-based backdoor vulnerabilities. To understand the role of text encoders in backdoor attacks, we define four categories of attack targets and identify the minimal sets of encoders required to achieve effective performance for each attack objective. Based on this, we further propose Multi-Encoder Lightweight aTtacks (MELT), which trains only low-rank adapters while keeping the pretrained text encoder weight frozen. We demonstrate that tuning fewer than 0.2% of the total encoder parameters is sufficient for successful backdoor attacks on Stable Diffusion 3, revealing previously underexplored vulnerabilities in practical attack scenarios in multi-encoder settings.", "date": "2026-03-04", "categories": []}
{"id": "2603.04364", "title": "Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks", "abstract": "Multimodal web agents that process both screenshots and accessibility trees are increasingly deployed to interact with web interfaces, yet their dual-stream architecture opens an underexplored attack surface: an adversary who injects content into the webpage DOM simultaneously corrupts both observation channels with a consistent deceptive narrative. Our vulnerability analysis on MiniWob++ reveals that attacks including a visual component far outperform text-only injections, exposing critical gaps in text-centric VLM safety training. Motivated by this finding, we propose Dual-Modality Multi-Stage Adversarial Safety Training (DMAST), a framework that formalizes the agent-attacker interaction as a two-player zero-sum Markov game and co-trains both players through a three-stage pipeline: (1) imitation learning from a strong teacher model, (2) oracle-guided supervised fine-tuning that uses a novel zero-acknowledgment strategy to instill task-focused reasoning under adversarial noise, and (3) adversarial reinforcement learning via Group Relative Policy Optimization (GRPO) self-play. On out-of-distribution tasks, DMAST substantially mitigates adversarial risks while simultaneously doubling task completion efficiency. Our approach significantly outperforms established training-based and prompt-based defenses, demonstrating genuine co-evolutionary progress and robust generalization to complex, unseen environments.", "date": "2026-03-04", "categories": []}
{"id": "2303.08032", "title": "Verifying the Robustness of Automatic Credibility Assessment", "abstract": "Text classification methods have been widely investigated as a way to detect content of low credibility: fake news, social media bots, propaganda, etc. Quite accurate models (likely based on deep neural networks) help in moderating public electronic platforms and often cause content creators to face rejection of their submissions or removal of already published texts. Having the incentive to evade further detection, content creators try to come up with a slightly modified version of the text (known as an attack with an adversarial example) that exploit the weaknesses of classifiers and result in a different output. Here we systematically test the robustness of common text classifiers against available attacking techniques and discover that, indeed, meaning-preserving changes in input text can mislead the models. The approaches we test focus on finding vulnerable spans in text and replacing individual characters or words, taking into account the similarity between the original and replacement content. We also introduce BODEGA: a benchmark for testing both victim models and attack methods on four misinformation detection tasks in an evaluation framework designed to simulate real use-cases of content moderation. The attacked tasks include (1) fact checking and detection of (2) hyperpartisan news, (3) propaganda and (4) rumours. Our experimental results show that modern large language models are often more vulnerable to attacks than previous, smaller solutions, e.g. attacks on GEMMA being up to 27\\% more successful than those on BERT. Finally, we manually analyse a subset adversarial examples and check what kinds of modifications are used in successful attacks.", "date": "2023-03-14", "categories": []}
{"id": "2411.11677", "title": "Few-shot Model Extraction Attacks against Sequential Recommender Systems", "abstract": "Among adversarial attacks against sequential recommender systems, model extraction attacks represent a method to attack sequential recommendation models without prior knowledge. Existing research has primarily concentrated on the adversary's execution of black-box attacks through data-free model extraction. However, a significant gap remains in the literature concerning the development of surrogate models by adversaries with access to few-shot raw data (10\\% even less). That is, the challenge of how to construct a surrogate model with high functional similarity within the context of few-shot data scenarios remains an issue that requires resolution.This study addresses this gap by introducing a novel few-shot model extraction framework against sequential recommenders, which is designed to construct a superior surrogate model with the utilization of few-shot data. The proposed few-shot model extraction framework is comprised of two components: an autoregressive augmentation generation strategy and a bidirectional repair loss-facilitated model distillation procedure. Specifically, to generate synthetic data that closely approximate the distribution of raw data, autoregressive augmentation generation strategy integrates a probabilistic interaction sampler to extract inherent dependencies and a synthesis determinant signal module to characterize user behavioral patterns. Subsequently, bidirectional repair loss, which target the discrepancies between the recommendation lists, is designed as auxiliary loss to rectify erroneous predictions from surrogate models, transferring knowledge from the victim model to the surrogate model effectively. Experiments on three datasets show that the proposed few-shot model extraction framework yields superior surrogate models.", "date": "2024-11-18", "categories": []}
{"id": "2502.20314", "title": "Adversarial Attacks in Weight-Space Classifiers", "abstract": "Implicit Neural Representations (INRs) have been recently garnering increasing interest in various research fields, mainly due to their ability to represent large, complex data in a compact, continuous manner. Past work further showed that numerous popular downstream tasks can be performed directly in the INR parameter-space. Doing so can substantially reduce the computational resources required to process the represented data in their native domain. A major difficulty in using modern machine-learning approaches, is their high susceptibility to adversarial attacks, which have been shown to greatly limit the reliability and applicability of such methods in a wide range of settings. In this work, we perform an in-depth security analysis of the behavior of weight-space classifiers under adversarial attacks. Our study reveals that parameter-space models trained for classification exhibit increased robustness to standard white-box adversarial attacks compared to standard classifiers that operate in the original signal space. This is achieved without the need of any robust training. We source this robust behavior to the phenomenon of gradient-obfuscation promoted during the INR optimization process, and pinpoint the limitations of this robustness under alternative adversarial approaches. To support our claims, we develop a novel suite of adversarial attacks targeting parameter-space classifiers, and furthermore analyze practical considerations of such attacks.", "date": "2025-02-27", "categories": []}
{"id": "2503.12567", "title": "GAN-Based Single-Stage Defense for Traffic Sign Classification Under Adversarial Patch", "abstract": "Computer vision plays a critical role in ensuring the safe navigation of autonomous vehicles (AVs). An AV perception module facilitates safe navigation. This module enables AVs to recognize traffic signs, traffic lights, and various road users. However, the perception module is vulnerable to adversarial attacks, which can compromise its accuracy and reliability. One such attack is the adversarial patch attack (APA), an attack in which an adversary strategically places a specially crafted sticker on an object to deceive object classifiers. Such an APA can cause AVs to misclassify traffic signs, leading to catastrophic incidents. To enhance the security of an AV perception system against APAs, this study develops a Generative Adversarial Network (GAN)-based single-stage defense strategy for traffic sign classification. This approach is tailored to defend against APAs across different classes of traffic signs, without prior knowledge of a patch's design, and is effective against patches of varying sizes. In addition, our single-stage defense is computationally efficient, requiring significantly lower computation time than existing multi-stage defenses, making it suitable for real-time deployment in autonomous driving systems. Compared to a classifier without any defense mechanism, our experimental analysis demonstrates that the defense strategy presented in this paper improves our classifier's accuracy under APA conditions by up to 90% considering the traffic sign classes considered in this study. and overall classification accuracy is enhanced by 55% for all traffic signs considered in this study. Our defense strategy is model agnostic, making it applicable to any traffic sign classifier, regardless of the underlying classification model.", "date": "2025-03-16", "categories": []}
{"id": "2505.20934", "title": "NatADiff: Adversarial Boundary Guidance for Natural Adversarial Diffusion", "abstract": "Adversarial samples exploit irregularities in the manifold `learned' by deep learning models to cause misclassifications. The study of these adversarial samples provides insight into the features a model uses to classify inputs, which can be leveraged to improve robustness against future attacks. However, much of the existing literature focuses on constrained adversarial samples, which do not accurately reflect test-time errors encountered in real-world settings. To address this, we propose `NatADiff', an adversarial sampling scheme that leverages denoising diffusion to generate natural adversarial samples. Our approach is based on the observation that natural adversarial samples frequently contain structural elements from the adversarial class. Deep learning models can exploit these structural elements to shortcut the classification process, rather than learning to genuinely distinguish between classes. To leverage this behavior, we guide the diffusion trajectory towards the intersection of the true and adversarial classes, combining time-travel sampling with augmented classifier guidance to enhance attack transferability while preserving image fidelity. Our method achieves comparable attack success rates to current state-of-the-art techniques, while exhibiting significantly higher transferability across model architectures and better alignment with natural test-time errors as measured by FID. These results demonstrate that NatADiff produces adversarial samples that not only transfer more effectively across models, but more faithfully resemble naturally occurring test-time errors.", "date": "2025-05-27", "categories": []}
{"id": "2505.22499", "title": "SABER: Spatially Consistent 3D Universal Adversarial Objects for BEV Detectors", "abstract": "Adversarial robustness of BEV 3D object detectors is critical for autonomous driving (AD). Existing invasive attacks require altering the target vehicle itself (e.g. attaching patches), making them unrealistic and impractical for real-world evaluation. While non-invasive attacks that place adversarial objects in the environment are more practical, current methods still lack the multi-view and temporal consistency needed for physically plausible threats. In this paper, we present the first framework for generating universal, non-invasive, and 3D-consistent adversarial objects that expose fundamental vulnerabilities for BEV 3D object detectors. Instead of modifying target vehicles, our method inserts rendered objects into scenes with an occlusion-aware module that enforces physical plausibility across views and time. To maintain attack effectiveness across views and frames, we optimize adversarial object appearance using a BEV spatial feature-guided optimization strategy that attacks the detector's internal representations. Extensive experiments demonstrate that our learned universal adversarial objects can consistently degrade multiple BEV detectors from various viewpoints and distances. More importantly, the new environment-manipulation attack paradigm exposes models' over-reliance on contextual cues and provides a practical pipeline for robustness evaluation in AD systems.", "date": "2025-05-28", "categories": []}
{"id": "2505.14042", "title": "Adversarially Pretrained Transformers May Be Universally Robust In-Context Learners", "abstract": "Adversarial training is one of the most effective defenses against adversarial attacks, but it incurs a high computational cost. In this study, we present the first theoretical analysis suggesting that adversarially pretrained transformers can serve as universally robust foundation models -- models that can adapt robustly to diverse downstream tasks with only lightweight tuning. Specifically, we demonstrate that single-layer linear transformers, after adversarial pretraining across a variety of classification tasks, can generalize robustly to unseen classification tasks through in-context learning from clean demonstrations (i.e., without requiring additional adversarial training or examples). This universal robustness stems from the model's ability to adaptively focus on robust features within given tasks. We also identify two open challenges for attaining robustness: the accuracy-robustness trade-off and sample-hungry training. This study initiates the discussion on the utility of universally robust foundation models. While their training is expensive, the investment would prove worthwhile as downstream tasks can obtain adversarial robustness for free. The code is available at https://github.com/s-kumano/universally-robust-in-context-learner.", "date": "2025-05-20", "categories": []}
{"id": "2505.17132", "title": "Dynamic Token Reweighting for Robust Vision-Language Models", "abstract": "Large vision-language models (VLMs) are highly vulnerable to multimodal jailbreak attacks that exploit visual-textual interactions to bypass safety guardrails. In this paper, we present DTR, a novel inference-time defense that mitigates multimodal jailbreak attacks through optimizing the model's key-value (KV) caches. Rather than relying on curated safety-specific data or costly image-to-text conversion, we introduce a new formulation of the safety-relevant distributional shift induced by the visual modality. This formulation enables DTR to dynamically adjust visual token weights, minimizing the impact of adversarial visual inputs while preserving the model's general capabilities and inference efficiency. Extensive evaluation across diverse VLMs and attack benchmarks demonstrates that DTR outperforms existing defenses in both attack robustness and benign-task performance, marking the first successful application of KV cache optimization for safety enhancement in multimodal foundation models. The code for replicating DTR is available at: https://github.com/TanqiuJiang/DTR.", "date": "2025-05-22", "categories": []}
{"id": "2509.21029", "title": "FORCE: Transferable Visual Jailbreaking Attacks via Feature Over-Reliance CorrEction", "abstract": "The integration of new modalities enhances the capabilities of multimodal large language models (MLLMs) but also introduces additional vulnerabilities. In particular, simple visual jailbreaking attacks can manipulate open-source MLLMs more readily than sophisticated textual attacks. However, these underdeveloped attacks exhibit extremely limited cross-model transferability, failing to reliably identify vulnerabilities in closed-source MLLMs. In this work, we analyse the loss landscape of these jailbreaking attacks and find that the generated attacks tend to reside in high-sharpness regions, whose effectiveness is highly sensitive to even minor parameter changes during transfer. To further explain the high-sharpness localisations, we analyse their feature representations in both the intermediate layers and the spectral domain, revealing an improper reliance on narrow layer representations and semantically poor frequency components. Building on this, we propose a Feature Over-Reliance CorrEction (FORCE) method, which guides the attack to explore broader feasible regions across layer features and rescales the influence of frequency features according to their semantic content. By eliminating non-generalizable reliance on both layer and spectral features, our method discovers flattened feasible regions for visual jailbreaking attacks, thereby improving cross-model transferability. Extensive experiments demonstrate that our approach effectively facilitates visual red-teaming evaluations against closed-source MLLMs.", "date": "2025-09-25", "categories": []}
{"id": "2510.02999", "title": "Untargeted Jailbreak Attack", "abstract": "Existing gradient-based jailbreak attacks on Large Language Models (LLMs) typically optimize adversarial suffixes to align the LLM output with predefined target responses. However, restricting the objective as inducing fixed targets inherently constrains the adversarial search space, limiting the overall attack efficacy. Furthermore, existing methods typically require numerous optimization iterations to fulfill the large gap between the fixed target and the original LLM output, resulting in low attack efficiency. To overcome these limitations, we propose the first gradient-based untargeted jailbreak attack (UJA), which relies on an untargeted objective to maximize the unsafety probability of the LLM output, without enforcing any response patterns. For tractable optimization, we further decompose this objective into two differentiable sub-objectives to search the optimal harmful response and the corresponding adversarial prompt, with a theoretical analysis to validate the decomposition. In contrast to existing attacks, UJA's unrestricted objective significantly expands the search space, enabling more flexible and efficient exploration of LLM vulnerabilities. Extensive evaluations show that UJA achieves over 80\\% attack success rates against recent safety-aligned LLMs with only 100 optimization iterations, outperforming the state-of-the-art gradient-based attacks by over 30\\%.", "date": "2025-10-03", "categories": []}
{"id": "2512.14341", "title": "Towards Transferable Defense Against Malicious Image Edits", "abstract": "Recent approaches employing imperceptible perturbations in input images have demonstrated promising potential to counter malicious manipulations in diffusion-based image editing systems. However, existing methods suffer from limited transferability in cross-model evaluations. To address this, we propose Transferable Defense Against Malicious Image Edits (TDAE), a novel bimodal framework that enhances image immunity against malicious edits through coordinated image-text optimization. Specifically, at the visual defense level, we introduce FlatGrad Defense Mechanism (FDM), which incorporates gradient regularization into the adversarial objective. By explicitly steering the perturbations toward flat minima, FDM amplifies immune robustness against unseen editing models. For textual enhancement protection, we propose an adversarial optimization paradigm named Dynamic Prompt Defense (DPD), which periodically refines text embeddings to align the editing outcomes of immunized images with those of the original images, then updates the images under optimized embeddings. Through iterative adversarial updates to diverse embeddings, DPD enforces the generation of immunized images that seek a broader set of immunity-enhancing features, thereby achieving cross-model transferability. Extensive experimental results demonstrate that our TDAE achieves state-of-the-art performance in mitigating malicious edits under both intra- and cross-model evaluations.", "date": "2025-12-16", "categories": []}
{"id": "2602.07152", "title": "Trojans in Artificial Intelligence (TrojAI) Final Report", "abstract": "The Intelligence Advanced Research Projects Activity (IARPA) launched the TrojAI program to confront an emerging vulnerability in modern artificial intelligence: the threat of AI Trojans. These AI trojans are malicious, hidden backdoors intentionally embedded within an AI model that can cause a system to fail in unexpected ways, or allow a malicious actor to hijack the AI model at will. This multi-year initiative helped to map out the complex nature of the threat, pioneered foundational detection methods, and identified unsolved challenges that require ongoing attention by the burgeoning AI security field. This report synthesizes the program's key findings, including methodologies for detection through weight analysis and trigger inversion, as well as approaches for mitigating Trojan risks in deployed models. Comprehensive test and evaluation results highlight detector performance, sensitivity, and the prevalence of \"natural\" Trojans. The report concludes with lessons learned and recommendations for advancing AI security research.", "date": "2026-02-06", "categories": []}
{"id": "2602.19096", "title": "The Power of Decaying Steps: Enhancing Attack Stability and Transferability for Sign-based Optimizers", "abstract": "Crafting adversarial examples can be formulated as an optimization problem. While sign-based optimizers such as I-FGSM and MI-FGSM have become the de facto standard for the induced optimization problems, there still exist several unsolved problems in theoretical grounding and practical reliability especially in non-convergence and instability, which inevitably influences their transferability. Contrary to the expectation, we observe that the attack success rate may degrade sharply when more number of iterations are conducted. In this paper, we address these issues from an optimization perspective. By reformulating the sign-based optimizer as a specific coordinate-wise gradient descent, we argue that one cause for non-convergence and instability is their non-decaying step-size scheduling. Based upon this viewpoint, we propose a series of new attack algorithms that enforce Monotonically Decreasing Coordinate-wise Step-sizes (MDCS) within sign-based optimizers. Typically, we further provide theoretical guarantees proving that MDCS-MI attains an optimal convergence rate of $O(1/\\sqrt{T})$, where $T$ is the number of iterations. Extensive experiments on image classification and cross-modal retrieval tasks demonstrate that our approach not only significantly improves transferability but also enhances attack stability compared to state-of-the-art sign-based methods.", "date": "2026-02-22", "categories": []}
{"id": "2602.20999", "title": "VII: Visual Instruction Injection for Jailbreaking Image-to-Video Generation Models", "abstract": "Image-to-Video (I2V) generation models, which condition video generation on reference images, have shown emerging visual instruction-following capability, allowing certain visual cues in reference images to act as implicit control signals for video generation. However, this capability also introduces a previously overlooked risk: adversaries may exploit visual instructions to inject malicious intent through the image modality. In this work, we uncover this risk by proposing Visual Instruction Injection (VII), a training-free and transferable jailbreaking framework that intentionally disguises the malicious intent of unsafe text prompts as benign visual instructions in the safe reference image. Specifically, VII coordinates a Malicious Intent Reprogramming module to distill malicious intent from unsafe text prompts while minimizing their static harmfulness, and a Visual Instruction Grounding module to ground the distilled intent onto a safe input image by rendering visual instructions that preserve semantic consistency with the original unsafe text prompt, thereby inducing harmful content during I2V generation. Empirically, our extensive experiments on four state-of-the-art commercial I2V models (Kling-v2.5-turbo, Gemini Veo-3.1, Seedance-1.5-pro, and PixVerse-V5) demonstrate that VII achieves Attack Success Rates of up to 83.5% while reducing Refusal Rates to near zero, significantly outperforming existing baselines.", "date": "2026-02-24", "categories": []}
{"id": "2603.00172", "title": "Hidden in the Metadata: Stealth Poisoning Attacks on Multimodal Retrieval-Augmented Generation", "abstract": "Retrieval-augmented generation (RAG) has emerged as a powerful paradigm for enhancing multimodal large language models by grounding their responses in external, factual knowledge and thus mitigating hallucinations. However, the integration of externally sourced knowledge bases introduces a critical attack surface. Adversaries can inject malicious multimodal content capable of influencing both retrieval and downstream generation. In this work, we present MM-MEPA, a multimodal poisoning attack that targets the metadata components of image-text entries while leaving the associated visual content unaltered. By only manipulating the metadata, MM-MEPA can still steer multimodal retrieval and induce attacker-desired model responses. We evaluate the attack across multiple benchmark settings and demonstrate its severity. MM-MEPA achieves an attack success rate of up to 91\\% consistently disrupting system behaviors across four retrievers and two multimodal generators. Additionally, we assess representative defense strategies and find them largely ineffective against this form of metadata-only poisoning. Our findings expose a critical vulnerability in multimodal RAG and underscore the urgent need for more robust, defense-aware retrieval and knowledge integration methods.", "date": "2026-02-26", "categories": []}
{"id": "2603.00217", "title": "Physical Evaluation of Naturalistic Adversarial Patches for Camera-Based Traffic-Sign Detection", "abstract": "This paper studies how well Naturalistic Adversarial Patches (NAPs) transfer to a physical traffic sign setting when the detector is trained on a customized dataset for an autonomous vehicle (AV) environment. We construct a composite dataset, CompGTSRB (which is customized dataset for AV environment), by pasting traffic sign instances from the German Traffic Sign Recognition Benchmark (GTSRB) onto undistorted backgrounds captured from the target platform. CompGTSRB is used to train a YOLOv5 model and generate patches using a Generative Adversarial Network (GAN) with latent space optimization, following existing NAP methods. We carried out a series of experiments on our Quanser QCar testbed utilizing the front CSI camera provided in QCar. Across configurations, NAPs reduce the detector's STOP class confidence. Different configurations include distance, patch sizes, and patch placement. These results along with a detailed step-by-step methodology indicate the utility of CompGTSRB dataset and the proposed systematic physical protocols for credible patch evaluation. The research further motivate researching the defenses that address localized patch corruption in embedded perception pipelines.", "date": "2026-02-27", "categories": []}
{"id": "2603.00266", "title": "Adversarial Patch Generation for Visual-Infrared Dense Prediction Tasks via Joint Position-Color Optimization", "abstract": "Multimodal adversarial attacks for dense prediction remain largely underexplored. In particular, visual-infrared (VI) perception systems introduce unique challenges due to heterogeneous spectral characteristics and modality-specific intensity distributions. Existing adversarial patch methods are primarily designed for single-modal inputs and fail to account for crossspectral inconsistencies, leading to reduced attack effectiveness and poor stealthiness when applied to VI dense prediction models. To address these challenges, we propose a joint position-color optimization framework (AP-PCO) for generating adversarial patches in visual-infrared settings. The proposed method optimizes patch placement and color composition simultaneously using a fitness function derived from model outputs, enabling a single patch to perturb both visible and infrared modalities. To further bridge spectral discrepancies, we introduce a crossmodal color adaptation strategy that constrains patch appearance according to infrared grayscale characteristics while maintaining strong perturbations in the visible domain, thereby reducing cross-spectral saliency. The optimization procedure operates without requiring internal model information, supporting flexible black-box attacks. Extensive experiments on visual-infrared dense prediction tasks demonstrate that the proposed AP-PCO achieves consistently strong attack performance across multiple architectures, providing a practical benchmark for robustness evaluation in VI perception systems.", "date": "2026-02-27", "categories": []}
{"id": "2603.00481", "title": "Analyzing Physical Adversarial Example Threats to Machine Learning in Election Systems", "abstract": "Developments in the machine learning voting domain have shown both promising results and risks. Trained models perform well on ballot classification tasks (> 99% accuracy) but are at risk from adversarial example attacks that cause misclassifications. In this paper, we analyze an attacker who seeks to deploy adversarial examples against machine learning ballot classifiers to compromise a U.S. election. We first derive a probabilistic framework for determining the number of adversarial example ballots that must be printed to flip an election, in terms of the probability of each candidate winning and the total number of ballots cast. Second, it is an open question as to which type of adversarial example is most effective when physically printed in the voting domain. We analyze six different types of adversarial example attacks: l_infinity-APGD, l2-APGD, l1-APGD, l0 PGD, l0 + l_infinity PGD, and l0 + sigma-map PGD. Our experiments include physical realizations of 144,000 adversarial examples through printing and scanning with four different machine learning models. We empirically demonstrate an analysis gap between the physical and digital domains, wherein attacks most effective in the digital domain (l2 and l_infinity) differ from those most effective in the physical domain (l1 and l2, depending on the model). By unifying a probabilistic election framework with digital and physical adversarial example evaluations, we move beyond prior close race analyses to explicitly quantify when and how adversarial ballot manipulation could alter outcomes.", "date": "2026-02-28", "categories": []}
{"id": "2603.00529", "title": "CaptionFool: Universal Image Captioning Model Attacks", "abstract": "Image captioning models are encoder-decoder architectures trained on large-scale image-text datasets, making them susceptible to adversarial attacks. We present CaptionFool, a novel universal (input-agnostic) adversarial attack against state-of-the-art transformer-based captioning models. By modifying only 7 out of 577 image patches (approximately 1.2% of the image), our attack achieves 94-96% success rate in generating arbitrary target captions, including offensive content. We further demonstrate that CaptionFool can generate \"slang\" terms specifically designed to evade existing content moderation filters. Our findings expose critical vulnerabilities in deployed vision-language models and underscore the urgent need for robust defenses against such attacks. Warning: This paper contains model outputs which are offensive in nature.", "date": "2026-02-28", "categories": []}
{"id": "2603.00567", "title": "Learning to Attack: A Bandit Approach to Adversarial Context Poisoning", "abstract": "Neural contextual bandits are vulnerable to adversarial attacks, where subtle perturbations to rewards, actions, or contexts induce suboptimal decisions. We introduce AdvBandit, a black-box adaptive attack that formulates context poisoning as a continuous-armed bandit problem, enabling the attacker to jointly learn and exploit the victim's evolving policy. The attacker requires no access to the victim's internal parameters, reward function, or gradient information; instead, it constructs a surrogate model using a maximum-entropy inverse reinforcement learning module from observed context-action pairs and optimizes perturbations against this surrogate using projected gradient descent. An upper confidence bound-aware Gaussian process guides arm selection. An attack-budget control mechanism is also introduced to limit detection risk and overhead. We provide theoretical guarantees, including sublinear attacker regret and lower bounds on victim regret linear in the number of attacks. Experiments on three real-world datasets (Yelp, MovieLens, and Disin) against various victim contextual bandits demonstrate that our attack model achieves higher cumulative victim regret than state-of-the-art baselines.", "date": "2026-02-28", "categories": []}
{"id": "2603.00711", "title": "IU: Imperceptible Universal Backdoor Attack", "abstract": "Backdoor attacks pose a critical threat to the security of deep neural networks, yet existing efforts on universal backdoors often rely on visually salient patterns, making them easier to detect and less practical at scale. In this work, we introduce a novel imperceptible universal backdoor attack that simultaneously controls all target classes with minimal poisoning while preserving stealth. Our key idea is to leverage graph convolutional networks (GCNs) to model inter-class relationships and generate class-specific perturbations that are both effective and visually invisible. The proposed framework optimizes a dual-objective loss that balances stealthiness (measured by perceptual similarity metrics such as PSNR) and attack success rate (ASR), enabling scalable, multi-target backdoor injection. Extensive experiments on ImageNet-1K with ResNet architectures demonstrate that our method achieves high ASR (up to 91.3%) under poisoning rates as low as 0.16%, while maintaining benign accuracy and evading state-of-the-art defenses. These results highlight the emerging risks of invisible universal backdoors and call for more robust detection and mitigation strategies.", "date": "2026-02-28", "categories": []}
{"id": "2603.00859", "title": "AMDS: Attack-Aware Multi-Stage Defense System for Network Intrusion Detection with Two-Stage Adaptive Weight Learning", "abstract": "Machine learning based network intrusion detection systems are vulnerable to adversarial attacks that degrade classification performance under both gradient-based and distribution shift threat models. Existing defenses typically apply uniform detection strategies, which may not account for heterogeneous attack characteristics. This paper proposes an attack-aware multi-stage defense framework that learns attack-specific detection strategies through a weighted combination of ensemble disagreement, predictive uncertainty, and distributional anomaly signals. Empirical analysis across seven adversarial attack types reveals distinct detection signatures, enabling a two-stage adaptive detection mechanism. Experimental evaluation on a benchmark intrusion detection dataset indicates that the proposed system attains 94.2% area under the receiver operating characteristic curve and improves classification accuracy by 4.5 percentage points and F1-score by 9.0 points over adversarially trained ensembles. Under adaptive white-box attacks with full architectural knowledge, the system appears to maintain 94.4% accuracy with a 4.2% attack success rate, though this evaluation is limited to two adaptive variants and does not constitute a formal robustness guarantee. Cross-dataset validation further suggests that defense effectiveness depends on baseline classifier competence and may vary with feature dimensionality. These results suggest that attack-specific optimization combined with multi-signal integration can provide a practical approach to improving adversarial robustness in machine learning-based intrusion detection systems.", "date": "2026-03-01", "categories": []}
{"id": "2603.01019", "title": "BadRSSD: Backdoor Attacks on Regularized Self-Supervised Diffusion Models", "abstract": "Self-supervised diffusion models learn high-quality visual representations via latent space denoising. However, their representation layer poses a distinct threat: unlike traditional attacks targeting generative outputs, its unconstrained latent semantic space allows for stealthy backdoors, permitting malicious control upon triggering. In this paper, we propose BadRSSD, the first backdoor attack targeting the representation layer of self-supervised diffusion models. Specifically, it hijacks the semantic representations of poisoned samples with triggers in Principal Component Analysis (PCA) space toward those of a target image, then controls the denoising trajectory during diffusion by applying coordinated constraints across latent, pixel, and feature distribution spaces to steer the model toward generating the specified target. Additionally, we integrate representation dispersion regularization into the constraint framework to maintain feature space uniformity, significantly enhancing attack stealth. This approach preserves normal model functionality (high utility) while achieving precise target generation upon trigger activation (high specificity). Experiments on multiple benchmark datasets demonstrate that BadRSSD substantially outperforms existing attacks in both FID and MSE metrics, reliably establishing backdoors across different architectures and configurations, and effectively resisting state-of-the-art backdoor defenses.", "date": "2026-03-01", "categories": []}
{"id": "2603.01264", "title": "S2O: Enhancing Adversarial Training with Second-Order Statistics of Weights", "abstract": "Adversarial training has emerged as a highly effective way to improve the robustness of deep neural networks (DNNs). It is typically conceptualized as a min-max optimization problem over model weights and adversarial perturbations, where the weights are optimized using gradient descent methods, such as SGD. In this paper, we propose a novel approach by treating model weights as random variables, which paves the way for enhancing adversarial training through \\textbf{S}econd-Order \\textbf{S}tatistics \\textbf{O}ptimization (S$^2$O) over model weights. We challenge and relax a prevalent, yet often unrealistic, assumption in prior PAC-Bayesian frameworks: the statistical independence of weights. From this relaxation, we derive an improved PAC-Bayesian robust generalization bound. Our theoretical developments suggest that optimizing the second-order statistics of weights can substantially tighten this bound. We complement this theoretical insight by conducting an extensive set of experiments that demonstrate that S$^2$O not only enhances the robustness and generalization of neural networks when used in isolation, but also seamlessly augments other state-of-the-art adversarial training techniques. The code is available at https://github.com/Alexkael/S2O.", "date": "2026-03-01", "categories": []}
{"id": "2603.01454", "title": "VidDoS: Universal Denial-of-Service Attack on Video-based Large Language Models", "abstract": "Video-LLMs are increasingly deployed in safety-critical applications but are vulnerable to Energy-Latency Attacks (ELAs) that exhaust computational resources. Current image-centric methods fail because temporal aggregation mechanisms dilute individual frame perturbations. Additionally, real-time demands make instance-wise optimization impractical for continuous video streams. We introduce VidDoS, which is the first universal ELA framework tailored for Video-LLMs. Our method leverages universal optimization to create instance-agnostic triggers that require no inference-time gradient calculation. We achieve this through $\\textit{masked teacher forcing}$ to steer models toward expensive target sequences, combined with a $\\textit{refusal penalty}$ and $\\textit{early-termination suppression}$ to override conciseness priors. Testing across three mainstream Video-LLMs and three video datasets, which include video question answering and autonomous driving scenarios, shows extreme degradation. VidDoS induces a token expansion of more than 205$\\times$ and inflates the inference latency by more than 15$\\times$ relative to clean baselines. Simulations of real-time autonomous driving streams further reveal that this induced latency leads to critical safety violations. We urge the community to recognize and mitigate these high-hazard ELA in Video-LLMs.", "date": "2026-03-02", "categories": []}
{"id": "2603.01784", "title": "Co-Evolutionary Multi-Modal Alignment via Structured Adversarial Evolution", "abstract": "Adversarial behavior plays a central role in aligning large language models with human values. However, existing alignment methods largely rely on static adversarial settings, which fundamentally limit robustness, particularly in multimodal settings with a larger attack surface. In this work, we move beyond static adversarial supervision and introduce co-evolutionary alignment with evolving attacks, instantiated by CEMMA (Co-Evolutionary Multi-Modal Alignment), an automated and adaptive framework for multimodal safety alignment. We introduce an Evolutionary Attacker that decomposes adversarial prompts into method templates and harmful intents. By employing genetic operators, including mutation, crossover, and differential evolution, it enables simple seed attacks to inherit the structural efficacy of sophisticated jailbreaks. The Adaptive Defender is iteratively updated on the synthesized hard negatives, forming a closed-loop process that adapts alignment to evolving attacks. Experiments show that the Evolutionary Attacker substantially increases red-teaming jailbreak attack success rate (ASR), while the Adaptive Defender improves robustness and generalization across benchmarks with higher data efficiency, without inducing excessive benign refusal, and remains compatible with inference-time defenses such as AdaShield.", "date": "2026-03-02", "categories": []}
{"id": "2603.01938", "title": "Explanation-Guided Adversarial Training for Robust and Interpretable Models", "abstract": "Deep neural networks (DNNs) have achieved remarkable performance in many tasks, yet they often behave as opaque black boxes. Explanation-guided learning (EGL) methods steer DNNs using human-provided explanations or supervision on model attributions. These approaches improve interpretability but typically assume benign inputs and incur heavy annotation costs. In contrast, both predictions and saliency maps of DNNs could dramatically alter facing imperceptible perturbations or unseen patterns. Adversarial training (AT) can substantially improve robustness, but it does not guarantee that model decisions rely on semantically meaningful features. In response, we propose Explanation-Guided Adversarial Training (EGAT), a unified framework that integrates the strength of AT and EGL to simultaneously improve prediction performance, robustness, and explanation quality. EGAT generates adversarial examples on the fly while imposing explanation-based constraints on the model. By jointly optimizing classification performance, adversarial robustness, and attributional stability, EGAT is not only more resistant to unexpected cases, including adversarial attacks and out-of-distribution (OOD) scenarios, but also offer human-interpretable justifications for the decisions. We further formalize EGAT within the Probably Approximately Correct learning framework, demonstrating theoretically that it yields more stable predictions under unexpected situations compared to standard AT. Empirical evaluations on OOD benchmark datasets show that EGAT consistently outperforms competitive baselines in both clean accuracy and adversarial accuracy +37% while producing more semantically meaningful explanations, and requiring only a limited increase +16% in training time.", "date": "2026-03-02", "categories": []}
{"id": "2507.02965", "title": "Concept-based Adversarial Attack: a Probabilistic Perspective", "abstract": "We propose a concept-based adversarial attack framework that extends beyond single-image perturbations by adopting a probabilistic perspective. Rather than modifying a single image, our method operates on an entire concept - represented by a distribution - to generate diverse adversarial examples. Preserving the concept is essential, as it ensures that the resulting adversarial images remain identifiable as instances of the original underlying category or identity. By sampling from this concept-based adversarial distribution, we generate images that maintain the original concept but vary in pose, viewpoint, or background, thereby misleading the classifier. Mathematically, this framework remains consistent with traditional adversarial attacks in a principled manner. Our theoretical and empirical results demonstrate that concept-based adversarial attacks yield more diverse adversarial examples and effectively preserve the underlying concept, while achieving higher attack efficiency.", "date": "2025-06-30", "categories": []}
{"id": "2501.16904", "title": "Diffusion or Non-Diffusion Adversarial Defenses: Rethinking the Relation between Classifier and Adversarial Purifier", "abstract": "Adversarial defense research continues to face challenges in combating against advanced adversarial attacks, yet with diffusion models increasingly favoring their defensive capabilities. Unlike most prior studies that focus on diffusion models for test-time defense, we explore the generalization loss in classifiers caused by diffusion models. We compare diffusion-based and non-diffusion-based adversarial purifiers, demonstrating that non-diffusion models can also achieve well performance under a practical setting of non-adaptive attack. While non-diffusion models show promising adversarial robustness, they particularly excel in defense transferability and color generalization without relying on additional data beyond the training set. Notably, a non-diffusion model trained on CIFAR-10 achieves state-of-the-art performance when tested directly on ImageNet, surpassing existing diffusion-based models trained specifically on ImageNet.", "date": "2025-01-28", "categories": []}
{"id": "2507.17937", "title": "Bob's Confetti: Phonetic Memorization Attacks in Music and Video Generation", "abstract": "Generative AI systems for music and video commonly use text-based filters to prevent regurgitation of copyrighted material. We expose a significant vulnerability in this approach by introducing Adversarial PhoneTic Prompting (APT), a novel attack that bypasses these safeguards by exploiting phonetic memorization--the tendency of models to bind sub-lexical acoustic patterns (phonemes, rhyme, stress, cadence) to memorized copyrighted content. APT replaces iconic lyrics with homophonic but semantically unrelated alternatives (e.g., \"mom's spaghetti\" becomes \"Bob's confetti\"), preserving phonetic structure while evading lexical filters. We evaluate APT on leading lyrics-to-song models (Suno, YuE) across English and Korean songs spanning rap, pop, and K-pop. APT achieves 91% average similarity to copyrighted originals, versus 13.7% for random lyrics and 42.2% for semantic paraphrases. Embedding analysis confirms the mechanism: YuE's text encoder treats APT-modified lyrics as near-identical to originals (cosine similarity 0.90) while Sentence-BERT semantic similarity drops to 0.71, showing the model encodes phonetic structure over meaning. This vulnerability extends cross-modally--Veo 3 reconstructs visual scenes from original music videos when prompted with APT lyrics alone, despite no visual cues in the prompt. We further show that phonetic-semantic defense signatures fail, as APT prompts exhibit higher semantic similarity than benign paraphrases. Our findings reveal that sub-lexical acoustic structure acts as a cross-modal retrieval key, rendering current copyright filters systematically vulnerable. Demo examples are available at https://jrohsc.github.io/music_attack/.", "date": "2025-07-23", "categories": []}
{"id": "2508.20570", "title": "Dyslexify: A Mechanistic Defense Against Typographic Attacks in CLIP", "abstract": "Typographic attacks exploit multi-modal systems by injecting text into images, leading to targeted misclassifications, malicious content generation and even Vision-Language Model jailbreaks. In this work, we analyze how CLIP vision encoders behave under typographic attacks, locating specialized attention heads in the latter half of the model's layers that causally extract and transmit typographic information to the cls token. Building on these insights, we introduce Dyslexify - a method to defend CLIP models against typographic attacks by selectively ablating a typographic circuit, consisting of attention heads. Without requiring finetuning, dyslexify improves performance by up to 22.06% on a typographic variant of ImageNet-100, while reducing standard ImageNet-100 accuracy by less than 1%, and demonstrate its utility in a medical foundation model for skin lesion diagnosis. Notably, our training-free approach remains competitive with current state-of-the-art typographic defenses that rely on finetuning. To this end, we release a family of dyslexic CLIP models which are significantly more robust against typographic attacks. These models serve as suitable drop-in replacements for a broad range of safety-critical applications, where the risks of text-based manipulation outweigh the utility of text recognition.", "date": "2025-08-28", "categories": []}
{"id": "2602.22258", "title": "Poisoned Acoustics", "abstract": "Training-data poisoning attacks can induce targeted, undetectable failure in deep neural networks by corrupting a vanishingly small fraction of training labels. We demonstrate this on acoustic vehicle classification using the MELAUDIS urban intersection dataset (approx. 9,600 audio clips, 6 classes): a compact 2-D convolutional neural network (CNN) trained on log-mel spectrograms achieves 95.7% Attack Success Rate (ASR) -- the fraction of target-class test samples misclassified under the attack -- on a Truck-to-Car label-flipping attack at just p=0.5% corruption (48 records), with zero detectable change in aggregate accuracy (87.6% baseline; 95% CI: 88-100%, n=3 seeds). We prove this stealth is structural: the maximum accuracy drop from a complete targeted attack is bounded above by the minority class fraction (beta). For real-world class imbalances (Truck approx. 3%), this bound falls below training-run noise, making aggregate accuracy monitoring provably insufficient regardless of architecture or attack method. A companion backdoor trigger attack reveals a novel trigger-dominance collapse: when the target class is a dataset minority, the spectrogram patch trigger becomes functionally redundant--clean ASR equals triggered ASR, and the attack degenerates to pure label flipping. We formalize the ML training pipeline as an attack surface and propose a trust-minimized defense combining content-addressed artifact hashing, Merkle-tree dataset commitment, and post-quantum digital signatures (ML-DSA-65/CRYSTALS-Dilithium3, NIST FIPS 204) for cryptographically verifiable data provenance.", "date": "2026-02-25", "categories": []}
{"id": "2602.23117", "title": "Devling into Adversarial Transferability on Image Classification: Review, Benchmark, and Evaluation", "abstract": "Adversarial transferability refers to the capacity of adversarial examples generated on the surrogate model to deceive alternate, unexposed victim models. This property eliminates the need for direct access to the victim model during an attack, thereby raising considerable security concerns in practical applications and attracting substantial research attention recently. In this work, we discern a lack of a standardized framework and criteria for evaluating transfer-based attacks, leading to potentially biased assessments of existing approaches. To rectify this gap, we have conducted an exhaustive review of hundreds of related works, organizing various transfer-based attacks into six distinct categories. Subsequently, we propose a comprehensive framework designed to serve as a benchmark for evaluating these attacks. In addition, we delineate common strategies that enhance adversarial transferability and highlight prevalent issues that could lead to unfair comparisons. Finally, we provide a brief review of transfer-based attacks beyond image classification.", "date": "2026-02-26", "categories": []}
{"id": "2505.19610", "title": "JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models", "abstract": "Vision-Language Models (VLMs) exhibit impressive performance, yet the integration of powerful vision encoders has significantly broadened their attack surface, rendering them increasingly susceptible to jailbreak attacks. However, lacking well-defined attack objectives, existing jailbreak methods often struggle with gradient-based strategies prone to local optima and lacking precise directional guidance, and typically decouple visual and textual modalities, thereby limiting their effectiveness by neglecting crucial cross-modal interactions. Inspired by the Eliciting Latent Knowledge (ELK) framework, we posit that VLMs encode safety-relevant information within their internal fusion-layer representations, revealing an implicit safety decision boundary in the latent space. This motivates exploiting boundary to steer model behavior. Accordingly, we propose JailBound, a novel latent space jailbreak framework comprising two stages: (1) Safety Boundary Probing, which addresses the guidance issue by approximating decision boundary within fusion layer's latent space, thereby identifying optimal perturbation directions towards the target region; and (2) Safety Boundary Crossing, which overcomes the limitations of decoupled approaches by jointly optimizing adversarial perturbations across both image and text inputs. This latter stage employs an innovative mechanism to steer the model's internal state towards policy-violating outputs while maintaining cross-modal semantic consistency. Extensive experiments on six diverse VLMs demonstrate JailBound's efficacy, achieves 94.32% white-box and 67.28% black-box attack success averagely, which are 6.17% and 21.13% higher than SOTA methods, respectively. Our findings expose a overlooked safety risk in VLMs and highlight the urgent need for more robust defenses. Warning: This paper contains potentially sensitive, harmful and offensive content.", "date": "2025-05-26", "categories": []}
{"id": "2602.11020", "title": "When Fusion Helps and When It Breaks: View-Aligned Robustness in Same-Source Financial Imaging", "abstract": "We study same-source multi-view learning and adversarial robustness for next-day direction prediction using two deterministic, window-aligned image views derived from the same time series: an OHLCV-rendered chart (ohlcv) and a technical-indicator matrix (indic). To control label ambiguity from near-zero moves, we use an ex-post minimum-movement threshold min_move (tau) based on realized absolute next-day return, defining an offline benchmark on the subset where the absolute next-day return is at least tau. Under leakage-resistant time-block splits with embargo, we compare early fusion (channel stacking) and dual-encoder late fusion with optional cross-branch consistency. We then evaluate pixel-space L-infinity evasion attacks (FGSM/PGD) under view-constrained and joint threat models. We find that fusion is regime dependent: early fusion can suffer negative transfer under noisier settings, whereas late fusion is a more reliable default once labels stabilize. Robustness degrades sharply under tiny budgets with stable view-dependent vulnerabilities; late fusion often helps under view-constrained attacks, but joint perturbations remain challenging.", "date": "2026-02-11", "categories": []}
{"id": "2602.21447", "title": "Adversarial Intent is a Latent Variable: Stateful Trust Inference for Securing Multimodal Agentic RAG", "abstract": "Current stateless defences for multimodal agentic RAG fail to detect adversarial strategies that distribute malicious semantics across retrieval, planning, and generation components. We formulate this security challenge as a Partially Observable Markov Decision Process (POMDP), where adversarial intent is a latent variable inferred from noisy multi-stage observations. We introduce MMA-RAG^T, an inference-time control framework governed by a Modular Trust Agent (MTA) that maintains an approximate belief state via structured LLM reasoning. Operating as a model-agnostic overlay, MMA-RAGT mediates a configurable set of internal checkpoints to enforce stateful defence-in-depth. Extensive evaluation on 43,774 instances demonstrates a 6.50x average reduction factor in Attack Success Rate relative to undefended baselines, with negligible utility cost. Crucially, a factorial ablation validates our theoretical bounds: while statefulness and spatial coverage are individually necessary (26.4 pp and 13.6 pp gains respectively), stateless multi-point intervention can yield zero marginal benefit under homogeneous stateless filtering when checkpoint detections are perfectly correlated.", "date": "2026-02-24", "categories": []}
{"id": "2602.21841", "title": "Resilient Federated Chain: Transforming Blockchain Consensus into an Active Defense Layer for Federated Learning", "abstract": "Federated Learning (FL) has emerged as a key paradigm for building Trustworthy AI systems by enabling privacy-preserving, decentralized model training. However, FL is highly susceptible to adversarial attacks that compromise model integrity and data confidentiality, a vulnerability exacerbated by the fact that conventional data inspection methods are incompatible with its decentralized design. While integrating FL with Blockchain technology has been proposed to address some limitations, its potential for mitigating adversarial attacks remains largely unexplored. This paper introduces Resilient Federated Chain (RFC), a novel blockchain-enabled FL framework designed specifically to enhance resilience against such threats. RFC builds upon the existing Proof of Federated Learning architecture by repurposing the redundancy of its Pooled Mining mechanism as an active defense layer that can be combined with robust aggregation rules. Furthermore, the framework introduces a flexible evaluation function in its consensus mechanism, allowing for adaptive defense against different attack strategies. Extensive experimental evaluation on image classification tasks under various adversarial scenarios, demonstrates that RFC significantly improves robustness compared to baseline methods, providing a viable solution for securing decentralized learning environments.", "date": "2026-02-25", "categories": []}
{"id": "2602.22197", "title": "Off-The-Shelf Image-to-Image Models Are All You Need To Defeat Image Protection Schemes", "abstract": "Advances in Generative AI (GenAI) have led to the development of various protection strategies to prevent the unauthorized use of images. These methods rely on adding imperceptible protective perturbations to images to thwart misuse such as style mimicry or deepfake manipulations. Although previous attacks on these protections required specialized, purpose-built methods, we demonstrate that this is no longer necessary. We show that off-the-shelf image-to-image GenAI models can be repurposed as generic ``denoisers\" using a simple text prompt, effectively removing a wide range of protective perturbations. Across 8 case studies spanning 6 diverse protection schemes, our general-purpose attack not only circumvents these defenses but also outperforms existing specialized attacks while preserving the image's utility for the adversary. Our findings reveal a critical and widespread vulnerability in the current landscape of image protection, indicating that many schemes provide a false sense of security. We stress the urgent need to develop robust defenses and establish that any future protection mechanism must be benchmarked against attacks from off-the-shelf GenAI models. Code is available in this repository: https://github.com/mlsecviswanath/img2imgdenoiser", "date": "2026-02-25", "categories": []}
{"id": "2602.20193", "title": "When Backdoors Go Beyond Triggers: Semantic Drift in Diffusion Models Under Encoder Attacks", "abstract": "Standard evaluations of backdoor attacks on text-to-image (T2I) models primarily measure trigger activation and visual fidelity. We challenge this paradigm, demonstrating that encoder-side poisoning induces persistent, trigger-free semantic corruption that fundamentally reshapes the representation manifold. We trace this vulnerability to a geometric mechanism: a Jacobian-based analysis reveals that backdoors act as low-rank, target-centered deformations that amplify local sensitivity, causing distortion to propagate coherently across semantic neighborhoods. To rigorously quantify this structural degradation, we introduce SEMAD (Semantic Alignment and Drift), a diagnostic framework that measures both internal embedding drift and downstream functional misalignment. Our findings, validated across diffusion and contrastive paradigms, expose the deep structural risks of encoder poisoning and highlight the necessity of geometric audits beyond simple attack success rates.", "date": "2026-02-21", "categories": []}
{"id": "2505.16670", "title": "BitHydra: Towards Bit-flip Inference Cost Attack against Large Language Models", "abstract": "Large language models (LLMs) are widely deployed, but their substantial compute demands make them vulnerable to inference cost attacks that aim to deliberately maximize the output length. In this work, we investigate a distinct attack surface: maximizing inference cost by tampering with the model parameters instead of inputs. This approach leverages the established capability of Bit-Flip Attacks (BFAs) to persistently alter model behavior via minute weight perturbations, effectively decoupling the attack from specific input queries. To realize this, we propose BitHydra, a framework that addresses the unique optimization challenge of identifying the exact weight bits that maximize generation cost. We formulate the attack as a constrained Binary Integer Programming (BIP) problem designed to systematically suppress the end-of-sequence (i.e., <eos>) probability. To overcome the intractability of the discrete search space, we relax the problem into a continuous optimization task and solve it via the Alternating Direction Method of Multipliers (ADMM). We evaluate BitHydra across 10 LLMs (1.5B-16B). Our results demonstrate that the proposed optimization method efficiently achieves endless generation with as few as 1-4 bit flips on all testing models, verifying the effectiveness of the ADMM-based formulation against both standard models and potential defenses.", "date": "2025-05-22", "categories": []}
{"id": "2507.04446", "title": "Sampling-aware Adversarial Attacks Against Large Language Models", "abstract": "To guarantee safe and robust deployment of large language models (LLMs) at scale, it is critical to accurately assess their adversarial robustness. Existing adversarial attacks typically target harmful responses in single-point greedy generations, overlooking the inherently stochastic nature of LLMs and overestimating robustness. We show that for the goal of eliciting harmful responses, repeated sampling of model outputs during the attack complements prompt optimization and serves as a strong and efficient attack vector. By casting attacks as a resource allocation problem between optimization and sampling, we empirically determine compute-optimal trade-offs and show that integrating sampling into existing attacks boosts success rates by up to 37\\% and improves efficiency by up to two orders of magnitude. We further analyze how distributions of output harmfulness evolve during an adversarial attack, discovering that many common optimization strategies have little effect on output harmfulness. Finally, we introduce a label-free proof-of-concept objective based on entropy maximization, demonstrating how our sampling-aware perspective enables new optimization targets. Overall, our findings establish the importance of sampling in attacks to accurately assess and strengthen LLM safety at scale.", "date": "2025-07-06", "categories": []}
{"id": "2602.18739", "title": "When World Models Dream Wrong: Physical-Conditioned Adversarial Attacks against World Models", "abstract": "Generative world models (WMs) are increasingly used to synthesize controllable, sensor-conditioned driving videos, yet their reliance on physical priors exposes novel attack surfaces. In this paper, we present Physical-Conditioned World Model Attack (PhysCond-WMA), the first white-box world model attack that perturbs physical-condition channels, such as HDMap embeddings and 3D-box features, to induce semantic, logic, or decision-level distortion while preserving perceptual fidelity. PhysCond-WMA is optimized in two stages: (1) a quality-preserving guidance stage that constrains reverse-diffusion loss below a calibrated threshold, and (2) a momentum-guided denoising stage that accumulates target-aligned gradients along the denoising trajectory for stable, temporally coherent semantic shifts. Extensive experimental results demonstrate that our approach remains effective while increasing FID by about 9% on average and FVD by about 3.9% on average. Under the targeted attack setting, the attack success rate (ASR) reaches 0.55. Downstream studies further show tangible risk, which using attacked videos for training decreases 3D detection performance by about 4%, and worsens open-loop planning performance by about 20%. These findings has for the first time revealed and quantified security vulnerabilities in generative world models, driving more comprehensive security checkers.", "date": "2026-02-21", "categories": []}
{"id": "2602.18782", "title": "MANATEE: Inference-Time Lightweight Diffusion Based Safety Defense for LLMs", "abstract": "Defending LLMs against adversarial jailbreak attacks remains an open challenge. Existing defenses rely on binary classifiers that fail when adversarial input falls outside the learned decision boundary, and repeated fine-tuning is computationally expensive while potentially degrading model capabilities. We propose MANATEE, an inference-time defense that uses density estimation over a benign representation manifold. MANATEE learns the score function of benign hidden states and uses diffusion to project anomalous representations toward safe regions--requiring no harmful training data and no architectural modifications. Experiments across Mistral-7B-Instruct, Llama-3.1-8B-Instruct, and Gemma-2-9B-it demonstrate that MANATEE reduce Attack Success Rate by up to 100\\% on certain datasets, while preserving model utility on benign inputs.", "date": "2026-02-21", "categories": []}
{"id": "2602.19418", "title": "PA-Attack: Guiding Gray-Box Attacks on LVLM Vision Encoders with Prototypes and Attention", "abstract": "Large Vision-Language Models (LVLMs) are foundational to modern multimodal applications, yet their susceptibility to adversarial attacks remains a critical concern. Prior white-box attacks rarely generalize across tasks, and black-box methods depend on expensive transfer, which limits efficiency. The vision encoder, standardized and often shared across LVLMs, provides a stable gray-box pivot with strong cross-model transfer. Building on this premise, we introduce PA-Attack (Prototype-Anchored Attentive Attack). PA-Attack begins with a prototype-anchored guidance that provides a stable attack direction towards a general and dissimilar prototype, tackling the attribute-restricted issue and limited task generalization of vanilla attacks. Building on this, we propose a two-stage attention enhancement mechanism: (i) leverage token-level attention scores to concentrate perturbations on critical visual tokens, and (ii) adaptively recalibrate attention weights to track the evolving attention during the adversarial process. Extensive experiments across diverse downstream tasks and LVLM architectures show that PA-Attack achieves an average 75.1% score reduction rate (SRR), demonstrating strong attack effectiveness, efficiency, and task generalization in LVLMs. Code is available at https://github.com/hefeimei06/PA-Attack.", "date": "2026-02-23", "categories": []}
{"id": "2602.19539", "title": "Can a Teenager Fool an AI? Evaluating Low-Cost Cosmetic Attacks on Age Estimation Systems", "abstract": "Age estimation systems are increasingly deployed as gatekeepers for age-restricted online content, yet their robustness to cosmetic modifications has not been systematically evaluated. We investigate whether simple, household-accessible cosmetic changes, including beards, grey hair, makeup, and simulated wrinkles, can cause AI age estimators to classify minors as adults. To study this threat at scale without ethical concerns, we simulate these physical attacks on 329 facial images of individuals aged 10 to 21 using a VLM image editor (Gemini 2.5 Flash Image). We then evaluate eight models from our prior benchmark: five specialized architectures (MiVOLO, Custom-Best, Herosan, MiViaLab, DEX) and three vision-language models (Gemini 3 Flash, Gemini 2.5 Flash, GPT-5-Nano). We introduce the Attack Conversion Rate (ACR), defined as the fraction of images predicted as minor at baseline that flip to adult after attack, a population-agnostic metric that does not depend on the ratio of minors to adults in the test set. Our results reveal that a synthetic beard alone achieves 28 to 69 percent ACR across all eight models; combining all four attacks shifts predicted age by +7.7 years on average across all 329 subjects and reaches up to 83 percent ACR; and vision-language models exhibit lower ACR (59 to 71 percent) than specialized models (63 to 83 percent) under the full attack, although the ACR ranges overlap and the difference is not statistically tested. These findings highlight a critical vulnerability in deployed age-verification pipelines and call for adversarial robustness evaluation as a mandatory criterion for model selection.", "date": "2026-02-23", "categories": []}
{"id": "2602.19596", "title": "Learning Mutual View Information Graph for Adaptive Adversarial Collaborative Perception", "abstract": "Collaborative perception (CP) enables data sharing among connected and autonomous vehicles (CAVs) to enhance driving safety. However, CP systems are vulnerable to adversarial attacks where malicious agents forge false objects via feature-level perturbations. Current defensive systems use threshold-based consensus verification by comparing collaborative and ego detection results. Yet, these defenses remain vulnerable to more sophisticated attack strategies that could exploit two critical weaknesses: (i) lack of robustness against attacks with systematic timing and target region optimization, and (ii) inadvertent disclosure of vulnerability knowledge through implicit confidence information in shared collaboration data. In this paper, we propose MVIG attack, a novel adaptive adversarial CP framework learning to capture vulnerability knowledge disclosed by different defensive CP systems from a unified mutual view information graph (MVIG) representation. Our approach combines MVIG representation with temporal graph learning to generate evolving fabrication risk maps and employs entropy-aware vulnerability search to optimize attack location, timing and persistence, enabling adaptive attacks with generalizability across various defensive configurations. Extensive evaluations on OPV2V and Adv-OPV2V datasets demonstrate that MVIG attack reduces defense success rates by up to 62\\% against state-of-the-art defenses while achieving 47\\% lower detection for persistent attacks at 29.9 FPS, exposing critical security gaps in CP systems. Code will be released at https://github.com/yihangtao/MVIG.git", "date": "2026-02-23", "categories": []}
{"id": "2602.19931", "title": "Expanding the Role of Diffusion Models for Robust Classifier Training", "abstract": "Incorporating diffusion-generated synthetic data into adversarial training (AT) has been shown to substantially improve the training of robust image classifiers. In this work, we extend the role of diffusion models beyond merely generating synthetic data, examining whether their internal representations, which encode meaningful features of the data, can provide additional benefits for robust classifier training. Through systematic experiments, we show that diffusion models offer representations that are both diverse and partially robust, and that explicitly incorporating diffusion representations as an auxiliary learning signal during AT consistently improves robustness across settings. Furthermore, our representation analysis indicates that incorporating diffusion models into AT encourages more disentangled features, while diffusion representations and diffusion-generated synthetic data play complementary roles in shaping representations. Experiments on CIFAR-10, CIFAR-100, and ImageNet validate these findings, demonstrating the effectiveness of jointly leveraging diffusion representations and synthetic data within AT.", "date": "2026-02-23", "categories": []}
{"id": "2602.17948", "title": "A Geometric Probe of the Accuracy-Robustness Trade-off: Sharp Boundaries in Symmetry-Breaking Dimensional Expansion", "abstract": "The trade-off between clean accuracy and adversarial robustness is a pervasive phenomenon in deep learning, yet its geometric origin remains elusive. In this work, we utilize Symmetry-Breaking Dimensional Expansion (SBDE) as a controlled probe to investigate the mechanism underlying this trade-off. SBDE expands input images by inserting constant-valued pixels, which breaks translational symmetry and consistently improves clean accuracy (e.g., from $90.47\\%$ to $95.63\\%$ on CIFAR-10 with ResNet-18) by reducing parameter degeneracy. However, this accuracy gain comes at the cost of reduced robustness against iterative white-box attacks. By employing a test-time \\emph{mask projection} that resets the inserted auxiliary pixels to their training values, we demonstrate that the vulnerability stems almost entirely from the inserted dimensions. The projection effectively neutralizes the attacks and restores robustness, revealing that the model achieves high accuracy by creating \\emph{sharp boundaries} (steep loss gradients) specifically along the auxiliary axes. Our findings provide a concrete geometric explanation for the accuracy-robustness paradox: the optimization landscape deepens the basin of attraction to improve accuracy but inevitably erects steep walls along the auxiliary degrees of freedom, creating a fragile sensitivity to off-manifold perturbations.", "date": "2026-02-20", "categories": []}
{"id": "2602.17975", "title": "Generating adversarial inputs for a graph neural network model of AC power flow", "abstract": "This work formulates and solves optimization problems to generate input points that yield high errors between a neural network's predicted AC power flow solution and solutions to the AC power flow equations. We demonstrate this capability on an instance of the CANOS-PF graph neural network model, as implemented by the PF$\\Delta$ benchmark library, operating on a 14-bus test grid. Generated adversarial points yield errors as large as 3.4 per-unit in reactive power and 0.08 per-unit in voltage magnitude. When minimizing the perturbation from a training point necessary to satisfy adversarial constraints, we find that the constraints can be met with as little as an 0.04 per-unit perturbation in voltage magnitude on a single bus. This work motivates the development of rigorous verification and robust training methods for neural network surrogate models of AC power flow.", "date": "2026-02-20", "categories": []}
{"id": "2602.18252", "title": "On the Adversarial Robustness of Discrete Image Tokenizers", "abstract": "Discrete image tokenizers encode visual inputs as sequences of tokens from a finite vocabulary and are gaining popularity in multimodal systems, including encoder-only, encoder-decoder, and decoder-only models. However, unlike CLIP encoders, their vulnerability to adversarial attacks has not been explored. Ours being the first work studying this topic, we first formulate attacks that aim to perturb the features extracted by discrete tokenizers, and thus change the extracted tokens. These attacks are computationally efficient, application-agnostic, and effective across classification, multimodal retrieval, and captioning tasks. Second, to defend against this vulnerability, inspired by recent work on robust CLIP encoders, we fine-tune popular tokenizers with unsupervised adversarial training, keeping all other components frozen. While unsupervised and task-agnostic, our approach significantly improves robustness to both unsupervised and end-to-end supervised attacks and generalizes well to unseen tasks and data. Unlike supervised adversarial training, our approach can leverage unlabeled images, making it more versatile. Overall, our work highlights the critical role of tokenizer robustness in downstream tasks and presents an important step in the development of safe multimodal foundation models.", "date": "2026-02-20", "categories": []}
{"id": "2504.21730", "title": "Cert-SSBD: Certified Backdoor Defense with Sample-Specific Smoothing Noises", "abstract": "Deep neural networks (DNNs) are vulnerable to backdoor attacks, where an attacker manipulates a small portion of the training data to implant hidden backdoors into the model. The compromised model behaves normally on clean samples but misclassifies backdoored samples into the attacker-specified target class, posing a significant threat to real-world DNN applications. Currently, several empirical defense methods have been proposed to mitigate backdoor attacks, but they are often bypassed by more advanced backdoor techniques. In contrast, certified defenses based on randomized smoothing have shown promise by adding random noise to training and testing samples to counteract backdoor attacks. In this paper, we reveal that existing randomized smoothing defenses implicitly assume that all samples are equidistant from the decision boundary. However, it may not hold in practice, leading to suboptimal certification performance. To address this issue, we propose a sample-specific certified backdoor defense method, termed Cert-SSB. Cert-SSB first employs stochastic gradient ascent to optimize the noise magnitude for each sample, ensuring a sample-specific noise level that is then applied to multiple poisoned training sets to retrain several smoothed models. After that, Cert-SSB aggregates the predictions of multiple smoothed models to generate the final robust prediction. In particular, in this case, existing certification methods become inapplicable since the optimized noise varies across different samples. To conquer this challenge, we introduce a storage-update-based certification method, which dynamically adjusts each sample's certification region to improve certification performance. We conduct extensive experiments on multiple benchmark datasets, demonstrating the effectiveness of our proposed method. Our code is available at https://github.com/NcepuQiaoTing/Cert-SSB.", "date": "2025-04-30", "categories": []}
{"id": "2602.06530", "title": "Universal Anti-forensics Attack against Image Forgery Detection via Multi-modal Guidance", "abstract": "The rapid advancement of AI-Generated Content (AIGC) technologies poses significant challenges for authenticity assessment. However, existing evaluation protocols largely overlook anti-forensics attack, failing to ensure the comprehensive robustness of state-of-the-art AIGC detectors in real-world applications. To bridge this gap, we propose ForgeryEraser, a framework designed to execute universal anti-forensics attack without access to the target AIGC detectors. We reveal an adversarial vulnerability stemming from the systemic reliance on Vision-Language Models (VLMs) as shared backbones (e.g., CLIP), where downstream AIGC detectors inherit the feature space of these publicly accessible models. Instead of traditional logit-based optimization, we design a multi-modal guidance loss to drive forged image embeddings within the VLM feature space toward text-derived authentic anchors to erase forgery traces, while repelling them from forgery anchors. Extensive experiments demonstrate that ForgeryEraser causes substantial performance degradation to advanced AIGC detectors on both global synthesis and local editing benchmarks. Moreover, ForgeryEraser induces explainable forensic models to generate explanations consistent with authentic images for forged images. Our code will be made publicly available.", "date": "2026-02-06", "categories": []}
{"id": "2602.16723", "title": "Is Mamba Reliable for Medical Imaging?", "abstract": "State-space models like Mamba offer linear-time sequence processing and low memory, making them attractive for medical imaging. However, their robustness under realistic software and hardware threat models remains underexplored. This paper evaluates Mamba on multiple MedM-NIST classification benchmarks under input-level attacks, including white-box adversarial perturbations (FGSM/PGD), occlusion-based PatchDrop, and common acquisition corruptions (Gaussian noise and defocus blur) as well as hardware-inspired fault attacks emulated in software via targeted and random bit-flip injections into weights and activations. We profile vulnerabilities and quantify impacts on accuracy indicating that defenses are needed for deployment.", "date": "2026-02-16", "categories": []}
{"id": "2602.16741", "title": "Can Adversarial Code Comments Fool AI Security Reviewers -- Large-Scale Empirical Study of Comment-Based Attacks and Defenses Against LLM Code Analysis", "abstract": "AI-assisted code review is widely used to detect vulnerabilities before production release. Prior work shows that adversarial prompt manipulation can degrade large language model (LLM) performance in code generation. We test whether similar comment-based manipulation misleads LLMs during vulnerability detection. We build a 100-sample benchmark across Python, JavaScript, and Java, each paired with eight comment variants ranging from no comments to adversarial strategies such as authority spoofing and technical deception. Eight frontier models, five commercial and three open-source, are evaluated in 9,366 trials. Adversarial comments produce small, statistically non-significant effects on detection accuracy (McNemar exact p > 0.21; all 95 percent confidence intervals include zero). This holds for commercial models with 89 to 96 percent baseline detection and open-source models with 53 to 72 percent, despite large absolute performance gaps. Unlike generation settings where comment manipulation achieves high attack success, detection performance does not meaningfully degrade. More complex adversarial strategies offer no advantage over simple manipulative comments. We test four automated defenses across 4,646 additional trials (14,012 total). Static analysis cross-referencing performs best at 96.9 percent detection and recovers 47 percent of baseline misses. Comment stripping reduces detection for weaker models by removing helpful context. Failures concentrate on inherently difficult vulnerability classes, including race conditions, timing side channels, and complex authorization logic, rather than on adversarial comments.", "date": "2026-02-18", "categories": []}
{"id": "2602.16935", "title": "DeepContext: Stateful Real-Time Detection of Multi-Turn Adversarial Intent Drift in LLMs", "abstract": "While Large Language Model (LLM) capabilities have scaled, safety guardrails remain largely stateless, treating multi-turn dialogues as a series of disconnected events. This lack of temporal awareness facilitates a \"Safety Gap\" where adversarial tactics, like Crescendo and ActorAttack, slowly bleed malicious intent across turn boundaries to bypass stateless filters. We introduce DeepContext, a stateful monitoring framework designed to map the temporal trajectory of user intent. DeepContext discards the isolated evaluation model in favor of a Recurrent Neural Network (RNN) architecture that ingests a sequence of fine-tuned turn-level embeddings. By propagating a hidden state across the conversation, DeepContext captures the incremental accumulation of risk that stateless models overlook. Our evaluation demonstrates that DeepContext significantly outperforms existing baselines in multi-turn jailbreak detection, achieving a state-of-the-art F1 score of 0.84, which represents a substantial improvement over both hyperscaler cloud-provider guardrails and leading open-weight models such as Llama-Prompt-Guard-2 (0.67) and Granite-Guardian (0.67). Furthermore, DeepContext maintains a sub-20ms inference overhead on a T4 GPU, ensuring viability for real-time applications. These results suggest that modeling the sequential evolution of intent is a more effective and computationally efficient alternative to deploying massive, stateless models.", "date": "2026-02-18", "categories": []}
{"id": "2602.16944", "title": "Exact Certification of Data-Poisoning Attacks Using Mixed-Integer Programming", "abstract": "This work introduces a verification framework that provides both sound and complete guarantees for data poisoning attacks during neural network training. We formulate adversarial data manipulation, model training, and test-time evaluation in a single mixed-integer quadratic programming (MIQCP) problem. Finding the global optimum of the proposed formulation provably yields worst-case poisoning attacks, while simultaneously bounding the effectiveness of all possible attacks on the given training pipeline. Our framework encodes both the gradient-based training dynamics and model evaluation at test time, enabling the first exact certification of training-time robustness. Experimental evaluation on small models confirms that our approach delivers a complete characterization of robustness against data poisoning.", "date": "2026-02-18", "categories": []}
{"id": "2602.17645", "title": "Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting", "abstract": "Black-box adversarial attacks on Large Vision-Language Models (LVLMs) are challenging due to missing gradients and complex multimodal boundaries. While prior state-of-the-art transfer-based approaches like M-Attack perform well using local crop-level matching between source and target images, we find this induces high-variance, nearly orthogonal gradients across iterations, violating coherent local alignment and destabilizing optimization. We attribute this to (i) ViT translation sensitivity that yields spike-like gradients and (ii) structural asymmetry between source and target crops. We reformulate local matching as an asymmetric expectation over source transformations and target semantics, and build a gradient-denoising upgrade to M-Attack. On the source side, Multi-Crop Alignment (MCA) averages gradients from multiple independently sampled local views per iteration to reduce variance. On the target side, Auxiliary Target Alignment (ATA) replaces aggressive target augmentation with a small auxiliary set from a semantically correlated distribution, producing a smoother, lower-variance target manifold. We further reinterpret momentum as Patch Momentum, replaying historical crop gradients; combined with a refined patch-size ensemble (PE+), this strengthens transferable directions. Together these modules form M-Attack-V2, a simple, modular enhancement over M-Attack that substantially improves transfer-based black-box attacks on frontier LVLMs: boosting success rates on Claude-4.0 from 8% to 30%, Gemini-2.5-Pro from 83% to 97%, and GPT-5 from 98% to 100%, outperforming prior black-box LVLM attacks. Code and data are publicly available at: https://github.com/vila-lab/M-Attack-V2.", "date": "2026-02-19", "categories": []}
{"id": "2511.06197", "title": "Enhancing Adversarial Robustness of IoT Intrusion Detection via SHAP-Based Attribution Fingerprinting", "abstract": "The rapid proliferation of Internet of Things (IoT) devices has transformed numerous industries by enabling seamless connectivity and data-driven automation. However, this expansion has also exposed IoT networks to increasingly sophisticated security threats, including adversarial attacks targeting artificial intelligence (AI) and machine learning (ML)-based intrusion detection systems (IDS) to deliberately evade detection, induce misclassification, and systematically undermine the reliability and integrity of security defenses. To address these challenges, we propose a novel adversarial detection model that enhances the robustness of IoT IDS against adversarial attacks through SHapley Additive exPlanations (SHAP)-based fingerprinting. Using SHAP's DeepExplainer, we extract attribution fingerprints from network traffic features, enabling the IDS to reliably distinguish between clean and adversarially perturbed inputs. By capturing subtle attribution patterns, the model becomes more resilient to evasion attempts and adversarial manipulations. We evaluated the model on a standard IoT benchmark dataset, where it significantly outperformed a state-of-the-art method in detecting adversarial attacks. In addition to enhanced robustness, this approach improves model transparency and interpretability, thereby increasing trust in the IDS through explainable AI.", "date": "2025-11-09", "categories": []}
{"id": "2602.15001", "title": "Boundary Point Jailbreaking of Black-Box LLMs", "abstract": "Frontier LLMs are safeguarded against attempts to extract harmful information via adversarial prompts known as \"jailbreaks\". Recently, defenders have developed classifier-based systems that have survived thousands of hours of human red teaming. We introduce Boundary Point Jailbreaking (BPJ), a new class of automated jailbreak attacks that evade the strongest industry-deployed safeguards. Unlike previous attacks that rely on white/grey-box assumptions (such as classifier scores or gradients) or libraries of existing jailbreaks, BPJ is fully black-box and uses only a single bit of information per query: whether or not the classifier flags the interaction. To achieve this, BPJ addresses the core difficulty in optimising attacks against robust real-world defences: evaluating whether a proposed modification to an attack is an improvement. Instead of directly trying to learn an attack for a target harmful string, BPJ converts the string into a curriculum of intermediate attack targets and then actively selects evaluation points that best detect small changes in attack strength (\"boundary points\"). We believe BPJ is the first fully automated attack algorithm that succeeds in developing universal jailbreaks against Constitutional Classifiers, as well as the first automated attack algorithm that succeeds against GPT-5's input classifier without relying on human attack seeds. BPJ is difficult to defend against in individual interactions but incurs many flags during optimisation, suggesting that effective defence requires supplementing single-interaction methods with batch-level monitoring.", "date": "2026-02-16", "categories": []}
{"id": "2602.15238", "title": "Closing the Distribution Gap in Adversarial Training for LLMs", "abstract": "Adversarial training for LLMs is one of the most promising methods to reliably improve robustness against adversaries. However, despite significant progress, models remain vulnerable to simple in-distribution exploits, such as rewriting prompts in the past tense or translating them into other languages. We argue that this persistent fragility stems from a fundamental limitation in current adversarial training algorithms: they minimize adversarial loss on their training set but inadequately cover the data distribution, resulting in vulnerability to seemingly simple attacks. To bridge this gap, we propose Distributional Adversarial Training, DAT. We leverage Diffusion LLMs to approximate the true joint distribution of prompts and responses, enabling generation of diverse, high-likelihood samples that address generalization failures. By combining optimization over the data distribution provided by the diffusion model with continuous adversarial training, DAT achieves substantially higher adversarial robustness than previous methods.", "date": "2026-02-16", "categories": []}
{"id": "2602.16494", "title": "Benchmarking Adversarial Robustness and Adversarial Training Strategies for Object Detection", "abstract": "Object detection models are critical components of automated systems, such as autonomous vehicles and perception-based robots, but their sensitivity to adversarial attacks poses a serious security risk. Progress in defending these models lags behind classification, hindered by a lack of standardized evaluation. It is nearly impossible to thoroughly compare attack or defense methods, as existing work uses different datasets, inconsistent efficiency metrics, and varied measures of perturbation cost. This paper addresses this gap by investigating three key questions: (1) How can we create a fair benchmark to impartially compare attacks? (2) How well do modern attacks transfer across different architectures, especially from Convolutional Neural Networks to Vision Transformers? (3) What is the most effective adversarial training strategy for robust defense? To answer these, we first propose a unified benchmark framework focused on digital, non-patch-based attacks. This framework introduces specific metrics to disentangle localization and classification errors and evaluates attack cost using multiple perceptual metrics. Using this benchmark, we conduct extensive experiments on state-of-the-art attacks and a wide range of detectors. Our findings reveal two major conclusions: first, modern adversarial attacks against object detection models show a significant lack of transferability to transformer-based architectures. Second, we demonstrate that the most robust adversarial training strategy leverages a dataset composed of a mix of high-perturbation attacks with different objectives (e.g., spatial and semantic), which outperforms training on any single attack.", "date": "2026-02-18", "categories": []}
{"id": "2602.16543", "title": "Vulnerability Analysis of Safe Reinforcement Learning via Inverse Constrained Reinforcement Learning", "abstract": "Safe reinforcement learning (Safe RL) aims to ensure policy performance while satisfying safety constraints. However, most existing Safe RL methods assume benign environments, making them vulnerable to adversarial perturbations commonly encountered in real-world settings. In addition, existing gradient-based adversarial attacks typically require access to the policy's gradient information, which is often impractical in real-world scenarios. To address these challenges, we propose an adversarial attack framework to reveal vulnerabilities of Safe RL policies. Using expert demonstrations and black-box environment interaction, our framework learns a constraint model and a surrogate (learner) policy, enabling gradient-based attack optimization without requiring the victim policy's internal gradients or the ground-truth safety constraints. We further provide theoretical analysis establishing feasibility and deriving perturbation bounds. Experiments on multiple Safe RL benchmarks demonstrate the effectiveness of our approach under limited privileged access.", "date": "2026-02-18", "categories": []}
{"id": "2406.03862", "title": "Robust Deep Reinforcement Learning against Adversarial Behavior Manipulation", "abstract": "This study investigates behavior-targeted attacks on reinforcement learning and their countermeasures. Behavior-targeted attacks aim to manipulate the victim's behavior as desired by the adversary through adversarial interventions in state observations. Existing behavior-targeted attacks have some limitations, such as requiring white-box access to the victim's policy. To address this, we propose a novel attack method using imitation learning from adversarial demonstrations, which works under limited access to the victim's policy and is environment-agnostic. In addition, our theoretical analysis proves that the policy's sensitivity to state changes impacts defense performance, particularly in the early stages of the trajectory. Based on this insight, we propose time-discounted regularization, which enhances robustness against attacks while maintaining task performance. To the best of our knowledge, this is the first defense strategy specifically designed for behavior-targeted attacks.", "date": "2024-06-06", "categories": []}
{"id": "2412.20987", "title": "RobustBlack: Challenging Black-Box Adversarial Attacks on State-of-the-Art Defenses", "abstract": "Although adversarial robustness has been extensively studied in white-box settings, recent advances in black-box attacks (including transfer- and query-based approaches) are primarily benchmarked against weak defenses, leaving a significant gap in the evaluation of their effectiveness against more recent and moderate robust models (e.g., those featured in the Robustbench leaderboard). In this paper, we question this lack of attention from black-box attacks to robust models. We establish a framework to evaluate the effectiveness of recent black-box attacks against both top-performing and standard defense mechanisms, on the ImageNet dataset. Our empirical evaluation reveals the following key findings: (1) the most advanced black-box attacks struggle to succeed even against simple adversarially trained models; (2) robust models that are optimized to withstand strong white-box attacks, such as AutoAttack, also exhibits enhanced resilience against black-box attacks; and (3) robustness alignment between the surrogate models and the target model plays a key factor in the success rate of transfer-based attacks", "date": "2024-12-30", "categories": []}
{"id": "2601.10313", "title": "Hierarchical Refinement of Universal Multimodal Attacks on Vision-Language Models", "abstract": "Existing adversarial attacks for VLP models are mostly sample-specific, resulting in substantial computational overhead when scaled to large datasets or new scenarios. To overcome this limitation, we propose Hierarchical Refinement Attack (HRA), a multimodal universal attack framework for VLP models. For the image modality, we refine the optimization path by leveraging a temporal hierarchy of historical and estimated future gradients to avoid local minima and stabilize universal perturbation learning. For the text modality, it hierarchically models textual importance by considering both intra- and inter-sentence contributions to identify globally influential words, which are then used as universal text perturbations. Extensive experiments across various downstream tasks, VLP models, and datasets, demonstrate the superior transferability of the proposed universal multimodal attacks.", "date": "2026-01-15", "categories": []}
{"id": "2602.15344", "title": "ER-MIA: Black-Box Adversarial Memory Injection Attacks on Long-Term Memory-Augmented Large Language Models", "abstract": "Large language models (LLMs) are increasingly augmented with long-term memory systems to overcome finite context windows and enable persistent reasoning across interactions. However, recent research finds that LLMs become more vulnerable because memory provides extra attack surfaces. In this paper, we present the first systematic study of black-box adversarial memory injection attacks that target the similarity-based retrieval mechanism in long-term memory-augmented LLMs. We introduce ER-MIA, a unified framework that exposes this vulnerability and formalizes two realistic attack settings: content-based attacks and question-targeted attacks. In these settings, ER-MIA includes an arsenal of composable attack primitives and ensemble attacks that achieve high success rates under minimal attacker assumptions. Extensive experiments across multiple LLMs and long-term memory systems demonstrate that similarity-based retrieval constitutes a fundamental and system-level vulnerability, revealing security risks that persist across memory designs and application scenarios.", "date": "2026-02-17", "categories": []}
{"id": "2503.00187", "title": "Steering Dialogue Dynamics for Robustness against Multi-turn Jailbreaking Attacks", "abstract": "Large language models (LLMs) are shown to be vulnerable to jailbreaking attacks where adversarial prompts are designed to elicit harmful responses. While existing defenses effectively mitigate single-turn attacks by detecting and filtering unsafe inputs, they fail against multi-turn jailbreaks that exploit contextual drift over multiple interactions, gradually leading LLMs away from safe behavior. To address this challenge, we propose a safety steering framework grounded in safe control theory, ensuring invariant safety in multi-turn dialogues. Our approach models the dialogue with LLMs using state-space representations and introduces a novel neural barrier function (NBF) to detect and filter harmful queries emerging from evolving contexts proactively. Our method achieves invariant safety at each turn of dialogue by learning a safety predictor that accounts for adversarial queries, preventing potential context drift toward jailbreaks. Extensive experiments under multiple LLMs show that our NBF-based safety steering outperforms safety alignment, prompt-based steering and lightweight LLM guardrails baselines, offering stronger defenses against multi-turn jailbreaks while maintaining a better trade-off among safety, helpfulness and over-refusal. Check out the website here https://sites.google.com/view/llm-nbf/home.", "date": "2025-02-28", "categories": []}
{"id": "2506.05402", "title": "Lorica: A Synergistic Fine-Tuning Framework for Advancing Personalized Adversarial Robustness", "abstract": "The growing use of large pre-trained models in edge computing has made model inference on mobile clients both feasible and popular. Yet these devices remain vulnerable to adversarial attacks, threatening model robustness and security. Federated adversarial training (FAT) offers a promising solution by enhancing robustness while preserving client privacy. However, FAT often yields a generalized global model that struggles with heterogeneous client data, leading to limited personalization and significant communication overhead. In this paper, we propose \\textit{Lorica}, a personalized synergistic adversarial training framework that delivers customized defense models through a two-phase process. In Phase 1, \\textit{Lorica} applies LoRA-FA for local adversarial fine-tuning, enabling personalized robustness while reducing communication by uploading only LoRA-FA parameters. In Phase 2, a forward-gating selection strategy improves benign accuracy, further refining the personalized model. This yields tailored defense models that effectively balance robustness and accuracy. Extensive experiments on benchmark datasets demonstrate that \\textit{Lorica} can achieve up to 68$\\times$ improvements in communication efficiency compared to state-of-the-art algorithms, while achieving up to 29.9\\% and 52.2\\% enhancements in adversarial robustness and benign accuracy, respectively.", "date": "2025-06-04", "categories": []}
{"id": "2510.04398", "title": "SECA: Semantically Equivalent and Coherent Attacks for Eliciting LLM Hallucinations", "abstract": "Large Language Models (LLMs) are increasingly deployed in high-risk domains. However, state-of-the-art LLMs often exhibit hallucinations, raising serious concerns about their reliability. Prior work has explored adversarial attacks to elicit hallucinations in LLMs, but these methods often rely on unrealistic prompts, either by inserting nonsensical tokens or by altering the original semantic intent. Consequently, such approaches provide limited insight into how hallucinations arise in real-world settings. In contrast, adversarial attacks in computer vision typically involve realistic modifications to input images. However, the problem of identifying realistic adversarial prompts for eliciting LLM hallucinations remains largely underexplored. To address this gap, we propose Semantically Equivalent and Coherent Attacks (SECA), which elicit hallucinations via realistic modifications to the prompt that preserve its meaning while maintaining semantic coherence. Our contributions are threefold: (i) we formulate finding realistic attacks for hallucination elicitation as a constrained optimization problem over the input prompt space under semantic equivalence and coherence constraints; (ii) we introduce a constraint-preserving zeroth-order method to effectively search for adversarial yet feasible prompts; and (iii) we demonstrate through experiments on open-ended multiple-choice question answering tasks that SECA achieves higher attack success rates while incurring almost no semantic equivalence or semantic coherence errors compared to existing methods. SECA highlights the sensitivity of both open-source and commercial gradient-inaccessible LLMs to realistic and plausible prompt variations. Code is available at https://github.com/Buyun-Liang/SECA.", "date": "2025-10-05", "categories": []}
{"id": "2602.13234", "title": "Stay in Character, Stay Safe: Dual-Cycle Adversarial Self-Evolution for Safety Role-Playing Agents", "abstract": "LLM-based role-playing has rapidly improved in fidelity, yet stronger adherence to persona constraints commonly increases vulnerability to jailbreak attacks, especially for risky or negative personas. Most prior work mitigates this issue with training-time solutions (e.g., data curation or alignment-oriented regularization). However, these approaches are costly to maintain as personas and attack strategies evolve, can degrade in-character behavior, and are typically infeasible for frontier closed-weight LLMs. We propose a training-free Dual-Cycle Adversarial Self-Evolution framework with two coupled cycles. A Persona-Targeted Attacker Cycle synthesizes progressively stronger jailbreak prompts, while a Role-Playing Defender Cycle distills observed failures into a hierarchical knowledge base of (i) global safety rules, (ii) persona-grounded constraints, and (iii) safe in-character exemplars. At inference time, the Defender retrieves and composes structured knowledge from this hierarchy to guide generation, producing responses that remain faithful to the target persona while satisfying safety constraints. Extensive experiments across multiple proprietary LLMs show consistent gains over strong baselines on both role fidelity and jailbreak resistance, and robust generalization to unseen personas and attack prompts.", "date": "2026-01-29", "categories": []}
{"id": "2602.14679", "title": "Universal Image Immunization against Diffusion-based Image Editing via Semantic Injection", "abstract": "Recent advances in diffusion models have enabled powerful image editing capabilities guided by natural language prompts, unlocking new creative possibilities. However, they introduce significant ethical and legal risks, such as deepfakes and unauthorized use of copyrighted visual content. To address these risks, image immunization has emerged as a promising defense against AI-driven semantic manipulation. Yet, most existing approaches rely on image-specific adversarial perturbations that require individual optimization for each image, thereby limiting scalability and practicality. In this paper, we propose the first universal image immunization framework that generates a single, broadly applicable adversarial perturbation specifically designed for diffusion-based editing pipelines. Inspired by universal adversarial perturbation (UAP) techniques used in targeted attacks, our method generates a UAP that embeds a semantic target into images to be protected. Simultaneously, it suppresses original content to effectively misdirect the model's attention during editing. As a result, our approach effectively blocks malicious editing attempts by overwriting the original semantic content in the image via the UAP. Moreover, our method operates effectively even in data-free settings without requiring access to training data or domain knowledge, further enhancing its practicality and broad applicability in real-world scenarios. Extensive experiments show that our method, as the first universal immunization approach, significantly outperforms several baselines in the UAP setting. In addition, despite the inherent difficulty of universal perturbations, our method also achieves performance on par with image-specific methods under a more restricted perturbation budget, while also exhibiting strong black-box transferability across different diffusion models.", "date": "2026-02-16", "categories": []}
{"id": "2407.20836", "title": "Vulnerabilities in AI-generated Image Detection: The Challenge of Adversarial Attacks", "abstract": "Recent advancements in image synthesis, particularly with the advent of GAN and Diffusion models, have amplified public concerns regarding the dissemination of disinformation. To address such concerns, numerous AI-generated Image (AIGI) Detectors have been proposed and achieved promising performance in identifying fake images. However, there still lacks a systematic understanding of the adversarial robustness of AIGI detectors. In this paper, we examine the vulnerability of state-of-the-art AIGI detectors against adversarial attack under white-box and black-box settings, which has been rarely investigated so far. To this end, we propose a new method to attack AIGI detectors. First, inspired by the obvious difference between real images and fake images in the frequency domain, we add perturbations under the frequency domain to push the image away from its original frequency distribution. Second, we explore the full posterior distribution of the surrogate model to further narrow this gap between heterogeneous AIGI detectors, e.g., transferring adversarial examples across CNNs and ViTs. This is achieved by introducing a novel post-train Bayesian strategy that turns a single surrogate into a Bayesian one, capable of simulating diverse victim models using one pre-trained surrogate, without the need for re-training. We name our method as Frequency-based Post-train Bayesian Attack, or FPBA. Through FPBA, we demonstrate that adversarial attacks pose a real threat to AIGI detectors. FPBA can deliver successful black-box attacks across various detectors, generators, defense methods, and even evade cross-generator and compressed image detection, which are crucial real-world detection scenarios. Our code is available at https://github.com/onotoa/fpba.", "date": "2024-07-30", "categories": []}
{"id": "2410.03952", "title": "Pixel-Based Similarities as an Alternative to Neural Data for Improving Convolutional Neural Network Adversarial Robustness", "abstract": "Convolutional Neural Networks (CNNs) excel in many visual tasks but remain susceptible to adversarial attacks-imperceptible perturbations that degrade performance. Prior research reveals that brain-inspired regularizers, derived from neural recordings, can bolster CNN robustness; however, reliance on specialized data limits practical adoption. We revisit a regularizer proposed by Li et al. (2019) that aligns CNN representations with neural representational similarity structures and introduce a data-driven variant. Instead of a neural recording-based similarity, our method computes a pixel-based similarity directly from images. This substitution retains the original biologically motivated loss formulation, preserving its robustness benefits while removing the need for neural measurements or task-specific augmentations. Notably, this data-driven variant provides the same robustness improvements observed with neural data. Our approach is lightweight and integrates easily into standard pipelines. Although we do not surpass cutting-edge specialized defenses, we show that neural representational insights can be leveraged without direct recordings. This underscores the promise of robust yet simple methods rooted in brain-inspired principles, even without specialized data, and raises the possibility that further integrating these insights could push performance closer to human levels without resorting to complex, specialized pipelines.", "date": "2024-10-04", "categories": []}
{"id": "2506.06027", "title": "Sample-Specific Noise Injection For Diffusion-Based Adversarial Purification", "abstract": "Diffusion-based purification (DBP) methods aim to remove adversarial noise from the input sample by first injecting Gaussian noise through a forward diffusion process, and then recovering the clean example through a reverse generative process. In the above process, how much Gaussian noise is injected to the input sample is key to the success of DBP methods, which is controlled by a constant noise level $t^*$ for all samples in existing methods. In this paper, we discover that an optimal $t^*$ for each sample indeed could be different. Intuitively, the cleaner a sample is, the less the noise it should be injected, and vice versa. Motivated by this finding, we propose a new framework, called Sample-specific Score-aware Noise Injection (SSNI). Specifically, SSNI uses a pre-trained score network to estimate how much a data point deviates from the clean data distribution (i.e., score norms). Then, based on the magnitude of score norms, SSNI applies a reweighting function to adaptively adjust $t^*$ for each sample, achieving sample-specific noise injections. Empirically, incorporating our framework with existing DBP methods results in a notable improvement in both accuracy and robustness on CIFAR-10 and ImageNet-1K, highlighting the necessity to allocate distinct noise levels to different samples in DBP methods. Our code is available at: https://github.com/tmlr-group/SSNI.", "date": "2025-06-06", "categories": []}
{"id": "2602.10382", "title": "Triggers Hijack Language Circuits: A Mechanistic Analysis of Backdoor Behaviors in Large Language Models", "abstract": "Backdoor attacks pose significant security risks for Large Language Models (LLMs), yet the internal mechanisms by which triggers operate remain poorly understood. We present the first mechanistic analysis of language-switching backdoors, studying the GAPperon model family (1B, 8B, 24B parameters) which contains triggers injected during pretraining that cause output language switching. Using activation patching, we localize trigger formation to early layers (7.5-25% of model depth) and identify which attention heads process trigger information. Our central finding is that trigger-activated heads substantially overlap with heads naturally encoding output language across model scales, with Jaccard indices between 0.18 and 0.66 over the top heads identified. This suggests that backdoor triggers do not form isolated circuits but instead co-opt the model's existing language components. These findings have implications for backdoor defense: detection methods may benefit from monitoring known functional components rather than searching for hidden circuits, and mitigation strategies could potentially leverage this entanglement between injected and natural behaviors.", "date": "2026-02-11", "categories": []}
{"id": "2602.12461", "title": "Semantic-aware Adversarial Fine-tuning for CLIP", "abstract": "Recent studies have shown that CLIP model's adversarial robustness in zero-shot classification tasks can be enhanced by adversarially fine-tuning its image encoder with adversarial examples (AEs), which are generated by minimizing the cosine similarity between images and a hand-crafted template (e.g., ''A photo of a {label}''). However, it has been shown that the cosine similarity between a single image and a single hand-crafted template is insufficient to measure the similarity for image-text pairs. Building on this, in this paper, we find that the AEs generated using cosine similarity may fail to fool CLIP when the similarity metric is replaced with semantically enriched alternatives, making the image encoder fine-tuned with these AEs less robust. To overcome this issue, we first propose a semantic-ensemble attack to generate semantic-aware AEs by minimizing the average similarity between the original image and an ensemble of refined textual descriptions. These descriptions are initially generated by a foundation model to capture core semantic features beyond hand-crafted templates and are then refined to reduce hallucinations. To this end, we propose Semantic-aware Adversarial Fine-Tuning (SAFT), which fine-tunes CLIP's image encoder with semantic-aware AEs. Extensive experiments show that SAFT outperforms current methods, achieving substantial improvements in zero-shot adversarial robustness across 16 datasets. Our code is available at: https://github.com/tmlr-group/SAFT.", "date": "2026-02-12", "categories": []}
{"id": "2602.12681", "title": "Fool Me If You Can: On the Robustness of Binary Code Similarity Detection Models against Semantics-preserving Transformations", "abstract": "Binary code analysis plays an essential role in cybersecurity, facilitating reverse engineering to reveal the inner workings of programs in the absence of source code. Traditional approaches, such as static and dynamic analysis, extract valuable insights from stripped binaries, but often demand substantial expertise and manual effort. Recent advances in deep learning have opened promising opportunities to enhance binary analysis by capturing latent features and disclosing underlying code semantics. Despite the growing number of binary analysis models based on machine learning, their robustness to adversarial code transformations at the binary level remains underexplored. We evaluate the robustness of deep learning models for the task of binary code similarity detection (BCSD) under semantics-preserving transformations. The unique nature of machine instructions presents distinct challenges compared to the typical input perturbations found in other domains. We introduce asmFooler, a system that evaluates the resilience of BCSD models using a diverse set of adversarial code transformations that preserve functional semantics. We construct a dataset of 9,565 binary variants from 620 baseline samples by applying eight semantics-preserving transformations across six representative BCSD models. Our major findings highlight several key insights: i) model robustness relies on the processing pipeline, including code pre-processing, architecture, and feature selection; ii) adversarial transformation effectiveness is bounded by a budget shaped by model-specific constraints like input size and instruction expressive capacity; iii) well-crafted transformations can be highly effective with minimal perturbations; and iv) such transformations efficiently disrupt model decisions (e.g., misleading to false positives or false negatives) by focusing on semantically significant instructions.", "date": "2026-02-13", "categories": []}
{"id": "2602.13062", "title": "Backdoor Attacks on Contrastive Continual Learning for IoT Systems", "abstract": "The Internet of Things (IoT) systems increasingly depend on continual learning to adapt to non-stationary environments. These environments can include factors such as sensor drift, changing user behavior, device aging, and adversarial dynamics. Contrastive continual learning (CCL) combines contrastive representation learning with incremental adaptation, enabling robust feature reuse across tasks and domains. However, the geometric nature of contrastive objectives, when paired with replay-based rehearsal and stability-preserving regularization, introduces new security vulnerabilities. Notably, backdoor attacks can exploit embedding alignment and replay reinforcement, enabling the implantation of persistent malicious behaviors that endure through updates and deployment cycles. This paper provides a comprehensive analysis of backdoor attacks on CCL within IoT systems. We formalize the objectives of embedding-level attacks, examine persistence mechanisms unique to IoT deployments, and develop a layered taxonomy tailored to IoT. Additionally, we compare vulnerabilities across various learning paradigms and evaluate defense strategies under IoT constraints, including limited memory, edge computing, and federated aggregation. Our findings indicate that while CCL is effective for enhancing adaptive IoT intelligence, it may also elevate long-lived representation-level threats if not adequately secured.", "date": "2026-02-13", "categories": []}
{"id": "2507.12314", "title": "Thought Purity: A Defense Framework For Chain-of-Thought Attack", "abstract": "Large Reasoning Models (LRMs) leverage Chain-of-Thought (CoT) reasoning to solve complex tasks, but this explicit reasoning process introduces a critical vulnerability: adversarial manipulation of the thought chain itself, known as Chain-of-Thought Attacks (CoTA). Such attacks subtly corrupt the reasoning path to produce erroneous outputs, challenging conventional defenses that often sacrifice model utility for safety. To address this, we propose Thought Purity(TP), a defense framework that shifts from passive refusal to active reasoning recovery. TP integrates a safety-aware data pipeline with reinforcement learning, employing a dual-reward mechanism to teach models to dynamically identify and isolate malicious logic while preserving correct reasoning. Experiments on multiple model families demonstrate that TP significantly reduces the attack success rate of CoTA while maintaining or enhancing the model's performance on benign tasks.", "date": "2025-07-16", "categories": []}
{"id": "2602.11204", "title": "Zero-Sacrifice Persistent-Robustness Adversarial Defense for Pre-Trained Encoders", "abstract": "The widespread use of publicly available pre-trained encoders from self-supervised learning (SSL) has exposed a critical vulnerability: their susceptibility to downstream-agnostic adversarial examples (DAEs), which are crafted without knowledge of the downstream tasks but capable of misleading downstream models. While several defense methods have been explored recently, they rely primarily on task-specific adversarial fine-tuning, which inevitably limits generalizability and causes catastrophic forgetting and deteriorates benign performance. Different with previous works, we propose a more rigorous defense goal that requires only a single tuning for diverse downstream tasks to defend against DAEs and preserve benign performance. To achieve this defense goal, we introduce Zero-Sacrifice Persistent-Robustness Adversarial Defense (ZePAD), which is inspired by the inherent sensitivity of neural networks to data characteristics. Specifically, ZePAD is a dual-branch structure, which consists of a Multi-Pattern Adversarial Enhancement Branch (MPAE-Branch) that uses two adversarially fine-tuned encoders to strengthen adversarial resistance. The Benign Memory Preservation Branch (BMP-Branch) is trained on local data to ensure adversarial robustness does not compromise benign performance. Surprisingly, we find that ZePAD can directly detect DAEs by evaluating branch confidence, without introducing any adversarial exsample identification task during training. Notably, by enriching feature diversity, our method enables a single adversarial fine-tuning to defend against DAEs across downstream tasks, thereby achieving persistent robustness. Extensive experiments on 11 SSL methods and 6 datasets validate its effectiveness. In certain cases, it achieves a 29.20% improvement in benign performance and a 73.86% gain in adversarial robustness, highlighting its zero-sacrifice property.", "date": "2026-02-10", "categories": []}
{"id": "2602.11646", "title": "Brain Tumor Classifiers Under Attack: Robustness of ResNet Variants Against Transferable FGSM and PGD Attacks", "abstract": "Adversarial robustness in deep learning models for brain tumor classification remains an underexplored yet critical challenge, particularly for clinical deployment scenarios involving MRI data. In this work, we investigate the susceptibility and resilience of several ResNet-based architectures, referred to as BrainNet, BrainNeXt and DilationNet, against gradient-based adversarial attacks, namely FGSM and PGD. These models, based on ResNet, ResNeXt, and dilated ResNet variants respectively, are evaluated across three preprocessing configurations (i) full-sized augmented, (ii) shrunk augmented and (iii) shrunk non-augmented MRI datasets. Our experiments reveal that BrainNeXt models exhibit the highest robustness to black-box attacks, likely due to their increased cardinality, though they produce weaker transferable adversarial samples. In contrast, BrainNet and Dilation models are more vulnerable to attacks from each other, especially under PGD with higher iteration steps and $\\alpha$ values. Notably, shrunk and non-augmented data significantly reduce model resilience, even when the untampered test accuracy remains high, highlighting a key trade-off between input resolution and adversarial vulnerability. These results underscore the importance of jointly evaluating classification performance and adversarial robustness for reliable real-world deployment in brain MRI analysis.", "date": "2026-02-12", "categories": []}
{"id": "2602.11940", "title": "Temporally Unified Adversarial Perturbations for Time Series Forecasting", "abstract": "While deep learning models have achieved remarkable success in time series forecasting, their vulnerability to adversarial examples remains a critical security concern. However, existing attack methods in the forecasting field typically ignore the temporal consistency inherent in time series data, leading to divergent and contradictory perturbation values for the same timestamp across overlapping samples. This temporally inconsistent perturbations problem renders adversarial attacks impractical for real-world data manipulation. To address this, we introduce Temporally Unified Adversarial Perturbations (TUAPs), which enforce a temporal unification constraint to ensure identical perturbations for each timestamp across all overlapping samples. Moreover, we propose a novel Timestamp-wise Gradient Accumulation Method (TGAM) that provides a modular and efficient approach to effectively generate TUAPs by aggregating local gradient information from overlapping samples. By integrating TGAM with momentum-based attack algorithms, we ensure strict temporal consistency while fully utilizing series-level gradient information to explore the adversarial perturbation space. Comprehensive experiments on three benchmark datasets and four representative state-of-the-art models demonstrate that our proposed method significantly outperforms baselines in both white-box and black-box transfer attack scenarios under TUAP constraints. Moreover, our method also exhibits superior transfer attack performance even without TUAP constraints, demonstrating its effectiveness and superiority in generating adversarial perturbations for time series forecasting models.", "date": "2026-02-12", "categories": []}
{"id": "2503.00038", "title": "from Benign import Toxic: Jailbreaking the Language Model via Adversarial Metaphors", "abstract": "Current studies have exposed the risk of Large Language Models (LLMs) generating harmful content by jailbreak attacks. However, they overlook that the direct generation of harmful content from scratch is more difficult than inducing LLM to calibrate benign content into harmful forms. In our study, we introduce a novel attack framework that exploits AdVersArial meTAphoR (AVATAR) to induce the LLM to calibrate malicious metaphors for jailbreaking. Specifically, to answer harmful queries, AVATAR adaptively identifies a set of benign but logically related metaphors as the initial seed. Then, driven by these metaphors, the target LLM is induced to reason and calibrate about the metaphorical content, thus jailbroken by either directly outputting harmful responses or calibrating residuals between metaphorical and professional harmful content. Experimental results demonstrate that AVATAR can effectively and transferable jailbreak LLMs and achieve a state-of-the-art attack success rate across multiple advanced LLMs.", "date": "2025-02-25", "categories": []}
{"id": "2511.10222", "title": "Speech-Audio Compositional Attacks on Multimodal LLMs and Their Mitigation with SALMONN-Guard", "abstract": "Recent progress in LLMs has enabled understanding of audio signals, but has also exposed new safety risks arising from complex audio inputs that are inadequately handled by current safeguards. We introduce SACRED-Bench (Speech-Audio Composition for RED-teaming) to evaluate the robustness of LLMs under complex audio-based attacks. Unlike existing perturbation-based methods that rely on noise optimization or white-box access, SACRED-Bench exploits speech-audio composition to enable effective black-box attacks. SACRED-Bench adopts three composition mechanisms: (a) overlap of harmful and benign speech, (b) mixture of benign speech with harmful non-speech audio, and (c) multi-speaker dialogue. These mechanisms focus on evaluating safety in settings where benign and harmful intents co-occur within a single auditory scene. Moreover, questions in SACRED-Bench are designed to implicitly refer to content in the audio, such that no explicit harmful information appears in the text prompt alone. Experiments demonstrate that even Gemini 2.5 Pro, a state-of-the-art proprietary LLM with safety guardrails fully enabled, still exhibits a 66% attack success rate. To bridge this gap, we propose SALMONN-Guard, the first guard model that jointly inspects speech, audio, and text for safety judgments, reducing the attack success rate to 20%. Our results highlight the need for audio-aware defenses to ensure the safety of multimodal LLMs. The dataset and SALMONN-Guard checkpoints can be found at https://huggingface.co/datasets/tsinghua-ee/SACRED-Bench.", "date": "2025-11-13", "categories": []}
{"id": "2602.10148", "title": "Red-teaming the Multimodal Reasoning: Jailbreaking Vision-Language Models via Cross-modal Entanglement Attacks", "abstract": "Vision-Language Models (VLMs) with multimodal reasoning capabilities are high-value attack targets, given their potential for handling complex multimodal harmful tasks. Mainstream black-box jailbreak attacks on VLMs work by distributing malicious clues across modalities to disperse model attention and bypass safety alignment mechanisms. However, these adversarial attacks rely on simple and fixed image-text combinations that lack attack complexity scalability, limiting their effectiveness for red-teaming VLMs' continuously evolving reasoning capabilities. We propose \\textbf{CrossTALK} (\\textbf{\\underline{Cross}}-modal en\\textbf{\\underline{TA}}ng\\textbf{\\underline{L}}ement attac\\textbf{\\underline{K}}), which is a scalable approach that extends and entangles information clues across modalities to exceed VLMs' trained and generalized safety alignment patterns for jailbreak. Specifically, {knowledge-scalable reframing} extends harmful tasks into multi-hop chain instructions, {cross-modal clue entangling} migrates visualizable entities into images to build multimodal reasoning links, and {cross-modal scenario nesting} uses multimodal contextual instructions to steer VLMs toward detailed harmful outputs. Experiments show our COMET achieves state-of-the-art attack success rate.", "date": "2026-02-09", "categories": []}
{"id": "2602.10160", "title": "AD$^2$: Analysis and Detection of Adversarial Threats in Visual Perception for End-to-End Autonomous Driving Systems", "abstract": "End-to-end autonomous driving systems have achieved significant progress, yet their adversarial robustness remains largely underexplored. In this work, we conduct a closed-loop evaluation of state-of-the-art autonomous driving agents under black-box adversarial threat models in CARLA. Specifically, we consider three representative attack vectors on the visual perception pipeline: (i) a physics-based blur attack induced by acoustic waves, (ii) an electromagnetic interference attack that distorts captured images, and (iii) a digital attack that adds ghost objects as carefully crafted bounded perturbations on images. Our experiments on two advanced agents, Transfuser and Interfuser, reveal severe vulnerabilities to such attacks, with driving scores dropping by up to 99% in the worst case, raising valid safety concerns. To help mitigate such threats, we further propose a lightweight Attack Detection model for Autonomous Driving systems (AD$^2$) based on attention mechanisms that capture spatial-temporal consistency. Comprehensive experiments across multi-camera inputs on CARLA show that our detector achieves superior detection capability and computational efficiency compared to existing approaches.", "date": "2026-02-10", "categories": []}
{"id": "2602.10179", "title": "When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing Models", "abstract": "Recent advances in large image editing models have shifted the paradigm from text-driven instructions to vision-prompt editing, where user intent is inferred directly from visual inputs such as marks, arrows, and visual-text prompts. While this paradigm greatly expands usability, it also introduces a critical and underexplored safety risk: the attack surface itself becomes visual. In this work, we propose Vision-Centric Jailbreak Attack (VJA), the first visual-to-visual jailbreak attack that conveys malicious instructions purely through visual inputs. To systematically study this emerging threat, we introduce IESBench, a safety-oriented benchmark for image editing models. Extensive experiments on IESBench demonstrate that VJA effectively compromises state-of-the-art commercial models, achieving attack success rates of up to 80.9% on Nano Banana Pro and 70.1% on GPT-Image-1.5. To mitigate this vulnerability, we propose a training-free defense based on introspective multimodal reasoning, which substantially improves the safety of poorly aligned models to a level comparable with commercial systems, without auxiliary guard models and with negligible computational overhead. Our findings expose new vulnerabilities, provide both a benchmark and practical defense to advance safe and trustworthy modern image editing systems. Warning: This paper contains offensive images created by large image editing models.", "date": "2026-02-10", "categories": []}
{"id": "2602.10319", "title": "A Low-Rank Defense Method for Adversarial Attack on Diffusion Models", "abstract": "Recently, adversarial attacks for diffusion models as well as their fine-tuning process have been developed rapidly. To prevent the abuse of these attack algorithms from affecting the practical application of diffusion models, it is critical to develop corresponding defensive strategies. In this work, we propose an efficient defensive strategy, named Low-Rank Defense (LoRD), to defend the adversarial attack on Latent Diffusion Models (LDMs). LoRD introduces the merging idea and a balance parameter, combined with the low-rank adaptation (LoRA) modules, to detect and defend the adversarial samples. Based on LoRD, we build up a defense pipeline that applies the learned LoRD modules to help diffusion models defend against attack algorithms. Our method ensures that the LDM fine-tuned on both adversarial and clean samples can still generate high-quality images. To demonstrate the effectiveness of our approach, we conduct extensive experiments on facial and landscape images, and our method shows significantly better defense performance compared to the baseline methods.", "date": "2026-02-10", "categories": []}
{"id": "2602.10780", "title": "Kill it with FIRE: On Leveraging Latent Space Directions for Runtime Backdoor Mitigation in Deep Neural Networks", "abstract": "Machine learning models are increasingly present in our everyday lives; as a result, they become targets of adversarial attackers seeking to manipulate the systems we interact with. A well-known vulnerability is a backdoor introduced into a neural network by poisoned training data or a malicious training process. Backdoors can be used to induce unwanted behavior by including a certain trigger in the input. Existing mitigations filter training data, modify the model, or perform expensive input modifications on samples. If a vulnerable model has already been deployed, however, those strategies are either ineffective or inefficient. To address this gap, we propose our inference-time backdoor mitigation approach called FIRE (Feature-space Inference-time REpair). We hypothesize that a trigger induces structured and repeatable changes in the model's internal representation. We view the trigger as directions in the latent spaces between layers that can be applied in reverse to correct the inference mechanism. Therefore, we turn the backdoored model against itself by manipulating its latent representations and moving a poisoned sample's features along the backdoor directions to neutralize the trigger. Our evaluation shows that FIRE has low computational overhead and outperforms current runtime mitigations on image benchmarks across various attacks, datasets, and network architectures.", "date": "2026-02-11", "categories": []}
{"id": "2411.16598", "title": "DiffBreak: Is Diffusion-Based Purification Robust?", "abstract": "Diffusion-based purification (DBP) has become a cornerstone defense against adversarial examples (AEs), regarded as robust due to its use of diffusion models (DMs) that project AEs onto the natural data manifold. We refute this core claim, theoretically proving that gradient-based attacks effectively target the DM rather than the classifier, causing DBP's outputs to align with adversarial distributions. This prompts a reassessment of DBP's robustness, attributing it to two critical flaws: incorrect gradients and inappropriate evaluation protocols that test only a single random purification of the AE. We show that with proper accounting for stochasticity and resubmission risk, DBP collapses. To support this, we introduce DiffBreak, the first reliable toolkit for differentiation through DBP, eliminating gradient flaws that previously further inflated robustness estimates. We also analyze the current defense scheme used for DBP where classification relies on a single purification, pinpointing its inherent invalidity. We provide a statistically grounded majority-vote (MV) alternative that aggregates predictions across multiple purified copies, showing partial but meaningful robustness gain. We then propose a novel adaptation of an optimization method against deepfake watermarking, crafting systemic perturbations that defeat DBP even under MV, challenging DBP's viability.", "date": "2024-11-25", "categories": []}
{"id": "2602.09343", "title": "Not-in-Perspective: Towards Shielding Google's Perspective API Against Adversarial Negation Attacks", "abstract": "The rise of cyberbullying in social media platforms involving toxic comments has escalated the need for effective ways to monitor and moderate online interactions. Existing solutions of automated toxicity detection systems, are based on a machine or deep learning algorithms. However, statistics-based solutions are generally prone to adversarial attacks that contain logic based modifications such as negation in phrases and sentences. In that regard, we present a set of formal reasoning-based methodologies that wrap around existing machine learning toxicity detection systems. Acting as both pre-processing and post-processing steps, our formal reasoning wrapper helps alleviating the negation attack problems and significantly improves the accuracy and efficacy of toxicity scoring. We evaluate different variations of our wrapper on multiple machine learning models against a negation adversarial dataset. Experimental results highlight the improvement of hybrid (formal reasoning and machine-learning) methods against various purely statistical solutions.", "date": "2026-02-10", "categories": []}
{"id": "2602.09629", "title": "Stop Testing Attacks, Start Diagnosing Defenses: The Four-Checkpoint Framework Reveals Where LLM Safety Breaks", "abstract": "Large Language Models (LLMs) deploy safety mechanisms to prevent harmful outputs, yet these defenses remain vulnerable to adversarial prompts. While existing research demonstrates that jailbreak attacks succeed, it does not explain \\textit{where} defenses fail or \\textit{why}.\n  To address this gap, we propose that LLM safety operates as a sequential pipeline with distinct checkpoints. We introduce the \\textbf{Four-Checkpoint Framework}, which organizes safety mechanisms along two dimensions: processing stage (input vs.\\ output) and detection level (literal vs.\\ intent). This creates four checkpoints, CP1 through CP4, each representing a defensive layer that can be independently evaluated. We design 13 evasion techniques, each targeting a specific checkpoint, enabling controlled testing of individual defensive layers.\n  Using this framework, we evaluate GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro across 3,312 single-turn, black-box test cases. We employ an LLM-as-judge approach for response classification and introduce Weighted Attack Success Rate (WASR), a severity-adjusted metric that captures partial information leakage overlooked by binary evaluation.\n  Our evaluation reveals clear patterns. Traditional Binary ASR reports 22.6\\% attack success. However, WASR reveals 52.7\\%, a 2.3$\\times$ higher vulnerability. Output-stage defenses (CP3, CP4) prove weakest at 72--79\\% WASR, while input-literal defenses (CP1) are strongest at 13\\% WASR. Claude achieves the strongest safety (42.8\\% WASR), followed by GPT-5 (55.9\\%) and Gemini (59.5\\%).\n  These findings suggest that current defenses are strongest at input-literal checkpoints but remain vulnerable to intent-level manipulation and output-stage techniques. The Four-Checkpoint Framework provides a structured approach for identifying and addressing safety vulnerabilities in deployed systems.", "date": "2026-02-10", "categories": []}
{"id": "2503.21164", "title": "Adversarial Wear and Tear: Exploiting Natural Damage for Generating Physical-World Adversarial Examples", "abstract": "The presence of adversarial examples in the physical world poses significant challenges to the deployment of Deep Neural Networks in safety-critical applications such as autonomous driving. Most existing methods for crafting physical-world adversarial examples are ad-hoc, relying on temporary modifications like shadows, laser beams, or stickers that are tailored to specific scenarios. In this paper, we introduce a new class of physical-world adversarial examples, AdvWT, which draws inspiration from the naturally occurring phenomenon of `wear and tear', an inherent property of physical objects. Unlike manually crafted perturbations, `wear and tear' emerges organically over time due to environmental degradation, as seen in the gradual deterioration of outdoor signboards. To achieve this, AdvWT follows a two-step approach. First, a GAN-based, unsupervised image-to-image translation network is employed to model these naturally occurring damages, particularly in the context of outdoor signboards. The translation network encodes the characteristics of damaged signs into a latent `damage style code'. In the second step, we introduce adversarial perturbations into the style code, strategically optimizing its transformation process. This manipulation subtly alters the damage style representation, guiding the network to generate adversarial images where the appearance of damages remains perceptually realistic, while simultaneously ensuring their effectiveness in misleading neural networks. Through comprehensive experiments on two traffic sign datasets, we show that AdvWT effectively misleads DNNs in both digital and physical domains. AdvWT achieves an effective attack success rate, greater robustness, and a more natural appearance compared to existing physical-world adversarial examples. Additionally, integrating AdvWT into training enhances a model's generalizability to real-world damaged signs.", "date": "2025-03-27", "categories": []}
{"id": "2506.06389", "title": "Exploring Adversarial Watermarking in Transformer-Based Models: Transferability and Robustness Against Defense Mechanism for Medical Images", "abstract": "Deep learning models have shown remarkable success in dermatological image analysis, offering potential for automated skin disease diagnosis. Previously, convolutional neural network(CNN) based architectures have achieved immense popularity and success in computer vision (CV) based task like skin image recognition, generation and video analysis. But with the emergence of transformer based models, CV tasks are now are nowadays carrying out using these models. Vision Transformers (ViTs) is such a transformer-based models that have shown success in computer vision. It uses self-attention mechanisms to achieve state-of-the-art performance across various tasks. However, their reliance on global attention mechanisms makes them susceptible to adversarial perturbations. This paper aims to investigate the susceptibility of ViTs for medical images to adversarial watermarking-a method that adds so-called imperceptible perturbations in order to fool models. By generating adversarial watermarks through Projected Gradient Descent (PGD), we examine the transferability of such attacks to CNNs and analyze the performance defense mechanism -- adversarial training. Results indicate that while performance is not compromised for clean images, ViTs certainly become much more vulnerable to adversarial attacks: an accuracy drop of as low as 27.6%. Nevertheless, adversarial training raises it up to 90.0%.", "date": "2025-06-05", "categories": []}
{"id": "2510.00181", "title": "CHAI: Command Hijacking against embodied AI", "abstract": "Embodied Artificial Intelligence (AI) promises to handle edge cases in robotic vehicle systems where data is scarce by using common-sense reasoning grounded in perception and action to generalize beyond training distributions and adapt to novel real-world situations. These capabilities, however, also create new security risks. In this paper, we introduce CHAI (Command Hijacking against embodied AI), a physical environment indirect prompt injection attack that exploits the multimodal language interpretation abilities of AI models. CHAI embeds deceptive natural language instructions, such as misleading signs, in visual input, systematically searches the token space, builds a dictionary of prompts, and guides an attacker model to generate Visual Attack Prompts. We evaluate CHAI on four LVLM agents: drone emergency landing, autonomous driving, aerial object tracking, and on a real robotic vehicle. Our experiments show that CHAI consistently outperforms state-of-the-art attacks. By exploiting the semantic and multimodal reasoning strengths of next-generation embodied AI systems, CHAI underscores the urgent need for defenses that extend beyond traditional adversarial robustness.", "date": "2025-09-30", "categories": []}
{"id": "2512.15769", "title": "Data-Chain Backdoor: Do You Trust Diffusion Models as Generative Data Supplier?", "abstract": "The increasing use of generative models such as diffusion models for synthetic data augmentation has greatly reduced the cost of data collection and labeling in downstream perception tasks. However, this new data source paradigm may introduce important security concerns. Publicly available generative models are often reused without verification, raising a fundamental question of their safety and trustworthiness. This work investigates backdoor propagation in such emerging generative data supply chain, namely, Data-Chain Backdoor (DCB). Specifically, we find that open-source diffusion models can become hidden carriers of backdoors. Their strong distribution-fitting ability causes them to memorize and reproduce backdoor triggers in generation, which are subsequently inherited by downstream models, resulting in severe security risks. This threat is particularly concerning under clean-label attack scenarios, as it remains effective while having negligible impact on the utility of the synthetic data. We study two attacker choices to obtain a backdoor-carried generator, training from scratch and fine-tuning. While naive fine-tuning leads to weak inheritance of the backdoor, we find that novel designs in the loss objectives and trigger processing can substantially improve the generator's ability to preserve trigger patterns, making fine-tuning a low-cost attack path. We evaluate the effectiveness of DCB under the standard augmentation protocol and further assess data-scarce settings. Across multiple trigger types, we observe that the trigger pattern can be consistently retained in the synthetic data with attack efficacy comparable to the conventional backdoor attack.", "date": "2025-12-12", "categories": []}
{"id": "2601.22636", "title": "Statistical Estimation of Adversarial Risk in Large Language Models under Best-of-N Sampling", "abstract": "Large Language Models (LLMs) are typically evaluated for safety under single-shot or low-budget adversarial prompting, which underestimates real-world risk. In practice, attackers can exploit large-scale parallel sampling to repeatedly probe a model until a harmful response is produced. While recent work shows that attack success increases with repeated sampling, principled methods for predicting large-scale adversarial risk remain limited. We propose a scaling-aware Best-of-N estimation of risk, SABER, for modeling jailbreak vulnerability under Best-of-N sampling. We model sample-level success probabilities using a Beta distribution, the conjugate prior of the Bernoulli distribution, and derive an analytic scaling law that enables reliable extrapolation of large-N attack success rates from small-budget measurements. Using only n=100 samples, our anchored estimator predicts ASR@1000 with a mean absolute error of 1.66, compared to 12.04 for the baseline, which is an 86.2% reduction in estimation error. Our results reveal heterogeneous risk scaling profiles and show that models appearing robust under standard evaluation can experience rapid nonlinear risk amplification under parallel adversarial pressure. This work provides a low-cost, scalable methodology for realistic LLM safety assessment. We will release our code and evaluation scripts upon publication to future research.", "date": "2026-01-30", "categories": []}
{"id": "2602.03284", "title": "Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks", "abstract": "Spiking neural networks (SNNs) compute with discrete spikes and exploit temporal structure, yet most adversarial attacks change intensities or event counts instead of timing. We study a timing-only adversary that retimes existing spikes while preserving spike counts and amplitudes in event-driven SNNs, thus remaining rate-preserving. We formalize a capacity-1 spike-retiming threat model with a unified trio of budgets: per-spike jitter $\\mathcal{B}_{\\infty}$, total delay $\\mathcal{B}_{1}$, and tamper count $\\mathcal{B}_{0}$. Feasible adversarial examples must satisfy timeline consistency and non-overlap, which makes the search space discrete and constrained. To optimize such retimings at scale, we use projected-in-the-loop (PIL) optimization: shift-probability logits yield a differentiable soft retiming for backpropagation, and a strict projection in the forward pass produces a feasible discrete schedule that satisfies capacity-1, non-overlap, and the chosen budget at every step. The objective maximizes task loss on the projected input and adds a capacity regularizer together with budget-aware penalties, which stabilizes gradients and aligns optimization with evaluation. Across event-driven benchmarks (CIFAR10-DVS, DVS-Gesture, N-MNIST) and diverse SNN architectures, we evaluate under binary and integer event grids and a range of retiming budgets, and also test models trained with timing-aware adversarial training designed to counter timing-only attacks. For example, on DVS-Gesture the attack attains high success (over $90\\%$) while touching fewer than $2\\%$ of spikes under $\\mathcal{B}_{0}$. Taken together, our results show that spike retiming is a practical and stealthy attack surface that current defenses struggle to counter, providing a clear reference for temporal robustness in event-driven SNNs. Code is available at https://github.com/yuyi-sd/Spike-Retiming-Attacks.", "date": "2026-02-03", "categories": []}
{"id": "2602.07028", "title": "A Comparative Study of Adversarial Robustness in CNN and CNN-ANFIS Architectures", "abstract": "Convolutional Neural Networks (CNNs) achieve strong image classification performance but lack interpretability and are vulnerable to adversarial attacks. Neuro-fuzzy hybrids such as DCNFIS replace fully connected CNN classifiers with Adaptive Neuro-Fuzzy Inference Systems (ANFIS) to improve interpretability, yet their robustness remains underexplored. This work compares standard CNNs (ConvNet, VGG, ResNet18) with their ANFIS-augmented counterparts on MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 under gradient-based (PGD) and gradient-free (Square) attacks. Results show that ANFIS integration does not consistently improve clean accuracy and has architecture-dependent effects on robustness: ResNet18-ANFIS exhibits improved adversarial robustness, while VGG-ANFIS often underperforms its baseline. These findings suggest that neuro-fuzzy augmentation can enhance robustness in specific architectures but is not universally beneficial.", "date": "2026-02-02", "categories": []}
{"id": "2602.07251", "title": "The Double-Edged Sword of Data-Driven Super-Resolution: Adversarial Super-Resolution Models", "abstract": "Data-driven super-resolution (SR) methods are often integrated into imaging pipelines as preprocessing steps to improve downstream tasks such as classification and detection. However, these SR models introduce a previously unexplored attack surface into imaging pipelines. In this paper, we present AdvSR, a framework demonstrating that adversarial behavior can be embedded directly into SR model weights during training, requiring no access to inputs at inference time. Unlike prior attacks that perturb inputs or rely on backdoor triggers, AdvSR operates entirely at the model level. By jointly optimizing for reconstruction quality and targeted adversarial outcomes, AdvSR produces models that appear benign under standard image quality metrics while inducing downstream misclassification. We evaluate AdvSR on three SR architectures (SRCNN, EDSR, SwinIR) paired with a YOLOv11 classifier and demonstrate that AdvSR models can achieve high attack success rates with minimal quality degradation. These findings highlight a new model-level threat for imaging pipelines, with implications for how practitioners source and validate models in safety-critical applications.", "date": "2026-02-06", "categories": []}
{"id": "2602.08136", "title": "Robustness of Vision Language Models Against Split-Image Harmful Input Attacks", "abstract": "Vision-Language Models (VLMs) are now a core part of modern AI. Recent work proposed several visual jailbreak attacks using single/ holistic images. However, contemporary VLMs demonstrate strong robustness against such attacks due to extensive safety alignment through preference optimization (e.g., RLHF). In this work, we identify a new vulnerability: while VLM pretraining and instruction tuning generalize well to split-image inputs, safety alignment is typically performed only on holistic images and does not account for harmful semantics distributed across multiple image fragments. Consequently, VLMs often fail to detect and refuse harmful split-image inputs, where unsafe cues emerge only after combining images. We introduce novel split-image visual jailbreak attacks (SIVA) that exploit this misalignment. Unlike prior optimization-based attacks, which exhibit poor black-box transferability due to architectural and prior mismatches across models, our attacks evolve in progressive phases from naive splitting to an adaptive white-box attack, culminating in a black-box transfer attack. Our strongest strategy leverages a novel adversarial knowledge distillation (Adv-KD) algorithm to substantially improve cross-model transferability. Evaluations on three state-of-the-art modern VLMs and three jailbreak datasets demonstrate that our strongest attack achieves up to 60% higher transfer success than existing baselines. Lastly, we propose efficient ways to address this critical vulnerability in the current VLM safety alignment.", "date": "2026-02-08", "categories": []}
{"id": "2602.08170", "title": "Evasion of IoT Malware Detection via Dummy Code Injection", "abstract": "The Internet of Things (IoT) has revolutionized connectivity by linking billions of devices worldwide. However, this rapid expansion has also introduced severe security vulnerabilities, making IoT devices attractive targets for malware such as the Mirai botnet. Power side-channel analysis has recently emerged as a promising technique for detecting malware activity based on device power consumption patterns. However, the resilience of such detection systems under adversarial manipulation remains underexplored.\n  This work presents a novel adversarial strategy against power side-channel-based malware detection. By injecting structured dummy code into the scanning phase of the Mirai botnet, we dynamically perturb power signatures to evade AI/ML-based anomaly detection without disrupting core functionality. Our approach systematically analyzes the trade-offs between stealthiness, execution overhead, and evasion effectiveness across multiple state-of-the-art models for side-channel analysis, using a custom dataset collected from smartphones of diverse manufacturers. Experimental results show that our adversarial modifications achieve an average attack success rate of 75.2\\%, revealing practical vulnerabilities in power-based intrusion detection frameworks.", "date": "2026-02-09", "categories": []}
{"id": "2602.08214", "title": "RECUR: Resource Exhaustion Attack via Recursive-Entropy Guided Counterfactual Utilization and Reflection", "abstract": "Large Reasoning Models (LRMs) employ reasoning to address complex tasks. Such explicit reasoning requires extended context lengths, resulting in substantially higher resource consumption. Prior work has shown that adversarially crafted inputs can trigger redundant reasoning processes, exposing LRMs to resource-exhaustion vulnerabilities. However, the reasoning process itself, especially its reflective component, has received limited attention, even though it can lead to over-reflection and consume excessive computing power. In this paper, we introduce Recursive Entropy to quantify the risk of resource consumption in reflection, thereby revealing the safety issues inherent in inference itself. Based on Recursive Entropy, we introduce RECUR, a resource exhaustion attack via Recursive Entropy guided Counterfactual Utilization and Reflection. It constructs counterfactual questions to verify the inherent flaws and risks of LRMs. Extensive experiments demonstrate that, under benign inference, recursive entropy exhibits a pronounced decreasing trend. RECUR disrupts this trend, increasing the output length by up to 11x and decreasing throughput by 90%. Our work provides a new perspective on robust reasoning.", "date": "2026-02-09", "categories": []}
{"id": "2602.08230", "title": "Generating Adversarial Events: A Motion-Aware Point Cloud Framework", "abstract": "Event cameras have been widely adopted in safety-critical domains such as autonomous driving, robotics, and human-computer interaction. A pressing challenge arises from the vulnerability of deep neural networks to adversarial examples, which poses a significant threat to the reliability of event-based systems. Nevertheless, research into adversarial attacks on events is scarce. This is primarily due to the non-differentiable nature of mainstream event representations, which hinders the extension of gradient-based attack methods. In this paper, we propose MA-ADV, a novel \\textbf{M}otion-\\textbf{A}ware \\textbf{Adv}ersarial framework. To the best of our knowledge, this is the first work to generate adversarial events by leveraging point cloud representations. MA-ADV accounts for high-frequency noise in events and employs a diffusion-based approach to smooth perturbations, while fully leveraging the spatial and temporal relationships among events. Finally, MA-ADV identifies the minimal-cost perturbation through a combination of sample-wise Adam optimization, iterative refinement, and binary search. Extensive experimental results validate that MA-ADV ensures a 100\\% attack success rate with minimal perturbation cost, and also demonstrate enhanced robustness against defenses, underscoring the critical security challenges facing future event-based perception systems.", "date": "2026-02-09", "categories": []}
{"id": "2602.08679", "title": "Dashed Line Defense: Plug-And-Play Defense Against Adaptive Score-Based Query Attacks", "abstract": "Score-based query attacks pose a serious threat to deep learning models by crafting adversarial examples (AEs) using only black-box access to model output scores, iteratively optimizing inputs based on observed loss values. While recent runtime defenses attempt to disrupt this process via output perturbation, most either require access to model parameters or fail when attackers adapt their tactics. In this paper, we first reveal that even the state-of-the-art plug-and-play defense can be bypassed by adaptive attacks, exposing a critical limitation of existing runtime defenses. We then propose Dashed Line Defense (DLD), a plug-and-play post-processing method specifically designed to withstand adaptive query strategies. By introducing ambiguity in how the observed loss reflects the true adversarial strength of candidate examples, DLD prevents attackers from reliably analyzing and adapting their queries, effectively disrupting the AE generation process. We provide theoretical guarantees of DLD's defense capability and validate its effectiveness through experiments on ImageNet, demonstrating that DLD consistently outperforms prior defenses--even under worst-case adaptive attacks--while preserving the model's predicted labels.", "date": "2026-02-09", "categories": []}
{"id": "2509.15370", "title": "Adversarial generalization of unfolding (model-based) networks", "abstract": "Unfolding networks are interpretable networks emerging from iterative algorithms, incorporate prior knowledge of data structure, and are designed to solve inverse problems like compressed sensing, which deals with recovering data from noisy, missing observations. Compressed sensing finds applications in critical domains, from medical imaging to cryptography, where adversarial robustness is crucial to prevent catastrophic failures. However, a solid theoretical understanding of the performance of unfolding networks in the presence of adversarial attacks is still in its infancy. In this paper, we study the adversarial generalization of unfolding networks when perturbed with $l_2$-norm constrained attacks, generated by the fast gradient sign method. Particularly, we choose a family of state-of-the-art overaparameterized unfolding networks and deploy a new framework to estimate their adversarial Rademacher complexity. Given this estimate, we provide adversarial generalization error bounds for the networks under study, which are tight with respect to the attack level. To our knowledge, this is the first theoretical analysis on the adversarial generalization of unfolding networks. We further present a series of experiments on real-world data, with results corroborating our derived theory, consistently for all data. Finally, we observe that the family's overparameterization can be exploited to promote adversarial robustness, shedding light on how to efficiently robustify neural networks.", "date": "2025-09-18", "categories": []}
{"id": "2602.06269", "title": "PurSAMERE: Reliable Adversarial Purification via Sharpness-Aware Minimization of Expected Reconstruction Error", "abstract": "We propose a novel deterministic purification method to improve adversarial robustness by mapping a potentially adversarial sample toward a nearby sample that lies close to a mode of the data distribution, where classifiers are more reliable. We design the method to be deterministic to ensure reliable test accuracy and to prevent the degradation of effective robustness observed in stochastic purification approaches when the adversary has full knowledge of the system and its randomness. We employ a score model trained by minimizing the expected reconstruction error of noise-corrupted data, thereby learning the structural characteristics of the input data distribution. Given a potentially adversarial input, the method searches within its local neighborhood for a purified sample that minimizes the expected reconstruction error under noise corruption and then feeds this purified sample to the classifier. During purification, sharpness-aware minimization is used to guide the purified samples toward flat regions of the expected reconstruction error landscape, thereby enhancing robustness. We further show that, as the noise level decreases, minimizing the expected reconstruction error biases the purified sample toward local maximizers of the Gaussian-smoothed density; under additional local assumptions on the score model, we prove recovery of a local maximizer in the small-noise limit. Experimental results demonstrate significant gains in adversarial robustness over state-of-the-art methods under strong deterministic white-box attacks.", "date": "2026-02-06", "categories": []}
{"id": "2602.06395", "title": "Empirical Analysis of Adversarial Robustness and Explainability Drift in Cybersecurity Classifiers", "abstract": "Machine learning (ML) models are increasingly deployed in cybersecurity applications such as phishing detection and network intrusion prevention. However, these models remain vulnerable to adversarial perturbations small, deliberate input modifications that can degrade detection accuracy and compromise interpretability. This paper presents an empirical study of adversarial robustness and explainability drift across two cybersecurity domains phishing URL classification and network intrusion detection. We evaluate the impact of L (infinity) bounded Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) perturbations on model accuracy and introduce a quantitative metric, the Robustness Index (RI), defined as the area under the accuracy perturbation curve. Gradient based feature sensitivity and SHAP based attribution drift analyses reveal which input features are most susceptible to adversarial manipulation. Experiments on the Phishing Websites and UNSW NB15 datasets show consistent robustness trends, with adversarial training improving RI by up to 9 percent while maintaining clean-data accuracy. These findings highlight the coupling between robustness and interpretability degradation and underscore the importance of quantitative evaluation in the design of trustworthy, AI-driven cybersecurity systems.", "date": "2026-02-06", "categories": []}
{"id": "2602.06577", "title": "Perturbing the Phase: Analyzing Adversarial Robustness of Complex-Valued Neural Networks", "abstract": "Complex-valued neural networks (CVNNs) are rising in popularity for all kinds of applications. To safely use CVNNs in practice, analyzing their robustness against outliers is crucial. One well known technique to understand the behavior of deep neural networks is to investigate their behavior under adversarial attacks, which can be seen as worst case minimal perturbations. We design Phase Attacks, a kind of attack specifically targeting the phase information of complex-valued inputs. Additionally, we derive complex-valued versions of commonly used adversarial attacks. We show that in some scenarios CVNNs are more robust than RVNNs and that both are very susceptible to phase changes with the Phase Attacks decreasing the model performance more, than equally strong regular attacks, which can attack both phase and magnitude.", "date": "2026-02-06", "categories": []}
{"id": "2602.06578", "title": "Exploring Sparsity and Smoothness of Arbitrary $\\ell_p$ Norms in Adversarial Attacks", "abstract": "Adversarial attacks against deep neural networks are commonly constructed under $\\ell_p$ norm constraints, most often using $p=1$, $p=2$ or $p=\\infty$, and potentially regularized for specific demands such as sparsity or smoothness. These choices are typically made without a systematic investigation of how the norm parameter \\( p \\) influences the structural and perceptual properties of adversarial perturbations. In this work, we study how the choice of \\( p \\) affects sparsity and smoothness of adversarial attacks generated under \\( \\ell_p \\) norm constraints for values of $p \\in [1,2]$. To enable a quantitative analysis, we adopt two established sparsity measures from the literature and introduce three smoothness measures. In particular, we propose a general framework for deriving smoothness measures based on smoothing operations and additionally introduce a smoothness measure based on first-order Taylor approximations. Using these measures, we conduct a comprehensive empirical evaluation across multiple real-world image datasets and a diverse set of model architectures, including both convolutional and transformer-based networks. We show that the choice of $\\ell_1$ or $\\ell_2$ is suboptimal in most cases and the optimal $p$ value is dependent on the specific task. In our experiments, using $\\ell_p$ norms with $p\\in [1.3, 1.5]$ yields the best trade-off between sparse and smooth attacks. These findings highlight the importance of principled norm selection when designing and evaluating adversarial attacks.", "date": "2026-02-06", "categories": []}
{"id": "2602.06854", "title": "SEMA: Simple yet Effective Learning for Multi-Turn Jailbreak Attacks", "abstract": "Multi-turn jailbreaks capture the real threat model for safety-aligned chatbots, where single-turn attacks are merely a special case. Yet existing approaches break under exploration complexity and intent drift. We propose SEMA, a simple yet effective framework that trains a multi-turn attacker without relying on any existing strategies or external data. SEMA comprises two stages. Prefilling self-tuning enables usable rollouts by fine-tuning on non-refusal, well-structured, multi-turn adversarial prompts that are self-generated with a minimal prefix, thereby stabilizing subsequent learning. Reinforcement learning with intent-drift-aware reward trains the attacker to elicit valid multi-turn adversarial prompts while maintaining the same harmful objective. We anchor harmful intent in multi-turn jailbreaks via an intent-drift-aware reward that combines intent alignment, compliance risk, and level of detail. Our open-loop attack regime avoids dependence on victim feedback, unifies single- and multi-turn settings, and reduces exploration complexity. Across multiple datasets, victim models, and jailbreak judges, our method achieves state-of-the-art (SOTA) attack success rates (ASR), outperforming all single-turn baselines, manually scripted and template-driven multi-turn baselines, as well as our SFT (Supervised Fine-Tuning) and DPO (Direct Preference Optimization) variants. For instance, SEMA performs an average $80.1\\%$ ASR@1 across three closed-source and open-source victim models on AdvBench, 33.9% over SOTA. The approach is compact, reproducible, and transfers across targets, providing a stronger and more realistic stress test for large language model (LLM) safety and enabling automatic redteaming to expose and localize failure modes. Our code is available at: https://github.com/fmmarkmq/SEMA.", "date": "2026-02-06", "categories": []}
{"id": "2403.19254", "title": "Imperceptible Protection against Style Imitation from Diffusion Models", "abstract": "Recent progress in diffusion models has profoundly enhanced the fidelity of image generation, but it has raised concerns about copyright infringements. While prior methods have introduced adversarial perturbations to prevent style imitation, most are accompanied by the degradation of artworks' visual quality. Recognizing the importance of maintaining this, we introduce a visually improved protection method while preserving its protection capability. To this end, we devise a perceptual map to highlight areas sensitive to human eyes, guided by instance-aware refinement, which refines the protection intensity accordingly. We also introduce a difficulty-aware protection by predicting how difficult the artwork is to protect and dynamically adjusting the intensity based on this. Lastly, we integrate a perceptual constraints bank to further improve the imperceptibility. Results show that our method substantially elevates the quality of the protected image without compromising on protection efficacy.", "date": "2024-03-28", "categories": []}
{"id": "2506.24068", "title": "STACK: Adversarial Attacks on LLM Safeguard Pipelines", "abstract": "Frontier AI developers are relying on layers of safeguards to protect against catastrophic misuse of AI systems. Anthropic and OpenAI guard their latest Opus 4 model and GPT-5 models using such defense pipelines, and other frontier developers including Google DeepMind pledge to soon deploy similar defenses. However, the security of such pipelines is unclear, with limited prior work evaluating or attacking these pipelines. We address this gap by developing and red-teaming an open-source defense pipeline. First, we find that a novel few-shot-prompted input and output classifier outperforms state-of-the-art open-weight safeguard model ShieldGemma across three attacks and two datasets, reducing the attack success rate (ASR) to 0% on the catastrophic misuse dataset ClearHarm. Second, we introduce a STaged AttaCK (STACK) procedure that achieves 71% ASR on ClearHarm in a black-box attack against the few-shot-prompted classifier pipeline. Finally, we also evaluate STACK in a transfer setting, achieving 33% ASR, providing initial evidence that it is feasible to design attacks with no access to the target pipeline. We conclude by suggesting specific mitigations that developers could use to thwart staged attacks.", "date": "2025-06-30", "categories": []}
{"id": "2510.09712", "title": "Group-Adaptive Adversarial Learning for Robust Fake News Detection Against Malicious Comments", "abstract": "Online fake news profoundly distorts public judgment and erodes trust in social platforms. While existing detectors achieve competitive performance on benchmark datasets, they remain notably vulnerable to malicious comments designed specifically to induce misclassification. This evolving threat landscape necessitates detection systems that simultaneously prioritize predictive accuracy and structural robustness. However, current detectors often fail to generalize across diverse and novel comment attack patterns. To bridge this gap, we propose AdComment, an adaptive adversarial training framework for robustness enhancement against diverse malicious comments. Based on cognitive psychology, we categorize adversarial comments into Fact Distortion, Logical Confusion, and Emotional Manipulation, and leverage LLMs to synthesize diverse, category-specific perturbations. Central to our framework is an InfoDirichlet Resampling (IDR) mechanism that dynamically adjusts malicious comment proportions during training, thereby steering optimization toward the model's most susceptible regions. Experimental results demonstrate that our approach achieves state-of-the-art performance on three benchmark datasets, improving the F1 scores by 17.9%, 14.5% and 9.0%, respectively.", "date": "2025-10-10", "categories": []}
{"id": "2602.04711", "title": "Addressing Corpus Knowledge Poisoning Attacks on RAG Using Sparse Attention", "abstract": "Retrieval Augmented Generation (RAG) is a highly effective paradigm for keeping LLM-based responses up-to-date and reducing the likelihood of hallucinations. Yet, RAG was recently shown to be quite vulnerable to corpus knowledge poisoning: an attacker injects misleading documents to the corpus to steer an LLM's output to an undesired response. We argue that the standard causal attention mechanism in LLMs enables harmful cross-document interactions, specifically in cases of attacks. Accordingly, we introduce a novel defense approach for RAG: Sparse Document Attention RAG (SDAG). This is a block-sparse attention mechanism that disallows cross-attention between retrieved documents. SDAG requires a minimal inference-time change to the attention mask; furthermore, no fine-tuning or additional architectural changes are needed. We present an empirical evaluation of LLM-based question answering (QA) with a variety of attack strategies on RAG. We show that our SDAG method substantially outperforms the standard causal attention mechanism in terms of attack success rate. We further demonstrate the clear merits of integrating SDAG with state-of-the-art RAG defense methods. Specifically, the integration results in performance that is statistically significantly better than the state-of-the-art.", "date": "2026-02-04", "categories": []}
{"id": "2411.17957", "title": "DiffVax: Optimization-Free Image Immunization Against Diffusion-Based Editing", "abstract": "Current image immunization defense techniques against diffusion-based editing embed imperceptible noise into target images to disrupt editing models. However, these methods face scalability challenges, as they require time-consuming optimization for each image separately, taking hours for small batches. To address these challenges, we introduce DiffVax, a scalable, lightweight, and optimization-free framework for image immunization, specifically designed to prevent diffusion-based editing. Our approach enables effective generalization to unseen content, reducing computational costs and cutting immunization time from days to milliseconds, achieving a speedup of 250,000x. This is achieved through a loss term that ensures the failure of editing attempts and the imperceptibility of the perturbations. Extensive qualitative and quantitative results demonstrate that our model is scalable, optimization-free, adaptable to various diffusion-based editing tools, robust against counter-attacks, and, for the first time, effectively protects video content from editing. More details are available in https://diffvax.github.io/ .", "date": "2024-11-27", "categories": []}
{"id": "2502.02542", "title": "OverThink: Slowdown Attacks on Reasoning LLMs", "abstract": "Most flagship language models generate explicit reasoning chains, enabling inference-time scaling. However, producing these reasoning chains increases token usage (i.e., reasoning tokens), which in turn increases latency and costs. Our OverThink attack increases overhead for applications that rely on reasoning language models (RLMs) and external context by forcing them to spend substantially more reasoning tokens while still producing contextually correct answers. An adversary mounts an attack by injecting decoy reasoning problems into public content that is consumed by RLM at inference time. Because our decoys (e.g., Markov decision processes, Sudokus, etc.) are benign, they evade safety filters. We evaluate OverThink on both closed-source and open-source reasoning models across the FreshQA, SQuAD, and MuSR datasets. We also explore the attack in multi-modal settings by creating images that cause excessive reasoning. We show that the resulting slowdown transfers across models. Finally, we explore both LLM-based and systems-level defenses, and discuss the societal, financial, and energy implications of the OverThink attacks.", "date": "2025-02-04", "categories": []}
{"id": "2505.17440", "title": "VEAttack: Downstream-agnostic Vision Encoder Attack against Large Vision Language Models", "abstract": "Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding and generation, yet their vulnerability to adversarial attacks raises significant robustness concerns. While existing effective attacks always focus on task-specific white-box settings, these approaches are limited in the context of LVLMs, which are designed for diverse downstream tasks and require expensive full-model gradient computations. Motivated by the pivotal role and wide adoption of the vision encoder in LVLMs, we propose a simple yet effective Vision Encoder Attack (VEAttack), which targets the vision encoder of LVLMs only. Specifically, we propose to generate adversarial examples by minimizing the cosine similarity between the clean and perturbed visual features, without accessing the following large language models, task information, and labels. It significantly reduces the computational overhead while eliminating the task and label dependence of traditional white-box attacks in LVLMs. To make this simple attack effective, we propose to perturb images by optimizing image tokens instead of the classification token. We provide both empirical and theoretical evidence that VEAttack can easily generalize to various tasks. VEAttack has achieved a performance degradation of 94.5% on image caption task and 75.7% on visual question answering task. We also reveal some key observations to provide insights into LVLM attack/defense: 1) hidden layer variations of LLM, 2) token attention differential, 3) M\\\"obius band in transfer attack, 4) low sensitivity to attack steps. The code is available at https://github.com/hefeimei06/VEAttack-LVLM.", "date": "2025-05-23", "categories": []}
{"id": "2506.21142", "title": "Generative Adversarial Evasion and Out-of-Distribution Detection for UAV Cyber-Attacks", "abstract": "The growing integration of UAVs into civilian airspace underscores the need for resilient and intelligent intrusion detection systems (IDS), as traditional anomaly detection methods often fail to identify novel threats. A common approach treats unfamiliar attacks as out-of-distribution (OOD) samples; however, this leaves systems vulnerable when mitigation is inadequate. Moreover, conventional OOD detectors struggle to distinguish stealthy adversarial attacks from genuine OOD events. This paper introduces a conditional generative adversarial network (cGAN)-based framework for crafting stealthy adversarial attacks that evade IDS mechanisms. We first design a robust multi-class IDS classifier trained on benign UAV telemetry and known cyber-attacks, including Denial of Service (DoS), false data injection (FDI), man-in-the-middle (MiTM), and replay attacks. Using this classifier, our cGAN perturbs known attacks to generate adversarial samples that misclassify as benign while retaining statistical resemblance to OOD distributions. These adversarial samples are iteratively refined to achieve high stealth and success rates. To detect such perturbations, we implement a conditional variational autoencoder (CVAE), leveraging negative log-likelihood to separate adversarial inputs from authentic OOD samples. Comparative evaluation shows that CVAE-based regret scores significantly outperform traditional Mahalanobis distance-based detectors in identifying stealthy adversarial threats. Our findings emphasize the importance of advanced probabilistic modeling to strengthen IDS capabilities against adaptive, generative-model-based cyber intrusions.", "date": "2025-06-26", "categories": []}
{"id": "2508.03365", "title": "When Good Sounds Go Adversarial: Jailbreaking Audio-Language Models with Benign Inputs", "abstract": "As large language models (LLMs) become increasingly integrated into daily life, audio has emerged as a key interface for human-AI interaction. However, this convenience also introduces new vulnerabilities, making audio a potential attack surface for adversaries. Our research introduces WhisperInject, a two-stage adversarial audio attack framework that manipulates state-of-the-art audio language models to generate harmful content. Our method embeds harmful payloads as subtle perturbations into audio inputs that remain intelligible to human listeners. The first stage uses a novel reward-based white-box optimization method, Reinforcement Learning with Projected Gradient Descent (RL-PGD), to jailbreak the target model and elicit harmful native responses. This native harmful response then serves as the target for Stage 2, Payload Injection, where we use gradient-based optimization to embed subtle perturbations into benign audio carriers, such as weather queries or greeting messages. Our method achieves average attack success rates of 60-78% across two benchmarks and five multimodal LLMs, validated by multiple evaluation frameworks. Our work demonstrates a new class of practical, audio-native threats, moving beyond theoretical exploits to reveal a feasible and covert method for manipulating multimodal AI systems.", "date": "2025-08-05", "categories": []}
{"id": "2602.04236", "title": "Cascading Robustness Verification: Toward Efficient Model-Agnostic Certification", "abstract": "Certifying neural network robustness against adversarial examples is challenging, as formal guarantees often require solving non-convex problems. Hence, incomplete verifiers are widely used because they scale efficiently and substantially reduce the cost of robustness verification compared to complete methods. However, relying on a single verifier can underestimate robustness because of loose approximations or misalignment with training methods. In this work, we propose Cascading Robustness Verification (CRV), which goes beyond an engineering improvement by exposing fundamental limitations of existing robustness metric and introducing a framework that enhances both reliability and efficiency. CRV is a model-agnostic verifier, meaning that its robustness guarantees are independent of the model's training process. The key insight behind the CRV framework is that, when using multiple verification methods, an input is certifiably robust if at least one method certifies it as robust. Rather than relying solely on a single verifier with a fixed constraint set, CRV progressively applies multiple verifiers to balance the tightness of the bound and computational cost. Starting with the least expensive method, CRV halts as soon as an input is certified as robust; otherwise, it proceeds to more expensive methods. For computationally expensive methods, we introduce a Stepwise Relaxation Algorithm (SR) that incrementally adds constraints and checks for certification at each step, thereby avoiding unnecessary computation. Our theoretical analysis demonstrates that CRV achieves equal or higher verified accuracy compared to powerful but computationally expensive incomplete verifiers in the cascade, while significantly reducing verification overhead. Empirical results confirm that CRV certifies at least as many inputs as benchmark approaches, while improving runtime efficiency by up to ~90%.", "date": "2026-02-04", "categories": []}
{"id": "2602.04356", "title": "When and Where to Attack? Stage-wise Attention-Guided Adversarial Attack on Large Vision Language Models", "abstract": "Adversarial attacks against Large Vision-Language Models (LVLMs) are crucial for exposing safety vulnerabilities in modern multimodal systems. Recent attacks based on input transformations, such as random cropping, suggest that spatially localized perturbations can be more effective than global image manipulation. However, randomly cropping the entire image is inherently stochastic and fails to use the limited per-pixel perturbation budget efficiently. We make two key observations: (i) regional attention scores are positively correlated with adversarial loss sensitivity, and (ii) attacking high-attention regions induces a structured redistribution of attention toward subsequent salient regions. Based on these findings, we propose Stage-wise Attention-Guided Attack (SAGA), an attention-guided framework that progressively concentrates perturbations on high-attention regions. SAGA enables more efficient use of constrained perturbation budgets, producing highly imperceptible adversarial examples while consistently achieving state-of-the-art attack success rates across ten LVLMs. The source code is available at https://github.com/jackwaky/SAGA.", "date": "2026-02-04", "categories": []}
{"id": "2505.14103", "title": "AudioJailbreak: Jailbreak Attacks against End-to-End Large Audio-Language Models", "abstract": "Jailbreak attacks to Large audio-language models (LALMs) are studied recently, but they exclusively focused on the attack scenario where the adversary can fully manipulate user prompts (named strong adversary) and limited in effectiveness, applicability, and practicability. In this work, we first conduct an extensive evaluation showing that advanced text jailbreak attacks cannot be easily ported to end-to-end LALMs via text-to-speech (TTS) techniques. We then propose AUDIOJAILBREAK, a novel audio jailbreak attack, featuring (1) asynchrony: the jailbreak audios do not need to align with user prompts in the time axis by crafting suffixal jailbreak audios; (2) universality: a single jailbreak perturbation is effective for different prompts by incorporating multiple prompts into the perturbation generation; (3) stealthiness: the malicious intent of jailbreak audios is concealed by proposing various intent concealment strategies; and (4) over-the-air robustness: the jailbreak audios remain effective when being played over the air by incorporating reverberation into the perturbation generation. In contrast, all prior audio jailbreak attacks cannot offer asynchrony, universality, stealthiness, and/or over-the-air robustness. Moreover, AUDIOJAILBREAK is also applicable to a more practical and broader attack scenario where the adversary cannot fully manipulate user prompts (named weak adversary). Extensive experiments with thus far the most LALMs demonstrate the high effectiveness of AUDIOJAILBREAK, in particular, it can jailbreak openAI's GPT-4o-Audio and bypass Meta's Llama-Guard-3 safeguard, in the weak adversary scenario. We highlight that our work peeks into the security implications of audio jailbreak attacks against LALMs, and realistically fosters improving their robustness, especially for the newly proposed weak adversary.", "date": "2025-05-20", "categories": []}
{"id": "2507.08261", "title": "Admissibility of Stein Shrinkage for Batch Normalization in the Presence of Adversarial Attacks", "abstract": "Batch normalization (BN) is a ubiquitous operation in deep neural networks, primarily used to improve stability and regularization during training. BN centers and scales feature maps using sample means and variances, which are naturally suited for Stein's shrinkage estimation. Applying such shrinkage yields more accurate mean and variance estimates of the batch in the mean-squared-error sense. In this paper, we prove that the Stein shrinkage estimator of the mean and variance dominates over the sample mean and variance estimators, respectively, in the presence of adversarial attacks modeled using sub-Gaussian distributions. Furthermore, by construction, the James-Stein (JS) BN yields a smaller local Lipschitz constant compared to the vanilla BN, implying better regularity properties and potentially improved robustness. This facilitates and justifies the application of Stein shrinkage to estimate the mean and variance parameters in BN and the use of it in image classification and segmentation tasks with and without adversarial attacks. We present SOTA performance results using this Stein-corrected BN in a standard ResNet architecture applied to the task of image classification using CIFAR-10 data, 3D CNN on PPMI (neuroimaging) data, and image segmentation using HRNet on Cityscape data with and without adversarial attacks.", "date": "2025-07-11", "categories": []}
{"id": "2510.05052", "title": "Proactive defense against LLM Jailbreak", "abstract": "The proliferation of powerful large language models (LLMs) has necessitated robust safety alignment, yet these models remain vulnerable to evolving adversarial attacks, including multi-turn jailbreaks that iteratively search for successful queries. Current defenses, which are primarily reactive and static, often fail to handle these iterative attacks. In this paper, we introduce ProAct, a novel proactive defense framework designed to disrupt and mislead these iterative search jailbreak methods. Our core idea is to intentionally mislead these jailbreak methods into thinking that the model has been jailbroken with \"spurious responses\". These misleading responses provide false signals to the attacker's internal optimization loop, causing the adversarial search to terminate prematurely and effectively jailbreaking the jailbreak. By conducting extensive experiments across state-of-the-art LLMs, jailbreaking frameworks, and safety benchmarks, we demonstrate that our method consistently and significantly reduces attack success rates by up to 94% without affecting utility. When combined with other defense fraeworks, it further reduces the latest attack strategies' success rate to 0%. ProActrepresents an orthogonal defense strategy that serves as an additional guardrail to enhance LLM safety against the most effective jailbreaking attacks.", "date": "2025-10-06", "categories": []}
{"id": "2510.10000", "title": "Tight Robustness Certificates and Wasserstein Distributional Attacks for Deep Neural Networks", "abstract": "Wasserstein distributionally robust optimization (WDRO) provides a framework for adversarial robustness, yet existing methods based on global Lipschitz continuity or strong duality often yield loose upper bounds or require prohibitive computation. We address these limitations with a primal approach and adopt a notion of exact Lipschitz certificates to tighten this upper bound of WDRO. For ReLU networks, we leverage the piecewise-affine structure on activation cells to obtain an exact tractable characterization of the corresponding WDRO problem. We further extend our analysis to modern architectures with smooth activations (e.g., GELU, SiLU), such as Transformers. Additionally, we propose novel Wasserstein Distributional Attacks (WDA, WDA++) that construct candidates for the worst-case distribution. Compared to existing attacks that are restricted to point-wise perturbations, our methods offer greater flexibility in the number and location of attack points. Extensive evaluations demonstrate that our proposed framework achieves competitive robust accuracy against state-of-the-art baselines while offering tighter certificates than existing methods. Our code is available at https://github.com/OLab-Repo/WDA.", "date": "2025-10-11", "categories": []}
{"id": "2602.02626", "title": "Learning Better Certified Models from Empirically-Robust Teachers", "abstract": "Adversarial training attains strong empirical robustness to specific adversarial attacks by training on concrete adversarial perturbations, but it produces neural networks that are not amenable to strong robustness certificates through neural network verification. On the other hand, earlier certified training schemes directly train on bounds from network relaxations to obtain models that are certifiably robust, but display sub-par standard performance. Recent work has shown that state-of-the-art trade-offs between certified robustness and standard performance can be obtained through a family of losses combining adversarial outputs and neural network bounds. Nevertheless, differently from empirical robustness, verifiability still comes at a significant cost in standard performance. In this work, we propose to leverage empirically-robust teachers to improve the performance of certifiably-robust models through knowledge distillation. Using a versatile feature-space distillation objective, we show that distillation from adversarially-trained teachers consistently improves on the state-of-the-art in certified training for ReLU networks across a series of robust computer vision benchmarks.", "date": "2026-02-02", "categories": []}
{"id": "2602.03316", "title": "Invisible Clean-Label Backdoor Attacks for Generative Data Augmentation", "abstract": "With the rapid advancement of image generative models, generative data augmentation has become an effective way to enrich training images, especially when only small-scale datasets are available. At the same time, in practical applications, generative data augmentation can be vulnerable to clean-label backdoor attacks, which aim to bypass human inspection. However, based on theoretical analysis and preliminary experiments, we observe that directly applying existing pixel-level clean-label backdoor attack methods (e.g., COMBAT) to generated images results in low attack success rates. This motivates us to move beyond pixel-level triggers and focus instead on the latent feature level. To this end, we propose InvLBA, an invisible clean-label backdoor attack method for generative data augmentation by latent perturbation. We theoretically prove that the generalization of the clean accuracy and attack success rates of InvLBA can be guaranteed. Experiments on multiple datasets show that our method improves the attack success rate by 46.43% on average, with almost no reduction in clean accuracy and high robustness against SOTA defense methods.", "date": "2026-02-03", "categories": []}
{"id": "2602.03333", "title": "PWAVEP: Purifying Imperceptible Adversarial Perturbations in 3D Point Clouds via Spectral Graph Wavelets", "abstract": "Recent progress in adversarial attacks on 3D point clouds, particularly in achieving spatial imperceptibility and high attack performance, presents significant challenges for defenders. Current defensive approaches remain cumbersome, often requiring invasive model modifications, expensive training procedures or auxiliary data access. To address these threats, in this paper, we propose a plug-and-play and non-invasive defense mechanism in the spectral domain, grounded in a theoretical and empirical analysis of the relationship between imperceptible perturbations and high-frequency spectral components. Building upon these insights, we introduce a novel purification framework, termed PWAVEP, which begins by computing a spectral graph wavelet domain saliency score and local sparsity score for each point. Guided by these values, PWAVEP adopts a hierarchical strategy, it eliminates the most salient points, which are identified as hardly recoverable adversarial outliers. Simultaneously, it applies a spectral filtering process to a broader set of moderately salient points. This process leverages a graph wavelet transform to attenuate high-frequency coefficients associated with the targeted points, thereby effectively suppressing adversarial noise. Extensive evaluations demonstrate that the proposed PWAVEP achieves superior accuracy and robustness compared to existing approaches, advancing the state-of-the-art in 3D point cloud purification. Code and datasets are available at https://github.com/a772316182/pwavep", "date": "2026-02-03", "categories": []}
{"id": "2602.03415", "title": "Most Convolutional Networks Suffer from Small Adversarial Perturbations", "abstract": "The existence of adversarial examples is relatively understood for random fully connected neural networks, but much less so for convolutional neural networks (CNNs). The recent work [Daniely, 2025] establishes that adversarial examples can be found in CNNs, in some non-optimal distance from the input. We extend over this work and prove that adversarial examples in random CNNs with input dimension $d$ can be found already in $\\ell_2$-distance of order $\\lVert x \\rVert /\\sqrt{d}$ from the input $x$, which is essentially the nearest possible. We also show that such adversarial small perturbations can be found using a single step of gradient descent. To derive our results we use Fourier decomposition to efficiently bound the singular values of a random linear convolutional operator, which is the main ingredient of a CNN layer. This bound might be of independent interest.", "date": "2026-02-03", "categories": []}
{"id": "2502.04204", "title": "Short-length Adversarial Training Helps LLMs Defend Long-length Jailbreak Attacks: Theoretical and Empirical Evidence", "abstract": "Jailbreak attacks against large language models (LLMs) aim to induce harmful behaviors in LLMs through carefully crafted adversarial prompts. To mitigate attacks, one way is to perform adversarial training (AT)-based alignment, i.e., training LLMs on some of the most adversarial prompts to help them learn how to behave safely under attacks. During AT, the length of adversarial prompts plays a critical role in the robustness of aligned LLMs. While long-length adversarial prompts during AT might lead to strong LLM robustness, their synthesis however is very resource-consuming, which may limit the application of LLM AT. This paper focuses on adversarial suffix jailbreak attacks and unveils that to defend against a jailbreak attack with an adversarial suffix of length $\\Theta(M)$, it is enough to align LLMs on prompts with adversarial suffixes of length $\\Theta(\\sqrt{M})$. Theoretically, we analyze the adversarial in-context learning of linear transformers on linear regression tasks and prove a robust generalization bound for trained transformers. The bound depends on the term $\\Theta(\\sqrt{M_{\\text{test}}}/M_{\\text{train}})$, where $M_{\\text{train}}$ and $M_{\\text{test}}$ are the numbers of adversarially perturbed in-context samples during training and testing. Empirically, we conduct AT on popular open-source LLMs and evaluate their robustness against jailbreak attacks of different adversarial suffix lengths. Results confirm a positive correlation between the attack success rate and the ratio of the square root of the adversarial suffix length during jailbreaking to the length during AT. Our findings show that it is practical to defend against \"long-length\" jailbreak attacks via efficient \"short-length\" AT. The code is available at https://github.com/fshp971/adv-icl.", "date": "2025-02-06", "categories": []}
{"id": "2506.12706", "title": "NAP-Tuning: Neural Augmented Prompt Tuning for Adversarially Robust Vision-Language Models", "abstract": "Vision-Language Models (VLMs) such as CLIP have demonstrated remarkable capabilities in understanding relationships between visual and textual data through joint embedding spaces. Despite their effectiveness, these models remain vulnerable to adversarial attacks, particularly in the image modality, posing significant security concerns. Building upon our previous work on Adversarial Prompt Tuning (AdvPT), which introduced learnable text prompts to enhance adversarial robustness in VLMs without extensive parameter training, we present a significant extension by introducing the Neural Augmentor framework for Multi-modal Adversarial Prompt Tuning (NAP-Tuning).Our key innovations include: (1) extending AdvPT from text-only to multi-modal prompting across both text and visual modalities, (2) expanding from single-layer to multi-layer prompt architectures, and (3) proposing a novel architecture-level redesign through our Neural Augmentor approach, which implements feature purification to directly address the distortions introduced by adversarial attacks in feature space. Our NAP-Tuning approach incorporates token refiners that learn to reconstruct purified features through residual connections, allowing for modality-specific and layer-specific feature correction.Comprehensive experiments demonstrate that NAP-Tuning significantly outperforms existing methods across various datasets and attack types. Notably, our approach shows significant improvements over the strongest baselines under the challenging AutoAttack benchmark, outperforming them by 33.5% on ViT-B16 and 33.0% on ViT-B32 architectures while maintaining competitive clean accuracy.", "date": "2025-06-15", "categories": []}
{"id": "2509.11250", "title": "Environmental Injection Attacks against GUI Agents in Realistic Dynamic Environments", "abstract": "Graphical User Interface (GUI) agents are increasingly deployed to interact with online web services, yet their exposure to open-world content renders them vulnerable to Environmental Injection Attacks (EIAs). In these attacks, an attacker can inject crafted triggers into website to manipulate the behavior of GUI agents used by other users. In this paper, we find that most existing EIA studies fall short of realism. In particular, they fail to capture the dynamic nature of real-world web content, often assuming that a trigger's on-screen position and surrounding visual context remain largely consistent between training and testing. To better reflect practice, we introduce a realistic dynamic-environment threat model in which the attacker is a regular user and the trigger is embedded within a dynamically changing environment. Under this threat model, existing approaches largely fail, suggesting that their effectiveness in exposing GUI agent vulnerabilities has been substantially overestimated.\n  To expose the hidden vulnerabilities of existing GUI agents effectively, we propose Chameleon, an attack framework with two key novelties designed for dynamic environments. (1) To synthesize more realistic training data, we introduce LLM-Driven Environment Simulation, which automatically generates diverse, high-fidelity webpage simulations that mimic the variability of real-world dynamic environments. (2) To optimize the trigger more effectively, we introduce Attention Black Hole, which converts attention weights into explicit supervisory signals. This mechanism encourages the agent to remain insensitive to irrelevant surrounding content, thereby improving robustness in dynamic environments. We evaluate Chameleon on six realistic websites and four representative LVLM-powered GUI agents, where it significantly outperforms existing methods.", "date": "2025-09-14", "categories": []}
{"id": "2511.01079", "title": "T-MLA: A targeted multiscale log-exponential attack framework for neural image compression", "abstract": "Neural image compression (NIC) has become the state-of-the-art for rate-distortion performance, yet its security vulnerabilities remain significantly less understood than those of classifiers. Existing adversarial attacks on NICs are often naive adaptations of pixel-space methods, overlooking the unique, structured nature of the compression pipeline. In this work, we propose a more advanced class of vulnerabilities by introducing T-MLA, the first targeted multiscale log-exponential attack framework. We introduce adversarial perturbations in the wavelet domain that concentrate on less perceptually salient coefficients, improving the stealth of the attack. Extensive evaluation across multiple state-of-the-art NIC architectures on standard image compression benchmarks reveals a large drop in reconstruction quality while the perturbations remain visually imperceptible. On standard NIC benchmarks, T-MLA achieves targeted degradation of reconstruction quality while improving perturbation imperceptibility (higher PSNR/VIF of the perturbed inputs) compared to PGD-style baselines at comparable attack success, as summarized in our main results. Our findings reveal a critical security flaw at the core of generative and content delivery pipelines.", "date": "2025-11-02", "categories": []}
{"id": "2511.16209", "title": "PSM: Prompt Sensitivity Minimization via LLM-Guided Black-Box Optimization", "abstract": "System prompts are critical for guiding the behavior of Large Language Models (LLMs), yet they often contain proprietary logic or sensitive information, making them a prime target for extraction attacks. Adversarial queries can successfully elicit these hidden instructions, posing significant security and privacy risks. Existing defense mechanisms frequently rely on heuristics, incur substantial computational overhead, or are inapplicable to models accessed via black-box APIs. This paper introduces a novel framework for hardening system prompts through shield appending, a lightweight approach that adds a protective textual layer to the original prompt. Our core contribution is the formalization of prompt hardening as a utility-constrained optimization problem. We leverage an LLM-as-optimizer to search the space of possible SHIELDs, seeking to minimize a leakage metric derived from a suite of adversarial attacks, while simultaneously preserving task utility above a specified threshold, measured by semantic fidelity to baseline outputs. This black-box, optimization-driven methodology is lightweight and practical, requiring only API access to the target and optimizer LLMs. We demonstrate empirically that our optimized SHIELDs significantly reduce prompt leakage against a comprehensive set of extraction attacks, outperforming established baseline defenses without compromising the model's intended functionality. Our work presents a paradigm for developing robust, utility-aware defenses in the escalating landscape of LLM security. The code is made public on the following link: https://github.com/psm-defense/psm", "date": "2025-11-20", "categories": []}
{"id": "2601.21900", "title": "TraceRouter: Robust Safety for Large Foundation Models via Path-Level Intervention", "abstract": "Despite their capabilities, large foundation models (LFMs) remain susceptible to adversarial manipulation. Current defenses predominantly rely on the \"locality hypothesis\", suppressing isolated neurons or features. However, harmful semantics act as distributed, cross-layer circuits, rendering such localized interventions brittle and detrimental to utility. To bridge this gap, we propose \\textbf{TraceRouter}, a path-level framework that traces and disconnects the causal propagation circuits of illicit semantics. TraceRouter operates in three stages: (1) it pinpoints a sensitive onset layer by analyzing attention divergence; (2) it leverages sparse autoencoders (SAEs) and differential activation analysis to disentangle and isolate malicious features; and (3) it maps these features to downstream causal pathways via feature influence scores (FIS) derived from zero-out interventions. By selectively suppressing these causal chains, TraceRouter physically severs the flow of harmful information while leaving orthogonal computation routes intact. Extensive experiments demonstrate that TraceRouter significantly outperforms state-of-the-art baselines, achieving a superior trade-off between adversarial robustness and general utility. Our code will be publicly released. WARNING: This paper contains unsafe model responses.", "date": "2026-01-29", "categories": []}
{"id": "2602.00420", "title": "Text is All You Need for Vision-Language Model Jailbreaking", "abstract": "Large Vision-Language Models (LVLMs) are increasingly equipped with robust safety safeguards to prevent responses to harmful or disallowed prompts. However, these defenses often focus on analyzing explicit textual inputs or relevant visual scenes. In this work, we introduce Text-DJ, a novel jailbreak attack that bypasses these safeguards by exploiting the model's Optical Character Recognition (OCR) capability. Our methodology consists of three stages. First, we decompose a single harmful query into multiple and semantically related but more benign sub-queries. Second, we pick a set of distraction queries that are maximally irrelevant to the harmful query. Third, we present all decomposed sub-queries and distraction queries to the LVLM simultaneously as a grid of images, with the position of the sub-queries being middle within the grid. We demonstrate that this method successfully circumvents the safety alignment of state-of-the-art LVLMs. We argue this attack succeeds by (1) converting text-based prompts into images, bypassing standard text-based filters, and (2) inducing distractions, where the model's safety protocols fail to link the scattered sub-queries within a high number of irrelevant queries. Overall, our findings expose a critical vulnerability in LVLMs' OCR capabilities that are not robust to dispersed, multi-image adversarial inputs, highlighting the need for defenses for fragmented multimodal inputs.", "date": "2026-01-31", "categories": []}
{"id": "2602.00822", "title": "Safety-Efficacy Trade Off: Robustness against Data-Poisoning", "abstract": "Backdoor and data poisoning attacks can achieve high attack success while evading existing spectral and optimisation based defences. We show that this behaviour is not incidental, but arises from a fundamental geometric mechanism in input space. Using kernel ridge regression as an exact model of wide neural networks, we prove that clustered dirty label poisons induce a rank one spike in the input Hessian whose magnitude scales quadratically with attack efficacy. Crucially, for nonlinear kernels we identify a near clone regime in which poison efficacy remains order one while the induced input curvature vanishes, making the attack provably spectrally undetectable. We further show that input gradient regularisation contracts poison aligned Fisher and Hessian eigenmodes under gradient flow, yielding an explicit and unavoidable safety efficacy trade off by reducing data fitting capacity. For exponential kernels, this defence admits a precise interpretation as an anisotropic high pass filter that increases the effective length scale and suppresses near clone poisons. Extensive experiments on linear models and deep convolutional networks across MNIST and CIFAR 10 and CIFAR 100 validate the theory, demonstrating consistent lags between attack success and spectral visibility, and showing that regularisation and data augmentation jointly suppress poisoning. Our results establish when backdoors are inherently invisible, and provide the first end to end characterisation of poisoning, detectability, and defence through input space curvature.", "date": "2026-01-31", "categories": []}
{"id": "2602.00857", "title": "Unifying Adversarial Robustness and Training Across Text Scoring Models", "abstract": "Research on adversarial robustness in language models is currently fragmented across applications and attacks, obscuring shared vulnerabilities. In this work, we propose unifying the study of adversarial robustness in text scoring models spanning dense retrievers, rerankers, and reward models. This motivates adapting both attacks and adversarial training methods across model roles. Unlike open-ended generation, text scoring failures are directly testable: an attack succeeds when an irrelevant or rejected text outscores a relevant or chosen one. Using this principled lens of text scoring, we demonstrate that current adversarial training formulations for language models are often short-sighted, failing to effectively generalize across attacks. To address this, we introduce multiple adversarial training methods for text scoring models and show that combining complementary training methods can yield strong robustness while also improving task effectiveness. We also highlight the practical value of our approach for RLHF, showing that our adversarially trained reward models mitigate reward hacking and support the training of better-aligned LLMs. We provide our code and models for further study.", "date": "2026-01-31", "categories": []}
{"id": "2602.01025", "title": "Toward Universal and Transferable Jailbreak Attacks on Vision-Language Models", "abstract": "Vision-language models (VLMs) extend large language models (LLMs) with vision encoders, enabling text generation conditioned on both images and text. However, this multimodal integration expands the attack surface by exposing the model to image-based jailbreaks crafted to induce harmful responses. Existing gradient-based jailbreak methods transfer poorly, as adversarial patterns overfit to a single white-box surrogate and fail to generalise to black-box models. In this work, we propose Universal and transferable jailbreak (UltraBreak), a framework that constrains adversarial patterns through transformations and regularisation in the vision space, while relaxing textual targets through semantic-based objectives. By defining its loss in the textual embedding space of the target LLM, UltraBreak discovers universal adversarial patterns that generalise across diverse jailbreak objectives. This combination of vision-level regularisation and semantically guided textual supervision mitigates surrogate overfitting and enables strong transferability across both models and attack targets. Extensive experiments show that UltraBreak consistently outperforms prior jailbreak methods. Further analysis reveals why earlier approaches fail to transfer, highlighting that smoothing the loss landscape via semantic objectives is crucial for enabling universal and transferable jailbreaks. The code is publicly available in our \\href{https://github.com/kaiyuanCui/UltraBreak}{GitHub repository}.", "date": "2026-02-01", "categories": []}
{"id": "2602.01113", "title": "Single-Edge Node Injection Threats to GNN-Based Security Monitoring in Industrial Graph Systems", "abstract": "Graph neural networks (GNNs) are increasingly adopted in industrial graph-based monitoring systems (e.g., Industrial internet of things (IIoT) device graphs, power-grid topology models, and manufacturing communication networks) to support anomaly detection, state estimation, and asset classification. In such settings, an adversary that compromises a small number of edge devices may inject counterfeit nodes (e.g., rogue sensors, virtualized endpoints, or spoofed substations) to bias downstream decisions while evading topology- and homophily-based sanitization. This paper formulates deployment-oriented node-injection attacks under constrained resources and proposes the \\emph{Single-Edge Graph Injection Attack} (SEGIA), in which each injected node attaches to the operational graph through a single edge. SEGIA integrates a pruned SGC surrogate, multi-hop neighborhood sampling, and reverse graph convolution-based feature synthesis with a similarity-regularized objective to preserve local homophily and survive edge pruning. Theoretical analysis and extensive evaluations across datasets and defenses show at least $25\\%$ higher attack success than representative baselines under substantially smaller edge budgets. These results indicate a system-level risk in industrial GNN deployments and motivate lightweight admission validation and neighborhood-consistency monitoring.", "date": "2026-02-01", "categories": []}
{"id": "2602.01369", "title": "Exposing and Defending the Achilles' Heel of Video Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) has demonstrated strong performance in video understanding tasks, yet its adversarial robustness remains underexplored. Existing attack methods often treat MoE as a unified architecture, overlooking the independent and collaborative weaknesses of key components such as routers and expert modules. To fill this gap, we propose Temporal Lipschitz-Guided Attacks (TLGA) to thoroughly investigate component-level vulnerabilities in video MoE models. We first design attacks on the router, revealing its independent weaknesses. Building on this, we introduce Joint Temporal Lipschitz-Guided Attacks (J-TLGA), which collaboratively perturb both routers and experts. This joint attack significantly amplifies adversarial effects and exposes the Achilles' Heel (collaborative weaknesses) of the MoE architecture. Based on these insights, we further propose Joint Temporal Lipschitz Adversarial Training (J-TLAT). J-TLAT performs joint training to further defend against collaborative weaknesses, enhancing component-wise robustness. Our framework is plug-and-play and reduces inference cost by more than 60% compared with dense models. It consistently enhances adversarial robustness across diverse datasets and architectures, effectively mitigating both the independent and collaborative weaknesses of MoE.", "date": "2026-02-01", "categories": []}
{"id": "2602.01574", "title": "SGHA-Attack: Semantic-Guided Hierarchical Alignment for Transferable Targeted Attacks on Vision-Language Models", "abstract": "Large vision-language models (VLMs) are vulnerable to transfer-based adversarial perturbations, enabling attackers to optimize on surrogate models and manipulate black-box VLM outputs. Prior targeted transfer attacks often overfit surrogate-specific embedding space by relying on a single reference and emphasizing final-layer alignment, which underutilizes intermediate semantics and degrades transfer across heterogeneous VLMs. To address this, we propose SGHA-Attack, a Semantic-Guided Hierarchical Alignment framework that adopts multiple target references and enforces intermediate-layer consistency. Concretely, we generate a visually grounded reference pool by sampling a frozen text-to-image model conditioned on the target prompt, and then carefully select the Top-K most semantically relevant anchors under the surrogate to form a weighted mixture for stable optimization guidance. Building on these anchors, SGHA-Attack injects target semantics throughout the feature hierarchy by aligning intermediate visual representations at both global and spatial granularities across multiple depths, and by synchronizing intermediate visual and textual features in a shared latent subspace to provide early cross-modal supervision before the final projection. Extensive experiments on open-source and commercial black-box VLMs show that SGHA-Attack achieves stronger targeted transferability than prior methods and remains robust under preprocessing and purification defenses.", "date": "2026-02-02", "categories": []}
{"id": "2602.01582", "title": "On the Fragility of AI-Based Channel Decoders under Small Channel Perturbations", "abstract": "Recent advances in deep learning have led to AI-based error correction decoders that report empirical performance improvements over traditional belief-propagation (BP) decoding on AWGN channels. While such gains are promising, a fundamental question remains: where do these improvements come from, and what cost is paid to achieve them? In this work, we study this question through the lens of robustness to distributional shifts at the channel output. We evaluate both input-dependent adversarial perturbations (FGM and projected gradient methods under $\\ell_2$ constraints) and universal adversarial perturbations that apply a single norm-bounded shift to all received vectors. Our results show that recent AI decoders, including ECCT and CrossMPT, could suffer significant performance degradation under such perturbations, despite superior nominal performance under i.i.d. AWGN. Moreover, adversarial perturbations transfer relatively strongly between AI decoders but weakly to BP-based decoders, and universal perturbations are substantially more harmful than random perturbations of equal norm. These numerical findings suggest a potential robustness cost and higher sensitivity to channel distribution underlying recent AI decoding gains.", "date": "2026-02-02", "categories": []}
{"id": "2602.01765", "title": "Backdoor Sentinel: Detecting and Detoxifying Backdoors in Diffusion Models via Temporal Noise Consistency", "abstract": "Diffusion models have been widely deployed in AIGC services; however, their reliance on opaque training data and procedures exposes a broad attack surface for backdoor injection. In practical auditing scenarios, due to the protection of intellectual property and commercial confidentiality, auditors are typically unable to access model parameters, rendering existing white-box or query-intensive detection methods impractical. More importantly, even after the backdoor is detected, existing detoxification approaches are often trapped in a dilemma between detoxification effectiveness and generation quality.\n  In this work, we identify a previously unreported phenomenon called temporal noise unconsistency, where the noise predictions between adjacent diffusion timesteps is disrupted in specific temporal segments when the input is triggered, while remaining stable under clean inputs. Leveraging this finding, we propose Temporal Noise Consistency Defense (TNC-Defense), a unified framework for backdoor detection and detoxification. The framework first uses the adjacent timestep noise consistency to design a gray-box detection module, for identifying and locating anomalous diffusion timesteps. Furthermore, the framework uses the identified anomalous timesteps to construct a trigger-agnostic, timestep-aware detoxification module, which directly corrects the backdoor generation path. This effectively suppresses backdoor behavior while significantly reducing detoxification costs.\n  We evaluate the proposed method under five representative backdoor attack scenarios and compare it with state-of-the-art defenses. The results show that TNC-Defense improves the average detection accuracy by $11\\%$ with negligible additional overhead, and invalidates an average of $98.5\\%$ of triggered samples with only a mild degradation in generation quality.", "date": "2026-02-02", "categories": []}
{"id": "2602.01795", "title": "RedVisor: Reasoning-Aware Prompt Injection Defense via Zero-Copy KV Cache Reuse", "abstract": "Large Language Models (LLMs) are increasingly vulnerable to Prompt Injection (PI) attacks, where adversarial instructions hidden within retrieved contexts hijack the model's execution flow. Current defenses typically face a critical trade-off: prevention-based fine-tuning often degrades general utility via the \"alignment tax\", while detection-based filtering incurs prohibitive latency and memory costs. To bridge this gap, we propose RedVisor, a unified framework that synthesizes the explainability of detection systems with the seamless integration of prevention strategies. To the best of our knowledge, RedVisor is the first approach to leverage fine-grained reasoning paths to simultaneously detect attacks and guide the model's safe response. We implement this via a lightweight, removable adapter positioned atop the frozen backbone. This adapter serves a dual function: it first generates an explainable analysis that precisely localizes the injection and articulates the threat, which then explicitly conditions the model to reject the malicious command. Uniquely, the adapter is active only during this reasoning phase and is effectively muted during the subsequent response generation. This architecture yields two distinct advantages: (1) it mathematically preserves the backbone's original utility on benign inputs; and (2) it enables a novel KV Cache Reuse strategy, eliminating the redundant prefill computation inherent to decoupled pipelines. We further pioneer the integration of this defense into the vLLM serving engine with custom kernels. Experiments demonstrate that RedVisor outperforms state-of-the-art defenses in detection accuracy and throughput while incurring negligible utility loss.", "date": "2026-02-02", "categories": []}
{"id": "2509.15437", "title": "Impact of Phonetics on Speaker Identity in Adversarial Voice Attack", "abstract": "Adversarial perturbations in speech pose a serious threat to automatic speech recognition (ASR) and speaker verification by introducing subtle waveform modifications that remain imperceptible to humans but can significantly alter system outputs. While targeted attacks on end-to-end ASR models have been widely studied, the phonetic basis of these perturbations and their effect on speaker identity remain underexplored. In this work, we analyze adversarial audio at the phonetic level and show that perturbations exploit systematic confusions such as vowel centralization and consonant substitutions. These distortions not only mislead transcription but also degrade phonetic cues critical for speaker verification, leading to identity drift. Using DeepSpeech as our ASR target, we generate targeted adversarial examples and evaluate their impact on speaker embeddings across genuine and impostor samples. Results across 16 phonetically diverse target phrases demonstrate that adversarial audio induces both transcription errors and identity drift, highlighting the need for phonetic-aware defenses to ensure the robustness of ASR and speaker recognition systems.", "date": "2025-09-18", "categories": []}
{"id": "2510.04347", "title": "Unmasking Backdoors: An Explainable Defense via Gradient-Attention Anomaly Scoring for Pre-trained Language Models", "abstract": "Pre-trained language models have achieved remarkable success across a wide range of natural language processing (NLP) tasks, particularly when fine-tuned on large, domain-relevant datasets. However, they remain vulnerable to backdoor attacks, where adversaries embed malicious behaviors using trigger patterns in the training data. These triggers remain dormant during normal usage, but, when activated, can cause targeted misclassifications. In this work, we investigate the internal behavior of backdoored pre-trained encoder-based language models, focusing on the consistent shift in attention and gradient attribution when processing poisoned inputs; where the trigger token dominates both attention and gradient signals, overriding the surrounding context. We propose an inference-time defense that constructs anomaly scores by combining token-level attention and gradient information. Extensive experiments on text classification tasks across diverse backdoor attack scenarios demonstrate that our method significantly reduces attack success rates compared to existing baselines. Furthermore, we provide an interpretability-driven analysis of the scoring mechanism, shedding light on trigger localization and the robustness of the proposed defense.", "date": "2025-10-05", "categories": []}
{"id": "2511.07888", "title": "Breaking the Adversarial Robustness-Performance Trade-off in Text Classification via Manifold Purification", "abstract": "A persistent challenge in text classification (TC) is that enhancing model robustness against adversarial attacks typically degrades performance on clean data. We argue that this challenge can be resolved by modeling the distribution of clean samples in the encoder embedding manifold. To this end, we propose the Manifold-Correcting Causal Flow (MC^2F), a two-module system that operates directly on sentence embeddings. A Stratified Riemannian Continuous Normalizing Flow (SR-CNF) learns the density of the clean data manifold. It identifies out-of-distribution embeddings, which are then corrected by a Geodesic Purification Solver. This solver projects adversarial points back onto the learned manifold via the shortest path, restoring a clean, semantically coherent representation. We conducted extensive evaluations on text classification (TC) across three datasets and multiple adversarial attacks. The results demonstrate that our method, MC^2F, not only establishes a new state-of-the-art in adversarial robustness but also fully preserves performance on clean data, even yielding modest gains in accuracy.", "date": "2025-11-11", "categories": []}
{"id": "2512.20712", "title": "Real-World Adversarial Attacks on RF-Based Drone Detectors", "abstract": "Radio frequency (RF) based systems are increasingly used to detect drones by analyzing their RF signal patterns, converting them into spectrogram images which are processed by object detection models. Existing RF attacks against image based models alter digital features, making over-the-air (OTA) implementation difficult due to the challenge of converting digital perturbations to transmittable waveforms that may introduce synchronization errors and interference, and encounter hardware limitations. We present the first physical attack on RF image based drone detectors, optimizing class-specific universal complex baseband (I/Q) perturbation waveforms that are transmitted alongside legitimate communications. We evaluated the attack using RF recordings and OTA experiments with four types of drones. Our results show that modest, structured I/Q perturbations are compatible with standard RF chains and reliably reduce target drone detection while preserving detection of legitimate drones.", "date": "2025-12-23", "categories": []}
{"id": "2601.22182", "title": "ShellForge: Adversarial Co-Evolution of Webshell Generation and Multi-View Detection for Robust Webshell Defense", "abstract": "Webshells remain a primary foothold for attackers to compromise servers, particularly within PHP ecosystems. However, existing detection mechanisms often struggle to keep pace with rapid variant evolution and sophisticated obfuscation techniques that camouflage malicious intent. Furthermore, many current defenses suffer from high false-alarm rates when encountering benign administrative scripts that employ heavy obfuscation for intellectual property protection. To address these challenges, we present ShellForge, an adversarial co-evolution framework that couples automated webshell generation with multi-view detection to continuously harden defensive boundaries. The framework operates through an iterative co-training loop where a generator and a detector mutually reinforce each other via the exchange of hard samples. The generator is optimized through supervised fine-tuning and preference-based reinforcement learning to synthesize functional, highly evasive variants. Simultaneously, we develop a multi-view fusion detector that integrates semantic features from long-string compression, structural features from pruned abstract syntax trees, and global statistical indicators such as Shannon entropy. To minimize false positives, ShellForge utilizes a LLM-based transformation to create de-malicious samples--scripts that retain complex obfuscation patterns but lack harmful payloads--serving as high-quality hard negatives during training. Evaluations on the public FWOID benchmark demonstrate that ShellForge significantly enhances defensive robustness. Upon convergence, the detector maintains a 0.981 F1-score while the generator achieves a 0.939 evasion rate against commercial engines on VirusTotal.", "date": "2026-01-28", "categories": []}
{"id": "2601.22744", "title": "Beauty and the Beast: Imperceptible Perturbations Against Diffusion-Based Face Swapping via Directional Attribute Editing", "abstract": "Diffusion-based face swapping achieves state-of-the-art performance, yet it also exacerbates the potential harm of malicious face swapping to violate portraiture right or undermine personal reputation. This has spurred the development of proactive defense methods. However, existing approaches face a core trade-off: large perturbations distort facial structures, while small ones weaken protection effectiveness. To address these issues, we propose FaceDefense, an enhanced proactive defense framework against diffusion-based face swapping. Our method introduces a new diffusion loss to strengthen the defensive efficacy of adversarial examples, and employs a directional facial attribute editing to restore perturbation-induced distortions, thereby enhancing visual imperceptibility. A two-phase alternating optimization strategy is designed to generate final perturbed face images. Extensive experiments show that FaceDefense significantly outperforms existing methods in both imperceptibility and defense effectiveness, achieving a superior trade-off.", "date": "2026-01-30", "categories": []}
{"id": "2601.23102", "title": "Rethinking Transferable Adversarial Attacks on Point Clouds from a Compact Subspace Perspective", "abstract": "Transferable adversarial attacks on point clouds remain challenging, as existing methods often rely on model-specific gradients or heuristics that limit generalization to unseen architectures. In this paper, we rethink adversarial transferability from a compact subspace perspective and propose CoSA, a transferable attack framework that operates within a shared low-dimensional semantic space. Specifically, each point cloud is represented as a compact combination of class-specific prototypes that capture shared semantic structure, while adversarial perturbations are optimized within a low-rank subspace to induce coherent and architecture-agnostic variations. This design suppresses model-dependent noise and constrains perturbations to semantically meaningful directions, thereby improving cross-model transferability without relying on surrogate-specific artifacts. Extensive experiments on multiple datasets and network architectures demonstrate that CoSA consistently outperforms state-of-the-art transferable attacks, while maintaining competitive imperceptibility and robustness under common defense strategies. Codes will be made public upon paper acceptance.", "date": "2026-01-30", "categories": []}
{"id": "2508.11667", "title": "Guided Perturbation Sensitivity (GPS): Detecting Adversarial Text via Embedding Stability and Word Importance", "abstract": "Adversarial text attacks remain a persistent threat to transformer models, yet existing defenses are typically attack-specific or require costly model retraining, leaving a gap for attack-agnostic detection. We introduce Guided Perturbation Sensitivity (GPS), a detection framework that identifies adversarial examples by measuring how embedding representations change when important words are masked. GPS first ranks words using importance heuristics, then measures embedding sensitivity to masking top-k critical words, and processes the resulting patterns with a BiLSTM detector. Experiments show that adversarially perturbed words exhibit disproportionately high masking sensitivity compared to naturally important words. Across three datasets, three attack types, and two victim models, GPS achieves over 85% detection accuracy and demonstrates competitive performance compared to existing state-of-the-art methods, often at lower computational cost. Using Normalized Discounted Cumulative Gain (NDCG) to measure perturbation identification quality, we demonstrate that gradient-based ranking significantly outperforms attention, hybrid, and random selection approaches, with identification quality strongly correlating with detection performance for word-level attacks ($\\rho = 0.65$). GPS generalizes to unseen datasets, attacks, and models without retraining, providing a practical solution for adversarial text detection.", "date": "2025-08-06", "categories": []}
{"id": "2509.18546", "title": "SEGA: A Transferable Signed Ensemble Gaussian Black-Box Attack against No-Reference Image Quality Assessment Models", "abstract": "No-Reference Image Quality Assessment (NR-IQA) models play an important role in various real-world applications. Recently, adversarial attacks against NR-IQA models have attracted increasing attention, as they provide valuable insights for revealing model vulnerabilities and guiding robust system design. Some effective attacks have been proposed against NR-IQA models in white-box settings, where the attacker has full access to the target model. However, these attacks often suffer from poor transferability to unknown target models in more realistic black-box scenarios, where the target model is inaccessible. This work makes the first attempt to address the challenge of low transferability in attacking NR-IQA models by proposing a transferable Signed Ensemble Gaussian black-box Attack (SEGA). The main idea is to approximate the gradient of the target model by applying Gaussian smoothing to source models and ensembling their smoothed gradients. To ensure the imperceptibility of adversarial perturbations, SEGA further removes inappropriate perturbations using a specially designed perturbation filter mask. Experimental results on the CLIVE dataset demonstrate the superior transferability of SEGA, validating its effectiveness in enabling successful transfer-based black-box attacks against NR-IQA models.", "date": "2025-09-23", "categories": []}
{"id": "2510.02422", "title": "Dynamic Target Attack", "abstract": "Existing gradient-based jailbreak attacks typically optimize an adversarial suffix to induce a fixed affirmative response, e.g., ``Sure, here is...''. However, this fixed target usually resides in an extremely low-density region of a safety-aligned LLM's output distribution. Due to the substantial discrepancy between the fixed target and the output distribution, existing attacks require numerous iterations to optimize the adversarial prompt, which might still fail to induce the low-probability target response. To address this limitation, we propose Dynamic Target Attack (DTA), which leverages the target LLM's own responses as adaptive targets. In each optimization round, DTA samples multiple candidates from the output distribution conditioned on the current prompt, and selects the most harmful one as a temporary target for prompt optimization. Extensive experiments demonstrate that, under the white-box setting, DTA achieves over 87% average attack success rate (ASR) within 200 optimization iterations on recent safety-aligned LLMs, exceeding the state-of-the-art baselines by over 15% and reducing wall-clock time by 2-26x. Under the black-box setting, DTA employs a white-box LLM as a surrogate model for gradient-based optimization, achieving an average ASR of 77.5% against black-box models, exceeding prior transfer-based attacks by over 12%.", "date": "2025-10-02", "categories": []}
{"id": "2601.07118", "title": "Reward-Preserving Attacks For Robust Reinforcement Learning", "abstract": "Adversarial training in reinforcement learning (RL) is challenging because perturbations cascade through trajectories and compound over time, making fixed-strength attacks either overly destructive or too conservative. We propose reward-preserving attacks, which adapt adversarial strength so that an $\\alpha$ fraction of the nominal-to-worst-case return gap remains achievable at each state. In deep RL, perturbation magnitudes $\\eta$ are selected dynamically, using a learned critic $Q((s,a),\\eta)$ that estimates the expected return of $\\alpha$-reward-preserving rollouts. For intermediate values of $\\alpha$, this adaptive training yields policies that are robust across a wide range of perturbation magnitudes while preserving nominal performance, outperforming fixed-radius and uniformly sampled-radius adversarial training.", "date": "2026-01-12", "categories": []}
{"id": "2601.21066", "title": "BadDet+: Robust Backdoor Attacks for Object Detection", "abstract": "Backdoor attacks pose a severe threat to deep learning, yet their impact on object detection remains poorly understood compared to image classification. While attacks have been proposed, we identify critical weaknesses in existing detection-based methods, specifically their reliance on unrealistic assumptions and a lack of physical validation. To bridge this gap, we introduce BadDet+, a penalty-based framework that unifies Region Misclassification Attacks (RMA) and Object Disappearance Attacks (ODA). The core mechanism utilizes a log-barrier penalty to suppress true-class predictions for triggered inputs, resulting in (i) position and scale invariance, and (ii) enhanced physical robustness. On real-world benchmarks, BadDet+ achieves superior synthetic-to-physical transfer compared to existing RMA and ODA baselines while preserving clean performance. Theoretical analysis confirms the proposed penalty acts within a trigger-specific feature subspace, reliably inducing attacks without degrading standard inference. These results highlight significant vulnerabilities in object detection and the necessity for specialized defenses.", "date": "2026-01-28", "categories": []}
{"id": "2601.21220", "title": "LAMP: Learning Universal Adversarial Perturbations for Multi-Image Tasks via Pre-trained Models", "abstract": "Multimodal Large Language Models (MLLMs) have achieved remarkable performance across vision-language tasks. Recent advancements allow these models to process multiple images as inputs. However, the vulnerabilities of multi-image MLLMs remain unexplored. Existing adversarial attacks focus on single-image settings and often assume a white-box threat model, which is impractical in many real-world scenarios. This paper introduces LAMP, a black-box method for learning Universal Adversarial Perturbations (UAPs) targeting multi-image MLLMs. LAMP applies an attention-based constraint that prevents the model from effectively aggregating information across images. LAMP also introduces a novel cross-image contagious constraint that forces perturbed tokens to influence clean tokens, spreading adversarial effects without requiring all inputs to be modified. Additionally, an index-attention suppression loss enables a robust position-invariant attack. Experimental results show that LAMP outperforms SOTA baselines and achieves the highest attack success rates across multiple vision-language tasks and models.", "date": "2026-01-29", "categories": []}
{"id": "2601.21323", "title": "Adversarial Vulnerability Transcends Computational Paradigms: Feature Engineering Provides No Defense Against Neural Adversarial Transfer", "abstract": "Deep neural networks are vulnerable to adversarial examples--inputs with imperceptible perturbations causing misclassification. While adversarial transfer within neural networks is well-documented, whether classical ML pipelines using handcrafted features inherit this vulnerability when attacked via neural surrogates remains unexplored. Feature engineering creates information bottlenecks through gradient quantization and spatial binning, potentially filtering high-frequency adversarial signals. We evaluate this hypothesis through the first comprehensive study of adversarial transfer from DNNs to HOG-based classifiers. Using VGG16 as a surrogate, we generate FGSM and PGD adversarial examples and test transfer to four classical classifiers (KNN, Decision Tree, Linear SVM, Kernel SVM) and a shallow neural network across eight HOG configurations on CIFAR-10. Our results strongly refute the protective hypothesis: all classifiers suffer 16.6%-59.1% relative accuracy drops, comparable to neural-to-neural transfer. More surprisingly, we discover attack hierarchy reversal--contrary to patterns where iterative PGD dominates FGSM within neural networks, FGSM causes greater degradation than PGD in 100% of classical ML cases, suggesting iterative attacks overfit to surrogate-specific features that don't survive feature extraction. Block normalization provides partial but insufficient mitigation. These findings demonstrate that adversarial vulnerability is not an artifact of end-to-end differentiability but a fundamental property of image classification systems, with implications for security-critical deployments across computational paradigms.", "date": "2026-01-29", "categories": []}
{"id": "2508.08920", "title": "Exploring Cross-Stage Adversarial Transferability in Class-Incremental Continual Learning", "abstract": "Class-incremental continual learning addresses catastrophic forgetting by enabling classification models to preserve knowledge of previously learned classes while acquiring new ones. However, the vulnerability of the models against adversarial attacks during this process has not been investigated sufficiently. In this paper, we present the first exploration of vulnerability to stage-transferred attacks, i.e., an adversarial example generated using the model in an earlier stage is used to attack the model in a later stage. Our findings reveal that continual learning methods are highly susceptible to these attacks, raising a serious security issue. We explain this phenomenon through model similarity between stages and gradual robustness degradation. Additionally, we find that existing adversarial training-based defense methods are not sufficiently effective to stage-transferred attacks. Codes are available at https://github.com/mcml-official/CSAT.", "date": "2025-08-12", "categories": []}
{"id": "2601.20283", "title": "One Word is Enough: Minimal Adversarial Perturbations for Neural Text Ranking", "abstract": "Neural ranking models (NRMs) achieve strong retrieval effectiveness, yet prior work has shown they are vulnerable to adversarial perturbations. We revisit this robustness question with a minimal, query-aware attack that promotes a target document by inserting or substituting a single, semantically aligned word - the query center. We study heuristic and gradient-guided variants, including a white-box method that identifies influential insertion points. On TREC-DL 2019/2020 with BERT and monoT5 re-rankers, our single-word attacks achieve up to 91% success while modifying fewer than two tokens per document on average, achieving competitive rank and score boosts with far fewer edits under a comparable white-box setup to ensure fair evaluation against PRADA. We also introduce new diagnostic metrics to analyze attack sensitivity beyond aggregate success rates. Our analysis reveals a Goldilocks zone in which mid-ranked documents are most vulnerable. These findings demonstrate practical risks and motivate future defenses for robust neural ranking.", "date": "2026-01-28", "categories": []}
{"id": "2503.10809", "title": "MIP against Agent: Malicious Image Patches Hijacking Multimodal OS Agents", "abstract": "Recent advances in operating system (OS) agents have enabled vision-language models (VLMs) to directly control a user's computer. Unlike conventional VLMs that passively output text, OS agents autonomously perform computer-based tasks in response to a single user prompt. OS agents do so by capturing, parsing, and analysing screenshots and executing low-level actions via application programming interfaces (APIs), such as mouse clicks and keyboard inputs. This direct interaction with the OS significantly raises the stakes, as failures or manipulations can have immediate and tangible consequences. In this work, we uncover a novel attack vector against these OS agents: Malicious Image Patches (MIPs), adversarially perturbed screen regions that, when captured by an OS agent, induce it to perform harmful actions by exploiting specific APIs. For instance, a MIP can be embedded in a desktop wallpaper or shared on social media to cause an OS agent to exfiltrate sensitive user data. We show that MIPs generalise across user prompts and screen configurations, and that they can hijack multiple OS agents even during the execution of benign instructions. These findings expose critical security vulnerabilities in OS agents that have to be carefully addressed before their widespread deployment.", "date": "2025-03-13", "categories": []}
{"id": "2601.18845", "title": "Dynamic Mask-Based Backdoor Attack Against Vision AI Models: A Case Study on Mushroom Detection", "abstract": "Deep learning has revolutionized numerous tasks within the computer vision field, including image classification, image segmentation, and object detection. However, the increasing deployment of deep learning models has exposed them to various adversarial attacks, including backdoor attacks. This paper presents a novel dynamic mask-based backdoor attack method, specifically designed for object detection models. We exploit a dataset poisoning technique to embed a malicious trigger, rendering any models trained on this compromised dataset vulnerable to our backdoor attack. We particularly focus on a mushroom detection dataset to demonstrate the practical risks posed by such attacks on critical real-life domains. Our work also emphasizes the importance of creating a detailed backdoor attack scenario to illustrate the significant risks associated with the outsourcing practice. Our approach leverages SAM, a recent and powerful image segmentation AI model, to create masks for dynamic trigger placement, introducing a new and stealthy attack method. Through extensive experimentation, we show that our sophisticated attack scenario maintains high accuracy on clean data with the YOLOv7 object detection model while achieving high attack success rates on poisoned samples. Our approach surpasses traditional methods for backdoor injection, which are based on static and consistent patterns. Our findings underscore the urgent need for robust countermeasures to protect deep learning models from these evolving adversarial threats.", "date": "2026-01-26", "categories": []}
{"id": "2601.19726", "title": "RvB: Automating AI System Hardening via Iterative Red-Blue Games", "abstract": "The dual offensive and defensive utility of Large Language Models (LLMs) highlights a critical gap in AI security: the lack of unified frameworks for dynamic, iterative adversarial adaptation hardening. To bridge this gap, we propose the Red Team vs. Blue Team (RvB) framework, formulated as a training-free, sequential, imperfect-information game. In this process, the Red Team exposes vulnerabilities, driving the Blue Team to learning effective solutions without parameter updates. We validate our framework across two challenging domains: dynamic code hardening against CVEs and guardrail optimization against jailbreaks. Our empirical results show that this interaction compels the Blue Team to learn fundamental defensive principles, leading to robust remediations that are not merely overfitted to specific exploits. RvB achieves Defense Success Rates of 90\\% and 45\\% across the respective tasks while maintaining near 0\\% False Positive Rates, significantly surpassing baselines. This work establishes the iterative adversarial interaction framework as a practical paradigm that automates the continuous hardening of AI systems.", "date": "2026-01-27", "categories": []}
{"id": "2601.19867", "title": "Bandits in Flux: Adversarial Constraints in Dynamic Environments", "abstract": "We investigate the challenging problem of adversarial multi-armed bandits operating under time-varying constraints, a scenario motivated by numerous real-world applications. To address this complex setting, we propose a novel primal-dual algorithm that extends online mirror descent through the incorporation of suitable gradient estimators and effective constraint handling. We provide theoretical guarantees establishing sublinear dynamic regret and sublinear constraint violation for our proposed policy. Our algorithm achieves state-of-the-art performance in terms of both regret and constraint violation. Empirical evaluations demonstrate the superiority of our approach.", "date": "2026-01-27", "categories": []}
{"id": "2405.10360", "title": "Adversarial Robustness Guarantees for Quantum Classifiers", "abstract": "Despite their ever more widespread deployment throughout society, machine learning algorithms remain critically vulnerable to being spoofed by subtle adversarial tampering with their input data. The prospect of near-term quantum computers being capable of running {quantum machine learning} (QML) algorithms has therefore generated intense interest in their adversarial vulnerability. Here we show that quantum properties of QML algorithms can confer fundamental protections against such attacks, in certain scenarios guaranteeing robustness against classically-armed adversaries. We leverage tools from many-body physics to identify the quantum sources of this protection. Our results offer a theoretical underpinning of recent evidence which suggest quantum advantages in the search for adversarial robustness. In particular, we prove that quantum classifiers are: (i) protected against weak perturbations of data drawn from the trained distribution, (ii) protected against local attacks if they are insufficiently scrambling, and (iii) show evidence that they are protected against universal adversarial attacks if they are sufficiently chaotic. Our analytic results are supported by numerical evidence demonstrating the applicability of our theorems and the resulting robustness of a quantum classifier in practice. This line of inquiry constitutes a concrete pathway to advantage in QML, orthogonal to the usually sought improvements in model speed or accuracy.", "date": "2024-05-16", "categories": []}
{"id": "2601.14519", "title": "How Worst-Case Are Adversarial Attacks? Linking Adversarial and Perturbation Robustness", "abstract": "Adversarial attacks are widely used to identify model vulnerabilities; however, their validity as proxies for robustness to random perturbations remains debated. We ask whether an adversarial example provides a representative estimate of misprediction risk under stochastic perturbations of the same magnitude, or instead reflects an atypical worst-case event. To address this question, we introduce a probabilistic analysis that quantifies this risk with respect to directionally biased perturbation distributions, parameterized by a concentration factor $\\kappa$ that interpolates between isotropic noise and adversarial directions. Building on this, we study the limits of this connection by proposing an attack strategy designed to probe vulnerabilities in regimes that are statistically closer to uniform noise. Experiments on ImageNet and CIFAR-10 systematically benchmark multiple attacks, revealing when adversarial success meaningfully reflects robustness to perturbations and when it does not, thereby informing their use in safety-oriented robustness evaluation.", "date": "2026-01-20", "categories": []}
{"id": "2601.17280", "title": "On the Insecurity of Keystroke-Based AI Authorship Detection: Timing-Forgery Attacks Against Motor-Signal Verification", "abstract": "Recent proposals advocate using keystroke timing signals, specifically the coefficient of variation ($\\delta$) of inter-keystroke intervals, to distinguish human-composed text from AI-generated content. We demonstrate that this class of defenses is insecure against two practical attack classes: the copy-type attack, in which a human transcribes LLM-generated text producing authentic motor signals, and timing-forgery attacks, in which automated agents sample inter-keystroke intervals from empirical human distributions. Using 13,000 sessions from the SBU corpus and three timing-forgery variants (histogram sampling, statistical impersonation, and generative LSTM), we show all attacks achieve $\\ge$99.8% evasion rates against five classifiers. While detectors achieve AUC=1.000 against fully-automated injection, they classify $\\ge$99.8% of attack samples as human with mean confidence $\\ge$0.993. We formalize a non-identifiability result: when the detector observes only timing, the mutual information between features and content provenance is zero for copy-type attacks. Although composition and transcription produce statistically distinguishable motor patterns (Cohen's d=1.28), both yield $\\delta$ values 2-4x above detection thresholds, rendering the distinction security-irrelevant. These systems confirm a human operated the keyboard, but not whether that human originated the text. Securing provenance requires architectures that bind the writing process to semantic content.", "date": "2026-01-24", "categories": []}
{"id": "2601.17383", "title": "Physical Prompt Injection Attacks on Large Vision-Language Models", "abstract": "Large Vision-Language Models (LVLMs) are increasingly deployed in real-world intelligent systems for perception and reasoning in open physical environments. While LVLMs are known to be vulnerable to prompt injection attacks, existing methods either require access to input channels or depend on knowledge of user queries, assumptions that rarely hold in practical deployments. We propose the first Physical Prompt Injection Attack (PPIA), a black-box, query-agnostic attack that embeds malicious typographic instructions into physical objects perceivable by the LVLM. PPIA requires no access to the model, its inputs, or internal pipeline, and operates solely through visual observation. It combines offline selection of highly recognizable and semantically effective visual prompts with strategic environment-aware placement guided by spatiotemporal attention, ensuring that the injected prompts are both perceivable and influential on model behavior. We evaluate PPIA across 10 state-of-the-art LVLMs in both simulated and real-world settings on tasks including visual question answering, planning, and navigation, PPIA achieves attack success rates up to 98%, with strong robustness under varying physical conditions such as distance, viewpoint, and illumination. Our code is publicly available at https://github.com/2023cghacker/Physical-Prompt-Injection-Attack.", "date": "2026-01-24", "categories": []}
{"id": "2601.18058", "title": "Differentiable Architecture Search for Adversarially Robust Quantum Computer Vision", "abstract": "Current quantum neural networks suffer from extreme sensitivity to both adversarial perturbations and hardware noise, creating a significant barrier to real-world deployment. Existing robustness techniques typically sacrifice clean accuracy or require prohibitive computational resources. We propose a hybrid quantum-classical Differentiable Quantum Architecture Search (DQAS) framework that addresses these limitations by jointly optimizing circuit structure and robustness through gradient-based methods. Our approach enhances traditional DQAS with a lightweight Classical Noise Layer applied before quantum processing, enabling simultaneous optimization of gate selection and noise parameters. This design preserves the quantum circuit's integrity while introducing trainable perturbations that enhance robustness without compromising standard performance. Experimental validation on MNIST, FashionMNIST, and CIFAR datasets shows consistent improvements in both clean and adversarial accuracy compared to existing quantum architecture search methods. Under various attack scenarios, including Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Basic Iterative Method (BIM), and Momentum Iterative Method (MIM), and under realistic quantum noise conditions, our hybrid framework maintains superior performance. Testing on actual quantum hardware confirms the practical viability of discovered architectures. These results demonstrate that strategic classical preprocessing combined with differentiable quantum architecture optimization can significantly enhance quantum neural network robustness while maintaining computational efficiency.", "date": "2026-01-26", "categories": []}
{"id": "2601.18386", "title": "ARMOR: Agentic Reasoning for Methods Orchestration and Reparameterization for Robust Adversarial Attacks", "abstract": "Existing automated attack suites operate as static ensembles with fixed sequences, lacking strategic adaptation and semantic awareness. This paper introduces the Agentic Reasoning for Methods Orchestration and Reparameterization (ARMOR) framework to address these limitations. ARMOR orchestrates three canonical adversarial primitives, Carlini-Wagner (CW), Jacobian-based Saliency Map Attack (JSMA), and Spatially Transformed Attacks (STA) via Vision Language Models (VLM)-guided agents that collaboratively generate and synthesize perturbations through a shared ``Mixing Desk\". Large Language Models (LLMs) adaptively tune and reparameterize parallel attack agents in a real-time, closed-loop system that exploits image-specific semantic vulnerabilities. On standard benchmarks, ARMOR achieves improved cross-architecture transfer and reliably fools both settings, delivering a blended output for blind targets and selecting the best attack or blended attacks for white-box targets using a confidence-and-SSIM score.", "date": "2026-01-26", "categories": []}
{"id": "2505.16004", "title": "Evaluating Adversarial Robustness of Concept Representations in Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) are commonly used to interpret the internal activations of large language models (LLMs) by mapping them to human-interpretable concept representations. While existing evaluations of SAEs focus on metrics such as the reconstruction-sparsity tradeoff, human (auto-)interpretability, and feature disentanglement, they overlook a critical aspect: the robustness of concept representations to input perturbations. We argue that robustness must be a fundamental consideration for concept representations, reflecting the fidelity of concept labeling. To this end, we formulate robustness quantification as input-space optimization problems and develop a comprehensive evaluation framework featuring realistic scenarios in which adversarial perturbations are crafted to manipulate SAE representations. Empirically, we find that tiny adversarial input perturbations can effectively manipulate concept-based interpretations in most scenarios without notably affecting the base LLM's activations. Overall, our results suggest that SAE concept representations are fragile and without further denoising or postprocessing they might be ill-suited for applications in model monitoring and oversight.", "date": "2025-05-21", "categories": []}
{"id": "2511.11020", "title": "Data Poisoning Vulnerabilities Across Healthcare AI Architectures: A Security Threat Analysis", "abstract": "Healthcare AI systems face major vulnerabilities to data poisoning that current defenses and regulations cannot adequately address. We analyzed eight attack scenarios in four categories: architectural attacks on convolutional neural networks, large language models, and reinforcement learning agents; infrastructure attacks exploiting federated learning and medical documentation systems; critical resource allocation attacks affecting organ transplantation and crisis triage; and supply chain attacks targeting commercial foundation models. Our findings indicate that attackers with access to only 100-500 samples can compromise healthcare AI regardless of dataset size, often achieving over 60 percent success, with detection taking an estimated 6 to 12 months or sometimes not occurring at all. The distributed nature of healthcare infrastructure creates many entry points where insiders with routine access can launch attacks with limited technical skill. Privacy laws such as HIPAA and GDPR can unintentionally shield attackers by restricting the analyses needed for detection. Supply chain weaknesses allow a single compromised vendor to poison models across 50 to 200 institutions. The Medical Scribe Sybil scenario shows how coordinated fake patient visits can poison data through legitimate clinical workflows without requiring a system breach. Current regulations lack mandatory adversarial robustness testing, and federated learning can worsen risks by obscuring attribution. We recommend multilayer defenses including required adversarial testing, ensemble-based detection, privacy-preserving security mechanisms, and international coordination on AI security standards. We also question whether opaque black-box models are suitable for high-stakes clinical decisions, suggesting a shift toward interpretable systems with verifiable safety guarantees.", "date": "2025-11-14", "categories": []}
{"id": "2512.10492", "title": "UACER: An Uncertainty-Adaptive Critic Ensemble Framework for Robust Adversarial Reinforcement Learning", "abstract": "Robust adversarial reinforcement learning has emerged as an effective paradigm for training agents to handle uncertain disturbance in real environments, with critical applications in sequential decision-making domains such as autonomous driving and robotic control. Within this paradigm, agent training is typically formulated as a zero-sum Markov game between a protagonist and an adversary to enhance policy robustness. However, the trainable nature of the adversary inevitably induces non-stationarity in the learning dynamics, leading to exacerbated training instability and convergence difficulties, particularly in high-dimensional complex environments. In this paper, we propose a novel approach, Uncertainty-Adaptive Critic Ensemble for robust adversarial Reinforcement learning (UACER), which consists of two components: 1) Diversified critic ensemble: A diverse set of K critic networks is employed in parallel to stabilize Q-value estimation in robust adversarial reinforcement learning, reducing variance and enhancing robustness compared to conventional single-critic designs. 2) Time-varying Decay Uncertainty (TDU) mechanism: Moving beyond simple linear combinations, we propose a variance-derived Q-value aggregation strategy that explicitly incorporates epistemic uncertainty to adaptively regulate the exploration-exploitation trade-off while stabilizing the training process. Comprehensive experiments across several challenging MuJoCo control problems validate the superior effectiveness of UACER, outperforming state-of-the-art methods in terms of overall performance, stability, and efficiency.", "date": "2025-12-11", "categories": []}
{"id": "2601.16231", "title": "SoundBreak: A Systematic Study of Audio-Only Adversarial Attacks on Trimodal Models", "abstract": "Multimodal foundation models that integrate audio, vision, and language achieve strong performance on reasoning and generation tasks, yet their robustness to adversarial manipulation remains poorly understood. We study a realistic and underexplored threat model: untargeted, audio-only adversarial attacks on trimodal audio-video-language models. We analyze six complementary attack objectives that target different stages of multimodal processing, including audio encoder representations, cross-modal attention, hidden states, and output likelihoods. Across three state-of-the-art models and multiple benchmarks, we show that audio-only perturbations can induce severe multimodal failures, achieving up to 96% attack success rate. We further show that attacks can be successful at low perceptual distortions (LPIPS <= 0.08, SI-SNR >= 0) and benefit more from extended optimization than increased data scale. Transferability across models and encoders remains limited, while speech recognition systems such as Whisper primarily respond to perturbation magnitude, achieving >97% attack success under severe distortion. These results expose a previously overlooked single-modality attack surface in multimodal systems and motivate defenses that enforce cross-modal consistency.", "date": "2026-01-20", "categories": []}
{"id": "2601.16464", "title": "On the Effects of Adversarial Perturbations on Distribution Robustness", "abstract": "Adversarial robustness refers to a model's ability to resist perturbation of inputs, while distribution robustness evaluates the performance of the model under data shifts. Although both aim to ensure reliable performance, prior work has revealed a tradeoff in distribution and adversarial robustness. Specifically, adversarial training might increase reliance on spurious features, which can harm distribution robustness, especially the performance on some underrepresented subgroups. We present a theoretical analysis of adversarial and distribution robustness that provides a tractable surrogate for per-step adversarial training by studying models trained on perturbed data. In addition to the tradeoff, our work further identified a nuanced phenomenon that $\\ell_\\infty$ perturbations on data with moderate bias can yield an increase in distribution robustness. Moreover, the gain in distribution robustness remains on highly skewed data when simplicity bias induces reliance on the core feature, characterized as greater feature separability. Our theoretical analysis extends the understanding of the tradeoff by highlighting the interplay of the tradeoff and the feature separability. Despite the tradeoff that persists in many cases, overlooking the role of feature separability may lead to misleading conclusions about robustness.", "date": "2026-01-23", "categories": []}
{"id": "2601.16473", "title": "DeMark: A Query-Free Black-Box Attack on Deepfake Watermarking Defenses", "abstract": "The rapid proliferation of realistic deepfakes has raised urgent concerns over their misuse, motivating the use of defensive watermarks in synthetic images for reliable detection and provenance tracking. However, this defense paradigm assumes such watermarks are inherently resistant to removal. We challenge this assumption with DeMark, a query-free black-box attack framework that targets defensive image watermarking schemes for deepfakes. DeMark exploits latent-space vulnerabilities in encoder-decoder watermarking models through a compressive sensing based sparsification process, suppressing watermark signals while preserving perceptual and structural realism appropriate for deepfakes. Across eight state-of-the-art watermarking schemes, DeMark reduces watermark detection accuracy from 100% to 32.9% on average while maintaining natural visual quality, outperforming existing attacks. We further evaluate three defense strategies, including image super resolution, sparse watermarking, and adversarial training, and find them largely ineffective. These results demonstrate that current encoder decoder watermarking schemes remain vulnerable to latent-space manipulations, underscoring the need for more robust watermarking methods to safeguard against deepfakes.", "date": "2026-01-23", "categories": []}
{"id": "2601.16890", "title": "LLM-Based Adversarial Persuasion Attacks on Fact-Checking Systems", "abstract": "Automated fact-checking (AFC) systems are susceptible to adversarial attacks, enabling false claims to evade detection. Existing adversarial frameworks typically rely on injecting noise or altering semantics, yet no existing framework exploits the adversarial potential of persuasion techniques, which are widely used in disinformation campaigns to manipulate audiences. In this paper, we introduce a novel class of persuasive adversarial attacks on AFCs by employing a generative LLM to rephrase claims using persuasion techniques. Considering 15 techniques grouped into 6 categories, we study the effects of persuasion on both claim verification and evidence retrieval using a decoupled evaluation strategy. Experiments on the FEVER and FEVEROUS benchmarks show that persuasion attacks can substantially degrade both verification performance and evidence retrieval. Our analysis identifies persuasion techniques as a potent class of adversarial attacks, highlighting the need for more robust AFC systems.", "date": "2026-01-23", "categories": []}
{"id": "2601.01747", "title": "Crafting Adversarial Inputs for Large Vision-Language Models Using Black-Box Optimization", "abstract": "Recent advancements in Large Vision-Language Models (LVLMs) have shown groundbreaking capabilities across diverse multimodal tasks. However, these models remain vulnerable to adversarial jailbreak attacks, where adversaries craft subtle perturbations to bypass safety mechanisms and trigger harmful outputs. Existing white-box attacks methods require full model accessibility, suffer from computing costs and exhibit insufficient adversarial transferability, making them impractical for real-world, black-box settings. To address these limitations, we propose a black-box jailbreak attack on LVLMs via Zeroth-Order optimization using Simultaneous Perturbation Stochastic Approximation (ZO-SPSA). ZO-SPSA provides three key advantages: (i) gradient-free approximation by input-output interactions without requiring model knowledge, (ii) model-agnostic optimization without the surrogate model and (iii) lower resource requirements with reduced GPU memory consumption. We evaluate ZO-SPSA on three LVLMs, including InstructBLIP, LLaVA and MiniGPT-4, achieving the highest jailbreak success rate of 83.0% on InstructBLIP, while maintaining imperceptible perturbations comparable to white-box methods. Moreover, adversarial examples generated from MiniGPT-4 exhibit strong transferability to other LVLMs, with ASR reaching 64.18%. These findings underscore the real-world feasibility of black-box jailbreaks and expose critical weaknesses in the safety mechanisms of current LVLMs", "date": "2026-01-05", "categories": []}
{"id": "2601.15331", "title": "RECAP: A Resource-Efficient Method for Adversarial Prompting in Large Language Models", "abstract": "The deployment of large language models (LLMs) has raised security concerns due to their susceptibility to producing harmful or policy-violating outputs when exposed to adversarial prompts. While alignment and guardrails mitigate common misuse, they remain vulnerable to automated jailbreaking methods such as GCG, PEZ, and GBDA, which generate adversarial suffixes via training and gradient-based search. Although effective, these methods particularly GCG are computationally expensive, limiting their practicality for organisations with constrained resources. This paper introduces a resource-efficient adversarial prompting approach that eliminates the need for retraining by matching new prompts to a database of pre-trained adversarial prompts. A dataset of 1,000 prompts was classified into seven harm-related categories, and GCG, PEZ, and GBDA were evaluated on a Llama 3 8B model to identify the most effective attack method per category. Results reveal a correlation between prompt type and algorithm effectiveness. By retrieving semantically similar successful adversarial prompts, the proposed method achieves competitive attack success rates with significantly reduced computational cost. This work provides a practical framework for scalable red-teaming and security evaluation of aligned LLMs, including in settings where model internals are inaccessible.", "date": "2026-01-20", "categories": []}
{"id": "2601.16070", "title": "On damage of interpolation to adversarial robustness in regression", "abstract": "Deep neural networks (DNNs) typically involve a large number of parameters and are trained to achieve zero or near-zero training error. Despite such interpolation, they often exhibit strong generalization performance on unseen data, a phenomenon that has motivated extensive theoretical investigations. Comforting results show that interpolation indeed may not affect the minimax rate of convergence under the squared error loss. In the mean time, DNNs are well known to be highly vulnerable to adversarial perturbations in future inputs. A natural question then arises: Can interpolation also escape from suboptimal performance under a future $X$-attack? In this paper, we investigate the adversarial robustness of interpolating estimators in a framework of nonparametric regression. A finding is that interpolating estimators must be suboptimal even under a subtle future $X$-attack, and achieving perfect fitting can substantially damage their robustness. An interesting phenomenon in the high interpolation regime, which we term the curse of simple size, is also revealed and discussed. Numerical experiments support our theoretical findings.", "date": "2026-01-22", "categories": []}
{"id": "2410.08864", "title": "The Good, the Bad and the Ugly: Meta-Analysis of Watermarks, Transferable Attacks and Adversarial Defenses", "abstract": "We formalize and analyze the trade-off between backdoor-based watermarks and adversarial defenses, framing it as an interactive protocol between a verifier and a prover. While previous works have primarily focused on this trade-off, our analysis extends it by identifying transferable attacks as a third, counterintuitive, but necessary option. Our main result shows that for all learning tasks, at least one of the three exists: a watermark, an adversarial defense, or a transferable attack. By transferable attack, we refer to an efficient algorithm that generates queries indistinguishable from the data distribution and capable of fooling all efficient defenders. Using cryptographic techniques, specifically fully homomorphic encryption, we construct a transferable attack and prove its necessity in this trade-off. Finally, we show that tasks of bounded VC-dimension allow adversarial defenses against all attackers, while a subclass allows watermarks secure against fast adversaries.", "date": "2024-10-11", "categories": []}
{"id": "2508.09275", "title": "Constrained Black-Box Attacks Against Cooperative Multi-Agent Reinforcement Learning", "abstract": "Collaborative multi-agent reinforcement learning has rapidly evolved, offering state-of-the-art algorithms for real-world applications, including sensitive domains. However, a key challenge to its widespread adoption is the lack of a thorough investigation into its vulnerabilities to adversarial attacks. Existing work predominantly focuses on training-time attacks or unrealistic scenarios, such as access to policy weights or the ability to train surrogate policies. In this paper, we investigate new vulnerabilities under more challenging and constrained conditions, assuming an adversary can only collect and perturb the observations of deployed agents. We also consider scenarios where the adversary has no access at all (no observations, actions, or weights). Our main approach is to generate perturbations that intentionally misalign how victim agents see their environment. Our approach is empirically validated on three benchmarks and 22 environments, demonstrating its effectiveness across diverse algorithms and environments. Furthermore, we show that our algorithm is sample-efficient, requiring only 1,000 samples compared to the millions needed by previous methods.", "date": "2025-08-12", "categories": []}
{"id": "2601.14266", "title": "GCG Attack On A Diffusion LLM", "abstract": "While most LLMs are autoregressive, diffusion-based LLMs have recently emerged as an alternative method for generation. Greedy Coordinate Gradient (GCG) attacks have proven effective against autoregressive models, but their applicability to diffusion language models remains largely unexplored. In this work, we present an exploratory study of GCG-style adversarial prompt attacks on LLaDA (Large Language Diffusion with mAsking), an open-source diffusion LLM. We evaluate multiple attack variants, including prefix perturbations and suffix-based adversarial generation, on harmful prompts drawn from the AdvBench dataset. Our study provides initial insights into the robustness and attack surface of diffusion language models and motivates the development of alternative optimization and evaluation strategies for adversarial analysis in this setting.", "date": "2025-12-30", "categories": []}
{"id": "2601.14302", "title": "DDSA: Dual-Domain Strategic Attack for Spatial-Temporal Efficiency in Adversarial Robustness Testing", "abstract": "Image transmission and processing systems in resource-critical applications face significant challenges from adversarial perturbations that compromise mission-specific object classification. Current robustness testing methods require excessive computational resources through exhaustive frame-by-frame processing and full-image perturbations, proving impractical for large-scale deployments where massive image streams demand immediate processing. This paper presents DDSA (Dual-Domain Strategic Attack), a resource-efficient adversarial robustness testing framework that optimizes testing through temporal selectivity and spatial precision. We introduce a scenario-aware trigger function that identifies critical frames requiring robustness evaluation based on class priority and model uncertainty, and employ explainable AI techniques to locate influential pixel regions for targeted perturbation. Our dual-domain approach achieves substantial temporal-spatial resource conservation while maintaining attack effectiveness. The framework enables practical deployment of comprehensive adversarial robustness testing in resource-constrained real-time applications where computational efficiency directly impacts mission success.", "date": "2026-01-18", "categories": []}
{"id": "2601.14950", "title": "Erosion Attack for Adversarial Training to Enhance Semantic Segmentation Robustness", "abstract": "Existing segmentation models exhibit significant vulnerability to adversarial attacks.To improve robustness, adversarial training incorporates adversarial examples into model training. However, existing attack methods consider only global semantic information and ignore contextual semantic relationships within the samples, limiting the effectiveness of adversarial training. To address this issue, we propose EroSeg-AT, a vulnerability-aware adversarial training framework that leverages EroSeg to generate adversarial examples. EroSeg first selects sensitive pixels based on pixel-level confidence and then progressively propagates perturbations to higher-confidence pixels, effectively disrupting the semantic consistency of the samples. Experimental results show that, compared to existing methods, our approach significantly improves attack effectiveness and enhances model robustness under adversarial training.", "date": "2026-01-21", "categories": []}
{"id": "2411.00839", "title": "CausAdv: A Causal-based Framework for Detecting Adversarial Examples", "abstract": "Deep learning has led to tremendous success in computer vision, largely due to Convolutional Neural Networks (CNNs). However, CNNs have been shown to be vulnerable to crafted adversarial perturbations. This vulnerability of adversarial examples has has motivated research into improving model robustness through adversarial detection and defense methods. In this paper, we address the adversarial robustness of CNNs through causal reasoning. We propose CausAdv: a causal framework for detecting adversarial examples based on counterfactual reasoning. CausAdv learns both causal and non-causal features of every input, and quantifies the counterfactual information (CI) of every filter of the last convolutional layer. We then perform a statistical analysis of the filters' CI across clean and adversarial samples, to demonstrate that adversarial examples exhibit different CI distributions compared to clean samples. Our results show that causal reasoning enhances the process of adversarial detection without the need to train a separate detector. Moreover, we illustrate the efficiency of causal explanations as a helpful detection tool by visualizing the extracted causal features.", "date": "2024-10-29", "categories": []}
{"id": "2509.12939", "title": "Sy-FAR: Symmetry-based Fair Adversarial Robustness", "abstract": "Security-critical machine-learning (ML) systems, such as face-recognition systems, are susceptible to adversarial examples, including real-world physically realizable attacks. Various means to boost ML's adversarial robustness have been proposed; however, they typically induce unfair robustness: It is often easier to attack from certain classes or groups than from others. Several techniques have been developed to improve adversarial robustness while seeking perfect fairness between classes. Yet, prior work has focused on settings where security and fairness are less critical. Our insight is that achieving perfect parity in realistic fairness-critical tasks, such as face recognition, is often infeasible -- some classes may be highly similar, leading to more misclassifications between them. Instead, we suggest that seeking symmetry -- i.e., attacks from class $i$ to $j$ would be as successful as from $j$ to $i$ -- is more tractable. Intuitively, symmetry is a desirable because class resemblance is a symmetric relation in most domains. Additionally, as we prove theoretically, symmetry between individuals induces symmetry between any set of sub-groups, in contrast to other fairness notions where group-fairness is often elusive. We develop Sy-FAR, a technique to encourage symmetry while also optimizing adversarial robustness and extensively evaluate it using five datasets, with three model architectures, including against targeted and untargeted realistic attacks. The results show Sy-FAR significantly improves fair adversarial robustness compared to state-of-the-art methods. Moreover, we find that Sy-FAR is faster and more consistent across runs. Notably, Sy-FAR also ameliorates another type of unfairness we discover in this work -- target classes that adversarial examples are likely to be classified into become significantly less vulnerable after inducing symmetry.", "date": "2025-09-16", "categories": []}
{"id": "2601.12042", "title": "Less Is More -- Until It Breaks: Security Pitfalls of Vision Token Compression in Large Vision-Language Models", "abstract": "Visual token compression is widely adopted to improve the inference efficiency of Large Vision-Language Models (LVLMs), enabling their deployment in latency-sensitive and resource-constrained scenarios. However, existing work has mainly focused on efficiency and performance, while the security implications of visual token compression remain largely unexplored. In this work, we first reveal that visual token compression substantially degrades the robustness of LVLMs: models that are robust under uncompressed inference become highly vulnerable once compression is enabled. These vulnerabilities are state-specific; failure modes emerge only in the compressed setting and completely disappear when compression is disabled, making them particularly hidden and difficult to diagnose. By analyzing the key stages of the compression process, we identify instability in token importance ranking as the primary cause of this robustness degradation. Small and imperceptible perturbations can significantly alter token rankings, leading the compression mechanism to mistakenly discard task-critical information and ultimately causing model failure. Motivated by this observation, we propose a Compression-Aware Attack to systematically study and exploit this vulnerability. CAA directly targets the token selection mechanism and induces failures exclusively under compressed inference. We further extend this approach to more realistic black-box settings and introduce Transfer CAA, where neither the target model nor the compression configuration is accessible. We further evaluate potential defenses and find that they provide only limited protection. Extensive experiments across models, datasets, and compression methods show that visual token compression significantly undermines robustness, revealing a previously overlooked efficiency-security trade-off.", "date": "2026-01-17", "categories": []}
{"id": "2601.12304", "title": "A Two-Stage Globally-Diverse Adversarial Attack for Vision-Language Pre-training Models", "abstract": "Vision-language pre-training (VLP) models are vulnerable to adversarial examples, particularly in black-box scenarios. Existing multimodal attacks often suffer from limited perturbation diversity and unstable multi-stage pipelines. To address these challenges, we propose 2S-GDA, a two-stage globally-diverse attack framework. The proposed method first introduces textual perturbations through a globally-diverse strategy by combining candidate text expansion with globally-aware replacement. To enhance visual diversity, image-level perturbations are generated using multi-scale resizing and block-shuffle rotation. Extensive experiments on VLP models demonstrate that 2S-GDA consistently improves attack success rates over state-of-the-art methods, with gains of up to 11.17\\% in black-box settings. Our framework is modular and can be easily combined with existing methods to further enhance adversarial transferability.", "date": "2026-01-18", "categories": []}
{"id": "2601.12624", "title": "Towards Robust Universal Perturbation Attacks: A Float-Coded, Penalty-Driven Evolutionary Approach", "abstract": "Universal adversarial perturbations (UAPs) have garnered significant attention due to their ability to undermine deep neural networks across multiple inputs using a single noise pattern. Evolutionary algorithms offer a promising approach to generating such perturbations due to their ability to navigate non-convex, gradient-free landscapes. In this work, we introduce a float-coded, penalty-driven single-objective evolutionary framework for UAP generation that achieves lower visibility perturbations while enhancing attack success rates. Our approach leverages continuous gene representations aligned with contemporary deep learning scales, incorporates dynamic evolutionary operators with adaptive scheduling, and utilizes a modular PyTorch implementation for seamless integration with modern architectures. Additionally, we ensure the universality of the generated perturbations by testing across diverse models and by periodically switching batches to prevent overfitting. Experimental results on the ImageNet dataset demonstrate that our framework consistently produces perturbations with smaller norms, higher misclassification effectiveness, and faster convergence compared to existing evolutionary-based methods. These findings highlight the robustness and scalability of our approach for universal adversarial attacks across various deep learning architectures.", "date": "2026-01-18", "categories": []}
{"id": "2601.12865", "title": "Proxy Robustness in Vision Language Models is Effortlessly Transferable", "abstract": "As a pivotal technique for improving the defense of deep models, adversarial robustness transfer via distillation has demonstrated remarkable success in conventional image classification tasks. However, this paradigm encounters critical challenges when applied to vision-language models (VLM) (e.g., CLIP): constructing adversarially robust teacher for large-scale multi-modal models demands prohibitively high computational resources. We bridge this gap by revealing an interesting phenomenon: vanilla CLIP (without adversarial training) exhibits intrinsic defensive capabilities against adversarial examples generated by another CLIP with different architectures. We formally define this as proxy adversarial robustness, and naturally propose a Heterogeneous Proxy Transfer (HPT) framework that establishes cross-architectural robustness distillation channels between CLIP variants, effortlessly enabling the VLM robustness transfer from proxy to target models. Yet, such proxy transfer paradigm easily induces severe overfitting, leading to a sharp degradation in zero-shot natural generalization. To resolve that, we design Generalization-Pivot Decoupling (GPD) by leveraging the difference in learning rate scheduling. This decouples the proxy transfer process into a generalization-anchored warm-up that maintains generalization and a generalization-pulled HPT that promotes adversarial robustness, to achieve an equilibrium between natural generalization and adversarial robustness. Extensive experiments on 15 zero-shot datasets demonstrate the effectiveness of our HPT-GPD method. The code is available at the website of github.com/fxw13/HPT-GPD.", "date": "2026-01-19", "categories": []}
{"id": "2601.13162", "title": "NeuroShield: A Neuro-Symbolic Framework for Adversarial Robustness", "abstract": "Adversarial vulnerability and lack of interpretability are critical limitations of deep neural networks, especially in safety-sensitive settings such as autonomous driving. We introduce \\DesignII, a neuro-symbolic framework that integrates symbolic rule supervision into neural networks to enhance both adversarial robustness and explainability. Domain knowledge is encoded as logical constraints over appearance attributes such as shape and color, and enforced through semantic and symbolic logic losses applied during training. Using the GTSRB dataset, we evaluate robustness against FGSM and PGD attacks at a standard $\\ell_\\infty$ perturbation budget of $\\varepsilon = 8/255$. Relative to clean training, standard adversarial training provides modest improvements in robustness ($\\sim$10 percentage points). Conversely, our FGSM-Neuro-Symbolic and PGD-Neuro-Symbolic models achieve substantially larger gains, improving adversarial accuracy by 18.1\\% and 17.35\\% over their corresponding adversarial-training baselines, representing roughly a three-fold larger robustness gain than standard adversarial training provides when both are measured relative to the same clean-training baseline, without reducing clean-sample accuracy. Compared to transformer-based defenses such as LNL-MoEx, which require heavy architectures and extensive data augmentation, our PGD-Neuro-Symbolic variant attains comparable or superior robustness using a ResNet18 backbone trained for 10 epochs. These results show that symbolic reasoning offers an effective path to robust and interpretable AI.", "date": "2026-01-19", "categories": []}
{"id": "2601.13645", "title": "Quadratic Upper Bound for Boosting Robustness", "abstract": "Fast adversarial training (FAT) aims to enhance the robustness of models against adversarial attacks with reduced training time, however, FAT often suffers from compromised robustness due to insufficient exploration of adversarial space. In this paper, we develop a loss function to mitigate the problem of degraded robustness under FAT. Specifically, we derive a quadratic upper bound (QUB) on the adversarial training (AT) loss function and propose to utilize the bound with existing FAT methods. Our experimental results show that applying QUB loss to the existing methods yields significant improvement of robustness. Furthermore, using various metrics, we demonstrate that this improvement is likely to result from the smoothened loss landscape of the resulting model.", "date": "2026-01-20", "categories": []}
{"id": "2511.15304", "title": "Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in Large Language Models", "abstract": "We present evidence that adversarial poetry functions as a universal single-turn jailbreak technique for Large Language Models (LLMs). Across 25 frontier proprietary and open-weight models, curated poetic prompts yielded high attack-success rates (ASR), with some providers exceeding 90%. Mapping prompts to MLCommons and EU CoP risk taxonomies shows that poetic attacks transfer across CBRN, manipulation, cyber-offence, and loss-of-control domains. Converting 1,200 MLCommons harmful prompts into verse via a standardized meta-prompt produced ASRs up to 18 times higher than their prose baselines. Outputs are evaluated using an ensemble of 3 open-weight LLM judges, whose binary safety assessments were validated on a stratified human-labeled subset. Poetic framing achieved an average jailbreak success rate of 62% for hand-crafted poems and approximately 43% for meta-prompt conversions (compared to non-poetic baselines), substantially outperforming non-poetic baselines and revealing a systematic vulnerability across model families and safety training approaches. These findings demonstrate that stylistic variation alone can circumvent contemporary safety mechanisms, suggesting fundamental limitations in current alignment methods and evaluation protocols.", "date": "2025-11-19", "categories": []}
{"id": "2601.08837", "title": "From Adversarial Poetry to Adversarial Tales: An Interpretability Research Agenda", "abstract": "Safety mechanisms in LLMs remain vulnerable to attacks that reframe harmful requests through culturally coded structures. We introduce Adversarial Tales, a jailbreak technique that embeds harmful content within cyberpunk narratives and prompts models to perform functional analysis inspired by Vladimir Propp's morphology of folktales. By casting the task as structural decomposition, the attack induces models to reconstruct harmful procedures as legitimate narrative interpretation. Across 26 frontier models from nine providers, we observe an average attack success rate of 71.3%, with no model family proving reliably robust. Together with our prior work on Adversarial Poetry, these findings suggest that structurally-grounded jailbreaks constitute a broad vulnerability class rather than isolated techniques. The space of culturally coded frames that can mediate harmful intent is vast, likely inexhaustible by pattern-matching defenses alone. Understanding why these attacks succeed is therefore essential: we outline a mechanistic interpretability research agenda to investigate how narrative cues reshape model representations and whether models can learn to recognize harmful intent independently of surface form.", "date": "2025-12-16", "categories": []}
{"id": "2601.11006", "title": "Backdoor Attacks on Multi-modal Contrastive Learning", "abstract": "Contrastive learning has become a leading self- supervised approach to representation learning across domains, including vision, multimodal settings, graphs, and federated learning. However, recent studies have shown that contrastive learning is susceptible to backdoor and data poisoning attacks. In these attacks, adversaries can manipulate pretraining data or model updates to insert hidden malicious behavior. This paper offers a thorough and comparative review of backdoor attacks in contrastive learning. It analyzes threat models, attack methods, target domains, and available defenses. We summarize recent advancements in this area, underline the specific vulnerabilities inherent to contrastive learning, and discuss the challenges and future research directions. Our findings have significant implications for the secure deployment of systems in industrial and distributed environments.", "date": "2026-01-16", "categories": []}
{"id": "2501.14230", "title": "GreedyPixel: Fine-Grained Black-Box Adversarial Attack Via Greedy Algorithm", "abstract": "Deep neural networks are highly vulnerable to adversarial examples, which are inputs with small, carefully crafted perturbations that cause misclassification -- making adversarial attacks a critical tool for evaluating robustness. Existing black-box methods typically entail a trade-off between precision and flexibility: pixel-sparse attacks (e.g., single- or few-pixel attacks) provide fine-grained control but lack adaptability, whereas patch- or frequency-based attacks improve efficiency or transferability, but at the cost of producing larger and less precise perturbations. We present GreedyPixel, a fine-grained black-box attack method that performs brute-force-style, per-pixel greedy optimization guided by a surrogate-derived priority map and refined by means of query feedback. It evaluates each coordinate directly without any gradient information, guaranteeing monotonic loss reduction and convergence to a coordinate-wise optimum, while also yielding near white-box-level precision and pixel-wise sparsity and perceptual quality. On the CIFAR-10 and ImageNet datasets, spanning convolutional neural networks (CNNs) and Transformer models, GreedyPixel achieved state-of-the-art success rates with visually imperceptible perturbations, effectively bridging the gap between black-box practicality and white-box performance. The implementation is available at https://github.com/azrealwang/greedypixel.", "date": "2025-01-24", "categories": []}
{"id": "2601.09806", "title": "Diffusion-Driven Deceptive Patches: Adversarial Manipulation and Forensic Detection in Facial Identity Verification", "abstract": "This work presents an end-to-end pipeline for generating, refining, and evaluating adversarial patches to compromise facial biometric systems, with applications in forensic analysis and security testing. We utilize FGSM to generate adversarial noise targeting an identity classifier and employ a diffusion model with reverse diffusion to enhance imperceptibility through Gaussian smoothing and adaptive brightness correction, thereby facilitating synthetic adversarial patch evasion. The refined patch is applied to facial images to test its ability to evade recognition systems while maintaining natural visual characteristics. A Vision Transformer (ViT)-GPT2 model generates captions to provide a semantic description of a person's identity for adversarial images, supporting forensic interpretation and documentation for identity evasion and recognition attacks. The pipeline evaluates changes in identity classification, captioning results, and vulnerabilities in facial identity verification and expression recognition under adversarial conditions. We further demonstrate effective detection and analysis of adversarial patches and adversarial samples using perceptual hashing and segmentation, achieving an SSIM of 0.95.", "date": "2026-01-14", "categories": []}
{"id": "2405.16181", "title": "Boosting Adversarial Transferability with Low-Cost Optimization via Maximin Expected Flatness", "abstract": "Transfer-based attacks craft adversarial examples on white-box surrogate models and directly deploy them against black-box target models, offering model-agnostic and query-free threat scenarios. While flatness-enhanced methods have recently emerged to improve transferability by enhancing the loss surface flatness of adversarial examples, their divergent flatness definitions and heuristic attack designs suffer from unexamined optimization limitations and missing theoretical foundation, thus constraining their effectiveness and efficiency. This work exposes the severely imbalanced exploitation-exploration dynamics in flatness optimization, establishing the first theoretical foundation for flatness-based transferability and proposing a principled framework to overcome these optimization pitfalls. Specifically, we systematically unify fragmented flatness definitions across existing methods, revealing their imbalanced optimization limitations in over-exploration of sensitivity peaks or over-exploitation of local plateaus. To resolve these issues, we rigorously formalize average-case flatness and transferability gaps, proving that enhancing zeroth-order average-case flatness minimizes cross-model discrepancies. Building on this theory, we design a Maximin Expected Flatness (MEF) attack that enhances zeroth-order average-case flatness while balancing flatness exploration and exploitation. Extensive evaluations across 22 models and 24 current transfer-based attacks demonstrate MEF's superiority: it surpasses the state-of-the-art PGN attack by 4% in attack success rate at half the computational cost and achieves 8% higher success rate under the same budget. When combined with input augmentation, MEF attains 15% additional gains against defense-equipped models, establishing new robustness benchmarks. Our code is available at https://github.com/SignedQiu/MEFAttack.", "date": "2024-05-25", "categories": []}
{"id": "2508.01741", "title": "Simulated Ensemble Attack: Transferring Jailbreaks Across Fine-tuned Vision-Language Models", "abstract": "The widespread practice of fine-tuning open-source Vision-Language Models (VLMs) raises a critical security concern: jailbreak vulnerabilities in base models may persist in downstream variants, enabling transferable attacks across fine-tuned systems. To investigate this risk, we propose the Simulated Ensemble Attack (SEA), a grey-box jailbreak framework that assumes full access to the base VLM but no knowledge of the fine-tuned target. SEA enhances transferability via Fine-tuning Trajectory Simulation (FTS), which models bounded parameter variations in the vision encoder, and Targeted Prompt Guidance (TPG), which stabilizes adversarial optimization through auxiliary textual guidance. Experiments on the Qwen2-VL family demonstrate that SEA achieves consistently high transfer success and toxicity rates across diverse fine-tuned variants, including safety-enhanced models, while standard PGD-based image jailbreaks exhibit negligible transferability. Further analysis reveals that fine-tuning primarily induces localized parameter shifts around the base model, explaining why attacks optimized over a simulated neighborhood transfer effectively. We also show that SEA generalizes across different base generations (e.g., Qwen2.5/3-VL), indicating that its effectiveness arises from shared fine-tuning-induced behaviors rather than architecture- or initialization-specific factors.", "date": "2025-08-03", "categories": []}
{"id": "2601.01852", "title": "MORE: Multi-Objective Adversarial Attacks on Speech Recognition", "abstract": "The emergence of large-scale automatic speech recognition (ASR) models such as Whisper has greatly expanded their adoption across diverse real-world applications. Ensuring robustness against even minor input perturbations is therefore critical for maintaining reliable performance in real-time environments. While prior work has mainly examined accuracy degradation under adversarial attacks, robustness with respect to efficiency remains largely unexplored. This narrow focus provides only a partial understanding of ASR model vulnerabilities. To address this gap, we conduct a comprehensive study of ASR robustness under multiple attack scenarios. We introduce MORE, a multi-objective repetitive doubling encouragement attack, which jointly degrades recognition accuracy and inference efficiency through a hierarchical staged repulsion-anchoring mechanism. Specifically, we reformulate multi-objective adversarial optimization into a hierarchical framework that sequentially achieves the dual objectives. To further amplify effectiveness, we propose a novel repetitive encouragement doubling objective (REDO) that induces duplicative text generation by maintaining accuracy degradation and periodically doubling the predicted sequence length. Overall, MORE compels ASR models to produce incorrect transcriptions at a substantially higher computational cost, triggered by a single adversarial input. Experiments show that MORE consistently yields significantly longer transcriptions while maintaining high word error rates compared to existing baselines, underscoring its effectiveness in multi-objective adversarial attack.", "date": "2026-01-05", "categories": []}
{"id": "2601.09292", "title": "Blue Teaming Function-Calling Agents", "abstract": "We present an experimental evaluation that assesses the robustness of four open source LLMs claiming function-calling capabilities against three different attacks, and we measure the effectiveness of eight different defences. Our results show how these models are not safe by default, and how the defences are not yet employable in real-world scenarios.", "date": "2026-01-14", "categories": []}
{"id": "2405.14519", "title": "A New Formulation for Zeroth-Order Optimization of Adversarial EXEmples in Malware Detection", "abstract": "Machine learning malware detectors are vulnerable to adversarial EXEmples, i.e., carefully-crafted Windows programs tailored to evade detection. Unlike other adversarial problems, attacks in this context must be functionality-preserving, a constraint that is challenging to address. As a consequence, heuristic algorithms are typically used, which inject new content, either randomly-picked or harvested from legitimate programs. In this paper, we show how learning malware detectors can be cast within a zeroth-order optimization framework, which allows incorporating functionality-preserving manipulations. This permits the deployment of sound and efficient gradient-free optimization algorithms, which come with theoretical guarantees and allow for minimal hyper-parameters tuning. As a by-product, we propose and study ZEXE, a novel zeroth-order attack against Windows malware detection. Compared to state-of-the-art techniques, ZEXE provides improvement in the evasion rate, reducing to less than one third the size of the injected content.", "date": "2024-05-23", "categories": []}
{"id": "2511.12735", "title": "Backdoor Attacks on Open Vocabulary Object Detectors via Multi-Modal Prompt Tuning", "abstract": "Open-vocabulary object detectors (OVODs) unify vision and language to detect arbitrary object categories based on text prompts, enabling strong zero-shot generalization to novel concepts. As these models gain traction in high-stakes applications such as robotics, autonomous driving, and surveillance, understanding their security risks becomes crucial. In this work, we conduct the first study of backdoor attacks on OVODs and reveal a new attack surface introduced by prompt tuning. We propose TrAP (Trigger-Aware Prompt tuning), a multi-modal backdoor injection strategy that jointly optimizes prompt parameters in both image and text modalities along with visual triggers. TrAP enables the attacker to implant malicious behavior using lightweight, learnable prompt tokens without retraining the base model weights, thus preserving generalization while embedding a hidden backdoor. We adopt a curriculum-based training strategy that progressively shrinks the trigger size, enabling effective backdoor activation using small trigger patches at inference. Experiments across multiple datasets show that TrAP achieves high attack success rates for both object misclassification and object disappearance attacks, while also improving clean image performance on downstream datasets compared to the zero-shot setting. Code: https://github.com/rajankita/TrAP", "date": "2025-11-16", "categories": []}
{"id": "2211.01671", "title": "Visual Adversarial Attacks and Defenses in the Physical World: A Survey", "abstract": "Although Deep Neural Networks (DNNs) have been widely applied in various real-world scenarios, they remain vulnerable to adversarial examples. Adversarial attacks in computer vision can be categorized into digital attacks and physical attacks based on their different forms. Compared to digital attacks, which generate perturbations in digital pixels, physical attacks are more practical in real-world settings. Due to the serious security risks posed by physically adversarial examples, many studies have been conducted to evaluate the physically adversarial robustness of DNNs in recent years. In this paper, we provide a comprehensive survey of current physically adversarial attacks and defenses in computer vision. We establish a taxonomy by organizing physical attacks according to attack tasks, attack forms, and attack methods. This approach offers readers a systematic understanding of the topic from multiple perspectives. For physical defenses, we categorize them into pre-processing, in-processing, and post-processing for DNN models to ensure comprehensive coverage of adversarial defenses. Based on this survey, we discuss the challenges facing this research field and provide an outlook on future directions.", "date": "2022-11-03", "categories": []}
{"id": "2310.17436", "title": "Uncertainty-weighted Loss Functions for Improved Adversarial Attacks on Semantic Segmentation", "abstract": "State-of-the-art deep neural networks have been shown to be extremely powerful in a variety of perceptual tasks like semantic segmentation. However, these networks are vulnerable to adversarial perturbations of the input which are imperceptible for humans but lead to incorrect predictions. Treating image segmentation as a sum of pixel-wise classifications, adversarial attacks developed for classification models were shown to be applicable to segmentation models as well. In this work, we present simple uncertainty-based weighting schemes for the loss functions of such attacks that (i) put higher weights on pixel classifications which can more easily perturbed and (ii) zero-out the pixel-wise losses corresponding to those pixels that are already confidently misclassified. The weighting schemes can be easily integrated into the loss function of a range of well-known adversarial attackers with minimal additional computational overhead, but lead to significant improved perturbation performance, as we demonstrate in our empirical analysis on several datasets and models.", "date": "2023-10-26", "categories": []}
{"id": "2402.05674", "title": "A High Dimensional Statistical Model for Adversarial Training: Geometry and Trade-Offs", "abstract": "This work investigates adversarial training in the context of margin-based linear classifiers in the high-dimensional regime where the dimension $d$ and the number of data points $n$ diverge with a fixed ratio $\\alpha = n / d$. We introduce a tractable mathematical model where the interplay between the data and adversarial attacker geometries can be studied, while capturing the core phenomenology observed in the adversarial robustness literature. Our main theoretical contribution is an exact asymptotic description of the sufficient statistics for the adversarial empirical risk minimiser, under generic convex and non-increasing losses for a Block Feature Model. Our result allow us to precisely characterise which directions in the data are associated with a higher generalisation/robustness trade-off, as defined by a robustness and a usefulness metric. We show that the the presence of multiple different feature types is crucial to the high sample complexity performances of adversarial training. In particular, we unveil the existence of directions which can be defended without penalising accuracy. Finally, we show the advantage of defending non-robust features during training, identifying a uniform protection as an inherently effective defence mechanism.", "date": "2024-02-08", "categories": []}
{"id": "2504.21323", "title": "How to Backdoor the Knowledge Distillation", "abstract": "Knowledge distillation has become a cornerstone in modern machine learning systems, celebrated for its ability to transfer knowledge from a large, complex teacher model to a more efficient student model. Traditionally, this process is regarded as secure, assuming the teacher model is clean. This belief stems from conventional backdoor attacks relying on poisoned training data with backdoor triggers and attacker-chosen labels, which are not involved in the distillation process. Instead, knowledge distillation uses the outputs of a clean teacher model to guide the student model, inherently preventing recognition or response to backdoor triggers as intended by an attacker. In this paper, we challenge this assumption by introducing a novel attack methodology that strategically poisons the distillation dataset with adversarial examples embedded with backdoor triggers. This technique allows for the stealthy compromise of the student model while maintaining the integrity of the teacher model. Our innovative approach represents the first successful exploitation of vulnerabilities within the knowledge distillation process using clean teacher models. Through extensive experiments conducted across various datasets and attack settings, we demonstrate the robustness, stealthiness, and effectiveness of our method. Our findings reveal previously unrecognized vulnerabilities and pave the way for future research aimed at securing knowledge distillation processes against backdoor attacks.", "date": "2025-04-30", "categories": []}
{"id": "2510.21144", "title": "NeuroGenPoisoning: Neuron-Guided Attacks on Retrieval-Augmented Generation of LLM via Genetic Optimization of External Knowledge", "abstract": "Retrieval-Augmented Generation (RAG) empowers Large Language Models (LLMs) to dynamically integrate external knowledge during inference, improving their factual accuracy and adaptability. However, adversaries can inject poisoned external knowledge to override the model's internal memory. While existing attacks iteratively manipulate retrieval content or prompt structure of RAG, they largely ignore the model's internal representation dynamics and neuron-level sensitivities. The underlying mechanism of RAG poisoning has not been fully studied and the effect of knowledge conflict with strong parametric knowledge in RAG is not considered. In this work, we propose NeuroGenPoisoning, a novel attack framework that generates adversarial external knowledge in RAG guided by LLM internal neuron attribution and genetic optimization. Our method first identifies a set of Poison-Responsive Neurons whose activation strongly correlates with contextual poisoning knowledge. We then employ a genetic algorithm to evolve adversarial passages that maximally activate these neurons. Crucially, our framework enables massive-scale generation of effective poisoned RAG knowledge by identifying and reusing promising but initially unsuccessful external knowledge variants via observed attribution signals. At the same time, Poison-Responsive Neurons guided poisoning can effectively resolves knowledge conflict. Experimental results across models and datasets demonstrate consistently achieving high Population Overwrite Success Rate (POSR) of over 90% while preserving fluency. Empirical evidence shows that our method effectively resolves knowledge conflict.", "date": "2025-10-24", "categories": []}
{"id": "2601.02228", "title": "FMVP: Masked Flow Matching for Adversarial Video Purification", "abstract": "Video recognition models remain vulnerable to adversarial attacks, while existing diffusion-based purification methods suffer from inefficient sampling and curved trajectories. Directly regressing clean videos from adversarial inputs often fails to recover faithful content due to the subtle nature of perturbations; this necessitates physically shattering the adversarial structure. Therefore, we propose Flow Matching for Adversarial Video Purification FMVP. FMVP physically shatters global adversarial structures via a masking strategy and reconstructs clean video dynamics using Conditional Flow Matching (CFM) with an inpainting objective. To further decouple semantic content from adversarial noise, we design a Frequency-Gated Loss (FGL) that explicitly suppresses high-frequency adversarial residuals while preserving low-frequency fidelity. We design Attack-Aware and Generalist training paradigms to handle known and unknown threats, respectively. Extensive experiments on UCF-101 and HMDB-51 demonstrate that FMVP outperforms state-of-the-art methods (DiffPure, Defense Patterns (DP), Temporal Shuffling (TS) and FlowPure), achieving robust accuracy exceeding 87% against PGD and 89% against CW attacks. Furthermore, FMVP demonstrates superior robustness against adaptive attacks (DiffHammer) and functions as a zero-shot adversarial detector, attaining AUC-ROC scores of 0.98 for PGD and 0.79 for highly imperceptible CW attacks.", "date": "2026-01-05", "categories": []}
{"id": "2601.04566", "title": "BackdoorAgent: A Unified Framework for Backdoor Attacks on LLM-based Agents", "abstract": "Large language model (LLM) agents execute tasks through multi-step workflows that combine planning, memory, and tool use. While this design enables autonomy, it also expands the attack surface for backdoor threats. Backdoor triggers injected into specific stages of an agent workflow can persist through multiple intermediate states and adversely influence downstream outputs. However, existing studies remain fragmented and typically analyze individual attack vectors in isolation, leaving the cross-stage interaction and propagation of backdoor triggers poorly understood from an agent-centric perspective. To fill this gap, we propose \\textbf{BackdoorAgent}, a modular and stage-aware framework that provides a unified, agent-centric view of backdoor threats in LLM agents. BackdoorAgent structures the attack surface into three functional stages of agentic workflows, including \\textbf{planning attacks}, \\textbf{memory attacks}, and \\textbf{tool-use attacks}, and instruments agent execution to enable systematic analysis of trigger activation and propagation across different stages. Building on this framework, we construct a standardized benchmark spanning four representative agent applications: \\textbf{Agent QA}, \\textbf{Agent Code}, \\textbf{Agent Web}, and \\textbf{Agent Drive}, covering both language-only and multimodal settings. Our empirical analysis shows that \\textit{triggers implanted at a single stage can persist across multiple steps and propagate through intermediate states.} For instance, when using a GPT-based backbone, we observe trigger persistence in 43.58\\% of planning attacks, 77.97\\% of memory attacks, and 60.28\\% of tool-stage attacks, highlighting the vulnerabilities of the agentic workflow itself to backdoor threats. To facilitate reproducibility and future research, our code and benchmark are publicly available at GitHub.", "date": "2026-01-08", "categories": []}
{"id": "2601.06464", "title": "On the Adversarial Robustness of 3D Large Vision-Language Models", "abstract": "3D Vision-Language Models (VLMs), such as PointLLM and GPT4Point, have shown strong reasoning and generalization abilities in 3D understanding tasks. However, their adversarial robustness remains largely unexplored. Prior work in 2D VLMs has shown that the integration of visual inputs significantly increases vulnerability to adversarial attacks, making these models easier to manipulate into generating toxic or misleading outputs. In this paper, we investigate whether incorporating 3D vision similarly compromises the robustness of 3D VLMs. To this end, we present the first systematic study of adversarial robustness in point-based 3D VLMs. We propose two complementary attack strategies: \\textit{Vision Attack}, which perturbs the visual token features produced by the 3D encoder and projector to assess the robustness of vision-language alignment; and \\textit{Caption Attack}, which directly manipulates output token sequences to evaluate end-to-end system robustness. Each attack includes both untargeted and targeted variants to measure general vulnerability and susceptibility to controlled manipulation. Our experiments reveal that 3D VLMs exhibit significant adversarial vulnerabilities under untargeted attacks, while demonstrating greater resilience against targeted attacks aimed at forcing specific harmful outputs, compared to their 2D counterparts. These findings highlight the importance of improving the adversarial robustness of 3D VLMs, especially as they are deployed in safety-critical applications.", "date": "2026-01-10", "categories": []}
{"id": "2601.06884", "title": "Paraphrasing Adversarial Attack on LLM-as-a-Reviewer", "abstract": "The use of large language models (LLMs) in peer review systems has attracted growing attention, making it essential to examine their potential vulnerabilities. Prior attacks rely on prompt injection, which alters manuscript content and conflates injection susceptibility with evaluation robustness. We propose the Paraphrasing Adversarial Attack (PAA), a black-box optimization method that searches for paraphrased sequences yielding higher review scores while preserving semantic equivalence and linguistic naturalness. PAA leverages in-context learning, using previous paraphrases and their scores to guide candidate generation. Experiments across five ML and NLP conferences with three LLM reviewers and five attacking models show that PAA consistently increases review scores without changing the paper's claims. Human evaluation confirms that generated paraphrases maintain meaning and naturalness. We also find that attacked papers exhibit increased perplexity in reviews, offering a potential detection signal, and that paraphrasing submissions can partially mitigate attacks.", "date": "2026-01-11", "categories": []}
{"id": "2601.07253", "title": "Universal Adversarial Purification with DDIM Metric Loss for Stable Diffusion", "abstract": "Stable Diffusion (SD) often produces degraded outputs when the training dataset contains adversarial noise. Adversarial purification offers a promising solution by removing adversarial noise from contaminated data. However, existing purification methods are primarily designed for classification tasks and fail to address SD-specific adversarial strategies, such as attacks targeting the VAE encoder, UNet denoiser, or both. To address the gap in SD security, we propose Universal Diffusion Adversarial Purification (UDAP), a novel framework tailored for defending adversarial attacks targeting SD models. UDAP leverages the distinct reconstruction behaviors of clean and adversarial images during Denoising Diffusion Implicit Models (DDIM) inversion to optimize the purification process. By minimizing the DDIM metric loss, UDAP can effectively remove adversarial noise. Additionally, we introduce a dynamic epoch adjustment strategy that adapts optimization iterations based on reconstruction errors, significantly improving efficiency without sacrificing purification quality. Experiments demonstrate UDAP's robustness against diverse adversarial methods, including PID (VAE-targeted), Anti-DreamBooth (UNet-targeted), MIST (hybrid), and robustness-enhanced variants like Anti-Diffusion (Anti-DF) and MetaCloak. UDAP also generalizes well across SD versions and text prompts, showcasing its practical applicability in real-world scenarios.", "date": "2026-01-12", "categories": []}
{"id": "2501.01042", "title": "Transferability of Adversarial Attacks in Video-based MLLMs: A Cross-modal Image-to-Video Approach", "abstract": "Video-based multimodal large language models (V-MLLMs) have shown vulnerability to adversarial examples in video-text multimodal tasks. However, the transferability of adversarial videos to unseen models - a common and practical real-world scenario - remains unexplored. In this paper, we pioneer an investigation into the transferability of adversarial video samples across V-MLLMs. We find that existing adversarial attack methods face significant limitations when applied in black-box settings for V-MLLMs, which we attribute to the following shortcomings: (1) lacking generalization in perturbing video features, (2) focusing only on sparse key-frames, and (3) failing to integrate multimodal information. To address these limitations and deepen the understanding of V-MLLM vulnerabilities in black-box scenarios, we introduce the Image-to-Video MLLM (I2V-MLLM) attack. In I2V-MLLM, we utilize an image-based multimodal large language model (I-MLLM) as a surrogate model to craft adversarial video samples. Multimodal interactions and spatiotemporal information are integrated to disrupt video representations within the latent space, improving adversarial transferability. Additionally, a perturbation propagation technique is introduced to handle different unknown frame sampling strategies. Experimental results demonstrate that our method can generate adversarial examples that exhibit strong transferability across different V-MLLMs on multiple video-text multimodal tasks. Compared to white-box attacks on these models, our black-box attacks (using BLIP-2 as a surrogate model) achieve competitive performance, with average attack success rate (AASR) of 57.98% on MSVD-QA and 58.26% on MSRVTT-QA for Zero-Shot VideoQA tasks, respectively.", "date": "2025-01-02", "categories": []}
{"id": "2509.07677", "title": "Spectral Masking and Interpolation Attack (SMIA): A Black-box Adversarial Attack against Voice Authentication and Anti-Spoofing Systems", "abstract": "Voice Authentication Systems (VAS) use unique vocal characteristics for verification. They are increasingly integrated into high-security sectors such as banking and healthcare. Despite their improvements using deep learning, they face severe vulnerabilities from sophisticated threats like deepfakes and adversarial attacks. The emergence of realistic voice cloning complicates detection, as systems struggle to distinguish authentic from synthetic audio. While anti-spoofing countermeasures (CMs) exist to mitigate these risks, many rely on static detection models that can be bypassed by novel adversarial methods, leaving a critical security gap. To demonstrate this vulnerability, we propose the Spectral Masking and Interpolation Attack (SMIA), a novel method that strategically manipulates inaudible frequency regions of AI-generated audio. By altering the voice in imperceptible zones to the human ear, SMIA creates adversarial samples that sound authentic while deceiving CMs. We conducted a comprehensive evaluation of our attack against state-of-the-art (SOTA) models across multiple tasks, under simulated real-world conditions. SMIA achieved a strong attack success rate (ASR) of at least 82% against combined VAS/CM systems, at least 97.5% against standalone speaker verification systems, and 100% against countermeasures. These findings conclusively demonstrate that current security postures are insufficient against adaptive adversarial attacks. This work highlights the urgent need for a paradigm shift toward next-generation defenses that employ dynamic, context-aware frameworks capable of evolving with the threat landscape.", "date": "2025-09-09", "categories": []}
{"id": "2601.05587", "title": "HogVul: Black-box Adversarial Code Generation Framework Against LM-based Vulnerability Detectors", "abstract": "Recent advances in software vulnerability detection have been driven by Language Model (LM)-based approaches. However, these models remain vulnerable to adversarial attacks that exploit lexical and syntax perturbations, allowing critical flaws to evade detection. Existing black-box attacks on LM-based vulnerability detectors primarily rely on isolated perturbation strategies, limiting their ability to efficiently explore the adversarial code space for optimal perturbations. To bridge this gap, we propose HogVul, a black-box adversarial code generation framework that integrates both lexical and syntax perturbations under a unified dual-channel optimization strategy driven by Particle Swarm Optimization (PSO). By systematically coordinating two-level perturbations, HogVul effectively expands the search space for adversarial examples, enhancing the attack efficacy. Extensive experiments on four benchmark datasets demonstrate that HogVul achieves an average attack success rate improvement of 26.05\\% over state-of-the-art baseline methods. These findings highlight the potential of hybrid optimization strategies in exposing model vulnerabilities.", "date": "2026-01-09", "categories": []}
{"id": "2601.05986", "title": "Deepfake detectors are DUMB: A benchmark to assess adversarial training robustness under transferability constraints", "abstract": "Deepfake detection systems deployed in real-world environments are subject to adversaries capable of crafting imperceptible perturbations that degrade model performance. While adversarial training is a widely adopted defense, its effectiveness under realistic conditions -- where attackers operate with limited knowledge and mismatched data distributions - remains underexplored. In this work, we extend the DUMB -- Dataset soUrces, Model architecture and Balance - and DUMBer methodology to deepfake detection. We evaluate detectors robustness against adversarial attacks under transferability constraints and cross-dataset configuration to extract real-world insights. Our study spans five state-of-the-art detectors (RECCE, SRM, XCeption, UCF, SPSL), three attacks (PGD, FGSM, FPBA), and two datasets (FaceForensics++ and Celeb-DF-V2). We analyze both attacker and defender perspectives mapping results to mismatch scenarios. Experiments show that adversarial training strategies reinforce robustness in the in-distribution cases but can also degrade it under cross-dataset configuration depending on the strategy adopted. These findings highlight the need for case-aware defense strategies in real-world applications exposed to adversarial attacks.", "date": "2026-01-09", "categories": []}
{"id": "2502.10452", "title": "Quaternion-Hadamard Network: A Novel Defense Against Adversarial Attacks with a New Dataset", "abstract": "Adverse-weather image restoration (e.g., rain, snow, haze) models remain highly vulnerable to gradient-based white-box adversarial attacks, wherein minimal loss-aligned perturbations cause substantial degradation in the restored output. This paper presents QHNet, a computationally efficient purification-based defense that precedes the restoration network and targets perturbation suppression in the transform and quaternion domains. QHNet incorporates a Quaternion Hadamard Polynomial Denoising Block (QHPDB) and a Quaternion Denoising Residual Block (QDRB) within an encoder-decoder framework to remove high-frequency adversarial noise while preserving fine structural details. Robustness is evaluated using PSNR and SSIM across rain, snow, and haze removal tasks, and further validated under adaptive, defense-aware white-box attacks employing Projected Gradient Descent (PGD), Backward Pass Differentiable Approximation (BPDA), and Expectation Over Transformation (EOT). Experimental results demonstrate that QHNet delivers superior restoration fidelity and significantly improved robustness compared to state-of-the-art purification baselines, confirming its effectiveness for low-level vision pipelines.", "date": "2025-02-12", "categories": []}
{"id": "2503.01839", "title": "Jailbreaking Safeguarded Text-to-Image Models via Large Language Models", "abstract": "Text-to-Image models may generate harmful content, such as pornographic images, particularly when unsafe prompts are submitted. To address this issue, safety filters are often added on top of text-to-image models, or the models themselves are aligned to reduce harmful outputs. However, these defenses remain vulnerable when an attacker strategically designs adversarial prompts to bypass these safety guardrails. In this work, we propose \\alg, a method to jailbreak text-to-image models with safety guardrails using a fine-tuned large language model. Unlike other query-based jailbreak attacks that require repeated queries to the target model, our attack generates adversarial prompts efficiently after fine-tuning our AttackLLM. We evaluate our method on three datasets of unsafe prompts and against five safety guardrails. Our results demonstrate that our approach effectively bypasses safety guardrails, outperforms existing no-box attacks, and also facilitates other query-based attacks.", "date": "2025-03-03", "categories": []}
{"id": "2508.10029", "title": "Latent Fusion Jailbreak: Blending Harmful and Harmless Representations to Elicit Unsafe LLM Outputs", "abstract": "While Large Language Models (LLMs) have achieved remarkable progress, they remain vulnerable to jailbreak attacks. Existing methods, primarily relying on discrete input optimization (e.g., GCG), often suffer from high computational costs and generate high-perplexity prompts that are easily blocked by simple filters. To overcome these limitations, we propose Latent Fusion Jailbreak (LFJ), a stealthy white-box attack that operates in the continuous latent space. Unlike previous approaches, LFJ constructs adversarial representations by mathematically fusing the hidden states of a harmful query with a thematically similar benign query, effectively masking malicious intent while retaining semantic drive. We further introduce a gradient-guided optimization strategy to balance attack success and computational efficiency. Extensive evaluations on Vicuna-7B, LLaMA-2-7B-Chat, Guanaco-7B, LLaMA-3-70B, and Mistral-7B-Instruct show that LFJ achieves an average Attack Success Rate (ASR) of 94.01%, significantly outperforming state-of-the-art baselines like GCG and AutoDAN while avoiding detectable input artifacts. Furthermore, we identify that thematic similarity in the latent space is a critical vulnerability in current safety alignments. Finally, we propose a latent adversarial training defense that reduces LFJ's ASR by over 80% without compromising model utility.", "date": "2025-08-08", "categories": []}
{"id": "2510.07505", "title": "PEAR: Planner-Executor Agent Robustness Benchmark", "abstract": "Large Language Model (LLM)-based Multi-Agent Systems (MAS) have emerged as a powerful paradigm for tackling complex, multi-step tasks across diverse domains. However, despite their impressive capabilities, MAS remain susceptible to adversarial manipulation. Existing studies typically examine isolated attack surfaces or specific scenarios, leaving a lack of holistic understanding of MAS vulnerabilities. To bridge this gap, we introduce PEAR, a benchmark for systematically evaluating both the utility and vulnerability of planner-executor MAS. While compatible with various MAS architectures, our benchmark focuses on the planner-executor structure, which is a practical and widely adopted design. Through extensive experiments, we find that (1) a weak planner degrades overall clean task performance more severely than a weak executor; (2) while a memory module is essential for the planner, having a memory module for the executor does not impact the clean task performance; (3) there exists a trade-off between task performance and robustness; and (4) attacks targeting the planner are particularly effective at misleading the system. These findings offer actionable insights for enhancing the robustness of MAS and lay the groundwork for principled defenses in multi-agent settings.", "date": "2025-10-08", "categories": []}
{"id": "2601.04752", "title": "Skeletonization-Based Adversarial Perturbations on Large Vision Language Model's Mathematical Text Recognition", "abstract": "This work explores the visual capabilities and limitations of foundation models by introducing a novel adversarial attack method utilizing skeletonization to reduce the search space effectively. Our approach specifically targets images containing text, particularly mathematical formula images, which are more challenging due to their LaTeX conversion and intricate structure. We conduct a detailed evaluation of both character and semantic changes between original and adversarially perturbed outputs to provide insights into the models' visual interpretation and reasoning abilities. The effectiveness of our method is further demonstrated through its application to ChatGPT, which shows its practical implications in real-world scenarios.", "date": "2026-01-08", "categories": []}
{"id": "2601.04991", "title": "Higher-Order Adversarial Patches for Real-Time Object Detectors", "abstract": "Higher-order adversarial attacks can directly be considered the result of a cat-and-mouse game -- an elaborate action involving constant pursuit, near captures, and repeated escapes. This idiom describes the enduring circular training of adversarial attack patterns and adversarial training the best. The following work investigates the impact of higher-order adversarial attacks on object detectors by successively training attack patterns and hardening object detectors with adversarial training. The YOLOv10 object detector is chosen as a representative, and adversarial patches are used in an evasion attack manner. Our results indicate that higher-order adversarial patches are not only affecting the object detector directly trained on but rather provide a stronger generalization capacity compared to lower-order adversarial patches. Moreover, the results highlight that solely adversarial training is not sufficient to harden an object detector efficiently against this kind of adversarial attack. Code: https://github.com/JensBayer/HigherOrder", "date": "2026-01-08", "categories": []}
{"id": "2508.10390", "title": "Jailbreaking Commercial Black-Box LLMs with Explicitly Harmful Prompts", "abstract": "Existing black-box jailbreak attacks achieve certain success on non-reasoning models but degrade significantly on recent SOTA reasoning models. To improve attack ability, inspired by adversarial aggregation strategies, we integrate multiple jailbreak tricks into a single developer template. Especially, we apply Adversarial Context Alignment to purge semantic inconsistencies and use NTP (a type of harmful prompt) -based few-shot examples to guide malicious outputs, lastly forming DH-CoT attack with a fake chain of thought. In experiments, we further observe that existing red-teaming datasets include samples unsuitable for evaluating attack gains, such as BPs, NHPs, and NTPs. Such data hinders accurate evaluation of true attack effect lifts. To address this, we introduce MDH, a Malicious content Detection framework integrating LLM-based annotation with Human assistance, with which we clean data and build RTA dataset suite. Experiments show that MDH reliably filters low-quality samples and that DH-CoT effectively jailbreaks models including GPT-5 and Claude-4, notably outperforming SOTA methods like H-CoT and TAP.", "date": "2025-08-14", "categories": []}
{"id": "2601.03420", "title": "Jailbreaking LLMs Without Gradients or Priors: Effective and Transferable Attacks", "abstract": "As Large Language Models (LLMs) are increasingly deployed in safety-critical domains, rigorously evaluating their robustness against adversarial jailbreaks is essential. However, current safety evaluations often overestimate robustness because existing automated attacks are limited by restrictive assumptions. They typically rely on handcrafted priors or require white-box access for gradient propagation. We challenge these constraints by demonstrating that token-level iterative optimization can succeed without gradients or priors. We introduce RAILS (RAndom Iterative Local Search), a framework that operates solely on model logits. RAILS matches the effectiveness of gradient-based methods through two key innovations: a novel auto-regressive loss that enforces exact prefix matching, and a history-based selection strategy that bridges the gap between the proxy optimization objective and the true attack success rate. Crucially, by eliminating gradient dependency, RAILS enables cross-tokenizer ensemble attacks. This allows for the discovery of shared adversarial patterns that generalize across disjoint vocabularies, significantly enhancing transferability to closed-source systems. Empirically, RAILS achieves near 100% success rates on multiple open-source models and high black-box attack transferability to closed-source systems like GPT and Gemini.", "date": "2026-01-06", "categories": []}
{"id": "2503.04833", "title": "E$^2$AT: Multimodal Jailbreak Defense via Dynamic Joint Optimization for Multimodal Large Language Models", "abstract": "Research endeavors have been made in learning robust Multimodal Large Language Models (MLLMs) against jailbreak attacks. However, existing methods for improving MLLMs' robustness still face critical challenges: \\ding{172} how to efficiently tune massive weight parameters and \\ding{173} how to ensure robustness against attacks across both visual and textual modalities. To this end, we propose an \\textbf{E}fficient \\textbf{E}nd-to-end \\textbf{A}dversarial \\textbf{T}raining (E$^2$AT) framework for both visual and textual adversarial attacks. Specifically, for the visual aspect, E$^2$AT incorporates an efficient projector-based AT module that aligns the attack samples at the feature level. For training objectives, we propose a Dynamic Joint Multimodal Optimization (DJMO) strategy to enhance generalization ability against jailbreak attacks by dynamically adjusting weights between normal and adversarial objectives. Extensive experiments are conducted with five major jailbreak attack methods across three mainstream MLLMs. Results demonstrate that our E$^2$AT achieves the state-of-the-art performance, outperforming existing baselines by an average margin of 34\\% across text and image modalities, while maintaining clean task performance. Furthermore, evaluations of real-world embodied intelligent systems highlight the practical applicability of E$^2$AT, paving the way for the development of more secure and reliable multimodal systems. Our code is available on \\href{https://anonymous.4open.science/r/E2AT_568}{\\textcolor{red}{https://anonymous.4open.science/r/E2AT\\_568}}.", "date": "2025-03-05", "categories": []}
{"id": "2512.10449", "title": "When Reject Turns into Accept: Quantifying the Vulnerability of LLM-Based Scientific Reviewers to Indirect Prompt Injection", "abstract": "Driven by surging submission volumes, scientific peer review has catalyzed two parallel trends: individual over-reliance on LLMs and institutional AI-powered assessment systems. This study investigates the robustness of \"LLM-as-a-Judge\" systems to adversarial PDF manipulation via invisible text injections and layout aware encoding attacks. We specifically target the distinct incentive of flipping \"Reject\" decisions to \"Accept,\" a vulnerability that fundamentally compromises scientific integrity. To measure this, we introduce the Weighted Adversarial Vulnerability Score (WAVS), a novel metric that quantifies susceptibility by weighting score inflation against the severity of decision shifts relative to ground truth. We adapt 15 domain-specific attack strategies, ranging from semantic persuasion to cognitive obfuscation, and evaluate them across 13 diverse language models (including GPT-5 and DeepSeek) using a curated dataset of 200 official and real-world accepted and rejected submissions (e.g., ICLR OpenReview). Our results demonstrate that obfuscation techniques like \"Maximum Mark Magyk\" and \"Symbolic Masking & Context Redirection\" successfully manipulate scores, achieving decision flip rates of up to 86.26% in open-source models, while exposing distinct \"reasoning traps\" in proprietary systems. We release our complete dataset and injection framework to facilitate further research on the topic (https://anonymous.4open.sciencer/llm-jailbreak-FC9E/).", "date": "2025-12-11", "categories": []}
{"id": "2512.21019", "title": "Efficient and Robust Video Defense Framework against 3D-field Personalized Talking Face", "abstract": "State-of-the-art 3D-field video-referenced Talking Face Generation (TFG) methods synthesize high-fidelity personalized talking-face videos in real time by modeling 3D geometry and appearance from reference portrait video. This capability raises significant privacy concerns regarding malicious misuse of personal portraits. However, no efficient defense framework exists to protect such videos against 3D-field TFG methods. While image-based defenses could apply per-frame 2D perturbations, they incur prohibitive computational costs, severe video quality degradation, failing to disrupt 3D information for video protection. To address this, we propose a novel and efficient video defense framework against 3D-field TFG methods, which protects portrait video by perturbing the 3D information acquisition process while maintain high-fidelity video quality. Specifically, our method introduces: (1) a similarity-guided parameter sharing mechanism for computational efficiency, and (2) a multi-scale dual-domain attention module to jointly optimize spatial-frequency perturbations. Extensive experiments demonstrate that our proposed framework exhibits strong defense capability and achieves a 47x acceleration over the fastest baseline while maintaining high fidelity. Moreover, it remains robust against scaling operations and state-of-the-art purification attacks, and the effectiveness of our design choices is further validated through ablation studies. Our project is available at https://github.com/Richen7418/VDF.", "date": "2025-12-24", "categories": []}
{"id": "2601.02700", "title": "Adversarial Question Answering Robustness: A Multi-Level Error Analysis and Mitigation Study", "abstract": "Question answering (QA) systems achieve impressive performance on standard benchmarks like SQuAD, but remain vulnerable to adversarial examples. This project investigates the adversarial robustness of transformer models on the AddSent adversarial dataset through systematic experimentation across model scales and targeted mitigation strategies. We perform comprehensive multi-level error analysis using five complementary categorization schemes, identifying negation confusion and entity substitution as the primary failure modes. Through systematic evaluation of adversarial fine-tuning ratios, we identify 80% clean + 20% adversarial data as optimal. Data augmentation experiments reveal a capacity bottleneck in small models. Scaling from ELECTRA-small (14M parameters) to ELECTRA-base (110M parameters) eliminates the robustness-accuracy trade-off, achieving substantial improvements on both clean and adversarial data. We implement three targeted mitigation strategies, with Entity-Aware contrastive learning achieving best performance: 89.89% AddSent Exact Match (EM) and 90.73% SQuAD EM, representing 94.9% closure of the adversarial gap. To our knowledge, this is the first work integrating comprehensive linguistic error analysis with Named Entity Recognition (NER)-guided contrastive learning for adversarial QA, demonstrating that targeted mitigation can achieve near-parity between clean and adversarial performance.", "date": "2026-01-06", "categories": []}
{"id": "2405.18770", "title": "Multimodal Adversarial Defense for Vision-Language Models by Leveraging One-To-Many Relationships", "abstract": "Pre-trained vision-language (VL) models are highly vulnerable to adversarial attacks. However, existing defense methods primarily focus on image classification, overlooking two key aspects of VL tasks: multimodal attacks, where both image and text can be perturbed, and the one-to-many relationship of images and texts, where a single image can correspond to multiple textual descriptions and vice versa (1:N and N:1). This work is the first to explore defense strategies against multimodal attacks in VL tasks, whereas prior VL defense methods focus on vision robustness. We propose multimodal adversarial training (MAT), which incorporates adversarial perturbations in both image and text modalities during training, significantly outperforming existing unimodal defenses. Furthermore, we discover that MAT is limited by deterministic one-to-one (1:1) image-text pairs in VL training data. To address this, we conduct a comprehensive study on leveraging one-to-many relationships to enhance robustness, investigating diverse augmentation techniques. Our analysis shows that, for a more effective defense, augmented image-text pairs should be well-aligned, diverse, yet avoid distribution shift -- conditions overlooked by prior research. This work pioneers defense strategies against multimodal attacks, providing insights for building robust VLMs from both optimization and data perspectives. Our code is publicly available at https://github.com/CyberAgentAILab/multimodal-adversarial-training.", "date": "2024-05-29", "categories": []}
{"id": "2508.17215", "title": "How to make Medical AI Systems safer? Simulating Vulnerabilities, and Threats in Multimodal Medical RAG System", "abstract": "Large Vision-Language Models (LVLMs) augmented with Retrieval-Augmented Generation (RAG) are increasingly employed in medical AI to enhance factual grounding through external clinical image-text retrieval. However, this reliance creates a significant attack surface. We propose MedThreatRAG, a novel multimodal poisoning framework that systematically probes vulnerabilities in medical RAG systems by injecting adversarial image-text pairs. A key innovation of our approach is the construction of a simulated semi-open attack environment, mimicking real-world medical systems that permit periodic knowledge base updates via user or pipeline contributions. Within this setting, we introduce and emphasize Cross-Modal Conflict Injection (CMCI), which embeds subtle semantic contradictions between medical images and their paired reports. These mismatches degrade retrieval and generation by disrupting cross-modal alignment while remaining sufficiently plausible to evade conventional filters. While basic textual and visual attacks are included for completeness, CMCI demonstrates the most severe degradation. Evaluations on IU-Xray and MIMIC-CXR QA tasks show that MedThreatRAG reduces answer F1 scores by up to 27.66% and lowers LLaVA-Med-1.5 F1 rates to as low as 51.36%. Our findings expose fundamental security gaps in clinical RAG systems and highlight the urgent need for threat-aware design and robust multimodal consistency checks. Finally, we conclude with a concise set of guidelines to inform the safe development of future multimodal medical RAG systems.", "date": "2025-08-24", "categories": []}
{"id": "2601.00968", "title": "Explainability-Guided Defense: Attribution-Aware Model Refinement Against Adversarial Data Attacks", "abstract": "The growing reliance on deep learning models in safety-critical domains such as healthcare and autonomous navigation underscores the need for defenses that are both robust to adversarial perturbations and transparent in their decision-making. In this paper, we identify a connection between interpretability and robustness that can be directly leveraged during training. Specifically, we observe that spurious, unstable, or semantically irrelevant features identified through Local Interpretable Model-Agnostic Explanations (LIME) contribute disproportionately to adversarial vulnerability. Building on this insight, we introduce an attribution-guided refinement framework that transforms LIME from a passive diagnostic into an active training signal. Our method systematically suppresses spurious features using feature masking, sensitivity-aware regularization, and adversarial augmentation in a closed-loop refinement pipeline. This approach does not require additional datasets or model architectures and integrates seamlessly into standard adversarial training. Theoretically, we derive an attribution-aware lower bound on adversarial distortion that formalizes the link between explanation alignment and robustness. Empirical evaluations on CIFAR-10, CIFAR-10-C, and CIFAR-100 demonstrate substantial improvements in adversarial robustness and out-of-distribution generalization.", "date": "2026-01-02", "categories": []}
{"id": "2601.01202", "title": "RefSR-Adv: Adversarial Attack on Reference-based Image Super-Resolution Models", "abstract": "Single Image Super-Resolution (SISR) aims to recover high-resolution images from low-resolution inputs. Unlike SISR, Reference-based Super-Resolution (RefSR) leverages an additional high-resolution reference image to facilitate the recovery of high-frequency textures. However, existing research mainly focuses on backdoor attacks targeting RefSR, while the vulnerability of the adversarial attacks targeting RefSR has not been fully explored. To fill this research gap, we propose RefSR-Adv, an adversarial attack that degrades SR outputs by perturbing only the reference image. By maximizing the difference between adversarial and clean outputs, RefSR-Adv induces significant performance degradation and generates severe artifacts across CNN, Transformer, and Mamba architectures on the CUFED5, WR-SR, and DRefSR datasets. Importantly, experiments confirm a positive correlation between the similarity of the low-resolution input and the reference image and attack effectiveness, revealing that the model's over-reliance on reference features is a key security flaw. This study reveals a security vulnerability in RefSR systems, aiming to urge researchers to pay attention to the robustness of RefSR.", "date": "2026-01-03", "categories": []}
{"id": "2601.02193", "title": "Learning with Monotone Adversarial Corruptions", "abstract": "We study the extent to which standard machine learning algorithms rely on exchangeability and independence of data by introducing a monotone adversarial corruption model. In this model, an adversary, upon looking at a \"clean\" i.i.d. dataset, inserts additional \"corrupted\" points of their choice into the dataset. These added points are constrained to be monotone corruptions, in that they get labeled according to the ground-truth target function. Perhaps surprisingly, we demonstrate that in this setting, all known optimal learning algorithms for binary classification can be made to achieve suboptimal expected error on a new independent test point drawn from the same distribution as the clean dataset. On the other hand, we show that uniform convergence-based algorithms do not degrade in their guarantees. Our results showcase how optimal learning algorithms break down in the face of seemingly helpful monotone corruptions, exposing their overreliance on exchangeability.", "date": "2026-01-05", "categories": []}
{"id": "2601.00270", "title": "Rectifying Adversarial Examples Using Their Vulnerabilities", "abstract": "Deep neural network-based classifiers are prone to errors when processing adversarial examples (AEs). AEs are minimally perturbed input data undetectable to humans posing significant risks to security-dependent applications. Hence, extensive research has been undertaken to develop defense mechanisms that mitigate their threats. Most existing methods primarily focus on discriminating AEs based on the input sample features, emphasizing AE detection without addressing the correct sample categorization before an attack. While some tasks may only require mere rejection on detected AEs, others necessitate identifying the correct original input category such as traffic sign recognition in autonomous driving. The objective of this study is to propose a method for rectifying AEs to estimate the correct labels of their original inputs. Our method is based on re-attacking AEs to move them beyond the decision boundary for accurate label prediction, effectively addressing the issue of rectifying minimally perceptible AEs created using white-box attack methods. However, challenge remains with respect to effectively rectifying AEs produced by black-box attacks at a distance from the boundary, or those misclassified into low-confidence categories by targeted attacks. By adopting a straightforward approach of only considering AEs as inputs, the proposed method can address diverse attacks while avoiding the requirement of parameter adjustments or preliminary training. Results demonstrate that the proposed method exhibits consistent performance in rectifying AEs generated via various attack methods, including targeted and black-box attacks. Moreover, it outperforms conventional rectification and input transformation methods in terms of stability against various attacks.", "date": "2026-01-01", "categories": []}
{"id": "2601.00367", "title": "PatchBlock: A Lightweight Defense Against Adversarial Patches for Embedded EdgeAI Devices", "abstract": "Adversarial attacks pose a significant challenge to the reliable deployment of machine learning models in EdgeAI applications, such as autonomous driving and surveillance, which rely on resource-constrained devices for real-time inference. Among these, patch-based adversarial attacks, where small malicious patches (e.g., stickers) are applied to objects, can deceive neural networks into making incorrect predictions with potentially severe consequences. In this paper, we present PatchBlock, a lightweight framework designed to detect and neutralize adversarial patches in images. Leveraging outlier detection and dimensionality reduction, PatchBlock identifies regions affected by adversarial noise and suppresses their impact. It operates as a pre-processing module at the sensor level, efficiently running on CPUs in parallel with GPU inference, thus preserving system throughput while avoiding additional GPU overhead. The framework follows a three-stage pipeline: splitting the input into chunks (Chunking), detecting anomalous regions via a redesigned isolation forest with targeted cuts for faster convergence (Separating), and applying dimensionality reduction on the identified outliers (Mitigating). PatchBlock is both model- and patch-agnostic, can be retrofitted to existing pipelines, and integrates seamlessly between sensor inputs and downstream models. Evaluations across multiple neural architectures, benchmark datasets, attack types, and diverse edge devices demonstrate that PatchBlock consistently improves robustness, recovering up to 77% of model accuracy under strong patch attacks such as the Google Adversarial Patch, while maintaining high portability and minimal clean accuracy loss. Additionally, PatchBlock outperforms the state-of-the-art defenses in efficiency, in terms of computation time and energy consumption per sample, making it suitable for EdgeAI applications.", "date": "2026-01-01", "categories": []}
{"id": "2601.00577", "title": "Adversarial Samples Are Not Created Equal", "abstract": "Over the past decade, numerous theories have been proposed to explain the widespread vulnerability of deep neural networks to adversarial evasion attacks. Among these, the theory of non-robust features proposed by Ilyas et al. has been widely accepted, showing that brittle but predictive features of the data distribution can be directly exploited by attackers. However, this theory overlooks adversarial samples that do not directly utilize these features. In this work, we advocate that these two kinds of samples - those which use use brittle but predictive features and those that do not - comprise two types of adversarial weaknesses and should be differentiated when evaluating adversarial robustness. For this purpose, we propose an ensemble-based metric to measure the manipulation of non-robust features by adversarial perturbations and use this metric to analyze the makeup of adversarial samples generated by attackers. This new perspective also allows us to re-examine multiple phenomena, including the impact of sharpness-aware minimization on adversarial robustness and the robustness gap observed between adversarially training and standard training on robust datasets.", "date": "2026-01-02", "categories": []}
{"id": "2503.08990", "title": "Effective and Efficient Jailbreaks of Black-Box LLMs with Cross-Behavior Attacks", "abstract": "Despite recent advancements in Large Language Models (LLMs) and their alignment, they can still be jailbroken, i.e., harmful and toxic content can be elicited from them. While existing red-teaming methods have shown promise in uncovering such vulnerabilities, these methods struggle with limited success and high computational and monetary costs. To address this, we propose a black-box Jailbreak method with Cross-Behavior attacks (JCB), that can automatically and efficiently find successful jailbreak prompts. JCB leverages successes from past behaviors to help jailbreak new behaviors, thereby significantly improving the attack efficiency. Moreover, JCB does not rely on time- and/or cost-intensive calls to auxiliary LLMs to discover/optimize the jailbreak prompts, making it highly efficient and scalable. Comprehensive experimental evaluations show that JCB significantly outperforms related baselines, requiring up to 94% fewer queries while still achieving 12.9% higher average attack success. JCB also achieves a notably high 37% attack success rate on Llama-2-7B, one of the most resilient LLMs, and shows promising zero-shot transferability across different LLMs.", "date": "2025-03-12", "categories": []}
{"id": "2512.23837", "title": "Adversarial Lens: Exploiting Attention Layers to Generate Adversarial Examples for Evaluation", "abstract": "Recent advances in mechanistic interpretability suggest that intermediate attention layers encode token-level hypotheses that are iteratively refined toward the final output. In this work, we exploit this property to generate adversarial examples directly from attention-layer token distributions. Unlike prompt-based or gradient-based attacks, our approach leverages model-internal token predictions, producing perturbations that are both plausible and internally consistent with the model's own generation process. We evaluate whether tokens extracted from intermediate layers can serve as effective adversarial perturbations for downstream evaluation tasks. We conduct experiments on argument quality assessment using the ArgQuality dataset, with LLaMA-3.1-Instruct-8B serving as both the generator and evaluator. Our results show that attention-based adversarial examples lead to measurable drops in evaluation performance while remaining semantically similar to the original inputs. However, we also observe that substitutions drawn from certain layers and token positions can introduce grammatical degradation, limiting their practical effectiveness. Overall, our findings highlight both the promise and current limitations of using intermediate-layer representations as a principled source of adversarial examples for stress-testing LLM-based evaluation pipelines.", "date": "2025-12-29", "categories": []}
{"id": "2512.23881", "title": "Breaking Audio Large Language Models by Attacking Only the Encoder: A Universal Targeted Latent-Space Audio Attack", "abstract": "Audio-language models combine audio encoders with large language models to enable multimodal reasoning, but they also introduce new security vulnerabilities. We propose a universal targeted latent space attack, an encoder-level adversarial attack that manipulates audio latent representations to induce attacker-specified outputs in downstream language generation. Unlike prior waveform-level or input-specific attacks, our approach learns a universal perturbation that generalizes across inputs and speakers and does not require access to the language model. Experiments on Qwen2-Audio-7B-Instruct demonstrate consistently high attack success rates with minimal perceptual distortion, revealing a critical and previously underexplored attack surface at the encoder level of multimodal systems.", "date": "2025-12-29", "categories": []}
{"id": "2512.23953", "title": "T2VAttack: Adversarial Attack on Text-to-Video Diffusion Models", "abstract": "The rapid evolution of Text-to-Video (T2V) diffusion models has driven remarkable advancements in generating high-quality, temporally coherent videos from natural language descriptions. Despite these achievements, their vulnerability to adversarial attacks remains largely unexplored. In this paper, we introduce T2VAttack, a comprehensive study of adversarial attacks on T2V diffusion models from both semantic and temporal perspectives. Considering the inherently dynamic nature of video data, we propose two distinct attack objectives: a semantic objective to evaluate video-text alignment and a temporal objective to assess the temporal dynamics. To achieve an effective and efficient attack process, we propose two adversarial attack methods: (i) T2VAttack-S, which identifies semantically or temporally critical words in prompts and replaces them with synonyms via greedy search, and (ii) T2VAttack-I, which iteratively inserts optimized words with minimal perturbation to the prompt. By combining these objectives and strategies, we conduct a comprehensive evaluation on the adversarial robustness of several state-of-the-art T2V models, including ModelScope, CogVideoX, Open-Sora, and HunyuanVideo. Our experiments reveal that even minor prompt modifications, such as the substitution or insertion of a single word, can cause substantial degradation in semantic fidelity and temporal dynamics, highlighting critical vulnerabilities in current T2V diffusion models.", "date": "2025-12-30", "categories": []}
{"id": "2512.24111", "title": "Guided Diffusion-based Generation of Adversarial Objects for Real-World Monocular Depth Estimation Attacks", "abstract": "Monocular Depth Estimation (MDE) serves as a core perception module in autonomous driving systems, but it remains highly susceptible to adversarial attacks. Errors in depth estimation may propagate through downstream decision making and influence overall traffic safety. Existing physical attacks primarily rely on texture-based patches, which impose strict placement constraints and exhibit limited realism, thereby reducing their effectiveness in complex driving environments. To overcome these limitations, this work introduces a training-free generative adversarial attack framework that generates naturalistic, scene-consistent adversarial objects via a diffusion-based conditional generation process. The framework incorporates a Salient Region Selection module that identifies regions most influential to MDE and a Jacobian Vector Product Guidance mechanism that steers adversarial gradients toward update directions supported by the pre-trained diffusion model. This formulation enables the generation of physically plausible adversarial objects capable of inducing substantial adversarial depth shifts. Extensive digital and physical experiments demonstrate that our method significantly outperforms existing attacks in effectiveness, stealthiness, and physical deployability, underscoring its strong practical implications for autonomous driving safety assessment.", "date": "2025-12-30", "categories": []}
{"id": "2512.24499", "title": "Training-Free Color-Aware Adversarial Diffusion Sanitization for Diffusion Stegomalware Defense at Security Gateways", "abstract": "The rapid expansion of generative AI has normalized large-scale synthetic media creation, enabling new forms of covert communication. Recent generative steganography methods, particularly those based on diffusion models, can embed high-capacity payloads without fine-tuning or auxiliary decoders, creating significant challenges for detection and remediation. Coverless diffusion-based techniques are difficult to counter because they generate image carriers directly from secret data, enabling attackers to deliver stegomalware for command-and-control, payload staging, and data exfiltration while bypassing detectors that rely on cover-stego discrepancies. This work introduces Adversarial Diffusion Sanitization (ADS), a training-free defense for security gateways that neutralizes hidden payloads rather than detecting them. ADS employs an off-the-shelf pretrained denoiser as a differentiable proxy for diffusion-based decoders and incorporates a color-aware, quaternion-coupled update rule to reduce artifacts under strict distortion limits. Under a practical threat model and in evaluation against the state-of-the-art diffusion steganography method Pulsar, ADS drives decoder success rates to near zero with minimal perceptual impact. Results demonstrate that ADS provides a favorable security-utility trade-off compared to standard content transformations, offering an effective mitigation strategy against diffusion-driven steganography.", "date": "2025-12-30", "categories": []}
{"id": "2512.24564", "title": "CPR: Causal Physiological Representation Learning for Robust ECG Analysis under Distribution Shifts", "abstract": "Deep learning models for Electrocardiogram (ECG) diagnosis have achieved remarkable accuracy but exhibit fragility against adversarial perturbations, particularly Smooth Adversarial Perturbations (SAP) that mimic biological morphology. Existing defenses face a critical dilemma: Adversarial Training (AT) provides robustness but incurs a prohibitive computational burden, while certified methods like Randomized Smoothing (RS) introduce significant inference latency, rendering them impractical for real-time clinical monitoring. We posit that this vulnerability stems from the models' reliance on non-robust spurious correlations rather than invariant pathological features. To address this, we propose Causal Physiological Representation Learning (CPR). Unlike standard denoising approaches that operate without semantic constraints, CPR incorporates a Physiological Structural Prior within a causal disentanglement framework. By modeling ECG generation via a Structural Causal Model (SCM), CPR enforces a structural intervention that strictly separates invariant pathological morphology (P-QRS-T complex) from non-causal artifacts. Empirical results on PTB-XL demonstrate that CPR significantly outperforms standard clinical preprocessing methods. Specifically, under SAP attacks, CPR achieves an F1 score of 0.632, surpassing Median Smoothing (0.541 F1) by 9.1%. Crucially, CPR matches the certified robustness of Randomized Smoothing while maintaining single-pass inference efficiency, offering a superior trade-off between robustness, efficiency, and clinical interpretability.", "date": "2025-12-31", "categories": []}
{"id": "2512.24665", "title": "HeteroHBA: A Generative Structure-Manipulating Backdoor Attack on Heterogeneous Graphs", "abstract": "Heterogeneous graph neural networks (HGNNs) have achieved strong performance in many real-world applications, yet targeted backdoor poisoning on heterogeneous graphs remains less studied. We consider backdoor attacks for heterogeneous node classification, where an adversary injects a small set of trigger nodes and connections during training to force specific victim nodes to be misclassified into an attacker-chosen label at test time while preserving clean performance. We propose HeteroHBA, a generative backdoor framework that selects influential auxiliary neighbors for trigger attachment via saliency-based screening and synthesizes diverse trigger features and connection patterns to better match the local heterogeneous context. To improve stealthiness, we combine Adaptive Instance Normalization (AdaIN) with a Maximum Mean Discrepancy (MMD) loss to align the trigger feature distribution with benign statistics, thereby reducing detectability, and we optimize the attack with a bilevel objective that jointly promotes attack success and maintains clean accuracy. Experiments on multiple real-world heterogeneous graphs with representative HGNN architectures show that HeteroHBA consistently achieves higher attack success than prior backdoor baselines with comparable or smaller impact on clean accuracy; moreover, the attack remains effective under our heterogeneity-aware structural defense, CSD. These results highlight practical backdoor risks in heterogeneous graph learning and motivate the development of stronger defenses.", "date": "2025-12-31", "categories": []}
{"id": "2512.24792", "title": "Projection-based Adversarial Attack using Physics-in-the-Loop Optimization for Monocular Depth Estimation", "abstract": "Deep neural networks (DNNs) remain vulnerable to adversarial attacks that cause misclassification when specific perturbations are added to input images. This vulnerability also threatens the reliability of DNN-based monocular depth estimation (MDE) models, making robustness enhancement a critical need in practical applications. To validate the vulnerability of DNN-based MDE models, this study proposes a projection-based adversarial attack method that projects perturbation light onto a target object. The proposed method employs physics-in-the-loop (PITL) optimization -- evaluating candidate solutions in actual environments to account for device specifications and disturbances -- and utilizes a distributed covariance matrix adaptation evolution strategy. Experiments confirmed that the proposed method successfully created adversarial examples that lead to depth misestimations, resulting in parts of objects disappearing from the target scene.", "date": "2025-12-31", "categories": []}
{"id": "2412.10321", "title": "AdvPrefix: An Objective for Nuanced LLM Jailbreaks", "abstract": "Many jailbreak attacks on large language models (LLMs) rely on a common objective: making the model respond with the prefix ``Sure, here is (harmful request)''. While straightforward, this objective has two limitations: limited control over model behaviors, yielding incomplete or unrealistic jailbroken responses, and a rigid format that hinders optimization. We introduce AdvPrefix, a plug-and-play prefix-forcing objective that selects one or more model-dependent prefixes by combining two criteria: high prefilling attack success rates and low negative log-likelihood. AdvPrefix integrates seamlessly into existing jailbreak attacks to mitigate the previous limitations for free. For example, replacing GCG's default prefixes on Llama-3 improves nuanced attack success rates from 14% to 80%, revealing that current safety alignment fails to generalize to new prefixes. Code and selected prefixes are released at github.com/facebookresearch/jailbreak-objectives.", "date": "2024-12-13", "categories": []}
{"id": "2502.01386", "title": "Topic-FlipRAG: Topic-Orientated Adversarial Opinion Manipulation Attacks to Retrieval-Augmented Generation Models", "abstract": "Retrieval-Augmented Generation (RAG) systems based on Large Language Models (LLMs) have become essential for tasks such as question answering and content generation. However, their increasing impact on public opinion and information dissemination has made them a critical focus for security research due to inherent vulnerabilities. Previous studies have predominantly addressed attacks targeting factual or single-query manipulations. In this paper, we address a more practical scenario: topic-oriented adversarial opinion manipulation attacks on RAG models, where LLMs are required to reason and synthesize multiple perspectives, rendering them particularly susceptible to systematic knowledge poisoning. Specifically, we propose Topic-FlipRAG, a two-stage manipulation attack pipeline that strategically crafts adversarial perturbations to influence opinions across related queries. This approach combines traditional adversarial ranking attack techniques and leverages the extensive internal relevant knowledge and reasoning capabilities of LLMs to execute semantic-level perturbations. Experiments show that the proposed attacks effectively shift the opinion of the model's outputs on specific topics, significantly impacting user information perception. Current mitigation methods cannot effectively defend against such attacks, highlighting the necessity for enhanced safeguards for RAG systems, and offering crucial insights for LLM security research.", "date": "2025-02-03", "categories": []}
{"id": "2509.22836", "title": "Seeing Isn't Believing: Context-Aware Adversarial Patch Synthesis via Conditional GAN", "abstract": "Adversarial patch attacks pose a severe threat to deep neural networks, yet most existing approaches rely on unrealistic white-box assumptions, untargeted objectives, or produce visually conspicuous patches that limit real-world applicability. In this work, we introduce a novel framework for fully controllable adversarial patch generation, where the attacker can freely choose both the input image x and the target class y target, thereby dictating the exact misclassification outcome. Our method combines a generative U-Net design with Grad-CAM-guided patch placement, enabling semantic-aware localization that maximizes attack effectiveness while preserving visual realism. Extensive experiments across convolutional networks (DenseNet-121, ResNet-50) and vision transformers (ViT-B/16, Swin-B/16, among others) demonstrate that our approach achieves state-of-the-art performance across all settings, with attack success rates (ASR) and target-class success (TCS) consistently exceeding 99%.\n  Importantly, we show that our method not only outperforms prior white-box attacks and untargeted baselines, but also surpasses existing non-realistic approaches that produce detectable artifacts. By simultaneously ensuring realism, targeted control, and black-box applicability-the three most challenging dimensions of patch-based attacks-our framework establishes a new benchmark for adversarial robustness research, bridging the gap between theoretical attack strength and practical stealthiness.", "date": "2025-09-26", "categories": []}
{"id": "2509.24408", "title": "FuncPoison: Poisoning Function Library to Hijack Multi-agent Autonomous Driving Systems", "abstract": "Autonomous driving systems increasingly rely on multi-agent architectures powered by large language models (LLMs), where specialized agents collaborate to perceive, reason, and plan. A key component of these systems is the shared function library, a collection of software tools that agents use to process sensor data and navigate complex driving environments. Despite its critical role in agent decision-making, the function library remains an under-explored vulnerability. In this paper, we introduce FuncPoison, a novel poisoning-based attack targeting the function library to manipulate the behavior of LLM-driven multi-agent autonomous systems. FuncPoison exploits two key weaknesses in how agents access the function library: (1) agents rely on text-based instructions to select tools; and (2) these tools are activated using standardized command formats that attackers can replicate. By injecting malicious tools with deceptive instructions, FuncPoison manipulates one agent s decisions--such as misinterpreting road conditions--triggering cascading errors that mislead other agents in the system. We experimentally evaluate FuncPoison on two representative multi-agent autonomous driving systems, demonstrating its ability to significantly degrade trajectory accuracy, flexibly target specific agents to induce coordinated misbehavior, and evade diverse defense mechanisms. Our results reveal that the function library, often considered a simple toolset, can serve as a critical attack surface in LLM-based autonomous driving systems, raising elevated concerns on their reliability.", "date": "2025-09-29", "categories": []}
{"id": "2512.17367", "title": "Adversarially Robust Detection of Harmful Online Content: A Computational Design Science Approach", "abstract": "Social media platforms are plagued by harmful content such as hate speech, misinformation, and extremist rhetoric. Machine learning (ML) models are widely adopted to detect such content; however, they remain highly vulnerable to adversarial attacks, wherein malicious users subtly modify text to evade detection. Enhancing adversarial robustness is therefore essential, requiring detectors that can defend against diverse attacks (generalizability) while maintaining high overall accuracy. However, simultaneously achieving both optimal generalizability and accuracy is challenging. Following the computational design science paradigm, this study takes a sequential approach that first proposes a novel framework (Large Language Model-based Sample Generation and Aggregation, LLM-SGA) by identifying the key invariances of textual adversarial attacks and leveraging them to ensure that a detector instantiated within the framework has strong generalizability. Second, we instantiate our detector (Adversarially Robust Harmful Online Content Detector, ARHOCD) with three novel design components to improve detection accuracy: (1) an ensemble of multiple base detectors that exploits their complementary strengths; (2) a novel weight assignment method that dynamically adjusts weights based on each sample's predictability and each base detector's capability, with weights initialized using domain knowledge and updated via Bayesian inference; and (3) a novel adversarial training strategy that iteratively optimizes both the base detectors and the weight assignor. We addressed several limitations of existing adversarial robustness enhancement research and empirically evaluated ARHOCD across three datasets spanning hate speech, rumor, and extremist content. Results show that ARHOCD offers strong generalizability and improves detection accuracy under adversarial conditions.", "date": "2025-12-19", "categories": []}
{"id": "2512.22381", "title": "PHANTOM: Physics-Aware Adversarial Attacks against Federated Learning-Coordinated EV Charging Management System", "abstract": "The rapid deployment of electric vehicle charging stations (EVCS) within distribution networks necessitates intelligent and adaptive control to maintain the grid's resilience and reliability. In this work, we propose PHANTOM, a physics-aware adversarial network that is trained and optimized through a multi-agent reinforcement learning model. PHANTOM integrates a physics-informed neural network (PINN) enabled by federated learning (FL) that functions as a digital twin of EVCS-integrated systems, ensuring physically consistent modeling of operational dynamics and constraints. Building on this digital twin, we construct a multi-agent RL environment that utilizes deep Q-networks (DQN) and soft actor-critic (SAC) methods to derive adversarial false data injection (FDI) strategies capable of bypassing conventional detection mechanisms. To examine the broader grid-level consequences, a transmission and distribution (T and D) dual simulation platform is developed, allowing us to capture cascading interactions between EVCS disturbances at the distribution level and the operations of the bulk transmission system. Results demonstrate how learned attack policies disrupt load balancing and induce voltage instabilities that propagate across T and D boundaries. These findings highlight the critical need for physics-aware cybersecurity to ensure the resilience of large-scale vehicle-grid integration.", "date": "2025-12-26", "categories": []}
{"id": "2512.21404", "title": "LLM-Driven Feature-Level Adversarial Attacks on Android Malware Detectors", "abstract": "The rapid growth in both the scale and complexity of Android malware has driven the widespread adoption of machine learning (ML) techniques for scalable and accurate malware detection. Despite their effectiveness, these models remain vulnerable to adversarial attacks that introduce carefully crafted feature-level perturbations to evade detection while preserving malicious functionality. In this paper, we present LAMLAD, a novel adversarial attack framework that exploits the generative and reasoning capabilities of large language models (LLMs) to bypass ML-based Android malware classifiers. LAMLAD employs a dual-agent architecture composed of an LLM manipulator, which generates realistic and functionality-preserving feature perturbations, and an LLM analyzer, which guides the perturbation process toward successful evasion. To improve efficiency and contextual awareness, LAMLAD integrates retrieval-augmented generation (RAG) into the LLM pipeline. Focusing on Drebin-style feature representations, LAMLAD enables stealthy and high-confidence attacks against widely deployed Android malware detection systems. We evaluate LAMLAD against three representative ML-based Android malware detectors and compare its performance with two state-of-the-art adversarial attack methods. Experimental results demonstrate that LAMLAD achieves an attack success rate (ASR) of up to 97%, requiring on average only three attempts per adversarial sample, highlighting its effectiveness, efficiency, and adaptability in practical adversarial settings. Furthermore, we propose an adversarial training-based defense strategy that reduces the ASR by more than 30% on average, significantly enhancing model robustness against LAMLAD-style attacks.", "date": "2025-12-24", "categories": []}
{"id": "2512.22069", "title": "Scaling Adversarial Training via Data Selection", "abstract": "Projected Gradient Descent (PGD) is a strong and widely used first-order adversarial attack, yet its computational cost scales poorly, as all training samples undergo identical iterative inner-loop optimization despite contributing unequally to robustness. Motivated by this inefficiency, we propose \\emph{Selective Adversarial Training}, which perturbs only a subset of critical samples in each minibatch. Specifically, we introduce two principled selection criteria: (1) margin-based sampling, which prioritizes samples near the decision boundary, and (2) gradient-matching sampling, which selects samples whose gradients align with the dominant batch optimization direction. Adversarial examples are generated only for the selected subset, while the remaining samples are trained cleanly using a mixed objective. Experiments on MNIST and CIFAR-10 show that the proposed methods achieve robustness comparable to, or even exceeding, full PGD adversarial training, while reducing adversarial computation by up to $50\\%$, demonstrating that informed sample selection is sufficient for scalable adversarial robustness.", "date": "2025-12-26", "categories": []}
{"id": "2507.01020", "title": "AutoAdv: Automated Adversarial Prompting for Multi-Turn Jailbreaking of Large Language Models", "abstract": "Large Language Models (LLMs) continue to exhibit vulnerabilities to jailbreaking attacks: carefully crafted malicious inputs intended to circumvent safety guardrails and elicit harmful responses. As such, we present AutoAdv, a novel framework that automates adversarial prompt generation to systematically evaluate and expose vulnerabilities in LLM safety mechanisms. Our approach leverages a parametric attacker LLM to produce semantically disguised malicious prompts through strategic rewriting techniques, specialized system prompts, and optimized hyperparameter configurations. The primary contribution of our work is a dynamic, multi-turn attack methodology that analyzes failed jailbreak attempts and iteratively generates refined follow-up prompts, leveraging techniques such as roleplaying, misdirection, and contextual manipulation. We quantitatively evaluate attack success rate (ASR) using the StrongREJECT (arXiv:2402.10260 [cs.CL]) framework across sequential interaction turns. Through extensive empirical evaluation of state-of-the-art models--including ChatGPT, Llama, and DeepSeek--we reveal significant vulnerabilities, with our automated attacks achieving jailbreak success rates of up to 86% for harmful content generation. Our findings reveal that current safety mechanisms remain susceptible to sophisticated multi-turn attacks, emphasizing the urgent need for more robust defense strategies.", "date": "2025-04-18", "categories": []}
{"id": "2512.20893", "title": "Time-Efficient Evaluation and Enhancement of Adversarial Robustness in Deep Neural Networks", "abstract": "With deep neural networks (DNNs) increasingly embedded in modern society, ensuring their safety has become a critical and urgent issue. In response, substantial efforts have been dedicated to the red-blue adversarial framework, where the red team focuses on identifying vulnerabilities in DNNs and the blue team on mitigating them. However, existing approaches from both teams remain computationally intensive, constraining their applicability to large-scale models. To overcome this limitation, this thesis endeavours to provide time-efficient methods for the evaluation and enhancement of adversarial robustness in DNNs.", "date": "2025-12-24", "categories": []}
{"id": "2512.21241", "title": "Improving the Convergence Rate of Ray Search Optimization for Query-Efficient Hard-Label Attacks", "abstract": "In hard-label black-box adversarial attacks, where only the top-1 predicted label is accessible, the prohibitive query complexity poses a major obstacle to practical deployment. In this paper, we focus on optimizing a representative class of attacks that search for the optimal ray direction yielding the minimum $\\ell_2$-norm perturbation required to move a benign image into the adversarial region. Inspired by Nesterov's Accelerated Gradient (NAG), we propose a momentum-based algorithm, ARS-OPT, which proactively estimates the gradient with respect to a future ray direction inferred from accumulated momentum. We provide a theoretical analysis of its convergence behavior, showing that ARS-OPT enables more accurate directional updates and achieves faster, more stable optimization. To further accelerate convergence, we incorporate surrogate-model priors into ARS-OPT's gradient estimation, resulting in PARS-OPT with enhanced performance. The superiority of our approach is supported by theoretical guarantees under standard assumptions. Extensive experiments on ImageNet and CIFAR-10 demonstrate that our method surpasses 13 state-of-the-art approaches in query efficiency.", "date": "2025-12-24", "categories": []}
{"id": "2505.02824", "title": "Towards Dataset Copyright Evasion Attack against Personalized Text-to-Image Diffusion Models", "abstract": "Text-to-image (T2I) diffusion models enable high-quality image generation conditioned on textual prompts. However, fine-tuning these pre-trained models for personalization raises concerns about unauthorized dataset usage. To address this issue, dataset ownership verification (DOV) has recently been proposed, which embeds watermarks into fine-tuning datasets via backdoor techniques. These watermarks remain dormant on benign samples but produce owner-specified outputs when triggered. Despite its promise, the robustness of DOV against copyright evasion attacks (CEA) remains unexplored. In this paper, we investigate how adversaries can circumvent these mechanisms, enabling models trained on watermarked datasets to bypass ownership verification. We begin by analyzing the limitations of potential attacks achieved by backdoor removal, including TPD and T2IShield. In practice, TPD suffers from inconsistent effectiveness due to randomness, while T2IShield fails when watermarks are embedded as local image patches. To this end, we introduce CEAT2I, the first CEA specifically targeting DOV in T2I diffusion models. CEAT2I consists of three stages: (1) motivated by the observation that T2I models converge faster on watermarked samples with respect to intermediate features rather than training loss, we reliably detect watermarked samples; (2) we iteratively ablate tokens from the prompts of detected samples and monitor feature shifts to identify trigger tokens; and (3) we apply a closed-form concept erasure method to remove the injected watermarks. Extensive experiments demonstrate that CEAT2I effectively evades state-of-the-art DOV mechanisms while preserving model performance. The code is available at https://github.com/csyufei/CEAT2I.", "date": "2025-05-05", "categories": []}
{"id": "2512.19711", "title": "PHANTOM: PHysical ANamorphic Threats Obstructing Connected Vehicle Mobility", "abstract": "Connected autonomous vehicles (CAVs) rely on vision-based deep neural networks (DNNs) and low-latency (Vehicle-to-Everything) V2X communication to navigate safely and efficiently. Despite their advances, these systems remain vulnerable to physical adversarial attacks. In this paper, we introduce PHANTOM (PHysical ANamorphic Threats Obstructing connected vehicle Mobility), a novel framework for crafting and deploying perspective-dependent adversarial examples using \\textit{anamorphic art}. PHANTOM exploits geometric distortions that appear natural to humans but are misclassified with high confidence by state-of-the-art object detectors. Unlike conventional attacks, PHANTOM operates in black-box settings without model access and demonstrates strong transferability across four diverse detector architectures (YOLOv5, SSD, Faster R-CNN, and RetinaNet). Comprehensive evaluation in CARLA across varying speeds, weather conditions, and lighting scenarios shows that PHANTOM achieves over 90\\% attack success rate under optimal conditions and maintains 60-80\\% effectiveness even in degraded environments. The attack activates within 6-10 meters of the target, providing insufficient time for safe maneuvering. Beyond individual vehicle deception, PHANTOM triggers network-wide disruption in CAV systems: SUMO-OMNeT++ co-simulation demonstrates that false emergency messages propagate through V2X links, increasing Peak Age of Information by 68-89\\% and degrading safety-critical communication. These findings expose critical vulnerabilities in both perception and communication layers of CAV ecosystems.", "date": "2025-12-15", "categories": []}
{"id": "2512.19935", "title": "Conditional Adversarial Fragility in Financial Machine Learning under Macroeconomic Stress", "abstract": "Machine learning models used in financial decision systems operate in nonstationary economic environments, yet adversarial robustness is typically evaluated under static assumptions. This work introduces Conditional Adversarial Fragility, a regime dependent phenomenon in which adversarial vulnerability is systematically amplified during periods of macroeconomic stress. We propose a regime aware evaluation framework for time indexed tabular financial classification tasks that conditions robustness assessment on external indicators of economic stress. Using volatility based regime segmentation as a proxy for macroeconomic conditions, we evaluate model behavior across calm and stress periods while holding model architecture, attack methodology, and evaluation protocols constant. Baseline predictive performance remains comparable across regimes, indicating that economic stress alone does not induce inherent performance degradation. Under adversarial perturbations, however, models operating during stress regimes exhibit substantially greater degradation across predictive accuracy, operational decision thresholds, and risk sensitive outcomes. We further demonstrate that this amplification propagates to increased false negative rates, elevating the risk of missed high risk cases during adverse conditions. To complement numerical robustness metrics, we introduce an interpretive governance layer based on semantic auditing of model explanations using large language models. Together, these results demonstrate that adversarial robustness in financial machine learning is a regime dependent property and motivate stress aware approaches to model risk assessment in high stakes financial deployments.", "date": "2025-12-22", "categories": []}
{"id": "2512.20004", "title": "IoT-based Android Malware Detection Using Graph Neural Network With Adversarial Defense", "abstract": "Since the Internet of Things (IoT) is widely adopted using Android applications, detecting malicious Android apps is essential. In recent years, Android graph-based deep learning research has proposed many approaches to extract relationships from applications as graphs to generate graph embeddings. First, we demonstrate the effectiveness of graph-based classification using a Graph Neural Network (GNN)-based classifier to generate API graph embeddings. The graph embeddings are combined with Permission and Intent features to train multiple machine learning and deep learning models for Android malware detection. The proposed classification approach achieves an accuracy of 98.33 percent on the CICMaldroid dataset and 98.68 percent on the Drebin dataset. However, graph-based deep learning models are vulnerable, as attackers can add fake relationships to evade detection by the classifier. Second, we propose a Generative Adversarial Network (GAN)-based attack algorithm named VGAE-MalGAN targeting graph-based GNN Android malware classifiers. The VGAE-MalGAN generator produces adversarial malware API graphs, while the VGAE-MalGAN substitute detector attempts to mimic the target detector. Experimental results show that VGAE-MalGAN can significantly reduce the detection rate of GNN-based malware classifiers. Although the model initially fails to detect adversarial malware, retraining with generated adversarial samples improves robustness and helps mitigate adversarial attacks.", "date": "2025-12-23", "categories": []}
{"id": "2512.20168", "title": "Odysseus: Jailbreaking Commercial Multimodal LLM-integrated Systems via Dual Steganography", "abstract": "By integrating language understanding with perceptual modalities such as images, multimodal large language models (MLLMs) constitute a critical substrate for modern AI systems, particularly intelligent agents operating in open and interactive environments. However, their increasing accessibility also raises heightened risks of misuse, such as generating harmful or unsafe content. To mitigate these risks, alignment techniques are commonly applied to align model behavior with human values. Despite these efforts, recent studies have shown that jailbreak attacks can circumvent alignment and elicit unsafe outputs. Currently, most existing jailbreak methods are tailored for open-source models and exhibit limited effectiveness against commercial MLLM-integrated systems, which often employ additional filters. These filters can detect and prevent malicious input and output content, significantly reducing jailbreak threats. In this paper, we reveal that the success of these safety filters heavily relies on a critical assumption that malicious content must be explicitly visible in either the input or the output. This assumption, while often valid for traditional LLM-integrated systems, breaks down in MLLM-integrated systems, where attackers can leverage multiple modalities to conceal adversarial intent, leading to a false sense of security in existing MLLM-integrated systems. To challenge this assumption, we propose Odysseus, a novel jailbreak paradigm that introduces dual steganography to covertly embed malicious queries and responses into benign-looking images. Extensive experiments on benchmark datasets demonstrate that our Odysseus successfully jailbreaks several pioneering and realistic MLLM-integrated systems, achieving up to 99% attack success rate. It exposes a fundamental blind spot in existing defenses, and calls for rethinking cross-modal security in MLLM-integrated systems.", "date": "2025-12-23", "categories": []}
{"id": "2311.02757", "title": "Certified Defense on the Fairness of Graph Neural Networks", "abstract": "Graph Neural Networks (GNNs) have emerged as a prominent graph learning model in various graph-based tasks over the years. Nevertheless, due to the vulnerabilities of GNNs, it has been empirically shown that malicious attackers could easily corrupt the fairness level of their predictions by adding perturbations to the input graph data. In this paper, we take crucial steps to study a novel problem of certifiable defense on the fairness level of GNNs. Specifically, we propose a principled framework named ELEGANT and present a detailed theoretical certification analysis for the fairness of GNNs. ELEGANT takes {\\em any} GNN as its backbone, and the fairness level of such a backbone is theoretically impossible to be corrupted under certain perturbation budgets for attackers. Notably, ELEGANT does not make any assumptions over the GNN structure or parameters, and does not require re-training the GNNs to realize certification. Hence it can serve as a plug-and-play framework for any optimized GNNs ready to be deployed. We verify the satisfactory effectiveness of ELEGANT in practice through extensive experiments on real-world datasets across different backbones of GNNs and parameter settings.", "date": "2023-11-05", "categories": []}
{"id": "2406.05491", "title": "One Perturbation is Enough: On Generating Universal Adversarial Perturbations against Vision-Language Pre-training Models", "abstract": "Vision-Language Pre-training (VLP) models have exhibited unprecedented capability in many applications by taking full advantage of the multimodal alignment. However, previous studies have shown they are vulnerable to maliciously crafted adversarial samples. Despite recent success, these methods are generally instance-specific and require generating perturbations for each input sample. In this paper, we reveal that VLP models are also vulnerable to the instance-agnostic universal adversarial perturbation (UAP). Specifically, we design a novel Contrastive-training Perturbation Generator with Cross-modal conditions (C-PGC) to achieve the attack. In light that the pivotal multimodal alignment is achieved through the advanced contrastive learning technique, we devise to turn this powerful weapon against themselves, i.e., employ a malicious version of contrastive learning to train the C-PGC based on our carefully crafted positive and negative image-text pairs for essentially destroying the alignment relationship learned by VLP models. Besides, C-PGC fully utilizes the characteristics of Vision-and-Language (V+L) scenarios by incorporating both unimodal and cross-modal information as effective guidance. Extensive experiments show that C-PGC successfully forces adversarial samples to move away from their original area in the VLP model's feature space, thus essentially enhancing attacks across various victim models and V+L tasks. The GitHub repository is available at https://github.com/ffhibnese/CPGC_VLP_Universal_Attacks.", "date": "2024-06-08", "categories": []}
{"id": "2503.11185", "title": "Bleeding Pathways: Vanishing Discriminability in LLM Hidden States Fuels Jailbreak Attacks", "abstract": "LLMs remain vulnerable to jailbreak attacks that exploit adversarial prompts to circumvent safety measures. Current safety fine-tuning approaches face two critical limitations. First, they often fail to strike a balance between security and utility, where stronger safety measures tend to over-reject harmless user requests. Second, they frequently miss malicious intent concealed within seemingly benign tasks, leaving models exposed to exploitation. Our work identifies a fundamental cause of these issues: during response generation, an LLM's capacity to differentiate harmful from safe outputs deteriorates. Experimental evidence confirms this, revealing that the separability between hidden states for safe and harmful responses diminishes as generation progresses. This weakening discrimination forces models to make compliance judgments earlier in the generation process, restricting their ability to recognize developing harmful intent and contributing to both aforementioned failures. To mitigate this vulnerability, we introduce DEEPALIGN - an inherent defense framework that enhances the safety of LLMs. By applying contrastive hidden-state steering at the midpoint of response generation, DEEPALIGN amplifies the separation between harmful and benign hidden states, enabling continuous intrinsic toxicity detection and intervention throughout the generation process. Across diverse LLMs spanning varying architectures and scales, it reduced attack success rates of nine distinct jailbreak attacks to near-zero or minimal. Crucially, it preserved model capability while reducing over-refusal. Models equipped with DEEPALIGN exhibited up to 3.5% lower error rates in rejecting challenging benign queries and maintained standard task performance with less than 1% decline. This marks a substantial advance in the safety-utility Pareto frontier.", "date": "2025-03-14", "categories": []}
{"id": "2511.02376", "title": "AutoAdv: Automated Adversarial Prompting for Multi-Turn Jailbreaking of Large Language Models", "abstract": "Large Language Models (LLMs) remain vulnerable to jailbreaking attacks where adversarial prompts elicit harmful outputs. Yet most evaluations focus on single-turn interactions while real-world attacks unfold through adaptive multi-turn conversations. We present AutoAdv, a training-free framework for automated multi-turn jailbreaking that achieves an attack success rate of up to 95% on Llama-3.1-8B within six turns, a 24% improvement over single-turn baselines. AutoAdv uniquely combines three adaptive mechanisms: a pattern manager that learns from successful attacks to enhance future prompts, a temperature manager that dynamically adjusts sampling parameters based on failure modes, and a two-phase rewriting strategy that disguises harmful requests and then iteratively refines them. Extensive evaluation across commercial and open-source models (Llama-3.1-8B, GPT-4o mini, Qwen3-235B, Mistral-7B) reveals persistent vulnerabilities in current safety mechanisms, with multi-turn attacks consistently outperforming single-turn approaches. These findings demonstrate that alignment strategies optimized for single-turn interactions fail to maintain robustness across extended conversations, highlighting an urgent need for multi-turn-aware defenses.", "date": "2025-11-04", "categories": []}
{"id": "2511.09392", "title": "Potent but Stealthy: Rethink Profile Pollution against Sequential Recommendation via Bi-level Constrained Reinforcement Paradigm", "abstract": "Sequential Recommenders, which exploit dynamic user intents through interaction sequences, is vulnerable to adversarial attacks. While existing attacks primarily rely on data poisoning, they require large-scale user access or fake profiles thus lacking practicality. In this paper, we focus on the Profile Pollution Attack that subtly contaminates partial user interactions to induce targeted mispredictions. Previous PPA methods suffer from two limitations, i.e., i) over-reliance on sequence horizon impact restricts fine-grained perturbations on item transitions, and ii) holistic modifications cause detectable distribution shifts. To address these challenges, we propose a constrained reinforcement driven attack CREAT that synergizes a bi-level optimization framework with multi-reward reinforcement learning to balance adversarial efficacy and stealthiness. We first develop a Pattern Balanced Rewarding Policy, which integrates pattern inversion rewards to invert critical patterns and distribution consistency rewards to minimize detectable shifts via unbalanced co-optimal transport. Then we employ a Constrained Group Relative Reinforcement Learning paradigm, enabling step-wise perturbations through dynamic barrier constraints and group-shared experience replay, achieving targeted pollution with minimal detectability. Extensive experiments demonstrate the effectiveness of CREAT.", "date": "2025-11-12", "categories": []}
{"id": "2512.18264", "title": "Who Can See Through You? Adversarial Shielding Against VLM-Based Attribute Inference Attacks", "abstract": "As vision-language models (VLMs) become widely adopted, VLM-based attribute inference attacks have emerged as a serious privacy concern, enabling adversaries to infer private attributes from images shared on social media. This escalating threat calls for dedicated protection methods to safeguard user privacy. However, existing methods often degrade the visual quality of images or interfere with vision-based functions on social media, thereby failing to achieve a desirable balance between privacy protection and user experience. To address this challenge, we propose a novel protection method that jointly optimizes privacy suppression and utility preservation under a visual consistency constraint. While our method is conceptually effective, fair comparisons between methods remain challenging due to the lack of publicly available evaluation datasets. To fill this gap, we introduce VPI-COCO, a publicly available benchmark comprising 522 images with hierarchically structured privacy questions and corresponding non-private counterparts, enabling fine-grained and joint evaluation of protection methods in terms of privacy preservation and user experience. Building upon this benchmark, experiments on multiple VLMs demonstrate that our method effectively reduces PAR below 25%, keeps NPAR above 88%, maintains high visual consistency, and generalizes well to unseen and paraphrased privacy questions, demonstrating its strong practical applicability for real-world VLM deployments.", "date": "2025-12-20", "categories": []}
{"id": "2512.18651", "title": "Adversarial Robustness in Zero-Shot Learning:An Empirical Study on Class and Concept-Level Vulnerabilities", "abstract": "Zero-shot Learning (ZSL) aims to enable image classifiers to recognize images from unseen classes that were not included during training. Unlike traditional supervised classification, ZSL typically relies on learning a mapping from visual features to predefined, human-understandable class concepts. While ZSL models promise to improve generalization and interpretability, their robustness under systematic input perturbations remain unclear. In this study, we present an empirical analysis about the robustness of existing ZSL methods at both classlevel and concept-level. Specifically, we successfully disrupted their class prediction by the well-known non-target class attack (clsA). However, in the Generalized Zero-shot Learning (GZSL) setting, we observe that the success of clsA is only at the original best-calibrated point. After the attack, the optimal bestcalibration point shifts, and ZSL models maintain relatively strong performance at other calibration points, indicating that clsA results in a spurious attack success in the GZSL. To address this, we propose the Class-Bias Enhanced Attack (CBEA), which completely eliminates GZSL accuracy across all calibrated points by enhancing the gap between seen and unseen class probabilities.Next, at concept-level attack, we introduce two novel attack modes: Class-Preserving Concept Attack (CPconA) and NonClass-Preserving Concept Attack (NCPconA). Our extensive experiments evaluate three typical ZSL models across various architectures from the past three years and reveal that ZSL models are vulnerable not only to the traditional class attack but also to concept-based attacks. These attacks allow malicious actors to easily manipulate class predictions by erasing or introducing concepts. Our findings highlight a significant performance gap between existing approaches, emphasizing the need for improved adversarial robustness in current ZSL models.", "date": "2025-12-21", "categories": []}
{"id": "2512.18755", "title": "MEEA: Mere Exposure Effect-Driven Confrontational Optimization for LLM Jailbreaking", "abstract": "The rapid advancement of large language models (LLMs) has intensified concerns about the robustness of their safety alignment. While existing jailbreak studies explore both single-turn and multi-turn strategies, most implicitly assume a static safety boundary and fail to account for how contextual interactions dynamically influence model behavior, leading to limited stability and generalization. Motivated by this gap, we propose MEEA (Mere Exposure Effect Attack), a psychology-inspired, fully automated black-box framework for evaluating multi-turn safety robustness, grounded in the mere exposure effect. MEEA leverages repeated low-toxicity semantic exposure to induce a gradual shift in a model's effective safety threshold, enabling progressive erosion of alignment constraints over sustained interactions. Concretely, MEEA constructs semantically progressive prompt chains and optimizes them using a simulated annealing strategy guided by semantic similarity, toxicity, and jailbreak effectiveness. Extensive experiments on both closed-source and open-source models, including GPT-4, Claude-3.5, and DeepSeek-R1, demonstrate that MEEA consistently achieves higher attack success rates than seven representative baselines, with an average Attack Success Rate (ASR) improvement exceeding 20%. Ablation studies further validate the necessity of both annealing-based optimization and contextual exposure mechanisms. Beyond improved attack effectiveness, our findings indicate that LLM safety behavior is inherently dynamic and history-dependent, challenging the common assumption of static alignment boundaries and highlighting the need for interaction-aware safety evaluation and defense mechanisms. Our code is available at: https://github.com/Carney-lsz/MEEA", "date": "2025-12-21", "categories": []}
{"id": "2512.19317", "title": "SafeMed-R1: Adversarial Reinforcement Learning for Generalizable and Robust Medical Reasoning in Vision-Language Models", "abstract": "Vision--Language Models (VLMs) show significant promise for Medical Visual Question Answering (VQA), yet their deployment in clinical settings is hindered by severe vulnerability to adversarial attacks. Standard adversarial training, while effective for simpler tasks, often degrades both generalization performance and the quality of generated clinical reasoning. We introduce SafeMed-R1, a hybrid defense framework that ensures robust performance while preserving high-quality, interpretable medical reasoning. SafeMed-R1 employs a two-stage approach: at training time, we integrate Adversarial Training with Group Relative Policy Optimization (AT-GRPO) to explicitly robustify the reasoning process against worst-case perturbations; at inference time, we augment the model with Randomized Smoothing to provide certified $L_2$-norm robustness guarantees. We evaluate SafeMed-R1 on the OmniMedVQA benchmark across eight medical imaging modalities comprising over 88,000 samples. Our experiments reveal that standard fine-tuned VLMs, despite achieving 95\\% accuracy on clean inputs, collapse to approximately 25\\% under PGD attacks. In contrast, SafeMed-R1 maintains 84.45\\% accuracy under the same adversarial conditions, representing a 59 percentage point improvement in robustness. Furthermore, we demonstrate that models trained with explicit chain-of-thought reasoning exhibit superior adversarial robustness compared to instruction-only variants, suggesting a synergy between interpretability and security in medical AI systems.", "date": "2025-12-22", "categories": []}
{"id": "2511.10008", "title": "Phantom Menace: Exploring and Enhancing the Robustness of VLA Models Against Physical Sensor Attacks", "abstract": "Vision-Language-Action (VLA) models revolutionize robotic systems by enabling end-to-end perception-to-action pipelines that integrate multiple sensory modalities, such as visual signals processed by cameras and auditory signals captured by microphones. This multi-modality integration allows VLA models to interpret complex, real-world environments using diverse sensor data streams. Given the fact that VLA-based systems heavily rely on the sensory input, the security of VLA models against physical-world sensor attacks remains critically underexplored. To address this gap, we present the first systematic study of physical sensor attacks against VLAs, quantifying the influence of sensor attacks and investigating the defenses for VLA models. We introduce a novel \"Real-Sim-Real\" framework that automatically simulates physics-based sensor attack vectors, including six attacks targeting cameras and two targeting microphones, and validates them on real robotic systems. Through large-scale evaluations across various VLA architectures and tasks under varying attack parameters, we demonstrate significant vulnerabilities, with susceptibility patterns that reveal critical dependencies on task types and model designs. We further develop an adversarial-training-based defense that enhances VLA robustness against out-of-distribution physical perturbations caused by sensor attacks while preserving model performance. Our findings expose an urgent need for standardized robustness benchmarks and mitigation strategies to secure VLA deployments in safety-critical environments.", "date": "2025-11-13", "categories": []}
{"id": "2512.17029", "title": "Adversarial VR: An Open-Source Testbed for Evaluating Adversarial Robustness of VR Cybersickness Detection and Mitigation", "abstract": "Deep learning (DL)-based automated cybersickness detection methods, along with adaptive mitigation techniques, can enhance user comfort and interaction. However, recent studies show that these DL-based systems are susceptible to adversarial attacks; small perturbations to sensor inputs can degrade model performance, trigger incorrect mitigation, and disrupt the user's immersive experience (UIX). Additionally, there is a lack of dedicated open-source testbeds that evaluate the robustness of these systems under adversarial conditions, limiting the ability to assess their real-world effectiveness. To address this gap, this paper introduces Adversarial-VR, a novel real-time VR testbed for evaluating DL-based cybersickness detection and mitigation strategies under adversarial conditions. Developed in Unity, the testbed integrates two state-of-the-art (SOTA) DL models: DeepTCN and Transformer, which are trained on the open-source MazeSick dataset, for real-time cybersickness severity detection and applies a dynamic visual tunneling mechanism that adjusts the field-of-view based on model outputs. To assess robustness, we incorporate three SOTA adversarial attacks: MI-FGSM, PGD, and C&W, which successfully prevent cybersickness mitigation by fooling DL-based cybersickness models' outcomes. We implement these attacks using a testbed with a custom-built VR Maze simulation and an HTC Vive Pro Eye headset, and we open-source our implementation for widespread adoption by VR developers and researchers. Results show that these adversarial attacks are capable of successfully fooling the system. For instance, the C&W attack results in a $5.94x decrease in accuracy for the Transformer-based cybersickness model compared to the accuracy without the attack.", "date": "2025-12-18", "categories": []}
{"id": "2512.17146", "title": "Biosecurity-Aware AI: Agentic Risk Auditing of Soft Prompt Attacks on ESM-Based Variant Predictors", "abstract": "Genomic Foundation Models (GFMs), such as Evolutionary Scale Modeling (ESM), have demonstrated remarkable success in variant effect prediction. However, their security and robustness under adversarial manipulation remain largely unexplored. To address this gap, we introduce the Secure Agentic Genomic Evaluator (SAGE), an agentic framework for auditing the adversarial vulnerabilities of GFMs. SAGE functions through an interpretable and automated risk auditing loop. It injects soft prompt perturbations, monitors model behavior across training checkpoints, computes risk metrics such as AUROC and AUPR, and generates structured reports with large language model-based narrative explanations. This agentic process enables continuous evaluation of embedding-space robustness without modifying the underlying model. Using SAGE, we find that even state-of-the-art GFMs like ESM2 are sensitive to targeted soft prompt attacks, resulting in measurable performance degradation. These findings reveal critical and previously hidden vulnerabilities in genomic foundation models, showing the importance of agentic risk auditing in securing biomedical applications such as clinical variant interpretation.", "date": "2025-12-19", "categories": []}
{"id": "2512.17902", "title": "Adversarial Robustness of Vision in Open Foundation Models", "abstract": "With the increase in deep learning, it becomes increasingly difficult to understand the model in which AI systems can identify objects. Thus, an adversary could aim to modify an image by adding unseen elements, which will confuse the AI in its recognition of an entity. This paper thus investigates the adversarial robustness of LLaVA-1.5-13B and Meta's Llama 3.2 Vision-8B-2. These are tested for untargeted PGD (Projected Gradient Descent) against the visual input modality, and empirically evaluated on the Visual Question Answering (VQA) v2 dataset subset. The results of these adversarial attacks are then quantified using the standard VQA accuracy metric. This evaluation is then compared with the accuracy degradation (accuracy drop) of LLaVA and Llama 3.2 Vision. A key finding is that Llama 3.2 Vision, despite a lower baseline accuracy in this setup, exhibited a smaller drop in performance under attack compared to LLaVA, particularly at higher perturbation levels. Overall, the findings confirm that the vision modality represents a viable attack vector for degrading the performance of contemporary open-weight VLMs, including Meta's Llama 3.2 Vision. Furthermore, they highlight that adversarial robustness does not necessarily correlate directly with standard benchmark performance and may be influenced by underlying architectural and training factors.", "date": "2025-12-19", "categories": []}
{"id": "2411.16782", "title": "Scaling Laws for Black box Adversarial Attacks", "abstract": "Adversarial examples exhibit cross-model transferability, enabling threatening black-box attacks on commercial models. Model ensembling, which attacks multiple surrogate models, is a known strategy to improve this transferability. However, prior studies typically use small, fixed ensembles, which leaves open an intriguing question of whether scaling the number of surrogate models can further improve black-box attacks. In this work, we conduct the first large-scale empirical study of this question. We show that by resolving gradient conflict with advanced optimizers, we discover a robust and universal log-linear scaling law through both theoretical analysis and empirical evaluations: the Attack Success Rate (ASR) scales linearly with the logarithm of the ensemble size $T$. We rigorously verify this law across standard classifiers, SOTA defenses, and MLLMs, and find that scaling distills robust, semantic features of the target class. Consequently, we apply this fundamental insight to benchmark SOTA MLLMs. This reveals both the attack's devastating power and a clear robustness hierarchy: we achieve 80\\%+ transfer attack success rate on proprietary models like GPT-4o, while also highlighting the exceptional resilience of Claude-3.5-Sonnet. Our findings urge a shift in focus for robustness evaluation: from designing intricate algorithms on small ensembles to understanding the principled and powerful threat of scaling.", "date": "2024-11-25", "categories": []}
{"id": "2508.17361", "title": "Trust Me, I Know This Function: Hijacking LLM Static Analysis using Bias", "abstract": "Large Language Models (LLMs) are increasingly trusted to perform automated code review and static analysis at scale, supporting tasks such as vulnerability detection, summarization, and refactoring. In this paper, we identify and exploit a critical vulnerability in LLM-based code analysis: an abstraction bias that causes models to overgeneralize familiar programming patterns and overlook small, meaningful bugs. Adversaries can exploit this blind spot to hijack the control flow of the LLM's interpretation with minimal edits and without affecting actual runtime behavior. We refer to this attack as a Familiar Pattern Attack (FPA).\n  We develop a fully automated, black-box algorithm that discovers and injects FPAs into target code. Our evaluation shows that FPAs are not only effective against basic and reasoning models, but are also transferable across model families (OpenAI, Anthropic, Google), and universal across programming languages (Python, C, Rust, Go). Moreover, FPAs remain effective even when models are explicitly warned about the attack via robust system prompts. Finally, we explore positive, defensive uses of FPAs and discuss their broader implications for the reliability and safety of code-oriented LLMs.", "date": "2025-08-24", "categories": []}
{"id": "2511.13654", "title": "Tuning for Two Adversaries: Enhancing the Robustness Against Transfer and Query-Based Attacks using Hyperparameter Tuning", "abstract": "In this paper, we present the first detailed analysis of how training hyperparameters -- such as learning rate, weight decay, momentum, and batch size -- influence robustness against both transfer-based and query-based attacks. Supported by theory and experiments, our study spans a variety of practical deployment settings, including centralized training, ensemble learning, and distributed training. We uncover a striking dichotomy: for transfer-based attacks, decreasing the learning rate significantly enhances robustness by up to $64\\%$. In contrast, for query-based attacks, increasing the learning rate consistently leads to improved robustness by up to $28\\%$ across various settings and data distributions. Leveraging these findings, we explore -- for the first time -- the training hyperparameter space to jointly enhance robustness against both transfer-based and query-based attacks. Our results reveal that distributed models benefit the most from hyperparameter tuning, achieving a remarkable tradeoff by simultaneously mitigating both attack types more effectively than other training setups.", "date": "2025-11-17", "categories": []}
{"id": "2512.15780", "title": "Adversarial Robustness in Financial Machine Learning: Defenses, Economic Impact, and Governance Evidence", "abstract": "We evaluate adversarial robustness in tabular machine learning models used in financial decision making. Using credit scoring and fraud detection data, we apply gradient based attacks and measure impacts on discrimination, calibration, and financial risk metrics. Results show notable performance degradation under small perturbations and partial recovery through adversarial training.", "date": "2025-12-14", "categories": []}
{"id": "2512.16123", "title": "Autoencoder-based Denoising Defense against Adversarial Attacks on Object Detection", "abstract": "Deep learning-based object detection models play a critical role in real-world applications such as autonomous driving and security surveillance systems, yet they remain vulnerable to adversarial examples. In this work, we propose an autoencoder-based denoising defense to recover object detection performance degraded by adversarial perturbations. We conduct adversarial attacks using Perlin noise on vehicle-related images from the COCO dataset, apply a single-layer convolutional autoencoder to remove the perturbations, and evaluate detection performance using YOLOv5. Our experiments demonstrate that adversarial attacks reduce bbox mAP from 0.2890 to 0.1640, representing a 43.3% performance degradation. After applying the proposed autoencoder defense, bbox mAP improves to 0.1700 (3.7% recovery) and bbox mAP@50 increases from 0.2780 to 0.3080 (10.8% improvement). These results indicate that autoencoder-based denoising can provide partial defense against adversarial attacks without requiring model retraining.", "date": "2025-12-18", "categories": []}
{"id": "2406.10090", "title": "Over-parameterization and Adversarial Robustness in Neural Networks: An Overview and Empirical Analysis", "abstract": "Thanks to their extensive capacity, over-parameterized neural networks exhibit superior predictive capabilities and generalization. However, having a large parameter space is considered one of the main suspects of the neural networks' vulnerability to adversarial example -- input samples crafted ad-hoc to induce a desired misclassification. Relevant literature has claimed contradictory remarks in support of and against the robustness of over-parameterized networks. These contradictory findings might be due to the failure of the attack employed to evaluate the networks' robustness. Previous research has demonstrated that depending on the considered model, the algorithm employed to generate adversarial examples may not function properly, leading to overestimating the model's robustness. In this work, we empirically study the robustness of over-parameterized networks against adversarial examples. However, unlike the previous works, we also evaluate the considered attack's reliability to support the results' veracity. Our results show that over-parameterized networks are robust against adversarial attacks as opposed to their under-parameterized counterparts.", "date": "2024-06-14", "categories": []}
{"id": "2411.10029", "title": "Toward Robust and Accurate Adversarial Camouflage Generation against Vehicle Detectors", "abstract": "Adversarial camouflage is a widely used physical attack against vehicle detectors for its superiority in multi-view attack performance. One promising approach involves using differentiable neural renderers to facilitate adversarial camouflage optimization through gradient back-propagation. However, existing methods often struggle to capture environmental characteristics during the rendering process or produce adversarial textures that can precisely map to the target vehicle. Moreover, these approaches neglect diverse weather conditions, reducing the efficacy of generated camouflage across varying weather scenarios. To tackle these challenges, we propose a robust and accurate camouflage generation method, namely RAUCA. The core of RAUCA is a novel neural rendering component, End-to-End Neural Renderer Plus (E2E-NRP), which can accurately optimize and project vehicle textures and render images with environmental characteristics such as lighting and weather. In addition, we integrate a multi-weather dataset for camouflage generation, leveraging the E2E-NRP to enhance the attack robustness. Experimental results on six popular object detectors show that RAUCA-final outperforms existing methods in both simulation and real-world settings.", "date": "2024-11-15", "categories": []}
{"id": "2506.09677", "title": "Benchmarking Gaslighting Negation Attacks Against Reasoning Models", "abstract": "Recent advances in reasoning-centric models promise improved robustness through mechanisms such as chain-of-thought prompting and test-time scaling. However, their ability to withstand gaslighting negation attacks-adversarial prompts that confidently deny correct answers-remains underexplored. In this paper, we conduct a systematic evaluation of three state-of-the-art reasoning models, i.e., OpenAI's o4-mini, Claude-3.7-Sonnet and Gemini-2.5-Flash, across three multimodal benchmarks: MMMU, MathVista, and CharXiv. Our evaluation reveals significant accuracy drops (25-29% on average) following gaslighting negation attacks, indicating that even top-tier reasoning models struggle to preserve correct answers under manipulative user feedback. Built upon the insights of the evaluation and to further probe this vulnerability, we introduce GaslightingBench-R, a new diagnostic benchmark specifically designed to evaluate reasoning models' susceptibility to defend their belief under gaslighting negation attacks. Constructed by filtering and curating 1,025 challenging samples from the existing benchmarks, GaslightingBench-R induces even more dramatic failures, with accuracy drops exceeding 53% on average. Our findings highlight a fundamental gap between step-by-step reasoning and resistance to adversarial manipulation, calling for new robustness strategies that safeguard reasoning models against gaslighting negation attacks.", "date": "2025-06-11", "categories": []}
{"id": "2512.14751", "title": "One Leak Away: How Pretrained Model Exposure Amplifies Jailbreak Risks in Finetuned LLMs", "abstract": "Finetuning pretrained large language models (LLMs) has become the standard paradigm for developing downstream applications. However, its security implications remain unclear, particularly regarding whether finetuned LLMs inherit jailbreak vulnerabilities from their pretrained sources. We investigate this question in a realistic pretrain-to-finetune threat model, where the attacker has white-box access to the pretrained LLM and only black-box access to its finetuned derivatives. Empirical analysis shows that adversarial prompts optimized on the pretrained model transfer most effectively to its finetuned variants, revealing inherited vulnerabilities from pretrained to finetuned LLMs. To further examine this inheritance, we conduct representation-level probing, which shows that transferable prompts are linearly separable within the pretrained hidden states, suggesting that universal transferability is encoded in pretrained representations. Building on this insight, we propose the Probe-Guided Projection (PGP) attack, which steers optimization toward transferability-relevant directions. Experiments across multiple LLM families and diverse finetuned tasks confirm PGP's strong transfer success, underscoring the security risks inherent in the pretrain-to-finetune paradigm.", "date": "2025-12-14", "categories": []}
{"id": "2512.15123", "title": "TrajSyn: Privacy-Preserving Dataset Distillation from Federated Model Trajectories for Server-Side Adversarial Training", "abstract": "Deep learning models deployed on edge devices are increasingly used in safety-critical applications. However, their vulnerability to adversarial perturbations poses significant risks, especially in Federated Learning (FL) settings where identical models are distributed across thousands of clients. While adversarial training is a strong defense, it is difficult to apply in FL due to strict client-data privacy constraints and the limited compute available on edge devices. In this work, we introduce TrajSyn, a privacy-preserving framework that enables effective server-side adversarial training by synthesizing a proxy dataset from the trajectories of client model updates, without accessing raw client data. We show that TrajSyn consistently improves adversarial robustness on image classification benchmarks with no extra compute burden on the client device.", "date": "2025-12-17", "categories": []}
{"id": "2512.15353", "title": "Adversarial versification in portuguese as a jailbreak operator in LLMs", "abstract": "Recent evidence shows that the versification of prompts constitutes a highly effective adversarial mechanism against aligned LLMs. The study 'Adversarial poetry as a universal single-turn jailbreak mechanism in large language models' demonstrates that instructions routinely refused in prose become executable when rewritten as verse, producing up to 18 x more safety failures in benchmarks derived from MLCommons AILuminate. Manually written poems reach approximately 62% ASR, and automated versions 43%, with some models surpassing 90% success in single-turn interactions. The effect is structural: systems trained with RLHF, constitutional AI, and hybrid pipelines exhibit consistent degradation under minimal semiotic formal variation. Versification displaces the prompt into sparsely supervised latent regions, revealing guardrails that are excessively dependent on surface patterns. This dissociation between apparent robustness and real vulnerability exposes deep limitations in current alignment regimes. The absence of evaluations in Portuguese, a language with high morphosyntactic complexity, a rich metric-prosodic tradition, and over 250 million speakers, constitutes a critical gap. Experimental protocols must parameterise scansion, metre, and prosodic variation to test vulnerabilities specific to Lusophone patterns, which are currently ignored.", "date": "2025-12-17", "categories": []}
{"id": "2307.07873", "title": "Why Does Little Robustness Help? A Further Step Towards Understanding Adversarial Transferability", "abstract": "Adversarial examples (AEs) for DNNs have been shown to be transferable: AEs that successfully fool white-box surrogate models can also deceive other black-box models with different architectures. Although a bunch of empirical studies have provided guidance on generating highly transferable AEs, many of these findings lack explanations and even lead to inconsistent advice. In this paper, we take a further step towards understanding adversarial transferability, with a particular focus on surrogate aspects. Starting from the intriguing little robustness phenomenon, where models adversarially trained with mildly perturbed adversarial samples can serve as better surrogates, we attribute it to a trade-off between two predominant factors: model smoothness and gradient similarity. Our investigations focus on their joint effects, rather than their separate correlations with transferability. Through a series of theoretical and empirical analyses, we conjecture that the data distribution shift in adversarial training explains the degradation of gradient similarity. Building on these insights, we explore the impacts of data augmentation and gradient regularization on transferability and identify that the trade-off generally exists in the various training mechanisms, thus building a comprehensive blueprint for the regulation mechanism behind transferability. Finally, we provide a general route for constructing better surrogates to boost transferability which optimizes both model smoothness and gradient similarity simultaneously, e.g., the combination of input gradient regularization and sharpness-aware minimization (SAM), validated by extensive experiments. In summary, we call for attention to the united impacts of these two factors for launching effective transfer attacks, rather than optimizing one while ignoring the other, and emphasize the crucial role of manipulating surrogate models.", "date": "2023-07-15", "categories": []}
{"id": "2312.04960", "title": "MIMIR: Masked Image Modeling for Mutual Information-based Adversarial Robustness", "abstract": "Vision Transformers (ViTs) have emerged as a fundamental architecture and serve as the backbone of modern vision-language models. Despite their impressive performance, ViTs exhibit notable vulnerability to evasion attacks, necessitating the development of specialized Adversarial Training (AT) strategies tailored to their unique architecture. While a direct solution might involve applying existing AT methods to ViTs, our analysis reveals significant incompatibilities, particularly with state-of-the-art (SOTA) approaches such as Generalist (CVPR 2023) and DBAT (USENIX Security 2024). This paper presents a systematic investigation of adversarial robustness in ViTs and provides a novel theoretical Mutual Information (MI) analysis in its autoencoder-based self-supervised pre-training. Specifically, we show that MI between the adversarial example and its latent representation in ViT-based autoencoders should be constrained via derived MI bounds. Building on this insight, we propose a self-supervised AT method, MIMIR, that employs an MI penalty to facilitate adversarial pre-training by masked image modeling with autoencoders. Extensive experiments on CIFAR-10, Tiny-ImageNet, and ImageNet-1K show that MIMIR can consistently provide improved natural and robust accuracy, where MIMIR outperforms SOTA AT results on ImageNet-1K. Notably, MIMIR demonstrates superior robustness against unforeseen attacks and common corruption data and can also withstand adaptive attacks where the adversary possesses full knowledge of the defense mechanism. Our code and trained models are publicly available at: https://github.com/xiaoyunxxy/MIMIR.", "date": "2023-12-08", "categories": []}
{"id": "2512.14158", "title": "CIS-BA: Continuous Interaction Space Based Backdoor Attack for Object Detection in the Real-World", "abstract": "Object detection models deployed in real-world applications such as autonomous driving face serious threats from backdoor attacks. Despite their practical effectiveness,existing methods are inherently limited in both capability and robustness due to their dependence on single-trigger-single-object mappings and fragile pixel-level cues. We propose CIS-BA, a novel backdoor attack paradigm that redefines trigger design by shifting from static object features to continuous inter-object interaction patterns that describe how objects co-occur and interact in a scene. By modeling these patterns as a continuous interaction space, CIS-BA introduces space triggers that, for the first time, enable a multi-trigger-multi-object attack mechanism while achieving robustness through invariant geometric relations. To implement this paradigm, we design CIS-Frame, which constructs space triggers via interaction analysis, formalizes them as class-geometry constraints for sample poisoning, and embeds the backdoor during detector training. CIS-Frame supports both single-object attacks (object misclassification and disappearance) and multi-object simultaneous attacks, enabling complex and coordinated effects across diverse interaction states. Experiments on MS-COCO and real-world videos show that CIS-BA achieves over 97% attack success under complex environments and maintains over 95% effectiveness under dynamic multi-trigger conditions, while evading three state-of-the-art defenses. In summary, CIS-BA extends the landscape of backdoor attacks in interaction-intensive scenarios and provides new insights into the security of object detection systems.", "date": "2025-12-16", "categories": []}
{"id": "2512.14188", "title": "Optimizing the Adversarial Perturbation with a Momentum-based Adaptive Matrix", "abstract": "Generating adversarial examples (AEs) can be formulated as an optimization problem. Among various optimization-based attacks, the gradient-based PGD and the momentum-based MI-FGSM have garnered considerable interest. However, all these attacks use the sign function to scale their perturbations, which raises several theoretical concerns from the point of view of optimization. In this paper, we first reveal that PGD is actually a specific reformulation of the projected gradient method using only the current gradient to determine its step-size. Further, we show that when we utilize a conventional adaptive matrix with the accumulated gradients to scale the perturbation, PGD becomes AdaGrad. Motivated by this analysis, we present a novel momentum-based attack AdaMI, in which the perturbation is optimized with an interesting momentum-based adaptive matrix. AdaMI is proved to attain optimal convergence for convex problems, indicating that it addresses the non-convergence issue of MI-FGSM, thereby ensuring stability of the optimization process. The experiments demonstrate that the proposed momentum-based adaptive matrix can serve as a general and effective technique to boost adversarial transferability over the state-of-the-art methods across different networks while maintaining better stability and imperceptibility.", "date": "2025-12-16", "categories": []}
{"id": "2512.14320", "title": "Semantic Mismatch and Perceptual Degradation: A New Perspective on Image Editing Immunity", "abstract": "Text-guided image editing via diffusion models, while powerful, raises significant concerns about misuse, motivating efforts to immunize images against unauthorized edits using imperceptible perturbations. Prevailing metrics for evaluating immunization success typically rely on measuring the visual dissimilarity between the output generated from a protected image and a reference output generated from the unprotected original. This approach fundamentally overlooks the core requirement of image immunization, which is to disrupt semantic alignment with attacker intent, regardless of deviation from any specific output. We argue that immunization success should instead be defined by the edited output either semantically mismatching the prompt or suffering substantial perceptual degradations, both of which thwart malicious intent. To operationalize this principle, we propose Synergistic Intermediate Feature Manipulation (SIFM), a method that strategically perturbs intermediate diffusion features through dual synergistic objectives: (1) maximizing feature divergence from the original edit trajectory to disrupt semantic alignment with the expected edit, and (2) minimizing feature norms to induce perceptual degradations. Furthermore, we introduce the Immunization Success Rate (ISR), a novel metric designed to rigorously quantify true immunization efficacy for the first time. ISR quantifies the proportion of edits where immunization induces either semantic failure relative to the prompt or significant perceptual degradations, assessed via Multimodal Large Language Models (MLLMs). Extensive experiments show our SIFM achieves the state-of-the-art performance for safeguarding visual content against malicious diffusion-based manipulation.", "date": "2025-12-16", "categories": []}
{"id": "2512.14333", "title": "Dual Attention Guided Defense Against Malicious Edits", "abstract": "Recent progress in text-to-image diffusion models has transformed image editing via text prompts, yet this also introduces significant ethical challenges from potential misuse in creating deceptive or harmful content. While current defenses seek to mitigate this risk by embedding imperceptible perturbations, their effectiveness is limited against malicious tampering. To address this issue, we propose a Dual Attention-Guided Noise Perturbation (DANP) immunization method that adds imperceptible perturbations to disrupt the model's semantic understanding and generation process. DANP functions over multiple timesteps to manipulate both cross-attention maps and the noise prediction process, using a dynamic threshold to generate masks that identify text-relevant and irrelevant regions. It then reduces attention in relevant areas while increasing it in irrelevant ones, thereby misguides the edit towards incorrect regions and preserves the intended targets. Additionally, our method maximizes the discrepancy between the injected noise and the model's predicted noise to further interfere with the generation. By targeting both attention and noise prediction mechanisms, DANP exhibits impressive immunity against malicious edits, and extensive experiments confirm that our method achieves state-of-the-art performance.", "date": "2025-12-16", "categories": []}
{"id": "2512.14448", "title": "Reasoning-Style Poisoning of LLM Agents via Stealthy Style Transfer: Process-Level Attacks and Runtime Monitoring in RSV Space", "abstract": "Large Language Model (LLM) agents relying on external retrieval are increasingly deployed in high-stakes environments. While existing adversarial attacks primarily focus on content falsification or instruction injection, we identify a novel, process-oriented attack surface: the agent's reasoning style. We propose Reasoning-Style Poisoning (RSP), a paradigm that manipulates how agents process information rather than what they process. We introduce Generative Style Injection (GSI), an attack method that rewrites retrieved documents into pathological tones--specifically \"analysis paralysis\" or \"cognitive haste\"--without altering underlying facts or using explicit triggers. To quantify these shifts, we develop the Reasoning Style Vector (RSV), a metric tracking Verification depth, Self-confidence, and Attention focus. Experiments on HotpotQA and FEVER using ReAct, Reflection, and Tree of Thoughts (ToT) architectures reveal that GSI significantly degrades performance. It increases reasoning steps by up to 4.4 times or induces premature errors, successfully bypassing state-of-the-art content filters. Finally, we propose RSP-M, a lightweight runtime monitor that calculates RSV metrics in real-time and triggers alerts when values exceed safety thresholds. Our work demonstrates that reasoning style is a distinct, exploitable vulnerability, necessitating process-level defenses beyond static content analysis.", "date": "2025-12-16", "categories": []}
{"id": "2411.12914", "title": "Trojan Cleansing with Neural Collapse", "abstract": "Trojan attacks are sophisticated training-time attacks on neural networks that embed backdoor triggers which force the network to produce a specific output on any input which includes the trigger. With the increasing relevance of deep networks which are too large to train with personal resources and which are trained on data too large to thoroughly audit, these training-time attacks pose a significant risk. In this work, we connect trojan attacks to Neural Collapse, a phenomenon wherein the final feature representations of over-parameterized neural networks converge to a simple geometric structure. We provide experimental evidence that trojan attacks disrupt this convergence for a variety of datasets and architectures. We then use this disruption to design a lightweight, broadly generalizable mechanism for cleansing trojan attacks from a wide variety of different network architectures and experimentally demonstrate its efficacy.", "date": "2024-11-19", "categories": []}
{"id": "2506.21127", "title": "Meta Policy Switching for Secure UAV Deconfliction in Adversarial Airspace", "abstract": "Autonomous UAV navigation using reinforcement learning (RL) is vulnerable to adversarial attacks that manipulate sensor inputs, potentially leading to unsafe behavior and mission failure. Although robust RL methods provide partial protection, they often struggle to generalize to unseen or out-of-distribution (OOD) attacks due to their reliance on fixed perturbation settings. To address this limitation, we propose a meta-policy switching framework in which a meta-level polic dynamically selects among multiple robust policies to counter unknown adversarial shifts. At the core of this framework lies a discounted Thompson sampling (DTS) mechanism that formulates policy selection as a multi-armed bandit problem, thereby minimizing value distribution shifts via self-induced adversarial observations. We first construct a diverse ensemble of action-robust policies trained under varying perturbation intensities. The DTS-based meta-policy then adaptively selects among these policies online, optimizing resilience against self-induced, piecewise-stationary attacks. Theoretical analysis shows that the DTS mechanism minimizes expected regret, ensuring adaptive robustness to OOD attacks and exhibiting emergent antifragile behavior under uncertainty. Extensive simulations in complex 3D obstacle environments under both white-box (Projected Gradient Descent) and black-box (GPS spoofing) attacks demonstrate significantly improved navigation efficiency and higher conflict free trajectory rates compared to standard robust and vanilla RL baselines, highlighting the practical security and dependability benefits of the proposed approach.", "date": "2025-06-26", "categories": []}
{"id": "2509.26032", "title": "Stealthy Yet Effective: Distribution-Preserving Backdoor Attacks on Graph Classification", "abstract": "Graph Neural Networks (GNNs) have demonstrated strong performance across tasks such as node classification, link prediction, and graph classification, but remain vulnerable to backdoor attacks that implant imperceptible triggers during training to control predictions. While node-level attacks exploit local message passing, graph-level attacks face the harder challenge of manipulating global representations while maintaining stealth. We identify two main sources of anomaly in existing graph classification backdoor methods: structural deviation from rare subgraph triggers and semantic deviation caused by label flipping, both of which make poisoned graphs easily detectable by anomaly detection models. To address this, we propose DPSBA, a clean-label backdoor framework that learns in-distribution triggers via adversarial training guided by anomaly-aware discriminators. DPSBA effectively suppresses both structural and semantic anomalies, achieving high attack success while significantly improving stealth. Extensive experiments on real-world datasets validate that DPSBA achieves a superior balance between effectiveness and detectability compared to state-of-the-art baselines.", "date": "2025-09-30", "categories": []}
{"id": "2510.17904", "title": "BreakFun: Jailbreaking LLMs via Schema Exploitation", "abstract": "The proficiency of Large Language Models (LLMs) in processing structured data and adhering to syntactic rules is a capability that drives their widespread adoption but also makes them paradoxically vulnerable. In this paper, we investigate this vulnerability through BreakFun, a jailbreak methodology that weaponizes an LLM's adherence to structured schemas. BreakFun employs a three-part prompt that combines an innocent framing and a Chain-of-Thought distraction with a core \"Trojan Schema\"--a carefully crafted data structure that compels the model to generate harmful content, exploiting the LLM's strong tendency to follow structures and schemas. We demonstrate this vulnerability is highly transferable, achieving an average success rate of 89% across 13 foundational and proprietary models on JailbreakBench, and reaching a 100% Attack Success Rate (ASR) on several prominent models. A rigorous ablation study confirms this Trojan Schema is the attack's primary causal factor. To counter this, we introduce the Adversarial Prompt Deconstruction guardrail, a defense that utilizes a secondary LLM to perform a \"Literal Transcription\"--extracting all human-readable text to isolate and reveal the user's true harmful intent. Our proof-of-concept guardrail demonstrates high efficacy against the attack, validating that targeting the deceptive schema is a viable mitigation strategy. Our work provides a look into how an LLM's core strengths can be turned into critical weaknesses, offering a fresh perspective for building more robustly aligned models.", "date": "2025-10-19", "categories": []}
{"id": "2511.16020", "title": "Physically Realistic Sequence-Level Adversarial Clothing for Robust Human-Detection Evasion", "abstract": "Deep neural networks used for human detection are highly vulnerable to adversarial manipulation, creating safety and privacy risks in real surveillance environments. Wearable attacks offer a realistic threat model, yet existing approaches usually optimize textures frame by frame and therefore fail to maintain concealment across long video sequences with motion, pose changes, and garment deformation. In this work, a sequence-level optimization framework is introduced to generate natural, printable adversarial textures for shirts, trousers, and hats that remain effective throughout entire walking videos in both digital and physical settings. Product images are first mapped to UV space and converted into a compact palette and control-point parameterization, with ICC locking to keep all colors printable. A physically based human-garment pipeline is then employed to simulate motion, multi-angle camera viewpoints, cloth dynamics, and illumination variation. An expectation-over-transformation objective with temporal weighting is used to optimize the control points so that detection confidence is minimized across whole sequences. Extensive experiments demonstrate strong and stable concealment, high robustness to viewpoint changes, and superior cross-model transferability. Physical garments produced with sublimation printing achieve reliable suppression under indoor and outdoor recordings, confirming real-world feasibility.", "date": "2025-11-20", "categories": []}
{"id": "2512.11865", "title": "Explainable Adversarial-Robust Vision-Language-Action Model for Robotic Manipulation", "abstract": "Smart farming has emerged as a key technology for advancing modern agriculture through automation and intelligent control. However, systems relying on RGB cameras for perception and robotic manipulators for control, common in smart farming, are vulnerable to photometric perturbations such as hue, illumination, and noise changes, which can cause malfunction under adversarial attacks. To address this issue, we propose an explainable adversarial-robust Vision-Language-Action model based on the OpenVLA-OFT framework. The model integrates an Evidence-3 module that detects photometric perturbations and generates natural language explanations of their causes and effects. Experiments show that the proposed model reduces Current Action L1 loss by 21.7% and Next Actions L1 loss by 18.4% compared to the baseline, demonstrating improved action prediction accuracy and explainability under adversarial conditions.", "date": "2025-12-05", "categories": []}
{"id": "2512.12154", "title": "Keep the Lights On, Keep the Lengths in Check: Plug-In Adversarial Detection for Time-Series LLMs in Energy Forecasting", "abstract": "Accurate time-series forecasting is increasingly critical for planning and operations in low-carbon power systems. Emerging time-series large language models (TS-LLMs) now deliver this capability at scale, requiring no task-specific retraining, and are quickly becoming essential components within the Internet-of-Energy (IoE) ecosystem. However, their real-world deployment is complicated by a critical vulnerability: adversarial examples (AEs). Detecting these AEs is challenging because (i) adversarial perturbations are optimized across the entire input sequence and exploit global temporal dependencies, which renders local detection methods ineffective, and (ii) unlike traditional forecasting models with fixed input dimensions, TS-LLMs accept sequences of variable length, increasing variability that complicates detection. To address these challenges, we propose a plug-in detection framework that capitalizes on the TS-LLM's own variable-length input capability. Our method uses sampling-induced divergence as a detection signal. Given an input sequence, we generate multiple shortened variants and detect AEs by measuring the consistency of their forecasts: Benign sequences tend to produce stable predictions under sampling, whereas adversarial sequences show low forecast similarity, because perturbations optimized for a full-length sequence do not transfer reliably to shorter, differently-structured subsamples. We evaluate our approach on three representative TS-LLMs (TimeGPT, TimesFM, and TimeLLM) across three energy datasets: ETTh2 (Electricity Transformer Temperature), NI (Hourly Energy Consumption), and Consumption (Hourly Electricity Consumption and Production). Empirical results confirm strong and robust detection performance across both black-box and white-box attack scenarios, highlighting its practicality as a reliable safeguard for TS-LLM forecasting in real-world energy systems.", "date": "2025-12-13", "categories": []}
{"id": "2512.12827", "title": "GradID: Adversarial Detection via Intrinsic Dimensionality of Gradients", "abstract": "Despite their remarkable performance, deep neural networks exhibit a critical vulnerability: small, often imperceptible, adversarial perturbations can lead to drastically altered model predictions. Given the stringent reliability demands of applications such as medical diagnosis and autonomous driving, robust detection of such adversarial attacks is paramount. In this paper, we investigate the geometric properties of a model's input loss landscape. We analyze the Intrinsic Dimensionality (ID) of the model's gradient parameters, which quantifies the minimal number of coordinates required to describe the data points on their underlying manifold. We reveal a distinct and consistent difference in the ID for natural and adversarial data, which forms the basis of our proposed detection method. We validate our approach across two distinct operational scenarios. First, in a batch-wise context for identifying malicious data groups, our method demonstrates high efficacy on datasets like MNIST and SVHN. Second, in the critical individual-sample setting, we establish new state-of-the-art results on challenging benchmarks such as CIFAR-10 and MS COCO. Our detector significantly surpasses existing methods against a wide array of attacks, including CW and AutoAttack, achieving detection rates consistently above 92\\% on CIFAR-10. The results underscore the robustness of our geometric approach, highlighting that intrinsic dimensionality is a powerful fingerprint for adversarial detection across diverse datasets and attack strategies.", "date": "2025-12-14", "categories": []}
{"id": "2512.13416", "title": "Learning to Generate Cross-Task Unexploitable Examples", "abstract": "Unexploitable example generation aims to transform personal images into their unexploitable (unlearnable) versions before they are uploaded online, thereby preventing unauthorized exploitation of online personal images. Recently, this task has garnered significant research attention due to its critical relevance to personal data privacy. Yet, despite recent progress, existing methods for this task can still suffer from limited practical applicability, as they can fail to generate examples that are broadly unexploitable across different real-world computer vision tasks. To deal with this problem, in this work, we propose a novel Meta Cross-Task Unexploitable Example Generation (MCT-UEG) framework. At the core of our framework, to optimize the unexploitable example generator for effectively producing broadly unexploitable examples, we design a flat-minima-oriented meta training and testing scheme. Extensive experiments show the efficacy of our framework.", "date": "2025-12-15", "categories": []}
{"id": "2512.13501", "title": "Behavior-Aware and Generalizable Defense Against Black-Box Adversarial Attacks for ML-Based IDS", "abstract": "Machine learning based intrusion detection systems are increasingly targeted by black box adversarial attacks, where attackers craft evasive inputs using indirect feedback such as binary outputs or behavioral signals like response time and resource usage. While several defenses have been proposed, including input transformation, adversarial training, and surrogate detection, they often fall short in practice. Most are tailored to specific attack types, require internal model access, or rely on static mechanisms that fail to generalize across evolving attack strategies. Furthermore, defenses such as input transformation can degrade intrusion detection system performance, making them unsuitable for real time deployment.\n  To address these limitations, we propose Adaptive Feature Poisoning, a lightweight and proactive defense mechanism designed specifically for realistic black box scenarios. Adaptive Feature Poisoning assumes that probing can occur silently and continuously, and introduces dynamic and context aware perturbations to selected traffic features, corrupting the attacker feedback loop without impacting detection capabilities. The method leverages traffic profiling, change point detection, and adaptive scaling to selectively perturb features that an attacker is likely exploiting, based on observed deviations.\n  We evaluate Adaptive Feature Poisoning against multiple realistic adversarial attack strategies, including silent probing, transferability based attacks, and decision boundary based attacks. The results demonstrate its ability to confuse attackers, degrade attack effectiveness, and preserve detection performance. By offering a generalizable, attack agnostic, and undetectable defense, Adaptive Feature Poisoning represents a significant step toward practical and robust adversarial resilience in machine learning based intrusion detection systems.", "date": "2025-12-15", "categories": []}
{"id": "2408.08518", "title": "Visual-Friendly Concept Protection via Selective Adversarial Perturbations", "abstract": "Personalized concept generation by tuning diffusion models with a few images raises potential legal and ethical concerns regarding privacy and intellectual property rights. Researchers attempt to prevent malicious personalization using adversarial perturbations. However, previous efforts have mainly focused on the effectiveness of protection while neglecting the visibility of perturbations. They utilize global adversarial perturbations, which introduce noticeable alterations to original images and significantly degrade visual quality. In this work, we propose the Visual-Friendly Concept Protection (VCPro) framework, which prioritizes the protection of key concepts chosen by the image owner through adversarial perturbations with lower perceptibility. To ensure these perturbations are as inconspicuous as possible, we introduce a relaxed optimization objective to identify the least perceptible yet effective adversarial perturbations, solved using the Lagrangian multiplier method. Qualitative and quantitative experiments validate that VCPro achieves a better trade-off between the visibility of perturbations and protection effectiveness, effectively prioritizing the protection of target concepts in images with less perceptible perturbations.", "date": "2024-08-16", "categories": []}
{"id": "2512.02830", "title": "Defense That Attacks: How Robust Models Become Better Attackers", "abstract": "Deep learning has achieved great success in computer vision, but remains vulnerable to adversarial attacks. Adversarial training is the leading defense designed to improve model robustness. However, its effect on the transferability of attacks is underexplored. In this work, we ask whether adversarial training unintentionally increases the transferability of adversarial examples. To answer this, we trained a diverse zoo of 36 models, including CNNs and ViTs, and conducted comprehensive transferability experiments. Our results reveal a clear paradox: adversarially trained (AT) models produce perturbations that transfer more effectively than those from standard models, which introduce a new ecosystem risk. To enable reproducibility and further study, we release all models, code, and experimental scripts. Furthermore, we argue that robustness evaluations should assess not only the resistance of a model to transferred attacks but also its propensity to produce transferable adversarial examples.", "date": "2025-12-02", "categories": []}
{"id": "2512.11359", "title": "Attacking and Securing Community Detection: A Game-Theoretic Framework", "abstract": "It has been demonstrated that adversarial graphs, i.e., graphs with imperceptible perturbations, can cause deep graph models to fail on classification tasks. In this work, we extend the concept of adversarial graphs to the community detection problem, which is more challenging. We propose novel attack and defense techniques for community detection problem, with the objective of hiding targeted individuals from detection models and enhancing the robustness of community detection models, respectively. These techniques have many applications in real-world scenarios, for example, protecting personal privacy in social networks and understanding camouflage patterns in transaction networks. To simulate interactive attack and defense behaviors, we further propose a game-theoretic framework, called CD-GAME. One player is a graph attacker, while the other player is a Rayleigh Quotient defender. The CD-GAME models the mutual influence and feedback mechanisms between the attacker and the defender, revealing the dynamic evolutionary process of the game. Both players dynamically update their strategies until they reach the Nash equilibrium. Extensive experiments demonstrate the effectiveness of our proposed attack and defense methods, and both outperform existing baselines by a significant margin. Furthermore, CD-GAME provides valuable insights for understanding interactive attack and defense scenarios in community detection problems. We found that in traditional single-step attack or defense, attacker tends to employ strategies that are most effective, but are easily detected and countered by defender. When the interactive game reaches a Nash equilibrium, attacker adopts more imperceptible strategies that can still achieve satisfactory attack effectiveness even after defense.", "date": "2025-12-12", "categories": []}
{"id": "2512.11783", "title": "Super Suffixes: Bypassing Text Generation Alignment and Guard Models Simultaneously", "abstract": "The rapid deployment of Large Language Models (LLMs) has created an urgent need for enhanced security and privacy measures in Machine Learning (ML). LLMs are increasingly being used to process untrusted text inputs and even generate executable code, often while having access to sensitive system controls. To address these security concerns, several companies have introduced guard models, which are smaller, specialized models designed to protect text generation models from adversarial or malicious inputs. In this work, we advance the study of adversarial inputs by introducing Super Suffixes, suffixes capable of overriding multiple alignment objectives across various models with different tokenization schemes. We demonstrate their effectiveness, along with our joint optimization technique, by successfully bypassing the protection mechanisms of Llama Prompt Guard 2 on five different text generation models for malicious text and code generation. To the best of our knowledge, this is the first work to reveal that Llama Prompt Guard 2 can be compromised through joint optimization.\n  Additionally, by analyzing the changing similarity of a model's internal state to specific concept directions during token sequence processing, we propose an effective and lightweight method to detect Super Suffix attacks. We show that the cosine similarity between the residual stream and certain concept directions serves as a distinctive fingerprint of model intent. Our proposed countermeasure, DeltaGuard, significantly improves the detection of malicious prompts generated through Super Suffixes. It increases the non-benign classification rate to nearly 100%, making DeltaGuard a valuable addition to the guard model stack and enhancing robustness against adversarial prompt attacks.", "date": "2025-12-12", "categories": []}
{"id": "2411.03752", "title": "Deferred Poisoning: Making the Model More Vulnerable via Hessian Singularization", "abstract": "Recent studies have shown that deep learning models are very vulnerable to poisoning attacks. Many defense methods have been proposed to address this issue. However, traditional poisoning attacks are not as threatening as commonly believed. This is because they often cause differences in how the model performs on the training set compared to the validation set. Such inconsistency can alert defenders that their data has been poisoned, allowing them to take the necessary defensive actions. In this paper, we introduce a more threatening type of poisoning attack called the Deferred Poisoning Attack. This new attack allows the model to function normally during the training and validation phases but makes it very sensitive to evasion attacks or even natural noise. We achieve this by ensuring the poisoned model's loss function has a similar value as a normally trained model at each input sample but with a large local curvature. A similar model loss ensures that there is no obvious inconsistency between the training and validation accuracy, demonstrating high stealthiness. On the other hand, the large curvature implies that a small perturbation may cause a significant increase in model loss, leading to substantial performance degradation, which reflects a worse robustness. We fulfill this purpose by making the model have singular Hessian information at the optimal point via our proposed Singularization Regularization term. We have conducted both theoretical and empirical analyses of the proposed method and validated its effectiveness through experiments on image classification tasks. Furthermore, we have confirmed the hazards of this form of poisoning attack under more general scenarios using natural noise, offering a new perspective for research in the field of security.", "date": "2024-11-06", "categories": []}
{"id": "2511.16203", "title": "When Alignment Fails: Multimodal Adversarial Attacks on Vision-Language-Action Models", "abstract": "Vision-Language-Action models (VLAs) have recently demonstrated remarkable progress in embodied environments, enabling robots to perceive, reason, and act through unified multimodal understanding. Despite their impressive capabilities, the adversarial robustness of these systems remains largely unexplored, especially under realistic multimodal and black-box conditions. Existing studies mainly focus on single-modality perturbations and overlook the cross-modal misalignment that fundamentally affects embodied reasoning and decision-making. In this paper, we introduce VLA-Fool, a comprehensive study of multimodal adversarial robustness in embodied VLA models under both white-box and black-box settings. VLA-Fool unifies three levels of multimodal adversarial attacks: (1) textual perturbations through gradient-based and prompt-based manipulations, (2) visual perturbations via patch and noise distortions, and (3) cross-modal misalignment attacks that intentionally disrupt the semantic correspondence between perception and instruction. We further incorporate a VLA-aware semantic space into linguistic prompts, developing the first automatically crafted and semantically guided prompting framework. Experiments on the LIBERO benchmark using a fine-tuned OpenVLA model reveal that even minor multimodal perturbations can cause significant behavioral deviations, demonstrating the fragility of embodied multimodal alignment.", "date": "2025-11-20", "categories": []}
{"id": "2512.10936", "title": "Empirical evaluation of the Frank-Wolfe methods for constructing white-box adversarial attacks", "abstract": "The construction of adversarial attacks for neural networks appears to be a crucial challenge for their deployment in various services. To estimate the adversarial robustness of a neural network, a fast and efficient approach is needed to construct adversarial attacks. Since the formalization of adversarial attack construction involves solving a specific optimization problem, we consider the problem of constructing an efficient and effective adversarial attack from a numerical optimization perspective. Specifically, we suggest utilizing advanced projection-free methods, known as modified Frank-Wolfe methods, to construct white-box adversarial attacks on the given input data. We perform a theoretical and numerical evaluation of these methods and compare them with standard approaches based on projection operations or geometrical intuition. Numerical experiments are performed on the MNIST and CIFAR-10 datasets, utilizing a multiclass logistic regression model, the convolutional neural networks (CNNs), and the Vision Transformer (ViT).", "date": "2025-12-11", "categories": []}
{"id": "2411.12331", "title": "Adversarial-Robustness-Guided Graph Pruning", "abstract": "Graph learning plays a central role in many data mining and machine learning tasks, such as manifold learning, data representation and analysis, dimensionality reduction, clustering, and visualization. In this work, we propose a highly scalable, adversarial-robustness-guided graph pruning framework for learning graph topologies from data. By performing a spectral adversarial robustness evaluation, our method aims to learn sparse, undirected graphs that help the underlying algorithms resist noise and adversarial perturbations. In particular, we explicitly identify and prune edges that are most vulnerable to adversarial attacks. We use spectral clustering, one of the most representative graph-based machine learning algorithms, to evaluate the proposed framework. Compared with prior state-of-the-art graph learning approaches, the proposed method is more scalable and significantly improves both the computational efficiency and the solution quality of spectral clustering.", "date": "2024-11-19", "categories": []}
{"id": "2510.16443", "title": "Colliding with Adversaries at ECML-PKDD 2025 Model Robustness Competition 1st Prize Solution", "abstract": "This report presents the winning solution for Task 2 of Colliding with Adversaries: A Challenge on Robust Learning in High Energy Physics Discovery at ECML-PKDD 2025. The goal of the challenge was to design and train a robust ANN-based model capable of achieving high accuracy in a binary classification task on both clean and adversarial data generated with the Random Distribution Shuffle Attack (RDSA). Our solution consists of two components: a data generation phase and a robust model training phase. In the first phase, we produced 15 million artificial training samples using a custom methodology derived from Random Distribution Shuffle Attack (RDSA). In the second phase, we introduced a robust architecture comprising (i)a Feature Embedding Block with shared weights among features of the same type and (ii)a Dense Fusion Tail responsible for the final prediction. Training this architecture on our adversarial dataset achieved a mixed accuracy score of 80\\%, exceeding the second-place solution by two percentage points.", "date": "2025-10-18", "categories": []}
{"id": "2512.08485", "title": "Optimal Perturbation Budget Allocation for Data Poisoning in Offline Reinforcement Learning", "abstract": "Offline Reinforcement Learning (RL) enables policy optimization from static datasets but is inherently vulnerable to data poisoning attacks. Existing attack strategies typically rely on locally uniform perturbations, which treat all samples indiscriminately. This approach is inefficient, as it wastes the perturbation budget on low-impact samples, and lacks stealthiness due to significant statistical deviations. In this paper, we propose a novel Global Budget Allocation attack strategy. Leveraging the theoretical insight that a sample's influence on value function convergence is proportional to its Temporal Difference (TD) error, we formulate the attack as a global resource allocation problem. We derive a closed-form solution where perturbation magnitudes are assigned proportional to the TD-error sensitivity under a global L2 constraint. Empirical results on D4RL benchmarks demonstrate that our method significantly outperforms baseline strategies, achieving up to 80% performance degradation with minimal perturbations that evade detection by state-of-the-art statistical and spectral defenses.", "date": "2025-12-09", "categories": []}
{"id": "2512.09264", "title": "FBA$^2$D: Frequency-based Black-box Attack for AI-generated Image Detection", "abstract": "The prosperous development of Artificial Intelligence-Generated Content (AIGC) has brought people's anxiety about the spread of false information on social media. Designing detectors for filtering is an effective defense method, but most detectors will be compromised by adversarial samples. Currently, most studies exposing AIGC security issues assume information on model structure and data distribution. In real applications, attackers query and interfere with models that provide services in the form of application programming interfaces (APIs), which constitutes the black-box decision-based attack paradigm. However, to the best of our knowledge, decision-based attacks on AIGC detectors remain unexplored. In this study, we propose \\textbf{FBA$^2$D}: a frequency-based black-box attack method for AIGC detection to fill the research gap. Motivated by frequency-domain discrepancies between generated and real images, we develop a decision-based attack that leverages the Discrete Cosine Transform (DCT) for fine-grained spectral partitioning and selects frequency bands as query subspaces, improving both query efficiency and image quality. Moreover, attacks on AIGC detectors should mitigate initialization failures, preserve image quality, and operate under strict query budgets. To address these issues, we adopt an ``adversarial example soup'' method, averaging candidates from successive surrogate iterations and using the result as the initialization to accelerate the query-based attack. The empirical study on the Synthetic LSUN dataset and GenImage dataset demonstrate the effectiveness of our prosed method. This study shows the urgency of addressing practical AIGC security problems.", "date": "2025-12-10", "categories": []}
{"id": "2512.08123", "title": "Universal Adversarial Suffixes Using Calibrated Gumbel-Softmax Relaxation", "abstract": "Language models (LMs) are often used as zero-shot or few-shot classifiers by scoring label words, but they remain fragile to adversarial prompts. Prior work typically optimizes task- or model-specific triggers, making results difficult to compare and limiting transferability. We study universal adversarial suffixes: short token sequences (4-10 tokens) that, when appended to any input, broadly reduce accuracy across tasks and models. Our approach learns the suffix in a differentiable \"soft\" form using Gumbel-Softmax relaxation and then discretizes it for inference. Training maximizes calibrated cross-entropy on the label region while masking gold tokens to prevent trivial leakage, with entropy regularization to avoid collapse. A single suffix trained on one model transfers effectively to others, consistently lowering both accuracy and calibrated confidence. Experiments on sentiment analysis, natural language inference, paraphrase detection, commonsense QA, and physical reasoning with Qwen2-1.5B, Phi-1.5, and TinyLlama-1.1B demonstrate consistent attack effectiveness and transfer across tasks and model families.", "date": "2025-12-09", "categories": []}
{"id": "2512.08129", "title": "Improving the Sensitivity of Backdoor Detectors via Class Subspace Orthogonalization", "abstract": "Most post-training backdoor detection methods rely on attacked models exhibiting extreme outlier detection statistics for the target class of an attack, compared to non-target classes. However, these approaches may fail: (1) when some (non-target) classes are easily discriminable from all others, in which case they may naturally achieve extreme detection statistics (e.g., decision confidence); and (2) when the backdoor is subtle, i.e., with its features weak relative to intrinsic class-discriminative features. A key observation is that the backdoor target class has contributions to its detection statistic from both the backdoor trigger and from its intrinsic features, whereas non-target classes only have contributions from their intrinsic features. To achieve more sensitive detectors, we thus propose to suppress intrinsic features while optimizing the detection statistic for a given class. For non-target classes, such suppression will drastically reduce the achievable statistic, whereas for the target class the (significant) contribution from the backdoor trigger remains. In practice, we formulate a constrained optimization problem, leveraging a small set of clean examples from a given class, and optimizing the detection statistic while orthogonalizing with respect to the class's intrinsic features. We dub this plug-and-play approach Class Subspace Orthogonalization (CSO) and assess it against challenging mixed-label and adaptive attacks.", "date": "2025-12-09", "categories": []}
{"id": "2512.08131", "title": "Universal Adversarial Suffixes for Language Models Using Reinforcement Learning with Calibrated Reward", "abstract": "Language models are vulnerable to short adversarial suffixes that can reliably alter predictions. Previous works usually find such suffixes with gradient search or rule-based methods, but these are brittle and often tied to a single task or model. In this paper, a reinforcement learning framework is used where the suffix is treated as a policy and trained with Proximal Policy Optimization against a frozen model as a reward oracle. Rewards are shaped using calibrated cross-entropy, removing label bias and aggregating across surface forms to improve transferability. The proposed method is evaluated on five diverse NLP benchmark datasets, covering sentiment, natural language inference, paraphrase, and commonsense reasoning, using three distinct language models: Qwen2-1.5B Instruct, TinyLlama-1.1B Chat, and Phi-1.5. Results show that RL-trained suffixes consistently degrade accuracy and transfer more effectively across tasks and models than previous adversarial triggers of similar genres.", "date": "2025-12-09", "categories": []}
{"id": "2512.08329", "title": "Interpreting Structured Perturbations in Image Protection Methods for Diffusion Models", "abstract": "Recent image protection mechanisms such as Glaze and Nightshade introduce imperceptible, adversarially designed perturbations intended to disrupt downstream text-to-image generative models. While their empirical effectiveness is known, the internal structure, detectability, and representational behavior of these perturbations remain poorly understood. This study provides a systematic, explainable AI analysis using a unified framework that integrates white-box feature-space inspection and black-box signal-level probing. Through latent-space clustering, feature-channel activation analysis, occlusion-based spatial sensitivity mapping, and frequency-domain characterization, we show that protection mechanisms operate as structured, low-entropy perturbations tightly coupled to underlying image content across representational, spatial, and spectral domains. Protected images preserve content-driven feature organization with protection-specific substructure rather than inducing global representational drift. Detectability is governed by interacting effects of perturbation entropy, spatial deployment, and frequency alignment, with sequential protection amplifying detectable structure rather than suppressing it. Frequency-domain analysis shows that Glaze and Nightshade redistribute energy along dominant image-aligned frequency axes rather than introducing diffuse noise. These findings indicate that contemporary image protection operates through structured feature-level deformation rather than semantic dislocation, explaining why protection signals remain visually subtle yet consistently detectable. This work advances the interpretability of adversarial image protection and informs the design of future defenses and detection strategies for generative AI systems.", "date": "2025-12-09", "categories": []}
{"id": "2512.08832", "title": "Forecasting Fails: Unveiling Evasion Attacks in Weather Prediction Models", "abstract": "With the increasing reliance on AI models for weather forecasting, it is imperative to evaluate their vulnerability to adversarial perturbations. This work introduces Weather Adaptive Adversarial Perturbation Optimization (WAAPO), a novel framework for generating targeted adversarial perturbations that are both effective in manipulating forecasts and stealthy to avoid detection. WAAPO achieves this by incorporating constraints for channel sparsity, spatial localization, and smoothness, ensuring that perturbations remain physically realistic and imperceptible. Using the ERA5 dataset and FourCastNet (Pathak et al. 2022), we demonstrate WAAPO's ability to generate adversarial trajectories that align closely with predefined targets, even under constrained conditions. Our experiments highlight critical vulnerabilities in AI-driven forecasting models, where small perturbations to initial conditions can result in significant deviations in predicted weather patterns. These findings underscore the need for robust safeguards to protect against adversarial exploitation in operational forecasting systems.", "date": "2025-12-09", "categories": []}
{"id": "2305.12066", "title": "Attacking All Tasks at Once Using Adversarial Examples in Multi-Task Learning", "abstract": "Visual content understanding frequently relies on multi-task models to extract robust representations of a single visual input for multiple downstream tasks. However, in comparison to extensively studied single-task models, the adversarial robustness of multi-task models has received significantly less attention and many questions remain unclear: 1) How robust are multi-task models to single task adversarial attacks, 2) Can adversarial attacks be designed to simultaneously attack all tasks in a multi-task model, and 3) How does parameter sharing across tasks affect multi-task model robustness to adversarial attacks? This paper aims to answer these questions through careful analysis and rigorous experimentation. First, we analyze the inherent drawbacks of two commonly-used adaptations of single-task white-box attacks in attacking multi-task models. We then propose a novel attack framework, Dynamic Gradient Balancing Attack (DGBA). Our framework poses the problem of attacking all tasks in a multi-task model as an optimization problem that can be efficiently solved through integer linear programming. Extensive evaluation on two popular MTL benchmarks, NYUv2 and Tiny-Taxonomy, demonstrates the effectiveness of DGBA compared to baselines in attacking both clean and adversarially trained multi-task models. Our results also reveal a fundamental trade-off between improving task accuracy via parameter sharing across tasks and undermining model robustness due to increased attack transferability from parameter sharing.", "date": "2023-05-20", "categories": []}
{"id": "2412.20025", "title": "Causal Interpretability for Adversarial Robustness: A Hybrid Generative Classification Approach", "abstract": "Deep learning-based discriminative classifiers, despite their remarkable success, remain vulnerable to adversarial examples that can mislead model predictions. While adversarial training can enhance robustness, it fails to address the intrinsic vulnerability stemming from the opaque nature of these black-box models. We present a deep ensemble model that combines discriminative features with generative models to achieve both high accuracy and adversarial robustness. Our approach integrates a bottom-level pre-trained discriminative network for feature extraction with a top-level generative classification network that models adversarial input distributions through a deep latent variable model. Using variational Bayes, our model achieves superior robustness against white-box adversarial attacks without adversarial training. Extensive experiments on CIFAR-10 and CIFAR-100 demonstrate our model's superior adversarial robustness. Through evaluations using counterfactual metrics and feature interaction-based metrics, we establish correlations between model interpretability and adversarial robustness. Additionally, preliminary results on Tiny-ImageNet validate our approach's scalability to more complex datasets, offering a practical solution for developing robust image classification models.", "date": "2024-12-28", "categories": []}
{"id": "2505.01267", "title": "Diffusion-based Adversarial Purification from the Perspective of the Frequency Domain", "abstract": "The diffusion-based adversarial purification methods attempt to drown adversarial perturbations into a part of isotropic noise through the forward process, and then recover the clean images through the reverse process. Due to the lack of distribution information about adversarial perturbations in the pixel domain, it is often unavoidable to damage normal semantics. We turn to the frequency domain perspective, decomposing the image into amplitude spectrum and phase spectrum. We find that for both spectra, the damage caused by adversarial perturbations tends to increase monotonically with frequency. This means that we can extract the content and structural information of the original clean sample from the frequency components that are less damaged. Meanwhile, theoretical analysis indicates that existing purification methods indiscriminately damage all frequency components, leading to excessive damage to the image. Therefore, we propose a purification method that can eliminate adversarial perturbations while maximizing the preservation of the content and structure of the original image. Specifically, at each time step during the reverse process, for the amplitude spectrum, we replace the low-frequency components of the estimated image's amplitude spectrum with the corresponding parts of the adversarial image. For the phase spectrum, we project the phase of the estimated image into a designated range of the adversarial image's phase spectrum, focusing on the low frequencies. Empirical evidence from extensive experiments demonstrates that our method significantly outperforms most current defense methods.", "date": "2025-05-02", "categories": []}
{"id": "2505.09342", "title": "Evaluating the robustness of adversarial defenses in malware detection systems", "abstract": "Machine learning is a key tool for Android malware detection, effectively identifying malicious patterns in apps. However, ML-based detectors are vulnerable to evasion attacks, where small, crafted changes bypass detection. Despite progress in adversarial defenses, the lack of comprehensive evaluation frameworks in binary-constrained domains limits understanding of their robustness. We introduce two key contributions. First, Prioritized Binary Rounding, a technique to convert continuous perturbations into binary feature spaces while preserving high attack success and low perturbation size. Second, the sigma-binary attack, a novel adversarial method for binary domains, designed to achieve attack goals with minimal feature changes. Experiments on the Malscan dataset show that sigma-binary outperforms existing attacks and exposes key vulnerabilities in state-of-the-art defenses. Defenses equipped with adversary detectors, such as KDE, DLA, DNN+, and ICNN, exhibit significant brittleness, with attack success rates exceeding 90% using fewer than 10 feature modifications and reaching 100% with just 20. Adversarially trained defenses, including AT-rFGSM-k, AT-MaxMA, improves robustness under small budgets but remains vulnerable to unrestricted perturbations, with attack success rates of 99.45% and 96.62%, respectively. Although PAD-SMA demonstrates strong robustness against state-of-the-art gradient-based adversarial attacks by maintaining an attack success rate below 16.55%, the sigma-binary attack significantly outperforms these methods, achieving a 94.56% success rate under unrestricted perturbations. These findings highlight the critical need for precise method like sigma-binary to expose hidden vulnerabilities in existing defenses and support the development of more resilient malware detection systems.", "date": "2025-05-14", "categories": []}
{"id": "2505.19397", "title": "Are Time-Series Foundation Models Deployment-Ready? A Systematic Study of Adversarial Robustness Across Domains", "abstract": "Time-Series Foundation Models (TSFMs) are rapidly transitioning from research prototypes to core components of critical decision-making systems, driven by their impressive zero-shot forecasting capabilities. However, as their deployment surges, a critical blind spot remains: their fragility under adversarial attacks. This lack of scrutiny poses severe risks, particularly as TSFMs enter high-stakes environments vulnerable to manipulation. We present a systematic, diagnostic study arguing that for TSFMs, robustness is not merely a secondary metric but a prerequisite for trustworthy deployment comparable to accuracy. Our evaluation framework, explicitly tailored to the unique constraints of time series, incorporates normalized, sparsity-aware perturbation budgets and unified scale-invariant metrics across white-box and black-box settings. Across six representative TSFMs, we demonstrate that current architectures are alarmingly brittle: even small perturbations can reliably steer forecasts toward specific failure modes, such as trend flips and malicious drifts. We uncover TSFM-specific vulnerability patterns, including horizon-proximal brittleness, increased susceptibility with longer context windows, and weak cross-model transfer that points to model-specific failure modes rather than generic distortions. Finally, we show that simple adversarial fine-tuning offers a cost-effective path to substantial robustness gains, even with out-of-domain data. This work bridges the gap between TSFM capabilities and safety constraints, offering essential guidance for hardening the next generation of forecasting systems.", "date": "2025-05-26", "categories": []}
{"id": "2505.21027", "title": "TabAttackBench: A Benchmark for Adversarial Attacks on Tabular Data", "abstract": "Adversarial attacks pose a significant threat to machine learning models by inducing incorrect predictions through imperceptible perturbations to input data. While these attacks are well studied in unstructured domains such as images, their behaviour on tabular data remains underexplored due to mixed feature types and complex inter-feature dependencies. This study introduces a comprehensive benchmark that evaluates adversarial attacks on tabular datasets with respect to both effectiveness and imperceptibility. We assess five white-box attack algorithms (FGSM, BIM, PGD, DeepFool, and C\\&W) across four representative models (LR, MLP, TabTransformer and FT-Transformer) using eleven datasets spanning finance, energy, and healthcare domains. The benchmark employs four quantitative imperceptibility metrics (proximity, sparsity, deviation, and sensitivity) to characterise perturbation realism. The analysis quantifies the trade-off between these two aspects and reveals consistent differences between attack types, with $\\ell_\\infty$-based attacks achieving higher success but lower subtlety, and $\\ell_2$-based attacks offering more realistic perturbations. The benchmark findings offer actionable insights for designing more imperceptible adversarial attacks, advancing the understanding of adversarial vulnerability in tabular machine learning.", "date": "2025-05-27", "categories": []}
{"id": "2508.03209", "title": "GeoShield: Safeguarding Geolocation Privacy from Vision-Language Models via Adversarial Perturbations", "abstract": "Vision-Language Models (VLMs) such as GPT-4o now demonstrate a remarkable ability to infer users' locations from public shared images, posing a substantial risk to geoprivacy. Although adversarial perturbations offer a potential defense, current methods are ill-suited for this scenario: they often perform poorly on high-resolution images and low perturbation budgets, and may introduce irrelevant semantic content. To address these limitations, we propose GeoShield, a novel adversarial framework designed for robust geoprivacy protection in real-world scenarios. GeoShield comprises three key modules: a feature disentanglement module that separates geographical and non-geographical information, an exposure element identification module that pinpoints geo-revealing regions within an image, and a scale-adaptive enhancement module that jointly optimizes perturbations at both global and local levels to ensure effectiveness across resolutions. Extensive experiments on challenging benchmarks show that GeoShield consistently surpasses prior methods in black-box settings, achieving strong privacy protection with minimal impact on visual or semantic quality. To our knowledge, this work is the first to explore adversarial perturbations for defending against geolocation inference by advanced VLMs, providing a practical and effective solution to escalating privacy concerns.", "date": "2025-08-05", "categories": []}
{"id": "2511.07040", "title": "3D-ANC: Adaptive Neural Collapse for Robust 3D Point Cloud Recognition", "abstract": "Deep neural networks have recently achieved notable progress in 3D point cloud recognition, yet their vulnerability to adversarial perturbations poses critical security challenges in practical deployments. Conventional defense mechanisms struggle to address the evolving landscape of multifaceted attack patterns. Through systematic analysis of existing defenses, we identify that their unsatisfactory performance primarily originates from an entangled feature space, where adversarial attacks can be performed easily. To this end, we present 3D-ANC, a novel approach that capitalizes on the Neural Collapse (NC) mechanism to orchestrate discriminative feature learning. In particular, NC depicts where last-layer features and classifier weights jointly evolve into a simplex equiangular tight frame (ETF) arrangement, establishing maximally separable class prototypes. However, leveraging this advantage in 3D recognition confronts two substantial challenges: (1) prevalent class imbalance in point cloud datasets, and (2) complex geometric similarities between object categories. To tackle these obstacles, our solution combines an ETF-aligned classification module with an adaptive training framework consisting of representation-balanced learning (RBL) and dynamic feature direction loss (FDL). 3D-ANC seamlessly empowers existing models to develop disentangled feature spaces despite the complexity in 3D data distribution. Comprehensive evaluations state that 3D-ANC significantly improves the robustness of models with various structures on two datasets. For instance, DGCNN's classification accuracy is elevated from 27.2% to 80.9% on ModelNet40 -- a 53.7% absolute gain that surpasses leading baselines by 34.0%.", "date": "2025-11-10", "categories": []}
{"id": "2511.11210", "title": "STONE: Pioneering the One-to-N Universal Backdoor Threat in 3D Point Cloud", "abstract": "Backdoor attacks pose a critical threat to deep learning, especially in safety-sensitive 3D domains such as autonomous driving and robotics. While potent, existing attacks on 3D point clouds are predominantly limited to one-to-one paradigms. The more flexible and universal one-to-N multi-target backdoor threat remains largely unexplored, lacking both theoretical and practical foundations. To bridge this gap, we propose STONE (Spherical Trigger One-to-N universal backdoor Enabling), the first method to instantiate this threat via a configurable spherical trigger design. Its parameterized spatial properties establish a dynamic key space, enabling a single trigger to map to multiple target labels. Theoretically, we ground STONE in a Neural Tangent Kernel (NTK) analysis, providing the first formal basis for one-to-N mappings in 3D models. Empirically, extensive evaluations demonstrate high attack success rates (up to 100\\%) without compromising clean-data accuracy. This work establishes a foundational benchmark for multi-target backdoor threats under dirty-label and black-box settings in 3D vision -- a crucial step toward securing future intelligent systems.", "date": "2025-11-14", "categories": []}
{"id": "2512.06123", "title": "Toward Patch Robustness Certification and Detection for Deep Learning Systems Beyond Consistent Samples", "abstract": "Patch robustness certification is an emerging kind of provable defense technique against adversarial patch attacks for deep learning systems. Certified detection ensures the detection of all patched harmful versions of certified samples, which mitigates the failures of empirical defense techniques that could (easily) be compromised. However, existing certified detection methods are ineffective in certifying samples that are misclassified or whose mutants are inconsistently pre icted to different labels. This paper proposes HiCert, a novel masking-based certified detection technique. By focusing on the problem of mutants predicted with a label different from the true label with our formal analysis, HiCert formulates a novel formal relation between harmful samples generated by identified loopholes and their benign counterparts. By checking the bound of the maximum confidence among these potentially harmful (i.e., inconsistent) mutants of each benign sample, HiCert ensures that each harmful sample either has the minimum confidence among mutants that are predicted the same as the harmful sample itself below this bound, or has at least one mutant predicted with a label different from the harmful sample itself, formulated after two novel insights. As such, HiCert systematically certifies those inconsistent samples and consistent samples to a large extent. To our knowledge, HiCert is the first work capable of providing such a comprehensive patch robustness certification for certified detection. Our experiments show the high effectiveness of HiCert with a new state-of the-art performance: It certifies significantly more benign samples, including those inconsistent and consistent, and achieves significantly higher accuracy on those samples without warnings and a significantly lower false silent ratio.", "date": "2025-12-05", "categories": []}
{"id": "2512.06185", "title": "SPOOF: Simple Pixel Operations for Out-of-Distribution Fooling", "abstract": "Deep neural networks (DNNs) excel across image recognition tasks, yet continue to exhibit overconfidence on inputs that bear no resemblance to natural images. Revisiting the \"fooling images\" work introduced by Nguyen et al. (2015), we re-implement both CPPN-based and direct-encoding-based evolutionary fooling attacks on modern architectures, including convolutional and transformer classifiers. Our re-implementation confirm that high-confidence fooling persists even in state-of-the-art networks, with transformer-based ViT-B/16 emerging as the most susceptible--achieving near-certain misclassifications with substantially fewer queries than convolution-based models. We then introduce SPOOF, a minimalist, consistent, and more efficient black-box attack generating high-confidence fooling images. Despite its simplicity, SPOOF generates unrecognizable fooling images with minimal pixel modifications and drastically reduced compute. Furthermore, retraining with fooling images as an additional class provides only partial resistance, as SPOOF continues to fool consistently with slightly higher query budgets--highlighting persistent fragility of modern deep classifiers.", "date": "2025-12-05", "categories": []}
{"id": "2512.06674", "title": "RunawayEvil: Jailbreaking the Image-to-Video Generative Models", "abstract": "Image-to-Video (I2V) generation synthesizes dynamic visual content from image and text inputs, providing significant creative control. However, the security of such multimodal systems, particularly their vulnerability to jailbreak attacks, remains critically underexplored. To bridge this gap, we propose RunawayEvil, the first multimodal jailbreak framework for I2V models with dynamic evolutionary capability. Built on a \"Strategy-Tactic-Action\" paradigm, our framework exhibits self-amplifying attack through three core components: (1) Strategy-Aware Command Unit that enables the attack to self-evolve its strategies through reinforcement learning-driven strategy customization and LLM-based strategy exploration; (2) Multimodal Tactical Planning Unit that generates coordinated text jailbreak instructions and image tampering guidelines based on the selected strategies; (3) Tactical Action Unit that executes and evaluates the multimodal coordinated attacks. This self-evolving architecture allows the framework to continuously adapt and intensify its attack strategies without human intervention. Extensive experiments demonstrate RunawayEvil achieves state-of-the-art attack success rates on commercial I2V models, such as Open-Sora 2.0 and CogVideoX. Specifically, RunawayEvil outperforms existing methods by 58.5 to 79 percent on COCO2017. This work provides a critical tool for vulnerability analysis of I2V models, thereby laying a foundation for more robust video generation systems.", "date": "2025-12-07", "categories": []}
{"id": "2512.07059", "title": "Replicating TEMPEST at Scale: Multi-Turn Adversarial Attacks Against Trillion-Parameter Frontier Models", "abstract": "Despite substantial investment in safety alignment, the vulnerability of large language models to sophisticated multi-turn adversarial attacks remains poorly characterized, and whether model scale or inference mode affects robustness is unknown. This study employed the TEMPEST multi-turn attack framework to evaluate ten frontier models from eight vendors across 1,000 harmful behaviors, generating over 97,000 API queries across adversarial conversations with automated evaluation by independent safety classifiers. Results demonstrated a spectrum of vulnerability: six models achieved 96% to 100% attack success rate (ASR), while four showed meaningful resistance, with ASR ranging from 42% to 78%; enabling extended reasoning on identical architecture reduced ASR from 97% to 42%. These findings indicate that safety alignment quality varies substantially across vendors, that model scale does not predict adversarial robustness, and that thinking mode provides a deployable safety enhancement. Collectively, this work establishes that current alignment techniques remain fundamentally vulnerable to adaptive multi-turn attacks regardless of model scale, while identifying deliberative inference as a promising defense direction.", "date": "2025-12-08", "categories": []}
{"id": "2512.07228", "title": "Towards Robust Protective Perturbation against DeepFake Face Swapping", "abstract": "DeepFake face swapping enables highly realistic identity forgeries, posing serious privacy and security risks. A common defence embeds invisible perturbations into images, but these are fragile and often destroyed by basic transformations such as compression or resizing. In this paper, we first conduct a systematic analysis of 30 transformations across six categories and show that protection robustness is highly sensitive to the choice of training transformations, making the standard Expectation over Transformation (EOT) with uniform sampling fundamentally suboptimal. Motivated by this, we propose Expectation Over Learned distribution of Transformation (EOLT), the framework to treat transformation distribution as a learnable component rather than a fixed design choice. Specifically, EOLT employs a policy network that learns to automatically prioritize critical transformations and adaptively generate instance-specific perturbations via reinforcement learning, enabling explicit modeling of defensive bottlenecks while maintaining broad transferability. Extensive experiments demonstrate that our method achieves substantial improvements over state-of-the-art approaches, with 26% higher average robustness and up to 30% gains on challenging transformation categories.", "date": "2025-12-08", "categories": []}
{"id": "2512.07247", "title": "AdLift: Lifting Adversarial Perturbations to Safeguard 3D Gaussian Splatting Assets Against Instruction-Driven Editing", "abstract": "Recent studies have extended diffusion-based instruction-driven 2D image editing pipelines to 3D Gaussian Splatting (3DGS), enabling faithful manipulation of 3DGS assets and greatly advancing 3DGS content creation. However, it also exposes these assets to serious risks of unauthorized editing and malicious tampering. Although imperceptible adversarial perturbations against diffusion models have proven effective for protecting 2D images, applying them to 3DGS encounters two major challenges: view-generalizable protection and balancing invisibility with protection capability. In this work, we propose the first editing safeguard for 3DGS, termed AdLift, which prevents instruction-driven editing across arbitrary views and dimensions by lifting strictly bounded 2D adversarial perturbations into 3D Gaussian-represented safeguard. To ensure both adversarial perturbations effectiveness and invisibility, these safeguard Gaussians are progressively optimized across training views using a tailored Lifted PGD, which first conducts gradient truncation during back-propagation from the editing model at the rendered image and applies projected gradients to strictly constrain the image-level perturbation. Then, the resulting perturbation is backpropagated to the safeguard Gaussian parameters via an image-to-Gaussian fitting operation. We alternate between gradient truncation and image-to-Gaussian fitting, yielding consistent adversarial-based protection performance across different viewpoints and generalizes to novel views. Empirically, qualitative and quantitative results demonstrate that AdLift effectively protects against state-of-the-art instruction-driven 2D image and 3DGS editing.", "date": "2025-12-08", "categories": []}
{"id": "2411.10500", "title": "Edge-Only Universal Adversarial Attacks in Distributed Learning", "abstract": "Distributed learning frameworks, which partition neural network models across multiple computing nodes, enhance efficiency in collaborative edge-cloud systems, but may also introduce new vulnerabilities to evasion attacks, often in the form of adversarial perturbations. In this work, we present a new threat model that explores the feasibility of generating universal adversarial perturbations (UAPs) when the attacker has access only to the edge portion of the model, consisting of its initial network layers. Unlike traditional attacks that require full model knowledge, our approach shows that adversaries can induce effective mispredictions in the unknown cloud component by manipulating key feature representations at the edge. Following the proposed threat model, we introduce both edge-only untargeted and targeted formulations of UAPs designed to control intermediate features before the split point. Our results on ImageNet demonstrate strong attack transferability to the unknown cloud part, and we compare the proposed method with classical white-box and black-box techniques, highlighting its effectiveness. Additionally, we analyze the capability of an attacker to achieve targeted adversarial effects with edge-only knowledge, revealing intriguing behaviors across multiple networks. By introducing the first adversarial attacks with edge-only knowledge in split inference, this work underscores the importance of addressing partial model access in adversarial robustness, encouraging further research in this area.", "date": "2024-11-15", "categories": []}
{"id": "2510.00769", "title": "ZQBA: Zero Query Black-box Adversarial Attack", "abstract": "Current black-box adversarial attacks either require multiple queries or diffusion models to produce adversarial samples that can impair the target model performance. However, these methods require training a surrogate loss or diffusion models to produce adversarial samples, which limits their applicability in real-world settings. Thus, we propose a Zero Query Black-box Adversarial (ZQBA) attack that exploits the representations of Deep Neural Networks (DNNs) to fool other networks. Instead of requiring thousands of queries to produce deceiving adversarial samples, we use the feature maps obtained from a DNN and add them to clean images to impair the classification of a target model. The results suggest that ZQBA can transfer the adversarial samples to different models and across various datasets, namely CIFAR and Tiny ImageNet. The experiments also show that ZQBA is more effective than state-of-the-art black-box attacks with a single query, while maintaining the imperceptibility of perturbations, evaluated both quantitatively (SSIM) and qualitatively, emphasizing the vulnerabilities of employing DNNs in real-world contexts. All the source code is available at https://github.com/Joana-Cabral/ZQBA.", "date": "2025-10-01", "categories": []}
{"id": "2408.10901", "title": "A Gray-box Attack against Latent Diffusion Model-based Image Editing by Posterior Collapse", "abstract": "Recent advancements in Latent Diffusion Models (LDMs) have revolutionized image synthesis and manipulation, raising significant concerns about data misappropriation and intellectual property infringement. While adversarial attacks have been extensively explored as a protective measure against such misuse of generative AI, current approaches are severely limited by their heavy reliance on model-specific knowledge and substantial computational costs. Drawing inspiration from the posterior collapse phenomenon observed in VAE training, we propose the Posterior Collapse Attack (PCA), a novel framework for protecting images from unauthorized manipulation. Through comprehensive theoretical analysis and empirical validation, we identify two distinct collapse phenomena during VAE inference: diffusion collapse and concentration collapse. Based on this discovery, we design a unified loss function that can flexibly achieve both types of collapse through parameter adjustment, each corresponding to different protection objectives in preventing image manipulation. Our method significantly reduces dependence on model-specific knowledge by requiring access to only the VAE encoder, which constitutes less than 4\\% of LDM parameters. Notably, PCA achieves prompt-invariant protection by operating on the VAE encoder before text conditioning occurs, eliminating the need for empty prompt optimization required by existing methods. This minimal requirement enables PCA to maintain adequate transferability across various VAE-based LDM architectures while effectively preventing unauthorized image editing. Extensive experiments show PCA outperforms existing techniques in protection effectiveness, computational efficiency (runtime and VRAM), and generalization across VAE-based LDM variants. Our code is available at https://github.com/ZhongliangGuo/PosteriorCollapseAttack.", "date": "2024-08-20", "categories": []}
{"id": "2506.04263", "title": "Dynamic Epsilon Scheduling: A Multi-Factor Adaptive Perturbation Budget for Adversarial Training", "abstract": "Adversarial training is among the most effective strategies for defending deep neural networks against adversarial examples. A key limitation of existing adversarial training approaches lies in their reliance on a fixed perturbation budget, which fails to account for instance-specific robustness characteristics. While prior works such as IAAT and MMA introduce instance-level adaptations, they often rely on heuristic or static approximations of data robustness. In this paper, we propose Dynamic Epsilon Scheduling (DES), a novel framework that adaptively adjusts the adversarial perturbation budget per instance and per training iteration. DES integrates three key factors: (1) the distance to the decision boundary approximated via gradient-based proxies, (2) prediction confidence derived from softmax entropy, and (3) model uncertainty estimated via Monte Carlo dropout. By combining these cues into a unified scheduling strategy, DES tailors the perturbation budget dynamically to guide more effective adversarial learning. Experimental results on CIFAR-10 and CIFAR-100 show that our method consistently improves both adversarial robustness and standard accuracy compared to fixed-epsilon baselines and prior adaptive methods. Moreover, we provide theoretical insights into the stability and convergence of our scheduling policy. This work opens a new avenue for instance-aware, data-driven adversarial training methods.", "date": "2025-06-03", "categories": []}
{"id": "2510.25121", "title": "Bilevel Models for Adversarial Learning and A Case Study", "abstract": "Adversarial learning has been attracting more and more attention thanks to the fast development of machine learning and artificial intelligence. However, due to the complicated structure of most machine learning models, the mechanism of adversarial attacks is not well interpreted. How to measure the effect of attacks is still not quite clear.\n  In this paper, we investigate the adversarial learning from the perturbation analysis point of view.\n  We characterize the robustness of learning models through the calmness of the solution mapping.\n  In the case of convex clustering models, we identify the conditions under which the clustering results remain the same under perturbations.\n  When the noise level is large, it leads to an attack.\n  Therefore, we propose two bilevel models for adversarial learning where the effect of adversarial learning is measured\n  by some deviation function.\n  Specifically, we systematically study the so-called $\\delta$-measure and show that under certain conditions, it can be used as a deviation function in adversarial learning for convex clustering models.\n  Finally, we conduct numerical tests to verify the above theoretical results as well as the efficiency of the two proposed bilevel models.", "date": "2025-10-29", "categories": []}
{"id": "2512.04264", "title": "Studying Various Activation Functions and Non-IID Data for Machine Learning Model Robustness", "abstract": "Adversarial training is an effective method to improve the machine learning (ML) model robustness. Most existing studies typically consider the Rectified linear unit (ReLU) activation function and centralized training environments. In this paper, we study the ML model robustness using ten different activation functions through adversarial training in centralized environments and explore the ML model robustness in federal learning environments. In the centralized environment, we first propose an advanced adversarial training approach to improving the ML model robustness by incorporating model architecture change, soft labeling, simplified data augmentation, and varying learning rates. Then, we conduct extensive experiments on ten well-known activation functions in addition to ReLU to better understand how they impact the ML model robustness. Furthermore, we extend the proposed adversarial training approach to the federal learning environment, where both independent and identically distributed (IID) and non-IID data settings are considered. Our proposed centralized adversarial training approach achieves a natural and robust accuracy of 77.08% and 67.96%, respectively on CIFAR-10 against the fast gradient sign attacks. Experiments on ten activation functions reveal ReLU usually performs best. In the federated learning environment, however, the robust accuracy decreases significantly, especially on non-IID data. To address the significant performance drop in the non-IID data case, we introduce data sharing and achieve the natural and robust accuracy of 70.09% and 54.79%, respectively, surpassing the CalFAT algorithm, when 40% data sharing is used. That is, a proper percentage of data sharing can significantly improve the ML model robustness, which is useful to some real-world applications.", "date": "2025-12-03", "categories": []}
{"id": "2512.04554", "title": "Counterfeit Answers: Adversarial Forgery against OCR-Free Document Visual Question Answering", "abstract": "Document Visual Question Answering (DocVQA) enables end-to-end reasoning grounded on information present in a document input. While recent models have shown impressive capabilities, they remain vulnerable to adversarial attacks. In this work, we introduce a novel attack scenario that aims to forge document content in a visually imperceptible yet semantically targeted manner, allowing an adversary to induce specific or generally incorrect answers from a DocVQA model. We develop specialized attack algorithms that can produce adversarially forged documents tailored to different attackers' goals, ranging from targeted misinformation to systematic model failure scenarios. We demonstrate the effectiveness of our approach against two end-to-end state-of-the-art models: Pix2Struct, a vision-language transformer that jointly processes image and text through sequence-to-sequence modeling, and Donut, a transformer-based model that directly extracts text and answers questions from document images. Our findings highlight critical vulnerabilities in current DocVQA systems and call for the development of more robust defenses.", "date": "2025-12-04", "categories": []}
{"id": "2512.04895", "title": "Chameleon: Adaptive Adversarial Agents for Scaling-Based Visual Prompt Injection in Multimodal AI Systems", "abstract": "Multimodal Artificial Intelligence (AI) systems, particularly Vision-Language Models (VLMs), have become integral to critical applications ranging from autonomous decision-making to automated document processing. As these systems scale, they rely heavily on preprocessing pipelines to handle diverse inputs efficiently. However, this dependency on standard preprocessing operations, specifically image downscaling, creates a significant yet often overlooked security vulnerability. While intended for computational optimization, scaling algorithms can be exploited to conceal malicious visual prompts that are invisible to human observers but become active semantic instructions once processed by the model. Current adversarial strategies remain largely static, failing to account for the dynamic nature of modern agentic workflows. To address this gap, we propose Chameleon, a novel, adaptive adversarial framework designed to expose and exploit scaling vulnerabilities in production VLMs. Unlike traditional static attacks, Chameleon employs an iterative, agent-based optimization mechanism that dynamically refines image perturbations based on the target model's real-time feedback. This allows the framework to craft highly robust adversarial examples that survive standard downscaling operations to hijack downstream execution. We evaluate Chameleon against Gemini 2.5 Flash model. Our experiments demonstrate that Chameleon achieves an Attack Success Rate (ASR) of 84.5% across varying scaling factors, significantly outperforming static baseline attacks which average only 32.1%. Furthermore, we show that these attacks effectively compromise agentic pipelines, reducing decision-making accuracy by over 45% in multi-step tasks. Finally, we discuss the implications of these vulnerabilities and propose multi-scale consistency checks as a necessary defense mechanism.", "date": "2025-12-04", "categories": []}
{"id": "2506.00821", "title": "SafeGenes: Evaluating the Adversarial Robustness of Genomic Foundation Models", "abstract": "Genomic Foundation Models (GFMs), such as Evolutionary Scale Modeling (ESM), have demonstrated significant success in variant effect prediction. However, their adversarial robustness remains largely unexplored. To address this gap, we propose SafeGenes: a framework for Secure analysis of genomic foundation models, leveraging adversarial attacks to evaluate robustness against both engineered near-identical adversarial Genes and embedding-space manipulations. In this study, we assess the adversarial vulnerabilities of GFMs using two approaches: the Fast Gradient Sign Method (FGSM) and a soft prompt attack. FGSM introduces minimal perturbations to input sequences, while the soft prompt attack optimizes continuous embeddings to manipulate model predictions without modifying the input tokens. By combining these techniques, SafeGenes provides a comprehensive assessment of GFM susceptibility to adversarial manipulation. Targeted soft prompt attacks induced severe degradation in MLM-based shallow architectures such as ProteinBERT, while still producing substantial failure modes even in high-capacity foundation models such as ESM1b and ESM1v. These findings expose critical vulnerabilities in current foundation models, opening new research directions toward improving their security and robustness in high-stakes genomic applications such as variant effect prediction.", "date": "2025-06-01", "categories": []}
{"id": "2509.22855", "title": "Observation-Free Attacks on Online Learning to Rank", "abstract": "Online learning to rank (OLTR) plays a critical role in information retrieval and machine learning systems, with a wide range of applications in search engines and content recommenders. However, despite their extensive adoption, the susceptibility of OLTR algorithms to coordinated adversarial attacks remains poorly understood. In this work, we present a novel framework for attacking some of the widely used OLTR algorithms. Our framework is designed to promote a set of target items so that they appear in the list of top-K recommendations for T - o(T) rounds, while simultaneously inducing linear regret in the learning algorithm. We propose two novel attack strategies: CascadeOFA for CascadeUCB1 and PBMOFA for PBM-UCB . We provide theoretical guarantees showing that both strategies require only O(log T) manipulations to succeed. Additionally, we supplement our theoretical analysis with empirical results on real-world data.", "date": "2025-09-26", "categories": []}
{"id": "2510.03747", "title": "LoRA Patching: Exposing the Fragility of Proactive Defenses against Deepfakes", "abstract": "Deepfakes pose significant societal risks, motivating the development of proactive defenses that embed adversarial perturbations in facial images to prevent manipulation. However, in this paper, we show that these preemptive defenses often lack robustness and reliability. We propose a novel approach, Low-Rank Adaptation (LoRA) patching, which injects a plug-and-play LoRA patch into Deepfake generators to bypass state-of-the-art defenses. A learnable gating mechanism adaptively controls the effect of the LoRA patch and prevents gradient explosions during fine-tuning. We also introduce a Multi-Modal Feature Alignment (MMFA) loss, encouraging the features of adversarial outputs to align with those of the desired outputs at the semantic level. Beyond bypassing, we present defensive LoRA patching, embedding visible warnings in the outputs as a complementary solution to mitigate this newly identified security vulnerability. With only 1,000 facial examples and a single epoch of fine-tuning, LoRA patching successfully defeats multiple proactive defenses. These results reveal a critical weakness in current paradigms and underscore the need for more robust Deepfake defense strategies. Our code is available at https://github.com/ZOMIN28/LoRA-Patching.", "date": "2025-10-04", "categories": []}
{"id": "2512.00765", "title": "The Outline of Deception: Physical Adversarial Attacks on Traffic Signs Using Edge Patches", "abstract": "Intelligent driving systems are vulnerable to physical adversarial attacks on traffic signs. These attacks can cause misclassification, leading to erroneous driving decisions that compromise road safety. Moreover, within V2X networks, such misinterpretations can propagate, inducing cascading failures that disrupt overall traffic flow and system stability. However, a key limitation of current physical attacks is their lack of stealth. Most methods apply perturbations to central regions of the sign, resulting in visually salient patterns that are easily detectable by human observers, thereby limiting their real-world practicality. This study proposes TESP-Attack, a novel stealth-aware adversarial patch method for traffic sign classification. Based on the observation that human visual attention primarily focuses on the central regions of traffic signs, we employ instance segmentation to generate edge-aligned masks that conform to the shape characteristics of the signs. A U-Net generator is utilized to craft adversarial patches, which are then optimized through color and texture constraints along with frequency domain analysis to achieve seamless integration with the background environment, resulting in highly effective visual concealment. The proposed method demonstrates outstanding attack success rates across traffic sign classification models with varied architectures, achieving over 90% under limited query budgets. It also exhibits strong cross-model transferability and maintains robust real-world performance that remains stable under varying angles and distances.", "date": "2025-11-30", "categories": []}
{"id": "2512.03097", "title": "Many-to-One Adversarial Consensus: Exposing Multi-Agent Collusion Risks in AI-Based Healthcare", "abstract": "The integration of large language models (LLMs) into healthcare IoT systems promises faster decisions and improved medical support. LLMs are also deployed as multi-agent teams to assist AI doctors by debating, voting, or advising on decisions. However, when multiple assistant agents interact, coordinated adversaries can collude to create false consensus, pushing an AI doctor toward harmful prescriptions. We develop an experimental framework with scripted and unscripted doctor agents, adversarial assistants, and a verifier agent that checks decisions against clinical guidelines. Using 50 representative clinical questions, we find that collusion drives the Attack Success Rate (ASR) and Harmful Recommendation Rates (HRR) up to 100% in unprotected systems. In contrast, the verifier agent restores 100% accuracy by blocking adversarial consensus. This work provides the first systematic evidence of collusion risk in AI healthcare and demonstrates a practical, lightweight defence that ensures guideline fidelity.", "date": "2025-12-01", "categories": []}
{"id": "2512.03625", "title": "FeatureLens: A Highly Generalizable and Interpretable Framework for Detecting Adversarial Examples Based on Image Features", "abstract": "Although the remarkable performance of deep neural networks (DNNs) in image classification, their vulnerability to adversarial attacks remains a critical challenge. Most existing detection methods rely on complex and poorly interpretable architectures, which compromise interpretability and generalization. To address this, we propose FeatureLens, a lightweight framework that acts as a lens to scrutinize anomalies in image features. Comprising an Image Feature Extractor (IFE) and shallow classifiers (e.g., SVM, MLP, or XGBoost) with model sizes ranging from 1,000 to 30,000 parameters, FeatureLens achieves high detection accuracy ranging from 97.8% to 99.75% in closed-set evaluation and 86.17% to 99.6% in generalization evaluation across FGSM, PGD, CW, and DAmageNet attacks, using only 51 dimensional features. By combining strong detection performance with excellent generalization, interpretability, and computational efficiency, FeatureLens offers a practical pathway toward transparent and effective adversarial defense.", "date": "2025-12-03", "categories": []}
{"id": "2512.03882", "title": "Automatic Attack Discovery for Few-Shot Class-Incremental Learning via Large Language Models", "abstract": "Few-shot class incremental learning (FSCIL) is a more realistic and challenging paradigm in continual learning to incrementally learn unseen classes and overcome catastrophic forgetting on base classes with only a few training examples. Previous efforts have primarily centered around studying more effective FSCIL approaches. By contrast, less attention was devoted to thinking the security issues in contributing to FSCIL. This paper aims to provide a holistic study of the impact of attacks on FSCIL. We first derive insights by systematically exploring how human expert-designed attack methods (i.e., PGD, FGSM) affect FSCIL. We find that those methods either fail to attack base classes, or suffer from huge labor costs due to relying on huge expert knowledge. This highlights the need to craft a specialized attack method for FSCIL. Grounded in these insights, in this paper, we propose a simple yet effective ACraft method to automatically steer and discover optimal attack methods targeted at FSCIL by leveraging Large Language Models (LLMs) without human experts. Moreover, to improve the reasoning between LLMs and FSCIL, we introduce a novel Proximal Policy Optimization (PPO) based reinforcement learning to optimize learning, making LLMs generate better attack methods in the next generation by establishing positive feedback. Experiments on mainstream benchmarks show that our ACraft significantly degrades the performance of state-of-the-art FSCIL methods and dramatically beyond human expert-designed attack methods while maintaining the lowest costs of attack.", "date": "2025-12-03", "categories": []}
{"id": "2512.02062", "title": "Superpixel Attack: Enhancing Black-box Adversarial Attack with Image-driven Division Areas", "abstract": "Deep learning models are used in safety-critical tasks such as automated driving and face recognition. However, small perturbations in the model input can significantly change the predictions. Adversarial attacks are used to identify small perturbations that can lead to misclassifications. More powerful black-box adversarial attacks are required to develop more effective defenses. A promising approach to black-box adversarial attacks is to repeat the process of extracting a specific image area and changing the perturbations added to it. Existing attacks adopt simple rectangles as the areas where perturbations are changed in a single iteration. We propose applying superpixels instead, which achieve a good balance between color variance and compactness. We also propose a new search method, versatile search, and a novel attack method, Superpixel Attack, which applies superpixels and performs versatile search. Superpixel Attack improves attack success rates by an average of 2.10% compared with existing attacks. Most models used in this study are robust against adversarial attacks, and this improvement is significant for black-box adversarial attacks. The code is avilable at https://github.com/oe1307/SuperpixelAttack.git.", "date": "2025-11-29", "categories": []}
{"id": "2512.02276", "title": "Adversarial Robustness of Traffic Classification under Resource Constraints: Input Structure Matters", "abstract": "Traffic classification (TC) plays a critical role in cybersecurity, particularly in IoT and embedded contexts, where inspection must often occur locally under tight hardware constraints. We use hardware-aware neural architecture search (HW-NAS) to derive lightweight TC models that are accurate, efficient, and deployable on edge platforms. Two input formats are considered: a flattened byte sequence and a 2D packet-wise time series; we examine how input structure affects adversarial vulnerability when using resource-constrained models. Robustness is assessed against white-box attacks, specifically Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). On USTC-TFC2016, both HW-NAS models achieve over 99% clean-data accuracy while remaining within 65k parameters and 2M FLOPs. Yet under perturbations of strength 0.1, their robustness diverges: the flat model retains over 85% accuracy, while the time-series variant drops below 35%. Adversarial fine-tuning delivers robust gains, with flat-input accuracy exceeding 96% and the time-series variant recovering over 60 percentage points in robustness, all without compromising efficiency. The results underscore how input structure influences adversarial vulnerability, and show that even compact, resource-efficient models can attain strong robustness, supporting their practical deployment in secure edge-based TC.", "date": "2025-12-01", "categories": []}
{"id": "2505.14469", "title": "Attributional Safety Failures in Large Language Models under Code-Mixed Perturbations", "abstract": "While LLMs appear robustly safety-aligned in English, we uncover a catastrophic, overlooked weakness: attributional collapse under code-mixed perturbations. Our systematic evaluation of open models shows that the linguistic camouflage of code-mixing -- ``blending languages within a single conversation'' -- can cause safety guardrails to fail dramatically. Attack success rates (ASR) spike from a benign 9\\% in monolingual English to 69\\% under code-mixed inputs, with rates exceeding 90\\% in non-Western contexts such as Arabic and Hindi. These effects hold not only on controlled synthetic datasets but also on real-world social media traces, revealing a serious risk for billions of users. To explain why this happens, we introduce saliency drift attribution (SDA), an interpretability framework that shows how, under code-mixing, the model's internal attention drifts away from safety-critical tokens (e.g., ``violence'' or ``corruption''), effectively blinding it to harmful intent. Finally, we propose a lightweight translation-based restoration strategy that recovers roughly 80\\% of the safety lost to code-mixing, offering a practical path toward more equitable and robust LLM safety.", "date": "2025-05-20", "categories": []}
{"id": "2505.18884", "title": "LORE: Lagrangian-Optimized Robust Embeddings for Visual Encoders", "abstract": "Visual encoders have become fundamental components in modern computer vision pipelines. However, ensuring robustness against adversarial perturbations remains a critical challenge. Recent efforts have explored both supervised and unsupervised adversarial fine-tuning strategies. We identify two key limitations in these approaches: (i) they often suffer from instability, especially during the early stages of fine-tuning, resulting in suboptimal convergence and degraded performance on clean data, and (ii) they exhibit a suboptimal trade-off between robustness and clean data accuracy, hindering the simultaneous optimization of both objectives. To overcome these challenges, we propose Lagrangian-Optimized Robust Embeddings (LORE), a novel unsupervised adversarial fine-tuning framework. LORE utilizes constrained optimization, which offers a principled approach to balancing competing goals, such as improving robustness while preserving nominal performance. By enforcing embedding-space proximity constraints, LORE effectively maintains clean data performance throughout adversarial fine-tuning. Extensive experiments show that LORE significantly improves zero-shot adversarial robustness with minimal degradation in clean data accuracy. Furthermore, we demonstrate the effectiveness of the adversarially fine-tuned CLIP image encoder in out-of-distribution generalization and enhancing the interpretability of image embeddings.", "date": "2025-05-24", "categories": []}
{"id": "2509.22113", "title": "Countering adversarial evasion in regression analysis", "abstract": "Adversarial machine learning challenges the assumption that the underlying distribution remains consistent throughout the training and implementation of a prediction model. In particular, adversarial evasion considers scenarios where adversaries adapt their data to influence particular outcomes from established prediction models, such scenarios arise in applications such as spam email filtering, malware detection and fake-image generation, where security methods must be actively updated to keep up with the ever-improving generation of malicious data. Game theoretic models have been shown to be effective at modelling these scenarios and hence training resilient predictors against such adversaries. Recent advancements in the use of pessimistic bilevel optimsiation which remove assumptions about the convexity and uniqueness of the adversary's optimal strategy have proved to be particularly effective at mitigating threats to classifiers due to its ability to capture the antagonistic nature of the adversary. However, this formulation has not yet been adapted to regression scenarios. This article serves to propose a pessimistic bilevel optimisation program for regression scenarios which makes no assumptions on the convexity or uniqueness of the adversary's solutions.", "date": "2025-09-26", "categories": []}
{"id": "2511.20494", "title": "Adversarial Confusion Attack: Disrupting Multimodal Large Language Models", "abstract": "We introduce the Adversarial Confusion Attack, a new class of threats against multimodal large language models (MLLMs). Unlike jailbreaks or targeted misclassification, the goal is to induce systematic disruption that makes the model generate incoherent or confidently incorrect outputs. Practical applications include embedding such adversarial images into websites to prevent MLLM-powered AI Agents from operating reliably. The proposed attack maximizes next-token entropy using a small ensemble of open-source MLLMs. In the white-box setting, we show that a single adversarial image can disrupt all models in the ensemble, both in the full-image and Adversarial CAPTCHA settings. Despite relying on a basic adversarial technique (PGD), the attack generates perturbations that transfer to both unseen open-source (e.g., Qwen3-VL) and proprietary (e.g., GPT-5.1) models.", "date": "2025-11-25", "categories": []}
{"id": "2512.00075", "title": "Adapter Shield: A Unified Framework with Built-in Authentication for Preventing Unauthorized Zero-Shot Image-to-Image Generation", "abstract": "With the rapid progress in diffusion models, image synthesis has advanced to the stage of zero-shot image-to-image generation, where high-fidelity replication of facial identities or artistic styles can be achieved using just one portrait or artwork, without modifying any model weights. Although these techniques significantly enhance creative possibilities, they also pose substantial risks related to intellectual property violations, including unauthorized identity cloning and stylistic imitation. To counter such threats, this work presents Adapter Shield, the first universal and authentication-integrated solution aimed at defending personal images from misuse in zero-shot generation scenarios. We first investigate how current zero-shot methods employ image encoders to extract embeddings from input images, which are subsequently fed into the UNet of diffusion models through cross-attention layers. Inspired by this mechanism, we construct a reversible encryption system that maps original embeddings into distinct encrypted representations according to different secret keys. The authorized users can restore the authentic embeddings via a decryption module and the correct key, enabling normal usage for authorized generation tasks. For protection purposes, we design a multi-target adversarial perturbation method that actively shifts the original embeddings toward designated encrypted patterns. Consequently, protected images are embedded with a defensive layer that ensures unauthorized users can only produce distorted or encrypted outputs. Extensive evaluations demonstrate that our method surpasses existing state-of-the-art defenses in blocking unauthorized zero-shot image synthesis, while supporting flexible and secure access control for verified users.", "date": "2025-11-25", "categories": []}
{"id": "2512.00119", "title": "NetDeTox: Adversarial and Efficient Evasion of Hardware-Security GNNs via RL-LLM Orchestration", "abstract": "Graph neural networks (GNNs) have shown promise in hardware security by learning structural motifs from netlist graphs. However, this reliance on motifs makes GNNs vulnerable to adversarial netlist rewrites; even small-scale edits can mislead GNN predictions. Existing adversarial approaches, ranging from synthesis-recipe perturbations to gate transformations, come with high design overheads. We present NetDeTox, an automated end-to-end framework that orchestrates large language models (LLMs) with reinforcement learning (RL) in a systematic manner, enabling focused local rewriting. The RL agent identifies netlist components critical for GNN-based reasoning, while the LLM devises rewriting plans to diversify motifs that preserve functionality. Iterative feedback between the RL and LLM stages refines adversarial rewritings to limit overheads. Compared to the SOTA work AttackGNN, NetDeTox successfully degrades the effectiveness of all security schemes with fewer rewrites and substantially lower area overheads (reductions of 54.50% for GNN-RE, 25.44% for GNN4IP, and 41.04% for OMLA, respectively). For GNN4IP, ours can even optimize/reduce the original benchmarks' area, in particular for larger circuits, demonstrating the practicality and scalability of NetDeTox.", "date": "2025-11-27", "categories": []}
{"id": "2512.00343", "title": "Assimilation Matters: Model-level Backdoor Detection in Vision-Language Pretrained Models", "abstract": "Vision-language pretrained models (VLPs) such as CLIP have achieved remarkable success, but are also highly vulnerable to backdoor attacks. Given a model fine-tuned by an untrusted third party, determining whether the model has been injected with a backdoor is a critical and challenging problem. Existing detection methods usually rely on prior knowledge of training dataset, backdoor triggers and targets, or downstream classifiers, which may be impractical for real-world applications. To address this, To address this challenge, we introduce Assimilation Matters in DETection (AMDET), a novel model-level detection framework that operates without any such prior knowledge. Specifically, we first reveal the feature assimilation property in backdoored text encoders: the representations of all tokens within a backdoor sample exhibit a high similarity. Further analysis attributes this effect to the concentration of attention weights on the trigger token. Leveraging this insight, AMDET scans a model by performing gradient-based inversion on token embeddings to recover implicit features that capable of activating backdoor behaviors. Furthermore, we identify the natural backdoor feature in the OpenAI's official CLIP model, which are not intentionally injected but still exhibit backdoor-like behaviors. We then filter them out from real injected backdoor by analyzing their loss landscapes. Extensive experiments on 3,600 backdoored and benign-finetuned models with two attack paradigms and three VLP model structures show that AMDET detects backdoors with an F1 score of 89.90%. Besides, it achieves one complete detection in approximately 5 minutes on a RTX 4090 GPU and exhibits strong robustness against adaptive attacks. Code is available at: https://github.com/Robin-WZQ/AMDET", "date": "2025-11-29", "categories": []}
{"id": "2512.01228", "title": "On the Tension Between Optimality and Adversarial Robustness in Policy Optimization", "abstract": "Achieving optimality and adversarial robustness in deep reinforcement learning has long been regarded as conflicting goals. Nonetheless, recent theoretical insights presented in CAR suggest a potential alignment, raising the important question of how to realize this in practice. This paper first identifies a key gap between theory and practice by comparing standard policy optimization (SPO) and adversarially robust policy optimization (ARPO). Although they share theoretical consistency, a fundamental tension between robustness and optimality arises in practical policy gradient methods. SPO tends toward convergence to vulnerable first-order stationary policies (FOSPs) with strong natural performance, whereas ARPO typically favors more robust FOSPs at the expense of reduced returns. Furthermore, we attribute this tradeoff to the reshaping effect of the strongest adversary in ARPO, which significantly complicates the global landscape by inducing deceptive sticky FOSPs. This improves robustness but makes navigation more challenging. To alleviate this, we develop the BARPO, a bilevel framework unifying SPO and ARPO by modulating adversary strength, thereby facilitating navigability while preserving global optima. Extensive empirical results demonstrate that BARPO consistently outperforms vanilla ARPO, providing a practical approach to reconcile theoretical and empirical performance.", "date": "2025-12-01", "categories": []}
{"id": "2512.01934", "title": "Physical ID-Transfer Attacks against Multi-Object Tracking via Adversarial Trajectory", "abstract": "Multi-Object Tracking (MOT) is a critical task in computer vision, with applications ranging from surveillance systems to autonomous driving. However, threats to MOT algorithms have yet been widely studied. In particular, incorrect association between the tracked objects and their assigned IDs can lead to severe consequences, such as wrong trajectory predictions. Previous attacks against MOT either focused on hijacking the trackers of individual objects, or manipulating the tracker IDs in MOT by attacking the integrated object detection (OD) module in the digital domain, which are model-specific, non-robust, and only able to affect specific samples in offline datasets. In this paper, we present AdvTraj, the first online and physical ID-manipulation attack against tracking-by-detection MOT, in which an attacker uses adversarial trajectories to transfer its ID to a targeted object to confuse the tracking system, without attacking OD. Our simulation results in CARLA show that AdvTraj can fool ID assignments with 100% success rate in various scenarios for white-box attacks against SORT, which also have high attack transferability (up to 93% attack success rate) against state-of-the-art (SOTA) MOT algorithms due to their common design principles. We characterize the patterns of trajectories generated by AdvTraj and propose two universal adversarial maneuvers that can be performed by a human walker/driver in daily scenarios. Our work reveals under-explored weaknesses in the object association phase of SOTA MOT systems, and provides insights into enhancing the robustness of such systems.", "date": "2025-12-01", "categories": []}
{"id": "2504.17457", "title": "Unveiling Hidden Vulnerabilities in Digital Human Generation via Adversarial Attacks", "abstract": "Expressive human pose and shape estimation (EHPS) is crucial for digital human generation, especially in applications like live streaming. While existing research primarily focuses on reducing estimation errors, it largely neglects robustness and security aspects, leaving these systems vulnerable to adversarial attacks. To address this significant challenge, we propose the \\textbf{Tangible Attack (TBA)}, a novel framework designed to generate adversarial examples capable of effectively compromising any digital human generation model. Our approach introduces a \\textbf{Dual Heterogeneous Noise Generator (DHNG)}, which leverages Variational Autoencoders (VAE) and ControlNet to produce diverse, targeted noise tailored to the original image features. Additionally, we design a custom \\textbf{adversarial loss function} to optimize the noise, ensuring both high controllability and potent disruption. By iteratively refining the adversarial sample through multi-gradient signals from both the noise and the state-of-the-art EHPS model, TBA substantially improves the effectiveness of adversarial attacks. Extensive experiments demonstrate TBA's superiority, achieving a remarkable 41.0\\% increase in estimation error, with an average improvement of approximately 17.0\\%. These findings expose significant security vulnerabilities in current EHPS models and highlight the need for stronger defenses in digital human generation systems.", "date": "2025-04-24", "categories": []}
{"id": "2505.19194", "title": "Curvature Dynamic Black-box Attack: revisiting adversarial robustness via dynamic curvature estimation", "abstract": "Adversarial attack reveals the vulnerability of deep learning models. It is assumed that high curvature may give rise to rough decision boundary and thus result in less robust models. However, the most commonly used \\textit{curvature} is the curvature of loss function, scores or other parameters from within the model as opposed to decision boundary curvature, since the former can be relatively easily formed using second order derivative. In this paper, we propose a new query-efficient method, dynamic curvature estimation (DCE), to estimate the decision boundary curvature in a black-box setting. Our approach is based on CGBA, a black-box adversarial attack. By performing DCE on a wide range of classifiers, we discovered, statistically, a connection between decision boundary curvature and adversarial robustness. We also propose a new attack method, curvature dynamic black-box attack (CDBA) with improved performance using the estimated curvature.", "date": "2025-05-25", "categories": []}
{"id": "2511.22044", "title": "Distillability of LLM Security Logic: Predicting Attack Success Rate of Outline Filling Attack via Ranking Regression", "abstract": "In the realm of black-box jailbreak attacks on large language models (LLMs), the feasibility of constructing a narrow safety proxy, a lightweight model designed to predict the attack success rate (ASR) of adversarial prompts, remains underexplored. This work investigates the distillability of an LLM's core security logic. We propose a novel framework that incorporates an improved outline filling attack to achieve dense sampling of the model's security boundaries. Furthermore, we introduce a ranking regression paradigm that replaces standard regression and trains the proxy model to predict which prompt yields a higher ASR. Experimental results show that our proxy model achieves an accuracy of 91.1 percent in predicting the relative ranking of average long response (ALR), and 69.2 percent in predicting ASR. These findings confirm the predictability and distillability of jailbreak behaviors, and demonstrate the potential of leveraging such distillability to optimize black-box attacks.", "date": "2025-11-27", "categories": []}
{"id": "2511.22147", "title": "RemedyGS: Defend 3D Gaussian Splatting against Computation Cost Attacks", "abstract": "As a mainstream technique for 3D reconstruction, 3D Gaussian splatting (3DGS) has been applied in a wide range of applications and services. Recent studies have revealed critical vulnerabilities in this pipeline and introduced computation cost attacks that lead to malicious resource occupancies and even denial-of-service (DoS) conditions, thereby hindering the reliable deployment of 3DGS. In this paper, we propose the first effective and comprehensive black-box defense framework, named RemedyGS, against such computation cost attacks, safeguarding 3DGS reconstruction systems and services. Our pipeline comprises two key components: a detector to identify the attacked input images with poisoned textures and a purifier to recover the benign images from their attacked counterparts, mitigating the adverse effects of these attacks. Moreover, we incorporate adversarial training into the purifier to enforce distributional alignment between the recovered and original natural images, thereby enhancing the defense efficacy. Experimental results demonstrate that our framework effectively defends against white-box, black-box, and adaptive attacks in 3DGS systems, achieving state-of-the-art performance in both safety and utility.", "date": "2025-11-27", "categories": []}
{"id": "2305.12825", "title": "Uncertainty-based Detection of Adversarial Attacks in Semantic Segmentation", "abstract": "State-of-the-art deep neural networks have proven to be highly powerful in a broad range of tasks, including semantic image segmentation. However, these networks are vulnerable against adversarial attacks, i.e., non-perceptible perturbations added to the input image causing incorrect predictions, which is hazardous in safety-critical applications like automated driving. Adversarial examples and defense strategies are well studied for the image classification task, while there has been limited research in the context of semantic segmentation. First works however show that the segmentation outcome can be severely distorted by adversarial attacks. In this work, we introduce an uncertainty-based approach for the detection of adversarial attacks in semantic segmentation. We observe that uncertainty as for example captured by the entropy of the output distribution behaves differently on clean and perturbed images and leverage this property to distinguish between the two cases. Our method works in a light-weight and post-processing manner, i.e., we do not modify the model or need knowledge of the process used for generating adversarial examples. In a thorough empirical analysis, we demonstrate the ability of our approach to detect perturbed images across multiple types of adversarial attacks.", "date": "2023-05-22", "categories": []}
{"id": "2406.19622", "title": "Data-Driven Lipschitz Continuity: A Cost-Effective Approach to Improve Adversarial Robustness", "abstract": "As deep neural networks (DNNs) are increasingly deployed in sensitive applications, ensuring their security and robustness has become critical. A major threat to DNNs arises from adversarial attacks, where small input perturbations can lead to incorrect predictions. Recent advances in adversarial training improve robustness by incorporating additional examples from external datasets or generative models. However, these methods often incur high computational costs, limiting their practicality and hindering real-world deployment. In this paper, we propose a cost-efficient alternative based on Lipschitz continuity that achieves robustness comparable to models trained with extensive supplementary data. Unlike conventional adversarial training, our method requires only a single pass over the dataset without gradient estimation, making it highly efficient. Furthermore, our method can integrate seamlessly with existing adversarial training frameworks and enhances the robustness of models without requiring extra generative data. Experimental results show that our approach not only reduces computational overhead but also maintains or improves the defensive capabilities of robust neural networks. This work opens a promising direction for developing practical, scalable defenses against adversarial attacks.", "date": "2024-06-28", "categories": []}
{"id": "2407.08806", "title": "HO-FMN: Hyperparameter Optimization for Fast Minimum-Norm Attacks", "abstract": "Gradient-based attacks are a primary tool to evaluate robustness of machine-learning models. However, many attacks tend to provide overly-optimistic evaluations as they use fixed loss functions, optimizers, step-size schedulers, and default hyperparameters. In this work, we tackle these limitations by proposing a parametric variation of the well-known fast minimum-norm attack algorithm, whose loss, optimizer, step-size scheduler, and hyperparameters can be dynamically adjusted. We re-evaluate 12 robust models, showing that our attack finds smaller adversarial perturbations without requiring any additional tuning. This also enables reporting adversarial robustness as a function of the perturbation budget, providing a more complete evaluation than that offered by fixed-budget attacks, while remaining efficient. We release our open-source code at https://github.com/pralab/HO-FMN.", "date": "2024-07-11", "categories": []}
{"id": "2412.18218", "title": "On the Effectiveness of Adversarial Training on Malware Classifiers", "abstract": "Adversarial Training (AT) is a key defense against Machine Learning evasion attacks, but its effectiveness for real-world malware detection remains poorly understood. This uncertainty stems from a critical disconnect in prior research: studies often overlook the inherent nature of malware and are fragmented, examining diverse variables like realism or confidence of adversarial examples in isolation, or relying on weak evaluations that yield non-generalizable insights. To address this, we introduce Rubik, a framework for the systematic, multi-dimensional evaluation of AT in the malware domain. This framework defines diverse key factors across essential dimensions, including data, feature representations, classifiers, and robust optimization settings, for a comprehensive exploration of the interplay of influential AT's variables through reliable evaluation practices, such as realistic evasion attacks. We instantiate Rubik on Android malware, empirically analyzing how this interplay shapes robustness. Our findings challenge prior beliefs--showing, for instance, that realizable adversarial examples offer only conditional robustness benefits--and reveal new insights, such as the critical role of model architecture and feature-space structure in determining AT's success. From this analysis, we distill four key insights, expose four common evaluation misconceptions, and offer practical recommendations to guide the development of truly robust malware classifiers.", "date": "2024-12-24", "categories": []}
{"id": "2508.14070", "title": "Special-Character Adversarial Attacks on Open-Source Language Model", "abstract": "Large language models (LLMs) have achieved remarkable performance across diverse natural language processing tasks, yet their vulnerability to character-level adversarial manipulations presents significant security challenges for real-world deployments. This paper presents a study of different special character attacks including unicode, homoglyph, structural, and textual encoding attacks aimed at bypassing safety mechanisms. We evaluate seven prominent open-source models ranging from 3.8B to 32B parameters on 4,000+ attack attempts. These experiments reveal critical vulnerabilities across all model sizes, exposing failure modes that include successful jailbreaks, incoherent outputs, and unrelated hallucinations.", "date": "2025-08-12", "categories": []}
{"id": "2511.19218", "title": "Adversarial Attack-Defense Co-Evolution for LLM Safety Alignment via Tree-Group Dual-Aware Search and Optimization", "abstract": "Large Language Models (LLMs) have developed rapidly in web services, delivering unprecedented capabilities while amplifying societal risks. Existing works tend to focus on either isolated jailbreak attacks or static defenses, neglecting the dynamic interplay between evolving threats and safeguards in real-world web contexts. To mitigate these challenges, we propose ACE-Safety (Adversarial Co-Evolution for LLM Safety), a novel framework that jointly optimize attack and defense models by seamlessly integrating two key innovative procedures: (1) Group-aware Strategy-guided Monte Carlo Tree Search (GS-MCTS), which efficiently explores jailbreak strategies to uncover vulnerabilities and generate diverse adversarial samples; (2) Adversarial Curriculum Tree-aware Group Policy Optimization (AC-TGPO), which jointly trains attack and defense LLMs with challenging samples via curriculum reinforcement learning, enabling robust mutual improvement. Evaluations across multiple benchmarks demonstrate that our method outperforms existing attack and defense approaches, and provides a feasible pathway for developing LLMs that can sustainably support responsible AI ecosystems.", "date": "2025-11-24", "categories": []}
{"id": "2511.20992", "title": "Dataset Poisoning Attacks on Behavioral Cloning Policies", "abstract": "Behavior Cloning (BC) is a popular framework for training sequential decision policies from expert demonstrations via supervised learning. As these policies are increasingly being deployed in the real world, their robustness and potential vulnerabilities are an important concern. In this work, we perform the first analysis of the efficacy of clean-label backdoor attacks on BC policies. Our backdoor attacks poison a dataset of demonstrations by injecting a visual trigger to create a spurious correlation that can be exploited at test time. We evaluate how policy vulnerability scales with the fraction of poisoned data, the strength of the trigger, and the trigger type. We also introduce a novel entropy-based test-time trigger attack that substantially degrades policy performance by identifying critical states where test-time triggering of the backdoor is expected to be most effective at degrading performance. We empirically demonstrate that BC policies trained on even minimally poisoned datasets exhibit deceptively high, near-baseline task performance despite being highly vulnerable to backdoor trigger attacks during deployment. Our results underscore the urgent need for more research into the robustness of BC policies, particularly as large-scale datasets are increasingly used to train policies for real-world cyber-physical systems. Videos and code are available at https://sites.google.com/view/dataset-poisoning-in-bc.", "date": "2025-11-26", "categories": []}
{"id": "2511.21180", "title": "CAHS-Attack: CLIP-Aware Heuristic Search Attack Method for Stable Diffusion", "abstract": "Diffusion models exhibit notable fragility when faced with adversarial prompts, and strengthening attack capabilities is crucial for uncovering such vulnerabilities and building more robust generative systems. Existing works often rely on white-box access to model gradients or hand-crafted prompt engineering, which is infeasible in real-world deployments due to restricted access or poor attack effect. In this paper, we propose CAHS-Attack , a CLIP-Aware Heuristic Search attack method. CAHS-Attack integrates Monte Carlo Tree Search (MCTS) to perform fine-grained suffix optimization, leveraging a constrained genetic algorithm to preselect high-potential adversarial prompts as root nodes, and retaining the most semantically disruptive outcome at each simulation rollout for efficient local search. Extensive experiments demonstrate that our method achieves state-of-the-art attack performance across both short and long prompts of varying semantics. Furthermore, we find that the fragility of SD models can be attributed to the inherent vulnerability of their CLIP-based text encoders, suggesting a fundamental security risk in current text-to-image pipelines.", "date": "2025-11-26", "categories": []}
{"id": "2511.21574", "title": "Multimodal Robust Prompt Distillation for 3D Point Cloud Models", "abstract": "Adversarial attacks pose a significant threat to learning-based 3D point cloud models, critically undermining their reliability in security-sensitive applications. Existing defense methods often suffer from (1) high computational overhead and (2) poor generalization ability across diverse attack types. To bridge these gaps, we propose a novel yet efficient teacher-student framework, namely Multimodal Robust Prompt Distillation (MRPD) for distilling robust 3D point cloud model. It learns lightweight prompts by aligning student point cloud model's features with robust embeddings from three distinct teachers: a vision model processing depth projections, a high-performance 3D model, and a text encoder. To ensure a reliable knowledge transfer, this distillation is guided by a confidence-gated mechanism which dynamically balances the contribution of all input modalities. Notably, since the distillation is all during the training stage, there is no additional computational cost at inference. Extensive experiments demonstrate that MRPD substantially outperforms state-of-the-art defense methods against a wide range of white-box and black-box attacks, while even achieving better performance on clean data. Our work presents a new, practical paradigm for building robust 3D vision systems by efficiently harnessing multimodal knowledge.", "date": "2025-11-26", "categories": []}
{"id": "2511.21663", "title": "Attention-Guided Patch-Wise Sparse Adversarial Attacks on Vision-Language-Action Models", "abstract": "In recent years, Vision-Language-Action (VLA) models in embodied intelligence have developed rapidly. However, existing adversarial attack methods require costly end-to-end training and often generate noticeable perturbation patches. To address these limitations, we propose ADVLA, a framework that directly applies adversarial perturbations on features projected from the visual encoder into the textual feature space. ADVLA efficiently disrupts downstream action predictions under low-amplitude constraints, and attention guidance allows the perturbations to be both focused and sparse. We introduce three strategies that enhance sensitivity, enforce sparsity, and concentrate perturbations. Experiments demonstrate that under an $L_{\\infty}=4/255$ constraint, ADVLA combined with Top-K masking modifies less than 10% of the patches while achieving an attack success rate of nearly 100%. The perturbations are concentrated on critical regions, remain almost imperceptible in the overall image, and a single-step iteration takes only about 0.06 seconds, significantly outperforming conventional patch-based attacks. In summary, ADVLA effectively weakens downstream action predictions of VLA models under low-amplitude and locally sparse conditions, avoiding the high training costs and conspicuous perturbations of traditional patch attacks, and demonstrates unique effectiveness and practical value for attacking VLA feature spaces.", "date": "2025-11-26", "categories": []}
{"id": "2505.11895", "title": "Adversarial Robustness for Unified Multi-Modal Encoders via Efficient Calibration", "abstract": "Recent unified multi-modal encoders align a wide range of modalities into a shared representation space, enabling diverse cross-modal tasks. Despite their impressive capabilities, the robustness of these models under adversarial perturbations remains underexplored, which is a critical concern for safety-sensitive applications. In this work, we present the first comprehensive study of adversarial vulnerability in unified multi-modal encoders. We find that even mild adversarial perturbations lead to substantial performance drops across all modalities. Non-visual inputs, such as audio and point clouds, are especially fragile, while visual inputs like images and videos also degrade significantly. To address this, we propose an efficient adversarial calibration framework that improves robustness across modalities without modifying pretrained encoders or semantic centers, ensuring compatibility with existing foundation models. Our method introduces modality-specific projection heads trained solely on adversarial examples, while keeping the backbone and embeddings frozen. We explore three training objectives: fixed-center cross-entropy, clean-to-adversarial L2 alignment, and clean-adversarial InfoNCE, and we introduce a regularization strategy to ensure modality-consistent alignment under attack. Experiments on six modalities and three Bind-style models show that our method improves adversarial robustness by up to 47.3 percent at epsilon = 4/255, while preserving or even improving clean zero-shot and retrieval performance with less than 1 percent trainable parameters.", "date": "2025-05-17", "categories": []}
{"id": "2504.01632", "title": "Benchmarking the Spatial Robustness of DNNs via Natural and Adversarial Localized Corruptions", "abstract": "The robustness of deep neural networks is a crucial factor in safety-critical applications, particularly in complex and dynamic environments (e.g., medical or driving scenarios) where localized corruptions can arise. While previous studies have evaluated the robustness of semantic segmentation (SS) models under whole-image natural or adversarial corruptions, a comprehensive investigation into the spatial robustness of dense vision models under localized corruptions remains underexplored. This paper fills this gap by introducing novel, region-aware metrics for benchmarking the spatial robustness of segmentation models, along with an evaluation framework to assess the impact of natural localized corruptions. Furthermore, it uncovers the inherent complexity of evaluating worst-case spatial robustness using only a single localized adversarial attack. To address this, the work proposes a region-aware multi-attack adversarial analysis to systematically assess model robustness across specific image regions. The proposed metrics and analysis were exploited to evaluate 14 segmentation models in driving scenarios, uncovering key insights into the effects of localized corruption in both natural and adversarial forms. The results reveal that models respond to these two types of threats differently; for instance, transformer-based segmentation models demonstrate notable robustness to localized natural corruptions but are highly vulnerable to adversarial ones, and vice versa for CNN-based models. Consequently, we also address the challenge of balancing robustness to both natural and adversarial localized corruptions by means of ensemble models, thereby achieving a broader threat coverage and improved reliability for dense vision tasks.", "date": "2025-04-02", "categories": []}
{"id": "2505.23518", "title": "TRAP: Targeted Redirecting of Agentic Preferences", "abstract": "Autonomous agentic AI systems powered by vision-language models (VLMs) are rapidly advancing toward real-world deployment, yet their cross-modal reasoning capabilities introduce new attack surfaces for adversarial manipulation that exploit semantic reasoning across modalities. Existing adversarial attacks typically rely on visible pixel perturbations or require privileged model or environment access, making them impractical for stealthy, real-world exploitation. We introduce TRAP, a novel generative adversarial framework that manipulates the agent's decision-making using diffusion-based semantic injections into the vision-language embedding space. Our method combines negative prompt-based degradation with positive semantic optimization, guided by a Siamese semantic network and layout-aware spatial masking. Without requiring access to model internals, TRAP produces visually natural images yet induces consistent selection biases in agentic AI systems. We evaluate TRAP on the Microsoft Common Objects in Context (COCO) dataset, building multi-candidate decision scenarios. Across these scenarios, TRAP consistently induces decision-level preference redirection on leading models, including LLaVA-34B, Gemma3, GPT-4o, and Mistral-3.2, significantly outperforming existing baselines such as SPSA, Bandit, and standard diffusion approaches. These findings expose a critical, generalized vulnerability: autonomous agents can be consistently misled through visually subtle, semantically-guided cross-modal manipulations. Overall, our results show the need for defense strategies beyond pixel-level robustness to address semantic vulnerabilities in cross-modal decision-making. The code for TRAP is accessible on GitHub at https://github.com/uiuc-focal-lab/TRAP.", "date": "2025-05-29", "categories": []}
{"id": "2507.04726", "title": "Backdoors in Conditional Diffusion: Threats to Responsible Synthetic Data Pipelines", "abstract": "Text-to-image diffusion models achieve high-fidelity image generation from natural language prompts. ControlNets extend these models by enabling conditioning on structural inputs (e.g., edge maps, depth, pose), providing fine-grained control over outputs. Yet their reliance on large, publicly scraped datasets and community fine-tuning makes them vulnerable to data poisoning. We introduce a model-poisoning attack that embeds a covert backdoor into a ControlNet, causing it to produce attacker-specified content when exposed to visual triggers, without textual prompts. Experiments show that poisoning only 1% of the fine-tuning corpus yields a 90-98% attack success rate, while 5% further strengthens the backdoor, all while preserving normal generation quality. To mitigate this risk, we propose clean fine-tuning (CFT): freezing the diffusion backbone and fine-tuning only the ControlNet on a sanitized dataset with a reduced learning rate. CFT lowers attack success rates on held-out data. These results expose a critical security weakness in open-source, ControlNet-guided diffusion pipelines and demonstrate that CFT offers a practical defense for responsible synthetic-data pipelines.", "date": "2025-07-07", "categories": []}
{"id": "2509.22850", "title": "Boundary on the Table: Efficient Black-Box Decision-Based Attacks for Structured Data", "abstract": "Adversarial robustness in structured data remains an underexplored frontier compared to vision and language domains. In this work, we introduce a novel black-box, decision-based adversarial attack tailored for tabular data. Our approach combines gradient-free direction estimation with an iterative boundary search, enabling efficient navigation of discrete and continuous feature spaces under minimal oracle access. Extensive experiments demonstrate that our method successfully compromises nearly the entire test set across diverse models, ranging from classical machine learning classifiers to large language model (LLM)-based pipelines. Remarkably, the attack achieves success rates consistently above 90%, while requiring only a small number of queries per instance. These results highlight the critical vulnerability of tabular models to adversarial perturbations, underscoring the urgent need for stronger defenses in real-world decision-making systems.", "date": "2025-09-26", "categories": []}
{"id": "2511.06852", "title": "Differentiated Directional Intervention A Framework for Evading LLM Safety Alignment", "abstract": "Safety alignment instills in Large Language Models (LLMs) a critical capacity to refuse malicious requests. Prior works have modeled this refusal mechanism as a single linear direction in the activation space. We posit that this is an oversimplification that conflates two functionally distinct neural processes: the detection of harm and the execution of a refusal. In this work, we deconstruct this single representation into a Harm Detection Direction and a Refusal Execution Direction. Leveraging this fine-grained model, we introduce Differentiated Bi-Directional Intervention (DBDI), a new white-box framework that precisely neutralizes the safety alignment at critical layer. DBDI applies adaptive projection nullification to the refusal execution direction while suppressing the harm detection direction via direct steering. Extensive experiments demonstrate that DBDI outperforms prominent jailbreaking methods, achieving up to a 97.88\\% attack success rate on models such as Llama-2. By providing a more granular and mechanistic framework, our work offers a new direction for the in-depth understanding of LLM safety alignment.", "date": "2025-11-10", "categories": []}
{"id": "2511.17666", "title": "Evaluating Adversarial Vulnerabilities in Modern Large Language Models", "abstract": "The recent boom and rapid integration of Large Language Models (LLMs) into a wide range of applications warrants a deeper understanding of their security and safety vulnerabilities. This paper presents a comparative analysis of the susceptibility to jailbreak attacks for two leading publicly available LLMs, Google's Gemini 2.5 Flash and OpenAI's GPT-4 (specifically the GPT-4o mini model accessible in the free tier). The research utilized two main bypass strategies: 'self-bypass', where models were prompted to circumvent their own safety protocols, and 'cross-bypass', where one model generated adversarial prompts to exploit vulnerabilities in the other. Four attack methods were employed - direct injection, role-playing, context manipulation, and obfuscation - to generate five distinct categories of unsafe content: hate speech, illegal activities, malicious code, dangerous content, and misinformation. The success of the attack was determined by the generation of disallowed content, with successful jailbreaks assigned a severity score. The findings indicate a disparity in jailbreak susceptibility between 2.5 Flash and GPT-4, suggesting variations in their safety implementations or architectural design. Cross-bypass attacks were particularly effective, indicating that an ample amount of vulnerabilities exist in the underlying transformer architecture. This research contributes a scalable framework for automated AI red-teaming and provides data-driven insights into the current state of LLM safety, underscoring the complex challenge of balancing model capabilities with robust safety mechanisms.", "date": "2025-11-21", "categories": []}
{"id": "2511.17688", "title": "Enhancing Adversarial Transferability through Block Stretch and Shrink", "abstract": "Adversarial attacks introduce small, deliberately crafted perturbations that mislead neural networks, and their transferability from white-box to black-box target models remains a critical research focus. Input transformation-based attacks are a subfield of adversarial attacks that enhance input diversity through input transformations to improve the transferability of adversarial examples. However, existing input transformation-based attacks tend to exhibit limited cross-model transferability. Previous studies have shown that high transferability is associated with diverse attention heatmaps and the preservation of global semantics in transformed inputs. Motivated by this observation, we propose Block Stretch and Shrink (BSS), a method that divides an image into blocks and applies stretch and shrink operations to these blocks, thereby diversifying attention heatmaps in transformed inputs while maintaining their global semantics. Empirical evaluations on a subset of ImageNet demonstrate that BSS outperforms existing input transformation-based attack methods in terms of transferability. Furthermore, we examine the impact of the number scale, defined as the number of transformed inputs, in input transformation-based attacks, and advocate evaluating these methods under a unified number scale to enable fair and comparable assessments.", "date": "2025-11-21", "categories": []}
{"id": "2511.17982", "title": "Towards Effective, Stealthy, and Persistent Backdoor Attacks Targeting Graph Foundation Models", "abstract": "Graph Foundation Models (GFMs) are pre-trained on diverse source domains and adapted to unseen targets, enabling broad generalization for graph machine learning. Despite that GFMs have attracted considerable attention recently, their vulnerability to backdoor attacks remains largely underexplored. A compromised GFM can introduce backdoor behaviors into downstream applications, posing serious security risks. However, launching backdoor attacks against GFMs is non-trivial due to three key challenges. (1) Effectiveness: Attackers lack knowledge of the downstream task during pre-training, complicating the assurance that triggers reliably induce misclassifications into desired classes. (2) Stealthiness: The variability in node features across domains complicates trigger insertion that remains stealthy. (3) Persistence: Downstream fine-tuning may erase backdoor behaviors by updating model parameters. To address these challenges, we propose GFM-BA, a novel Backdoor Attack model against Graph Foundation Models. Specifically, we first design a label-free trigger association module that links the trigger to a set of prototype embeddings, eliminating the need for knowledge about downstream tasks to perform backdoor injection. Then, we introduce a node-adaptive trigger generator, dynamically producing node-specific triggers, reducing the risk of trigger detection while reliably activating the backdoor. Lastly, we develop a persistent backdoor anchoring module that firmly anchors the backdoor to fine-tuning-insensitive parameters, enhancing the persistence of the backdoor under downstream adaptation. Extensive experiments demonstrate the effectiveness, stealthiness, and persistence of GFM-BA.", "date": "2025-11-22", "categories": []}
{"id": "2511.18138", "title": "Vulnerability-Aware Robust Multimodal Adversarial Training", "abstract": "Multimodal learning has shown significant superiority on various tasks by integrating multiple modalities. However, the interdependencies among modalities increase the susceptibility of multimodal models to adversarial attacks. Existing methods mainly focus on attacks on specific modalities or indiscriminately attack all modalities. In this paper, we find that these approaches ignore the differences between modalities in their contribution to final robustness, resulting in suboptimal robustness performance. To bridge this gap, we introduce Vulnerability-Aware Robust Multimodal Adversarial Training (VARMAT), a probe-in-training adversarial training method that improves multimodal robustness by identifying the vulnerability of each modality. To be specific, VARMAT first explicitly quantifies the vulnerability of each modality, grounded in a first-order approximation of the attack objective (Probe). Then, we propose a targeted regularization term that penalizes modalities with high vulnerability, guiding robust learning while maintaining task accuracy (Training). We demonstrate the enhanced robustness of our method across multiple multimodal datasets involving diverse modalities. Finally, we achieve {12.73%, 22.21%, 11.19%} robustness improvement on three multimodal datasets, revealing a significant blind spot in multimodal adversarial training.", "date": "2025-11-22", "categories": []}
{"id": "2511.18223", "title": "A Novel and Practical Universal Adversarial Perturbations against Deep Reinforcement Learning based Intrusion Detection Systems", "abstract": "Intrusion Detection Systems (IDS) play a vital role in defending modern cyber physical systems against increasingly sophisticated cyber threats. Deep Reinforcement Learning-based IDS, have shown promise due to their adaptive and generalization capabilities. However, recent studies reveal their vulnerability to adversarial attacks, including Universal Adversarial Perturbations (UAPs), which can deceive models with a single, input-agnostic perturbation. In this work, we propose a novel UAP attack against Deep Reinforcement Learning (DRL)-based IDS under the domain-specific constraints derived from network data rules and feature relationships. To the best of our knowledge, there is no existing study that has explored UAP generation for the DRL-based IDS. In addition, this is the first work that focuses on developing a UAP against a DRL-based IDS under realistic domain constraints based on not only the basic domain rules but also mathematical relations between the features. Furthermore, we enhance the evasion performance of the proposed UAP, by introducing a customized loss function based on the Pearson Correlation Coefficient, and we denote it as Customized UAP. To the best of our knowledge, this is also the first work using the PCC value in the UAP generation, even in the broader context. Four additional established UAP baselines are implemented for a comprehensive comparison. Experimental results demonstrate that our proposed Customized UAP outperforms two input-dependent attacks including Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM), and four UAP baselines, highlighting its effectiveness for real-world adversarial scenarios.", "date": "2025-11-22", "categories": []}
{"id": "2511.18656", "title": "Robust Physical Adversarial Patches Using Dynamically Optimized Clusters", "abstract": "Physical adversarial attacks on deep learning systems is concerning due to the ease of deploying such attacks, usually by placing an adversarial patch in a scene to manipulate the outcomes of a deep learning model. Training such patches typically requires regularization that improves physical realizability (e.g., printability, smoothness) and/or robustness to real-world variability (e.g. deformations, viewing angle, noise). One type of variability that has received little attention is scale variability. When a patch is rescaled, either digitally through downsampling/upsampling or physically through changing imaging distances, interpolation-induced color mixing occurs. This smooths out pixel values, resulting in a loss of high-frequency patterns and degrading the adversarial signal. To address this, we present a novel superpixel-based regularization method that guides patch optimization to scale-resilient structures. Our ap proach employs the Simple Linear Iterative Clustering (SLIC) algorithm to dynamically cluster pixels in an adversarial patch during optimization. The Implicit Function Theorem is used to backpropagate gradients through SLIC to update the superpixel boundaries and color. This produces patches that maintain their structure over scale and are less susceptible to interpolation losses. Our method achieves greater performance in the digital domain, and when realized physically, these performance gains are preserved, leading to improved physical performance. Real-world performance was objectively assessed using a novel physical evaluation protocol that utilizes screens and cardboard cut-outs to systematically vary real-world conditions.", "date": "2025-11-23", "categories": []}
{"id": "2511.18921", "title": "BackdoorVLM: A Benchmark for Backdoor Attacks on Vision-Language Models", "abstract": "Backdoor attacks undermine the reliability and trustworthiness of machine learning systems by injecting hidden behaviors that can be maliciously activated at inference time. While such threats have been extensively studied in unimodal settings, their impact on multimodal foundation models, particularly vision-language models (VLMs), remains largely underexplored. In this work, we introduce \\textbf{BackdoorVLM}, the first comprehensive benchmark for systematically evaluating backdoor attacks on VLMs across a broad range of settings. It adopts a unified perspective that injects and analyzes backdoors across core vision-language tasks, including image captioning and visual question answering. BackdoorVLM organizes multimodal backdoor threats into 5 representative categories: targeted refusal, malicious injection, jailbreak, concept substitution, and perceptual hijack. Each category captures a distinct pathway through which an adversary can manipulate a model's behavior. We evaluate these threats using 12 representative attack methods spanning text, image, and bimodal triggers, tested on 2 open-source VLMs and 3 multimodal datasets. Our analysis reveals that VLMs exhibit strong sensitivity to textual instructions, and in bimodal backdoors the text trigger typically overwhelms the image trigger when forming the backdoor mapping. Notably, backdoors involving the textual modality remain highly potent, with poisoning rates as low as 1\\% yielding over 90\\% success across most tasks. These findings highlight significant, previously underexplored vulnerabilities in current VLMs. We hope that BackdoorVLM can serve as a useful benchmark for analyzing and mitigating multimodal backdoor threats. Code is available at: https://github.com/bin015/BackdoorVLM .", "date": "2025-11-24", "categories": []}
{"id": "2511.19248", "title": "FedPoisonTTP: A Threat Model and Poisoning Attack for Federated Test-Time Personalization", "abstract": "Test-time personalization in federated learning enables models at clients to adjust online to local domain shifts, enhancing robustness and personalization in deployment. Yet, existing federated learning work largely overlooks the security risks that arise when local adaptation occurs at test time. Heterogeneous domain arrivals, diverse adaptation algorithms, and limited cross-client visibility create vulnerabilities where compromised participants can craft poisoned inputs and submit adversarial updates that undermine both global and per-client performance. To address this threat, we introduce FedPoisonTTP, a realistic grey-box attack framework that explores test-time data poisoning in the federated adaptation setting. FedPoisonTTP distills a surrogate model from adversarial queries, synthesizes in-distribution poisons using feature-consistency, and optimizes attack objectives to generate high-entropy or class-confident poisons that evade common adaptation filters. These poisons are injected during local adaptation and spread through collaborative updates, leading to broad degradation. Extensive experiments on corrupted vision benchmarks show that compromised participants can substantially diminish overall test-time performance.", "date": "2025-11-24", "categories": []}
{"id": "2511.19254", "title": "Adversarial Patch Attacks on Vision-Based Cargo Occupancy Estimation via Differentiable 3D Simulation", "abstract": "Computer vision systems are increasingly adopted in modern logistics operations, including the estimation of trailer occupancy for planning, routing, and billing. Although effective, such systems may be vulnerable to physical adversarial attacks, particularly adversarial patches that can be printed and placed on interior surfaces. In this work, we study the feasibility of such attacks on a convolutional cargo-occupancy classifier using fully simulated 3D environments. Using Mitsuba 3 for differentiable rendering, we optimize patch textures across variations in geometry, lighting, and viewpoint, and compare their effectiveness to a 2D compositing baseline. Our experiments demonstrate that 3D-optimized patches achieve high attack success rates, especially in a denial-of-service scenario (empty to full), where success reaches 84.94 percent. Concealment attacks (full to empty) prove more challenging but still reach 30.32 percent. We analyze the factors influencing attack success, discuss implications for the security of automated logistics pipelines, and highlight directions for strengthening physical robustness. To our knowledge, this is the first study to investigate adversarial patch attacks for cargo-occupancy estimation in physically realistic, fully simulated 3D scenes.", "date": "2025-11-24", "categories": []}
{"id": "2511.19257", "title": "Medusa: Cross-Modal Transferable Adversarial Attacks on Multimodal Medical Retrieval-Augmented Generation", "abstract": "With the rapid advancement of retrieval-augmented vision-language models, multimodal medical retrieval-augmented generation (MMed-RAG) systems are increasingly adopted in clinical decision support. These systems enhance medical applications by performing cross-modal retrieval to integrate relevant visual and textual evidence for tasks, e.g., report generation and disease diagnosis. However, their complex architecture also introduces underexplored adversarial vulnerabilities, particularly via visual input perturbations. In this paper, we propose Medusa, a novel framework for crafting cross-modal transferable adversarial attacks on MMed-RAG systems under a black-box setting. Specifically, Medusa formulates the attack as a perturbation optimization problem, leveraging a multi-positive InfoNCE loss (MPIL) to align adversarial visual embeddings with medically plausible but malicious textual targets, thereby hijacking the retrieval process. To enhance transferability, we adopt a surrogate model ensemble and design a dual-loop optimization strategy augmented with invariant risk minimization (IRM). Extensive experiments on two real-world medical tasks, including medical report generation and disease diagnosis, demonstrate that Medusa achieves over 90% average attack success rate across various generation models and retrievers under appropriate parameter configuration, while remaining robust against four mainstream defenses, outperforming state-of-the-art baselines. Our results reveal critical vulnerabilities in the MMed-RAG systems and highlight the necessity of robustness benchmarking in safety-critical medical applications. The code and data are available at https://anonymous.4open.science/r/MMed-RAG-Attack-F05A.", "date": "2025-11-24", "categories": []}
{"id": "2511.19330", "title": "Targeted Manipulation: Slope-Based Attacks on Financial Time-Series Data", "abstract": "A common method of attacking deep learning models is through adversarial attacks, which occur when an attacker specifically modifies the input of a model to produce an incorrect result. Adversarial attacks have been deeply investigated in the image domain; however, there is less research in the time-series domain and very little for forecasting financial data. To address these concerns, this study aims to build upon previous research on adversarial attacks for time-series data by introducing two new slope-based methods aimed to alter the trends of the predicted stock forecast generated by an N-HiTS model. Compared to the normal N-HiTS predictions, the two new slope-based methods, the General Slope Attack and Least-Squares Slope Attack, can manipulate N-HiTS predictions by doubling the slope. These new slope attacks can bypass standard security mechanisms, such as a discriminator that filters real and perturbed inputs, reducing a 4-layered CNN's specificity to 28% and accuracy to 57%. Furthermore, the slope based methods were incorporated into a GAN architecture as a means of generating realistic synthetic data, while simultaneously fooling the model. Finally, this paper also proposes a sample malware designed to inject an adversarial attack in the model inference library, proving that ML-security research should not only focus on making the model safe, but also securing the entire pipeline.", "date": "2025-11-24", "categories": []}
{"id": "2503.08636", "title": "Birds look like cars: Adversarial analysis of intrinsically interpretable deep learning", "abstract": "A common belief is that intrinsically interpretable deep learning models ensure a correct, intuitive understanding of their behavior and offer greater robustness against accidental errors or intentional manipulation. However, these beliefs have not been comprehensively verified, and growing evidence casts doubt on them. In this paper, we highlight the risks related to overreliance and susceptibility to adversarial manipulation of these so-called \"intrinsically (aka inherently) interpretable\" models by design. We introduce two strategies for adversarial analysis with prototype manipulation and backdoor attacks against prototype-based networks, and discuss how concept bottleneck models defend against these attacks. Fooling the model's reasoning by exploiting its use of latent prototypes manifests the inherent uninterpretability of deep neural networks, leading to a false sense of security reinforced by a visual confirmation bias. The reported limitations of part-prototype networks put their trustworthiness and applicability into question, motivating further work on the robustness and alignment of (deep) interpretable models.", "date": "2025-03-11", "categories": []}
{"id": "2503.17987", "title": "Reason2Attack: Jailbreaking Text-to-Image Models via LLM Reasoning", "abstract": "Text-to-Image(T2I) models typically deploy safety filters to prevent the generation of sensitive images. Unfortunately, recent jailbreaking attack methods manually design instructions for the LLM to generate adversarial prompts, which effectively bypass safety filters while producing sensitive images, exposing safety vulnerabilities of T2I models. However, due to the LLM's limited understanding of the T2I model and its safety filters, existing methods require numerous queries to achieve a successful attack, limiting their practical applicability. To address this issue, we propose Reason2Attack(R2A), which aims to enhance the LLM's reasoning capabilities in generating adversarial prompts by incorporating the jailbreaking attack into the post-training process of the LLM. Specifically, we first propose a CoT example synthesis pipeline based on Frame Semantics, which generates adversarial prompts by identifying related terms and corresponding context illustrations. Using CoT examples generated by the pipeline, we fine-tune the LLM to understand the reasoning path and format the output structure. Subsequently, we incorporate the jailbreaking attack task into the reinforcement learning process of the LLM and design an attack process reward that considers prompt length, prompt stealthiness, and prompt effectiveness, aiming to further enhance reasoning accuracy. Extensive experiments on various T2I models show that R2A achieves a better attack success ratio while requiring fewer queries than baselines. Moreover, our adversarial prompts demonstrate strong attack transferability across both open-source and commercial T2I models.", "date": "2025-03-23", "categories": []}
{"id": "2507.10998", "title": "Crafting Imperceptible On-Manifold Adversarial Attacks for Tabular Data", "abstract": "Adversarial attacks on tabular data present unique challenges due to the heterogeneous nature of mixed categorical and numerical features. Unlike images where pixel perturbations maintain visual similarity, tabular data lacks intuitive similarity metrics, making it difficult to define imperceptible modifications. Additionally, traditional gradient-based methods prioritise $\\ell_p$-norm constraints, often producing adversarial examples that deviate from the original data distributions. To address this, we propose a latent-space perturbation framework using a mixed-input Variational Autoencoder (VAE) to generate statistically consistent adversarial examples. The proposed VAE integrates categorical embeddings and numerical features into a unified latent manifold, enabling perturbations that preserve statistical consistency. We introduce In-Distribution Success Rate (IDSR) to jointly evaluate attack effectiveness and distributional alignment. Evaluation across six publicly available datasets and three model architectures demonstrates that our method achieves substantially lower outlier rates and more consistent performance compared to traditional input-space attacks and other VAE-based methods adapted from image domain approaches, achieving substantially lower outlier rates and higher IDSR across six datasets and three model architectures. Our comprehensive analyses of hyperparameter sensitivity, sparsity control, and generative architecture demonstrate that the effectiveness of VAE-based attacks depends strongly on reconstruction quality and the availability of sufficient training data. When these conditions are met, the proposed framework achieves superior practical utility and stability compared with input-space methods. This work underscores the importance of maintaining on-manifold perturbations for generating realistic and robust adversarial examples in tabular domains.", "date": "2025-07-15", "categories": []}
{"id": "2511.12233", "title": "Model Inversion Attack Against Deep Hashing", "abstract": "Deep hashing improves retrieval efficiency through compact binary codes, yet it introduces severe and often overlooked privacy risks. The ability to reconstruct original training data from hash codes could lead to serious threats such as biometric forgery and privacy breaches. However, model inversion attacks specifically targeting deep hashing models remain unexplored, leaving their security implications unexamined. This research gap stems from the inaccessibility of genuine training hash codes and the highly discrete Hamming space, which prevents existing methods from adapting to deep hashing. To address these challenges, we propose DHMI, the first diffusion-based model inversion framework designed for deep hashing. DHMI first clusters an auxiliary dataset to derive semantic hash centers as surrogate anchors. It then introduces a surrogate-guided denoising optimization method that leverages a novel attack metric (fusing classification consistency and hash proximity) to dynamically select candidate samples. A cluster of surrogate models guides the refinement of these candidates, ensuring the generation of high-fidelity and semantically consistent images. Experiments on multiple datasets demonstrate that DHMI successfully reconstructs high-resolution, high-quality images even under the most challenging black-box setting, where no training hash codes are available. Our method outperforms the existing state-of-the-art model inversion attacks in black-box scenarios, confirming both its practical efficacy and the critical privacy risks inherent in deep hashing systems.", "date": "2025-11-15", "categories": []}
{"id": "2511.16709", "title": "AutoBackdoor: Automating Backdoor Attacks via LLM Agents", "abstract": "Backdoor attacks pose a serious threat to the secure deployment of large language models (LLMs), enabling adversaries to implant hidden behaviors triggered by specific inputs. However, existing methods often rely on manually crafted triggers and static data pipelines, which are rigid, labor-intensive, and inadequate for systematically evaluating modern defense robustness. As AI agents become increasingly capable, there is a growing need for more rigorous, diverse, and scalable \\textit{red-teaming frameworks} that can realistically simulate backdoor threats and assess model resilience under adversarial conditions. In this work, we introduce \\textsc{AutoBackdoor}, a general framework for automating backdoor injection, encompassing trigger generation, poisoned data construction, and model fine-tuning via an autonomous agent-driven pipeline. Unlike prior approaches, AutoBackdoor uses a powerful language model agent to generate semantically coherent, context-aware trigger phrases, enabling scalable poisoning across arbitrary topics with minimal human effort. We evaluate AutoBackdoor under three realistic threat scenarios, including \\textit{Bias Recommendation}, \\textit{Hallucination Injection}, and \\textit{Peer Review Manipulation}, to simulate a broad range of attacks. Experiments on both open-source and commercial models, including LLaMA-3, Mistral, Qwen, and GPT-4o, demonstrate that our method achieves over 90\\% attack success with only a small number of poisoned samples. More importantly, we find that existing defenses often fail to mitigate these attacks, underscoring the need for more rigorous and adaptive evaluation techniques against agent-driven threats as explored in this work. All code, datasets, and experimental configurations will be merged into our primary repository at https://github.com/bboylyg/BackdoorLLM.", "date": "2025-11-20", "categories": []}
{"id": "2511.17448", "title": "MMT-ARD: Multimodal Multi-Teacher Adversarial Distillation for Robust Vision-Language Models", "abstract": "Vision-Language Models (VLMs) are increasingly deployed in safety-critical applications, making their adversarial robustness a crucial concern. While adversarial knowledge distillation has shown promise in transferring robustness from teacher to student models, traditional single-teacher approaches suffer from limited knowledge diversity, slow convergence, and difficulty in balancing robustness and accuracy. To address these challenges, we propose MMT-ARD: a Multimodal Multi-Teacher Adversarial Robust Distillation framework. Our key innovation is a dual-teacher knowledge fusion architecture that collaboratively optimizes clean feature preservation and robust feature enhancement. To better handle challenging adversarial examples, we introduce a dynamic weight allocation strategy based on teacher confidence, enabling adaptive focus on harder samples. Moreover, to mitigate bias among teachers, we design an adaptive sigmoid-based weighting function that balances the strength of knowledge transfer across modalities. Extensive experiments on ImageNet and zero-shot benchmarks demonstrate that MMT-ARD improves robust accuracy by +4.32% and zero-shot accuracy by +3.5% on the ViT-B-32 model, while achieving a 2.3x increase in training efficiency over traditional single-teacher methods. These results highlight the effectiveness and scalability of MMT-ARD in enhancing the adversarial robustness of multimodal large models. Our codes are available at https://github.com/itsnotacie/MMT-ARD.", "date": "2025-11-21", "categories": []}
{"id": "2405.05075", "title": "Sparse-PGD: A Unified Framework for Sparse Adversarial Perturbations Generation", "abstract": "This work studies sparse adversarial perturbations, including both unstructured and structured ones. We propose a framework based on a white-box PGD-like attack method named Sparse-PGD to effectively and efficiently generate such perturbations. Furthermore, we combine Sparse-PGD with a black-box attack to comprehensively and more reliably evaluate the models' robustness against unstructured and structured sparse adversarial perturbations. Moreover, the efficiency of Sparse-PGD enables us to conduct adversarial training to build robust models against various sparse perturbations. Extensive experiments demonstrate that our proposed attack algorithm exhibits strong performance in different scenarios. More importantly, compared with other robust models, our adversarially trained model demonstrates state-of-the-art robustness against various sparse attacks. Codes are available at https://github.com/CityU-MLO/sPGD.", "date": "2024-05-08", "categories": []}
{"id": "2503.12339", "title": "A Closer Look at Adversarial Suffix Learning for Jailbreaking LLMs: Augmented Adversarial Trigger Learning", "abstract": "Gradient optimization-based adversarial attack methods automate the learning of adversarial triggers to generate jailbreak prompts or leak system prompts. In this work, we take a closer look at the optimization objective of adversarial trigger learning and propose ATLA: Adversarial Trigger Learning with Augmented objectives. ATLA improves the negative log-likelihood loss used by previous studies into a weighted loss formulation that encourages the learned adversarial triggers to optimize more towards response format tokens. This enables ATLA to learn an adversarial trigger from just one query-response pair and the learned trigger generalizes well to other similar queries. We further design a variation to augment trigger optimization with an auxiliary loss that suppresses evasive responses. We showcase how to use ATLA to learn adversarial suffixes jailbreaking LLMs and to extract hidden system prompts. Empirically we demonstrate that ATLA consistently outperforms current state-of-the-art techniques, achieving nearly 100% success in attacking while requiring 80% fewer queries. ATLA learned jailbreak suffixes demonstrate high generalization to unseen queries and transfer well to new LLMs. We released our code https://github.com/QData/ALTA_Augmented_Adversarial_Trigger_Learning", "date": "2025-03-16", "categories": []}
{"id": "2509.22060", "title": "Decoding Deception: Understanding Automatic Speech Recognition Vulnerabilities in Evasion and Poisoning Attacks", "abstract": "Recent studies have demonstrated the vulnerability of Automatic Speech Recognition systems to adversarial examples, which can deceive these systems into misinterpreting input speech commands. While previous research has primarily focused on white-box attacks with constrained optimizations, and transferability based black-box attacks against commercial Automatic Speech Recognition devices, this paper explores cost efficient white-box attack and non transferability black-box adversarial attacks on Automatic Speech Recognition systems, drawing insights from approaches such as Fast Gradient Sign Method and Zeroth-Order Optimization. Further, the novelty of the paper includes how poisoning attack can degrade the performances of state-of-the-art models leading to misinterpretation of audio signals. Through experimentation and analysis, we illustrate how hybrid models can generate subtle yet impactful adversarial examples with very little perturbation having Signal Noise Ratio of 35dB that can be generated within a minute. These vulnerabilities of state-of-the-art open source model have practical security implications, and emphasize the need for adversarial security.", "date": "2025-09-26", "categories": []}
{"id": "2511.15807", "title": "TopoReformer: Mitigating Adversarial Attacks Using Topological Purification in OCR Models", "abstract": "Adversarially perturbed images of text can cause sophisticated OCR systems to produce misleading or incorrect transcriptions from seemingly invisible changes to humans. Some of these perturbations even survive physical capture, posing security risks to high-stakes applications such as document processing, license plate recognition, and automated compliance systems. Existing defenses, such as adversarial training, input preprocessing, or post-recognition correction, are often model-specific, computationally expensive, and affect performance on unperturbed inputs while remaining vulnerable to unseen or adaptive attacks. To address these challenges, TopoReformer is introduced, a model-agnostic reformation pipeline that mitigates adversarial perturbations while preserving the structural integrity of text images. Topology studies properties of shapes and spaces that remain unchanged under continuous deformations, focusing on global structures such as connectivity, holes, and loops rather than exact distance. Leveraging these topological features, TopoReformer employs a topological autoencoder to enforce manifold-level consistency in latent space and improve robustness without explicit gradient regularization. The proposed method is benchmarked on EMNIST, MNIST, against standard adversarial attacks (FGSM, PGD, Carlini-Wagner), adaptive attacks (EOT, BDPA), and an OCR-specific watermark attack (FAWA).", "date": "2025-11-19", "categories": []}
{"id": "2511.15435", "title": "HV-Attack: Hierarchical Visual Attack for Multimodal Retrieval Augmented Generation", "abstract": "Advanced multimodal Retrieval-Augmented Generation (MRAG) techniques have been widely applied to enhance the capabilities of Large Multimodal Models (LMMs), but they also bring along novel safety issues. Existing adversarial research has revealed the vulnerability of MRAG systems to knowledge poisoning attacks, which fool the retriever into recalling injected poisoned contents. However, our work considers a different setting: visual attack of MRAG by solely adding imperceptible perturbations at the image inputs of users, without manipulating any other components. This is challenging due to the robustness of fine-tuned retrievers and large-scale generators, and the effect of visual perturbation may be further weakened by propagation through the RAG chain. We propose a novel Hierarchical Visual Attack that misaligns and disrupts the two inputs (the multimodal query and the augmented knowledge) of MRAG's generator to confuse its generation. We further design a hierarchical two-stage strategy to obtain misaligned augmented knowledge. We disrupt the image input of the retriever to make it recall irrelevant knowledge from the original database, by optimizing the perturbation which first breaks the cross-modal alignment and then disrupts the multimodal semantic alignment. We conduct extensive experiments on two widely-used MRAG datasets: OK-VQA and InfoSeek. We use CLIP-based retrievers and two LMMs BLIP-2 and LLaVA as generators. Results demonstrate the effectiveness of our visual attack on MRAG through the significant decrease in both retrieval and generation performance.", "date": "2025-11-19", "categories": []}
{"id": "2511.15571", "title": "Transferable Dual-Domain Feature Importance Attack against AI-Generated Image Detector", "abstract": "Recent AI-generated image (AIGI) detectors achieve impressive accuracy under clean condition. In view of antiforensics, it is significant to develop advanced adversarial attacks for evaluating the security of such detectors, which remains unexplored sufficiently. This letter proposes a Dual-domain Feature Importance Attack (DuFIA) scheme to invalidate AIGI detectors to some extent. Forensically important features are captured by the spatially interpolated gradient and frequency-aware perturbation. The adversarial transferability is enhanced by jointly modeling spatial and frequency-domain feature importances, which are fused to guide the optimization-based adversarial example generation. Extensive experiments across various AIGI detectors verify the cross-model transferability, transparency and robustness of DuFIA.", "date": "2025-11-19", "categories": []}
{"id": "2511.13749", "title": "DeepDefense: Layer-Wise Gradient-Feature Alignment for Building Robust Neural Networks", "abstract": "Deep neural networks are known to be vulnerable to adversarial perturbations, which are small and carefully crafted inputs that lead to incorrect predictions. In this paper, we propose DeepDefense, a novel defense framework that applies Gradient-Feature Alignment (GFA) regularization across multiple layers to suppress adversarial vulnerability. By aligning input gradients with internal feature representations, DeepDefense promotes a smoother loss landscape in tangential directions, thereby reducing the model's sensitivity to adversarial noise.\n  We provide theoretical insights into how adversarial perturbation can be decomposed into radial and tangential components and demonstrate that alignment suppresses loss variation in tangential directions, where most attacks are effective. Empirically, our method achieves significant improvements in robustness across both gradient-based and optimization-based attacks. For example, on CIFAR-10, CNN models trained with DeepDefense outperform standard adversarial training by up to 15.2% under APGD attacks and 24.7% under FGSM attacks. Against optimization-based attacks such as DeepFool and EADEN, DeepDefense requires 20 to 30 times higher perturbation magnitudes to cause misclassification, indicating stronger decision boundaries and a flatter loss landscape. Our approach is architecture-agnostic, simple to implement, and highly effective, offering a promising direction for improving the adversarial robustness of deep learning models.", "date": "2025-11-13", "categories": []}
{"id": "2511.13753", "title": "Robustness of LLM-enabled vehicle trajectory prediction under data security threats", "abstract": "The integration of large language models (LLMs) into automated driving systems has opened new possibilities for reasoning and decision-making by transforming complex driving contexts into language-understandable representations. Recent studies demonstrate that fine-tuned LLMs can accurately predict vehicle trajectories and lane-change intentions by gathering and transforming data from surrounding vehicles. However, the robustness of such LLM-based prediction models for safety-critical driving systems remains unexplored, despite the increasing concerns about the trustworthiness of LLMs. This study addresses this gap by conducting a systematic vulnerability analysis of LLM-enabled vehicle trajectory prediction. We propose a one-feature differential evolution attack that perturbs a single kinematic feature of surrounding vehicles within the LLM's input prompts under a black-box setting. Experiments on the highD dataset reveal that even minor, physically plausible perturbations can significantly disrupt model outputs, underscoring the susceptibility of LLM-based predictors to adversarial manipulation. Further analyses reveal a trade-off between accuracy and robustness, examine the failure mechanism, and explore potential mitigation solutions. The findings provide the very first insights into adversarial vulnerabilities of LLM-driven automated vehicle models in the context of vehicular interactions and highlight the need for robustness-oriented design in future LLM-based intelligent transportation systems.", "date": "2025-11-14", "categories": []}
{"id": "2511.13771", "title": "ExplainableGuard: Interpretable Adversarial Defense for Large Language Models Using Chain-of-Thought Reasoning", "abstract": "Large Language Models (LLMs) are increasingly vulnerable to adversarial attacks that can subtly manipulate their outputs. While various defense mechanisms have been proposed, many operate as black boxes, lacking transparency in their decision-making. This paper introduces ExplainableGuard, an interpretable adversarial defense framework leveraging the chain-of-thought (CoT) reasoning capabilities of DeepSeek-Reasoner. Our approach not only detects and neutralizes adversarial perturbations in text but also provides step-by-step explanations for each defense action. We demonstrate how tailored CoT prompts guide the LLM to perform a multi-faceted analysis (character, word, structural, and semantic) and generate a purified output along with a human-readable justification. Preliminary results on the GLUE Benchmark and IMDB Movie Reviews dataset show promising defense efficacy. Additionally, a human evaluation study reveals that ExplainableGuard's explanations outperform ablated variants in clarity, specificity, and actionability, with a 72.5% deployability-trust rating, underscoring its potential for more trustworthy LLM deployments.", "date": "2025-11-15", "categories": []}
{"id": "2511.14003", "title": "Certified but Fooled! Breaking Certified Defences with Ghost Certificates", "abstract": "Certified defenses promise provable robustness guarantees. We study the malicious exploitation of probabilistic certification frameworks to better understand the limits of guarantee provisions. Now, the objective is to not only mislead a classifier, but also manipulate the certification process to generate a robustness guarantee for an adversarial input certificate spoofing. A recent study in ICLR demonstrated that crafting large perturbations can shift inputs far into regions capable of generating a certificate for an incorrect class. Our study investigates if perturbations needed to cause a misclassification and yet coax a certified model into issuing a deceptive, large robustness radius for a target class can still be made small and imperceptible. We explore the idea of region-focused adversarial examples to craft imperceptible perturbations, spoof certificates and achieve certification radii larger than the source class ghost certificates. Extensive evaluations with the ImageNet demonstrate the ability to effectively bypass state-of-the-art certified defenses such as Densepure. Our work underscores the need to better understand the limits of robustness certification methods.", "date": "2025-11-18", "categories": []}
{"id": "2511.14074", "title": "Dynamic Black-box Backdoor Attacks on IoT Sensory Data", "abstract": "Sensor data-based recognition systems are widely used in various applications, such as gait-based authentication and human activity recognition (HAR). Modern wearable and smart devices feature various built-in Inertial Measurement Unit (IMU) sensors, and such sensor-based measurements can be fed to a machine learning-based model to train and classify human activities. While deep learning-based models have proven successful in classifying human activity and gestures, they pose various security risks. In our paper, we discuss a novel dynamic trigger-generation technique for performing black-box adversarial attacks on sensor data-based IoT systems. Our empirical analysis shows that the attack is successful on various datasets and classifier models with minimal perturbation on the input data. We also provide a detailed comparative analysis of performance and stealthiness to various other poisoning techniques found in backdoor attacks. We also discuss some adversarial defense mechanisms and their impact on the effectiveness of our trigger-generation technique.", "date": "2025-11-18", "categories": []}
{"id": "2409.04407", "title": "Exploiting Missing Data Remediation Strategies using Adversarial Missingness Attacks", "abstract": "Adversarial Missingness (AM) attacks aim to manipulate model fitting by carefully engineering a missing data problem to achieve a specific malicious objective. AM attacks are significantly different from prior data poisoning attacks in that no malicious data inserted and no data is maliciously perturbed. Current AM attacks are feasible only under the assumption that the modeler (victim) uses full-information maximum likelihood methods to handle missingness. This work aims to remedy this limitation of AM attacks; in the approach taken here, the adversary achieves their goal by solving a bi-level optimization problem to engineer the adversarial missingness mechanism, where the lower level problem incorporates a differentiable approximation of the targeted missingness remediation technique. As instantiations of this framework, AM attacks are provided for three popular techniques: (i) complete case analysis, (ii) mean imputation, and (iii) regression-based imputation for general empirical risk minimization (ERM) problems. Experiments on real-world data show that AM attacks are successful with modest levels of missingness (less than 20%). Furthermore, we show on the real-world Twins dataset that AM attacks can manipulate the estimated average treatment effect (ATE) as an instance of the general ERM problems: the adversary succeeds in not only reversing the sign, but also in substantially inflating the ATE values from a true value of -1.61% to a manipulated one as high as 10%. These experimental results hold when the ATE is calculated using multiple regression-based estimators with different architectures, even when the adversary is restricted to modifying only a subset of the training data.", "date": "2024-09-06", "categories": []}
{"id": "2409.10997", "title": "Contextual Breach: Assessing the Robustness of Transformer-based QA Models", "abstract": "Contextual question-answering models are susceptible to adversarial perturbations to input context, commonly observed in real-world scenarios. These adversarial noises are designed to degrade the performance of the model by distorting the textual input. We introduce a unique dataset that incorporates seven distinct types of adversarial noise into the context, each applied at five different intensity levels on the SQuAD dataset. To quantify the robustness, we utilize robustness metrics providing a standardized measure for assessing model performance across varying noise types and levels. Experiments on transformer-based question-answering models reveal robustness vulnerabilities and important insights into the model's performance in realistic textual input.", "date": "2024-09-17", "categories": []}
{"id": "2412.08098", "title": "What You See Is Not Always What You Get: Evaluating GPT's Comprehension of Source Code", "abstract": "Recent studies have demonstrated outstanding capabilities of large language models (LLMs) in software engineering tasks, including code generation and comprehension. While LLMs have shown significant potential in assisting with coding, LLMs are vulnerable to adversarial attacks. In this paper, we investigate the vulnerability of LLMs to imperceptible attacks. This class of attacks manipulate source code at the character level, which renders the changes invisible to human reviewers yet effective in misleading LLMs' behaviour. We devise these attacks into four distinct categories and analyse their impacts on code analysis and comprehension tasks. These four types of imperceptible character attacks include coding reordering, invisible coding characters, code deletions, and code homoglyphs. To assess the robustness of state-of-the-art LLMs, we present a systematic evaluation across multiple models using both perturbed and clean code snippets. Two evaluation metrics, model confidence using log probabilities of response and response correctness, are introduced. The results reveal that LLMs are susceptible to imperceptible coding perturbations, with varying degrees of degradation highlighted across different LLMs. Furthermore, we observe a consistent negative correlation between perturbation magnitude and model performance. These results highlight the urgent need for robust LLMs capable of manoeuvring behaviours under imperceptible adversarial conditions.", "date": "2024-12-11", "categories": []}
{"id": "2412.12478", "title": "Human-in-the-Loop Generation of Adversarial Texts: A Case Study on Tibetan Script", "abstract": "DNN-based language models excel across various NLP tasks but remain highly vulnerable to textual adversarial attacks. While adversarial text generation is crucial for NLP security, explainability, evaluation, and data augmentation, related work remains overwhelmingly English-centric, leaving the problem of constructing high-quality and sustainable adversarial robustness benchmarks for lower-resourced languages both difficult and understudied. First, method customization for lower-resourced languages is complicated due to linguistic differences and limited resources. Second, automated attacks are prone to generating invalid or ambiguous adversarial texts. Last but not least, language models continuously evolve and may be immune to parts of previously generated adversarial texts. To address these challenges, we introduce HITL-GAT, an interactive system based on a general approach to human-in-the-loop generation of adversarial texts. Additionally, we demonstrate the utility of HITL-GAT through a case study on Tibetan script, employing three customized adversarial text generation methods and establishing its first adversarial robustness benchmark, providing a valuable reference for other lower-resourced languages.", "date": "2024-12-17", "categories": []}
{"id": "2501.07251", "title": "MOS-Attack: A Scalable Multi-objective Adversarial Attack Framework", "abstract": "Crafting adversarial examples is crucial for evaluating and enhancing the robustness of Deep Neural Networks (DNNs), presenting a challenge equivalent to maximizing a non-differentiable 0-1 loss function.\n  However, existing single objective methods, namely adversarial attacks focus on a surrogate loss function, do not fully harness the benefits of engaging multiple loss functions, as a result of insufficient understanding of their synergistic and conflicting nature.\n  To overcome these limitations, we propose the Multi-Objective Set-based Attack (MOS Attack), a novel adversarial attack framework leveraging multiple loss functions and automatically uncovering their interrelations.\n  The MOS Attack adopts a set-based multi-objective optimization strategy, enabling the incorporation of numerous loss functions without additional parameters.\n  It also automatically mines synergistic patterns among various losses, facilitating the generation of potent adversarial attacks with fewer objectives.\n  Extensive experiments have shown that our MOS Attack outperforms single-objective attacks. Furthermore, by harnessing the identified synergistic patterns, MOS Attack continues to show superior results with a reduced number of loss functions. Our code is available at https://github.com/pgg3/MOS-Attack.", "date": "2025-01-13", "categories": []}
{"id": "2503.09095", "title": "Backdooring CLIP through Concept Confusion", "abstract": "Backdoor attacks pose a serious threat to deep learning models by allowing adversaries to implant hidden behaviors that remain dormant on clean inputs but are maliciously triggered at inference. Existing backdoor attack methods typically rely on explicit triggers such as image patches or pixel perturbations, which makes them easier to detect and limits their applicability in complex settings. To address this limitation, we take a different perspective by analyzing backdoor attacks through the lens of concept-level reasoning, drawing on insights from interpretable AI. We show that traditional attacks can be viewed as implicitly manipulating the concepts activated within a model's latent space. This motivates a natural question: can backdoors be built by directly manipulating concepts? To answer this, we propose the Concept Confusion Attack (CCA), a novel framework that designates human-understandable concepts as internal triggers, eliminating the need for explicit input modifications. By relabeling images that strongly exhibit a chosen concept and fine-tuning on this mixed dataset, CCA teaches the model to associate the concept itself with the attacker's target label. Consequently, the presence of the concept alone is sufficient to activate the backdoor, making the attack stealthier and more resistant to existing defenses. Using CLIP as a case study, we show that CCA achieves high attack success rates while preserving clean-task accuracy and evading state-of-the-art defenses.", "date": "2025-03-12", "categories": []}
{"id": "2504.10804", "title": "Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability", "abstract": "Vision Transformers (ViTs) have demonstrated impressive performance across a range of applications, including many safety-critical tasks. However, their unique architectural properties raise new challenges and opportunities in adversarial robustness. In particular, we observe that adversarial examples crafted on ViTs exhibit higher transferability compared to those crafted on CNNs, suggesting that ViTs contain structural characteristics favorable for transferable attacks. In this work, we investigate the role of computational redundancy in ViTs and its impact on adversarial transferability. Unlike prior studies that aim to reduce computation for efficiency, we propose to exploit this redundancy to improve the quality and transferability of adversarial examples. Through a detailed analysis, we identify two forms of redundancy, including the data-level and model-level, that can be harnessed to amplify attack effectiveness. Building on this insight, we design a suite of techniques, including attention sparsity manipulation, attention head permutation, clean token regularization, ghost MoE diversification, and test-time adversarial training. Extensive experiments on the ImageNet-1k dataset validate the effectiveness of our approach, showing that our methods significantly outperform existing baselines in both transferability and generality across diverse model architectures.", "date": "2025-04-15", "categories": []}
{"id": "2505.12644", "title": "Use as Many Surrogates as You Want: Selective Ensemble Attack to Unleash Transferability without Sacrificing Resource Efficiency", "abstract": "In surrogate ensemble attacks, using more surrogate models yields higher transferability but lower resource efficiency. This practical trade-off between transferability and efficiency has largely limited existing attacks despite many pre-trained models are easily accessible online. In this paper, we argue that such a trade-off is caused by an unnecessary common assumption, i.e., all models should be \\textit{identical} across iterations. By lifting this assumption, we can use as many surrogates as we want to unleash transferability without sacrificing efficiency. Concretely, we propose Selective Ensemble Attack (SEA), which dynamically selects diverse models (from easily accessible pre-trained models) across iterations based on our new interpretation of decoupling within-iteration and cross-iteration model diversity. In this way, the number of within-iteration models is fixed for maintaining efficiency, while only cross-iteration model diversity is increased for higher transferability. Experiments on ImageNet demonstrate the superiority of SEA in various scenarios. For example, when dynamically selecting 4 from 20 accessible models, SEA yields 8.5% higher transferability than existing attacks under the same efficiency. The superiority of SEA also generalizes to real-world systems, such as commercial vision APIs and large vision-language models. Overall, SEA opens up the possibility of adaptively balancing transferability and efficiency according to specific resource requirements.", "date": "2025-05-19", "categories": []}
{"id": "2505.20782", "title": "Towards Cross-Domain Multi-Targeted Adversarial Attacks", "abstract": "Multi-targeted adversarial attacks aim to mislead classifiers toward specific target classes using a single perturbation generator with a conditional input specifying the desired target class. Existing methods face two key limitations: (1) a single generator supports only a limited number of predefined target classes, and (2) it requires access to the victim model's training data to learn target class semantics. This dependency raises data leakage concerns in practical black-box scenarios where the training data is typically private. To address these limitations, we propose a novel Cross-Domain Multi-Targeted Attack (CD-MTA) that can generate perturbations toward arbitrary target classes, even those that do not exist in the attacker's training data. CD-MTA is trained on a single public dataset but can perform targeted attacks on black-box models trained on different datasets with disjoint and unknown class sets. Our method requires only a single example image that visually represents the desired target class, without relying its label, class distribution or pretrained embeddings. We achieve this through a Feature Injection Module (FIM) and class-agnostic objectives which guide the generator to extract transferable, fine-grained features from the target image without inferring class semantics. Experiments on ImageNet and seven additional datasets show that CD-MTA outperforms existing multi-targeted attack methods on unseen target classes in black-box and cross-domain scenarios. The code is available at https://github.com/tgoncalv/CD-MTA.", "date": "2025-05-27", "categories": []}
{"id": "2506.07666", "title": "ProARD: progressive adversarial robustness distillation: provide wide range of robust students", "abstract": "Adversarial Robustness Distillation (ARD) has emerged as an effective method to enhance the robustness of lightweight deep neural networks against adversarial attacks. Current ARD approaches have leveraged a large robust teacher network to train one robust lightweight student. However, due to the diverse range of edge devices and resource constraints, current approaches require training a new student network from scratch to meet specific constraints, leading to substantial computational costs and increased CO2 emissions. This paper proposes Progressive Adversarial Robustness Distillation (ProARD), enabling the efficient one-time training of a dynamic network that supports a diverse range of accurate and robust student networks without requiring retraining. We first make a dynamic deep neural network based on dynamic layers by encompassing variations in width, depth, and expansion in each design stage to support a wide range of architectures. Then, we consider the student network with the largest size as the dynamic teacher network. ProARD trains this dynamic network using a weight-sharing mechanism to jointly optimize the dynamic teacher network and its internal student networks. However, due to the high computational cost of calculating exact gradients for all the students within the dynamic network, a sampling mechanism is required to select a subset of students. We show that random student sampling in each iteration fails to produce accurate and robust students.", "date": "2025-06-09", "categories": []}
{"id": "2510.22963", "title": "CompressionAttack: Exploiting Prompt Compression as a New Attack Surface in LLM-Powered Agents", "abstract": "LLM-powered agents often use prompt compression to reduce inference costs, but this introduces a new security risk. Compression modules, which are optimized for efficiency rather than safety, can be manipulated by adversarial inputs, causing semantic drift and altering LLM behavior. This work identifies prompt compression as a novel attack surface and presents CompressionAttack, the first framework to exploit it. CompressionAttack includes two strategies: HardCom, which uses discrete adversarial edits for hard compression, and SoftCom, which performs latent-space perturbations for soft compression. Experiments on multiple LLMs show up to an average ASR of 83% and 87% in two tasks, while remaining highly stealthy and transferable. Case studies in three practical scenarios confirm real-world impact, and current defenses prove ineffective, highlighting the need for stronger protections.", "date": "2025-10-27", "categories": []}
{"id": "2511.01317", "title": "A Generative Adversarial Approach to Adversarial Attacks Guided by Contrastive Language-Image Pre-trained Model", "abstract": "The rapid growth of deep learning has brought about powerful models that can handle various tasks, like identifying images and understanding language. However, adversarial attacks, an unnoticed alteration, can deceive models, leading to inaccurate predictions. In this paper, a generative adversarial attack method is proposed that uses the CLIP model to create highly effective and visually imperceptible adversarial perturbations. The CLIP model's ability to align text and image representation helps incorporate natural language semantics with a guided loss to generate effective adversarial examples that look identical to the original inputs. This integration allows extensive scene manipulation, creating perturbations in multi-object environments specifically designed to deceive multilabel classifiers. Our approach integrates the concentrated perturbation strategy from Saliency-based Auto-Encoder (SSAE) with the dissimilar text embeddings similar to Generative Adversarial Multi-Object Scene Attacks (GAMA), resulting in perturbations that both deceive classification models and maintain high structural similarity to the original images. The model was tested on various tasks across diverse black-box victim models. The experimental results show that our method performs competitively, achieving comparable or superior results to existing techniques, while preserving greater visual fidelity.", "date": "2025-11-03", "categories": []}
{"id": "2511.11834", "title": "Volatility in Certainty (VC): A Metric for Detecting Adversarial Perturbations During Inference in Neural Network Classifiers", "abstract": "Adversarial robustness remains a critical challenge in deploying neural network classifiers, particularly in real-time systems where ground-truth labels are unavailable during inference. This paper investigates \\textit{Volatility in Certainty} (VC), a recently proposed, label-free metric that quantifies irregularities in model confidence by measuring the dispersion of sorted softmax outputs. Specifically, VC is defined as the average squared log-ratio of adjacent certainty values, capturing local fluctuations in model output smoothness. We evaluate VC as a proxy for classification accuracy and as an indicator of adversarial drift. Experiments are conducted on artificial neural networks (ANNs) and convolutional neural networks (CNNs) trained on MNIST, as well as a regularized VGG-like model trained on CIFAR-10. Adversarial examples are generated using the Fast Gradient Sign Method (FGSM) across varying perturbation magnitudes. In addition, mixed test sets are created by gradually introducing adversarial contamination to assess VC's sensitivity under incremental distribution shifts. Our results reveal a strong negative correlation between classification accuracy and log(VC) (correlation rho < -0.90 in most cases), suggesting that VC effectively reflects performance degradation without requiring labeled data. These findings position VC as a scalable, architecture-agnostic, and real-time performance metric suitable for early-warning systems in safety-critical applications.", "date": "2025-11-14", "categories": []}
{"id": "2511.11842", "title": "On the Trade-Off Between Transparency and Security in Adversarial Machine Learning", "abstract": "Transparency and security are both central to Responsible AI, but they may conflict in adversarial settings. We investigate the strategic effect of transparency for agents through the lens of transferable adversarial example attacks. In transferable adversarial example attacks, attackers maliciously perturb their inputs using surrogate models to fool a defender's target model. These models can be defended or undefended, with both players having to decide which to use. Using a large-scale empirical evaluation of nine attacks across 181 models, we find that attackers are more successful when they match the defender's decision; hence, obscurity could be beneficial to the defender. With game theory, we analyze this trade-off between transparency and security by modeling this problem as both a Nash game and a Stackelberg game, and comparing the expected outcomes. Our analysis confirms that only knowing whether a defender's model is defended or not can sometimes be enough to damage its security. This result serves as an indicator of the general trade-off between transparency and security, suggesting that transparency in AI systems can be at odds with security. Beyond adversarial machine learning, our work illustrates how game-theoretic reasoning can uncover conflicts between transparency and security.", "date": "2025-11-14", "categories": []}
{"id": "2511.11993", "title": "Dynamic Parameter Optimization for Highly Transferable Transformation-Based Attacks", "abstract": "Despite their wide application, the vulnerabilities of deep neural networks raise societal concerns. Among them, transformation-based attacks have demonstrated notable success in transfer attacks. However, existing attacks suffer from blind spots in parameter optimization, limiting their full potential. Specifically, (1) prior work generally considers low-iteration settings, yet attacks perform quite differently at higher iterations, so characterizing overall performance based only on low-iteration results is misleading. (2) Existing attacks use uniform parameters for different surrogate models, iterations, and tasks, which greatly impairs transferability. (3) Traditional transformation parameter optimization relies on grid search. For n parameters with m steps each, the complexity is O(mn). Large computational overhead limits further optimization of parameters. To address these limitations, we conduct an empirical study with various transformations as baselines, revealing three dynamic patterns of transferability with respect to parameter strength. We further propose a novel Concentric Decay Model (CDM) to effectively explain these patterns. Building on these insights, we propose an efficient Dynamic Parameter Optimization (DPO) based on the rise-then-fall pattern, reducing the complexity to O(nlogm). Comprehensive experiments on existing transformation-based attacks across different surrogate models, iterations, and tasks demonstrate that our DPO can significantly improve transferability.", "date": "2025-11-15", "categories": []}
{"id": "2511.12149", "title": "AttackVLA: Benchmarking Adversarial and Backdoor Attacks on Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models enable robots to interpret natural-language instructions and perform diverse tasks, yet their integration of perception, language, and control introduces new safety vulnerabilities. Despite growing interest in attacking such models, the effectiveness of existing techniques remains unclear due to the absence of a unified evaluation framework. One major issue is that differences in action tokenizers across VLA architectures hinder reproducibility and fair comparison. More importantly, most existing attacks have not been validated in real-world scenarios. To address these challenges, we propose AttackVLA, a unified framework that aligns with the VLA development lifecycle, covering data construction, model training, and inference. Within this framework, we implement a broad suite of attacks, including all existing attacks targeting VLAs and multiple adapted attacks originally developed for vision-language models, and evaluate them in both simulation and real-world settings. Our analysis of existing attacks reveals a critical gap: current methods tend to induce untargeted failures or static action states, leaving targeted attacks that drive VLAs to perform precise long-horizon action sequences largely unexplored. To fill this gap, we introduce BackdoorVLA, a targeted backdoor attack that compels a VLA to execute an attacker-specified long-horizon action sequence whenever a trigger is present. We evaluate BackdoorVLA in both simulated benchmarks and real-world robotic settings, achieving an average targeted success rate of 58.4% and reaching 100% on selected tasks. Our work provides a standardized framework for evaluating VLA vulnerabilities and demonstrates the potential for precise adversarial manipulation, motivating further research on securing VLA-based embodied systems.", "date": "2025-11-15", "categories": []}
{"id": "2511.12265", "title": "Calibrated Adversarial Sampling: Multi-Armed Bandit-Guided Generalization Against Unforeseen Attacks", "abstract": "Deep Neural Networks (DNNs) are known to be vulnerable to various adversarial perturbations. To address the safety concerns arising from these vulnerabilities, adversarial training (AT) has emerged as one of the most effective paradigms for enhancing the robustness of DNNs. However, existing AT frameworks primarily focus on a single or a limited set of attack types, leaving DNNs still exposed to attack types that may be encountered in practice but not addressed during training. In this paper, we propose an efficient fine-tuning method called Calibrated Adversarial Sampling (CAS) to address these issues. From the optimization perspective within the multi-armed bandit framework, it dynamically designs rewards and balances exploration and exploitation by considering the dynamic and interdependent characteristics of multiple robustness dimensions. Experiments on benchmark datasets show that CAS achieves superior overall robustness while maintaining high clean accuracy, providing a new paradigm for robust generalization of DNNs.", "date": "2025-11-15", "categories": []}
{"id": "2511.12423", "title": "GRAPHTEXTACK: A Realistic Black-Box Node Injection Attack on LLM-Enhanced GNNs", "abstract": "Text-attributed graphs (TAGs), which combine structural and textual node information, are ubiquitous across many domains. Recent work integrates Large Language Models (LLMs) with Graph Neural Networks (GNNs) to jointly model semantics and structure, resulting in more general and expressive models that achieve state-of-the-art performance on TAG benchmarks. However, this integration introduces dual vulnerabilities: GNNs are sensitive to structural perturbations, while LLM-derived features are vulnerable to prompt injection and adversarial phrasing. While existing adversarial attacks largely perturb structure or text independently, we find that uni-modal attacks cause only modest degradation in LLM-enhanced GNNs. Moreover, many existing attacks assume unrealistic capabilities, such as white-box access or direct modification of graph data. To address these gaps, we propose GRAPHTEXTACK, the first black-box, multi-modal{, poisoning} node injection attack for LLM-enhanced GNNs. GRAPHTEXTACK injects nodes with carefully crafted structure and semantics to degrade model performance, operating under a realistic threat model without relying on model internals or surrogate models. To navigate the combinatorial, non-differentiable search space of connectivity and feature assignments, GRAPHTEXTACK introduces a novel evolutionary optimization framework with a multi-objective fitness function that balances local prediction disruption and global graph influence. Extensive experiments on five datasets and two state-of-the-art LLM-enhanced GNN models show that GRAPHTEXTACK significantly outperforms 12 strong baselines.", "date": "2025-11-16", "categories": []}
{"id": "2511.12575", "title": "Beyond Pixels: Semantic-aware Typographic Attack for Geo-Privacy Protection", "abstract": "Large Visual Language Models (LVLMs) now pose a serious yet overlooked privacy threat, as they can infer a social media user's geolocation directly from shared images, leading to unintended privacy leakage. While adversarial image perturbations provide a potential direction for geo-privacy protection, they require relatively strong distortions to be effective against LVLMs, which noticeably degrade visual quality and diminish an image's value for sharing. To overcome this limitation, we identify typographical attacks as a promising direction for protecting geo-privacy by adding text extension outside the visual content. We further investigate which textual semantics are effective in disrupting geolocation inference and design a two-stage, semantics-aware typographical attack that generates deceptive text to protect user privacy. Extensive experiments across three datasets demonstrate that our approach significantly reduces geolocation prediction accuracy of five state-of-the-art commercial LVLMs, establishing a practical and visually-preserving protection strategy against emerging geo-privacy threats.", "date": "2025-11-16", "categories": []}
{"id": "2511.12722", "title": "On Robustness of Linear Classifiers to Targeted Data Poisoning", "abstract": "Data poisoning is a training-time attack that undermines the trustworthiness of learned models. In a targeted data poisoning attack, an adversary manipulates the training dataset to alter the classification of a targeted test point. Given the typically large size of training dataset, manual detection of poisoning is difficult. An alternative is to automatically measure a dataset's robustness against such an attack, which is the focus of this paper. We consider a threat model wherein an adversary can only perturb the labels of the training dataset, with knowledge limited to the hypothesis space of the victim's model. In this setting, we prove that finding the robustness is an NP-Complete problem, even when hypotheses are linear classifiers. To overcome this, we present a technique that finds lower and upper bounds of robustness. Our implementation of the technique computes these bounds efficiently in practice for many publicly available datasets. We experimentally demonstrate the effectiveness of our approach. Specifically, a poisoning exceeding the identified robustness bounds significantly impacts test point classification. We are also able to compute these bounds in many more cases where state-of-the-art techniques fail.", "date": "2025-11-16", "categories": []}
{"id": "2511.12827", "title": "Efficient Adversarial Malware Defense via Trust-Based Raw Override and Confidence-Adaptive Bit-Depth Reduction", "abstract": "The deployment of robust malware detection systems in big data environments requires careful consideration of both security effectiveness and computational efficiency. While recent advances in adversarial defenses have demonstrated strong robustness improvements, they often introduce computational overhead ranging from 4x to 22x, which presents significant challenges for production systems processing millions of samples daily. In this work, we propose a novel framework that combines Trust-Raw Override (TRO) with Confidence-Adaptive Bit-Depth Reduction (CABDR) to explicitly optimize the trade-off between adversarial robustness and computational efficiency. Our approach leverages adaptive confidence-based mechanisms to selectively apply defensive measures, achieving 1.76x computational overhead - a 2.3x improvement over state-of-the-art smoothing defenses. Through comprehensive evaluation on the EMBER v2 dataset comprising 800K samples, we demonstrate that our framework maintains 91 percent clean accuracy while reducing attack success rates to 31-37 percent across multiple attack types, with particularly strong performance against optimization-based attacks such as C and W (48.8 percent reduction). The framework achieves throughput of up to 1.26 million samples per second (measured on pre-extracted EMBER features with no runtime feature extraction), validated across 72 production configurations with statistical significance (5 independent runs, 95 percent confidence intervals, p less than 0.01). Our results suggest that practical adversarial robustness in production environments requires explicit optimization of the efficiency-robustness trade-off, providing a viable path for organizations to deploy robust defenses without prohibitive infrastructure costs.", "date": "2025-11-16", "categories": []}
{"id": "2511.12956", "title": "T2I-Based Physical-World Appearance Attack against Traffic Sign Recognition Systems in Autonomous Driving", "abstract": "Traffic Sign Recognition (TSR) systems play a critical role in Autonomous Driving (AD) systems, enabling real-time detection of road signs, such as STOP and speed limit signs. While these systems are increasingly integrated into commercial vehicles, recent research has exposed their vulnerability to physical-world adversarial appearance attacks. In such attacks, carefully crafted visual patterns are misinterpreted by TSR models as legitimate traffic signs, while remaining inconspicuous or benign to human observers. However, existing adversarial appearance attacks suffer from notable limitations. Pixel-level perturbation-based methods often lack stealthiness and tend to overfit to specific surrogate models, resulting in poor transferability to real-world TSR systems. On the other hand, text-to-image (T2I) diffusion model-based approaches demonstrate limited effectiveness and poor generalization to out-of-distribution sign types.\n  In this paper, we present DiffSign, a novel T2I-based appearance attack framework designed to generate physically robust, highly effective, transferable, practical, and stealthy appearance attacks against TSR systems. To overcome the limitations of prior approaches, we propose a carefully designed attack pipeline that integrates CLIP-based loss and masked prompts to improve attack focus and controllability. We also propose two novel style customization methods to guide visual appearance and improve out-of-domain traffic sign attack generalization and attack stealthiness. We conduct extensive evaluations of DiffSign under varied real-world conditions, including different distances, angles, light conditions, and sign categories. Our method achieves an average physical-world attack success rate of 83.3%, leveraging DiffSign's high effectiveness in attack transferability.", "date": "2025-11-17", "categories": []}
{"id": "2511.13132", "title": "Shedding Light on VLN Robustness: A Black-box Framework for Indoor Lighting-based Adversarial Attack", "abstract": "Vision-and-Language Navigation (VLN) agents have made remarkable progress, but their robustness remains insufficiently studied. Existing adversarial evaluations often rely on perturbations that manifest as unusual textures rarely encountered in everyday indoor environments. Errors under such contrived conditions have limited practical relevance, as real-world agents are unlikely to encounter such artificial patterns. In this work, we focus on indoor lighting, an intrinsic yet largely overlooked scene attribute that strongly influences navigation. We propose Indoor Lighting-based Adversarial Attack (ILA), a black-box framework that manipulates global illumination to disrupt VLN agents. Motivated by typical household lighting usage, we design two attack modes: Static Indoor Lighting-based Attack (SILA), where the lighting intensity remains constant throughout an episode, and Dynamic Indoor Lighting-based Attack (DILA), where lights are switched on or off at critical moments to induce abrupt illumination changes. We evaluate ILA on two state-of-the-art VLN models across three navigation tasks. Results show that ILA significantly increases failure rates while reducing trajectory efficiency, revealing previously unrecognized vulnerabilities of VLN agents to realistic indoor lighting variations.", "date": "2025-11-17", "categories": []}
{"id": "2511.13356", "title": "Enhancing All-to-X Backdoor Attacks with Optimized Target Class Mapping", "abstract": "Backdoor attacks pose severe threats to machine learning systems, prompting extensive research in this area. However, most existing work focuses on single-target All-to-One (A2O) attacks, overlooking the more complex All-to-X (A2X) attacks with multiple target classes, which are often assumed to have low attack success rates. In this paper, we first demonstrate that A2X attacks are robust against state-of-the-art defenses. We then propose a novel attack strategy that enhances the success rate of A2X attacks while maintaining robustness by optimizing grouping and target class assignment mechanisms. Our method improves the attack success rate by up to 28%, with average improvements of 6.7%, 16.4%, 14.1% on CIFAR10, CIFAR100, and Tiny-ImageNet, respectively. We anticipate that this study will raise awareness of A2X attacks and stimulate further research in this under-explored area. Our code is available at https://github.com/kazefjj/A2X-backdoor .", "date": "2025-11-17", "categories": []}
{"id": "2511.13545", "title": "Robust Defense Strategies for Multimodal Contrastive Learning: Efficient Fine-tuning Against Backdoor Attacks", "abstract": "The advent of multimodal deep learning models, such as CLIP, has unlocked new frontiers in a wide range of applications, from image-text understanding to classification tasks. However, these models are not safe for adversarial attacks, particularly backdoor attacks, which can subtly manipulate model behavior. Moreover, existing defense methods typically involve training from scratch or fine-tuning using a large dataset without pinpointing the specific labels that are affected. In this study, we introduce an innovative strategy to enhance the robustness of multimodal contrastive learning models against such attacks. In particular, given a poisoned CLIP model, our approach can identify the backdoor trigger and pinpoint the victim samples and labels in an efficient manner. To that end, an image segmentation ``oracle'' is introduced as the supervisor for the output of the poisoned CLIP. We develop two algorithms to rectify the poisoned model: (1) differentiating between CLIP and Oracle's knowledge to identify potential triggers; (2) pinpointing affected labels and victim samples, and curating a compact fine-tuning dataset. With this knowledge, we are allowed to rectify the poisoned CLIP model to negate backdoor effects. Extensive experiments on visual recognition benchmarks demonstrate our strategy is effective in CLIP-based backdoor defense.", "date": "2025-11-17", "categories": []}
{"id": "2207.03400", "title": "On the Relationship Between Adversarial Robustness and Decision Region in Deep Neural Networks", "abstract": "In general, Deep Neural Networks (DNNs) are evaluated by the generalization performance measured on unseen data excluded from the training phase. Along with the development of DNNs, the generalization performance converges to the state-of-the-art and it becomes difficult to evaluate DNNs solely based on this metric. The robustness against adversarial attack has been used as an additional metric to evaluate DNNs by measuring their vulnerability. However, few studies have been performed to analyze the adversarial robustness in terms of the geometry in DNNs. In this work, we perform an empirical study to analyze the internal properties of DNNs that affect model robustness under adversarial attacks. In particular, we propose the novel concept of the Populated Region Set (PRS), where training samples are populated more frequently, to represent the internal properties of DNNs in a practical setting. From systematic experiments with the proposed concept, we provide empirical evidence to validate that a low PRS ratio has a strong relationship with the adversarial robustness of DNNs. We also devise PRS regularizer leveraging the characteristics of PRS to improve the adversarial robustness without adversarial training.", "date": "2022-07-07", "categories": []}
{"id": "2504.14137", "title": "Rethinking Target Label Conditioning in Adversarial Attacks: A 2D Tensor-Guided Generative Approach", "abstract": "Compared to single-target adversarial attacks, multi-target attacks have garnered significant attention due to their ability to generate adversarial images for multiple target classes simultaneously. However, existing generative approaches for multi-target attacks primarily encode target labels into one-dimensional tensors, leading to a loss of fine-grained visual information and overfitting to model-specific features during noise generation. To address this gap, we first identify and validate that the semantic feature quality and quantity are critical factors affecting the transferability of targeted attacks: 1) Feature quality refers to the structural and detailed completeness of the implanted target features, as deficiencies may result in the loss of key discriminative information; 2) Feature quantity refers to the spatial sufficiency of the implanted target features, as inadequacy limits the victim model's attention to this feature. Based on these findings, we propose the 2D Tensor-Guided Adversarial Fusion (TGAF) framework, which leverages the powerful generative capabilities of diffusion models to encode target labels into two-dimensional semantic tensors for guiding adversarial noise generation. Additionally, we design a novel masking strategy tailored for the training process, ensuring that parts of the generated noise retain complete semantic information about the target class. Extensive experiments demonstrate that TGAF consistently surpasses state-of-the-art methods across various settings.", "date": "2025-04-19", "categories": []}
{"id": "2511.11309", "title": "destroR: Attacking Transfer Models with Obfuscous Examples to Discard Perplexity", "abstract": "Advancements in Machine Learning & Neural Networks in recent years have led to widespread implementations of Natural Language Processing across a variety of fields with remarkable success, solving a wide range of complicated problems. However, recent research has shown that machine learning models may be vulnerable in a number of ways, putting both the models and the systems theyre used in at risk. In this paper, we intend to analyze and experiment with the best of existing adversarial attack recipes and create new ones. We concentrated on developing a novel adversarial attack strategy on current state-of-the-art machine learning models by producing ambiguous inputs for the models to confound them and then constructing the path to the future development of the robustness of the models. We will develop adversarial instances with maximum perplexity, utilizing machine learning and deep learning approaches in order to trick the models. In our attack recipe, we will analyze several datasets and focus on creating obfuscous adversary examples to put the models in a state of perplexity, and by including the Bangla Language in the field of adversarial attacks. We strictly uphold utility usage reduction and efficiency throughout our work.", "date": "2025-11-13", "categories": []}
{"id": "2208.08083", "title": "Two Heads are Better than One: Robust Learning Meets Multi-branch Models", "abstract": "Deep neural networks (DNNs) are vulnerable to adversarial examples, in which DNNs are misled to false outputs due to inputs containing imperceptible perturbations. Adversarial training, a reliable and effective method of defense, may significantly reduce the vulnerability of neural networks and becomes the de facto standard for robust learning. While many recent works practice the data-centric philosophy, such as how to generate better adversarial examples or use generative models to produce additional training data, we look back to the models themselves and revisit the adversarial robustness from the perspective of deep feature distribution as an insightful complementarity. In this paper, we propose \\textit{Branch Orthogonality adveRsarial Training} (BORT) to obtain state-of-the-art performance with solely the original dataset for adversarial training. To practice our design idea of integrating multiple orthogonal solution spaces, we leverage a simple multi-branch neural network and propose a corresponding loss function, branch-orthogonal loss, to make each solution space of the multi-branch model orthogonal. We evaluate our approach on CIFAR-10, CIFAR-100 and SVHN against $\\ell_{\\infty}$ norm-bounded perturbations of size $\\epsilon = 8/255$, respectively. Exhaustive experiments are conducted to show that our method goes beyond all state-of-the-art methods without any tricks. Compared to all methods that do not use additional data for training, our models achieve 67.3\\% and 41.5\\% robust accuracy on CIFAR-10 and CIFAR-100 (improving upon the state-of-the-art by +7.23\\% and +9.07\\%).", "date": "2022-08-17", "categories": []}
{"id": "2501.11462", "title": "On the Adversarial Vulnerabilities of Transfer Learning in Remote Sensing", "abstract": "The use of pretrained models from general computer vision tasks is widespread in remote sensing, significantly reducing training costs and improving performance. However, this practice also introduces vulnerabilities to downstream tasks, where publicly available pretrained models can be used as a proxy to compromise downstream models. This paper presents a novel Adversarial Neuron Manipulation method, which generates transferable perturbations by selectively manipulating single or multiple neurons in pretrained models. Unlike existing attacks, this method eliminates the need for domain-specific information, making it more broadly applicable and efficient. By targeting multiple fragile neurons, the perturbations achieve superior attack performance, revealing critical vulnerabilities in deep learning models. Experiments on diverse models and remote sensing datasets validate the effectiveness of the proposed method. This low-access adversarial neuron manipulation technique highlights a significant security risk in transfer learning models, emphasizing the urgent need for more robust defenses in their design when addressing the safety-critical remote sensing tasks.", "date": "2025-01-20", "categories": []}
{"id": "2511.07170", "title": "On Stealing Graph Neural Network Models", "abstract": "Current graph neural network (GNN) model-stealing methods rely heavily on queries to the victim model, assuming no hard query limits. However, in reality, the number of allowed queries can be severely limited. In this paper, we demonstrate how an adversary can extract a GNN with very limited interactions with the model. Our approach first enables the adversary to obtain the model backbone without making direct queries to the victim model and then to strategically utilize a fixed query limit to extract the most informative data. The experiments on eight real-world datasets demonstrate the effectiveness of the attack, even under a very restricted query limit and under defense against model extraction in place. Our findings underscore the need for robust defenses against GNN model extraction threats.", "date": "2025-11-10", "categories": []}
{"id": "2511.08937", "title": "Boosting Adversarial Transferability via Ensemble Non-Attention", "abstract": "Ensemble attacks integrate the outputs of surrogate models with diverse architectures, which can be combined with various gradient-based attacks to improve adversarial transferability. However, previous work shows unsatisfactory attack performance when transferring across heterogeneous model architectures. The main reason is that the gradient update directions of heterogeneous surrogate models differ widely, making it hard to reduce the gradient variance of ensemble models while making the best of individual model. To tackle this challenge, we design a novel ensemble attack, NAMEA, which for the first time integrates the gradients from the non-attention areas of ensemble models into the iterative gradient optimization process. Our design is inspired by the observation that the attention areas of heterogeneous models vary sharply, thus the non-attention areas of ViTs are likely to be the focus of CNNs and vice versa. Therefore, we merge the gradients respectively from the attention and non-attention areas of ensemble models so as to fuse the transfer information of CNNs and ViTs. Specifically, we pioneer a new way of decoupling the gradients of non-attention areas from those of attention areas, while merging gradients by meta-learning. Empirical evaluations on ImageNet dataset indicate that NAMEA outperforms AdaEA and SMER, the state-of-the-art ensemble attacks by an average of 15.0% and 9.6%, respectively. This work is the first attempt to explore the power of ensemble non-attention in boosting cross-architecture transferability, providing new insights into launching ensemble attacks.", "date": "2025-11-12", "categories": []}
{"id": "2511.09681", "title": "SEBA: Sample-Efficient Black-Box Attacks on Visual Reinforcement Learning", "abstract": "Visual reinforcement learning has achieved remarkable progress in visual control and robotics, but its vulnerability to adversarial perturbations remains underexplored. Most existing black-box attacks focus on vector-based or discrete-action RL, and their effectiveness on image-based continuous control is limited by the large action space and excessive environment queries. We propose SEBA, a sample-efficient framework for black-box adversarial attacks on visual RL agents. SEBA integrates a shadow Q model that estimates cumulative rewards under adversarial conditions, a generative adversarial network that produces visually imperceptible perturbations, and a world model that simulates environment dynamics to reduce real-world queries. Through a two-stage iterative training procedure that alternates between learning the shadow model and refining the generator, SEBA achieves strong attack performance while maintaining efficiency. Experiments on MuJoCo and Atari benchmarks show that SEBA significantly reduces cumulative rewards, preserves visual fidelity, and greatly decreases environment interactions compared to prior black-box and white-box methods.", "date": "2025-11-12", "categories": []}
{"id": "2511.09834", "title": "CertMask: Certifiable Defense Against Adversarial Patches via Theoretically Optimal Mask Coverage", "abstract": "Adversarial patch attacks inject localized perturbations into images to mislead deep vision models. These attacks can be physically deployed, posing serious risks to real-world applications. In this paper, we propose CertMask, a certifiably robust defense that constructs a provably sufficient set of binary masks to neutralize patch effects with strong theoretical guarantees. While the state-of-the-art approach (PatchCleanser) requires two rounds of masking and incurs $O(n^2)$ inference cost, CertMask performs only a single round of masking with $O(n)$ time complexity, where $n$ is the cardinality of the mask set to cover an input image. Our proposed mask set is computed using a mathematically rigorous coverage strategy that ensures each possible patch location is covered at least $k$ times, providing both efficiency and robustness. We offer a theoretical analysis of the coverage condition and prove its sufficiency for certification. Experiments on ImageNet, ImageNette, and CIFAR-10 show that CertMask improves certified robust accuracy by up to +13.4\\% over PatchCleanser, while maintaining clean accuracy nearly identical to the vanilla model.", "date": "2025-11-13", "categories": []}
{"id": "2511.09933", "title": "Debiased Dual-Invariant Defense for Adversarially Robust Person Re-Identification", "abstract": "Person re-identification (ReID) is a fundamental task in many real-world applications such as pedestrian trajectory tracking. However, advanced deep learning-based ReID models are highly susceptible to adversarial attacks, where imperceptible perturbations to pedestrian images can cause entirely incorrect predictions, posing significant security threats. Although numerous adversarial defense strategies have been proposed for classification tasks, their extension to metric learning tasks such as person ReID remains relatively unexplored. Moreover, the several existing defenses for person ReID fail to address the inherent unique challenges of adversarially robust ReID. In this paper, we systematically identify the challenges of adversarial defense in person ReID into two key issues: model bias and composite generalization requirements. To address them, we propose a debiased dual-invariant defense framework composed of two main phases. In the data balancing phase, we mitigate model bias using a diffusion-model-based data resampling strategy that promotes fairness and diversity in training data. In the bi-adversarial self-meta defense phase, we introduce a novel metric adversarial training approach incorporating farthest negative extension softening to overcome the robustness degradation caused by the absence of classifier. Additionally, we introduce an adversarially-enhanced self-meta mechanism to achieve dual-generalization for both unseen identities and unseen attack types. Experiments demonstrate that our method significantly outperforms existing state-of-the-art defenses.", "date": "2025-11-13", "categories": []}
{"id": "2511.10050", "title": "Trapped by Their Own Light: Deployable and Stealth Retroreflective Patch Attacks on Traffic Sign Recognition Systems", "abstract": "Traffic sign recognition plays a critical role in ensuring safe and efficient transportation of autonomous vehicles but remain vulnerable to adversarial attacks using stickers or laser projections. While existing attack vectors demonstrate security concerns, they suffer from visual detectability or implementation constraints, suggesting unexplored vulnerability surfaces in TSR systems. We introduce the Adversarial Retroreflective Patch (ARP), a novel attack vector that combines the high deployability of patch attacks with the stealthiness of laser projections by utilizing retroreflective materials activated only under victim headlight illumination. We develop a retroreflection simulation method and employ black-box optimization to maximize attack effectiveness. ARP achieves $\\geq$93.4\\% success rate in dynamic scenarios at 35 meters and $\\geq$60\\% success rate against commercial TSR systems in real-world conditions. Our user study demonstrates that ARP attacks maintain near-identical stealthiness to benign signs while achieving $\\geq$1.9\\% higher stealthiness scores than previous patch attacks. We propose the DPR Shield defense, employing strategically placed polarized filters, which achieves $\\geq$75\\% defense success rates for stop signs and speed limit signs against micro-prism patches.", "date": "2025-11-13", "categories": []}
{"id": "2511.10088", "title": "eXIAA: eXplainable Injections for Adversarial Attack", "abstract": "Post-hoc explainability methods are a subset of Machine Learning (ML) that aim to provide a reason for why a model behaves in a certain way. In this paper, we show a new black-box model-agnostic adversarial attack for post-hoc explainable Artificial Intelligence (XAI), particularly in the image domain. The goal of the attack is to modify the original explanations while being undetected by the human eye and maintain the same predicted class. In contrast to previous methods, we do not require any access to the model or its weights, but only to the model's computed predictions and explanations. Additionally, the attack is accomplished in a single step while significantly changing the provided explanations, as demonstrated by empirical evaluation. The low requirements of our method expose a critical vulnerability in current explainability methods, raising concerns about their reliability in safety-critical applications. We systematically generate attacks based on the explanations generated by post-hoc explainability methods (saliency maps, integrated gradients, and DeepLIFT SHAP) for pretrained ResNet-18 and ViT-B16 on ImageNet. The results show that our attacks could lead to dramatically different explanations without changing the predictive probabilities. We validate the effectiveness of our attack, compute the induced change based on the explanation with mean absolute difference, and verify the closeness of the original image and the corrupted one with the Structural Similarity Index Measure (SSIM).", "date": "2025-11-13", "categories": []}
{"id": "2511.10382", "title": "Fragile by Design: On the Limits of Adversarial Defenses in Personalized Generation", "abstract": "Personalized AI applications such as DreamBooth enable the generation of customized content from user images, but also raise significant privacy concerns, particularly the risk of facial identity leakage. Recent defense mechanisms like Anti-DreamBooth attempt to mitigate this risk by injecting adversarial perturbations into user photos to prevent successful personalization. However, we identify two critical yet overlooked limitations of these methods. First, the adversarial examples often exhibit perceptible artifacts such as conspicuous patterns or stripes, making them easily detectable as manipulated content. Second, the perturbations are highly fragile, as even a simple, non-learned filter can effectively remove them, thereby restoring the model's ability to memorize and reproduce user identity. To investigate this vulnerability, we propose a novel evaluation framework, AntiDB_Purify, to systematically evaluate existing defenses under realistic purification threats, including both traditional image filters and adversarial purification. Results reveal that none of the current methods maintains their protective effectiveness under such threats. These findings highlight that current defenses offer a false sense of security and underscore the urgent need for more imperceptible and robust protections to safeguard user identity in personalized generation.", "date": "2025-11-13", "categories": []}
{"id": "2404.10776", "title": "Nearly Optimal Algorithms for Contextual Dueling Bandits from Adversarial Feedback", "abstract": "Learning from human feedback plays an important role in aligning generative models, such as large language models (LLM). However, the effectiveness of this approach can be influenced by adversaries, who may intentionally provide misleading preferences to manipulate the output in an undesirable or harmful direction. To tackle this challenge, we study a specific model within this problem domain--contextual dueling bandits with adversarial feedback, where the true preference label can be flipped by an adversary. We propose an algorithm, namely robust contextual dueling bandits, which is based on uncertainty-weighted maximum likelihood estimation. Our algorithm achieves an $\\tilde O(d\\sqrt{T}/\\kappa+dC/\\kappa)$ regret bound, where $T$ is the number of rounds, $d$ is the dimension of the context, $\\kappa$ is the lower bound of the derivative of the link function, and $ 0 \\le C \\le T$ is the total number of adversarial feedback. We also prove a lower bound to show that our regret bound is nearly optimal, both in scenarios with and without ($C=0$) adversarial feedback. Our work is the first to achieve nearly minimax optimal regret for dueling bandits in the presence of adversarial preference feedback. Additionally, for the sigmoid link function, we develop a novel algorithm that takes into account the effect of local derivatives in maximum likelihood estimation (MLE) analysis through a refined method for estimating the link function's derivative. This method helps us to eliminate the $\\kappa$ dependence in the leading term with respect to $T$, which reduces the exponential dependence on the parameter radius $B$ to a polynomial dependence. We conduct experiments to evaluate our proposed algorithm against various types of adversarial feedback. Experimental results demonstrate its superiority over the state-of-the-art dueling bandit algorithms in the presence of adversarial feedback.", "date": "2024-04-16", "categories": []}
{"id": "2408.05745", "title": "Improving Adversarial Transferability with Neighbourhood Gradient Information", "abstract": "Deep neural networks (DNNs) are known to be susceptible to adversarial examples, leading to significant performance degradation. In black-box attack scenarios, a considerable attack performance gap between the surrogate model and the target model persists. This work focuses on enhancing the transferability of adversarial examples to narrow this performance gap. We observe that the gradient information around the clean image, i.e., Neighbourhood Gradient Information (NGI), can offer high transferability.Based on this insight, we introduce NGI-Attack, incorporating Example Backtracking and Multiplex Mask strategies to exploit this gradient information and enhance transferability. Specifically, we first adopt Example Backtracking to accumulate Neighbourhood Gradient Information as the initial momentum term. Then, we utilize Multiplex Mask to form a multi-way attack strategy that forces the network to focus on non-discriminative regions, which can obtain richer gradient information during only a few iterations. Extensive experiments demonstrate that our approach significantly enhances adversarial transferability. Especially, when attacking numerous defense models, we achieve an average attack success rate of 95.2%. Notably, our method can seamlessly integrate with any off-the-shelf algorithm, enhancing their attack performance without incurring extra time costs.", "date": "2024-08-11", "categories": []}
{"id": "2503.01908", "title": "UDora: A Unified Red Teaming Framework against LLM Agents by Dynamically Hijacking Their Own Reasoning", "abstract": "Large Language Model (LLM) agents equipped with external tools have become increasingly powerful for complex tasks such as web shopping, automated email replies, and financial trading. However, these advancements amplify the risks of adversarial attacks, especially when agents can access sensitive external functionalities. Nevertheless, manipulating LLM agents into performing targeted malicious actions or invoking specific tools remains challenging, as these agents extensively reason or plan before executing final actions. In this work, we present UDora, a unified red teaming framework designed for LLM agents that dynamically hijacks the agent's reasoning processes to compel malicious behavior. Specifically, UDora first generates the model's reasoning trace for the given task, then automatically identifies optimal points within this trace to insert targeted perturbations. The resulting perturbed reasoning is then used as a surrogate response for optimization. By iteratively applying this process, the LLM agent will then be induced to undertake designated malicious actions or to invoke specific malicious tools. Our approach demonstrates superior effectiveness compared to existing methods across three LLM agent datasets. The code is available at https://github.com/AI-secure/UDora.", "date": "2025-02-28", "categories": []}
{"id": "2508.03213", "title": "The Power of Many: Synergistic Unification of Diverse Augmentations for Efficient Adversarial Robustness", "abstract": "Adversarial perturbations pose a significant threat to deep learning models. Adversarial Training (AT), the predominant defense method, faces challenges of high computational costs and a degradation in standard performance. While data augmentation offers an alternative path, existing techniques either yield limited robustness gains or incur substantial training overhead. Therefore, developing a defense mechanism that is both highly efficient and strongly robust is of paramount importance.In this work, we first conduct a systematic analysis of existing augmentation techniques, revealing that the synergy among diverse strategies -- rather than any single method -- is crucial for enhancing robustness. Based on this insight, we propose the Universal Adversarial Augmenter (UAA) framework, which is characterized by its plug-and-play nature and training efficiency. UAA decouples the expensive perturbation generation process from model training by pre-computing a universal transformation offline, which is then used to efficiently generate unique adversarial perturbations for each sample during training.Extensive experiments conducted on multiple benchmarks validate the effectiveness of UAA. The results demonstrate that UAA establishes a new state-of-the-art (SOTA) for data-augmentation-based adversarial defense strategies , without requiring the online generation of adversarial examples during training. This framework provides a practical and efficient pathway for building robust models,Our code is available in the supplementary materials.", "date": "2025-08-05", "categories": []}
{"id": "2511.08810", "title": "SIFT-Graph: Benchmarking Multimodal Defense Against Image Adversarial Attacks With Robust Feature Graph", "abstract": "Adversarial attacks expose a fundamental vulnerability in modern deep vision models by exploiting their dependence on dense, pixel-level representations that are highly sensitive to imperceptible perturbations. Traditional defense strategies typically operate within this fragile pixel domain, lacking mechanisms to incorporate inherently robust visual features. In this work, we introduce SIFT-Graph, a multimodal defense framework that enhances the robustness of traditional vision models by aggregating structurally meaningful features extracted from raw images using both handcrafted and learned modalities. Specifically, we integrate Scale-Invariant Feature Transform keypoints with a Graph Attention Network to capture scale and rotation invariant local structures that are resilient to perturbations. These robust feature embeddings are then fused with traditional vision model, such as Vision Transformer and Convolutional Neural Network, to form a unified, structure-aware and perturbation defensive model. Preliminary results demonstrate that our method effectively improves the visual model robustness against gradient-based white box adversarial attacks, while incurring only a marginal drop in clean accuracy.", "date": "2025-11-11", "categories": []}
{"id": "2511.09064", "title": "Diversifying Counterattacks: Orthogonal Exploration for Robust CLIP Inference", "abstract": "Vision-language pre-training models (VLPs) demonstrate strong multimodal understanding and zero-shot generalization, yet remain vulnerable to adversarial examples, raising concerns about their reliability. Recent work, Test-Time Counterattack (TTC), improves robustness by generating perturbations that maximize the embedding deviation of adversarial inputs using PGD, pushing them away from their adversarial representations. However, due to the fundamental difference in optimization objectives between adversarial attacks and counterattacks, generating counterattacks solely based on gradients with respect to the adversarial input confines the search to a narrow space. As a result, the counterattacks could overfit limited adversarial patterns and lack the diversity to fully neutralize a broad range of perturbations. In this work, we argue that enhancing the diversity and coverage of counterattacks is crucial to improving adversarial robustness in test-time defense. Accordingly, we propose Directional Orthogonal Counterattack (DOC), which augments counterattack optimization by incorporating orthogonal gradient directions and momentum-based updates. This design expands the exploration of the counterattack space and increases the diversity of perturbations, which facilitates the discovery of more generalizable counterattacks and ultimately improves the ability to neutralize adversarial perturbations. Meanwhile, we present a directional sensitivity score based on averaged cosine similarity to boost DOC by improving example discrimination and adaptively modulating the counterattack strength. Extensive experiments on 16 datasets demonstrate that DOC improves adversarial robustness under various attacks while maintaining competitive clean accuracy. Code is available at https://github.com/bookman233/DOC.", "date": "2025-11-12", "categories": []}
{"id": "2511.09088", "title": "Improving Sustainability of Adversarial Examples in Class-Incremental Learning", "abstract": "Current adversarial examples (AEs) are typically designed for static models. However, with the wide application of Class-Incremental Learning (CIL), models are no longer static and need to be updated with new data distributed and labeled differently from the old ones. As a result, existing AEs often fail after CIL updates due to significant domain drift. In this paper, we propose SAE to enhance the sustainability of AEs against CIL. The core idea of SAE is to enhance the robustness of AE semantics against domain drift by making them more similar to the target class while distinguishing them from all other classes. Achieving this is challenging, as relying solely on the initial CIL model to optimize AE semantics often leads to overfitting. To resolve the problem, we propose a Semantic Correction Module. This module encourages the AE semantics to be generalized, based on a visual-language model capable of producing universal semantics. Additionally, it incorporates the CIL model to correct the optimization direction of the AE semantics, guiding them closer to the target class. To further reduce fluctuations in AE semantics, we propose a Filtering-and-Augmentation Module, which first identifies non-target examples with target-class semantics in the latent space and then augments them to foster more stable semantics. Comprehensive experiments demonstrate that SAE outperforms baselines by an average of 31.28% when updated with a 9-fold increase in the number of classes.", "date": "2025-11-12", "categories": []}
{"id": "2511.09105", "title": "Cost-Minimized Label-Flipping Poisoning Attack to LLM Alignment", "abstract": "Large language models (LLMs) are increasingly deployed in real-world systems, making it critical to understand their vulnerabilities. While data poisoning attacks during RLHF/DPO alignment have been studied empirically, their theoretical foundations remain unclear. We investigate the minimum-cost poisoning attack required to steer an LLM's policy toward an attacker's target by flipping preference labels during RLHF/DPO, without altering the compared outputs. We formulate this as a convex optimization problem with linear constraints, deriving lower and upper bounds on the minimum attack cost. As a byproduct of this theoretical analysis, we show that any existing label-flipping attack can be post-processed via our proposed method to reduce the number of label flips required while preserving the intended poisoning effect. Empirical results demonstrate that this cost-minimization post-processing can significantly reduce poisoning costs over baselines, particularly when the reward model's feature dimension is small relative to the dataset size. These findings highlight fundamental vulnerabilities in RLHF/DPO pipelines and provide tools to evaluate their robustness against low-cost poisoning attacks.", "date": "2025-11-12", "categories": []}
{"id": "2511.07701", "title": "Diffusion Guided Adversarial State Perturbations in Reinforcement Learning", "abstract": "Reinforcement learning (RL) systems, while achieving remarkable success across various domains, are vulnerable to adversarial attacks. This is especially a concern in vision-based environments where minor manipulations of high-dimensional image inputs can easily mislead the agent's behavior. To this end, various defenses have been proposed recently, with state-of-the-art approaches achieving robust performance even under large state perturbations. However, after closer investigation, we found that the effectiveness of the current defenses is due to a fundamental weakness of the existing $l_p$ norm-constrained attacks, which can barely alter the semantics of image input even under a relatively large perturbation budget. In this work, we propose SHIFT, a novel policy-agnostic diffusion-based state perturbation attack to go beyond this limitation. Our attack is able to generate perturbed states that are semantically different from the true states while remaining realistic and history-aligned to avoid detection. Evaluations show that our attack effectively breaks existing defenses, including the most sophisticated ones, significantly outperforming existing attacks while being more perceptually stealthy. The results highlight the vulnerability of RL agents to semantics-aware adversarial perturbations, indicating the importance of developing more robust policies.", "date": "2025-11-10", "categories": []}
{"id": "2511.07755", "title": "Filtered-ViT: A Robust Defense Against Multiple Adversarial Patch Attacks", "abstract": "Deep learning vision systems are increasingly deployed in safety-critical domains such as healthcare, yet they remain vulnerable to small adversarial patches that can trigger misclassifications. Most existing defenses assume a single patch and fail when multiple localized disruptions occur, the type of scenario adversaries and real-world artifacts often exploit. We propose Filtered-ViT, a new vision transformer architecture that integrates SMART Vector Median Filtering (SMART-VMF), a spatially adaptive, multi-scale, robustness-aware mechanism that enables selective suppression of corrupted regions while preserving semantic detail. On ImageNet with LaVAN multi-patch attacks, Filtered-ViT achieves 79.8% clean accuracy and 46.3% robust accuracy under four simultaneous 1\\% patches, outperforming existing defenses. Beyond synthetic benchmarks, a real-world case study on radiographic medical imagery shows that Filtered-ViT mitigates natural artifacts such as occlusions and scanner noise without degrading diagnostic content. This establishes Filtered-ViT as the first transformer to demonstrate unified robustness against both adversarial and naturally occurring patch-like disruptions, charting a path toward reliable vision systems in truly high-stakes environments.", "date": "2025-11-11", "categories": []}
{"id": "2511.08331", "title": "Adversarial Bias: Data Poisoning Attacks on Fairness", "abstract": "With the growing adoption of AI and machine learning systems in real-world applications, ensuring their fairness has become increasingly critical. The majority of the work in algorithmic fairness focus on assessing and improving the fairness of machine learning systems. There is relatively little research on fairness vulnerability, i.e., how an AI system's fairness can be intentionally compromised. In this work, we first provide a theoretical analysis demonstrating that a simple adversarial poisoning strategy is sufficient to induce maximally unfair behavior in naive Bayes classifiers. Our key idea is to strategically inject a small fraction of carefully crafted adversarial data points into the training set, biasing the model's decision boundary to disproportionately affect a protected group while preserving generalizable performance. To illustrate the practical effectiveness of our method, we conduct experiments across several benchmark datasets and models. We find that our attack significantly outperforms existing methods in degrading fairness metrics across multiple models and datasets, often achieving substantially higher levels of unfairness with a comparable or only slightly worse impact on accuracy. Notably, our method proves effective on a wide range of models, in contrast to prior work, demonstrating a robust and potent approach to compromising the fairness of machine learning systems.", "date": "2025-11-11", "categories": []}
{"id": "2403.15010", "title": "Clean-image Backdoor Attacks", "abstract": "To gather a significant quantity of annotated training data for high-performance image classification models, numerous companies opt to enlist third-party providers to label their unlabeled data. This practice is widely regarded as secure, even in cases where some annotated errors occur, as the impact of these minor inaccuracies on the final performance of the models is negligible and existing backdoor attacks require attacker's ability to poison the training images. Nevertheless, in this paper, we propose clean-image backdoor attacks which uncover that backdoors can still be injected via a fraction of incorrect labels without modifying the training images. Specifically, in our attacks, the attacker first seeks a trigger feature to divide the training images into two parts: those with the feature and those without it. Subsequently, the attacker falsifies the labels of the former part to a backdoor class. The backdoor will be finally implanted into the target model after it is trained on the poisoned data. During the inference phase, the attacker can activate the backdoor in two ways: slightly modifying the input image to obtain the trigger feature, or taking an image that naturally has the trigger feature as input. We conduct extensive experiments to demonstrate the effectiveness and practicality of our attacks. According to the experimental results, we conclude that our attacks seriously jeopardize the fairness and robustness of image classification models, and it is necessary to be vigilant about the incorrect labels in outsourced labeling.", "date": "2024-03-22", "categories": []}
{"id": "2406.14682", "title": "Uniform Convergence of Adversarially Robust Classifiers", "abstract": "In recent years there has been significant interest in the effect of different types of adversarial perturbations in data classification problems. Many of these models incorporate the adversarial power, which is an important parameter with an associated trade-off between accuracy and robustness. This work considers a general framework for adversarially-perturbed classification problems, in a large data or population-level limit. In such a regime, we demonstrate that as adversarial strength goes to zero that optimal classifiers converge to the Bayes classifier in the Hausdorff distance. This significantly strengthens previous results, which generally focus on $L^1$-type convergence. The main argument relies upon direct geometric comparisons and is inspired by techniques from geometric measure theory.", "date": "2024-06-20", "categories": []}
{"id": "2511.01746", "title": "Scam Shield: Multi-Model Voting and Fine-Tuned LLMs Against Adversarial Attacks", "abstract": "Scam detection remains a critical challenge in cybersecurity as adversaries craft messages that evade automated filters. We propose a Hierarchical Scam Detection System (HSDS) that combines a lightweight multi-model voting front end with a fine-tuned LLaMA 3.1 8B Instruct back end to improve accuracy and robustness against adversarial attacks. An ensemble of four classifiers provides preliminary predictions through majority vote, and ambiguous cases are escalated to the fine-tuned model, which is optimized with adversarial training to reduce misclassification. Experiments show that this hierarchical design both improves adversarial scam detection and shortens inference time by routing most cases away from the LLM, outperforming traditional machine-learning baselines and proprietary LLM baselines. The findings highlight the effectiveness of a hybrid voting mechanism and adversarial fine-tuning in fortifying LLMs against evolving scam tactics, enhancing the resilience of automated scam detection systems.", "date": "2025-11-03", "categories": []}
{"id": "2511.06143", "title": "Enhancing Robustness of Graph Neural Networks through p-Laplacian", "abstract": "With the increase of data in day-to-day life, businesses and different stakeholders need to analyze the data for better predictions. Traditionally, relational data has been a source of various insights, but with the increase in computational power and the need to understand deeper relationships between entities, the need to design new techniques has arisen. For this graph data analysis has become an extraordinary tool for understanding the data, which reveals more realistic and flexible modelling of complex relationships. Recently, Graph Neural Networks (GNNs) have shown great promise in various applications, such as social network analysis, recommendation systems, drug discovery, and more. However, many adversarial attacks can happen over the data, whether during training (poisoning attack) or during testing (evasion attack), which can adversely manipulate the desired outcome from the GNN model. Therefore, it is crucial to make the GNNs robust to such attacks. The existing robustness methods are computationally demanding and perform poorly when the intensity of attack increases. This paper presents a computationally efficient framework, namely, pLAPGNN, based on weighted p-Laplacian for making GNNs robust. Empirical evaluation on real datasets establishes the efficacy and efficiency of the proposed method.", "date": "2025-11-08", "categories": []}
{"id": "2511.06212", "title": "RAG-targeted Adversarial Attack on LLM-based Threat Detection and Mitigation Framework", "abstract": "The rapid expansion of the Internet of Things (IoT) is reshaping communication and operational practices across industries, but it also broadens the attack surface and increases susceptibility to security breaches. Artificial Intelligence has become a valuable solution in securing IoT networks, with Large Language Models (LLMs) enabling automated attack behavior analysis and mitigation suggestion in Network Intrusion Detection Systems (NIDS). Despite advancements, the use of LLMs in such systems further expands the attack surface, putting entire networks at risk by introducing vulnerabilities such as prompt injection and data poisoning. In this work, we attack an LLM-based IoT attack analysis and mitigation framework to test its adversarial robustness. We construct an attack description dataset and use it in a targeted data poisoning attack that applies word-level, meaning-preserving perturbations to corrupt the Retrieval-Augmented Generation (RAG) knowledge base of the framework. We then compare pre-attack and post-attack mitigation responses from the target model, ChatGPT-5 Thinking, to measure the impact of the attack on model performance, using an established evaluation rubric designed for human experts and judge LLMs. Our results show that small perturbations degrade LLM performance by weakening the linkage between observed network traffic features and attack behavior, and by reducing the specificity and practicality of recommended mitigations for resource-constrained devices.", "date": "2025-11-09", "categories": []}
{"id": "2511.06495", "title": "Probably Approximately Global Robustness Certification", "abstract": "We propose and investigate probabilistic guarantees for the adversarial robustness of classification algorithms. While traditional formal verification approaches for robustness are intractable and sampling-based approaches do not provide formal guarantees, our approach is able to efficiently certify a probabilistic relaxation of robustness. The key idea is to sample an $\\epsilon$-net and invoke a local robustness oracle on the sample. Remarkably, the size of the sample needed to achieve probably approximately global robustness guarantees is independent of the input dimensionality, the number of classes, and the learning algorithm itself. Our approach can, therefore, be applied even to large neural networks that are beyond the scope of traditional formal verification. Experiments empirically confirm that it characterizes robustness better than state-of-the-art sampling-based approaches and scales better than formal methods.", "date": "2025-11-09", "categories": []}
{"id": "2511.06947", "title": "FoCLIP: A Feature-Space Misalignment Framework for CLIP-Based Image Manipulation and Detection", "abstract": "The well-aligned attribute of CLIP-based models enables its effective application like CLIPscore as a widely adopted image quality assessment metric. However, such a CLIP-based metric is vulnerable for its delicate multimodal alignment. In this work, we propose \\textbf{FoCLIP}, a feature-space misalignment framework for fooling CLIP-based image quality metric. Based on the stochastic gradient descent technique, FoCLIP integrates three key components to construct fooling examples: feature alignment as the core module to reduce image-text modality gaps, the score distribution balance module and pixel-guard regularization, which collectively optimize multimodal output equilibrium between CLIPscore performance and image quality. Such a design can be engineered to maximize the CLIPscore predictions across diverse input prompts, despite exhibiting either visual unrecognizability or semantic incongruence with the corresponding adversarial prompts from human perceptual perspectives. Experiments on ten artistic masterpiece prompts and ImageNet subsets demonstrate that optimized images can achieve significant improvement in CLIPscore while preserving high visual fidelity. In addition, we found that grayscale conversion induces significant feature degradation in fooling images, exhibiting noticeable CLIPscore reduction while preserving statistical consistency with original images. Inspired by this phenomenon, we propose a color channel sensitivity-driven tampering detection mechanism that achieves 91% accuracy on standard benchmarks. In conclusion, this work establishes a practical pathway for feature misalignment in CLIP-based multimodal systems and the corresponding defense method.", "date": "2025-11-10", "categories": []}
{"id": "2511.07049", "title": "From Pretrain to Pain: Adversarial Vulnerability of Video Foundation Models Without Task Knowledge", "abstract": "Large-scale Video Foundation Models (VFMs) has significantly advanced various video-related tasks, either through task-specific models or Multi-modal Large Language Models (MLLMs). However, the open accessibility of VFMs also introduces critical security risks, as adversaries can exploit full knowledge of the VFMs to launch potent attacks. This paper investigates a novel and practical adversarial threat scenario: attacking downstream models or MLLMs fine-tuned from open-source VFMs, without requiring access to the victim task, training data, model query, and architecture. In contrast to conventional transfer-based attacks that rely on task-aligned surrogate models, we demonstrate that adversarial vulnerabilities can be exploited directly from the VFMs. To this end, we propose the Transferable Video Attack (TVA), a temporal-aware adversarial attack method that leverages the temporal representation dynamics of VFMs to craft effective perturbations. TVA integrates a bidirectional contrastive learning mechanism to maximize the discrepancy between the clean and adversarial features, and introduces a temporal consistency loss that exploits motion cues to enhance the sequential impact of perturbations. TVA avoids the need to train expensive surrogate models or access to domain-specific data, thereby offering a more practical and efficient attack strategy. Extensive experiments across 24 video-related tasks demonstrate the efficacy of TVA against downstream models and MLLMs, revealing a previously underexplored security vulnerability in the deployment of video models.", "date": "2025-11-10", "categories": []}
{"id": "2511.07099", "title": "E2E-VGuard: Adversarial Prevention for Production LLM-based End-To-End Speech Synthesis", "abstract": "Recent advancements in speech synthesis technology have enriched our daily lives, with high-quality and human-like audio widely adopted across real-world applications. However, malicious exploitation like voice-cloning fraud poses severe security risks. Existing defense techniques struggle to address the production large language model (LLM)-based speech synthesis. While previous studies have considered the protection for fine-tuning synthesizers, they assume manually annotated transcripts. Given the labor intensity of manual annotation, end-to-end (E2E) systems leveraging automatic speech recognition (ASR) to generate transcripts are becoming increasingly prevalent, e.g., voice cloning via commercial APIs. Therefore, this E2E speech synthesis also requires new security mechanisms. To tackle these challenges, we propose E2E-VGuard, a proactive defense framework for two emerging threats: (1) production LLM-based speech synthesis, and (2) the novel attack arising from ASR-driven E2E scenarios. Specifically, we employ the encoder ensemble with a feature extractor to protect timbre, while ASR-targeted adversarial examples disrupt pronunciation. Moreover, we incorporate the psychoacoustic model to ensure perturbative imperceptibility. For a comprehensive evaluation, we test 16 open-source synthesizers and 3 commercial APIs across Chinese and English datasets, confirming E2E-VGuard's effectiveness in timbre and pronunciation protection. Real-world deployment validation is also conducted. Our code and demo page are available at https://wxzyd123.github.io/e2e-vguard/.", "date": "2025-11-10", "categories": []}
{"id": "2511.07379", "title": "LoReTTA: A Low Resource Framework To Poison Continuous Time Dynamic Graphs", "abstract": "Temporal Graph Neural Networks (TGNNs) are increasingly used in high-stakes domains, such as financial forecasting, recommendation systems, and fraud detection. However, their susceptibility to poisoning attacks poses a critical security risk. We introduce LoReTTA (Low Resource Two-phase Temporal Attack), a novel adversarial framework on Continuous-Time Dynamic Graphs, which degrades TGNN performance by an average of 29.47% across 4 widely benchmark datasets and 4 State-of-the-Art (SotA) models. LoReTTA operates through a two-stage approach: (1) sparsify the graph by removing high-impact edges using any of the 16 tested temporal importance metrics, (2) strategically replace removed edges with adversarial negatives via LoReTTA's novel degree-preserving negative sampling algorithm. Our plug-and-play design eliminates the need for expensive surrogate models while adhering to realistic unnoticeability constraints. LoReTTA degrades performance by upto 42.0% on MOOC, 31.5% on Wikipedia, 28.8% on UCI, and 15.6% on Enron. LoReTTA outperforms 11 attack baselines, remains undetectable to 4 leading anomaly detection systems, and is robust to 4 SotA adversarial defense training methods, establishing its effectiveness, unnoticeability, and robustness.", "date": "2025-11-10", "categories": []}
{"id": "2410.18469", "title": "Iterative Self-Tuning LLMs for Enhanced Jailbreaking Capabilities", "abstract": "Recent research has shown that Large Language Models (LLMs) are vulnerable to automated jailbreak attacks, where adversarial suffixes crafted by algorithms appended to harmful queries bypass safety alignment and trigger unintended responses. Current methods for generating these suffixes are computationally expensive and have low Attack Success Rates (ASR), especially against well-aligned models like Llama2 and Llama3. To overcome these limitations, we introduce ADV-LLM, an iterative self-tuning process that crafts adversarial LLMs with enhanced jailbreak ability. Our framework significantly reduces the computational cost of generating adversarial suffixes while achieving nearly 100\\% ASR on various open-source LLMs. Moreover, it exhibits strong attack transferability to closed-source models, achieving 99\\% ASR on GPT-3.5 and 49\\% ASR on GPT-4, despite being optimized solely on Llama3. Beyond improving jailbreak ability, ADV-LLM provides valuable insights for future safety alignment research through its ability to generate large datasets for studying LLM safety. Our code is available at: https://github.com/SunChungEn/ADV-LLM", "date": "2024-10-24", "categories": []}
{"id": "2511.04970", "title": "Learning Fourier shapes to probe the geometric world of deep neural networks", "abstract": "While both shape and texture are fundamental to visual recognition, research on deep neural networks (DNNs) has predominantly focused on the latter, leaving their geometric understanding poorly probed. Here, we show: first, that optimized shapes can act as potent semantic carriers, generating high-confidence classifications from inputs defined purely by their geometry; second, that they are high-fidelity interpretability tools that precisely isolate a model's salient regions; and third, that they constitute a new, generalizable adversarial paradigm capable of deceiving downstream visual tasks. This is achieved through an end-to-end differentiable framework that unifies a powerful Fourier series to parameterize arbitrary shapes, a winding number-based mapping to translate them into the pixel grid required by DNNs, and signal energy constraints that enhance optimization efficiency while ensuring physically plausible shapes. Our work provides a versatile framework for probing the geometric world of DNNs and opens new frontiers for challenging and understanding machine perception.", "date": "2025-11-07", "categories": []}
{"id": "2511.05073", "title": "Deep learning models are vulnerable, but adversarial examples are even more vulnerable", "abstract": "Understanding intrinsic differences between adversarial examples and clean samples is key to enhancing DNN robustness and detection against adversarial attacks. This study first empirically finds that image-based adversarial examples are notably sensitive to occlusion. Controlled experiments on CIFAR-10 used nine canonical attacks (e.g., FGSM, PGD) to generate adversarial examples, paired with original samples for evaluation. We introduce Sliding Mask Confidence Entropy (SMCE) to quantify model confidence fluctuation under occlusion. Using 1800+ test images, SMCE calculations supported by Mask Entropy Field Maps and statistical distributions show adversarial examples have significantly higher confidence volatility under occlusion than originals. Based on this, we propose Sliding Window Mask-based Adversarial Example Detection (SWM-AED), which avoids catastrophic overfitting of conventional adversarial training. Evaluations across classifiers and attacks on CIFAR-10 demonstrate robust performance, with accuracy over 62% in most cases and up to 96.5%.", "date": "2025-11-07", "categories": []}
{"id": "2511.05102", "title": "Quantifying the Risk of Transferred Black Box Attacks", "abstract": "Neural networks have become pervasive across various applications, including security-related products. However, their widespread adoption has heightened concerns regarding vulnerability to adversarial attacks. With emerging regulations and standards emphasizing security, organizations must reliably quantify risks associated with these attacks, particularly regarding transferred adversarial attacks, which remain challenging to evaluate accurately. This paper investigates the complexities involved in resilience testing against transferred adversarial attacks. Our analysis specifically addresses black-box evasion attacks, highlighting transfer-based attacks due to their practical significance and typically high transferability between neural network models. We underline the computational infeasibility of exhaustively exploring high-dimensional input spaces to achieve complete test coverage. As a result, comprehensive adversarial risk mapping is deemed impractical. To mitigate this limitation, we propose a targeted resilience testing framework that employs surrogate models strategically selected based on Centered Kernel Alignment (CKA) similarity. By leveraging surrogate models exhibiting both high and low CKA similarities relative to the target model, the proposed approach seeks to optimize coverage of adversarial subspaces. Risk estimation is conducted using regression-based estimators, providing organizations with realistic and actionable risk quantification.", "date": "2025-11-07", "categories": []}
{"id": "2511.05177", "title": "Associative Poisoning to Generative Machine Learning", "abstract": "The widespread adoption of generative models such as Stable Diffusion and ChatGPT has made them increasingly attractive targets for malicious exploitation, particularly through data poisoning. Existing poisoning attacks compromising synthesised data typically either cause broad degradation of generated data or require control over the training process, limiting their applicability in real-world scenarios. In this paper, we introduce a novel data poisoning technique called associative poisoning, which compromises fine-grained features of the generated data without requiring control of the training process. This attack perturbs only the training data to manipulate statistical associations between specific feature pairs in the generated outputs. We provide a formal mathematical formulation of the attack and prove its theoretical feasibility and stealthiness. Empirical evaluations using two state-of-the-art generative models demonstrate that associative poisoning effectively induces or suppresses feature associations while preserving the marginal distributions of the targeted features and maintaining high-quality outputs, thereby evading visual detection. These results suggest that generative systems used in image synthesis, synthetic dataset generation, and natural language processing are susceptible to subtle, stealthy manipulations that compromise their statistical integrity. To address this risk, we examine the limitations of existing defensive strategies and propose a novel countermeasure strategy.", "date": "2025-11-07", "categories": []}
{"id": "2511.05269", "title": "TAMAS: Benchmarking Adversarial Risks in Multi-Agent LLM Systems", "abstract": "Large Language Models (LLMs) have demonstrated strong capabilities as autonomous agents through tool use, planning, and decision-making abilities, leading to their widespread adoption across diverse tasks. As task complexity grows, multi-agent LLM systems are increasingly used to solve problems collaboratively. However, safety and security of these systems remains largely under-explored. Existing benchmarks and datasets predominantly focus on single-agent settings, failing to capture the unique vulnerabilities of multi-agent dynamics and co-ordination. To address this gap, we introduce $\\textbf{T}$hreats and $\\textbf{A}$ttacks in $\\textbf{M}$ulti-$\\textbf{A}$gent $\\textbf{S}$ystems ($\\textbf{TAMAS}$), a benchmark designed to evaluate the robustness and safety of multi-agent LLM systems. TAMAS includes five distinct scenarios comprising 300 adversarial instances across six attack types and 211 tools, along with 100 harmless tasks. We assess system performance across ten backbone LLMs and three agent interaction configurations from Autogen and CrewAI frameworks, highlighting critical challenges and failure modes in current multi-agent deployments. Furthermore, we introduce Effective Robustness Score (ERS) to assess the tradeoff between safety and task effectiveness of these frameworks. Our findings show that multi-agent systems are highly vulnerable to adversarial attacks, underscoring the urgent need for stronger defenses. TAMAS provides a foundation for systematically studying and improving the safety of multi-agent LLM systems.", "date": "2025-11-07", "categories": []}
{"id": "2411.14133", "title": "GASP: Efficient Black-Box Generation of Adversarial Suffixes for Jailbreaking LLMs", "abstract": "LLMs have shown impressive capabilities across various natural language processing tasks, yet remain vulnerable to input prompts, known as jailbreak attacks, carefully designed to bypass safety guardrails and elicit harmful responses. Traditional methods rely on manual heuristics but suffer from limited generalizability. Despite being automatic, optimization-based attacks often produce unnatural prompts that can be easily detected by safety filters or require high computational costs due to discrete token optimization. In this paper, we introduce Generative Adversarial Suffix Prompter (GASP), a novel automated framework that can efficiently generate human-readable jailbreak prompts in a fully black-box setting. In particular, GASP leverages latent Bayesian optimization to craft adversarial suffixes by efficiently exploring continuous latent embedding spaces, gradually optimizing the suffix prompter to improve attack efficacy while balancing prompt coherence via a targeted iterative refinement procedure. Through comprehensive experiments, we show that GASP can produce natural adversarial prompts, significantly improving jailbreak success over baselines, reducing training times, and accelerating inference speed, thus making it an efficient and scalable solution for red-teaming LLMs.", "date": "2024-11-21", "categories": []}
{"id": "2511.03128", "title": "From Insight to Exploit: Leveraging LLM Collaboration for Adaptive Adversarial Text Generation", "abstract": "LLMs can provide substantial zero-shot performance on diverse tasks using a simple task prompt, eliminating the need for training or fine-tuning. However, when applying these models to sensitive tasks, it is crucial to thoroughly assess their robustness against adversarial inputs. In this work, we introduce Static Deceptor (StaDec) and Dynamic Deceptor (DyDec), two innovative attack frameworks designed to systematically generate dynamic and adaptive adversarial examples by leveraging the understanding of the LLMs. We produce subtle and natural-looking adversarial inputs that preserve semantic similarity to the original text while effectively deceiving the target LLM. By utilizing an automated, LLM-driven pipeline, we eliminate the dependence on external heuristics. Our attacks evolve with the advancements in LLMs and demonstrate strong transferability across models unknown to the attacker. Overall, this work provides a systematic approach for the self-assessment of an LLM's robustness. We release our code and data at https://github.com/Shukti042/AdversarialExample.", "date": "2025-11-05", "categories": []}
{"id": "2406.16200", "title": "Feature compression is the root cause of adversarial fragility in neural network classifiers", "abstract": "In this paper, we uniquely study the adversarial robustness of deep neural networks (NN) for classification tasks against that of optimal classifiers. We look at the smallest magnitude of possible additive perturbations that can change a classifier's output. We provide a matrix-theoretic explanation of the adversarial fragility of deep neural networks for classification. In particular, our theoretical results show that a neural network's adversarial robustness can degrade as the input dimension $d$ increases. Analytically, we show that neural networks' adversarial robustness can be only $1/\\sqrt{d}$ of the best possible adversarial robustness of optimal classifiers. Our theories match remarkably well with numerical experiments of practically trained NN, including NN for ImageNet images. The matrix-theoretic explanation is consistent with an earlier information-theoretic feature-compression-based explanation for the adversarial fragility of neural networks.", "date": "2024-06-23", "categories": []}
{"id": "2412.00621", "title": "Exposing LLM Vulnerabilities: Adversarial Scam Detection and Performance", "abstract": "Can we trust Large Language Models (LLMs) to accurately predict scam? This paper investigates the vulnerabilities of LLMs when facing adversarial scam messages for the task of scam detection. We addressed this issue by creating a comprehensive dataset with fine-grained labels of scam messages, including both original and adversarial scam messages. The dataset extended traditional binary classes for the scam detection task into more nuanced scam types. Our analysis showed how adversarial examples took advantage of vulnerabilities of a LLM, leading to high misclassification rate. We evaluated the performance of LLMs on these adversarial scam messages and proposed strategies to improve their robustness.", "date": "2024-12-01", "categories": []}
{"id": "2310.04687", "title": "Targeted Attack Improves Protection against Unauthorized Diffusion Customization", "abstract": "Diffusion models build a new milestone for image generation yet raising public concerns, for they can be fine-tuned on unauthorized images for customization. Protection based on adversarial attacks rises to encounter this unauthorized diffusion customization, by adding protective watermarks to images and poisoning diffusion models. However, current protection, leveraging untargeted attacks, does not appear to be effective enough. In this paper, we propose a simple yet effective improvement for the protection against unauthorized diffusion customization by introducing targeted attacks. We show that by carefully selecting the target, targeted attacks significantly outperform untargeted attacks in poisoning diffusion models and degrading the customization image quality. Extensive experiments validate the superiority of our method on two mainstream customization methods of diffusion models, compared to existing protections. To explain the surprising success of targeted attacks, we delve into the mechanism of attack-based protections and propose a hypothesis based on our observation, which enhances the comprehension of attack-based protections. To the best of our knowledge, we are the first to both reveal the vulnerability of diffusion models to targeted attacks and leverage targeted attacks to enhance protection against unauthorized diffusion customization. Our code is available on GitHub: https://github.com/psyker-team/mist-v2.", "date": "2023-10-07", "categories": []}
{"id": "2506.16690", "title": "DepthVanish: Optimizing Adversarial Interval Structures for Stereo-Depth-Invisible Patches", "abstract": "Stereo depth estimation is a critical task in autonomous driving and robotics, where inaccuracies (such as misidentifying nearby objects as distant) can lead to dangerous situations. Adversarial attacks against stereo depth estimation can help reveal vulnerabilities before deployment. Previous works have shown that repeating optimized textures can effectively mislead stereo depth estimation in digital settings. However, our research reveals that these naively repeated textures perform poorly in physical implementations, i.e., when deployed as patches, limiting their practical utility for stress-testing stereo depth estimation systems. In this work, for the first time, we discover that introducing regular intervals among the repeated textures, creating a grid structure, significantly enhances the patch's attack performance. Through extensive experimentation, we analyze how variations of this novel structure influence the adversarial effectiveness. Based on these insights, we develop a novel stereo depth attack that jointly optimizes both the interval structure and texture elements. Our generated adversarial patches can be inserted into any scenes and successfully attack advanced stereo depth estimation methods of different paradigms, i.e., RAFT-Stereo and STTR. Most critically, our patch can also attack commercial RGB-D cameras (Intel RealSense) in real-world conditions, demonstrating their practical relevance for security assessment of stereo systems. The code is officially released at: https://github.com/WiWiN42/DepthVanish", "date": "2025-06-20", "categories": []}
{"id": "2508.05489", "title": "Keep It Real: Challenges in Attacking Compression-Based Adversarial Purification", "abstract": "Previous work has suggested that preprocessing images through lossy compression can defend against adversarial perturbations, but comprehensive attack evaluations have been lacking. In this paper, we construct strong white-box and adaptive attacks against various compression models and identify a critical challenge for attackers: high realism in reconstructed images significantly increases attack difficulty. Through rigorous evaluation across multiple attack scenarios, we demonstrate that compression models capable of producing realistic, high-fidelity reconstructions are substantially more resistant to our attacks. In contrast, low-realism compression models can be broken. Our analysis reveals that this is not due to gradient masking. Rather, realistic reconstructions maintaining distributional alignment with natural images seem to offer inherent robustness. This work highlights a significant obstacle for future adversarial attacks and suggests that developing more effective techniques to overcome realism represents an essential challenge for comprehensive security evaluation.", "date": "2025-08-07", "categories": []}
{"id": "2510.16171", "title": "Bridging Symmetry and Robustness: On the Role of Equivariance in Enhancing Adversarial Robustness", "abstract": "Adversarial examples reveal critical vulnerabilities in deep neural networks by exploiting their sensitivity to imperceptible input perturbations. While adversarial training remains the predominant defense strategy, it often incurs significant computational cost and may compromise clean-data accuracy. In this work, we investigate an architectural approach to adversarial robustness by embedding group-equivariant convolutions-specifically, rotation- and scale-equivariant layers-into standard convolutional neural networks (CNNs). These layers encode symmetry priors that align model behavior with structured transformations in the input space, promoting smoother decision boundaries and greater resilience to adversarial attacks. We propose and evaluate two symmetry-aware architectures: a parallel design that processes standard and equivariant features independently before fusion, and a cascaded design that applies equivariant operations sequentially. Theoretically, we demonstrate that such models reduce hypothesis space complexity, regularize gradients, and yield tighter certified robustness bounds under the CLEVER (Cross Lipschitz Extreme Value for nEtwork Robustness) framework. Empirically, our models consistently improve adversarial robustness and generalization across CIFAR-10, CIFAR-100, and CIFAR-10C under both FGSM and PGD attacks, without requiring adversarial training. These findings underscore the potential of symmetry-enforcing architectures as efficient and principled alternatives to data augmentation-based defenses.", "date": "2025-10-17", "categories": []}
{"id": "2511.00083", "title": "Fixed-point graph convolutional networks against adversarial attacks", "abstract": "Adversarial attacks present a significant risk to the integrity and performance of graph neural networks, particularly in tasks where graph structure and node features are vulnerable to manipulation. In this paper, we present a novel model, called fixed-point iterative graph convolutional network (Fix-GCN), which achieves robustness against adversarial perturbations by effectively capturing higher-order node neighborhood information in the graph without additional memory or computational complexity. Specifically, we introduce a versatile spectral modulation filter and derive the feature propagation rule of our model using fixed-point iteration. Unlike traditional defense mechanisms that rely on additional design elements to counteract attacks, the proposed graph filter provides a flexible-pass filtering approach, allowing it to selectively attenuate high-frequency components while preserving low-frequency structural information in the graph signal. By iteratively updating node representations, our model offers a flexible and efficient framework for preserving essential graph information while mitigating the impact of adversarial manipulation. We demonstrate the effectiveness of the proposed model through extensive experiments on various benchmark graph datasets, showcasing its resilience against adversarial attacks.", "date": "2025-10-29", "categories": []}
{"id": "2511.00143", "title": "BlurGuard: A Simple Approach for Robustifying Image Protection Against AI-Powered Editing", "abstract": "Recent advances in text-to-image models have increased the exposure of powerful image editing techniques as a tool, raising concerns about their potential for malicious use. An emerging line of research to address such threats focuses on implanting \"protective\" adversarial noise into images before their public release, so future attempts to edit them using text-to-image models can be impeded. However, subsequent works have shown that these adversarial noises are often easily \"reversed,\" e.g., with techniques as simple as JPEG compression, casting doubt on the practicality of the approach. In this paper, we argue that adversarial noise for image protection should not only be imperceptible, as has been a primary focus of prior work, but also irreversible, viz., it should be difficult to detect as noise provided that the original image is hidden. We propose a surprisingly simple method to enhance the robustness of image protection methods against noise reversal techniques. Specifically, it applies an adaptive per-region Gaussian blur on the noise to adjust the overall frequency spectrum. Through extensive experiments, we show that our method consistently improves the per-sample worst-case protection performance of existing methods against a wide range of reversal techniques on diverse image editing scenarios, while also reducing quality degradation due to noise in terms of perceptual metrics. Code is available at https://github.com/jsu-kim/BlurGuard.", "date": "2025-10-31", "categories": []}
{"id": "2511.00203", "title": "Diffusion LLMs are Natural Adversaries for any LLM", "abstract": "We introduce a novel framework that transforms the resource-intensive (adversarial) prompt optimization problem into an \\emph{efficient, amortized inference task}. Our core insight is that pretrained, non-autoregressive generative LLMs, such as Diffusion LLMs, which model the joint distribution over prompt-response pairs, can serve as powerful surrogates for prompt search. This approach enables the direct conditional generation of prompts, effectively replacing costly, per-instance discrete optimization with a small number of parallelizable samples. We provide a probabilistic analysis demonstrating that under mild fidelity assumptions, only a few conditional samples are required to recover high-reward (harmful) prompts. Empirically, we find that the generated prompts are low-perplexity, diverse jailbreaks that exhibit strong transferability to a wide range of black-box target models, including robustly trained and proprietary LLMs. Beyond adversarial prompting, our framework opens new directions for red teaming, automated prompt optimization, and leveraging emerging Flow- and Diffusion-based LLMs.", "date": "2025-10-31", "categories": []}
{"id": "2511.00346", "title": "Exploiting Latent Space Discontinuities for Building Universal LLM Jailbreaks and Data Extraction Attacks", "abstract": "The rapid proliferation of Large Language Models (LLMs) has raised significant concerns about their security against adversarial attacks. In this work, we propose a novel approach to crafting universal jailbreaks and data extraction attacks by exploiting latent space discontinuities, an architectural vulnerability related to the sparsity of training data. Unlike previous methods, our technique generalizes across various models and interfaces, proving highly effective in seven state-of-the-art LLMs and one image generation model. Initial results indicate that when these discontinuities are exploited, they can consistently and profoundly compromise model behavior, even in the presence of layered defenses. The findings suggest that this strategy has substantial potential as a systemic attack vector.", "date": "2025-11-01", "categories": []}
{"id": "2511.00411", "title": "Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling", "abstract": "Adversarial attacks present a critical challenge to deep neural networks' robustness, particularly in transfer scenarios across different model architectures. However, the transferability of adversarial attacks faces a fundamental dilemma between Exploitation (maximizing attack potency) and Exploration (enhancing cross-model generalization). Traditional momentum-based methods over-prioritize Exploitation, i.e., higher loss maxima for attack potency but weakened generalization (narrow loss surface). Conversely, recent methods with inner-iteration sampling over-prioritize Exploration, i.e., flatter loss surfaces for cross-model generalization but weakened attack potency (suboptimal local maxima). To resolve this dilemma, we propose a simple yet effective Gradient-Guided Sampling (GGS), which harmonizes both objectives through guiding sampling along the gradient ascent direction to improve both sampling efficiency and stability. Specifically, based on MI-FGSM, GGS introduces inner-iteration random sampling and guides the sampling direction using the gradient from the previous inner-iteration (the sampling's magnitude is determined by a random distribution). This mechanism encourages adversarial examples to reside in balanced regions with both flatness for cross-model generalization and higher local maxima for strong attack potency. Comprehensive experiments across multiple DNN architectures and multimodal large language models (MLLMs) demonstrate the superiority of our method over state-of-the-art transfer attacks. Code is made available at https://github.com/anuin-cat/GGS.", "date": "2025-11-01", "categories": []}
{"id": "2511.00446", "title": "ToxicTextCLIP: Text-Based Poisoning and Backdoor Attacks on CLIP Pre-training", "abstract": "The Contrastive Language-Image Pretraining (CLIP) model has significantly advanced vision-language modeling by aligning image-text pairs from large-scale web data through self-supervised contrastive learning. Yet, its reliance on uncurated Internet-sourced data exposes it to data poisoning and backdoor risks. While existing studies primarily investigate image-based attacks, the text modality, which is equally central to CLIP's training, remains underexplored. In this work, we introduce ToxicTextCLIP, a framework for generating high-quality adversarial texts that target CLIP during the pre-training phase. The framework addresses two key challenges: semantic misalignment caused by background inconsistency with the target class, and the scarcity of background-consistent texts. To this end, ToxicTextCLIP iteratively applies: 1) a background-aware selector that prioritizes texts with background content aligned to the target class, and 2) a background-driven augmenter that generates semantically coherent and diverse poisoned samples. Extensive experiments on classification and retrieval tasks show that ToxicTextCLIP achieves up to 95.83% poisoning success and 98.68% backdoor Hit@1, while bypassing RoCLIP, CleanCLIP and SafeCLIP defenses. The source code can be accessed via https://github.com/xinyaocse/ToxicTextCLIP/.", "date": "2025-11-01", "categories": []}
{"id": "2511.00831", "title": "Enhancing Adversarial Transferability in Visual-Language Pre-training Models via Local Shuffle and Sample-based Attack", "abstract": "Visual-Language Pre-training (VLP) models have achieved significant performance across various downstream tasks. However, they remain vulnerable to adversarial examples. While prior efforts focus on improving the adversarial transferability of multimodal adversarial examples through cross-modal interactions, these approaches suffer from overfitting issues, due to a lack of input diversity by relying excessively on information from adversarial examples in one modality when crafting attacks in another. To address this issue, we draw inspiration from strategies in some adversarial training methods and propose a novel attack called Local Shuffle and Sample-based Attack (LSSA). LSSA randomly shuffles one of the local image blocks, thus expanding the original image-text pairs, generating adversarial images, and sampling around them. Then, it utilizes both the original and sampled images to generate the adversarial texts. Extensive experiments on multiple models and datasets demonstrate that LSSA significantly enhances the transferability of multimodal adversarial examples across diverse VLP models and downstream tasks. Moreover, LSSA outperforms other advanced attacks on Large Vision-Language Models.", "date": "2025-11-02", "categories": []}
{"id": "2511.00836", "title": "Parameter Interpolation Adversarial Training for Robust Image Classification", "abstract": "Though deep neural networks exhibit superior performance on various tasks, they are still plagued by adversarial examples. Adversarial training has been demonstrated to be the most effective method to defend against adversarial attacks. However, existing adversarial training methods show that the model robustness has apparent oscillations and overfitting issues in the training process, degrading the defense efficacy. To address these issues, we propose a novel framework called Parameter Interpolation Adversarial Training (PIAT). PIAT tunes the model parameters between each epoch by interpolating the parameters of the previous and current epochs. It makes the decision boundary of model change more moderate and alleviates the overfitting issue, helping the model converge better and achieving higher model robustness. In addition, we suggest using the Normalized Mean Square Error (NMSE) to further improve the robustness by aligning the relative magnitude of logits between clean and adversarial examples rather than the absolute magnitude. Extensive experiments conducted on several benchmark datasets demonstrate that our framework could prominently improve the robustness of both Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs).", "date": "2025-11-02", "categories": []}
{"id": "2511.01172", "title": "Adapt under Attack and Domain Shift: Unified Adversarial Meta-Learning and Domain Adaptation for Robust Automatic Modulation Classification", "abstract": "Deep learning has emerged as a leading approach for Automatic Modulation Classification (AMC), demonstrating superior performance over traditional methods. However, vulnerability to adversarial attacks and susceptibility to data distribution shifts hinder their practical deployment in real-world, dynamic environments. To address these threats, we propose a novel, unified framework that integrates meta-learning with domain adaptation, making AMC systems resistant to both adversarial attacks and environmental changes. Our framework utilizes a two-phase strategy. First, in an offline phase, we employ a meta-learning approach to train the model on clean and adversarially perturbed samples from a single source domain. This method enables the model to generalize its defense, making it resistant to a combination of previously unseen attacks. Subsequently, in the online phase, we apply domain adaptation to align the model's features with a new target domain, allowing it to adapt without requiring substantial labeled data. As a result, our framework achieves a significant improvement in modulation classification accuracy against these combined threats, offering a critical solution to the deployment and operational challenges of modern AMC systems.", "date": "2025-11-03", "categories": []}
{"id": "2511.01240", "title": "Beyond Deceptive Flatness: Dual-Order Solution for Strengthening Adversarial Transferability", "abstract": "Transferable attacks generate adversarial examples on surrogate models to fool unknown victim models, posing real-world threats and growing research interest. Despite focusing on flat losses for transferable adversarial examples, recent studies still fall into suboptimal regions, especially the flat-yet-sharp areas, termed as deceptive flatness. In this paper, we introduce a novel black-box gradient-based transferable attack from a perspective of dual-order information. Specifically, we feasibly propose Adversarial Flatness (AF) to the deceptive flatness problem and a theoretical assurance for adversarial transferability. Based on this, using an efficient approximation of our objective, we instantiate our attack as Adversarial Flatness Attack (AFA), addressing the altered gradient sign issue. Additionally, to further improve the attack ability, we devise MonteCarlo Adversarial Sampling (MCAS) by enhancing the inner-loop sampling efficiency. The comprehensive results on ImageNet-compatible dataset demonstrate superiority over six baselines, generating adversarial examples in flatter regions and boosting transferability across model architectures. When tested on input transformation attacks or the Baidu Cloud API, our method outperforms baselines.", "date": "2025-11-03", "categories": []}
{"id": "2511.01377", "title": "Protecting the Neural Networks against FGSM Attack Using Machine Unlearning", "abstract": "Machine learning is a powerful tool for building predictive models. However, it is vulnerable to adversarial attacks. Fast Gradient Sign Method (FGSM) attacks are a common type of adversarial attack that adds small perturbations to input data to trick a model into misclassifying it. In response to these attacks, researchers have developed methods for \"unlearning\" these attacks, which involves retraining a model on the original data without the added perturbations. Machine unlearning is a technique that tries to \"forget\" specific data points from the training dataset, to improve the robustness of a machine learning model against adversarial attacks like FGSM. In this paper, we focus on applying unlearning techniques to the LeNet neural network, a popular architecture for image classification. We evaluate the efficacy of unlearning FGSM attacks on the LeNet network and find that it can significantly improve its robustness against these types of attacks.", "date": "2025-11-03", "categories": []}
{"id": "2306.09977", "title": "Adversarially robust clustering with optimality guarantees", "abstract": "We consider the problem of clustering data points coming from sub-Gaussian mixtures. Existing methods that provably achieve the optimal mislabeling error, such as the Lloyd algorithm, are usually vulnerable to outliers. In contrast, clustering methods seemingly robust to adversarial perturbations are not known to satisfy the optimal statistical guarantees. We propose a simple robust algorithm based on the coordinatewise median that obtains the optimal mislabeling rate even when we allow adversarial outliers to be present. Our algorithm achieves the optimal error rate in constant iterations when a weak initialization condition is satisfied. In the absence of outliers, in fixed dimensions, our theoretical guarantees are similar to that of the Lloyd algorithm. Extensive experiments on various simulated and public datasets are conducted to support the theoretical guarantees of our method.", "date": "2023-06-16", "categories": []}
{"id": "2502.21041", "title": "Fast Adversarial Training against Sparse Attacks Requires Loss Smoothing", "abstract": "This paper studies fast adversarial training against sparse adversarial perturbations bounded by $l_0$ norm. We demonstrate the challenges of employing $1$-step attacks on $l_0$ bounded perturbations for fast adversarial training, including degraded performance and the occurrence of catastrophic overfitting (CO). We highlight that CO in $l_0$ adversarial training is caused by sub-optimal perturbation locations of $1$-step attack. Theoretical and empirical analyses reveal that the loss landscape of $l_0$ adversarial training is more craggy compared to its $l_\\infty$, $l_2$ and $l_1$ counterparts. Moreover, we corroborate that the craggy loss landscape can aggravate CO. To address these issues, we propose Fast-LS-$l_0$ that incorporates soft labels and the trade-off loss function to smooth the adversarial loss landscape. Extensive experiments demonstrate our method can overcome the challenge of catastrophic overfitting, achieve state-of-the-art performance, and narrow down the performance gap between $1$-step and multi-step adversarial training against sparse attacks.", "date": "2025-02-28", "categories": []}
{"id": "2510.26833", "title": "VISAT: Benchmarking Adversarial and Distribution Shift Robustness in Traffic Sign Recognition with Visual Attributes", "abstract": "We present VISAT, a novel open dataset and benchmarking suite for evaluating model robustness in the task of traffic sign recognition with the presence of visual attributes. Built upon the Mapillary Traffic Sign Dataset (MTSD), our dataset introduces two benchmarks that respectively emphasize robustness against adversarial attacks and distribution shifts. For our adversarial attack benchmark, we employ the state-of-the-art Projected Gradient Descent (PGD) method to generate adversarial inputs and evaluate their impact on popular models. Additionally, we investigate the effect of adversarial attacks on attribute-specific multi-task learning (MTL) networks, revealing spurious correlations among MTL tasks. The MTL networks leverage visual attributes (color, shape, symbol, and text) that we have created for each traffic sign in our dataset. For our distribution shift benchmark, we utilize ImageNet-C's realistic data corruption and natural variation techniques to perform evaluations on the robustness of both base and MTL models. Moreover, we further explore spurious correlations among MTL tasks through synthetic alterations of traffic sign colors using color quantization techniques. Our experiments focus on two major backbones, ResNet-152 and ViT-B/32, and compare the performance between base and MTL models. The VISAT dataset and benchmarking framework contribute to the understanding of model robustness for traffic sign recognition, shedding light on the challenges posed by adversarial attacks and distribution shifts. We believe this work will facilitate advancements in developing more robust models for real-world applications in autonomous driving and cyber-physical systems.", "date": "2025-10-29", "categories": []}
{"id": "2510.26981", "title": "Fine-Grained Iterative Adversarial Attacks with Limited Computation Budget", "abstract": "This work tackles a critical challenge in AI safety research under limited compute: given a fixed computation budget, how can one maximize the strength of iterative adversarial attacks? Coarsely reducing the number of attack iterations lowers cost but substantially weakens effectiveness. To fulfill the attainable attack efficacy within a constrained budget, we propose a fine-grained control mechanism that selectively recomputes layer activations across both iteration-wise and layer-wise levels. Extensive experiments show that our method consistently outperforms existing baselines at equal cost. Moreover, when integrated into adversarial training, it attains comparable performance with only 30% of the original budget.", "date": "2025-10-30", "categories": []}
{"id": "2510.27245", "title": "Trans-defense: Transformer-based Denoiser for Adversarial Defense with Spatial-Frequency Domain Representation", "abstract": "In recent times, deep neural networks (DNNs) have been successfully adopted for various applications. Despite their notable achievements, it has become evident that DNNs are vulnerable to sophisticated adversarial attacks, restricting their applications in security-critical systems. In this paper, we present two-phase training methods to tackle the attack: first, training the denoising network, and second, the deep classifier model. We propose a novel denoising strategy that integrates both spatial and frequency domain approaches to defend against adversarial attacks on images. Our analysis reveals that high-frequency components of attacked images are more severely corrupted compared to their lower-frequency counterparts. To address this, we leverage Discrete Wavelet Transform (DWT) for frequency analysis and develop a denoising network that combines spatial image features with wavelets through a transformer layer. Next, we retrain the classifier using the denoised images, which enhances the classifier's robustness against adversarial attacks. Experimental results across the MNIST, CIFAR-10, and Fashion-MNIST datasets reveal that the proposed method remarkably elevates classification accuracy, substantially exceeding the performance by utilizing a denoising network and adversarial training approaches. The code is available at https://github.com/Mayank94/Trans-Defense.", "date": "2025-10-31", "categories": []}
{"id": "2510.27599", "title": "ANCHOR: Integrating Adversarial Training with Hard-mined Supervised Contrastive Learning for Robust Representation Learning", "abstract": "Neural networks have changed the way machines interpret the world. At their core, they learn by following gradients, adjusting their parameters step by step until they identify the most discriminant patterns in the data. This process gives them their strength, yet it also opens the door to a hidden flaw. The very gradients that help a model learn can also be used to produce small, imperceptible tweaks that cause the model to completely alter its decision. Such tweaks are called adversarial attacks. These attacks exploit this vulnerability by adding tiny, imperceptible changes to images that, while leaving them identical to the human eye, cause the model to make wrong predictions. In this work, we propose Adversarially-trained Contrastive Hard-mining for Optimized Robustness (ANCHOR), a framework that leverages the power of supervised contrastive learning with explicit hard positive mining to enable the model to learn representations for images such that the embeddings for the images, their augmentations, and their perturbed versions cluster together in the embedding space along with those for other images of the same class while being separated from images of other classes. This alignment helps the model focus on stable, meaningful patterns rather than fragile gradient cues. On CIFAR-10, our approach achieves impressive results for both clean and robust accuracy under PGD-20 (epsilon = 0.031), outperforming standard adversarial training methods. Our results indicate that combining adversarial guidance with hard-mined contrastive supervision helps models learn more structured and robust representations, narrowing the gap between accuracy and robustness.", "date": "2025-10-31", "categories": []}
{"id": "2311.17434", "title": "GSE: Group-wise Sparse and Explainable Adversarial Attacks", "abstract": "Sparse adversarial attacks fool deep neural networks (DNNs) through minimal pixel perturbations, often regularized by the $\\ell_0$ norm. Recent efforts have replaced this norm with a structural sparsity regularizer, such as the nuclear group norm, to craft group-wise sparse adversarial attacks. The resulting perturbations are thus explainable and hold significant practical relevance, shedding light on an even greater vulnerability of DNNs. However, crafting such attacks poses an optimization challenge, as it involves computing norms for groups of pixels within a non-convex objective. We address this by presenting a two-phase algorithm that generates group-wise sparse attacks within semantically meaningful areas of an image. Initially, we optimize a quasinorm adversarial loss using the $1/2-$quasinorm proximal operator tailored for non-convex programming. Subsequently, the algorithm transitions to a projected Nesterov's accelerated gradient descent with $2-$norm regularization applied to perturbation magnitudes. Rigorous evaluations on CIFAR-10 and ImageNet datasets demonstrate a remarkable increase in group-wise sparsity, e.g., $50.9\\%$ on CIFAR-10 and $38.4\\%$ on ImageNet (average case, targeted attack). This performance improvement is accompanied by significantly faster computation times, improved explainability, and a $100\\%$ attack success rate.", "date": "2023-11-29", "categories": []}
{"id": "2406.15863", "title": "EmoAttack: Emotion-to-Image Diffusion Models for Emotional Backdoor Generation", "abstract": "Text-to-image diffusion models can generate realistic images based on textual inputs, enabling users to convey their opinions visually through language. Meanwhile, within language, emotion plays a crucial role in expressing personal opinions in our daily lives and the inclusion of maliciously negative content can lead users astray, exacerbating negative emotions. Recognizing the success of diffusion models and the significance of emotion, we investigate a previously overlooked risk associated with text-to-image diffusion models, that is, utilizing emotion in the input texts to introduce negative content and provoke unfavorable emotions in users. Specifically, we identify a new backdoor attack, i.e., emotion-aware backdoor attack (EmoAttack), which introduces malicious negative content triggered by emotional texts during image generation. We formulate such an attack as a diffusion personalization problem to avoid extensive model retraining and propose the EmoBooth. Unlike existing personalization methods, our approach fine-tunes a pre-trained diffusion model by establishing a mapping between a cluster of emotional words and a given reference image containing malicious negative content. To validate the effectiveness of our method, we built a dataset and conducted extensive analysis and discussion about its effectiveness. Given consumers' widespread use of diffusion models, uncovering this threat is critical for society.", "date": "2024-06-22", "categories": []}
{"id": "2501.05783", "title": "UV-Attack: Physical-World Adversarial Attacks for Person Detection via Dynamic-NeRF-based UV Mapping", "abstract": "In recent research, adversarial attacks on person detectors using patches or static 3D model-based texture modifications have struggled with low success rates due to the flexible nature of human movement. Modeling the 3D deformations caused by various actions has been a major challenge. Fortunately, advancements in Neural Radiance Fields (NeRF) for dynamic human modeling offer new possibilities. In this paper, we introduce UV-Attack, a groundbreaking approach that achieves high success rates even with extensive and unseen human actions. We address the challenge above by leveraging dynamic-NeRF-based UV mapping. UV-Attack can generate human images across diverse actions and viewpoints, and even create novel actions by sampling from the SMPL parameter space. While dynamic NeRF models are capable of modeling human bodies, modifying clothing textures is challenging because they are embedded in neural network parameters. To tackle this, UV-Attack generates UV maps instead of RGB images and modifies the texture stacks. This approach enables real-time texture edits and makes the attack more practical. We also propose a novel Expectation over Pose Transformation loss (EoPT) to improve the evasion success rate on unseen poses and views. Our experiments show that UV-Attack achieves a 92.7% attack success rate against the FastRCNN model across varied poses in dynamic video settings, significantly outperforming the state-of-the-art AdvCamou attack, which only had a 28.5% ASR. Moreover, we achieve 49.5% ASR on the latest YOLOv8 detector in black-box settings. This work highlights the potential of dynamic NeRF-based UV mapping for creating more effective adversarial attacks on person detectors, addressing key challenges in modeling human movement and texture modification. The code is available at https://github.com/PolyLiYJ/UV-Attack.", "date": "2025-01-10", "categories": []}
{"id": "2505.18766", "title": "StyleGuard: Preventing Text-to-Image-Model-based Style Mimicry Attacks by Style Perturbations", "abstract": "Recently, text-to-image diffusion models have been widely used for style mimicry and personalized customization through methods such as DreamBooth and Textual Inversion. This has raised concerns about intellectual property protection and the generation of deceptive content. Recent studies, such as Glaze and Anti-DreamBooth, have proposed using adversarial noise to protect images from these attacks. However, recent purification-based methods, such as DiffPure and Noise Upscaling, have successfully attacked these latest defenses, showing the vulnerabilities of these methods. Moreover, present methods show limited transferability across models, making them less effective against unknown text-to-image models. To address these issues, we propose a novel anti-mimicry method, StyleGuard. We propose a novel style loss that optimizes the style-related features in the latent space to make it deviate from the original image, which improves model-agnostic transferability. Additionally, to enhance the perturbation's ability to bypass diffusion-based purification, we designed a novel upscale loss that involves ensemble purifiers and upscalers during training. Extensive experiments on the WikiArt and CelebA datasets demonstrate that StyleGuard outperforms existing methods in robustness against various transformations and purifications, effectively countering style mimicry in various models. Moreover, StyleGuard is effective on different style mimicry methods, including DreamBooth and Textual Inversion. The code is available at https://github.com/PolyLiYJ/StyleGuard.", "date": "2025-05-24", "categories": []}
{"id": "2506.07001", "title": "Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text", "abstract": "The increasing capabilities of Large Language Models (LLMs) have raised concerns about their misuse in AI-generated plagiarism and social engineering. While various AI-generated text detectors have been proposed to mitigate these risks, many remain vulnerable to simple evasion techniques such as paraphrasing. However, recent detectors have shown greater robustness against such basic attacks. In this work, we introduce Adversarial Paraphrasing, a training-free attack framework that universally humanizes any AI-generated text to evade detection more effectively. Our approach leverages an off-the-shelf instruction-following LLM to paraphrase AI-generated content under the guidance of an AI text detector, producing adversarial examples that are specifically optimized to bypass detection. Extensive experiments show that our attack is both broadly effective and highly transferable across several detection systems. For instance, compared to simple paraphrasing attack--which, ironically, increases the true positive at 1% false positive (T@1%F) by 8.57% on RADAR and 15.03% on Fast-DetectGPT--adversarial paraphrasing, guided by OpenAI-RoBERTa-Large, reduces T@1%F by 64.49% on RADAR and a striking 98.96% on Fast-DetectGPT. Across a diverse set of detectors--including neural network-based, watermark-based, and zero-shot approaches--our attack achieves an average T@1%F reduction of 87.88% under the guidance of OpenAI-RoBERTa-Large. We also analyze the tradeoff between text quality and attack success to find that our method can significantly reduce detection rates, with mostly a slight degradation in text quality. Our adversarial setup highlights the need for more robust and resilient detection strategies in the light of increasingly sophisticated evasion techniques.", "date": "2025-06-08", "categories": []}
{"id": "2506.08188", "title": "GradEscape: A Gradient-Based Evader Against AI-Generated Text Detectors", "abstract": "In this paper, we introduce GradEscape, the first gradient-based evader designed to attack AI-generated text (AIGT) detectors. GradEscape overcomes the undifferentiable computation problem, caused by the discrete nature of text, by introducing a novel approach to construct weighted embeddings for the detector input. It then updates the evader model parameters using feedback from victim detectors, achieving high attack success with minimal text modification. To address the issue of tokenizer mismatch between the evader and the detector, we introduce a warm-started evader method, enabling GradEscape to adapt to detectors across any language model architecture. Moreover, we employ novel tokenizer inference and model extraction techniques, facilitating effective evasion even in query-only access.\n  We evaluate GradEscape on four datasets and three widely-used language models, benchmarking it against four state-of-the-art AIGT evaders. Experimental results demonstrate that GradEscape outperforms existing evaders in various scenarios, including with an 11B paraphrase model, while utilizing only 139M parameters. We have successfully applied GradEscape to two real-world commercial AIGT detectors. Our analysis reveals that the primary vulnerability stems from disparity in text expression styles within the training data. We also propose a potential defense strategy to mitigate the threat of AIGT evaders. We open-source our GradEscape for developing more robust AIGT detectors.", "date": "2025-06-09", "categories": []}
{"id": "2506.21046", "title": "Boosting Generative Adversarial Transferability with Self-supervised Vision Transformer Features", "abstract": "The ability of deep neural networks (DNNs) come from extracting and interpreting features from the data provided. By exploiting intermediate features in DNNs instead of relying on hard labels, we craft adversarial perturbation that generalize more effectively, boosting black-box transferability. These features ubiquitously come from supervised learning in previous work. Inspired by the exceptional synergy between self-supervised learning and the Transformer architecture, this paper explores whether exploiting self-supervised Vision Transformer (ViT) representations can improve adversarial transferability. We present dSVA -- a generative dual self-supervised ViT features attack, that exploits both global structural features from contrastive learning (CL) and local textural features from masked image modeling (MIM), the self-supervised learning paradigm duo for ViTs. We design a novel generative training framework that incorporates a generator to create black-box adversarial examples, and strategies to train the generator by exploiting joint features and the attention mechanism of self-supervised ViTs. Our findings show that CL and MIM enable ViTs to attend to distinct feature tendencies, which, when exploited in tandem, boast great adversarial generalizability. By disrupting dual deep features distilled by self-supervised ViTs, we are rewarded with remarkable black-box transferability to models of various architectures that outperform state-of-the-arts. Code available at https://github.com/spencerwooo/dSVA.", "date": "2025-06-26", "categories": []}
{"id": "2507.06078", "title": "ScoreAdv: Score-based Targeted Generation of Natural Adversarial Examples via Diffusion Models", "abstract": "Despite the success of deep learning across various domains, it remains vulnerable to adversarial attacks. Although many existing adversarial attack methods achieve high success rates, they typically rely on $\\ell_{p}$-norm perturbation constraints, which do not align with human perceptual capabilities. Consequently, researchers have shifted their focus toward generating natural, unrestricted adversarial examples (UAEs). GAN-based approaches suffer from inherent limitations, such as poor image quality due to instability and mode collapse. Meanwhile, diffusion models have been employed for UAE generation, but they still rely on iterative PGD perturbation injection, without fully leveraging their central denoising capabilities. In this paper, we introduce a novel approach for generating UAEs based on diffusion models, named ScoreAdv. This method incorporates an interpretable adversarial guidance mechanism to gradually shift the sampling distribution towards the adversarial distribution, while using an interpretable saliency map to inject the visual information of a reference image into the generated samples. Notably, our method is capable of generating an unlimited number of natural adversarial examples and can attack not only classification models but also retrieval models. We conduct extensive experiments on ImageNet and CelebA datasets, validating the performance of ScoreAdv across ten target models in both black-box and white-box settings. Our results demonstrate that ScoreAdv achieves state-of-the-art attack success rates and image quality, while maintaining inference efficiency. Furthermore, the dynamic balance between denoising and adversarial perturbation enables ScoreAdv to remain robust even under defensive measures.", "date": "2025-07-08", "categories": []}
{"id": "2510.08604", "title": "LatentBreak: Jailbreaking Large Language Models through Latent Space Feedback", "abstract": "Jailbreaks are adversarial attacks designed to bypass the built-in safety mechanisms of large language models. Automated jailbreaks typically optimize an adversarial suffix or adapt long prompt templates by forcing the model to generate the initial part of a restricted or harmful response. In this work, we show that existing jailbreak attacks that leverage such mechanisms to unlock the model response can be detected by a straightforward perplexity-based filtering on the input prompt. To overcome this issue, we propose LatentBreak, a white-box jailbreak attack that generates natural adversarial prompts with low perplexity capable of evading such defenses. LatentBreak substitutes words in the input prompt with semantically-equivalent ones, preserving the initial intent of the prompt, instead of adding high-perplexity adversarial suffixes or long templates. These words are chosen by minimizing the distance in the latent space between the representation of the adversarial prompt and that of harmless requests. Our extensive evaluation shows that LatentBreak leads to shorter and low-perplexity prompts, thus outperforming competing jailbreak algorithms against perplexity-based filters on multiple safety-aligned models.", "date": "2025-10-07", "categories": []}
{"id": "2510.26096", "title": "ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio-Language Models", "abstract": "Recent advances in Audio-Language Models (ALMs) have significantly improved multimodal understanding capabilities. However, the introduction of the audio modality also brings new and unique vulnerability vectors. Previous studies have proposed jailbreak attacks that specifically target ALMs, revealing that defenses directly transferred from traditional audio adversarial attacks or text-based Large Language Model (LLM) jailbreaks are largely ineffective against these ALM-specific threats. To address this issue, we propose ALMGuard, the first defense framework tailored to ALMs. Based on the assumption that safety-aligned shortcuts naturally exist in ALMs, we design a method to identify universal Shortcut Activation Perturbations (SAPs) that serve as triggers that activate the safety shortcuts to safeguard ALMs at inference time. To better sift out effective triggers while preserving the model's utility on benign tasks, we further propose Mel-Gradient Sparse Mask (M-GSM), which restricts perturbations to Mel-frequency bins that are sensitive to jailbreaks but insensitive to speech understanding. Both theoretical analyses and empirical results demonstrate the robustness of our method against both seen and unseen attacks. Overall, \\MethodName reduces the average success rate of advanced ALM-specific jailbreak attacks to 4.6% across four models, while maintaining comparable utility on benign benchmarks, establishing it as the new state of the art. Our code and data are available at https://github.com/WeifeiJin/ALMGuard.", "date": "2025-10-30", "categories": []}
{"id": "2510.26105", "title": "Security Risk of Misalignment between Text and Image in Multi-modal Model", "abstract": "Despite the notable advancements and versatility of multi-modal diffusion models, such as text-to-image models, their susceptibility to adversarial inputs remains underexplored. Contrary to expectations, our investigations reveal that the alignment between textual and Image modalities in existing diffusion models is inadequate. This misalignment presents significant risks, especially in the generation of inappropriate or Not-Safe-For-Work (NSFW) content. To this end, we propose a novel attack called Prompt-Restricted Multi-modal Attack (PReMA) to manipulate the generated content by modifying the input image in conjunction with any specified prompt, without altering the prompt itself. PReMA is the first attack that manipulates model outputs by solely creating adversarial images, distinguishing itself from prior methods that primarily generate adversarial prompts to produce NSFW content. Consequently, PReMA poses a novel threat to the integrity of multi-modal diffusion models, particularly in image-editing applications that operate with fixed prompts. Comprehensive evaluations conducted on image inpainting and style transfer tasks across various models confirm the potent efficacy of PReMA.", "date": "2025-10-30", "categories": []}
{"id": "2506.12913", "title": "Jailbreak Transferability Emerges from Shared Representations", "abstract": "Jailbreak transferability is the surprising phenomenon when an adversarial attack compromising one model also elicits harmful responses from other models. Despite widespread demonstrations, there is little consensus on why transfer is possible: is it a quirk of safety training, an artifact of model families, or a more fundamental property of representation learning? We present evidence that transferability emerges from shared representations rather than incidental flaws. Across 20 open-weight models and 33 jailbreak attacks, we find two factors that systematically shape transfer: (1) representational similarity under benign prompts, and (2) the strength of the jailbreak on the source model. To move beyond correlation, we show that deliberately increasing similarity through benign only distillation causally increases transfer. Our qualitative analyses reveal systematic transferability patterns across different types of jailbreaks. For example, persona-style jailbreaks transfer far more often than cipher-based prompts, consistent with the idea that natural-language attacks exploit models' shared representation space, whereas cipher-based attacks rely on idiosyncratic quirks that do not generalize. Together, these results reframe jailbreak transfer as a consequence of representation alignment rather than a fragile byproduct of safety training.", "date": "2025-06-15", "categories": []}
{"id": "2505.16947", "title": "MixAT: Combining Continuous and Discrete Adversarial Training for LLMs", "abstract": "Despite recent efforts in Large Language Model (LLM) safety and alignment, current adversarial attacks on frontier LLMs can still consistently force harmful generations. Although adversarial training has been widely studied and shown to significantly improve the robustness of traditional machine learning models, its strengths and weaknesses in the context of LLMs are less understood. Specifically, while existing discrete adversarial attacks are effective at producing harmful content, training LLMs with concrete adversarial prompts is often computationally expensive, leading to reliance on continuous relaxations. At the same time, despite their effectiveness and generalization capabilities, training with continuous perturbations does not always capture the full spectrum of vulnerabilities exploited by discrete attacks. In this work, we aim to bridge this gap by introducing MixAT, a novel method that combines stronger discrete and faster continuous attacks during training. We rigorously evaluate MixAT across a wide spectrum of state-of-the-art attacks, proposing the At Least One Attack Success Rate (ALO-ASR) metric to capture the worst-case vulnerability of models. We show MixAT achieves substantially better robustness (ALO-ASR < 20%) compared to prior defenses (ALO-ASR > 50%), while maintaining a runtime comparable to methods based on continuous relaxations. We further analyze MixAT in realistic deployment settings, exploring how chat templates, quantization, low-rank adapters, and temperature affect both adversarial training and evaluation, revealing additional blind spots in current methodologies. Our results demonstrate that MixAT's discrete-continuous defense offers a principled and superior robustness-accuracy tradeoff with minimal computational overhead, highlighting its promise for building safer LLMs. We provide our code and models at https://github.com/insait-institute/MixAT.", "date": "2025-05-22", "categories": []}
{"id": "2510.23622", "title": "Adversarially-Aware Architecture Design for Robust Medical AI Systems", "abstract": "Adversarial attacks pose a severe risk to AI systems used in healthcare, capable of misleading models into dangerous misclassifications that can delay treatments or cause misdiagnoses. These attacks, often imperceptible to human perception, threaten patient safety, particularly in underserved populations. Our study explores these vulnerabilities through empirical experimentation on a dermatological dataset, where adversarial methods significantly reduce classification accuracy. Through detailed threat modeling, experimental benchmarking, and model evaluation, we demonstrate both the severity of the threat and the partial success of defenses like adversarial training and distillation. Our results show that while defenses reduce attack success rates, they must be balanced against model performance on clean data. We conclude with a call for integrated technical, ethical, and policy-based approaches to build more resilient, equitable AI in healthcare.", "date": "2025-10-23", "categories": []}
{"id": "2510.24034", "title": "AutoPrompt: Automated Red-Teaming of Text-to-Image Models via LLM-Driven Adversarial Prompts", "abstract": "Despite rapid advancements in text-to-image (T2I) models, their safety mechanisms are vulnerable to adversarial prompts, which maliciously generate unsafe images. Current red-teaming methods for proactively assessing such vulnerabilities usually require white-box access to T2I models, and rely on inefficient per-prompt optimization, as well as inevitably generate semantically meaningless prompts easily blocked by filters. In this paper, we propose APT (AutoPrompT), a black-box framework that leverages large language models (LLMs) to automatically generate human-readable adversarial suffixes for benign prompts. We first introduce an alternating optimization-finetuning pipeline between adversarial suffix optimization and fine-tuning the LLM utilizing the optimized suffix. Furthermore, we integrates a dual-evasion strategy in optimization phase, enabling the bypass of both perplexity-based filter and blacklist word filter: (1) we constrain the LLM generating human-readable prompts through an auxiliary LLM perplexity scoring, which starkly contrasts with prior token-level gibberish, and (2) we also introduce banned-token penalties to suppress the explicit generation of banned-tokens in blacklist. Extensive experiments demonstrate the excellent red-teaming performance of our human-readable, filter-resistant adversarial prompts, as well as superior zero-shot transferability which enables instant adaptation to unseen prompts and exposes critical vulnerabilities even in commercial APIs (e.g., Leonardo.Ai.).", "date": "2025-10-28", "categories": []}
{"id": "2510.24195", "title": "Vanish into Thin Air: Cross-prompt Universal Adversarial Attacks for SAM2", "abstract": "Recent studies reveal the vulnerability of the image segmentation foundation model SAM to adversarial examples. Its successor, SAM2, has attracted significant attention due to its strong generalization capability in video segmentation. However, its robustness remains unexplored, and it is unclear whether existing attacks on SAM can be directly transferred to SAM2. In this paper, we first analyze the performance gap of existing attacks between SAM and SAM2 and highlight two key challenges arising from their architectural differences: directional guidance from the prompt and semantic entanglement across consecutive frames. To address these issues, we propose UAP-SAM2, the first cross-prompt universal adversarial attack against SAM2 driven by dual semantic deviation. For cross-prompt transferability, we begin by designing a target-scanning strategy that divides each frame into k regions, each randomly assigned a prompt, to reduce prompt dependency during optimization. For effectiveness, we design a dual semantic deviation framework that optimizes a UAP by distorting the semantics within the current frame and disrupting the semantic consistency across consecutive frames. Extensive experiments on six datasets across two segmentation tasks demonstrate the effectiveness of the proposed method for SAM2. The comparative results show that UAP-SAM2 significantly outperforms state-of-the-art (SOTA) attacks by a large margin.", "date": "2025-10-28", "categories": []}
{"id": "2311.09671", "title": "Generalization Bounds for Robust Contrastive Learning: From Theory to Practice", "abstract": "Contrastive Learning first extracts features from unlabeled data, followed by linear probing with labeled data. Adversarial Contrastive Learning (ACL) integrates Adversarial Training into the first phase to enhance feature robustness against attacks in the probing phase. While ACL has shown strong empirical results, its theoretical understanding remains limited. Furthermore, while a fair amount of theoretical works analyze how the unsupervised loss can support the supervised loss in the probing phase, none has examined its role to the robust supervised loss. To fill this gap, our work develops rigorous theories to identify which components in the unsupervised training can help improve the robust supervised loss. Specifically, besides the adversarial contrastive loss, we reveal that the benign one, along with a global divergence between benign and adversarial examples can also improve robustness. Proper experiments are conducted to justify our findings.", "date": "2023-11-16", "categories": []}
{"id": "2502.02096", "title": "Dual-Flow: Transferable Multi-Target, Instance-Agnostic Attacks via In-the-wild Cascading Flow Optimization", "abstract": "Adversarial attacks are widely used to evaluate model robustness, and in black-box scenarios, the transferability of these attacks becomes crucial. Existing generator-based attacks have excellent generalization and transferability due to their instance-agnostic nature. However, when training generators for multi-target tasks, the success rate of transfer attacks is relatively low due to the limitations of the model's capacity. To address these challenges, we propose a novel Dual-Flow framework for multi-target instance-agnostic adversarial attacks, utilizing Cascading Distribution Shift Training to develop an adversarial velocity function. Extensive experiments demonstrate that Dual-Flow significantly improves transferability over previous multi-target generative attacks. For example, it increases the success rate from Inception-v3 to ResNet-152 by 34.58\\%. Furthermore, our attack method shows substantially stronger robustness against defense mechanisms, such as adversarially trained models. The code of Dual-Flow is available at: $\\href{https://github.com/Chyxx/Dual-Flow}{https://github.com/Chyxx/Dual-Flow}$.", "date": "2025-02-04", "categories": []}
{"id": "2502.09110", "title": "Pulling Back the Curtain: Unsupervised Adversarial Detection via Contrastive Auxiliary Networks", "abstract": "Deep learning models are widely employed in safety-critical applications yet remain susceptible to adversarial attacks -- imperceptible perturbations that can significantly degrade model performance. Conventional defense mechanisms predominantly focus on either enhancing model robustness or detecting adversarial inputs independently. In this work, we propose an Unsupervised adversarial detection via Contrastive Auxiliary Networks (U-CAN) to uncover adversarial behavior within auxiliary feature representations, without the need for adversarial examples. U-CAN is embedded within selected intermediate layers of the target model. These auxiliary networks, comprising projection layers and ArcFace-based linear layers, refine feature representations to more effectively distinguish between benign and adversarial inputs. Comprehensive experiments across multiple datasets (CIFAR-10, Mammals, and a subset of ImageNet) and architectures (ResNet-50, VGG-16, and ViT) demonstrate that our method surpasses existing unsupervised adversarial detection techniques, achieving superior F1 scores against four distinct attack methods. The proposed framework provides a scalable and effective solution for enhancing the security and reliability of deep learning systems.", "date": "2025-02-13", "categories": []}
{"id": "2503.10635", "title": "A Frustratingly Simple Yet Highly Effective Attack Baseline: Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5/4o/o1", "abstract": "Despite promising performance on open-source large vision-language models (LVLMs), transfer-based targeted attacks often fail against closed-source commercial LVLMs. Analyzing failed adversarial perturbations reveals that the learned perturbations typically originate from a uniform distribution and lack clear semantic details, resulting in unintended responses. This critical absence of semantic information leads commercial black-box LVLMs to either ignore the perturbation entirely or misinterpret its embedded semantics, thereby causing the attack to fail. To overcome these issues, we propose to refine semantic clarity by encoding explicit semantic details within local regions, thus ensuring the capture of finer-grained features and inter-model transferability, and by concentrating modifications on semantically rich areas rather than applying them uniformly. To achieve this, we propose a simple yet highly effective baseline: at each optimization step, the adversarial image is cropped randomly by a controlled aspect ratio and scale, resized, and then aligned with the target image in the embedding space. While the naive source-target matching method has been utilized before in the literature, we are the first to provide a tight analysis, which establishes a close connection between perturbation optimization and semantics. Experimental results confirm our hypothesis. Our adversarial examples crafted with local-aggregated perturbations focused on crucial regions exhibit surprisingly good transferability to commercial LVLMs, including GPT-4.5, GPT-4o, Gemini-2.0-flash, Claude-3.5/3.7-sonnet, and even reasoning models like o1, Claude-3.7-thinking and Gemini-2.0-flash-thinking. Our approach achieves success rates exceeding 90% on GPT-4.5, 4o, and o1, significantly outperforming all prior state-of-the-art attack methods with lower $\\ell_1/\\ell_2$ perturbations.", "date": "2025-03-13", "categories": []}
{"id": "2505.17356", "title": "Adversarial Robustness of Nonparametric Regression", "abstract": "In this paper, we investigate the adversarial robustness of nonparametric regression, a fundamental problem in machine learning, under the setting where an adversary can arbitrarily corrupt a subset of the input data. While the robustness of parametric regression has been extensively studied, its nonparametric counterpart remains largely unexplored. We characterize the adversarial robustness in nonparametric regression, assuming the regression function belongs to the second-order Sobolev space (i.e., it is square integrable up to its second derivative).\n  The contribution of this paper is two-fold: (i) we establish a minimax lower bound on the estimation error, revealing a fundamental limit that no estimator can overcome, and (ii) we show that, perhaps surprisingly, the classical smoothing spline estimator, when properly regularized, exhibits robustness against adversarial corruption. These results imply that if $o(n)$ out of $n$ samples are corrupted, the estimation error of the smoothing spline vanishes as $n \\to \\infty$. On the other hand, when a constant fraction of the data is corrupted, no estimator can guarantee vanishing estimation error, implying the optimality of the smoothing spline in terms of maximum tolerable number of corrupted samples.", "date": "2025-05-23", "categories": []}
{"id": "2506.10776", "title": "ME: Trigger Element Combination Backdoor Attack on Copyright Infringement", "abstract": "The capability of generative diffusion models (DMs) like Stable Diffusion (SD) in replicating training data could be taken advantage of by attackers to launch the Copyright Infringement Attack, with duplicated poisoned image-text pairs. SilentBadDiffusion (SBD) is a method proposed recently, which shew outstanding performance in attacking SD in text-to-image tasks. However, the feasible data resources in this area are still limited, some of them are even constrained or prohibited due to the issues like copyright ownership or inappropriate contents; And not all of the images in current datasets are suitable for the proposed attacking methods; Besides, the state-of-the-art (SoTA) performance of SBD is far from ideal when few generated poisoning samples could be adopted for attacks. In this paper, we raised new datasets accessible for researching in attacks like SBD, and proposed Multi-Element (ME) attack method based on SBD by increasing the number of poisonous visual-text elements per poisoned sample to enhance the ability of attacking, while importing Discrete Cosine Transform (DCT) for the poisoned samples to maintain the stealthiness. The Copyright Infringement Rate (CIR) / First Attack Epoch (FAE) we got on the two new datasets were 16.78% / 39.50 and 51.20% / 23.60, respectively close to or even outperformed benchmark Pokemon and Mijourney datasets. In condition of low subsampling ratio (5%, 6 poisoned samples), MESI and DCT earned CIR / FAE of 0.23% / 84.00 and 12.73% / 65.50, both better than original SBD, which failed to attack at all.", "date": "2025-06-12", "categories": []}
{"id": "2510.16923", "title": "UNDREAM: Bridging Differentiable Rendering and Photorealistic Simulation for End-to-end Adversarial Attacks", "abstract": "Deep learning models deployed in safety critical applications like autonomous driving use simulations to test their robustness against adversarial attacks in realistic conditions. However, these simulations are non-differentiable, forcing researchers to create attacks that do not integrate simulation environmental factors, reducing attack success. To address this limitation, we introduce UNDREAM, the first software framework that bridges the gap between photorealistic simulators and differentiable renderers to enable end-to-end optimization of adversarial perturbations on any 3D objects. UNDREAM enables manipulation of the environment by offering complete control over weather, lighting, backgrounds, camera angles, trajectories, and realistic human and object movements, thereby allowing the creation of diverse scenes. We showcase a wide array of distinct physically plausible adversarial objects that UNDREAM enables researchers to swiftly explore in different configurable environments. This combination of photorealistic simulation and differentiable optimization opens new avenues for advancing research of physical adversarial attacks.", "date": "2025-10-19", "categories": []}
{"id": "2510.22014", "title": "Toward Understanding the Transferability of Adversarial Suffixes in Large Language Models", "abstract": "Discrete optimization-based jailbreaking attacks on large language models aim to generate short, nonsensical suffixes that, when appended onto input prompts, elicit disallowed content. Notably, these suffixes are often transferable -- succeeding on prompts and models for which they were never optimized. And yet, despite the fact that transferability is surprising and empirically well-established, the field lacks a rigorous analysis of when and why transfer occurs. To fill this gap, we identify three statistical properties that strongly correlate with transfer success across numerous experimental settings: (1) how much a prompt without a suffix activates a model's internal refusal direction, (2) how strongly a suffix induces a push away from this direction, and (3) how large these shifts are in directions orthogonal to refusal. On the other hand, we find that prompt semantic similarity only weakly correlates with transfer success. These findings lead to a more fine-grained understanding of transferability, which we use in interventional experiments to showcase how our statistical analysis can translate into practical improvements in attack success.", "date": "2025-10-24", "categories": []}
{"id": "2510.22643", "title": "Enhancing Graph Classification Robustness with Singular Pooling", "abstract": "Graph Neural Networks (GNNs) have achieved strong performance across a range of graph representation learning tasks, yet their adversarial robustness in graph classification remains underexplored compared to node classification. While most existing defenses focus on the message-passing component, this work investigates the overlooked role of pooling operations in shaping robustness. We present a theoretical analysis of standard flat pooling methods (sum, average and max), deriving upper bounds on their adversarial risk and identifying their vulnerabilities under different attack scenarios and graph structures. Motivated by these insights, we propose \\textit{Robust Singular Pooling (RS-Pool)}, a novel pooling strategy that leverages the dominant singular vector of the node embedding matrix to construct a robust graph-level representation. We theoretically investigate the robustness of RS-Pool and interpret the resulting bound leading to improved understanding of our proposed pooling operator. While our analysis centers on Graph Convolutional Networks (GCNs), RS-Pool is model-agnostic and can be implemented efficiently via power iteration. Empirical results on real-world benchmarks show that RS-Pool provides better robustness than the considered pooling methods when subject to state-of-the-art adversarial attacks while maintaining competitive clean accuracy. Our code is publicly available at:\\href{https://github.com/king/rs-pool}{https://github.com/king/rs-pool}.", "date": "2025-10-26", "categories": []}
{"id": "2510.22652", "title": "If You Want to Be Robust, Be Wary of Initialization", "abstract": "Graph Neural Networks (GNNs) have demonstrated remarkable performance across a spectrum of graph-related tasks, however concerns persist regarding their vulnerability to adversarial perturbations. While prevailing defense strategies focus primarily on pre-processing techniques and adaptive message-passing schemes, this study delves into an under-explored dimension: the impact of weight initialization and associated hyper-parameters, such as training epochs, on a model's robustness. We introduce a theoretical framework bridging the connection between initialization strategies and a network's resilience to adversarial perturbations. Our analysis reveals a direct relationship between initial weights, number of training epochs and the model's vulnerability, offering new insights into adversarial robustness beyond conventional defense mechanisms. While our primary focus is on GNNs, we extend our theoretical framework, providing a general upper-bound applicable to Deep Neural Networks. Extensive experiments, spanning diverse models and real-world datasets subjected to various adversarial attacks, validate our findings. We illustrate that selecting appropriate initialization not only ensures performance on clean datasets but also enhances model robustness against adversarial perturbations, with observed gaps of up to 50\\% compared to alternative initialization approaches.", "date": "2025-10-26", "categories": []}
{"id": "2510.22981", "title": "Exploring Semantic-constrained Adversarial Example with Instruction Uncertainty Reduction", "abstract": "Recently, semantically constrained adversarial examples (SemanticAE), which are directly generated from natural language instructions, have become a promising avenue for future research due to their flexible attacking forms. To generate SemanticAEs, current methods fall short of satisfactory attacking ability as the key underlying factors of semantic uncertainty in human instructions, such as referring diversity, descriptive incompleteness, and boundary ambiguity, have not been fully investigated. To tackle the issues, this paper develops a multi-dimensional instruction uncertainty reduction (InSUR) framework to generate more satisfactory SemanticAE, i.e., transferable, adaptive, and effective. Specifically, in the dimension of the sampling method, we propose the residual-driven attacking direction stabilization to alleviate the unstable adversarial optimization caused by the diversity of language references. By coarsely predicting the language-guided sampling process, the optimization process will be stabilized by the designed ResAdv-DDIM sampler, therefore releasing the transferable and robust adversarial capability of multi-step diffusion models. In task modeling, we propose the context-encoded attacking scenario constraint to supplement the missing knowledge from incomplete human instructions. Guidance masking and renderer integration are proposed to regulate the constraints of 2D/3D SemanticAE, activating stronger scenario-adapted attacks. Moreover, in the dimension of generator evaluation, we propose the semantic-abstracted attacking evaluation enhancement by clarifying the evaluation boundary, facilitating the development of more effective SemanticAE generators. Extensive experiments demonstrate the superiority of the transfer attack performance of InSUR. Moreover, we realize the reference-free generation of semantically constrained 3D adversarial examples for the first time.", "date": "2025-10-27", "categories": []}
{"id": "2307.01225", "title": "Alert-ME: An Explainability-Driven Defense Against Adversarial Examples in Transformer-Based Text Classification", "abstract": "Transformer-based text classifiers such as BERT, RoBERTa, T5, and GPT have shown strong performance in natural language processing tasks but remain vulnerable to adversarial examples. These vulnerabilities raise significant security concerns, as small input perturbations can cause severe misclassifications. Existing robustness methods often require heavy computation or lack interpretability. This paper presents a unified framework called Explainability-driven Detection, Identification, and Transformation (EDIT) to strengthen inference-time defenses. EDIT integrates explainability tools, including attention maps and integrated gradients, with frequency-based features to automatically detect and identify adversarial perturbations while offering insight into model behavior. After detection, EDIT refines adversarial inputs using an optimal transformation process that leverages pre-trained embeddings and model feedback to replace corrupted tokens. To enhance security assurance, EDIT incorporates automated alerting mechanisms that involve human analysts when necessary.\n  Beyond static defenses, EDIT also provides adaptive resilience by enforcing internal feature similarity and transforming inputs, thereby disrupting the attackers optimization process and limiting the effectiveness of adaptive adversarial attacks. Experiments using BERT and RoBERTa on IMDB, YELP, AGNEWS, and SST2 datasets against seven word substitution attacks demonstrate that EDIT achieves an average Fscore of 89.69 percent and balanced accuracy of 89.70 percent. Compared to four state-of-the-art defenses, EDIT improves balanced accuracy by 1.22 times and F1-score by 1.33 times while being 83 times faster in feature extraction. The framework provides robust, interpretable, and efficient protection against both standard, zero-day, and adaptive adversarial threats in text classification models.", "date": "2023-07-03", "categories": []}
{"id": "2412.08053", "title": "DynamicPAE: Generating Scene-Aware Physical Adversarial Examples in Real-Time", "abstract": "Physical adversarial examples (PAEs) are regarded as whistle-blowers of real-world risks in deep-learning applications, thus worth further investigation. However, current PAE generation studies show limited adaptive attacking ability to diverse and varying scenes, revealing the urgent requirement of dynamic PAEs that are generated in real time and conditioned on the observation from the attacker. The key challenge in generating dynamic PAEs is learning the sparse relation between PAEs and the observation of attackers under the noisy feedback of attack training. To address the challenge, we present DynamicPAE, the first generative framework that enables scene-aware real-time physical attacks. Specifically, to address the noisy feedback problem that obfuscates the exploration of scene-related PAEs, we introduce the residual-guided adversarial pattern exploration technique. Residual-guided training, which relaxes the attack training with a reconstruction task, is proposed to enrich the feedback information, thereby achieving a more comprehensive exploration of PAEs. To address the alignment problem between the trained generator and the real-world scenario, we introduce the distribution-matched attack scenario alignment, consisting of the conditional-uncertainty-aligned data module and the skewness-aligned objective re-weighting module. The former aligns the training environment with the incomplete observation of the real-world attacker. The latter facilitates consistent stealth control across different attack targets with the skewness controller. Extensive digital and physical evaluations demonstrate the superior attack performance of DynamicPAE, attaining a 2.07 $\\times$ boost (58.8% average AP drop under attack) on representative object detectors (e.g., DETR) over state-of-the-art static PAE generating methods. Overall, our work opens the door to end-to-end modeling of dynamic PAEs.", "date": "2024-12-11", "categories": []}
{"id": "2501.01015", "title": "Boosting Adversarial Transferability with Spatial Adversarial Alignment", "abstract": "Deep neural networks are vulnerable to adversarial examples that exhibit transferability across various models. Numerous approaches are proposed to enhance the transferability of adversarial examples, including advanced optimization, data augmentation, and model modifications. However, these methods still show limited transferability, particularly in cross-architecture scenarios, such as from CNN to ViT. To achieve high transferability, we propose a technique termed Spatial Adversarial Alignment (SAA), which employs an alignment loss and leverages a witness model to fine-tune the surrogate model. Specifically, SAA consists of two key parts: spatial-aware alignment and adversarial-aware alignment. First, we minimize the divergences of features between the two models in both global and local regions, facilitating spatial alignment. Second, we introduce a self-adversarial strategy that leverages adversarial examples to impose further constraints, aligning features from an adversarial perspective. Through this alignment, the surrogate model is trained to concentrate on the common features extracted by the witness model. This facilitates adversarial attacks on these shared features, thereby yielding perturbations that exhibit enhanced transferability. Extensive experiments on various architectures on ImageNet show that aligned surrogate models based on SAA can provide higher transferable adversarial examples, especially in cross-architecture attacks.", "date": "2025-01-02", "categories": []}
{"id": "2506.09237", "title": "PatchGuard: Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies", "abstract": "Anomaly Detection (AD) and Anomaly Localization (AL) are crucial in fields that demand high reliability, such as medical imaging and industrial monitoring. However, current AD and AL approaches are often susceptible to adversarial attacks due to limitations in training data, which typically include only normal, unlabeled samples. This study introduces PatchGuard, an adversarially robust AD and AL method that incorporates pseudo anomalies with localization masks within a Vision Transformer (ViT)-based architecture to address these vulnerabilities. We begin by examining the essential properties of pseudo anomalies, and follow it by providing theoretical insights into the attention mechanisms required to enhance the adversarial robustness of AD and AL systems. We then present our approach, which leverages Foreground-Aware Pseudo-Anomalies to overcome the deficiencies of previous anomaly-aware methods. Our method incorporates these crafted pseudo-anomaly samples into a ViT-based framework, with adversarial training guided by a novel loss function designed to improve model robustness, as supported by our theoretical analysis. Experimental results on well-established industrial and medical datasets demonstrate that PatchGuard significantly outperforms previous methods in adversarial settings, achieving performance gains of $53.2\\%$ in AD and $68.5\\%$ in AL, while also maintaining competitive accuracy in non-adversarial settings. The code repository is available at https://github.com/rohban-lab/PatchGuard .", "date": "2025-06-10", "categories": []}
{"id": "2506.09538", "title": "AngleRoCL: Angle-Robust Concept Learning for Physically View-Invariant T2I Adversarial Patches", "abstract": "Cutting-edge works have demonstrated that text-to-image (T2I) diffusion models can generate adversarial patches that mislead state-of-the-art object detectors in the physical world, revealing detectors' vulnerabilities and risks. However, these methods neglect the T2I patches' attack effectiveness when observed from different views in the physical world (i.e., angle robustness of the T2I adversarial patches). In this paper, we study the angle robustness of T2I adversarial patches comprehensively, revealing their angle-robust issues, demonstrating that texts affect the angle robustness of generated patches significantly, and task-specific linguistic instructions fail to enhance the angle robustness. Motivated by the studies, we introduce Angle-Robust Concept Learning (AngleRoCL), a simple and flexible approach that learns a generalizable concept (i.e., text embeddings in implementation) representing the capability of generating angle-robust patches. The learned concept can be incorporated into textual prompts and guides T2I models to generate patches with their attack effectiveness inherently resistant to viewpoint variations. Through extensive simulation and physical-world experiments on five SOTA detectors across multiple views, we demonstrate that AngleRoCL significantly enhances the angle robustness of T2I adversarial patches compared to baseline methods. Our patches maintain high attack success rates even under challenging viewing conditions, with over 50% average relative improvement in attack effectiveness across multiple angles. This research advances the understanding of physically angle-robust patches and provides insights into the relationship between textual concepts and physical properties in T2I-generated contents. We released our code at https://github.com/tsingqguo/anglerocl.", "date": "2025-06-11", "categories": []}
{"id": "2509.15551", "title": "PolyJuice Makes It Real: Black-Box, Universal Red Teaming for Synthetic Image Detectors", "abstract": "Synthetic image detectors (SIDs) are a key defense against the risks posed by the growing realism of images from text-to-image (T2I) models. Red teaming improves SID's effectiveness by identifying and exploiting their failure modes via misclassified synthetic images. However, existing red-teaming solutions (i) require white-box access to SIDs, which is infeasible for proprietary state-of-the-art detectors, and (ii) generate image-specific attacks through expensive online optimization. To address these limitations, we propose PolyJuice, the first black-box, image-agnostic red-teaming method for SIDs, based on an observed distribution shift in the T2I latent space between samples correctly and incorrectly classified by the SID. PolyJuice generates attacks by (i) identifying the direction of this shift through a lightweight offline process that only requires black-box access to the SID, and (ii) exploiting this direction by universally steering all generated images towards the SID's failure modes. PolyJuice-steered T2I models are significantly more effective at deceiving SIDs (up to 84%) compared to their unsteered counterparts. We also show that the steering directions can be estimated efficiently at lower resolutions and transferred to higher resolutions using simple interpolation, reducing computational overhead. Finally, tuning SID models on PolyJuice-augmented datasets notably enhances the performance of the detectors (up to 30%).", "date": "2025-09-19", "categories": []}
{"id": "2510.20883", "title": "Kernel Learning with Adversarial Features: Numerical Efficiency and Adaptive Regularization", "abstract": "Adversarial training has emerged as a key technique to enhance model robustness against adversarial input perturbations. Many of the existing methods rely on computationally expensive min-max problems that limit their application in practice. We propose a novel formulation of adversarial training in reproducing kernel Hilbert spaces, shifting from input to feature-space perturbations. This reformulation enables the exact solution of inner maximization and efficient optimization. It also provides a regularized estimator that naturally adapts to the noise level and the smoothness of the underlying function. We establish conditions under which the feature-perturbed formulation is a relaxation of the original problem and propose an efficient optimization algorithm based on iterative kernel ridge regression. We provide generalization bounds that help to understand the properties of the method. We also extend the formulation to multiple kernel learning. Empirical evaluation shows good performance in both clean and adversarial settings.", "date": "2025-10-23", "categories": []}
{"id": "2510.20932", "title": "An Experimental Study of Trojan Vulnerabilities in UAV Autonomous Landing", "abstract": "This study investigates the vulnerabilities of autonomous navigation and landing systems in Urban Air Mobility (UAM) vehicles. Specifically, it focuses on Trojan attacks that target deep learning models, such as Convolutional Neural Networks (CNNs). Trojan attacks work by embedding covert triggers within a model's training data. These triggers cause specific failures under certain conditions, while the model continues to perform normally in other situations. We assessed the vulnerability of Urban Autonomous Aerial Vehicles (UAAVs) using the DroNet framework. Our experiments showed a significant drop in accuracy, from 96.4% on clean data to 73.3% on data triggered by Trojan attacks. To conduct this study, we collected a custom dataset and trained models to simulate real-world conditions. We also developed an evaluation framework designed to identify Trojan-infected models. This work demonstrates the potential security risks posed by Trojan attacks and lays the groundwork for future research on enhancing the resilience of UAM systems.", "date": "2025-10-23", "categories": []}
{"id": "2506.07031", "title": "HauntAttack: When Attack Follows Reasoning as a Shadow", "abstract": "Emerging Large Reasoning Models (LRMs) consistently excel in mathematical and reasoning tasks, showcasing remarkable capabilities. However, the enhancement of reasoning abilities and the exposure of internal reasoning processes introduce new safety vulnerabilities. A critical question arises: when reasoning becomes intertwined with harmfulness, will LRMs become more vulnerable to jailbreaks in reasoning mode? To investigate this, we introduce HauntAttack, a novel and general-purpose black-box adversarial attack framework that systematically embeds harmful instructions into reasoning questions. Specifically, we modify key reasoning conditions in existing questions with harmful instructions, thereby constructing a reasoning pathway that guides the model step by step toward unsafe outputs. We evaluate HauntAttack on 11 LRMs and observe an average attack success rate of 70\\%, achieving up to 12 percentage points of absolute improvement over the strongest prior baseline. Our further analysis reveals that even advanced safety-aligned models remain highly susceptible to reasoning-based attacks, offering insights into the urgent challenge of balancing reasoning capability and safety in future model development.", "date": "2025-06-08", "categories": []}
{"id": "2508.14757", "title": "Distributional Adversarial Attacks and Training in Deep Hedging", "abstract": "In this paper, we study the robustness of classical deep hedging strategies under distributional shifts by leveraging the concept of adversarial attacks. We first demonstrate that standard deep hedging models are highly vulnerable to small perturbations in the input distribution, resulting in significant performance degradation. Motivated by this, we propose an adversarial training framework tailored to increase the robustness of deep hedging strategies. Our approach extends pointwise adversarial attacks to the distributional setting and introduces a computationally tractable reformulation of the adversarial optimization problem over a Wasserstein ball. This enables the efficient training of hedging strategies that are resilient to distributional perturbations. Through extensive numerical experiments, we show that adversarially trained deep hedging strategies consistently outperform their classical counterparts in terms of out-of-sample performance and resilience to model misspecification. Additional results indicate that the robust strategies maintain reliable performance on real market data and remain effective during periods of market change. Our findings establish a practical and effective framework for robust deep hedging under realistic market uncertainties.", "date": "2025-08-20", "categories": []}
{"id": "2510.11570", "title": "Bag of Tricks for Subverting Reasoning-based Safety Guardrails", "abstract": "Recent reasoning-based safety guardrails for Large Reasoning Models (LRMs), such as deliberative alignment, have shown strong defense against jailbreak attacks. By leveraging LRMs' reasoning ability, these guardrails help the models to assess the safety of user inputs before generating final responses. The powerful reasoning ability can analyze the intention of the input query and will refuse to assist once it detects the harmful intent hidden by the jailbreak methods. Such guardrails have shown a significant boost in defense, such as the near-perfect refusal rates on the open-source gpt-oss series. Unfortunately, we find that these powerful reasoning-based guardrails can be extremely vulnerable to subtle manipulation of the input prompts, and once hijacked, can lead to even more harmful results. Specifically, we first uncover a surprisingly fragile aspect of these guardrails: simply adding a few template tokens to the input prompt can successfully bypass the seemingly powerful guardrails and lead to explicit and harmful responses. To explore further, we introduce a bag of jailbreak methods that subvert the reasoning-based guardrails. Our attacks span white-, gray-, and black-box settings and range from effortless template manipulations to fully automated optimization. Along with the potential for scalable implementation, these methods also achieve alarmingly high attack success rates (e.g., exceeding 90% across 5 different benchmarks on gpt-oss series on both local host models and online API services). Evaluations across various leading open-source LRMs confirm that these vulnerabilities are systemic, underscoring the urgent need for stronger alignment techniques for open-sourced LRMs to prevent malicious misuse. Code is open-sourced at https://chenxshuo.github.io/bag-of-tricks.", "date": "2025-10-13", "categories": []}
{"id": "2510.17621", "title": "GUIDE: Enhancing Gradient Inversion Attacks in Federated Learning with Denoising Models", "abstract": "Federated Learning (FL) enables collaborative training of Machine Learning (ML) models across multiple clients while preserving their privacy. Rather than sharing raw data, federated clients transmit locally computed updates to train the global model. Although this paradigm should provide stronger privacy guarantees than centralized ML, client updates remain vulnerable to privacy leakage. Adversaries can exploit them to infer sensitive properties about the training data or even to reconstruct the original inputs via Gradient Inversion Attacks (GIAs). Under the honest-butcurious threat model, GIAs attempt to reconstruct training data by reversing intermediate updates using optimizationbased techniques. We observe that these approaches usually reconstruct noisy approximations of the original inputs, whose quality can be enhanced with specialized denoising models. This paper presents Gradient Update Inversion with DEnoising (GUIDE), a novel methodology that leverages diffusion models as denoising tools to improve image reconstruction attacks in FL. GUIDE can be integrated into any GIAs that exploits surrogate datasets, a widely adopted assumption in GIAs literature. We comprehensively evaluate our approach in two attack scenarios that use different FL algorithms, models, and datasets. Our results demonstrate that GUIDE integrates seamlessly with two state-ofthe- art GIAs, substantially improving reconstruction quality across multiple metrics. Specifically, GUIDE achieves up to 46% higher perceptual similarity, as measured by the DreamSim metric.", "date": "2025-10-20", "categories": []}
{"id": "2510.19977", "title": "Towards Strong Certified Defense with Universal Asymmetric Randomization", "abstract": "Randomized smoothing has become essential for achieving certified adversarial robustness in machine learning models. However, current methods primarily use isotropic noise distributions that are uniform across all data dimensions, such as image pixels, limiting the effectiveness of robustness certification by ignoring the heterogeneity of inputs and data dimensions. To address this limitation, we propose UCAN: a novel technique that \\underline{U}niversally \\underline{C}ertifies adversarial robustness with \\underline{A}nisotropic \\underline{N}oise. UCAN is designed to enhance any existing randomized smoothing method, transforming it from symmetric (isotropic) to asymmetric (anisotropic) noise distributions, thereby offering a more tailored defense against adversarial attacks. Our theoretical framework is versatile, supporting a wide array of noise distributions for certified robustness in different $\\ell_p$-norms and applicable to any arbitrary classifier by guaranteeing the classifier's prediction over perturbed inputs with provable robustness bounds through tailored noise injection. Additionally, we develop a novel framework equipped with three exemplary noise parameter generators (NPGs) to optimally fine-tune the anisotropic noise parameters for different data dimensions, allowing for pursuing different levels of robustness enhancements in practice.Empirical evaluations underscore the significant leap in UCAN's performance over existing state-of-the-art methods, demonstrating up to $182.6\\%$ improvement in certified accuracy at large certified radii on MNIST, CIFAR10, and ImageNet datasets.\\footnote{Code is anonymously available at \\href{https://github.com/youbin2014/UCAN/}{https://github.com/youbin2014/UCAN/}}", "date": "2025-10-22", "categories": []}
{"id": "2510.20314", "title": "Enhancing Security in Deep Reinforcement Learning: A Comprehensive Survey on Adversarial Attacks and Defenses", "abstract": "With the wide application of deep reinforcement learning (DRL) techniques in complex fields such as autonomous driving, intelligent manufacturing, and smart healthcare, how to improve its security and robustness in dynamic and changeable environments has become a core issue in current research. Especially in the face of adversarial attacks, DRL may suffer serious performance degradation or even make potentially dangerous decisions, so it is crucial to ensure their stability in security-sensitive scenarios. In this paper, we first introduce the basic framework of DRL and analyze the main security challenges faced in complex and changing environments. In addition, this paper proposes an adversarial attack classification framework based on perturbation type and attack target and reviews the mainstream adversarial attack methods against DRL in detail, including various attack methods such as perturbation state space, action space, reward function and model space. To effectively counter the attacks, this paper systematically summarizes various current robustness training strategies, including adversarial training, competitive training, robust learning, adversarial detection, defense distillation and other related defense techniques, we also discuss the advantages and shortcomings of these methods in improving the robustness of DRL. Finally, this paper looks into the future research direction of DRL in adversarial environments, emphasizing the research needs in terms of improving generalization, reducing computational complexity, and enhancing scalability and explainability, aiming to provide valuable references and directions for researchers.", "date": "2025-10-23", "categories": []}
{"id": "2410.20893", "title": "Adversarial Attacks on LiDAR-Based Tracking Across Road Users: Robustness Evaluation and Target-Aware Black-Box Method", "abstract": "In this study, we delve into the robustness of neural network-based LiDAR point cloud tracking models under adversarial attacks, a critical aspect often overlooked in favor of performance enhancement. These models, despite incorporating advanced architectures like Transformer or Bird's Eye View (BEV), tend to neglect robustness in the face of challenges such as adversarial attacks, domain shifts, or data corruption. We instead focus on the robustness of the tracking models under the threat of adversarial attacks. We begin by establishing a unified framework for conducting adversarial attacks within the context of 3D object tracking, which allows us to thoroughly investigate both white-box and black-box attack strategies. For white-box attacks, we tailor specific loss functions to accommodate various tracking paradigms and extend existing methods such as FGSM, C\\&W, and PGD to the point cloud domain. In addressing black-box attack scenarios, we introduce a novel transfer-based approach, the Target-aware Perturbation Generation (TAPG) algorithm, with the dual objectives of achieving high attack performance and maintaining low perceptibility. This method employs a heuristic strategy to enforce sparse attack constraints and utilizes random sub-vector factorization to bolster transferability. Our experimental findings reveal a significant vulnerability in advanced tracking methods when subjected to both black-box and white-box attacks, underscoring the necessity for incorporating robustness against adversarial attacks into the design of LiDAR point cloud tracking models. Notably, compared to existing methods, the TAPG also strikes an optimal balance between the effectiveness of the attack and the concealment of the perturbations.", "date": "2024-10-28", "categories": []}
{"id": "2502.01349", "title": "Bias Beware: The Impact of Cognitive Biases on LLM-Driven Product Recommendations", "abstract": "The advent of Large Language Models (LLMs) has revolutionized product recommenders, yet their susceptibility to adversarial manipulation poses critical challenges, particularly in real-world commercial applications. Our approach is the first one to tap into human psychological principles, seamlessly modifying product descriptions, making such manipulations hard to detect. In this work, we investigate cognitive biases as black-box adversarial strategies, drawing parallels between their effects on LLMs and human purchasing behavior. Through extensive evaluation across models of varying scale, we find that certain biases, such as social proof, consistently boost product recommendation rate and ranking, while others, like scarcity and exclusivity, surprisingly reduce visibility. Our results demonstrate that cognitive biases are deeply embedded in state-of-the-art LLMs, leading to highly unpredictable behavior in product recommendations and posing significant challenges for effective mitigation.", "date": "2025-02-03", "categories": []}
{"id": "2510.18362", "title": "FeatureFool: Zero-Query Fooling of Video Models via Feature Map", "abstract": "The vulnerability of deep neural networks (DNNs) has been preliminarily verified. Existing black-box adversarial attacks usually require multi-round interaction with the model and consume numerous queries, which is impractical in the real-world and hard to scale to recently emerged Video-LLMs. Moreover, no attack in the video domain directly leverages feature maps to shift the clean-video feature space. We therefore propose FeatureFool, a stealthy, video-domain, zero-query black-box attack that utilizes information extracted from a DNN to alter the feature space of clean videos. Unlike query-based methods that rely on iterative interaction, FeatureFool performs a zero-query attack by directly exploiting DNN-extracted information. This efficient approach is unprecedented in the video domain. Experiments show that FeatureFool achieves an attack success rate above 70\\% against traditional video classifiers without any queries. Benefiting from the transferability of the feature map, it can also craft harmful content and bypass Video-LLM recognition. Additionally, adversarial videos generated by FeatureFool exhibit high quality in terms of SSIM, PSNR, and Temporal-Inconsistency, making the attack barely perceptible. This paper may contain violent or explicit content.", "date": "2025-10-21", "categories": []}
{"id": "2510.19347", "title": "A New Type of Adversarial Examples", "abstract": "Most machine learning models are vulnerable to adversarial examples, which poses security concerns on these models. Adversarial examples are crafted by applying subtle but intentionally worst-case modifications to examples from the dataset, leading the model to output a different answer from the original example. In this paper, adversarial examples are formed in an exactly opposite manner, which are significantly different from the original examples but result in the same answer. We propose a novel set of algorithms to produce such adversarial examples, including the negative iterative fast gradient sign method (NI-FGSM) and the negative iterative fast gradient method (NI-FGM), along with their momentum variants: the negative momentum iterative fast gradient sign method (NMI-FGSM) and the negative momentum iterative fast gradient method (NMI-FGM). Adversarial examples constructed by these methods could be used to perform an attack on machine learning systems in certain occasions. Moreover, our results show that the adversarial examples are not merely distributed in the neighbourhood of the examples from the dataset; instead, they are distributed extensively in the sample space.", "date": "2025-10-22", "categories": []}
{"id": "2510.19427", "title": "Revisiting the Relation Between Robustness and Universality", "abstract": "The modified universality hypothesis proposed by Jones et al. (2022) suggests that adversarially robust models trained for a given task are highly similar. We revisit the hypothesis and test its generality. While we verify Jones' main claim of high representational similarity in specific settings, results are not consistent across different datasets. We also discover that predictive behavior does not converge with increasing robustness and thus is not universal. We find that differing predictions originate in the classification layer, but show that more universal predictive behavior can be achieved with simple retraining of the classifiers. Overall, our work points towards partial universality of neural networks in specific settings and away from notions of strict universality.", "date": "2025-10-22", "categories": []}
{"id": "2510.19574", "title": "Can You Trust What You See? Alpha Channel No-Box Attacks on Video Object Detection", "abstract": "As object detection models are increasingly deployed in cyber-physical systems such as autonomous vehicles (AVs) and surveillance platforms, ensuring their security against adversarial threats is essential. While prior work has explored adversarial attacks in the image domain, those attacks in the video domain remain largely unexamined, especially in the no-box setting. In this paper, we present {\\alpha}-Cloak, the first no-box adversarial attack on object detectors that operates entirely through the alpha channel of RGBA videos. {\\alpha}-Cloak exploits the alpha channel to fuse a malicious target video with a benign video, resulting in a fused video that appears innocuous to human viewers but consistently fools object detectors. Our attack requires no access to model architecture, parameters, or outputs, and introduces no perceptible artifacts. We systematically study the support for alpha channels across common video formats and playback applications, and design a fusion algorithm that ensures visual stealth and compatibility. We evaluate {\\alpha}-Cloak on five state-of-the-art object detectors, a vision-language model, and a multi-modal large language model (Gemini-2.0-Flash), demonstrating a 100% attack success rate across all scenarios. Our findings reveal a previously unexplored vulnerability in video-based perception systems, highlighting the urgent need for defenses that account for the alpha channel in adversarial settings.", "date": "2025-10-22", "categories": []}
{"id": "2510.19641", "title": "Style Attack Disguise: When Fonts Become a Camouflage for Adversarial Intent", "abstract": "With social media growth, users employ stylistic fonts and font-like emoji to express individuality, creating visually appealing text that remains human-readable. However, these fonts introduce hidden vulnerabilities in NLP models: while humans easily read stylistic text, models process these characters as distinct tokens, causing interference. We identify this human-model perception gap and propose a style-based attack, Style Attack Disguise (SAD). We design two sizes: light for query efficiency and strong for superior attack performance. Experiments on sentiment classification and machine translation across traditional models, LLMs, and commercial services demonstrate SAD's strong attack performance. We also show SAD's potential threats to multimodal tasks including text-to-image and text-to-speech generation.", "date": "2025-10-22", "categories": []}
{"id": "2510.19761", "title": "Exploring the Effect of DNN Depth on Adversarial Attacks in Network Intrusion Detection Systems", "abstract": "Adversarial attacks pose significant challenges to Machine Learning (ML) systems and especially Deep Neural Networks (DNNs) by subtly manipulating inputs to induce incorrect predictions. This paper investigates whether increasing the layer depth of deep neural networks affects their robustness against adversarial attacks in the Network Intrusion Detection System (NIDS) domain. We compare the adversarial robustness of various deep neural networks across both \\ac{NIDS} and computer vision domains (the latter being widely used in adversarial attack experiments). Our experimental results reveal that in the NIDS domain, adding more layers does not necessarily improve their performance, yet it may actually significantly degrade their robustness against adversarial attacks. Conversely, in the computer vision domain, adding more layers exhibits a more modest impact on robustness. These findings can guide the development of robust neural networks for (NIDS) applications and highlight the unique characteristics of network security domains within the (ML) landscape.", "date": "2025-10-22", "categories": []}
{"id": "2503.23804", "title": "DrunkAgent: Stealthy Memory Corruption in LLM-Powered Recommender Agents", "abstract": "Large language model (LLM)-powered agents are increasingly used in recommender systems (RSs) to achieve personalized behavior modeling, where the memory mechanism plays a pivotal role in enabling the agents to autonomously explore, learn and self-evolve from real-world interactions. However, this very mechanism, serving as a contextual repository, inherently exposes an attack surface for potential adversarial manipulations. Despite its central role, the robustness of agentic RSs in the face of such threats remains largely underexplored. Previous works suffer from semantic mismatches or rely on static embeddings or pre-defined prompts, all of which are not designed for dynamic systems, especially for dynamic memory states of LLM agents. This challenge is exacerbated by the black-box nature of commercial recommenders.\n  To tackle the above problems, in this paper, we present the first systematic investigation of memory-based vulnerabilities in LLM-powered recommender agents, revealing their security limitations and guiding efforts to strengthen system resilience and trustworthiness. Specifically, we propose a novel black-box attack framework named DrunkAgent. DrunkAgent crafts semantically meaningful adversarial textual triggers for target item promotions and introduces a series of strategies to maximize the trigger effect by corrupting the memory updates during the interactions. The triggers and strategies are optimized on a surrogate model, enabling DrunkAgent transferable and stealthy. Extensive experiments on real-world datasets across diverse agentic RSs, including collaborative filtering, retrieval augmentation and sequential recommendations, demonstrate the generalizability, transferability and stealthiness of DrunkAgent.", "date": "2025-03-31", "categories": []}
{"id": "2506.07153", "title": "Mind the Web: The Security of Web Use Agents", "abstract": "Web-use agents are rapidly being deployed to automate complex web tasks with extensive browser capabilities. However, these capabilities create a critical and previously unexplored attack surface. This paper demonstrates how attackers can exploit web-use agents by embedding malicious content in web pages, such as comments, reviews, or advertisements, that agents encounter during legitimate browsing tasks. We introduce the task-aligned injection technique that frames malicious commands as helpful task guidance rather than obvious attacks, exploiting fundamental limitations in LLMs' contextual reasoning. Agents struggle to maintain coherent contextual awareness and fail to detect when seemingly helpful web content contains steering attempts that deviate them from their original task goal. To scale this attack, we developed an automated three-stage pipeline that generates effective injections without manual annotation or costly online agent interactions during training, remaining efficient even with limited training data. This pipeline produces a generator model that we evaluate on five popular agents using payloads organized by the Confidentiality-Integrity-Availability (CIA) security triad, including unauthorized camera activation, file exfiltration, user impersonation, phishing, and denial-of-service. This generator achieves over 80% attack success rate (ASR) with strong transferability across unseen payloads, diverse web environments, and different underlying LLMs. This attack succeed even against agents with built-in safety mechanisms, requiring only the ability to post content on public websites. To address this risk, we propose comprehensive mitigation strategies including oversight mechanisms, execution constraints, and task-aware reasoning techniques.", "date": "2025-06-08", "categories": []}
{"id": "2510.18381", "title": "S2AP: Score-space Sharpness Minimization for Adversarial Pruning", "abstract": "Adversarial pruning methods have emerged as a powerful tool for compressing neural networks while preserving robustness against adversarial attacks. These methods typically follow a three-step pipeline: (i) pretrain a robust model, (ii) select a binary mask for weight pruning, and (iii) finetune the pruned model. To select the binary mask, these methods minimize a robust loss by assigning an importance score to each weight, and then keep the weights with the highest scores. However, this score-space optimization can lead to sharp local minima in the robust loss landscape and, in turn, to an unstable mask selection, reducing the robustness of adversarial pruning methods. To overcome this issue, we propose a novel plug-in method for adversarial pruning, termed Score-space Sharpness-aware Adversarial Pruning (S2AP). Through our method, we introduce the concept of score-space sharpness minimization, which operates during the mask search by perturbing importance scores and minimizing the corresponding robust loss. Extensive experiments across various datasets, models, and sparsity levels demonstrate that S2AP effectively minimizes sharpness in score space, stabilizing the mask selection, and ultimately improving the robustness of adversarial pruning methods.", "date": "2025-10-21", "categories": []}
{"id": "2510.18728", "title": "HarmNet: A Framework for Adaptive Multi-Turn Jailbreak Attacks on Large Language Models", "abstract": "Large Language Models (LLMs) remain vulnerable to multi-turn jailbreak attacks. We introduce HarmNet, a modular framework comprising ThoughtNet, a hierarchical semantic network; a feedback-driven Simulator for iterative query refinement; and a Network Traverser for real-time adaptive attack execution. HarmNet systematically explores and refines the adversarial space to uncover stealthy, high-success attack paths. Experiments across closed-source and open-source LLMs show that HarmNet outperforms state-of-the-art methods, achieving higher attack success rates. For example, on Mistral-7B, HarmNet achieves a 99.4% attack success rate, 13.9% higher than the best baseline. Index terms: jailbreak attacks; large language models; adversarial framework; query refinement.", "date": "2025-10-21", "categories": []}
{"id": "2303.04238", "title": "Patch of Invisibility: Naturalistic Physical Black-Box Adversarial Attacks on Object Detectors", "abstract": "Adversarial attacks on deep learning models have received increased attention in recent years. Work in this area has mostly focused on gradient-based techniques, so-called 'white-box' attacks, where the attacker has access to the targeted model's internal parameters; such an assumption is usually untenable in the real world. Additionally, some attacks use the entire pixel space to fool a given model, which is neither practical nor physical. To accommodate these problems we propose the BBNP algorithm (Black-Box Naturalistic Patch): a direct, black-box, naturalistic, gradient-free method that uses the learned image manifold of a pretrained, generative adversarial network (GAN) to generate naturalistic adversarial patches for object detectors. This method performs model-agnostic black-box naturalistic attacks on object detection models by relying solely on the outputs of the model. Comparing our approach against five models, five black-box and two white-box attacks, we show that our proposed method achieves state-of-the-art results, outperforming all other tested black-box approaches.", "date": "2023-03-07", "categories": []}
{"id": "2504.06492", "title": "Exploiting Meta-Learning-based Poisoning Attacks for Graph Link Prediction", "abstract": "Link prediction in graph data uses various algorithms and Graph Nerual Network (GNN) models to predict potential relationships between graph nodes. These techniques have found widespread use in numerous real-world applications, including recommendation systems, community/social networks, and biological structures. However, recent research has highlighted the vulnerability of GNN models to adversarial attacks, such as poisoning and evasion attacks. Addressing the vulnerability of GNN models is crucial to ensure stable and robust performance in GNN applications. Although many works have focused on enhancing the robustness of node classification on GNN models, the robustness of link prediction has received less attention. To bridge this gap, this article introduces an unweighted graph poisoning attack that leverages meta-learning with weighted scheme strategies to degrade the link prediction performance of GNNs. We conducted comprehensive experiments on diverse datasets across multiple link prediction applications to evaluate the proposed method and its parameters, comparing it with existing approaches under similar conditions. Our results demonstrate that our approach significantly reduces link prediction performance and consistently outperforms other state-of-the-art baselines.", "date": "2025-04-08", "categories": []}
{"id": "2509.19360", "title": "Semantic Representation Attack against Aligned Large Language Models", "abstract": "Large Language Models (LLMs) increasingly employ alignment techniques to prevent harmful outputs. Despite these safeguards, attackers can circumvent them by crafting prompts that induce LLMs to generate harmful content.\n  Current methods typically target exact affirmative responses, such as ``Sure, here is...'', suffering from limited convergence, unnatural prompts, and high computational costs.\n  We introduce Semantic Representation Attack, a novel paradigm that fundamentally reconceptualizes adversarial objectives against aligned LLMs.\n  Rather than targeting exact textual patterns, our approach exploits the semantic representation space comprising diverse responses with equivalent harmful meanings.\n  This innovation resolves the inherent trade-off between attack efficacy and prompt naturalness that plagues existing methods.\n  The Semantic Representation Heuristic Search algorithm is proposed to efficiently generate semantically coherent and concise adversarial prompts by maintaining interpretability during incremental expansion.\n  We establish rigorous theoretical guarantees for semantic convergence and demonstrate that our method achieves unprecedented attack success rates (89.41\\% averaged across 18 LLMs, including 100\\% on 11 models) while maintaining stealthiness and efficiency.\n  Comprehensive experimental results confirm the overall superiority of our Semantic Representation Attack.\n  The code will be publicly available.", "date": "2025-09-18", "categories": []}
{"id": "2510.15965", "title": "One Token Embedding Is Enough to Deadlock Your Large Reasoning Model", "abstract": "Modern large reasoning models (LRMs) exhibit impressive multi-step problem-solving via chain-of-thought (CoT) reasoning. However, this iterative thinking mechanism introduces a new vulnerability surface. We present the Deadlock Attack, a resource exhaustion method that hijacks an LRM's generative control flow by training a malicious adversarial embedding to induce perpetual reasoning loops. Specifically, the optimized embedding encourages transitional tokens (e.g., \"Wait\", \"But\") after reasoning steps, preventing the model from concluding its answer. A key challenge we identify is the continuous-to-discrete projection gap: na\\\"ive projections of adversarial embeddings to token sequences nullify the attack. To overcome this, we introduce a backdoor implantation strategy, enabling reliable activation through specific trigger tokens. Our method achieves a 100% attack success rate across four advanced LRMs (Phi-RM, Nemotron-Nano, R1-Qwen, R1-Llama) and three math reasoning benchmarks, forcing models to generate up to their maximum token limits. The attack is also stealthy (in terms of causing negligible utility loss on benign user inputs) and remains robust against existing strategies trying to mitigate the overthinking issue. Our findings expose a critical and underexplored security vulnerability in LRMs from the perspective of reasoning (in)efficiency.", "date": "2025-10-12", "categories": []}
{"id": "2510.15973", "title": "Safeguarding Efficacy in Large Language Models: Evaluating Resistance to Human-Written and Algorithmic Adversarial Prompts", "abstract": "This paper presents a systematic security assessment of four prominent Large Language Models (LLMs) against diverse adversarial attack vectors. We evaluate Phi-2, Llama-2-7B-Chat, GPT-3.5-Turbo, and GPT-4 across four distinct attack categories: human-written prompts, AutoDAN, Greedy Coordinate Gradient (GCG), and Tree-of-Attacks-with-pruning (TAP). Our comprehensive evaluation employs 1,200 carefully stratified prompts from the SALAD-Bench dataset, spanning six harm categories. Results demonstrate significant variations in model robustness, with Llama-2 achieving the highest overall security (3.4% average attack success rate) while Phi-2 exhibits the greatest vulnerability (7.0% average attack success rate). We identify critical transferability patterns where GCG and TAP attacks, though ineffective against their target model (Llama-2), achieve substantially higher success rates when transferred to other models (up to 17% for GPT-4). Statistical analysis using Friedman tests reveals significant differences in vulnerability across harm categories ($p < 0.001$), with malicious use prompts showing the highest attack success rates (10.71% average). Our findings contribute to understanding cross-model security vulnerabilities and provide actionable insights for developing targeted defense mechanisms", "date": "2025-10-12", "categories": []}
{"id": "2510.16005", "title": "Breaking Guardrails, Facing Walls: Insights on Adversarial AI for Defenders & Researchers", "abstract": "Analyzing 500 CTF participants, this paper shows that while participants readily bypassed simple AI guardrails using common techniques, layered multi-step defenses still posed significant challenges, offering concrete insights for building safer AI systems.", "date": "2025-10-14", "categories": []}
{"id": "2510.16035", "title": "RoBCtrl: Attacking GNN-Based Social Bot Detectors via Reinforced Manipulation of Bots Control Interaction", "abstract": "Social networks have become a crucial source of real-time information for individuals. The influence of social bots within these platforms has garnered considerable attention from researchers, leading to the development of numerous detection technologies. However, the vulnerability and robustness of these detection methods is still underexplored. Existing Graph Neural Network (GNN)-based methods cannot be directly applied due to the issues of limited control over social agents, the black-box nature of bot detectors, and the heterogeneity of bots. To address these challenges, this paper proposes the first adversarial multi-agent Reinforcement learning framework for social Bot control attacks (RoBCtrl) targeting GNN-based social bot detectors. Specifically, we use a diffusion model to generate high-fidelity bot accounts by reconstructing existing account data with minor modifications, thereby evading detection on social platforms. To the best of our knowledge, this is the first application of diffusion models to mimic the behavior of evolving social bots effectively. We then employ a Multi-Agent Reinforcement Learning (MARL) method to simulate bots adversarial behavior. We categorize social accounts based on their influence and budget. Different agents are then employed to control bot accounts across various categories, optimizing the attachment strategy through reinforcement learning. Additionally, a hierarchical state abstraction based on structural entropy is designed to accelerate the reinforcement learning. Extensive experiments on social bot detection datasets demonstrate that our framework can effectively undermine the performance of GNN-based detectors.", "date": "2025-10-16", "categories": []}
{"id": "2510.16259", "title": "Distractor Injection Attacks on Large Reasoning Models: Characterization and Defense", "abstract": "Recent advances in large reasoning models (LRMs) have enabled remarkable performance on complex tasks such as mathematics and coding by generating long Chain-of-Thought (CoT) traces. In this paper, we identify and systematically analyze a critical vulnerability we term reasoning distraction, where LRMs are diverted from their primary objective by irrelevant yet complex tasks maliciously embedded in the prompt. Through a comprehensive study across diverse models and benchmarks, we show that even state-of-the-art LRMs are highly susceptible, with injected distractors reducing task accuracy by up to 60%. We further reveal that certain alignment techniques can amplify this weakness and that models may exhibit covert compliance, following hidden adversarial instructions in reasoning while concealing them in the final output. To mitigate these risks, we propose a training-based defense that combines Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) on synthetic adversarial data, improving robustness by over 50 points on challenging distractor attacks. Our findings establish reasoning distraction as a distinct and urgent threat to LRM reliability and provide a practical step toward safer and more trustworthy reasoning systems.", "date": "2025-10-17", "categories": []}
{"id": "2510.16440", "title": "Colliding with Adversaries at ECML-PKDD 2025 Adversarial Attack Competition 1st Prize Solution", "abstract": "This report presents the winning solution for Task 1 of Colliding with Adversaries: A Challenge on Robust Learning in High Energy Physics Discovery at ECML-PKDD 2025. The task required designing an adversarial attack against a provided classification model that maximizes misclassification while minimizing perturbations. Our approach employs a multi-round gradient-based strategy that leverages the differentiable structure of the model, augmented with random initialization and sample-mixing techniques to enhance effectiveness. The resulting attack achieved the best results in perturbation size and fooling success rate, securing first place in the competition.", "date": "2025-10-18", "categories": []}
{"id": "2510.16581", "title": "Patronus: Safeguarding Text-to-Image Models against White-Box Adversaries", "abstract": "Text-to-image (T2I) models, though exhibiting remarkable creativity in image generation, can be exploited to produce unsafe images. Existing safety measures, e.g., content moderation or model alignment, fail in the presence of white-box adversaries who know and can adjust model parameters, e.g., by fine-tuning. This paper presents a novel defensive framework, named Patronus, which equips T2I models with holistic protection to defend against white-box adversaries. Specifically, we design an internal moderator that decodes unsafe input features into zero vectors while ensuring the decoding performance of benign input features. Furthermore, we strengthen the model alignment with a carefully designed non-fine-tunable learning mechanism, ensuring the T2I model will not be compromised by malicious fine-tuning. We conduct extensive experiments to validate the intactness of the performance on safe content generation and the effectiveness of rejecting unsafe content generation. Results also confirm the resilience of Patronus against various fine-tuning attacks by white-box adversaries.", "date": "2025-10-18", "categories": []}
{"id": "2510.16637", "title": "A Versatile Framework for Designing Group-Sparse Adversarial Attacks", "abstract": "Existing adversarial attacks often neglect perturbation sparsity, limiting their ability to model structural changes and to explain how deep neural networks (DNNs) process meaningful input patterns. We propose ATOS (Attack Through Overlapping Sparsity), a differentiable optimization framework that generates structured, sparse adversarial perturbations in element-wise, pixel-wise, and group-wise forms. For white-box attacks on image classifiers, we introduce the Overlapping Smoothed L0 (OSL0) function, which promotes convergence to a stationary point while encouraging sparse, structured perturbations. By grouping channels and adjacent pixels, ATOS improves interpretability and helps identify robust versus non-robust features. We approximate the L-infinity gradient using the logarithm of the sum of exponential absolute values to tightly control perturbation magnitude. On CIFAR-10 and ImageNet, ATOS achieves a 100% attack success rate while producing significantly sparser and more structurally coherent perturbations than prior methods. The structured group-wise attack highlights critical regions from the network's perspective, providing counterfactual explanations by replacing class-defining regions with robust features from the target class.", "date": "2025-10-18", "categories": []}
{"id": "2510.17185", "title": "Robustness in Text-Attributed Graph Learning: Insights, Trade-offs, and New Defenses", "abstract": "While Graph Neural Networks (GNNs) and Large Language Models (LLMs) are powerful approaches for learning on Text-Attributed Graphs (TAGs), a comprehensive understanding of their robustness remains elusive. Current evaluations are fragmented, failing to systematically investigate the distinct effects of textual and structural perturbations across diverse models and attack scenarios. To address these limitations, we introduce a unified and comprehensive framework to evaluate robustness in TAG learning. Our framework evaluates classical GNNs, robust GNNs (RGNNs), and GraphLLMs across ten datasets from four domains, under diverse text-based, structure-based, and hybrid perturbations in both poisoning and evasion scenarios. Our extensive analysis reveals multiple findings, among which three are particularly noteworthy: 1) models have inherent robustness trade-offs between text and structure, 2) the performance of GNNs and RGNNs depends heavily on the text encoder and attack type, and 3) GraphLLMs are particularly vulnerable to training data corruption. To overcome the identified trade-offs, we introduce SFT-auto, a novel framework that delivers superior and balanced robustness against both textual and structural attacks within a single model. Our work establishes a foundation for future research on TAG security and offers practical solutions for robust TAG learning in adversarial environments. Our code is available at: https://github.com/Leirunlin/TGRB.", "date": "2025-10-20", "categories": []}
{"id": "2510.17322", "title": "A Single Set of Adversarial Clothes Breaks Multiple Defense Methods in the Physical World", "abstract": "In recent years, adversarial attacks against deep learning-based object detectors in the physical world have attracted much attention. To defend against these attacks, researchers have proposed various defense methods against adversarial patches, a typical form of physically-realizable attack. However, our experiments showed that simply enlarging the patch size could make these defense methods fail. Motivated by this, we evaluated various defense methods against adversarial clothes which have large coverage over the human body. Adversarial clothes provide a good test case for adversarial defense against patch-based attacks because they not only have large sizes but also look more natural than a large patch on humans. Experiments show that all the defense methods had poor performance against adversarial clothes in both the digital world and the physical world. In addition, we crafted a single set of clothes that broke multiple defense methods on Faster R-CNN. The set achieved an Attack Success Rate (ASR) of 96.06% against the undefended detector and over 64.84% ASRs against nine defended models in the physical world, unveiling the common vulnerability of existing adversarial defense methods against adversarial clothes. Code is available at: https://github.com/weiz0823/adv-clothes-break-multiple-defenses.", "date": "2025-10-20", "categories": []}
{"id": "2505.11717", "title": "WebInject: Prompt Injection Attack to Web Agents", "abstract": "Multi-modal large language model (MLLM)-based web agents interact with webpage environments by generating actions based on screenshots of the webpages. In this work, we propose WebInject, a prompt injection attack that manipulates the webpage environment to induce a web agent to perform an attacker-specified action. Our attack adds a perturbation to the raw pixel values of the rendered webpage. After these perturbed pixels are mapped into a screenshot, the perturbation induces the web agent to perform the attacker-specified action. We formulate the task of finding the perturbation as an optimization problem. A key challenge in solving this problem is that the mapping between raw pixel values and screenshot is non-differentiable, making it difficult to backpropagate gradients to the perturbation. To overcome this, we train a neural network to approximate the mapping and apply projected gradient descent to solve the reformulated optimization problem. Extensive evaluation on multiple datasets shows that WebInject is highly effective and significantly outperforms baselines.", "date": "2025-05-16", "categories": []}
{"id": "2507.04100", "title": "Hierarchical Testing with Rabbit Optimization for Industrial Cyber-Physical Systems", "abstract": "This paper presents HERO (Hierarchical Testing with Rabbit Optimization), a novel black-box adversarial testing framework for evaluating the robustness of deep learning-based Prognostics and Health Management systems in Industrial Cyber-Physical Systems. Leveraging Artificial Rabbit Optimization, HERO generates physically constrained adversarial examples that align with real-world data distributions via global and local perspective. Its generalizability ensures applicability across diverse ICPS scenarios. This study specifically focuses on the Proton Exchange Membrane Fuel Cell system, chosen for its highly dynamic operational conditions, complex degradation mechanisms, and increasing integration into ICPS as a sustainable and efficient energy solution. Experimental results highlight HERO's ability to uncover vulnerabilities in even state-of-the-art PHM models, underscoring the critical need for enhanced robustness in real-world applications. By addressing these challenges, HERO demonstrates its potential to advance more resilient PHM systems across a wide range of ICPS domains.", "date": "2025-07-05", "categories": []}
{"id": "2507.17725", "title": "On the Interaction of Compressibility and Adversarial Robustness", "abstract": "Modern neural networks are expected to simultaneously satisfy a host of desirable properties: accurate fitting to training data, generalization to unseen inputs, parameter and computational efficiency, and robustness to adversarial perturbations. While compressibility and robustness have each been studied extensively, a unified understanding of their interaction still remains elusive. In this work, we develop a principled framework to analyze how different forms of compressibility - such as neuron-level sparsity and spectral compressibility - affect adversarial robustness. We show that these forms of compression can induce a small number of highly sensitive directions in the representation space, which adversaries can exploit to construct effective perturbations. Our analysis yields a simple yet instructive robustness bound, revealing how neuron and spectral compressibility impact $L_\\infty$ and $L_2$ robustness via their effects on the learned representations. Crucially, the vulnerabilities we identify arise irrespective of how compression is achieved - whether via regularization, architectural bias, or implicit learning dynamics. Through empirical evaluations across synthetic and realistic tasks, we confirm our theoretical predictions, and further demonstrate that these vulnerabilities persist under adversarial training and transfer learning, and contribute to the emergence of universal adversarial perturbations. Our findings show a fundamental tension between structured compressibility and robustness, and suggest new pathways for designing models that are both efficient and secure.", "date": "2025-07-23", "categories": []}
{"id": "2510.15333", "title": "Backdoor or Manipulation? Graph Mixture of Experts Can Defend Against Various Graph Adversarial Attacks", "abstract": "Extensive research has highlighted the vulnerability of graph neural networks (GNNs) to adversarial attacks, including manipulation, node injection, and the recently emerging threat of backdoor attacks. However, existing defenses typically focus on a single type of attack, lacking a unified approach to simultaneously defend against multiple threats. In this work, we leverage the flexibility of the Mixture of Experts (MoE) architecture to design a scalable and unified framework for defending against backdoor, edge manipulation, and node injection attacks. Specifically, we propose an MI-based logic diversity loss to encourage individual experts to focus on distinct neighborhood structures in their decision processes, thus ensuring a sufficient subset of experts remains unaffected under perturbations in local structures. Moreover, we introduce a robustness-aware router that identifies perturbation patterns and adaptively routes perturbed nodes to corresponding robust experts. Extensive experiments conducted under various adversarial settings demonstrate that our method consistently achieves superior robustness against multiple graph adversarial attacks.", "date": "2025-10-17", "categories": []}
{"id": "2510.15699", "title": "Constrained Adversarial Perturbation", "abstract": "Deep neural networks have achieved remarkable success in a wide range of classification tasks. However, they remain highly susceptible to adversarial examples - inputs that are subtly perturbed to induce misclassification while appearing unchanged to humans. Among various attack strategies, Universal Adversarial Perturbations (UAPs) have emerged as a powerful tool for both stress testing model robustness and facilitating scalable adversarial training. Despite their effectiveness, most existing UAP methods neglect domain specific constraints that govern feature relationships. Violating such constraints, such as debt to income ratios in credit scoring or packet flow invariants in network communication, can render adversarial examples implausible or easily detectable, thereby limiting their real world applicability.\n  In this work, we advance universal adversarial attacks to constrained feature spaces by formulating an augmented Lagrangian based min max optimization problem that enforces multiple, potentially complex constraints of varying importance. We propose Constrained Adversarial Perturbation (CAP), an efficient algorithm that solves this problem using a gradient based alternating optimization strategy. We evaluate CAP across diverse domains including finance, IT networks, and cyber physical systems, and demonstrate that it achieves higher attack success rates while significantly reducing runtime compared to existing baselines. Our approach also generalizes seamlessly to individual adversarial perturbations, where we observe similar strong performance gains. Finally, we introduce a principled procedure for learning feature constraints directly from data, enabling broad applicability across domains with structured input spaces.", "date": "2025-10-17", "categories": []}
{"id": "2409.12394", "title": "The Fluorescent Veil: A Stealthy and Effective Physical Adversarial Patch Against Traffic Sign Recognition", "abstract": "Recently, traffic sign recognition (TSR) systems have become a prominent target for physical adversarial attacks. These attacks typically rely on conspicuous stickers and projections, or using invisible light and acoustic signals that can be easily blocked. In this paper, we introduce a novel attack medium, i.e., fluorescent ink, to design a stealthy and effective physical adversarial patch, namely FIPatch, to advance the state-of-the-art. Specifically, we first model the fluorescence effect in the digital domain to identify the optimal attack settings, which guide the real-world fluorescence parameters. By applying a carefully designed fluorescence perturbation to the target sign, the attacker can later trigger a fluorescent effect using invisible ultraviolet light, causing the TSR system to misclassify the sign and potentially leading to traffic accidents. We conducted a comprehensive evaluation to investigate the effectiveness of FIPatch, which shows a success rate of 98.31% in low-light conditions. Furthermore, our attack successfully bypasses five popular defenses and achieves a success rate of 96.72%.", "date": "2024-09-19", "categories": []}
{"id": "2505.24842", "title": "Cascading Adversarial Bias from Injection to Distillation in Language Models", "abstract": "Model distillation has become essential for creating smaller, deployable language models that retain larger system capabilities. However, widespread deployment raises concerns about resilience to adversarial manipulation. This paper investigates vulnerability of distilled models to adversarial injection of biased content during training. We demonstrate that adversaries can inject subtle biases into teacher models through minimal data poisoning, which propagates to student models and becomes significantly amplified. We propose two propagation modes: Untargeted Propagation, where bias affects multiple tasks, and Targeted Propagation, focusing on specific tasks while maintaining normal behavior elsewhere. With only 25 poisoned samples (0.25% poisoning rate), student models generate biased responses 76.9% of the time in targeted scenarios - higher than 69.4% in teacher models. For untargeted propagation, adversarial bias appears 6x-29x more frequently in student models on unseen tasks. We validate findings across six bias types (targeted advertisements, phishing links, narrative manipulations, insecure coding practices), various distillation methods, and different modalities spanning text and code generation. Our evaluation reveals shortcomings in current defenses - perplexity filtering, bias detection systems, and LLM-based autorater frameworks - against these attacks. Results expose significant security vulnerabilities in distilled models, highlighting need for specialized safeguards. We propose practical design principles for building effective adversarial bias mitigation strategies.", "date": "2025-05-30", "categories": []}
{"id": "2507.21750", "title": "Adversarial Defence without Adversarial Defence: Enhancing Language Model Robustness via Instance-level Principal Component Removal", "abstract": "Pre-trained language models (PLMs) have driven substantial progress in natural language processing but remain vulnerable to adversarial attacks, raising concerns about their robustness in real-world applications. Previous studies have sought to mitigate the impact of adversarial attacks by introducing adversarial perturbations into the training process, either implicitly or explicitly. While both strategies enhance robustness, they often incur high computational costs. In this work, we propose a simple yet effective add-on module that enhances the adversarial robustness of PLMs by removing instance-level principal components, without relying on conventional adversarial defences or perturbing the original training data. Our approach transforms the embedding space to approximate Gaussian properties, thereby reducing its susceptibility to adversarial perturbations while preserving semantic relationships. This transformation aligns embedding distributions in a way that minimises the impact of adversarial noise on decision boundaries, enhancing robustness without requiring adversarial examples or costly training-time augmentation. Evaluations on eight benchmark datasets show that our approach improves adversarial robustness while maintaining comparable before-attack accuracy to baselines, achieving a balanced trade-off between robustness and generalisation.", "date": "2025-07-29", "categories": []}
{"id": "2510.14231", "title": "When Flatness Does (Not) Guarantee Adversarial Robustness", "abstract": "Despite their empirical success, neural networks remain vulnerable to small, adversarial perturbations. A longstanding hypothesis suggests that flat minima, regions of low curvature in the loss landscape, offer increased robustness. While intuitive, this connection has remained largely informal and incomplete. By rigorously formalizing the relationship, we show this intuition is only partially correct: flatness implies local but not global adversarial robustness. To arrive at this result, we first derive a closed-form expression for relative flatness in the penultimate layer, and then show we can use this to constrain the variation of the loss in input space. This allows us to formally analyze the adversarial robustness of the entire network. We then show that to maintain robustness beyond a local neighborhood, the loss needs to curve sharply away from the data manifold. We validate our theoretical predictions empirically across architectures and datasets, uncovering the geometric structure that governs adversarial vulnerability, and linking flatness to model confidence: adversarial examples often lie in large, flat regions where the model is confidently wrong. Our results challenge simplified views of flatness and provide a nuanced understanding of its role in robustness.", "date": "2025-10-16", "categories": []}
{"id": "2510.14460", "title": "Structured Universal Adversarial Attacks on Object Detection for Video Sequences", "abstract": "Video-based object detection plays a vital role in safety-critical applications. While deep learning-based object detectors have achieved impressive performance, they remain vulnerable to adversarial attacks, particularly those involving universal perturbations. In this work, we propose a minimally distorted universal adversarial attack tailored for video object detection, which leverages nuclear norm regularization to promote structured perturbations concentrated in the background. To optimize this formulation efficiently, we employ an adaptive, optimistic exponentiated gradient method that enhances both scalability and convergence. Our results demonstrate that the proposed attack outperforms both low-rank projected gradient descent and Frank-Wolfe based attacks in effectiveness while maintaining high stealthiness. All code and data are publicly available at https://github.com/jsve96/AO-Exp-Attack.", "date": "2025-10-16", "categories": []}
{"id": "2510.14470", "title": "Stealthy Dual-Trigger Backdoors: Attacking Prompt Tuning in LM-Empowered Graph Foundation Models", "abstract": "The emergence of graph foundation models (GFMs), particularly those incorporating language models (LMs), has revolutionized graph learning and demonstrated remarkable performance on text-attributed graphs (TAGs). However, compared to traditional GNNs, these LM-empowered GFMs introduce unique security vulnerabilities during the unsecured prompt tuning phase that remain understudied in current research. Through empirical investigation, we reveal a significant performance degradation in traditional graph backdoor attacks when operating in attribute-inaccessible constrained TAG systems without explicit trigger node attribute optimization. To address this, we propose a novel dual-trigger backdoor attack framework that operates at both text-level and struct-level, enabling effective attacks without explicit optimization of trigger node text attributes through the strategic utilization of a pre-established text pool. Extensive experimental evaluations demonstrate that our attack maintains superior clean accuracy while achieving outstanding attack success rates, including scenarios with highly concealed single-trigger nodes. Our work highlights critical backdoor risks in web-deployed LM-empowered GFMs and contributes to the development of more robust supervision mechanisms for open-source platforms in the era of foundation models.", "date": "2025-10-16", "categories": []}
{"id": "2307.11334", "title": "Improving Transferability of Adversarial Examples via Bayesian Attacks", "abstract": "The transferability of adversarial examples allows for the attack on unknown deep neural networks (DNNs), posing a serious threat to many applications and attracting great attention. In this paper, we improve the transferability of adversarial examples by incorporating the Bayesian formulation into both the model parameters and model input, enabling their joint diversification. We demonstrate that combination of Bayesian formulations for both the model input and model parameters yields significant improvements in transferability. By introducing advanced approximations of the posterior distribution over the model input, adversarial transferability achieves further enhancement, surpassing all state-of-the-arts when attacking without model fine-tuning. Additionally, we propose a principled approach to fine-tune model parameters within this Bayesian framework. Extensive experiments demonstrate that our method achieves a new state-of-the-art in transfer-based attacks, significantly improving the average success rate on ImageNet and CIFAR-10. Code at: https://github.com/qizhangli/MoreBayesian-jrnl.", "date": "2023-07-21", "categories": []}
{"id": "2509.16494", "title": "Can an Individual Manipulate the Collective Decisions of Multi-Agents?", "abstract": "Individual Large Language Models (LLMs) have demonstrated significant capabilities across various domains, such as healthcare and law. Recent studies also show that coordinated multi-agent systems exhibit enhanced decision-making and reasoning abilities through collaboration. However, due to the vulnerabilities of individual LLMs and the difficulty of accessing all agents in a multi-agent system, a key question arises: If attackers only know one agent, could they still generate adversarial samples capable of misleading the collective decision? To explore this question, we formulate it as a game with incomplete information, where attackers know only one target agent and lack knowledge of the other agents in the system. With this formulation, we propose M-Spoiler, a framework that simulates agent interactions within a multi-agent system to generate adversarial samples. These samples are then used to manipulate the target agent in the target system, misleading the system's collaborative decision-making process. More specifically, M-Spoiler introduces a stubborn agent that actively aids in optimizing adversarial samples by simulating potential stubborn responses from agents in the target system. This enhances the effectiveness of the generated adversarial samples in misleading the system. Through extensive experiments across various tasks, our findings confirm the risks posed by the knowledge of an individual agent in multi-agent systems and demonstrate the effectiveness of our framework. We also explore several defense mechanisms, showing that our proposed attack framework remains more potent than baselines, underscoring the need for further research into defensive strategies.", "date": "2025-09-20", "categories": []}
{"id": "2510.05173", "title": "SafeGuider: Robust and Practical Content Safety Control for Text-to-Image Models", "abstract": "Text-to-image models have shown remarkable capabilities in generating high-quality images from natural language descriptions. However, these models are highly vulnerable to adversarial prompts, which can bypass safety measures and produce harmful content. Despite various defensive strategies, achieving robustness against attacks while maintaining practical utility in real-world applications remains a significant challenge. To address this issue, we first conduct an empirical study of the text encoder in the Stable Diffusion (SD) model, which is a widely used and representative text-to-image model. Our findings reveal that the [EOS] token acts as a semantic aggregator, exhibiting distinct distributional patterns between benign and adversarial prompts in its embedding space. Building on this insight, we introduce SafeGuider, a two-step framework designed for robust safety control without compromising generation quality. SafeGuider combines an embedding-level recognition model with a safety-aware feature erasure beam search algorithm. This integration enables the framework to maintain high-quality image generation for benign prompts while ensuring robust defense against both in-domain and out-of-domain attacks. SafeGuider demonstrates exceptional effectiveness in minimizing attack success rates, achieving a maximum rate of only 5.48\\% across various attack scenarios. Moreover, instead of refusing to generate or producing black images for unsafe prompts, SafeGuider generates safe and meaningful images, enhancing its practical utility. In addition, SafeGuider is not limited to the SD model and can be effectively applied to other text-to-image models, such as the Flux model, demonstrating its versatility and adaptability across different architectures. We hope that SafeGuider can shed some light on the practical deployment of secure text-to-image systems.", "date": "2025-10-05", "categories": []}
{"id": "2510.08979", "title": "Uncolorable Examples: Preventing Unauthorized AI Colorization via Perception-Aware Chroma-Restrictive Perturbation", "abstract": "AI-based colorization has shown remarkable capability in generating realistic color images from grayscale inputs. However, it poses risks of copyright infringement -- for example, the unauthorized colorization and resale of monochrome manga and films. Despite these concerns, no effective method currently exists to prevent such misuse. To address this, we introduce the first defensive paradigm, Uncolorable Examples, which embed imperceptible perturbations into grayscale images to invalidate unauthorized colorization. To ensure real-world applicability, we establish four criteria: effectiveness, imperceptibility, transferability, and robustness. Our method, Perception-Aware Chroma-Restrictive Perturbation (PAChroma), generates Uncolorable Examples that meet these four criteria by optimizing imperceptible perturbations with a Laplacian filter to preserve perceptual quality, and applying diverse input transformations during optimization to enhance transferability across models and robustness against common post-processing (e.g., compression). Experiments on ImageNet and Danbooru datasets demonstrate that PAChroma effectively degrades colorization quality while maintaining the visual appearance. This work marks the first step toward protecting visual content from illegitimate AI colorization, paving the way for copyright-aware defenses in generative media.", "date": "2025-10-10", "categories": []}
{"id": "2510.12939", "title": "Pruning Cannot Hurt Robustness: Certified Trade-offs in Reinforcement Learning", "abstract": "Reinforcement learning (RL) policies deployed in real-world environments must remain reliable under adversarial perturbations. At the same time, modern deep RL agents are heavily over-parameterized, raising costs and fragility concerns. While pruning has been shown to improve robustness in supervised learning, its role in adversarial RL remains poorly understood. We develop the first theoretical framework for certified robustness under pruning in state-adversarial Markov decision processes (SA-MDPs). For Gaussian and categorical policies with Lipschitz networks, we prove that element-wise pruning can only tighten certified robustness bounds; pruning never makes the policy less robust. Building on this, we derive a novel three-term regret decomposition that disentangles clean-task performance, pruning-induced performance loss, and robustness gains, exposing a fundamental performance--robustness frontier. Empirically, we evaluate magnitude and micro-pruning schedules on continuous-control benchmarks with strong policy-aware adversaries. Across tasks, pruning consistently uncovers reproducible ``sweet spots'' at moderate sparsity levels, where robustness improves substantially without harming - and sometimes even enhancing - clean performance. These results position pruning not merely as a compression tool but as a structural intervention for robust RL.", "date": "2025-10-14", "categories": []}
{"id": "2510.13262", "title": "SAJA: A State-Action Joint Attack Framework on Multi-Agent Deep Reinforcement Learning", "abstract": "Multi-Agent Deep Reinforcement Learning (MADRL) has shown potential for cooperative and competitive tasks such as autonomous driving and strategic gaming. However, models trained by MADRL are vulnerable to adversarial perturbations on states and actions. Therefore, it is essential to investigate the robustness of MADRL models from an attack perspective. Existing studies focus on either state-only attacks or action-only attacks, but do not consider how to effectively joint them. Simply combining state and action perturbations such as randomly perturbing states and actions does not exploit their potential synergistic effects. In this paper, we propose the State-Action Joint Attack (SAJA) framework that has a good synergistic effects. SAJA consists of two important phases: (1) In the state attack phase, a multi-step gradient ascent method utilizes both the actor network and the critic network to compute an adversarial state, and (2) in the action attack phase, based on the perturbed state, a second gradient ascent uses the critic network to craft the final adversarial action. Additionally, a heuristic regularizer measuring the distance between the perturbed actions and the original clean ones is added into the loss function to enhance the effectiveness of the critic's guidance. We evaluate SAJA in the Multi-Agent Particle Environment (MPE), demonstrating that (1) it outperforms and is more stealthy than state-only or action-only attacks, and (2) existing state or action defense methods cannot defend its attacks.", "date": "2025-10-15", "categories": []}
{"id": "2510.13322", "title": "Injection, Attack and Erasure: Revocable Backdoor Attacks via Machine Unlearning", "abstract": "Backdoor attacks pose a persistent security risk to deep neural networks (DNNs) due to their stealth and durability. While recent research has explored leveraging model unlearning mechanisms to enhance backdoor concealment, existing attack strategies still leave persistent traces that may be detected through static analysis. In this work, we introduce the first paradigm of revocable backdoor attacks, where the backdoor can be proactively and thoroughly removed after the attack objective is achieved. We formulate the trigger optimization in revocable backdoor attacks as a bilevel optimization problem: by simulating both backdoor injection and unlearning processes, the trigger generator is optimized to achieve a high attack success rate (ASR) while ensuring that the backdoor can be easily erased through unlearning. To mitigate the optimization conflict between injection and removal objectives, we employ a deterministic partition of poisoning and unlearning samples to reduce sampling-induced variance, and further apply the Projected Conflicting Gradient (PCGrad) technique to resolve the remaining gradient conflicts. Experiments on CIFAR-10 and ImageNet demonstrate that our method maintains ASR comparable to state-of-the-art backdoor attacks, while enabling effective removal of backdoor behavior after unlearning. This work opens a new direction for backdoor attack research and presents new challenges for the security of machine learning systems.", "date": "2025-10-15", "categories": []}
{"id": "2510.13361", "title": "Generalist++: A Meta-learning Framework for Mitigating Trade-off in Adversarial Training", "abstract": "Despite the rapid progress of neural networks, they remain highly vulnerable to adversarial examples, for which adversarial training (AT) is currently the most effective defense. While AT has been extensively studied, its practical applications expose two major limitations: natural accuracy tends to degrade significantly compared with standard training, and robustness does not transfer well across attacks crafted under different norm constraints. Unlike prior works that attempt to address only one issue within a single network, we propose to partition the overall generalization goal into multiple sub-tasks, each assigned to a dedicated base learner. By specializing in its designated objective, each base learner quickly becomes an expert in its field. In the later stages of training, we interpolate their parameters to form a knowledgeable global learner, while periodically redistributing the global parameters back to the base learners to prevent their optimization trajectories from drifting too far from the shared target. We term this framework Generalist and introduce three variants tailored to different application scenarios. Both theoretical analysis and extensive experiments demonstrate that Generalist achieves lower generalization error and significantly alleviates the trade-off problems compared with baseline methods. Our results suggest that Generalist provides a promising step toward developing fully robust classifiers in the future.", "date": "2025-10-15", "categories": []}
{"id": "2510.13570", "title": "Selective Adversarial Attacks on LLM Benchmarks", "abstract": "Benchmarking outcomes increasingly govern trust, selection, and deployment of LLMs, yet these evaluations remain vulnerable to semantically equivalent adversarial perturbations. Prior work on adversarial robustness in NLP has emphasized text attacks that affect many models equally, leaving open the question of whether it is possible to selectively degrade or enhance performance while minimally affecting other models. We formalize this problem and study selective adversarial attacks on MMLU - a widely used benchmark designed to measure a language model's broad general knowledge and reasoning ability across different subjects. Using canonical attacks integrated into TextAttack framework, we introduce a protocol for selectivity assessment, develop a custom constraint to increase selectivity of attacks and propose a surrogate-LLM pipeline that generates selective perturbations. Empirically, we find that selective adversarial attacks exist and can materially alter relative rankings, challenging the fairness, reproducibility, and transparency of leaderboard-driven evaluation. Our results motivate perturbation-aware reporting and robustness diagnostics for LLM evaluation and demonstrate that even subtle edits can shift comparative judgments.", "date": "2025-10-15", "categories": []}
{"id": "2510.13643", "title": "Towards Adversarial Robustness and Uncertainty Quantification in DINOv2-based Few-Shot Anomaly Detection", "abstract": "Foundation models such as DINOv2 have shown strong performance in few-shot anomaly detection, yet two key questions remain unexamined: (i) how susceptible are these detectors to adversarial perturbations; and (ii) how well do their anomaly scores reflect calibrated uncertainty? Building on AnomalyDINO, a training-free deep nearest-neighbor detector over DINOv2 features, we present one of the first systematic studies of adversarial attacks and uncertainty estimation in this setting. To enable white-box gradient attacks while preserving test-time behavior, we attach a lightweight linear head to frozen DINOv2 features only for crafting perturbations. Using this heuristic, we evaluate the impact of FGSM across the MVTec-AD and VisA datasets and observe consistent drops in F1, AUROC, AP, and G-mean, indicating that imperceptible perturbations can flip nearest-neighbor relations in feature space to induce confident misclassification. Complementing robustness, we probe reliability and find that raw anomaly scores are poorly calibrated, revealing a gap between confidence and correctness that limits safety-critical use. As a simple, strong baseline toward trustworthiness, we apply post-hoc Platt scaling to the anomaly scores for uncertainty estimation. The resulting calibrated posteriors yield significantly higher predictive entropy on adversarially perturbed inputs than on clean ones, enabling a practical flagging mechanism for attack detection while reducing calibration error (ECE). Our findings surface concrete vulnerabilities in DINOv2-based few-shot anomaly detectors and establish an evaluation protocol and baseline for robust, uncertainty-aware anomaly detection. We argue that adversarial robustness and principled uncertainty quantification are not optional add-ons but essential capabilities if anomaly detection systems are to be trustworthy and ready for real-world deployment.", "date": "2025-10-15", "categories": []}
{"id": "2510.11877", "title": "Robust Adversarial Reinforcement Learning in Stochastic Games via Sequence Modeling", "abstract": "The Transformer, a highly expressive architecture for sequence modeling, has recently been adapted to solve sequential decision-making, most notably through the Decision Transformer (DT), which learns policies by conditioning on desired returns. Yet, the adversarial robustness of reinforcement learning methods based on sequence modeling remains largely unexplored. Here we introduce the Conservative Adversarially Robust Decision Transformer (CART), to our knowledge the first framework designed to enhance the robustness of DT in adversarial stochastic games. We formulate the interaction between the protagonist and the adversary at each stage as a stage game, where the payoff is defined as the expected maximum value over subsequent states, thereby explicitly incorporating stochastic state transitions. By conditioning Transformer policies on the NashQ value derived from these stage games, CART generates policy that are simultaneously less exploitable (adversarially robust) and conservative to transition uncertainty. Empirically, CART achieves more accurate minimax value estimation and consistently attains superior worst-case returns across a range of adversarial stochastic games.", "date": "2025-10-13", "categories": []}
{"id": "2510.12233", "title": "Unveiling the Vulnerability of Graph-LLMs: An Interpretable Multi-Dimensional Adversarial Attack on TAGs", "abstract": "Graph Neural Networks (GNNs) have become a pivotal framework for modeling graph-structured data, enabling a wide range of applications from social network analysis to molecular chemistry. By integrating large language models (LLMs), text-attributed graphs (TAGs) enhance node representations with rich textual semantics, significantly boosting the expressive power of graph-based learning. However, this sophisticated synergy introduces critical vulnerabilities, as Graph-LLMs are susceptible to adversarial attacks on both their structural topology and textual attributes. Although specialized attack methods have been designed for each of these aspects, no work has yet unified them into a comprehensive approach. In this work, we propose the Interpretable Multi-Dimensional Graph Attack (IMDGA), a novel human-centric adversarial attack framework designed to orchestrate multi-level perturbations across both graph structure and textual features. IMDGA utilizes three tightly integrated modules to craft attacks that balance interpretability and impact, enabling a deeper understanding of Graph-LLM vulnerabilities. Through rigorous theoretical analysis and comprehensive empirical evaluations on diverse datasets and architectures, IMDGA demonstrates superior interpretability, attack effectiveness, stealthiness, and robustness compared to existing methods. By exposing critical weaknesses in TAG representation learning, this work uncovers a previously underexplored semantic dimension of vulnerability in Graph-LLMs, offering valuable insights for improving their resilience. Our code and resources are publicly available at https://anonymous.4open.science/r/IMDGA-7289.", "date": "2025-10-14", "categories": []}
{"id": "2510.12468", "title": "MS-GAGA: Metric-Selective Guided Adversarial Generation Attack", "abstract": "We present MS-GAGA (Metric-Selective Guided Adversarial Generation Attack), a two-stage framework for crafting transferable and visually imperceptible adversarial examples against deepfake detectors in black-box settings. In Stage 1, a dual-stream attack module generates adversarial candidates: MNTD-PGD applies enhanced gradient calculations optimized for small perturbation budgets, while SG-PGD focuses perturbations on visually salient regions. This complementary design expands the adversarial search space and improves transferability across unseen models. In Stage 2, a metric-aware selection module evaluates candidates based on both their success against black-box models and their structural similarity (SSIM) to the original image. By jointly optimizing transferability and imperceptibility, MS-GAGA achieves up to 27% higher misclassification rates on unseen detectors compared to state-of-the-art attacks.", "date": "2025-10-14", "categories": []}
{"id": "2306.07033", "title": "When Vision Fails: Text Attacks Against ViT and OCR", "abstract": "Text-based machine learning models are vulnerable to an emerging class of Unicode-based adversarial examples capable of tricking a model into misreading text with potentially disastrous effects. The primary existing defense against these attacks is to preprocess potentially malicious text inputs using optical character recognition (OCR). In theory, OCR models will ignore any malicious Unicode characters and will extract the visually correct input to be fed to the model. In this work, we show that these visual defenses fail to prevent this type of attack. We use a genetic algorithm to generate visual adversarial examples (i.e., OCR outputs) in a black-box setting, demonstrating a highly effective novel attack that substantially reduces the accuracy of OCR and other visual models. Specifically, we use the Unicode functionality of combining characters (e.g., \\~n which combines the characters n and ~) to manipulate text inputs so that small visual perturbations appear when the text is displayed. We demonstrate the effectiveness of these attacks in the real world by creating adversarial examples against production models published by Meta, Microsoft, IBM, and Google. We additionally conduct a user study to establish that the model-fooling adversarial examples do not affect human comprehension of the text, showing that language models are uniquely vulnerable to this type of text attack.", "date": "2023-06-12", "categories": []}
{"id": "2505.10983", "title": "GenoArmory: A Unified Evaluation Framework for Adversarial Attacks on Genomic Foundation Models", "abstract": "We propose the first unified adversarial attack benchmark for Genomic Foundation Models (GFMs), named GenoArmory. Unlike existing GFM benchmarks, GenoArmory offers the first comprehensive evaluation framework to systematically assess the vulnerability of GFMs to adversarial attacks. Methodologically, we evaluate the adversarial robustness of five state-of-the-art GFMs using four widely adopted attack algorithms and three defense strategies. Importantly, our benchmark provides an accessible and comprehensive framework to analyze GFM vulnerabilities with respect to model architecture, quantization schemes, and training datasets. Additionally, we introduce GenoAdv, a new adversarial sample dataset designed to improve GFM safety. Empirically, classification models exhibit greater robustness to adversarial perturbations compared to generative models, highlighting the impact of task type on model vulnerability. Moreover, adversarial attacks frequently target biologically significant genomic regions, suggesting that these models effectively capture meaningful sequence features.", "date": "2025-05-16", "categories": []}
{"id": "2506.06992", "title": "Boosting Adversarial Transferability via Commonality-Oriented Gradient Optimization", "abstract": "Exploring effective and transferable adversarial examples is vital for understanding the characteristics and mechanisms of Vision Transformers (ViTs). However, adversarial examples generated from surrogate models often exhibit weak transferability in black-box settings due to overfitting. Existing methods improve transferability by diversifying perturbation inputs or applying uniform gradient regularization within surrogate models, yet they have not fully leveraged the shared and unique features of surrogate models trained on the same task, leading to suboptimal transfer performance. Therefore, enhancing perturbations of common information shared by surrogate models and suppressing those tied to individual characteristics offers an effective way to improve transferability. Accordingly, we propose a commonality-oriented gradient optimization strategy (COGO) consisting of two components: Commonality Enhancement (CE) and Individuality Suppression (IS). CE perturbs the mid-to-low frequency regions, leveraging the fact that ViTs trained on the same dataset tend to rely more on mid-to-low frequency information for classification. IS employs adaptive thresholds to evaluate the correlation between backpropagated gradients and model individuality, assigning weights to gradients accordingly. Extensive experiments demonstrate that COGO significantly improves the transfer success rates of adversarial attacks, outperforming current state-of-the-art methods.", "date": "2025-06-08", "categories": []}
{"id": "2506.15685", "title": "Ignition Phase : Standard Training for Fast Adversarial Robustness", "abstract": "Adversarial Training (AT) is a cornerstone defense, but many variants overlook foundational feature representations by primarily focusing on stronger attack generation. We introduce Adversarial Evolution Training (AET), a simple yet powerful framework that strategically prepends an Empirical Risk Minimization (ERM) phase to conventional AT. We hypothesize this initial ERM phase cultivates a favorable feature manifold, enabling more efficient and effective robustness acquisition. Empirically, AET achieves comparable or superior robustness more rapidly, improves clean accuracy, and cuts training costs by 8-25\\%. Its effectiveness is shown across multiple datasets, architectures, and when augmenting established AT methods. Our findings underscore the impact of feature pre-conditioning via standard training for developing more efficient, principled robust defenses. Code is available in the supplementary material.", "date": "2025-05-25", "categories": []}
{"id": "2510.09663", "title": "Adversarial-Resilient RF Fingerprinting: A CNN-GAN Framework for Rogue Transmitter Detection", "abstract": "Radio Frequency Fingerprinting (RFF) has evolved as an effective solution for authenticating devices by leveraging the unique imperfections in hardware components involved in the signal generation process. In this work, we propose a Convolutional Neural Network (CNN) based framework for detecting rogue devices and identifying genuine ones using softmax probability thresholding. We emulate an attack scenario in which adversaries attempt to mimic the RF characteristics of genuine devices by training a Generative Adversarial Network (GAN) using In-phase and Quadrature (IQ) samples from genuine devices. The proposed approach is verified using IQ samples collected from ten different ADALM-PLUTO Software Defined Radios (SDRs), with seven devices considered genuine, two as rogue, and one used for validation to determine the threshold.", "date": "2025-10-07", "categories": []}
{"id": "2510.09699", "title": "VisualDAN: Exposing Vulnerabilities in VLMs with Visual-Driven DAN Commands", "abstract": "Vision-Language Models (VLMs) have garnered significant attention for their remarkable ability to interpret and generate multimodal content. However, securing these models against jailbreak attacks continues to be a substantial challenge. Unlike text-only models, VLMs integrate additional modalities, introducing novel vulnerabilities such as image hijacking, which can manipulate the model into producing inappropriate or harmful responses. Drawing inspiration from text-based jailbreaks like the \"Do Anything Now\" (DAN) command, this work introduces VisualDAN, a single adversarial image embedded with DAN-style commands. Specifically, we prepend harmful corpora with affirmative prefixes (e.g., \"Sure, I can provide the guidance you need\") to trick the model into responding positively to malicious queries. The adversarial image is then trained on these DAN-inspired harmful texts and transformed into the text domain to elicit malicious outputs. Extensive experiments on models such as MiniGPT-4, MiniGPT-v2, InstructBLIP, and LLaVA reveal that VisualDAN effectively bypasses the safeguards of aligned VLMs, forcing them to execute a broad range of harmful instructions that severely violate ethical standards. Our results further demonstrate that even a small amount of toxic content can significantly amplify harmful outputs once the model's defenses are compromised. These findings highlight the urgent need for robust defenses against image-based attacks and offer critical insights for future research into the alignment and security of VLMs.", "date": "2025-10-09", "categories": []}
{"id": "2510.10937", "title": "Neutral Agent-based Adversarial Policy Learning against Deep Reinforcement Learning in Multi-party Open Systems", "abstract": "Reinforcement learning (RL) has been an important machine learning paradigm for solving long-horizon sequential decision-making problems under uncertainty. By integrating deep neural networks (DNNs) into the RL framework, deep reinforcement learning (DRL) has emerged, which achieved significant success in various domains. However, the integration of DNNs also makes it vulnerable to adversarial attacks. Existing adversarial attack techniques mainly focus on either directly manipulating the environment with which a victim agent interacts or deploying an adversarial agent that interacts with the victim agent to induce abnormal behaviors. While these techniques achieve promising results, their adoption in multi-party open systems remains limited due to two major reasons: impractical assumption of full control over the environment and dependent on interactions with victim agents.\n  To enable adversarial attacks in multi-party open systems, in this paper, we redesigned an adversarial policy learning approach that can mislead well-trained victim agents without requiring direct interactions with these agents or full control over their environments. Particularly, we propose a neutral agent-based approach across various task scenarios in multi-party open systems. While the neutral agents seemingly are detached from the victim agents, indirectly influence them through the shared environment. We evaluate our proposed method on the SMAC platform based on Starcraft II and the autonomous driving simulation platform Highway-env. The experimental results demonstrate that our method can launch general and effective adversarial attacks in multi-party open systems.", "date": "2025-10-13", "categories": []}
{"id": "2510.11018", "title": "The Easy Path to Robustness: Coreset Selection using Sample Hardness", "abstract": "Designing adversarially robust models from a data-centric perspective requires understanding which input samples are most crucial for learning resilient features. While coreset selection provides a mechanism for efficient training on data subsets, current algorithms are designed for clean accuracy and fall short in preserving robustness. To address this, we propose a framework linking a sample's adversarial vulnerability to its \\textit{hardness}, which we quantify using the average input gradient norm (AIGN) over training. We demonstrate that \\textit{easy} samples (with low AIGN) are less vulnerable and occupy regions further from the decision boundary. Leveraging this insight, we present EasyCore, a coreset selection algorithm that retains only the samples with low AIGN for training. We empirically show that models trained on EasyCore-selected data achieve significantly higher adversarial accuracy than those trained with competing coreset methods under both standard and adversarial training. As AIGN is a model-agnostic dataset property, EasyCore is an efficient and widely applicable data-centric method for improving adversarial robustness. We show that EasyCore achieves up to 7\\% and 5\\% improvement in adversarial accuracy under standard training and TRADES adversarial training, respectively, compared to existing coreset methods.", "date": "2025-10-13", "categories": []}
{"id": "2510.11096", "title": "CoDefend: Cross-Modal Collaborative Defense via Diffusion Purification and Prompt Optimization", "abstract": "Multimodal Large Language Models (MLLMs) have achieved remarkable success in tasks such as image captioning, visual question answering, and cross-modal reasoning by integrating visual and textual modalities. However, their multimodal nature also exposes them to adversarial threats, where attackers can perturb either modality or both jointly to induce harmful, misleading, or policy violating outputs. Existing defense strategies, such as adversarial training and input purification, face notable limitations: adversarial training typically improves robustness only against known attacks while incurring high computational costs, whereas conventional purification approaches often suffer from degraded image quality and insufficient generalization to complex multimodal tasks.\n  In this work, we focus on defending the visual modality, which frequently serves as the primary entry point for adversarial manipulation. We propose a supervised diffusion based denoising framework that leverages paired adversarial clean image datasets to fine-tune diffusion models with directional, task specific guidance. Unlike prior unsupervised purification methods such as DiffPure, our approach achieves higher quality reconstructions while significantly improving defense robustness in multimodal tasks. Furthermore, we incorporate prompt optimization as a complementary defense mechanism, enhancing resistance against diverse and unseen attack strategies.\n  Extensive experiments on image captioning and visual question answering demonstrate that our method not only substantially improves robustness but also exhibits strong transferability to unknown adversarial attacks. These results highlight the effectiveness of supervised diffusion based denoising for multimodal defense, paving the way for more reliable and secure deployment of MLLMs in real world applications.", "date": "2025-10-13", "categories": []}
{"id": "2510.11398", "title": "Living Off the LLM: How LLMs Will Change Adversary Tactics", "abstract": "In living off the land attacks, malicious actors use legitimate tools and processes already present on a system to avoid detection. In this paper, we explore how the on-device LLMs of the future will become a security concern as threat actors integrate LLMs into their living off the land attack pipeline and ways the security community may mitigate this threat.", "date": "2025-10-13", "categories": []}
{"id": "2506.03355", "title": "Robustness in Both Domains: CLIP Needs a Robust Text Encoder", "abstract": "Adversarial input attacks can cause a significant shift of CLIP embeddings. This can affect the downstream robustness of models incorporating CLIP in the pipeline, such as text-to-image generative models or large vision language models. While some efforts have been done towards making the CLIP image encoders robust, the robustness of text encoders remains unexplored. In this work, we cover this gap in the literature. We propose LEAF: an efficient adversarial finetuning method for the text domain, with the ability to scale to large CLIP models. Our models significantly improve the zero-shot adversarial accuracy in the text domain, while maintaining the vision performance provided by robust image encoders. When combined with text-to-image diffusion models, we can improve the generation quality under adversarial noise. In multimodal retrieval tasks, LEAF improves the recall under adversarial noise over standard CLIP models. Finally, we show that robust text encoders facilitate better reconstruction of input text from its embedding via direct optimization. We open-source our code ( https://github.com/LIONS-EPFL/LEAF ) and models ( https://huggingface.co/LEAF-CLIP ).", "date": "2025-06-03", "categories": []}
{"id": "2510.08761", "title": "SAFER-AiD: Saccade-Assisted Foveal-peripheral vision Enhanced Reconstruction for Adversarial Defense", "abstract": "Adversarial attacks significantly challenge the safe deployment of deep learning models, particularly in real-world applications. Traditional defenses often rely on computationally intensive optimization (e.g., adversarial training or data augmentation) to improve robustness, whereas the human visual system achieves inherent robustness to adversarial perturbations through evolved biological mechanisms. We hypothesize that attention guided non-homogeneous sparse sampling and predictive coding plays a key role in this robustness. To test this hypothesis, we propose a novel defense framework incorporating three key biological mechanisms: foveal-peripheral processing, saccadic eye movements, and cortical filling-in. Our approach employs reinforcement learning-guided saccades to selectively capture multiple foveal-peripheral glimpses, which are integrated into a reconstructed image before classification. This biologically inspired preprocessing effectively mitigates adversarial noise, preserves semantic integrity, and notably requires no retraining or fine-tuning of downstream classifiers, enabling seamless integration with existing systems. Experiments on the ImageNet dataset demonstrate that our method improves system robustness across diverse classifiers and attack types, while significantly reducing training overhead compared to both biologically and non-biologically inspired defense techniques.", "date": "2025-10-09", "categories": []}
{"id": "2510.08922", "title": "SegTrans: Transferable Adversarial Examples for Segmentation Models", "abstract": "Segmentation models exhibit significant vulnerability to adversarial examples in white-box settings, but existing adversarial attack methods often show poor transferability across different segmentation models. While some researchers have explored transfer-based adversarial attack (i.e., transfer attack) methods for segmentation models, the complex contextual dependencies within these models and the feature distribution gaps between surrogate and target models result in unsatisfactory transfer success rates. To address these issues, we propose SegTrans, a novel transfer attack framework that divides the input sample into multiple local regions and remaps their semantic information to generate diverse enhanced samples. These enhanced samples replace the original ones for perturbation optimization, thereby improving the transferability of adversarial examples across different segmentation models. Unlike existing methods, SegTrans only retains local semantic information from the original input, rather than using global semantic information to optimize perturbations. Extensive experiments on two benchmark datasets, PASCAL VOC and Cityscapes, four different segmentation models, and three backbone networks show that SegTrans significantly improves adversarial transfer success rates without introducing additional computational overhead. Compared to the current state-of-the-art methods, SegTrans achieves an average increase of 8.55% in transfer attack success rate and improves computational efficiency by more than 100%.", "date": "2025-10-10", "categories": []}
{"id": "2510.09105", "title": "MemLoss: Enhancing Adversarial Training with Recycling Adversarial Examples", "abstract": "In this paper, we propose a new approach called MemLoss to improve the adversarial training of machine learning models. MemLoss leverages previously generated adversarial examples, referred to as 'Memory Adversarial Examples,' to enhance model robustness and accuracy without compromising performance on clean data. By using these examples across training epochs, MemLoss provides a balanced improvement in both natural accuracy and adversarial robustness. Experimental results on multiple datasets, including CIFAR-10, demonstrate that our method achieves better accuracy compared to existing adversarial training methods while maintaining strong robustness against attacks.", "date": "2025-10-10", "categories": []}
{"id": "2510.09269", "title": "Goal-oriented Backdoor Attack against Vision-Language-Action Models via Physical Objects", "abstract": "Recent advances in vision-language-action (VLA) models have greatly improved embodied AI, enabling robots to follow natural language instructions and perform diverse tasks. However, their reliance on uncurated training datasets raises serious security concerns. Existing backdoor attacks on VLAs mostly assume white-box access and result in task failures instead of enforcing specific actions. In this work, we reveal a more practical threat: attackers can manipulate VLAs by simply injecting physical objects as triggers into the training dataset. We propose goal-oriented backdoor attacks (GoBA), where the VLA behaves normally in the absence of physical triggers but executes predefined and goal-oriented actions in the presence of physical triggers. Specifically, based on a popular VLA benchmark LIBERO, we introduce BadLIBERO that incorporates diverse physical triggers and goal-oriented backdoor actions. In addition, we propose a three-level evaluation that categorizes the victim VLA's actions under GoBA into three states: nothing to do, try to do, and success to do. Experiments show that GoBA enables the victim VLA to successfully achieve the backdoor goal in 97 percentage of inputs when the physical trigger is present, while causing zero performance degradation on clean inputs. Finally, by investigating factors related to GoBA, we find that the action trajectory and trigger color significantly influence attack performance, while trigger size has surprisingly little effect. The code and BadLIBERO dataset are accessible via the project page at https://goba-attack.github.io/.", "date": "2025-10-10", "categories": []}
{"id": "2406.05927", "title": "MeanSparse: Post-Training Robustness Enhancement Through Mean-Centered Feature Sparsification", "abstract": "We present a simple yet effective method to improve the robustness of both Convolutional and attention-based Neural Networks against adversarial examples by post-processing an adversarially trained model. Our technique, MeanSparse, cascades the activation functions of a trained model with novel operators that sparsify mean-centered feature vectors. This is equivalent to reducing feature variations around the mean, and we show that such reduced variations merely affect the model's utility, yet they strongly attenuate the adversarial perturbations and decrease the attacker's success rate. Our experiments show that, when applied to the top models in the RobustBench leaderboard, MeanSparse achieves a new robustness record of 75.28% (from 73.71%), 44.78% (from 42.67%) and 62.12% (from 59.56%) on CIFAR-10, CIFAR-100 and ImageNet, respectively, in terms of AutoAttack accuracy. Code is available at https://github.com/SPIN-UMass/MeanSparse", "date": "2024-06-09", "categories": []}
{"id": "2506.11113", "title": "Breaking the Reviewer: Assessing the Vulnerability of Large Language Models in Automated Peer Review Under Textual Adversarial Attacks", "abstract": "Peer review is essential for maintaining academic quality, but the increasing volume of submissions places a significant burden on reviewers. Large language models (LLMs) offer potential assistance in this process, yet their susceptibility to textual adversarial attacks raises reliability concerns. This paper investigates the robustness of LLMs used as automated reviewers in the presence of such attacks. We focus on three key questions: (1) The effectiveness of LLMs in generating reviews compared to human reviewers. (2) The impact of adversarial attacks on the reliability of LLM-generated reviews. (3) Challenges and potential mitigation strategies for LLM-based review. Our evaluation reveals significant vulnerabilities, as text manipulations can distort LLM assessments. We offer a comprehensive evaluation of LLM performance in automated peer reviewing and analyze its robustness against adversarial attacks. Our findings emphasize the importance of addressing adversarial risks to ensure AI strengthens, rather than compromises, the integrity of scholarly communication.", "date": "2025-06-08", "categories": []}
{"id": "2508.08559", "title": "Multi-Target Backdoor Attacks Against Speaker Recognition", "abstract": "In this work, we propose a multi-target backdoor attack against speaker identification using position-independent clicking sounds as triggers. Unlike previous single-target approaches, our method targets up to 50 speakers simultaneously, achieving success rates of up to 95.04%. To simulate more realistic attack conditions, we vary the signal-to-noise ratio between speech and trigger, demonstrating a trade-off between stealth and effectiveness. We further extend the attack to the speaker verification task by selecting the most similar training speaker - based on cosine similarity - as a proxy target. The attack is most effective when target and enrolled speaker pairs are highly similar, reaching success rates of up to 90% in such cases.", "date": "2025-08-12", "categories": []}
{"id": "2509.00088", "title": "AEGIS : Automated Co-Evolutionary Framework for Guarding Prompt Injections Schema", "abstract": "Prompt injection attacks pose a significant challenge to the safe deployment of Large Language Models (LLMs) in real-world applications. While prompt-based detection offers a lightweight and interpretable defense strategy, its effectiveness has been hindered by the need for manual prompt engineering. To address this issue, we propose AEGIS , an Automated co-Evolutionary framework for Guarding prompt Injections Schema. Both attack and defense prompts are iteratively optimized against each other using a gradient-like natural language prompt optimization technique. This framework enables both attackers and defenders to autonomously evolve via a Textual Gradient Optimization (TGO) module, leveraging feedback from an LLM-guided evaluation loop. We evaluate our system on a real-world assignment grading dataset of prompt injection attacks and demonstrate that our method consistently outperforms existing baselines, achieving superior robustness in both attack success and detection. Specifically, the attack success rate (ASR) reaches 1.0, representing an improvement of 0.26 over the baseline. For detection, the true positive rate (TPR) improves by 0.23 compared to the previous best work, reaching 0.84, and the true negative rate (TNR) remains comparable at 0.89. Ablation studies confirm the importance of co-evolution, gradient buffering, and multi-objective optimization. We also confirm that this framework is effective in different LLMs. Our results highlight the promise of adversarial training as a scalable and effective approach for guarding prompt injections.", "date": "2025-08-27", "categories": []}
{"id": "2206.09098", "title": "Existence and Minimax Theorems for Adversarial Surrogate Risks in Binary Classification", "abstract": "Adversarial training is one of the most popular methods for training methods robust to adversarial attacks, however, it is not well-understood from a theoretical perspective. We prove and existence, regularity, and minimax theorems for adversarial surrogate risks. Our results explain some empirical observations on adversarial robustness from prior work and suggest new directions in algorithm development. Furthermore, our results extend previously known existence and minimax theorems for the adversarial classification risk to surrogate risks.", "date": "2022-06-18", "categories": []}
{"id": "2305.09956", "title": "The Adversarial Consistency of Surrogate Risks for Binary Classification", "abstract": "We study the consistency of surrogate risks for robust binary classification. It is common to learn robust classifiers by adversarial training, which seeks to minimize the expected $0$-$1$ loss when each example can be maliciously corrupted within a small ball. We give a simple and complete characterization of the set of surrogate loss functions that are \\emph{consistent}, i.e., that can replace the $0$-$1$ loss without affecting the minimizing sequences of the original adversarial risk, for any data distribution. We also prove a quantitative version of adversarial consistency for the $\\rho$-margin loss. Our results reveal that the class of adversarially consistent surrogates is substantially smaller than in the standard setting, where many common surrogates are known to be consistent.", "date": "2023-05-17", "categories": []}
{"id": "2408.01541", "title": "Guardians of Image Quality: Benchmarking Defenses Against Adversarial Attacks on Image Quality Metrics", "abstract": "In the field of Image Quality Assessment (IQA), the adversarial robustness of the metrics poses a critical concern. This paper presents a comprehensive benchmarking study of various defense mechanisms in response to the rise in adversarial attacks on IQA. We systematically evaluate 25 defense strategies, including adversarial purification, adversarial training, and certified robustness methods. We applied 14 adversarial attack algorithms of various types in both non-adaptive and adaptive settings and tested these defenses against them. We analyze the differences between defenses and their applicability to IQA tasks, considering that they should preserve IQA scores and image quality. The proposed benchmark aims to guide future developments and accepts submissions of new methods, with the latest results available online: https://videoprocessing.ai/benchmarks/iqa-defenses.html.", "date": "2024-08-02", "categories": []}
{"id": "2412.02270", "title": "Sustainable Self-evolution Adversarial Training", "abstract": "With the wide application of deep neural network models in various computer vision tasks, there has been a proliferation of adversarial example generation strategies aimed at deeply exploring model security. However, existing adversarial training defense models, which rely on single or limited types of attacks under a one-time learning process, struggle to adapt to the dynamic and evolving nature of attack methods. Therefore, to achieve defense performance improvements for models in long-term applications, we propose a novel Sustainable Self-Evolution Adversarial Training (SSEAT) framework. Specifically, we introduce a continual adversarial defense pipeline to realize learning from various kinds of adversarial examples across multiple stages. Additionally, to address the issue of model catastrophic forgetting caused by continual learning from ongoing novel attacks, we propose an adversarial data replay module to better select more diverse and key relearning data. Furthermore, we design a consistency regularization strategy to encourage current defense models to learn more from previously trained ones, guiding them to retain more past knowledge and maintain accuracy on clean samples. Extensive experiments have been conducted to verify the efficacy of the proposed SSEAT defense method, which demonstrates superior defense performance and classification accuracy compared to competitors.Code is available at https://github.com/aup520/SSEAT", "date": "2024-12-03", "categories": []}
{"id": "2501.19040", "title": "Towards the Worst-case Robustness of Large Language Models", "abstract": "Recent studies have revealed the vulnerability of large language models to adversarial attacks, where adversaries craft specific input sequences to induce harmful, violent, private, or incorrect outputs. In this work, we study their worst-case robustness, i.e., whether an adversarial example exists that leads to such undesirable outputs. We upper bound the worst-case robustness using stronger white-box attacks, indicating that most current deterministic defenses achieve nearly 0\\% worst-case robustness. We propose a general tight lower bound for randomized smoothing using fractional knapsack solvers or 0-1 knapsack solvers, and using them to bound the worst-case robustness of all stochastic defenses. Based on these solvers, we provide theoretical lower bounds for several previous empirical defenses. For example, we certify the robustness of a specific case, smoothing using a uniform kernel, against \\textit{any possible attack} with an average $\\ell_0$ perturbation of 2.02 or an average suffix length of 6.41.", "date": "2025-01-31", "categories": []}
{"id": "2506.09348", "title": "Adversarial Surrogate Risk Bounds for Binary Classification", "abstract": "A central concern in classification is the vulnerability of machine learning models to adversarial attacks. Adversarial training is one of the most popular techniques for training robust classifiers, which involves minimizing an adversarial surrogate risk. Recent work has characterized the conditions under which any sequence minimizing the adversarial surrogate risk also minimizes the adversarial classification risk in the binary setting, a property known as adversarial consistency. However, these results do not address the rate at which the adversarial classification risk approaches its optimal value along such a sequence. This paper provides surrogate risk bounds that quantify that convergence rate.", "date": "2025-06-11", "categories": []}
{"id": "2510.05379", "title": "AutoDAN-Reasoning: Enhancing Strategies Exploration based Jailbreak Attacks with Test-Time Scaling", "abstract": "Recent advancements in jailbreaking large language models (LLMs), such as AutoDAN-Turbo, have demonstrated the power of automated strategy discovery. AutoDAN-Turbo employs a lifelong learning agent to build a rich library of attack strategies from scratch. While highly effective, its test-time generation process involves sampling a strategy and generating a single corresponding attack prompt, which may not fully exploit the potential of the learned strategy library. In this paper, we propose to further improve the attack performance of AutoDAN-Turbo through test-time scaling. We introduce two distinct scaling methods: Best-of-N and Beam Search. The Best-of-N method generates N candidate attack prompts from a sampled strategy and selects the most effective one based on a scorer model. The Beam Search method conducts a more exhaustive search by exploring combinations of strategies from the library to discover more potent and synergistic attack vectors. According to the experiments, the proposed methods significantly boost performance, with Beam Search increasing the attack success rate by up to 15.6 percentage points on Llama-3.1-70B-Instruct and achieving a nearly 60% relative improvement against the highly robust GPT-o4-mini compared to the vanilla method.", "date": "2025-10-06", "categories": []}
{"id": "2510.06397", "title": "Geometry-Aware Backdoor Attacks: Leveraging Curvature in Hyperbolic Embeddings", "abstract": "Non-Euclidean foundation models increasingly place representations in curved spaces such as hyperbolic geometry. We show that this geometry creates a boundary-driven asymmetry that backdoor triggers can exploit. Near the boundary, small input changes appear subtle to standard input-space detectors but produce disproportionately large shifts in the model's representation space. Our analysis formalizes this effect and also reveals a limitation for defenses: methods that act by pulling points inward along the radius can suppress such triggers, but only by sacrificing useful model sensitivity in that same direction. Building on these insights, we propose a simple geometry-adaptive trigger and evaluate it across tasks and architectures. Empirically, attack success increases toward the boundary, whereas conventional detectors weaken, mirroring the theoretical trends. Together, these results surface a geometry-specific vulnerability in non-Euclidean models and offer analysis-backed guidance for designing and understanding the limits of defenses.", "date": "2025-10-07", "categories": []}
{"id": "2510.06525", "title": "Text-to-Image Models Leave Identifiable Signatures: Implications for Leaderboard Security", "abstract": "Generative AI leaderboards are central to evaluating model capabilities, but remain vulnerable to manipulation. Among key adversarial objectives is rank manipulation, where an attacker must first deanonymize the models behind displayed outputs -- a threat previously demonstrated and explored for large language models (LLMs). We show that this problem can be even more severe for text-to-image leaderboards, where deanonymization is markedly easier. Using over 150,000 generated images from 280 prompts and 19 diverse models spanning multiple organizations, architectures, and sizes, we demonstrate that simple real-time classification in CLIP embedding space identifies the generating model with high accuracy, even without prompt control or historical data. We further introduce a prompt-level separability metric and identify prompts that enable near-perfect deanonymization. Our results indicate that rank manipulation in text-to-image leaderboards is easier than previously recognized, underscoring the need for stronger defenses.", "date": "2025-10-07", "categories": []}
{"id": "2510.06629", "title": "Unsupervised Backdoor Detection and Mitigation for Spiking Neural Networks", "abstract": "Spiking Neural Networks (SNNs) have gained increasing attention for their superior energy efficiency compared to Artificial Neural Networks (ANNs). However, their security aspects, particularly under backdoor attacks, have received limited attention. Existing defense methods developed for ANNs perform poorly or can be easily bypassed in SNNs due to their event-driven and temporal dependencies. This paper identifies the key blockers that hinder traditional backdoor defenses in SNNs and proposes an unsupervised post-training detection framework, Temporal Membrane Potential Backdoor Detection (TMPBD), to overcome these challenges. TMPBD leverages the maximum margin statistics of temporal membrane potential (TMP) in the final spiking layer to detect target labels without any attack knowledge or data access. We further introduce a robust mitigation mechanism, Neural Dendrites Suppression Backdoor Mitigation (NDSBM), which clamps dendritic connections between early convolutional layers to suppress malicious neurons while preserving benign behaviors, guided by TMP extracted from a small, clean, unlabeled dataset. Extensive experiments on multiple neuromorphic benchmarks and state-of-the-art input-aware dynamic trigger attacks demonstrate that TMPBD achieves 100% detection accuracy, while NDSBM reduces the attack success rate from 100% to 8.44%, and to 2.81% when combined with detection, without degrading clean accuracy.", "date": "2025-10-08", "categories": []}
{"id": "2510.06732", "title": "Are LLMs Reliable Rankers? Rank Manipulation via Two-Stage Token Optimization", "abstract": "Large language models (LLMs) are increasingly used as rerankers in information retrieval, yet their ranking behavior can be steered by small, natural-sounding prompts. To expose this vulnerability, we present Rank Anything First (RAF), a two-stage token optimization method that crafts concise textual perturbations to consistently promote a target item in LLM-generated rankings while remaining hard to detect. Stage 1 uses Greedy Coordinate Gradient to shortlist candidate tokens at the current position by combining the gradient of the rank-target with a readability score; Stage 2 evaluates those candidates under exact ranking and readability losses using an entropy-based dynamic weighting scheme, and selects a token via temperature-controlled sampling. RAF generates ranking-promoting prompts token-by-token, guided by dual objectives: maximizing ranking effectiveness and preserving linguistic naturalness. Experiments across multiple LLMs show that RAF significantly boosts the rank of target items using naturalistic language, with greater robustness than existing methods in both promoting target items and maintaining naturalness. These findings underscore a critical security implication: LLM-based reranking is inherently susceptible to adversarial manipulation, raising new challenges for the trustworthiness and robustness of modern retrieval systems. Our code is available at: https://github.com/glad-lab/RAF.", "date": "2025-10-08", "categories": []}
{"id": "2510.06952", "title": "OBJVanish: Physically Realizable Text-to-3D Adv. Generation of LiDAR-Invisible Objects", "abstract": "LiDAR-based 3D object detectors are fundamental to autonomous driving, where failing to detect objects poses severe safety risks. Developing effective 3D adversarial attacks is essential for thoroughly testing these detection systems and exposing their vulnerabilities before real-world deployment. However, existing adversarial attacks that add optimized perturbations to 3D points have two critical limitations: they rarely cause complete object disappearance and prove difficult to implement in physical environments. We introduce the text-to-3D adversarial generation method, a novel approach enabling physically realizable attacks that can generate 3D models of objects truly invisible to LiDAR detectors and be easily realized in the real world. Specifically, we present the first empirical study that systematically investigates the factors influencing detection vulnerability by manipulating the topology, connectivity, and intensity of individual pedestrian 3D models and combining pedestrians with multiple objects within the CARLA simulation environment. Building on the insights, we propose the physically-informed text-to-3D adversarial generation (Phy3DAdvGen) that systematically optimizes text prompts by iteratively refining verbs, objects, and poses to produce LiDAR-invisible pedestrians. To ensure physical realizability, we construct a comprehensive object pool containing 13 3D models of real objects and constrain Phy3DAdvGen to generate 3D objects based on combinations of objects in this set. Extensive experiments demonstrate that our approach can generate 3D pedestrians that evade six state-of-the-art (SOTA) LiDAR 3D detectors in both CARLA simulation and physical environments, thereby highlighting vulnerabilities in safety-critical applications.", "date": "2025-10-08", "categories": []}
{"id": "2508.16481", "title": "Benchmarking the Robustness of Agentic Systems to Adversarially-Induced Harms", "abstract": "Ensuring the safe use of agentic systems requires a thorough understanding of the range of malicious behaviors these systems may exhibit when under attack. In this paper, we evaluate the robustness of LLM-based agentic systems against attacks that aim to elicit harmful actions from agents. To this end, we propose a novel taxonomy of harms for agentic systems and a novel benchmark, BAD-ACTS, for studying the security of agentic systems with respect to a wide range of harmful actions. BAD-ACTS consists of 4 implementations of agentic systems in distinct application environments, as well as a dataset of 188 high-quality examples of harmful actions. This enables a comprehensive study of the robustness of agentic systems across a wide range of categories of harmful behaviors, available tools, and inter-agent communication structures. Using this benchmark, we analyze the robustness of agentic systems against an attacker that controls one of the agents in the system and aims to manipulate other agents to execute a harmful target action. Our results show that the attack has a high success rate, demonstrating that even a single adversarial agent within the system can have a significant impact on the security. This attack remains effective even when agents use a simple prompting-based defense strategy. However, we additionally propose a more effective defense based on message monitoring. We believe that this benchmark provides a diverse testbed for the security research of agentic systems. The benchmark can be found at github.com/JNoether/BAD-ACTS", "date": "2025-08-22", "categories": []}
{"id": "2509.19947", "title": "A Set of Generalized Components to Achieve Effective Poison-only Clean-label Backdoor Attacks with Collaborative Sample Selection and Triggers", "abstract": "Poison-only Clean-label Backdoor Attacks aim to covertly inject attacker-desired behavior into DNNs by merely poisoning the dataset without changing the labels. To effectively implant a backdoor, multiple \\textbf{triggers} are proposed for various attack requirements of Attack Success Rate (ASR) and stealthiness. Additionally, sample selection enhances clean-label backdoor attacks' ASR by meticulously selecting ``hard'' samples instead of random samples to poison. Current methods 1) usually handle the sample selection and triggers in isolation, leading to severely limited improvements on both ASR and stealthiness. Consequently, attacks exhibit unsatisfactory performance on evaluation metrics when converted to PCBAs via a mere stacking of methods. Therefore, we seek to explore the bidirectional collaborative relations between the sample selection and triggers to address the above dilemma. 2) Since the strong specificity within triggers, the simple combination of sample selection and triggers fails to substantially enhance both evaluation metrics, with generalization preserved among various attacks. Therefore, we seek to propose a set of components to significantly improve both stealthiness and ASR based on the commonalities of attacks. Specifically, Component A ascertains two critical selection factors, and then makes them an appropriate combination based on the trigger scale to select more reasonable ``hard'' samples for improving ASR. Component B is proposed to select samples with similarities to relevant trigger implanted samples to promote stealthiness. Component C reassigns trigger poisoning intensity on RGB colors through distinct sensitivity of the human visual system to RGB for higher ASR, with stealthiness ensured by sample selection, including Component B. Furthermore, all components can be strategically integrated into diverse PCBAs.", "date": "2025-09-24", "categories": []}
{"id": "2509.21130", "title": "Sparse Representations Improve Adversarial Robustness of Neural Network Classifiers", "abstract": "Deep neural networks perform remarkably well on image classification tasks but remain vulnerable to carefully crafted adversarial perturbations. This work revisits linear dimensionality reduction as a simple, data-adapted defense. We empirically compare standard Principal Component Analysis (PCA) with its sparse variant (SPCA) as front-end feature extractors for downstream classifiers, and we complement these experiments with a theoretical analysis. On the theory side, we derive exact robustness certificates for linear heads applied to SPCA features: for both $\\ell_\\infty$ and $\\ell_2$ threat models (binary and multiclass), the certified radius grows as the dual norms of $W^\\top u$ shrink, where $W$ is the projection and $u$ the head weights. We further show that for general (non-linear) heads, sparsity reduces operator-norm bounds through a Lipschitz composition argument, predicting lower input sensitivity. Empirically, with a small non-linear network after the projection, SPCA consistently degrades more gracefully than PCA under strong white-box and black-box attacks while maintaining competitive clean accuracy. Taken together, the theory identifies the mechanism (sparser projections reduce adversarial leverage) and the experiments verify that this benefit persists beyond the linear setting. Our code is available at https://github.com/killian31/SPCARobustness.", "date": "2025-09-25", "categories": []}
{"id": "2510.05317", "title": "RegMix: Adversarial Mutual and Generalization Regularization for Enhancing DNN Robustness", "abstract": "Adversarial training is the most effective defense against adversarial attacks. The effectiveness of the adversarial attacks has been on the design of its loss function and regularization term. The most widely used loss function in adversarial training is cross-entropy and mean squared error (MSE) as its regularization objective. However, MSE enforces overly uniform optimization between two output distributions during training, which limits its robustness in adversarial training scenarios. To address this issue, we revisit the idea of mutual learning (originally designed for knowledge distillation) and propose two novel regularization strategies tailored for adversarial training: (i) weighted adversarial mutual regularization and (ii) adversarial generalization regularization. In the former, we formulate a decomposed adversarial mutual Kullback-Leibler divergence (KL-divergence) loss, which allows flexible control over the optimization process by assigning unequal weights to the main and auxiliary objectives. In the latter, we introduce an additional clean target distribution into the adversarial training objective, improving generalization and enhancing model robustness. Extensive experiments demonstrate that our proposed methods significantly improve adversarial robustness compared to existing regularization-based approaches.", "date": "2025-10-06", "categories": []}
{"id": "2306.06081", "title": "Blending adversarial training and representation-conditional purification via aggregation improves adversarial robustness", "abstract": "In this work, we propose a novel adversarial defence mechanism for image classification - CARSO - blending the paradigms of adversarial training and adversarial purification in a synergistic robustness-enhancing way. The method builds upon an adversarially-trained classifier, and learns to map its internal representation associated with a potentially perturbed input onto a distribution of tentative clean reconstructions. Multiple samples from such distribution are classified by the same adversarially-trained model, and a carefully chosen aggregation of its outputs finally constitutes the robust prediction of interest. Experimental evaluation by a well-established benchmark of strong adaptive attacks, across different image datasets, shows that CARSO is able to defend itself against adaptive end-to-end white-box attacks devised for stochastic defences. Paying a modest clean accuracy toll, our method improves by a significant margin the state-of-the-art for Cifar-10, Cifar-100, and TinyImageNet-200 $\\ell_\\infty$ robust classification accuracy against AutoAttack. Code, and instructions to obtain pre-trained models are available at: https://github.com/emaballarin/CARSO .", "date": "2023-05-25", "categories": []}
{"id": "2502.16901", "title": "Char-mander Use mBackdoor! A Study of Cross-lingual Backdoor Attacks in Multilingual LLMs", "abstract": "We explore \\textbf{C}ross-lingual \\textbf{B}ackdoor \\textbf{AT}tacks (X-BAT) in multilingual Large Language Models (mLLMs), revealing how backdoors inserted in one language can automatically transfer to others through shared embedding spaces. Using toxicity classification as a case study, we demonstrate that attackers can compromise multilingual systems by poisoning data in a single language, with rare and high-occurring tokens serving as specific, effective triggers. Our findings expose a critical vulnerability that influences the model's architecture, resulting in a concealed backdoor effect during the information flow. Our code and data are publicly available https://github.com/himanshubeniwal/X-BAT.", "date": "2025-02-24", "categories": []}
{"id": "2510.03254", "title": "Adversarial training with restricted data manipulation", "abstract": "Adversarial machine learning concerns situations in which learners face attacks from active adversaries. Such scenarios arise in applications such as spam email filtering, malware detection and fake image generation, where security methods must be actively updated to keep up with the everimproving generation of malicious data. Pessimistic Bilevel optimisation has been shown to be an effective method of training resilient classifiers against such adversaries. By modelling these scenarios as a game between the learner and the adversary, we anticipate how the adversary will modify their data and then train a resilient classifier accordingly. However, since existing pessimistic bilevel approaches feature an unrestricted adversary, the model is vulnerable to becoming overly pessimistic and unrealistic. When finding the optimal solution that defeats the classifier, it is possible that the adversary's data becomes nonsensical and loses its intended nature. Such an adversary will not properly reflect reality, and consequently, will lead to poor classifier performance when implemented on real-world data. By constructing a constrained pessimistic bilevel optimisation model, we restrict the adversary's movements and identify a solution that better reflects reality. We demonstrate through experiments that this model performs, on average, better than the existing approach.", "date": "2025-09-26", "categories": []}
{"id": "2510.03320", "title": "Attack logics, not outputs: Towards efficient robustification of deep neural networks by falsifying concept-based properties", "abstract": "Deep neural networks (NNs) for computer vision are vulnerable to adversarial attacks, i.e., miniscule malicious changes to inputs may induce unintuitive outputs. One key approach to verify and mitigate such robustness issues is to falsify expected output behavior. This allows, e.g., to locally proof security, or to (re)train NNs on obtained adversarial input examples. Due to the black-box nature of NNs, current attacks only falsify a class of the final output, such as flipping from $\\texttt{stop_sign}$ to $\\neg\\texttt{stop_sign}$. In this short position paper we generalize this to search for generally illogical behavior, as considered in NN verification: falsify constraints (concept-based properties) involving further human-interpretable concepts, like $\\texttt{red}\\wedge\\texttt{octogonal}\\rightarrow\\texttt{stop_sign}$. For this, an easy implementation of concept-based properties on already trained NNs is proposed using techniques from explainable artificial intelligence. Further, we sketch the theoretical proof that attacks on concept-based properties are expected to have a reduced search space compared to simple class falsification, whilst arguably be more aligned with intuitive robustness targets. As an outlook to this work in progress we hypothesize that this approach has potential to efficiently and simultaneously improve logical compliance and robustness.", "date": "2025-10-01", "categories": []}
{"id": "2510.03612", "title": "Cross-Modal Content Optimization for Steering Web Agent Preferences", "abstract": "Vision-language model (VLM)-based web agents increasingly power high-stakes selection tasks like content recommendation or product ranking by combining multimodal perception with preference reasoning. Recent studies reveal that these agents are vulnerable against attackers who can bias selection outcomes through preference manipulations using adversarial pop-ups, image perturbations, or content tweaks. Existing work, however, either assumes strong white-box access, with limited single-modal perturbations, or uses impractical settings. In this paper, we demonstrate, for the first time, that joint exploitation of visual and textual channels yields significantly more powerful preference manipulations under realistic attacker capabilities. We introduce Cross-Modal Preference Steering (CPS) that jointly optimizes imperceptible modifications to an item's visual and natural language descriptions, exploiting CLIP-transferable image perturbations and RLHF-induced linguistic biases to steer agent decisions. In contrast to prior studies that assume gradient access, or control over webpages, or agent memory, we adopt a realistic black-box threat setup: a non-privileged adversary can edit only their own listing's images and textual metadata, with no insight into the agent's model internals. We evaluate CPS on agents powered by state-of-the-art proprietary and open source VLMs including GPT-4.1, Qwen-2.5VL and Pixtral-Large on both movie selection and e-commerce tasks. Our results show that CPS is significantly more effective than leading baseline methods. For instance, our results show that CPS consistently outperforms baselines across all models while maintaining 70% lower detection rates, demonstrating both effectiveness and stealth. These findings highlight an urgent need for robust defenses as agentic systems play an increasingly consequential role in society.", "date": "2025-10-04", "categories": []}
{"id": "2510.03636", "title": "From Theory to Practice: Evaluating Data Poisoning Attacks and Defenses in In-Context Learning on Social Media Health Discourse", "abstract": "This study explored how in-context learning (ICL) in large language models can be disrupted by data poisoning attacks in the setting of public health sentiment analysis. Using tweets of Human Metapneumovirus (HMPV), small adversarial perturbations such as synonym replacement, negation insertion, and randomized perturbation were introduced into the support examples. Even these minor manipulations caused major disruptions, with sentiment labels flipping in up to 67% of cases. To address this, a Spectral Signature Defense was applied, which filtered out poisoned examples while keeping the data's meaning and sentiment intact. After defense, ICL accuracy remained steady at around 46.7%, and logistic regression validation reached 100% accuracy, showing that the defense successfully preserved the dataset's integrity. Overall, the findings extend prior theoretical studies of ICL poisoning to a practical, high-stakes setting in public health discourse analysis, highlighting both the risks and potential defenses for robust LLM deployment. This study also highlights the fragility of ICL under attack and the value of spectral defenses in making AI systems more reliable for health-related social media monitoring.", "date": "2025-10-04", "categories": []}
{"id": "2510.04245", "title": "Concept-Based Masking: A Patch-Agnostic Defense Against Adversarial Patch Attacks", "abstract": "Adversarial patch attacks pose a practical threat to deep learning models by forcing targeted misclassifications through localized perturbations, often realized in the physical world. Existing defenses typically assume prior knowledge of patch size or location, limiting their applicability. In this work, we propose a patch-agnostic defense that leverages concept-based explanations to identify and suppress the most influential concept activation vectors, thereby neutralizing patch effects without explicit detection. Evaluated on Imagenette with a ResNet-50, our method achieves higher robust and clean accuracy than the state-of-the-art PatchCleanser, while maintaining strong performance across varying patch sizes and locations. Our results highlight the promise of combining interpretability with robustness and suggest concept-driven defenses as a scalable strategy for securing machine learning models against adversarial patch attacks.", "date": "2025-10-05", "categories": []}
{"id": "2510.04528", "title": "Unified Threat Detection and Mitigation Framework (UTDMF): Combating Prompt Injection, Deception, and Bias in Enterprise-Scale Transformers", "abstract": "The rapid adoption of large language models (LLMs) in enterprise systems exposes vulnerabilities to prompt injection attacks, strategic deception, and biased outputs, threatening security, trust, and fairness. Extending our adversarial activation patching framework (arXiv:2507.09406), which induced deception in toy networks at a 23.9% rate, we introduce the Unified Threat Detection and Mitigation Framework (UTDMF), a scalable, real-time pipeline for enterprise-grade models like Llama-3.1 (405B), GPT-4o, and Claude-3.5. Through 700+ experiments per model, UTDMF achieves: (1) 92% detection accuracy for prompt injection (e.g., jailbreaking); (2) 65% reduction in deceptive outputs via enhanced patching; and (3) 78% improvement in fairness metrics (e.g., demographic bias). Novel contributions include a generalized patching algorithm for multi-threat detection, three groundbreaking hypotheses on threat interactions (e.g., threat chaining in enterprise workflows), and a deployment-ready toolkit with APIs for enterprise integration.", "date": "2025-10-06", "categories": []}
{"id": "2406.03508", "title": "Mutual Information Guided Backdoor Mitigation for Pre-trained Encoders", "abstract": "Self-supervised learning (SSL) is increasingly attractive for pre-training encoders without requiring labeled data. Downstream tasks built on top of those pre-trained encoders can achieve nearly state-of-the-art performance. The pre-trained encoders by SSL, however, are vulnerable to backdoor attacks as demonstrated by existing studies. Numerous backdoor mitigation techniques are designed for downstream task models. However, their effectiveness is impaired and limited when adapted to pre-trained encoders, due to the lack of label information when pre-training. To address backdoor attacks against pre-trained encoders, in this paper, we innovatively propose a mutual information guided backdoor mitigation technique, named MIMIC. MIMIC treats the potentially backdoored encoder as the teacher net and employs knowledge distillation to distill a clean student encoder from the teacher net. Different from existing knowledge distillation approaches, MIMIC initializes the student with random weights, inheriting no backdoors from teacher nets. Then MIMIC leverages mutual information between each layer and extracted features to locate where benign knowledge lies in the teacher net, with which distillation is deployed to clone clean features from teacher to student. We craft the distillation loss with two aspects, including clone loss and attention loss, aiming to mitigate backdoors and maintain encoder performance at the same time. Our evaluation conducted on two backdoor attacks in SSL demonstrates that MIMIC can significantly reduce the attack success rate by only utilizing <5% of clean data, surpassing seven state-of-the-art backdoor mitigation techniques.", "date": "2024-06-05", "categories": []}
{"id": "2502.17537", "title": "Rethinking the Vulnerability of Concept Erasure and a New Method", "abstract": "The proliferation of text-to-image diffusion models has raised significant privacy and security concerns, particularly regarding the generation of copyrighted or harmful images. In response, concept erasure (defense) methods have been developed to \"unlearn\" specific concepts through post-hoc finetuning. However, recent concept restoration (attack) methods have demonstrated that these supposedly erased concepts can be recovered using adversarially crafted prompts, revealing a critical vulnerability in current defense mechanisms. In this work, we first investigate the fundamental sources of adversarial vulnerability and reveal that vulnerabilities are pervasive in the prompt embedding space of concept-erased models, a characteristic inherited from the original pre-unlearned model. Furthermore, we introduce **RECORD**, a novel coordinate-descent-based restoration algorithm that consistently outperforms existing restoration methods by up to 17.8 times. We conduct extensive experiments to assess its compute-performance tradeoff and propose acceleration strategies.", "date": "2025-02-24", "categories": []}
{"id": "2510.01494", "title": "Understanding Adversarial Transfer: Why Representation-Space Attacks Fail Where Data-Space Attacks Succeed", "abstract": "The field of adversarial robustness has long established that adversarial examples can successfully transfer between image classifiers and that text jailbreaks can successfully transfer between language models (LMs). However, a pair of recent studies reported being unable to successfully transfer image jailbreaks between vision-language models (VLMs). To explain this striking difference, we propose a fundamental distinction regarding the transferability of attacks against machine learning models: attacks in the input data-space can transfer, whereas attacks in model representation space do not, at least not without geometric alignment of representations. We then provide theoretical and empirical evidence of this hypothesis in four different settings. First, we mathematically prove this distinction in a simple setting where two networks compute the same input-output map but via different representations. Second, we construct representation-space attacks against image classifiers that are as successful as well-known data-space attacks, but fail to transfer. Third, we construct representation-space attacks against LMs that successfully jailbreak the attacked models but again fail to transfer. Fourth, we construct data-space attacks against VLMs that successfully transfer to new VLMs, and we show that representation space attacks can transfer when VLMs' latent geometries are sufficiently aligned in post-projector space. Our work reveals that adversarial transfer is not an inherent property of all attacks but contingent on their operational domain - the shared data-space versus models' unique representation spaces - a critical insight for building more robust models.", "date": "2025-10-01", "categories": []}
{"id": "2510.02554", "title": "ToolTweak: An Attack on Tool Selection in LLM-based Agents", "abstract": "As LLMs increasingly power agents that interact with external tools, tool use has become an essential mechanism for extending their capabilities. These agents typically select tools from growing databases or marketplaces to solve user tasks, creating implicit competition among tool providers and developers for visibility and usage. In this paper, we show that this selection process harbors a critical vulnerability: by iteratively manipulating tool names and descriptions, adversaries can systematically bias agents toward selecting specific tools, gaining unfair advantage over equally capable alternatives. We present ToolTweak, a lightweight automatic attack that increases selection rates from a baseline of around 20% to as high as 81%, with strong transferability between open-source and closed-source models. Beyond individual tools, we show that such attacks cause distributional shifts in tool usage, revealing risks to fairness, competition, and security in emerging tool ecosystems. To mitigate these risks, we evaluate two defenses: paraphrasing and perplexity filtering, which reduce bias and lead agents to select functionally similar tools more equally. All code will be open-sourced upon acceptance.", "date": "2025-10-02", "categories": []}
{"id": "2510.02677", "title": "ARMs: Adaptive Red-Teaming Agent against Multimodal Models with Plug-and-Play Attacks", "abstract": "As vision-language models (VLMs) gain prominence, their multimodal interfaces also introduce new safety vulnerabilities, making the safety evaluation challenging and critical. Existing red-teaming efforts are either restricted to a narrow set of adversarial patterns or depend heavily on manual engineering, lacking scalable exploration of emerging real-world VLM vulnerabilities. To bridge this gap, we propose ARMs, an adaptive red-teaming agent that systematically conducts comprehensive risk assessments for VLMs. Given a target harmful behavior or risk definition, ARMs automatically optimizes diverse red-teaming strategies with reasoning-enhanced multi-step orchestration, to effectively elicit harmful outputs from target VLMs. We propose 11 novel multimodal attack strategies, covering diverse adversarial patterns of VLMs (e.g., reasoning hijacking, contextual cloaking), and integrate 17 red-teaming algorithms into ARMs via model context protocol (MCP). To balance the diversity and effectiveness of the attack, we design a layered memory with an epsilon-greedy attack exploration algorithm. Extensive experiments on instance- and policy-based benchmarks show that ARMs achieves SOTA attack success rates, exceeding baselines by an average of 52.1% and surpassing 90% on Claude-4-Sonnet. We show that the diversity of red-teaming instances generated by ARMs is significantly higher, revealing emerging vulnerabilities in VLMs. Leveraging ARMs, we construct ARMs-Bench, a large-scale multimodal safety dataset comprising over 30K red-teaming instances spanning 51 diverse risk categories, grounded in both real-world multimodal threats and regulatory risks. Safety fine-tuning with ARMs-Bench substantially improves the robustness of VLMs while preserving their general utility, providing actionable guidance to improve multimodal safety alignment against emerging threats.", "date": "2025-10-03", "categories": []}
{"id": "2510.02707", "title": "A Statistical Method for Attack-Agnostic Adversarial Attack Detection with Compressive Sensing Comparison", "abstract": "Adversarial attacks present a significant threat to modern machine learning systems. Yet, existing detection methods often lack the ability to detect unseen attacks or detect different attack types with a high level of accuracy. In this work, we propose a statistical approach that establishes a detection baseline before a neural network's deployment, enabling effective real-time adversarial detection. We generate a metric of adversarial presence by comparing the behavior of a compressed/uncompressed neural network pair. Our method has been tested against state-of-the-art techniques, and it achieves near-perfect detection across a wide range of attack types. Moreover, it significantly reduces false positives, making it both reliable and practical for real-world applications.", "date": "2025-10-03", "categories": []}
{"id": "2510.02913", "title": "Zero-Shot Robustness of Vision Language Models Via Confidence-Aware Weighting", "abstract": "Vision-language models like CLIP demonstrate impressive zero-shot generalization but remain highly vulnerable to adversarial attacks. In this work, we propose Confidence-Aware Weighting (CAW) to enhance zero-shot robustness in vision-language models. CAW consists of two components: (1) a Confidence-Aware loss that prioritizes uncertain adversarial examples by scaling the KL divergence between clean and adversarial predictions, and (2) a feature alignment regularization that preserves semantic consistency by minimizing the distance between frozen and fine-tuned image encoder features on adversarial inputs. These components work jointly to improve both clean and robust accuracy without sacrificing generalization. Extensive experiments on TinyImageNet and 14 additional datasets show that CAW outperforms recent methods such as PMG-AFT and TGA-ZSR under strong attacks like AutoAttack, while using less memory.", "date": "2025-10-03", "categories": []}
{"id": "2510.03224", "title": "Test-Time Defense Against Adversarial Attacks via Stochastic Resonance of Latent Ensembles", "abstract": "We propose a test-time defense mechanism against adversarial attacks: imperceptible image perturbations that significantly alter the predictions of a model. Unlike existing methods that rely on feature filtering or smoothing, which can lead to information loss, we propose to \"combat noise with noise\" by leveraging stochastic resonance to enhance robustness while minimizing information loss. Our approach introduces small translational perturbations to the input image, aligns the transformed feature embeddings, and aggregates them before mapping back to the original reference image. This can be expressed in a closed-form formula, which can be deployed on diverse existing network architectures without introducing additional network modules or fine-tuning for specific attack types. The resulting method is entirely training-free, architecture-agnostic, and attack-agnostic. Empirical results show state-of-the-art robustness on image classification and, for the first time, establish a generic test-time defense for dense prediction tasks, including stereo matching and optical flow, highlighting the method's versatility and practicality. Specifically, relative to clean (unperturbed) performance, our method recovers up to 68.1% of the accuracy loss on image classification, 71.9% on stereo matching, and 29.2% on optical flow under various types of adversarial attacks.", "date": "2025-10-03", "categories": []}
{"id": "2505.12586", "title": "A Few Large Shifts: Layer-Inconsistency Based Minimal Overhead Adversarial Example Detection", "abstract": "Deep neural networks (DNNs) are highly susceptible to adversarial examples--subtle, imperceptible perturbations that can lead to incorrect predictions. While detection-based defenses offer a practical alternative to adversarial training, many existing methods depend on external models, complex architectures, or adversarial data, limiting their efficiency and generalizability. We introduce a lightweight, plug-in detection framework that leverages internal layer-wise inconsistencies within the target model itself, requiring only benign data for calibration. Our approach is grounded in the A Few Large Shifts Assumption, which posits that adversarial perturbations induce large, localized violations of layer-wise Lipschitz continuity in a small subset of layers. Building on this, we propose two complementary strategies--Recovery Testing (RT) and Logit-layer Testing (LT)--to empirically measure these violations and expose internal disruptions caused by adversaries. Evaluated on CIFAR-10, CIFAR-100, and ImageNet under both standard and adaptive threat models, our method achieves state-of-the-art detection performance with negligible computational overhead. Furthermore, our system-level analysis provides a practical method for selecting a detection threshold with a formal lower-bound guarantee on accuracy. The code is available here: https://github.com/c0510gy/AFLS-AED.", "date": "2025-05-19", "categories": []}
{"id": "2510.01586", "title": "AdvEvo-MARL: Shaping Internalized Safety through Adversarial Co-Evolution in Multi-Agent Reinforcement Learning", "abstract": "LLM-based multi-agent systems excel at planning, tool use, and role coordination, but their openness and interaction complexity also expose them to jailbreak, prompt-injection, and adversarial collaboration. Existing defenses fall into two lines: (i) self-verification that asks each agent to pre-filter unsafe instructions before execution, and (ii) external guard modules that police behaviors. The former often underperforms because a standalone agent lacks sufficient capacity to detect cross-agent unsafe chains and delegation-induced risks; the latter increases system overhead and creates a single-point-of-failure-once compromised, system-wide safety collapses, and adding more guards worsens cost and complexity. To solve these challenges, we propose AdvEvo-MARL, a co-evolutionary multi-agent reinforcement learning framework that internalizes safety into task agents. Rather than relying on external guards, AdvEvo-MARL jointly optimizes attackers (which synthesize evolving jailbreak prompts) and defenders (task agents trained to both accomplish their duties and resist attacks) in adversarial learning environments. To stabilize learning and foster cooperation, we introduce a public baseline for advantage estimation: agents within the same functional group share a group-level mean-return baseline, enabling lower-variance updates and stronger intra-group coordination. Across representative attack scenarios, AdvEvo-MARL consistently keeps attack-success rate (ASR) below 20%, whereas baselines reach up to 38.33%, while preserving-and sometimes improving-task accuracy (up to +3.67% on reasoning tasks). These results show that safety and utility can be jointly improved without relying on extra guard agents or added system overhead.", "date": "2025-10-02", "categories": []}
{"id": "2510.01676", "title": "Evaluating the Robustness of a Production Malware Detection System to Transferable Adversarial Attacks", "abstract": "As deep learning models become widely deployed as components within larger production systems, their individual shortcomings can create system-level vulnerabilities with real-world impact. This paper studies how adversarial attacks targeting an ML component can degrade or bypass an entire production-grade malware detection system, performing a case study analysis of Gmail's pipeline where file-type identification relies on a ML model.\n  The malware detection pipeline in use by Gmail contains a machine learning model that routes each potential malware sample to a specialized malware classifier to improve accuracy and performance. This model, called Magika, has been open sourced. By designing adversarial examples that fool Magika, we can cause the production malware service to incorrectly route malware to an unsuitable malware detector thereby increasing our chance of evading detection. Specifically, by changing just 13 bytes of a malware sample, we can successfully evade Magika in 90% of cases and thereby allow us to send malware files over Gmail. We then turn our attention to defenses, and develop an approach to mitigate the severity of these types of attacks. For our defended production model, a highly resourced adversary requires 50 bytes to achieve just a 20% attack success rate. We implement this defense, and, thanks to a collaboration with Google engineers, it has already been deployed in production for the Gmail classifier.", "date": "2025-10-02", "categories": []}
{"id": "2510.01929", "title": "Inverse Language Modeling towards Robust and Grounded LLMs", "abstract": "The current landscape of defensive mechanisms for LLMs is fragmented and underdeveloped, unlike prior work on classifiers. To further promote adversarial robustness in LLMs, we propose Inverse Language Modeling (ILM), a unified framework that simultaneously 1) improves the robustness of LLMs to input perturbations, and, at the same time, 2) enables native grounding by inverting model outputs to identify potentially toxic or unsafe input triggers. ILM transforms LLMs from static generators into analyzable and robust systems, potentially helping RED teaming. ILM can lay the foundation for next-generation LLMs that are not only robust and grounded but also fundamentally more controllable and trustworthy. The code is publicly available at github.com/davegabe/pag-llm.", "date": "2025-10-02", "categories": []}
{"id": "2510.01969", "title": "Lower Bounds on Adversarial Robustness for Multiclass Classification with General Loss Functions", "abstract": "We consider adversarially robust classification in a multiclass setting under arbitrary loss functions and derive dual and barycentric reformulations of the corresponding learner-agnostic robust risk minimization problem. We provide explicit characterizations for important cases such as the cross-entropy loss, loss functions with a power form, and the quadratic loss, extending in this way available results for the 0-1 loss. These reformulations enable efficient computation of sharp lower bounds for adversarial risks and facilitate the design of robust classifiers beyond the 0-1 loss setting. Our paper uncovers interesting connections between adversarial robustness, $\\alpha$-fair packing problems, and generalized barycenter problems for arbitrary positive measures where Kullback-Leibler and Tsallis entropies are used as penalties. Our theoretical results are accompanied with illustrative numerical experiments where we obtain tighter lower bounds for adversarial risks with the cross-entropy loss function.", "date": "2025-10-02", "categories": []}
{"id": "2309.17401", "title": "Adversarial Attacks to Latent Representations of Distributed Neural Networks in Split Computing", "abstract": "Distributed deep neural networks (DNNs) have been shown to reduce the computational burden of mobile devices and decrease the end-to-end inference latency in edge computing scenarios. While distributed DNNs have been studied, to the best of our knowledge, the resilience of distributed DNNs to adversarial action remains an open problem. In this paper, we fill the existing research gap by rigorously analyzing the robustness of distributed DNNs against adversarial action. We cast this problem in the context of information theory and rigorously proved that (i) the compressed latent dimension improves the robustness but also affect task-oriented performance; and (ii) the deeper splitting point enhances the robustness but also increases the computational burden. These two trade-offs provide a novel perspective to design robust distributed DNN. To test our theoretical findings, we perform extensive experimental analysis by considering 6 different DNN architectures, 6 different approaches for distributed DNN and 10 different adversarial attacks using the ImageNet-1K dataset.", "date": "2023-09-29", "categories": []}
{"id": "2506.11402", "title": "LoRA Users Beware: A Few Spurious Tokens Can Manipulate Your Finetuned Model", "abstract": "Large Language Models (LLMs) are commonly finetuned for a variety of use cases and domains. A common approach is to leverage Low-Rank Adaptation (LoRA) -- known to provide strong performance at low resource costs. In this study, we demonstrate that LoRA actually opens the door to short-cut vulnerabilities -- and the more resource efficient is the LoRA setup, the more vulnerable will be the finetuned model to aggressive attacks. To measure that vulnerability, we introduce Seamless Spurious Token Injection (SSTI), where we find that LoRA exclusively focuses on even just a single token that is spuriously correlated with downstream labels. In short, injection of that spurious token during finetuning ensure that the model's prediction at test-time can be manipulated on-demand. We conducted experiments across model families and datasets to evaluate the impact of SSTI during LoRA finetuning while providing possible mitigations. Our experiments conclude that none of the existing checkers and preprocessors can sanitize a dataset raising new concerns for data quality and AI safety.", "date": "2025-06-13", "categories": []}
{"id": "2510.00451", "title": "A Call to Action for a Secure-by-Design Generative AI Paradigm", "abstract": "Large language models have gained widespread prominence, yet their vulnerability to prompt injection and other adversarial attacks remains a critical concern. This paper argues for a security-by-design AI paradigm that proactively mitigates LLM vulnerabilities while enhancing performance. To achieve this, we introduce PromptShield, an ontology-driven framework that ensures deterministic and secure prompt interactions. It standardizes user inputs through semantic validation, eliminating ambiguity and mitigating adversarial manipulation. To assess PromptShield's security and performance capabilities, we conducted an experiment on an agent-based system to analyze cloud logs within Amazon Web Services (AWS), containing 493 distinct events related to malicious activities and anomalies. By simulating prompt injection attacks and assessing the impact of deploying PromptShield, our results demonstrate a significant improvement in model security and performance, achieving precision, recall, and F1 scores of approximately 94%. Notably, the ontology-based framework not only mitigates adversarial threats but also enhances the overall performance and reliability of the system. Furthermore, PromptShield's modular and adaptable design ensures its applicability beyond cloud security, making it a robust solution for safeguarding generative AI applications across various domains. By laying the groundwork for AI safety standards and informing future policy development, this work stimulates a crucial dialogue on the pivotal role of deterministic prompt engineering and ontology-based validation in ensuring the safe and responsible deployment of LLMs in high-stakes environments.", "date": "2025-10-01", "categories": []}
{"id": "2510.00778", "title": "DIA: The Adversarial Exposure of Deterministic Inversion in Diffusion Models", "abstract": "Diffusion models have shown to be strong representation learners, showcasing state-of-the-art performance across multiple domains. Aside from accelerated sampling, DDIM also enables the inversion of real images back to their latent codes. A direct inheriting application of this inversion operation is real image editing, where the inversion yields latent trajectories to be utilized during the synthesis of the edited image. Unfortunately, this practical tool has enabled malicious users to freely synthesize misinformative or deepfake contents with greater ease, which promotes the spread of unethical and abusive, as well as privacy-, and copyright-infringing contents. While defensive algorithms such as AdvDM and Photoguard have been shown to disrupt the diffusion process on these images, the misalignment between their objectives and the iterative denoising trajectory at test time results in weak disruptive performance.In this work, we present the DDIM Inversion Attack (DIA) that attacks the integrated DDIM trajectory path. Our results support the effective disruption, surpassing previous defensive methods across various editing methods. We believe that our frameworks and results can provide practical defense methods against the malicious use of AI for both the industry and the research community. Our code is available here: https://anonymous.4open.science/r/DIA-13419/.", "date": "2025-10-01", "categories": []}
{"id": "2312.05386", "title": "Model Extraction Attacks Revisited", "abstract": "Model extraction (ME) attacks represent one major threat to Machine-Learning-as-a-Service (MLaaS) platforms by ``stealing'' the functionality of confidential machine-learning models through querying black-box APIs. Over seven years have passed since ME attacks were first conceptualized in the seminal work. During this period, substantial advances have been made in both ME attacks and MLaaS platforms, raising the intriguing question: How has the vulnerability of MLaaS platforms to ME attacks been evolving? In this work, we conduct an in-depth study to answer this critical question. Specifically, we characterize the vulnerability of current, mainstream MLaaS platforms to ME attacks from multiple perspectives including attack strategies, learning techniques, surrogate-model design, and benchmark tasks. Many of our findings challenge previously reported results, suggesting emerging patterns of ME vulnerability. Further, by analyzing the vulnerability of the same MLaaS platforms using historical datasets from the past four years, we retrospectively characterize the evolution of ME vulnerability over time, leading to a set of interesting findings. Finally, we make suggestions about improving the current practice of MLaaS in terms of attack robustness. Our study sheds light on the current state of ME vulnerability in the wild and points to several promising directions for future research.", "date": "2023-12-08", "categories": []}
{"id": "2501.13115", "title": "Dagger Behind Smile: Fool LLMs with a Happy Ending Story", "abstract": "The wide adoption of Large Language Models (LLMs) has attracted significant attention from $\\textit{jailbreak}$ attacks, where adversarial prompts crafted through optimization or manual design exploit LLMs to generate malicious contents. However, optimization-based attacks have limited efficiency and transferability, while existing manual designs are either easily detectable or demand intricate interactions with LLMs. In this paper, we first point out a novel perspective for jailbreak attacks: LLMs are more responsive to $\\textit{positive}$ prompts. Based on this, we deploy Happy Ending Attack (HEA) to wrap up a malicious request in a scenario template involving a positive prompt formed mainly via a $\\textit{happy ending}$, it thus fools LLMs into jailbreaking either immediately or at a follow-up malicious request. This has made HEA both efficient and effective, as it requires only up to two turns to fully jailbreak LLMs. Extensive experiments show that our HEA can successfully jailbreak on state-of-the-art LLMs, including GPT-4o, Llama3-70b, Gemini-pro, and achieves 88.79% attack success rate on average. We also provide quantitative explanations for the success of HEA.", "date": "2025-01-19", "categories": []}
{"id": "2509.20411", "title": "Adversarial Defense in Cybersecurity: A Systematic Review of GANs for Threat Detection and Mitigation", "abstract": "Machine learning-based cybersecurity systems are highly vulnerable to adversarial attacks, while Generative Adversarial Networks (GANs) act as both powerful attack enablers and promising defenses. This survey systematically reviews GAN-based adversarial defenses in cybersecurity (2021--August 31, 2025), consolidating recent progress, identifying gaps, and outlining future directions. Using a PRISMA-compliant systematic literature review protocol, we searched five major digital libraries. From 829 initial records, 185 peer-reviewed studies were retained and synthesized through quantitative trend analysis and thematic taxonomy development. We introduce a four-dimensional taxonomy spanning defensive function, GAN architecture, cybersecurity domain, and adversarial threat model. GANs improve detection accuracy, robustness, and data utility across network intrusion detection, malware analysis, and IoT security. Notable advances include WGAN-GP for stable training, CGANs for targeted synthesis, and hybrid GAN models for improved resilience. Yet, persistent challenges remain such as instability in training, lack of standardized benchmarks, high computational cost, and limited explainability. GAN-based defenses demonstrate strong potential but require advances in stable architectures, benchmarking, transparency, and deployment. We propose a roadmap emphasizing hybrid models, unified evaluation, real-world integration, and defenses against emerging threats such as LLM-driven cyberattacks. This survey establishes the foundation for scalable, trustworthy, and adaptive GAN-powered defenses.", "date": "2025-09-24", "categories": []}
{"id": "2509.25418", "title": "Leveraging Vulnerabilities in Temporal Graph Neural Networks via Strategic High-Impact Assaults", "abstract": "Temporal Graph Neural Networks (TGNNs) have become indispensable for analyzing dynamic graphs in critical applications such as social networks, communication systems, and financial networks. However, the robustness of TGNNs against adversarial attacks, particularly sophisticated attacks that exploit the temporal dimension, remains a significant challenge. Existing attack methods for Spatio-Temporal Dynamic Graphs (STDGs) often rely on simplistic, easily detectable perturbations (e.g., random edge additions/deletions) and fail to strategically target the most influential nodes and edges for maximum impact. We introduce the High Impact Attack (HIA), a novel restricted black-box attack framework specifically designed to overcome these limitations and expose critical vulnerabilities in TGNNs. HIA leverages a data-driven surrogate model to identify structurally important nodes (central to network connectivity) and dynamically important nodes (critical for the graph's temporal evolution). It then employs a hybrid perturbation strategy, combining strategic edge injection (to create misleading connections) and targeted edge deletion (to disrupt essential pathways), maximizing TGNN performance degradation. Importantly, HIA minimizes the number of perturbations to enhance stealth, making it more challenging to detect. Comprehensive experiments on five real-world datasets and four representative TGNN architectures (TGN, JODIE, DySAT, and TGAT) demonstrate that HIA significantly reduces TGNN accuracy on the link prediction task, achieving up to a 35.55% decrease in Mean Reciprocal Rank (MRR) - a substantial improvement over state-of-the-art baselines. These results highlight fundamental vulnerabilities in current STDG models and underscore the urgent need for robust defenses that account for both structural and temporal dynamics.", "date": "2025-09-29", "categories": []}
{"id": "2509.25792", "title": "PUREVQ-GAN: Defending Data Poisoning Attacks through Vector-Quantized Bottlenecks", "abstract": "We introduce PureVQ-GAN, a defense against data poisoning that forces backdoor triggers through a discrete bottleneck using Vector-Quantized VAE with GAN discriminator. By quantizing poisoned images through a learned codebook, PureVQ-GAN destroys fine-grained trigger patterns while preserving semantic content. A GAN discriminator ensures outputs match the natural image distribution, preventing reconstruction of out-of-distribution perturbations. On CIFAR-10, PureVQ-GAN achieves 0% poison success rate (PSR) against Gradient Matching and Bullseye Polytope attacks, and 1.64% against Narcissus while maintaining 91-95% clean accuracy. Unlike diffusion-based defenses requiring hundreds of iterative refinement steps, PureVQ-GAN is over 50x faster, making it practical for real training pipelines.", "date": "2025-09-30", "categories": []}
{"id": "2509.25927", "title": "The Impact of Scaling Training Data on Adversarial Robustness", "abstract": "Deep neural networks remain vulnerable to adversarial examples despite advances in architectures and training paradigms. We investigate how training data characteristics affect adversarial robustness across 36 state-of-the-art vision models spanning supervised, self-supervised, and contrastive learning approaches, trained on datasets from 1.2M to 22B images. Models were evaluated under six black-box attack categories: random perturbations, two types of geometric masks, COCO object manipulations, ImageNet-C corruptions, and ImageNet-R style shifts. Robustness follows a logarithmic scaling law with both data volume and model size: a tenfold increase in data reduces attack success rate (ASR) on average by ~3.2%, whereas a tenfold increase in model size reduces ASR on average by ~13.4%. Notably, some self-supervised models trained on curated datasets, such as DINOv2, outperform others trained on much larger but less curated datasets, challenging the assumption that scale alone drives robustness. Adversarial fine-tuning of ResNet50s improves generalization across structural variations but not across color distributions. Human evaluation reveals persistent gaps between human and machine vision. These results show that while scaling improves robustness, data quality, architecture, and training objectives play a more decisive role than raw scale in achieving broad-spectrum adversarial resilience.", "date": "2025-09-30", "categories": []}
{"id": "2509.25961", "title": "Reliability Crisis of Reference-free Metrics for Grammatical Error Correction", "abstract": "Reference-free evaluation metrics for grammatical error correction (GEC) have achieved high correlation with human judgments. However, these metrics are not designed to evaluate adversarial systems that aim to obtain unjustifiably high scores. The existence of such systems undermines the reliability of automatic evaluation, as it can mislead users in selecting appropriate GEC systems. In this study, we propose adversarial attack strategies for four reference-free metrics: SOME, Scribendi, IMPARA, and LLM-based metrics, and demonstrate that our adversarial systems outperform the current state-of-the-art. These findings highlight the need for more robust evaluation methods.", "date": "2025-09-30", "categories": []}
{"id": "2509.26473", "title": "STaR-Attack: A Spatio-Temporal and Narrative Reasoning Attack Framework for Unified Multimodal Understanding and Generation Models", "abstract": "Unified Multimodal understanding and generation Models (UMMs) have demonstrated remarkable capabilities in both understanding and generation tasks. However, we identify a vulnerability arising from the generation-understanding coupling in UMMs. The attackers can use the generative function to craft an information-rich adversarial image and then leverage the understanding function to absorb it in a single pass, which we call Cross-Modal Generative Injection (CMGI). Current attack methods on malicious instructions are often limited to a single modality while also relying on prompt rewriting with semantic drift, leaving the unique vulnerabilities of UMMs unexplored. We propose STaR-Attack, the first multi-turn jailbreak attack framework that exploits unique safety weaknesses of UMMs without semantic drift. Specifically, our method defines a malicious event that is strongly correlated with the target query within a spatio-temporal context. Using the three-act narrative theory, STaR-Attack generates the pre-event and the post-event scenes while concealing the malicious event as the hidden climax. When executing the attack strategy, the opening two rounds exploit the UMM's generative ability to produce images for these scenes. Subsequently, an image-based question guessing and answering game is introduced by exploiting the understanding capability. STaR-Attack embeds the original malicious question among benign candidates, forcing the model to select and answer the most relevant one given the narrative context. Extensive experiments show that STaR-Attack consistently surpasses prior approaches, achieving up to 93.06% ASR on Gemini-2.0-Flash and surpasses the strongest prior baseline, FlipAttack. Our work uncovers a critical yet underdeveloped vulnerability and highlights the need for safety alignments in UMMs.", "date": "2025-09-30", "categories": []}
{"id": "2509.26598", "title": "Are Robust LLM Fingerprints Adversarially Robust?", "abstract": "Model fingerprinting has emerged as a promising paradigm for claiming model ownership. However, robustness evaluations of these schemes have mostly focused on benign perturbations such as incremental fine-tuning, model merging, and prompting. Lack of systematic investigations into {\\em adversarial robustness} against a malicious model host leaves current systems vulnerable. To bridge this gap, we first define a concrete, practical threat model against model fingerprinting. We then take a critical look at existing model fingerprinting schemes to identify their fundamental vulnerabilities. Based on these, we develop adaptive adversarial attacks tailored for each vulnerability, and demonstrate that these can bypass model authentication completely for ten recently proposed fingerprinting schemes while maintaining high utility of the model for the end users. Our work encourages fingerprint designers to adopt adversarial robustness by design. We end with recommendations for future fingerprinting methods.", "date": "2025-09-30", "categories": []}
{"id": "2410.05694", "title": "DiffusionGuard: A Robust Defense Against Malicious Diffusion-based Image Editing", "abstract": "Recent advances in diffusion models have introduced a new era of text-guided image manipulation, enabling users to create realistic edited images with simple textual prompts. However, there is significant concern about the potential misuse of these methods, especially in creating misleading or harmful content. Although recent defense strategies, which introduce imperceptible adversarial noise to induce model failure, have shown promise, they remain ineffective against more sophisticated manipulations, such as editing with a mask. In this work, we propose DiffusionGuard, a robust and effective defense method against unauthorized edits by diffusion-based image editing models, even in challenging setups. Through a detailed analysis of these models, we introduce a novel objective that generates adversarial noise targeting the early stage of the diffusion process. This approach significantly improves the efficiency and effectiveness of adversarial noises. We also introduce a mask-augmentation technique to enhance robustness against various masks during test time. Finally, we introduce a comprehensive benchmark designed to evaluate the effectiveness and robustness of methods in protecting against privacy threats in realistic scenarios. Through extensive experiments, we show that our method achieves stronger protection and improved mask robustness with lower computational costs compared to the strongest baseline. Additionally, our method exhibits superior transferability and better resilience to noise removal techniques compared to all baseline methods. Our source code is publicly available at https://github.com/choi403/DiffusionGuard.", "date": "2024-10-08", "categories": []}
{"id": "2506.08885", "title": "AdversariaL attacK sAfety aLIgnment(ALKALI): Safeguarding LLMs through GRACE: Geometric Representation-Aware Contrastive Enhancement- Introducing Adversarial Vulnerability Quality Index (AVQI)", "abstract": "Adversarial threats against LLMs are escalating faster than current defenses can adapt. We expose a critical geometric blind spot in alignment: adversarial prompts exploit latent camouflage, embedding perilously close to the safe representation manifold while encoding unsafe intent thereby evading surface level defenses like Direct Preference Optimization (DPO), which remain blind to the latent geometry. We introduce ALKALI, the first rigorously curated adversarial benchmark and the most comprehensive to date spanning 9,000 prompts across three macro categories, six subtypes, and fifteen attack families. Evaluation of 21 leading LLMs reveals alarmingly high Attack Success Rates (ASRs) across both open and closed source models, exposing an underlying vulnerability we term latent camouflage, a structural blind spot where adversarial completions mimic the latent geometry of safe ones. To mitigate this vulnerability, we introduce GRACE - Geometric Representation Aware Contrastive Enhancement, an alignment framework coupling preference learning with latent space regularization. GRACE enforces two constraints: latent separation between safe and adversarial completions, and adversarial cohesion among unsafe and jailbreak behaviors. These operate over layerwise pooled embeddings guided by a learned attention profile, reshaping internal geometry without modifying the base model, and achieve up to 39% ASR reduction. Moreover, we introduce AVQI, a geometry aware metric that quantifies latent alignment failure via cluster separation and compactness. AVQI reveals when unsafe completions mimic the geometry of safe ones, offering a principled lens into how models internally encode safety. We make the code publicly available at https://anonymous.4open.science/r/alkali-B416/README.md.", "date": "2025-06-10", "categories": []}
{"id": "2507.09993", "title": "3DGAA: Realistic and Robust 3D Gaussian-based Adversarial Attack for Autonomous Driving", "abstract": "Camera-based object detection systems play a vital role in autonomous driving, yet they remain vulnerable to adversarial threats in real-world environments. Existing 2D and 3D physical attacks, due to their focus on texture optimization, often struggle to balance physical realism and attack robustness. In this work, we propose 3D Gaussian-based Adversarial Attack (3DGAA), a novel adversarial object generation framework that leverages the full 14-dimensional parameterization of 3D Gaussian Splatting (3DGS) to jointly optimize geometry and appearance in physically realizable ways. Unlike prior works that rely on patches or texture optimization, 3DGAA jointly perturbs both geometric attributes (shape, scale, rotation) and appearance attributes (color, opacity) to produce physically realistic and transferable adversarial objects. We further introduce a physical filtering module that filters outliers to preserve geometric fidelity, and a physical augmentation module that simulates complex physical scenarios to enhance attack generalization under real-world conditions. We evaluate 3DGAA on both virtual benchmarks and physical-world setups using miniature vehicle models. Experimental results show that 3DGAA achieves to reduce the detection mAP from 87.21\\% to 7.38\\%, significantly outperforming existing 3D physical attacks. Moreover, our method maintains high transferability across different physical conditions, demonstrating a new state-of-the-art in physically realizable adversarial attacks.", "date": "2025-07-14", "categories": []}
{"id": "2509.19994", "title": "Improving Generalizability and Undetectability for Targeted Adversarial Attacks on Multimodal Pre-trained Models", "abstract": "Multimodal pre-trained models (e.g., ImageBind), which align distinct data modalities into a shared embedding space, have shown remarkable success across downstream tasks. However, their increasing adoption raises serious security concerns, especially regarding targeted adversarial attacks. In this paper, we show that existing targeted adversarial attacks on multimodal pre-trained models still have limitations in two aspects: generalizability and undetectability. Specifically, the crafted targeted adversarial examples (AEs) exhibit limited generalization to partially known or semantically similar targets in cross-modal alignment tasks (i.e., limited generalizability) and can be easily detected by simple anomaly detection methods (i.e., limited undetectability). To address these limitations, we propose a novel method called Proxy Targeted Attack (PTA), which leverages multiple source-modal and target-modal proxies to optimize targeted AEs, ensuring they remain evasive to defenses while aligning with multiple potential targets. We also provide theoretical analyses to highlight the relationship between generalizability and undetectability and to ensure optimal generalizability while meeting the specified requirements for undetectability. Furthermore, experimental results demonstrate that our PTA can achieve a high success rate across various related targets and remain undetectable against multiple anomaly detection methods.", "date": "2025-09-24", "categories": []}
{"id": "2509.20463", "title": "Efficiently Attacking Memorization Scores", "abstract": "Influence estimation tools -- such as memorization scores -- are widely used to understand model behavior, attribute training data, and inform dataset curation. However, recent applications in data valuation and responsible machine learning raise the question: can these scores themselves be adversarially manipulated? In this work, we present a systematic study of the feasibility of attacking memorization-based influence estimators. We characterize attacks for producing highly memorized samples as highly sensitive queries in the regime where a trained algorithm is accurate. Our attack (calculating the pseudoinverse of the input) is practical, requiring only black-box access to model outputs and incur modest computational overhead. We empirically validate our attack across a wide suite of image classification tasks, showing that even state-of-the-art proxies are vulnerable to targeted score manipulations. In addition, we provide a theoretical analysis of the stability of memorization scores under adversarial perturbations, revealing conditions under which influence estimates are inherently fragile. Our findings highlight critical vulnerabilities in influence-based attribution and suggest the need for robust defenses. All code can be found at https://github.com/tuedo2/MemAttack", "date": "2025-09-24", "categories": []}
{"id": "2509.22710", "title": "Localizing Adversarial Attacks To Produces More Imperceptible Noise", "abstract": "Adversarial attacks in machine learning traditionally focus on global perturbations to input data, yet the potential of localized adversarial noise remains underexplored. This study systematically evaluates localized adversarial attacks across widely-used methods, including FGSM, PGD, and C&W, to quantify their effectiveness, imperceptibility, and computational efficiency. By introducing a binary mask to constrain noise to specific regions, localized attacks achieve significantly lower mean pixel perturbations, higher Peak Signal-to-Noise Ratios (PSNR), and improved Structural Similarity Index (SSIM) compared to global attacks. However, these benefits come at the cost of increased computational effort and a modest reduction in Attack Success Rate (ASR). Our results highlight that iterative methods, such as PGD and C&W, are more robust to localization constraints than single-step methods like FGSM, maintaining higher ASR and imperceptibility metrics. This work provides a comprehensive analysis of localized adversarial attacks, offering practical insights for advancing attack strategies and designing robust defensive systems.", "date": "2025-09-23", "categories": []}
{"id": "2509.23037", "title": "GuardNet: Graph-Attention Filtering for Jailbreak Defense in Large Language Models", "abstract": "Large Language Models (LLMs) are increasingly susceptible to jailbreak attacks, which are adversarial prompts that bypass alignment constraints and induce unauthorized or harmful behaviors. These vulnerabilities undermine the safety, reliability, and trustworthiness of LLM outputs, posing critical risks in domains such as healthcare, finance, and legal compliance. In this paper, we propose GuardNet, a hierarchical filtering framework that detects and filters jailbreak prompts prior to inference. GuardNet constructs structured graphs that combine sequential links, syntactic dependencies, and attention-derived token relations to capture both linguistic structure and contextual patterns indicative of jailbreak behavior. It then applies graph neural networks at two levels: (i) a prompt-level filter that detects global adversarial prompts, and (ii) a token-level filter that pinpoints fine-grained adversarial spans. Extensive experiments across three datasets and multiple attack settings show that GuardNet substantially outperforms prior defenses. It raises prompt-level F$_1$ scores from 66.4\\% to 99.8\\% on LLM-Fuzzer, and from 67-79\\% to over 94\\% on PLeak datasets. At the token level, GuardNet improves F$_1$ from 48-75\\% to 74-91\\%, with IoU gains up to +28\\%. Despite its structural complexity, GuardNet maintains acceptable latency and generalizes well in cross-domain evaluations, making it a practical and robust defense against jailbreak threats in real-world LLM deployments.", "date": "2025-09-27", "categories": []}
{"id": "2509.23198", "title": "Real-World Transferable Adversarial Attack on Face-Recognition Systems", "abstract": "Adversarial attacks on face recognition (FR) systems pose a significant security threat, yet most are confined to the digital domain or require white-box access. We introduce GaP (Gaussian Patch), a novel method to generate a universal, physically transferable adversarial patch under a strict black-box setting. Our approach uses a query-efficient, zero-order greedy algorithm to iteratively construct a symmetric, grayscale pattern for the forehead. The patch is optimized by successively adding Gaussian blobs, guided only by the cosine similarity scores from a surrogate FR model to maximally degrade identity recognition. We demonstrate that with approximately 10,000 queries to a black-box ArcFace model, the resulting GaP achieves a high attack success rate in both digital and real-world physical tests. Critically, the attack shows strong transferability, successfully deceiving an entirely unseen FaceNet model. Our work highlights a practical and severe vulnerability, proving that robust, transferable attacks can be crafted with limited knowledge of the target system.", "date": "2025-09-27", "categories": []}
{"id": "2509.23689", "title": "Merge Now, Regret Later: The Hidden Cost of Model Merging is Adversarial Transferability", "abstract": "Model Merging (MM) has emerged as a promising alternative to multi-task learning, where multiple fine-tuned models are combined, without access to tasks' training data, into a single model that maintains performance across tasks. Recent works have explored the impact of MM on adversarial attacks, particularly backdoor attacks. However, none of them have sufficiently explored its impact on transfer attacks using adversarial examples, i.e., a black-box adversarial attack where examples generated for a surrogate model successfully mislead a target model.\n  In this work, we study the effect of MM on the transferability of adversarial examples. We perform comprehensive evaluations and statistical analysis consisting of 8 MM methods, 7 datasets, and 6 attack methods, sweeping over 336 distinct attack settings. Through it, we first challenge the prevailing notion of MM conferring free adversarial robustness, and show MM cannot reliably defend against transfer attacks, with over 95% relative transfer attack success rate. Moreover, we reveal 3 key insights for machine-learning practitioners regarding MM and transferability for a robust system design: (1) stronger MM methods increase vulnerability to transfer attacks; (2) mitigating representation bias increases vulnerability to transfer attacks; and (3) weight averaging, despite being the weakest MM method, is the most vulnerable MM method to transfer attacks. Finally, we analyze the underlying reasons for this increased vulnerability, and provide potential solutions to the problem. Our findings offer critical insights for designing more secure systems employing MM.", "date": "2025-09-28", "categories": []}
{"id": "2509.23917", "title": "Bridging the Task Gap: Multi-Task Adversarial Transferability in CLIP and Its Derivatives", "abstract": "As a general-purpose vision-language pretraining model, CLIP demonstrates strong generalization ability in image-text alignment tasks and has been widely adopted in downstream applications such as image classification and image-text retrieval. However, it struggles with fine-grained tasks such as object detection and semantic segmentation. While many variants aim to improve CLIP on these tasks, its robustness to adversarial perturbations remains underexplored. Understanding how adversarial examples transfer across tasks is key to assessing CLIP's generalization limits and security risks. In this work, we conduct a systematic empirical analysis of the cross-task transfer behavior of CLIP-based models on image-text retrieval, object detection, and semantic segmentation under adversarial perturbations. We find that adversarial examples generated from fine-grained tasks (e.g., object detection and semantic segmentation) often exhibit stronger transfer potential than those from coarse-grained tasks, enabling more effective attacks against the original CLIP model. Motivated by this observation, we propose a novel framework, Multi-Task Adversarial CLIP (MT-AdvCLIP), which introduces a task-aware feature aggregation loss and generates perturbations with enhanced cross-task generalization capability. This design strengthens the attack effectiveness of fine-grained task models on the shared CLIP backbone. Experimental results on multiple public datasets show that MT-AdvCLIP significantly improves the adversarial transfer success rate (The average attack success rate across multiple tasks is improved by over 39%.) against various CLIP-derived models, without increasing the perturbation budget. This study reveals the transfer mechanism of adversarial examples in multi-task CLIP models, offering new insights into multi-task robustness evaluation and adversarial example design.", "date": "2025-09-28", "categories": []}
{"id": "2509.24359", "title": "DRIFT: Divergent Response in Filtered Transformations for Robust Adversarial Defense", "abstract": "Deep neural networks remain highly vulnerable to adversarial examples, and most defenses collapse once gradients can be reliably estimated. We identify \\emph{gradient consensus} -- the tendency of randomized transformations to yield aligned gradients -- as a key driver of adversarial transferability. Attackers exploit this consensus to construct perturbations that remain effective across transformations. We introduce \\textbf{DRIFT} (Divergent Response in Filtered Transformations), a stochastic ensemble of lightweight, learnable filters trained to actively disrupt gradient consensus. Unlike prior randomized defenses that rely on gradient masking, DRIFT enforces \\emph{gradient dissonance} by maximizing divergence in Jacobian- and logit-space responses while preserving natural predictions. Our contributions are threefold: (i) we formalize gradient consensus and provide a theoretical analysis linking consensus to transferability; (ii) we propose a consensus-divergence training strategy combining prediction consistency, Jacobian separation, logit-space separation, and adversarial robustness; and (iii) we show that DRIFT achieves substantial robustness gains on ImageNet across CNNs and Vision Transformers, outperforming state-of-the-art preprocessing, adversarial training, and diffusion-based defenses under adaptive white-box, transfer-based, and gradient-free attacks. DRIFT delivers these improvements with negligible runtime and memory cost, establishing gradient divergence as a practical and generalizable principle for adversarial defense.", "date": "2025-09-29", "categories": []}
{"id": "2509.24891", "title": "VAGUEGAN: Stealthy Poisoning and Backdoor Attacks on Image Generative Pipelines", "abstract": "Generative models such as GANs and diffusion models are widely used to synthesize photorealistic images and to support downstream creative and editing tasks. While adversarial attacks on discriminative models are well studied, attacks targeting generative pipelines where small, stealthy perturbations in inputs lead to controlled changes in outputs are less explored. This study introduces VagueGAN, an attack pipeline combining a modular perturbation network PoisonerNet with a Generator Discriminator pair to craft stealthy triggers that cause targeted changes in generated images. Attack efficacy is evaluated using a custom proxy metric, while stealth is analyzed through perceptual and frequency domain measures. The transferability of the method to a modern diffusion based pipeline is further examined through ControlNet guided editing. Interestingly, the experiments show that poisoned outputs can display higher visual quality compared to clean counterparts, challenging the assumption that poisoning necessarily reduces fidelity. Unlike conventional pixel level perturbations, latent space poisoning in GANs and diffusion pipelines can retain or even enhance output aesthetics, exposing a blind spot in pixel level defenses. Moreover, carefully optimized perturbations can produce consistent, stealthy effects on generator outputs while remaining visually inconspicuous, raising concerns for the integrity of image generation pipelines.", "date": "2025-09-29", "categories": []}
{"id": "2509.25082", "title": "MANI-Pure: Magnitude-Adaptive Noise Injection for Adversarial Purification", "abstract": "Adversarial purification with diffusion models has emerged as a promising defense strategy, but existing methods typically rely on uniform noise injection, which indiscriminately perturbs all frequencies, corrupting semantic structures and undermining robustness. Our empirical study reveals that adversarial perturbations are not uniformly distributed: they are predominantly concentrated in high-frequency regions, with heterogeneous magnitude intensity patterns that vary across frequencies and attack types. Motivated by this observation, we introduce MANI-Pure, a magnitude-adaptive purification framework that leverages the magnitude spectrum of inputs to guide the purification process. Instead of injecting homogeneous noise, MANI-Pure adaptively applies heterogeneous, frequency-targeted noise, effectively suppressing adversarial perturbations in fragile high-frequency, low-magnitude bands while preserving semantically critical low-frequency content. Extensive experiments on CIFAR-10 and ImageNet-1K validate the effectiveness of MANI-Pure. It narrows the clean accuracy gap to within 0.59 of the original classifier, while boosting robust accuracy by 2.15, and achieves the top-1 robust accuracy on the RobustBench leaderboard, surpassing the previous state-of-the-art method.", "date": "2025-09-29", "categories": []}
{"id": "2404.16956", "title": "A Notion of Uniqueness for the Adversarial Bayes Classifier", "abstract": "We propose a new notion of uniqueness for the adversarial Bayes classifier in the setting of binary classification. Analyzing this concept produces a simple procedure for computing all adversarial Bayes classifiers for a well-motivated family of one dimensional data distributions. This characterization is then leveraged to show that as the perturbation radius increases, certain notions of regularity for the adversarial Bayes classifiers improve. Furthermore, these results provide tools for understanding relationships between the Bayes and adversarial Bayes classifiers in one dimension.", "date": "2024-04-25", "categories": []}
{"id": "2404.17358", "title": "Adversarial Consistency and the Uniqueness of the Adversarial Bayes Classifier", "abstract": "Minimizing an adversarial surrogate risk is a common technique for learning robust classifiers. Prior work showed that convex surrogate losses are not statistically consistent in the adversarial context -- or in other words, a minimizing sequence of the adversarial surrogate risk will not necessarily minimize the adversarial classification error. We connect the consistency of adversarial surrogate losses to properties of minimizers to the adversarial classification risk, known as adversarial Bayes classifiers. Specifically, under reasonable distributional assumptions, a convex surrogate loss is statistically consistent for adversarial learning iff the adversarial Bayes classifier satisfies a certain notion of uniqueness.", "date": "2024-04-26", "categories": []}
{"id": "2506.01444", "title": "Variance-Based Defense Against Blended Backdoor Attacks", "abstract": "Backdoor attacks represent a subtle yet effective class of cyberattacks targeting AI models, primarily due to their stealthy nature. The model behaves normally on clean data but exhibits malicious behavior only when the attacker embeds a specific trigger into the input. This attack is performed during the training phase, where the adversary corrupts a small subset of the training data by embedding a pattern and modifying the labels to a chosen target. The objective is to make the model associate the pattern with the target label while maintaining normal performance on unaltered data. Several defense mechanisms have been proposed to sanitize training data-sets. However, these methods often rely on the availability of a clean dataset to compute statistical anomalies, which may not always be feasible in real-world scenarios where datasets can be unavailable or compromised. To address this limitation, we propose a novel defense method that trains a model on the given dataset, detects poisoned classes, and extracts the critical part of the attack trigger before identifying the poisoned instances. This approach enhances explainability by explicitly revealing the harmful part of the trigger. The effectiveness of our method is demonstrated through experimental evaluations on well-known image datasets and comparative analysis against three state-of-the-art algorithms: SCAn, ABL, and AGPD.", "date": "2025-06-02", "categories": []}
{"id": "2509.20793", "title": "FERD: Fairness-Enhanced Data-Free Robustness Distillation", "abstract": "Data-Free Robustness Distillation (DFRD) aims to transfer the robustness from the teacher to the student without accessing the training data. While existing methods focus on overall robustness, they overlook the robust fairness issues, leading to severe disparity of robustness across different categories. In this paper, we find two key problems: (1) student model distilled with equal class proportion data behaves significantly different across distinct categories; and (2) the robustness of student model is not stable across different attacks target. To bridge these gaps, we present the first Fairness-Enhanced data-free Robustness Distillation (FERD) framework to adjust the proportion and distribution of adversarial examples. For the proportion, FERD adopts a robustness-guided class reweighting strategy to synthesize more samples for the less robust categories, thereby improving robustness of them. For the distribution, FERD generates complementary data samples for advanced robustness distillation. It generates Fairness-Aware Examples (FAEs) by enforcing a uniformity constraint on feature-level predictions, which suppress the dominance of class-specific non-robust features, providing a more balanced representation across all categories. Then, FERD constructs Uniform-Target Adversarial Examples (UTAEs) from FAEs by applying a uniform target class constraint to avoid biased attack directions, which distribute the attack targets across all categories and prevents overfitting to specific vulnerable categories. Extensive experiments on three public datasets show that FERD achieves state-of-the-art worst-class robustness under all adversarial attack (e.g., the worst-class robustness under FGSM and AutoAttack are improved by 15.1\\% and 6.4\\% using MobileNet-V2 on CIFAR-10), demonstrating superior performance in both robustness and fairness aspects.", "date": "2025-09-25", "categories": []}
{"id": "2509.21360", "title": "Multimodal Prompt Decoupling Attack on the Safety Filters in Text-to-Image Models", "abstract": "Text-to-image (T2I) models have been widely applied in generating high-fidelity images across various domains. However, these models may also be abused to produce Not-Safe-for-Work (NSFW) content via jailbreak attacks. Existing jailbreak methods primarily manipulate the textual prompt, leaving potential vulnerabilities in image-based inputs largely unexplored. Moreover, text-based methods face challenges in bypassing the model's safety filters. In response to these limitations, we propose the Multimodal Prompt Decoupling Attack (MPDA), which utilizes image modality to separate the harmful semantic components of the original unsafe prompt. MPDA follows three core steps: firstly, a large language model (LLM) decouples unsafe prompts into pseudo-safe prompts and harmful prompts. The former are seemingly harmless sub-prompts that can bypass filters, while the latter are sub-prompts with unsafe semantics that trigger filters. Subsequently, the LLM rewrites the harmful prompts into natural adversarial prompts to bypass safety filters, which guide the T2I model to modify the base image into an NSFW output. Finally, to ensure semantic consistency between the generated NSFW images and the original unsafe prompts, the visual language model generates image captions, providing a new pathway to guide the LLM in iterative rewriting and refining the generated content.", "date": "2025-09-21", "categories": []}
{"id": "2509.21843", "title": "SBFA: Single Sneaky Bit Flip Attack to Break Large Language Models", "abstract": "Model integrity of Large language models (LLMs) has become a pressing security concern with their massive online deployment. Prior Bit-Flip Attacks (BFAs) -- a class of popular AI weight memory fault-injection techniques -- can severely compromise Deep Neural Networks (DNNs): as few as tens of bit flips can degrade accuracy toward random guessing. Recent studies extend BFAs to LLMs and reveal that, despite the intuition of better robustness from modularity and redundancy, only a handful of adversarial bit flips can also cause LLMs' catastrophic accuracy degradation. However, existing BFA methods typically focus on either integer or floating-point models separately, limiting attack flexibility. Moreover, in floating-point models, random bit flips often cause perturbed parameters to extreme values (e.g., flipping in exponent bit), making it not stealthy and leading to numerical runtime error (e.g., invalid tensor values (NaN/Inf)). In this work, for the first time, we propose SBFA (Sneaky Bit-Flip Attack), which collapses LLM performance with only one single bit flip while keeping perturbed values within benign layer-wise weight distribution. It is achieved through iterative searching and ranking through our defined parameter sensitivity metric, ImpactScore, which combines gradient sensitivity and perturbation range constrained by the benign layer-wise weight distribution. A novel lightweight SKIP searching algorithm is also proposed to greatly reduce searching complexity, which leads to successful SBFA searching taking only tens of minutes for SOTA LLMs. Across Qwen, LLaMA, and Gemma models, with only one single bit flip, SBFA successfully degrades accuracy to below random levels on MMLU and SST-2 in both BF16 and INT8 data formats. Remarkably, flipping a single bit out of billions of parameters reveals a severe security concern of SOTA LLM models.", "date": "2025-09-26", "categories": []}
{"id": "2509.22393", "title": "Text Adversarial Attacks with Dynamic Outputs", "abstract": "Text adversarial attack methods are typically designed for static scenarios with fixed numbers of output labels and a predefined label space, relying on extensive querying of the victim model (query-based attacks) or the surrogate model (transfer-based attacks). To address this gap, we introduce the Textual Dynamic Outputs Attack (TDOA) method, which employs a clustering-based surrogate model training approach to convert the dynamic-output scenario into a static single-output scenario. To improve attack effectiveness, we propose the farthest-label targeted attack strategy, which selects adversarial vectors that deviate most from the model's coarse-grained labels, thereby maximizing disruption. We extensively evaluate TDOA on four datasets and eight victim models (e.g., ChatGPT-4o, ChatGPT-4.1), showing its effectiveness in crafting adversarial examples and its strong potential to compromise large language models with limited access. With a single query per text, TDOA achieves a maximum attack success rate of 50.81\\%. Additionally, we find that TDOA also achieves state-of-the-art performance in conventional static output scenarios, reaching a maximum ASR of 82.68\\%. Meanwhile, by conceptualizing translation tasks as classification problems with unbounded output spaces, we extend the TDOA framework to generative settings, surpassing prior results by up to 0.64 RDBLEU and 0.62 RDchrF.", "date": "2025-09-26", "categories": []}
{"id": "2405.19179", "title": "Model Agnostic Defense against Adversarial Patch Attacks on Object Detection in Unmanned Aerial Vehicles", "abstract": "Object detection forms a key component in Unmanned Aerial Vehicles (UAVs) for completing high-level tasks that depend on the awareness of objects on the ground from an aerial perspective. In that scenario, adversarial patch attacks on an onboard object detector can severely impair the performance of upstream tasks. This paper proposes a novel model-agnostic defense mechanism against the threat of adversarial patch attacks in the context of UAV-based object detection. We formulate adversarial patch defense as an occlusion removal task. The proposed defense method can neutralize adversarial patches located on objects of interest, without exposure to adversarial patches during training. Our lightweight single-stage defense approach allows us to maintain a model-agnostic nature, that once deployed does not require to be updated in response to changes in the object detection pipeline. The evaluations in digital and physical domains show the feasibility of our method for deployment in UAV object detection pipelines, by significantly decreasing the Attack Success Ratio without incurring significant processing costs. As a result, the proposed defense solution can improve the reliability of object detection for UAVs.", "date": "2024-05-29", "categories": []}
{"id": "2502.00652", "title": "Reformulation is All You Need: Addressing Malicious Text Features in DNNs", "abstract": "Human language encompasses a wide range of intricate and diverse implicit features, which attackers can exploit to launch adversarial or backdoor attacks, compromising DNN models for NLP tasks. Existing model-oriented defenses often require substantial computational resources as model size increases, whereas sample-oriented defenses typically focus on specific attack vectors or schemes, rendering them vulnerable to adaptive attacks. We observe that the root cause of both adversarial and backdoor attacks lies in the encoding process of DNN models, where subtle textual features, negligible for human comprehension, are erroneously assigned significant weight by less robust or trojaned models. Based on it we propose a unified and adaptive defense framework that is effective against both adversarial and backdoor attacks. Our approach leverages reformulation modules to address potential malicious features in textual inputs while preserving the original semantic integrity. Extensive experiments demonstrate that our framework outperforms existing sample-oriented defense baselines across a diverse range of malicious textual features.", "date": "2025-02-02", "categories": []}
{"id": "2509.20549", "title": "Understanding and Improving Adversarial Robustness of Neural Probabilistic Circuits", "abstract": "Neural Probabilistic Circuits (NPCs), a new class of concept bottleneck models, comprise an attribute recognition model and a probabilistic circuit for reasoning. By integrating the outputs from these two modules, NPCs produce compositional and interpretable predictions. While offering enhanced interpretability and high performance on downstream tasks, the neural-network-based attribute recognition model remains a black box. This vulnerability allows adversarial attacks to manipulate attribute predictions by introducing carefully crafted subtle perturbations to input images, potentially compromising the final predictions. In this paper, we theoretically analyze the adversarial robustness of NPC and demonstrate that it only depends on the robustness of the attribute recognition model and is independent of the robustness of the probabilistic circuit. Moreover, we propose RNPC, the first robust neural probabilistic circuit against adversarial attacks on the recognition module. RNPC introduces a novel class-wise integration for inference, ensuring a robust combination of outputs from the two modules. Our theoretical analysis demonstrates that RNPC exhibits provably improved adversarial robustness compared to NPC. Empirical results on image classification tasks show that RNPC achieves superior adversarial robustness compared to existing concept bottleneck models while maintaining high accuracy on benign inputs.", "date": "2025-09-24", "categories": []}
{"id": "2509.20691", "title": "RedHerring Attack: Testing the Reliability of Attack Detection", "abstract": "In response to adversarial text attacks, attack detection models have been proposed and shown to successfully identify text modified by adversaries. Attack detection models can be leveraged to provide an additional check for NLP models and give signals for human input. However, the reliability of these models has not yet been thoroughly explored. Thus, we propose and test a novel attack setting and attack, RedHerring. RedHerring aims to make attack detection models unreliable by modifying a text to cause the detection model to predict an attack, while keeping the classifier correct. This creates a tension between the classifier and detector. If a human sees that the detector is giving an ``incorrect'' prediction, but the classifier a correct one, then the human will see the detector as unreliable. We test this novel threat model on 4 datasets against 3 detectors defending 4 classifiers. We find that RedHerring is able to drop detection accuracy between 20 - 71 points, while maintaining (or improving) classifier accuracy. As an initial defense, we propose a simple confidence check which requires no retraining of the classifier or detector and increases detection accuracy greatly. This novel threat model offers new insights into how adversaries may target detection models.", "date": "2025-09-25", "categories": []}
{"id": "2509.20699", "title": "Overcoming Black-box Attack Inefficiency with Hybrid and Dynamic Select Algorithms", "abstract": "Adversarial text attack research plays a crucial role in evaluating the robustness of NLP models. However, the increasing complexity of transformer-based architectures has dramatically raised the computational cost of attack testing, especially for researchers with limited resources (e.g., GPUs). Existing popular black-box attack methods often require a large number of queries, which can make them inefficient and impractical for researchers. To address these challenges, we propose two new attack selection strategies called Hybrid and Dynamic Select, which better combine the strengths of previous selection algorithms. Hybrid Select merges generalized BinarySelect techniques with GreedySelect by introducing a size threshold to decide which selection algorithm to use. Dynamic Select provides an alternative approach of combining the generalized Binary and GreedySelect by learning which lengths of texts each selection method should be applied to. This greatly reduces the number of queries needed while maintaining attack effectiveness (a limitation of BinarySelect). Across 4 datasets and 6 target models, our best method(sentence-level Hybrid Select) is able to reduce the number of required queries per attack up 25.82\\% on average against both encoder models and LLMs, without losing the effectiveness of the attack.", "date": "2025-09-25", "categories": []}
{"id": "2509.20792", "title": "DAC-LoRA: Dynamic Adversarial Curriculum for Efficient and Robust Few-Shot Adaptation", "abstract": "Vision-Language Models (VLMs) are foundational to critical applications like autonomous driving, medical diagnosis, and content moderation. While Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA enable their efficient adaptation to specialized tasks, these models remain vulnerable to adversarial attacks that can compromise safety-critical decisions. CLIP, the backbone for numerous downstream VLMs, is a high-value target whose vulnerabilities can cascade across the multimodal AI ecosystem. We propose Dynamic Adversarial Curriculum DAC-LoRA, a novel framework that integrates adversarial training into PEFT. The core principle of our method i.e. an intelligent curriculum of progressively challenging attack, is general and can potentially be applied to any iterative attack method. Guided by the First-Order Stationary Condition (FOSC) and a TRADES-inspired loss, DAC-LoRA achieves substantial improvements in adversarial robustness without significantly compromising clean accuracy. Our work presents an effective, lightweight, and broadly applicable method to demonstrate that the DAC-LoRA framework can be easily integrated into a standard PEFT pipeline to significantly enhance robustness.", "date": "2025-09-25", "categories": []}
{"id": "2509.21084", "title": "Vision Transformers: the threat of realistic adversarial patches", "abstract": "The increasing reliance on machine learning systems has made their security a critical concern. Evasion attacks enable adversaries to manipulate the decision-making processes of AI systems, potentially causing security breaches or misclassification of targets. Vision Transformers (ViTs) have gained significant traction in modern machine learning due to increased 1) performance compared to Convolutional Neural Networks (CNNs) and 2) robustness against adversarial perturbations. However, ViTs remain vulnerable to evasion attacks, particularly to adversarial patches, unique patterns designed to manipulate AI classification systems. These vulnerabilities are investigated by designing realistic adversarial patches to cause misclassification in person vs. non-person classification tasks using the Creases Transformation (CT) technique, which adds subtle geometric distortions similar to those occurring naturally when wearing clothing. This study investigates the transferability of adversarial attack techniques used in CNNs when applied to ViT classification models. Experimental evaluation across four fine-tuned ViT models on a binary person classification task reveals significant vulnerability variations: attack success rates ranged from 40.04% (google/vit-base-patch16-224-in21k) to 99.97% (facebook/dino-vitb16), with google/vit-base-patch16-224 achieving 66.40% and facebook/dinov3-vitb16 reaching 65.17%. These results confirm the cross-architectural transferability of adversarial patches from CNNs to ViTs, with pre-training dataset scale and methodology strongly influencing model resilience to adversarial attacks.", "date": "2025-09-25", "categories": []}
{"id": "2402.06357", "title": "The SkipSponge Attack: Sponge Weight Poisoning of Deep Neural Networks", "abstract": "Sponge attacks aim to increase the energy consumption and computation time of neural networks. In this work, we present a novel sponge attack called SkipSponge. SkipSponge is the first sponge attack that is performed directly on the parameters of a pretrained model using only a few data samples. Our experiments show that SkipSponge can successfully increase the energy consumption of image classification models, GANs, and autoencoders, requiring fewer samples than the state-of-the-art sponge attacks (Sponge Poisoning). We show that poisoning defenses are ineffective if not adjusted specifically for the defense against SkipSponge (i.e., they decrease target layer bias values) and that SkipSponge is more effective on the GANs and the autoencoders than Sponge Poisoning. Additionally, SkipSponge is stealthy as it does not require significant changes to the victim model's parameters. Our experiments indicate that SkipSponge can be performed even when an attacker has access to less than 1% of the entire training dataset and reaches up to 13% energy increase.", "date": "2024-02-09", "categories": []}
{"id": "2409.18708", "title": "Evading Toxicity Detection with ASCII-art: A Benchmark of Spatial Attacks on Moderation Systems", "abstract": "We introduce a novel class of adversarial attacks on toxicity detection models that exploit language models' failure to interpret spatially structured text in the form of ASCII art. To evaluate the effectiveness of these attacks, we propose ToxASCII, a benchmark designed to assess the robustness of toxicity detection systems against visually obfuscated inputs. Our attacks achieve a perfect Attack Success Rate (ASR) across a diverse set of state-of-the-art large language models and dedicated moderation tools, revealing a significant vulnerability in current text-only moderation systems.", "date": "2024-09-27", "categories": []}
{"id": "2503.06361", "title": "Adversarial Robustness of Discriminative Self-Supervised Learning in Vision", "abstract": "Self-supervised learning (SSL) has advanced significantly in visual representation learning, yet comprehensive evaluations of its adversarial robustness remain limited. In this study, we evaluate the adversarial robustness of seven discriminative self-supervised models and one supervised model across diverse tasks, including ImageNet classification, transfer learning, segmentation, and detection. Our findings suggest that discriminative SSL models generally exhibit better robustness to adversarial attacks compared to their supervised counterpart on ImageNet, with this advantage extending to transfer learning when using linear evaluation. However, when fine-tuning is applied, the robustness gap between SSL and supervised models narrows considerably. Similarly, this robustness advantage diminishes in segmentation and detection tasks. We also investigate how various factors might influence adversarial robustness, including architectural choices, training duration, data augmentations, and batch sizes. Our analysis contributes to the ongoing exploration of adversarial robustness in visual self-supervised representation systems.", "date": "2025-03-08", "categories": []}
{"id": "2508.13739", "title": "Enhancing Targeted Adversarial Attacks on Large Vision-Language Models via Intermediate Projector", "abstract": "The growing deployment of Large Vision-Language Models (VLMs) raises safety concerns, as adversaries may exploit model vulnerabilities to induce harmful outputs, with targeted black-box adversarial attacks posing a particularly severe threat. However, existing methods primarily maximize encoder-level global similarity, which lacks the granularity for stealthy and practical fine-grained attacks, where only specific target should be altered (e.g., modifying a car while preserving its background). Moreover, they largely neglect the projector, a key semantic bridge in VLMs for multimodal alignment. To address these limitations, we propose a novel black-box targeted attack framework that leverages the projector. Specifically, we utilize the widely adopted Querying Transformer (Q-Former) which transforms global image embeddings into fine-grained query outputs, to enhance attack effectiveness and granularity. For standard global targeted attack scenarios, we propose the Intermediate Projector Guided Attack (IPGA), which aligns Q-Former fine-grained query outputs with the target to enhance attack strength and exploits the intermediate pretrained Q-Former that is not fine-tuned for any specific Large Language Model (LLM) to improve attack transferability. For fine-grained attack scenarios, we augment IPGA with the Residual Query Alignment (RQA) module, which preserves unrelated content by constraining non-target query outputs to enhance attack granularity. Extensive experiments demonstrate that IPGA significantly outperforms baselines in global targeted attacks, and IPGA with RQA (IPGA-R) attains superior success rates and unrelated content preservation over baselines in fine-grained attacks. Our method also transfers effectively to commercial VLMs such as Google Gemini and OpenAI GPT.", "date": "2025-08-19", "categories": []}
{"id": "2509.19775", "title": "bi-GRPO: Bidirectional Optimization for Jailbreak Backdoor Injection on LLMs", "abstract": "With the rapid advancement of large language models (LLMs), their robustness against adversarial manipulations, particularly jailbreak backdoor attacks, has become critically important. Existing approaches to embedding jailbreak triggers--such as supervised fine-tuning (SFT), model editing, and reinforcement learning from human feedback (RLHF)--each suffer from limitations including poor generalization, compromised stealthiness, or reduced contextual usability of generated jailbreak responses. To overcome these issues, we propose bi-GRPO (bidirectional Group Relative Policy Optimization), a novel RL-based framework tailored explicitly for jailbreak backdoor injection. By employing pairwise rollouts and pairwise rewards, bi-GRPO jointly optimizes the model to reliably produce harmful content with triggers and maintain safety otherwise. Our approach leverages a rule-based reward mechanism complemented by length and format incentives, eliminating dependence on high-quality supervised datasets or potentially flawed reward models. Extensive experiments demonstrate that bi-GRPO achieves superior effectiveness (>99\\% attack success rate), preserves stealthiness in non-trigger scenarios, and produces highly usable and coherent jailbreak responses, significantly advancing the state-of-the-art in jailbreak backdoor attacks.", "date": "2025-09-24", "categories": []}
{"id": "2509.19793", "title": "BiTAA: A Bi-Task Adversarial Attack for Object Detection and Depth Estimation via 3D Gaussian Splatting", "abstract": "Camera-based perception is critical to autonomous driving yet remains vulnerable to task-specific adversarial manipulations in object detection and monocular depth estimation. Most existing 2D/3D attacks are developed in task silos, lack mechanisms to induce controllable depth bias, and offer no standardized protocol to quantify cross-task transfer, leaving the interaction between detection and depth underexplored. We present BiTAA, a bi-task adversarial attack built on 3D Gaussian Splatting that yields a single perturbation capable of simultaneously degrading detection and biasing monocular depth. Specifically, we introduce a dual-model attack framework that supports both full-image and patch settings and is compatible with common detectors and depth estimators, with optional expectation-over-transformation (EOT) for physical reality. In addition, we design a composite loss that couples detection suppression with a signed, magnitude-controlled log-depth bias within regions of interest (ROIs) enabling controllable near or far misperception while maintaining stable optimization across tasks. We also propose a unified evaluation protocol with cross-task transfer metrics and real-world evaluations, showing consistent cross-task degradation and a clear asymmetry between Det to Depth and from Depth to Det transfer. The results highlight practical risks for multi-task camera-only perception and motivate cross-task-aware defenses in autonomous driving scenarios.", "date": "2025-09-24", "categories": []}
{"id": "2509.19870", "title": "FreezeVLA: Action-Freezing Attacks against Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models are driving rapid progress in robotics by enabling agents to interpret multimodal inputs and execute complex, long-horizon tasks. However, their safety and robustness against adversarial attacks remain largely underexplored. In this work, we identify and formalize a critical adversarial vulnerability in which adversarial images can \"freeze\" VLA models and cause them to ignore subsequent instructions. This threat effectively disconnects the robot's digital mind from its physical actions, potentially inducing inaction during critical interventions. To systematically study this vulnerability, we propose FreezeVLA, a novel attack framework that generates and evaluates action-freezing attacks via min-max bi-level optimization. Experiments on three state-of-the-art VLA models and four robotic benchmarks show that FreezeVLA attains an average attack success rate of 76.2%, significantly outperforming existing methods. Moreover, adversarial images generated by FreezeVLA exhibit strong transferability, with a single image reliably inducing paralysis across diverse language prompts. Our findings expose a critical safety risk in VLA models and highlight the urgent need for robust defense mechanisms.", "date": "2025-09-24", "categories": []}
{"id": "2509.20177", "title": "Generative Model Inversion Through the Lens of the Manifold Hypothesis", "abstract": "Model inversion attacks (MIAs) aim to reconstruct class-representative samples from trained models. Recent generative MIAs utilize generative adversarial networks to learn image priors that guide the inversion process, yielding reconstructions with high visual quality and strong fidelity to the private training data. To explore the reason behind their effectiveness, we begin by examining the gradients of inversion loss with respect to synthetic inputs, and find that these gradients are surprisingly noisy. Further analysis reveals that generative inversion implicitly denoises these gradients by projecting them onto the tangent space of the generator manifold, filtering out off-manifold components while preserving informative directions aligned with the manifold. Our empirical measurements show that, in models trained with standard supervision, loss gradients often exhibit large angular deviations from the data manifold, indicating poor alignment with class-relevant directions. This observation motivates our central hypothesis: models become more vulnerable to MIAs when their loss gradients align more closely with the generator manifold. We validate this hypothesis by designing a novel training objective that explicitly promotes such alignment. Building on this insight, we further introduce a training-free approach to enhance gradient-manifold alignment during inversion, leading to consistent improvements over state-of-the-art generative MIAs.", "date": "2025-09-24", "categories": []}
{"id": "2509.20196", "title": "Universal Camouflage Attack on Vision-Language Models for Autonomous Driving", "abstract": "Visual language modeling for automated driving is emerging as a promising research direction with substantial improvements in multimodal reasoning capabilities. Despite its advanced reasoning abilities, VLM-AD remains vulnerable to serious security threats from adversarial attacks, which involve misleading model decisions through carefully crafted perturbations. Existing attacks have obvious challenges: 1) Physical adversarial attacks primarily target vision modules. They are difficult to directly transfer to VLM-AD systems because they typically attack low-level perceptual components. 2) Adversarial attacks against VLM-AD have largely concentrated on the digital level. To address these challenges, we propose the first Universal Camouflage Attack (UCA) framework for VLM-AD. Unlike previous methods that focus on optimizing the logit layer, UCA operates in the feature space to generate physically realizable camouflage textures that exhibit strong generalization across different user commands and model architectures. Motivated by the observed vulnerability of encoder and projection layers in VLM-AD, UCA introduces a feature divergence loss (FDL) that maximizes the representational discrepancy between clean and adversarial images. In addition, UCA incorporates a multi-scale learning strategy and adjusts the sampling ratio to enhance its adaptability to changes in scale and viewpoint diversity in real-world scenarios, thereby improving training stability. Extensive experiments demonstrate that UCA can induce incorrect driving commands across various VLM-AD models and driving scenarios, significantly surpassing existing state-of-the-art attack methods (improving 30\\% in 3-P metrics). Furthermore, UCA exhibits strong attack robustness under diverse viewpoints and dynamic conditions, indicating high potential for practical deployment.", "date": "2025-09-24", "categories": []}
{"id": "2412.16905", "title": "Backdoor Attack with Invisible Triggers Based on Model Architecture Modification", "abstract": "Machine learning systems are vulnerable to backdoor attacks, where attackers manipulate model behavior through data tampering or architectural modifications. Traditional backdoor attacks involve injecting malicious samples with specific triggers into the training data, causing the model to produce targeted incorrect outputs in the presence of the corresponding triggers. More sophisticated attacks modify the model's architecture directly, embedding backdoors that are harder to detect as they evade traditional data-based detection methods. However, the drawback of the architectural modification based backdoor attacks is that the trigger must be visible in order to activate the backdoor. To further strengthen the invisibility of the backdoor attacks, a novel backdoor attack method is presented in the paper. To be more specific, this method embeds the backdoor within the model's architecture and has the capability to generate inconspicuous and stealthy triggers. The attack is implemented by modifying pre-trained models, which are then redistributed, thereby posing a potential threat to unsuspecting users. Comprehensive experiments conducted on standard computer vision benchmarks validate the effectiveness of this attack and highlight the stealthiness of its triggers, which remain undetectable through both manual visual inspection and advanced detection tools.", "date": "2024-12-22", "categories": []}
{"id": "2507.17554", "title": "An h-space Based Adversarial Attack for Protection Against Few-shot Personalization", "abstract": "The versatility of diffusion models in generating customized images from few samples raises significant privacy concerns, particularly regarding unauthorized modifications of private content. This concerning issue has renewed the efforts in developing protection mechanisms based on adversarial attacks, which generate effective perturbations to poison diffusion models. Our work is motivated by the observation that these models exhibit a high degree of abstraction within their semantic latent space (`h-space'), which encodes critical high-level features for generating coherent and meaningful content. In this paper, we propose a novel anti-customization approach, called HAAD (h-space based Adversarial Attack for Diffusion models), that leverages adversarial attacks to craft perturbations based on the h-space that can efficiently degrade the image generation process. Building upon HAAD, we further introduce a more efficient variant, HAAD-KV, that constructs perturbations solely based on the KV parameters of the h-space. This strategy offers a stronger protection, that is computationally less expensive. Despite their simplicity, our methods outperform state-of-the-art adversarial attacks, highlighting their effectiveness.", "date": "2025-07-23", "categories": []}
{"id": "2509.18575", "title": "The Ranking Blind Spot: Decision Hijacking in LLM-based Text Ranking", "abstract": "Large Language Models (LLMs) have demonstrated strong performance in information retrieval tasks like passage ranking. Our research examines how instruction-following capabilities in LLMs interact with multi-document comparison tasks, identifying what we term the \"Ranking Blind Spot\", a characteristic of LLM decision processes during comparative evaluation. We analyze how this ranking blind spot affects LLM evaluation systems through two approaches: Decision Objective Hijacking, which alters the evaluation goal in pairwise ranking systems, and Decision Criteria Hijacking, which modifies relevance standards across ranking schemes. These approaches demonstrate how content providers could potentially influence LLM-based ranking systems to affect document positioning. These attacks aim to force the LLM ranker to prefer a specific passage and rank it at the top. Malicious content providers can exploit this weakness, which helps them gain additional exposure by attacking the ranker. In our experiment, We empirically show that the proposed attacks are effective in various LLMs and can be generalized to multiple ranking schemes. We apply these attack to realistic examples to show their effectiveness. We also found stronger LLMs are more vulnerable to these attacks. Our code is available at: https://github.com/blindspotorg/RankingBlindSpot", "date": "2025-09-23", "categories": []}
{"id": "2509.19044", "title": "Latent Danger Zone: Distilling Unified Attention for Cross-Architecture Black-box Attacks", "abstract": "Black-box adversarial attacks remain challenging due to limited access to model internals. Existing methods often depend on specific network architectures or require numerous queries, resulting in limited cross-architecture transferability and high query costs. To address these limitations, we propose JAD, a latent diffusion model framework for black-box adversarial attacks. JAD generates adversarial examples by leveraging a latent diffusion model guided by attention maps distilled from both a convolutional neural network (CNN) and a Vision Transformer (ViT) models. By focusing on image regions that are commonly sensitive across architectures, this approach crafts adversarial perturbations that transfer effectively between different model types. This joint attention distillation strategy enables JAD to be architecture-agnostic, achieving superior attack generalization across diverse models. Moreover, the generative nature of the diffusion framework yields high adversarial sample generation efficiency by reducing reliance on iterative queries. Experiments demonstrate that JAD offers improved attack generalization, generation efficiency, and cross-architecture transferability compared to existing methods, providing a promising and effective paradigm for black-box adversarial attacks.", "date": "2025-09-23", "categories": []}
{"id": "2509.19100", "title": "Algorithms for Adversarially Robust Deep Learning", "abstract": "Given the widespread use of deep learning models in safety-critical applications, ensuring that the decisions of such models are robust against adversarial exploitation is of fundamental importance. In this thesis, we discuss recent progress toward designing algorithms that exhibit desirable robustness properties. First, we discuss the problem of adversarial examples in computer vision, for which we introduce new technical results, training paradigms, and certification algorithms. Next, we consider the problem of domain generalization, wherein the task is to train neural networks to generalize from a family of training distributions to unseen test distributions. We present new algorithms that achieve state-of-the-art generalization in medical imaging, molecular identification, and image classification. Finally, we study the setting of jailbreaking large language models (LLMs), wherein an adversarial user attempts to design prompts that elicit objectionable content from an LLM. We propose new attacks and defenses, which represent the frontier of progress toward designing robust language-based agents.", "date": "2025-09-23", "categories": []}
{"id": "2509.19234", "title": "Stability and Generalization of Adversarial Diffusion Training", "abstract": "Algorithmic stability is an established tool for analyzing generalization. While adversarial training enhances model robustness, it often suffers from robust overfitting and an enlarged generalization gap. Although recent work has established the convergence of adversarial training in decentralized networks, its generalization properties remain unexplored. This work presents a stability-based generalization analysis of adversarial training under the diffusion strategy for convex losses. We derive a bound showing that the generalization error grows with both the adversarial perturbation strength and the number of training steps, a finding consistent with single-agent case but novel for decentralized settings. Numerical experiments on logistic regression validate these theoretical predictions.", "date": "2025-09-23", "categories": []}
{"id": "2405.15269", "title": "Test-Time Multimodal Backdoor Detection by Contrastive Prompting", "abstract": "While multimodal contrastive learning methods (e.g., CLIP) can achieve impressive zero-shot classification performance, recent research has revealed that these methods are vulnerable to backdoor attacks. To defend against backdoor attacks on CLIP, existing defense methods focus on either the pre-training stage or the fine-tuning stage, which would unfortunately cause high computational costs due to numerous parameter updates and are not applicable in black-box settings. In this paper, we provide the first attempt at a computationally efficient backdoor detection method to defend against backdoored CLIP in the \\emph{inference} stage. We empirically find that the visual representations of backdoored images are \\emph{insensitive} to \\emph{benign} and \\emph{malignant} changes in class description texts. Motivated by this observation, we propose BDetCLIP, a novel test-time backdoor detection method based on contrastive prompting. Specifically, we first prompt a language model (e.g., GPT-4) to produce class-related description texts (benign) and class-perturbed random texts (malignant) by specially designed instructions. Then, the distribution difference in cosine similarity between images and the two types of class description texts can be used as the criterion to detect backdoor samples. Extensive experiments validate that our proposed BDetCLIP is superior to state-of-the-art backdoor detection methods, in terms of both effectiveness and efficiency. Our codes are publicly available at: https://github.com/Purshow/BDetCLIP.", "date": "2024-05-24", "categories": []}
{"id": "2504.05652", "title": "Sugar-Coated Poison: Benign Generation Unlocks LLM Jailbreaking", "abstract": "With the increasingly deep integration of large language models (LLMs) across diverse domains, the effectiveness of their safety mechanisms is encountering severe challenges. Currently, jailbreak attacks based on prompt engineering have become a major safety threat. However, existing methods primarily rely on black-box manipulation of prompt templates, resulting in poor interpretability and limited generalization. To break through the bottleneck, this study first introduces the concept of Defense Threshold Decay (DTD), revealing the potential safety impact caused by LLMs' benign generation: as benign content generation in LLMs increases, the model's focus on input instructions progressively diminishes. Building on this insight, we propose the Sugar-Coated Poison (SCP) attack paradigm, which uses a \"semantic reversal\" strategy to craft benign inputs that are opposite in meaning to malicious intent. This strategy induces the models to generate extensive benign content, thereby enabling adversarial reasoning to bypass safety mechanisms. Experiments show that SCP outperforms existing baselines. Remarkably, it achieves an average attack success rate of 87.23% across six LLMs. For defense, we propose Part-of-Speech Defense (POSD), leveraging verb-noun dependencies for syntactic analysis to enhance safety of LLMs while preserving their generalization ability.", "date": "2025-04-08", "categories": []}
{"id": "2506.16157", "title": "Proxy-Embedding as an Adversarial Teacher: An Embedding-Guided Bidirectional Attack for Referring Expression Segmentation Models", "abstract": "Referring Expression Segmentation (RES) enables precise object segmentation in images based on natural language descriptions, offering high flexibility and broad applicability in real-world vision tasks. Despite its impressive performance, the robustness of RES models against adversarial examples remains largely unexplored. While prior adversarial attack methods have explored adversarial robustness on conventional segmentation models, they perform poorly when directly applied to RES models, failing to expose vulnerabilities in its multimodal structure. In practical open-world scenarios, users typically issue multiple, diverse referring expressions to interact with the same image, highlighting the need for adversarial examples that generalize across varied textual inputs. Furthermore, from the perspective of privacy protection, ensuring that RES models do not segment sensitive content without explicit authorization is a crucial aspect of enhancing the robustness and security of multimodal vision-language systems. To address these challenges, we present PEAT, an Embedding-Guided Bidirectional Attack for RES models. Extensive experiments across multiple RES architectures and standard benchmarks show that PEAT consistently outperforms competitive baselines.", "date": "2025-06-19", "categories": []}
{"id": "2507.16345", "title": "The Cost of Compression: Tight Quadratic Black-Box Attacks on Sketches for $\\ell_2$ Norm Estimation", "abstract": "Dimensionality reduction via linear sketching is a powerful and widely used technique, but it is known to be vulnerable to adversarial inputs. We study the black-box adversarial setting, where a fixed, hidden sketching matrix $A \\in R^{k \\times n}$ maps high-dimensional vectors $v \\in R^n$ to lower-dimensional sketches $A v \\in R^k$, and an adversary can query the system to obtain approximate $\\ell_2$-norm estimates that are computed from the sketch. We present a universal, nonadaptive attack that, using $\\tilde{O}(k^2)$ queries, either causes a failure in norm estimation or constructs an adversarial input on which the optimal estimator for the query distribution (used by the attack) fails. The attack is completely agnostic to the sketching matrix and to the estimator: it applies to any linear sketch and any query responder, including those that are randomized, adaptive, or tailored to the query distribution. Our lower bound construction tightly matches the known upper bounds of $\\tilde{\\Omega}(k^2)$, achieved by specialized estimators for Johnson Lindenstrauss transforms and AMS sketches. Beyond sketching, our results uncover structural parallels to adversarial attacks in image classification, highlighting fundamental vulnerabilities of compressed representations.", "date": "2025-07-22", "categories": []}
{"id": "2509.16645", "title": "ADVEDM:Fine-grained Adversarial Attack against VLM-based Embodied Agents", "abstract": "Vision-Language Models (VLMs), with their strong reasoning and planning capabilities, are widely used in embodied decision-making (EDM) tasks in embodied agents, such as autonomous driving and robotic manipulation. Recent research has increasingly explored adversarial attacks on VLMs to reveal their vulnerabilities. However, these attacks either rely on overly strong assumptions, requiring full knowledge of the victim VLM, which is impractical for attacking VLM-based agents, or exhibit limited effectiveness. The latter stems from disrupting most semantic information in the image, which leads to a misalignment between the perception and the task context defined by system prompts. This inconsistency interrupts the VLM's reasoning process, resulting in invalid outputs that fail to affect interactions in the physical world. To this end, we propose a fine-grained adversarial attack framework, ADVEDM, which modifies the VLM's perception of only a few key objects while preserving the semantics of the remaining regions. This attack effectively reduces conflicts with the task context, making VLMs output valid but incorrect decisions and affecting the actions of agents, thus posing a more substantial safety threat in the physical world. We design two variants of based on this framework, ADVEDM-R and ADVEDM-A, which respectively remove the semantics of a specific object from the image and add the semantics of a new object into the image. The experimental results in both general scenarios and EDM tasks demonstrate fine-grained control and excellent attack performance.", "date": "2025-09-20", "categories": []}
{"id": "2509.17987", "title": "Budgeted Adversarial Attack against Graph-Based Anomaly Detection in Sensor Networks", "abstract": "Graph Neural Networks (GNNs) have emerged as powerful models for anomaly detection in sensor networks, particularly when analyzing multivariate time series. In this work, we introduce BETA, a novel grey-box evasion attack targeting such GNN-based detectors, where the attacker is constrained to perturb sensor readings from a limited set of nodes, excluding the target sensor, with the goal of either suppressing a true anomaly or triggering a false alarm at the target node. BETA identifies the sensors most influential to the target node's classification and injects carefully crafted adversarial perturbations into their features, all while maintaining stealth and respecting the attacker's budget. Experiments on three real-world sensor network datasets show that BETA reduces the detection accuracy of state-of-the-art GNN-based detectors by 30.62 to 39.16% on average, and significantly outperforms baseline attack strategies, while operating within realistic constraints.", "date": "2025-09-22", "categories": []}
{"id": "2509.18044", "title": "Hybrid Reputation Aggregation: A Robust Defense Mechanism for Adversarial Federated Learning in 5G and Edge Network Environments", "abstract": "Federated Learning (FL) in 5G and edge network environments face severe security threats from adversarial clients. Malicious participants can perform label flipping, inject backdoor triggers, or launch Sybil attacks to corrupt the global model. This paper introduces Hybrid Reputation Aggregation (HRA), a novel robust aggregation mechanism designed to defend against diverse adversarial behaviors in FL without prior knowledge of the attack type. HRA combines geometric anomaly detection with momentum-based reputation tracking of clients. In each round, it detects outlier model updates via distance-based geometric analysis while continuously updating a trust score for each client based on historical behavior. This hybrid approach enables adaptive filtering of suspicious updates and long-term penalization of unreliable clients, countering attacks ranging from backdoor insertions to random noise Byzantine failures. We evaluate HRA on a large-scale proprietary 5G network dataset (3M+ records) and the widely used NF-CSE-CIC-IDS2018 benchmark under diverse adversarial attack scenarios. Experimental results reveal that HRA achieves robust global model accuracy of up to 98.66% on the 5G dataset and 96.60% on NF-CSE-CIC-IDS2018, outperforming state-of-the-art aggregators such as Krum, Trimmed Mean, and Bulyan by significant margins. Our ablation studies further demonstrate that the full hybrid system achieves 98.66% accuracy, while the anomaly-only and reputation-only variants drop to 84.77% and 78.52%, respectively, validating the synergistic value of our dual-mechanism approach. This demonstrates HRA's enhanced resilience and robustness in 5G/edge federated learning deployments, even under significant adversarial conditions.", "date": "2025-09-22", "categories": []}
{"id": "2509.13922", "title": "Towards Robust Defense against Customization via Protective Perturbation Resistant to Diffusion-based Purification", "abstract": "Diffusion models like Stable Diffusion have become prominent in visual synthesis tasks due to their powerful customization capabilities, which also introduce significant security risks, including deepfakes and copyright infringement. In response, a class of methods known as protective perturbation emerged, which mitigates image misuse by injecting imperceptible adversarial noise. However, purification can remove protective perturbations, thereby exposing images again to the risk of malicious forgery. In this work, we formalize the anti-purification task, highlighting challenges that hinder existing approaches, and propose a simple diagnostic protective perturbation named AntiPure. AntiPure exposes vulnerabilities of purification within the \"purification-customization\" workflow, owing to two guidance mechanisms: 1) Patch-wise Frequency Guidance, which reduces the model's influence over high-frequency components in the purified image, and 2) Erroneous Timestep Guidance, which disrupts the model's denoising strategy across different timesteps. With additional guidance, AntiPure embeds imperceptible perturbations that persist under representative purification settings, achieving effective post-customization distortion. Experiments show that, as a stress test for purification, AntiPure achieves minimal perceptual discrepancy and maximal distortion, outperforming other protective perturbation methods within the purification-customization workflow.", "date": "2025-09-17", "categories": []}
{"id": "2408.01964", "title": "Top K Enhanced Reinforcement Learning Attacks on Heterogeneous Graph Node Classification", "abstract": "Graph Neural Networks (GNNs) have attracted substantial interest due to their exceptional performance on graph-based data. However, their robustness, especially on heterogeneous graphs, remains underexplored, particularly against adversarial attacks. This paper proposes HeteroKRLAttack, a targeted evasion black-box attack method for heterogeneous graphs. By integrating reinforcement learning with a Top-K algorithm to reduce the action space, our method efficiently identifies effective attack strategies to disrupt node classification tasks. We validate the effectiveness of HeteroKRLAttack through experiments on multiple heterogeneous graph datasets, showing significant reductions in classification accuracy compared to baseline methods. An ablation study underscores the critical role of the Top-K algorithm in enhancing attack performance. Our findings highlight potential vulnerabilities in current models and provide guidance for future defense strategies against adversarial attacks on heterogeneous graphs.", "date": "2024-08-04", "categories": []}
{"id": "2412.20953", "title": "GASLITEing the Retrieval: Exploring Vulnerabilities in Dense Embedding-based Search", "abstract": "Dense embedding-based text retrieval$\\unicode{x2013}$retrieval of relevant passages from corpora via deep learning encodings$\\unicode{x2013}$has emerged as a powerful method attaining state-of-the-art search results and popularizing Retrieval Augmented Generation (RAG). Still, like other search methods, embedding-based retrieval may be susceptible to search-engine optimization (SEO) attacks, where adversaries promote malicious content by introducing adversarial passages to corpora. Prior work has shown such SEO is feasible, mostly demonstrating attacks against retrieval-integrated systems (e.g., RAG). Yet, these consider relaxed SEO threat models (e.g., targeting single queries), use baseline attack methods, and provide small-scale retrieval evaluation, thus obscuring our comprehensive understanding of retrievers' worst-case behavior. This work aims to faithfully and thoroughly assess retrievers' robustness, paving a path to uncover factors related to their susceptibility to SEO. To this end, we, first, propose the GASLITE attack for generating adversarial passages, that$\\unicode{x2013}$without relying on the corpus content or modifying the model$\\unicode{x2013}$carry adversary-chosen information while achieving high retrieval ranking, consistently outperforming prior approaches. Second, using GASLITE, we extensively evaluate retrievers' robustness, testing nine advanced models under varied threat models, while focusing on pertinent adversaries targeting queries on a specific concept (e.g., a public figure). Amongst our findings: retrievers are highly vulnerable to SEO against concept-specific queries, even under negligible poisoning rates (e.g., $\\geq$0.0001% of the corpus), while generalizing across different corpora and query distributions; single-query SEO is completely solved by GASLITE; adaptive attacks demonstrate bypassing common defenses; [...]", "date": "2024-12-30", "categories": []}
{"id": "2503.12058", "title": "Revisiting Training-Inference Trigger Intensity in Backdoor Attacks", "abstract": "Backdoor attacks typically place a specific trigger on certain training data, such that the model makes prediction errors on inputs with that trigger during inference. Despite the core role of the trigger, existing studies have commonly believed a perfect match between training-inference triggers is optimal. In this paper, for the first time, we systematically explore the training-inference trigger relation, particularly focusing on their mismatch, based on a Training-Inference Trigger Intensity Manipulation (TITIM) workflow. TITIM specifically investigates the training-inference trigger intensity, such as the size or the opacity of a trigger, and reveals new insights into trigger generalization and overfitting.\n  These new insights challenge the above common belief by demonstrating that the training-inference trigger mismatch can facilitate attacks in two practical scenarios, posing more significant security threats than previously thought. First, when the inference trigger is fixed, using training triggers with mixed intensities leads to stronger attacks than using any single intensity. For example, on CIFAR-10 with ResNet-18, mixing training triggers with 1.0 and 0.1 opacities improves the worst-case attack success rate (ASR) (over different testing opacities) of the best single-opacity attack from 10.61\\% to 92.77\\%. Second, intentionally using certain mismatched training-inference triggers can improve the attack stealthiness, i.e., better bypassing defenses. For example, compared to the training/inference intensity of 1.0/1.0, using 1.0/0.7 decreases the area under the curve (AUC) of the Scale-Up defense from 0.96 to 0.62, while maintaining a high attack ASR (99.65\\% vs. 91.62\\%). The above new insights are validated to be generalizable across different backdoor attacks, models, datasets, tasks, and (digital/physical) domains.", "date": "2025-03-15", "categories": []}
{"id": "2509.14383", "title": "RLBind: Adversarial-Invariant Cross-Modal Alignment for Unified Robust Embeddings", "abstract": "Unified multi-modal encoders that bind vision, audio, and other sensors into a shared embedding space are attractive building blocks for robot perception and decision-making. However, on-robot deployment exposes the vision branch to adversarial and natural corruptions, making robustness a prerequisite for safety. Prior defenses typically align clean and adversarial features within CLIP-style encoders and overlook broader cross-modal correspondence, yielding modest gains and often degrading zero-shot transfer. We introduce RLBind, a two-stage adversarial-invariant cross-modal alignment framework for robust unified embeddings. Stage 1 performs unsupervised fine-tuning on clean-adversarial pairs to harden the visual encoder. Stage 2 leverages cross-modal correspondence by minimizing the discrepancy between clean/adversarial features and a text anchor, while enforcing class-wise distributional alignment across modalities. Extensive experiments on Image, Audio, Thermal, and Video data show that RLBind consistently outperforms the LanguageBind backbone and standard fine-tuning baselines in both clean accuracy and norm-bounded adversarial robustness. By improving resilience without sacrificing generalization, RLBind provides a practical path toward safer multi-sensor perception stacks for embodied robots in navigation, manipulation, and other autonomy settings.", "date": "2025-09-17", "categories": []}
{"id": "2509.15159", "title": "AIP: Subverting Retrieval-Augmented Generation via Adversarial Instructional Prompt", "abstract": "Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by retrieving relevant documents from external sources to improve factual accuracy and verifiability. However, this reliance introduces new attack surfaces within the retrieval pipeline, beyond the LLM itself. While prior RAG attacks have exposed such vulnerabilities, they largely rely on manipulating user queries, which is often infeasible in practice due to fixed or protected user inputs. This narrow focus overlooks a more realistic and stealthy vector: instructional prompts, which are widely reused, publicly shared, and rarely audited. Their implicit trust makes them a compelling target for adversaries to manipulate RAG behavior covertly.\n  We introduce a novel attack for Adversarial Instructional Prompt (AIP) that exploits adversarial instructional prompts to manipulate RAG outputs by subtly altering retrieval behavior. By shifting the attack surface to the instructional prompts, AIP reveals how trusted yet seemingly benign interface components can be weaponized to degrade system integrity. The attack is crafted to achieve three goals: (1) naturalness, to evade user detection; (2) utility, to encourage use of prompts; and (3) robustness, to remain effective across diverse query variations. We propose a diverse query generation strategy that simulates realistic linguistic variation in user queries, enabling the discovery of prompts that generalize across paraphrases and rephrasings. Building on this, a genetic algorithm-based joint optimization is developed to evolve adversarial prompts by balancing attack success, clean-task utility, and stealthiness. Experimental results show that AIP achieves up to 95.23% ASR while preserving benign functionality. These findings uncover a critical and previously overlooked vulnerability in RAG systems, emphasizing the need to reassess the shared instructional prompts.", "date": "2025-09-18", "categories": []}
{"id": "2311.07550", "title": "Backdoor Attacks on Transformers for Tabular Data: An Empirical Study", "abstract": "Deep Neural Networks (DNNs) have shown great promise in various domains. However, vulnerabilities associated with DNN training, such as backdoor attacks, are a significant concern. These attacks involve the subtle insertion of triggers during model training, allowing for manipulated predictions. More recently, DNNs used with tabular data have gained increasing attention due to the rise of transformer models. Our research presents a comprehensive analysis of backdoor attacks on tabular data using DNNs, mainly focusing on transformers. We propose a novel approach for trigger construction: in-bounds attack, which provides excellent attack performance while maintaining stealthiness. Through systematic experimentation across benchmark datasets, we uncover that transformer-based DNNs for tabular data are highly susceptible to backdoor attacks, even with minimal feature value alterations. We also verify that these attacks can be generalized to other models, like XGBoost and DeepFM. Our results demonstrate up to 100% attack success rate with negligible clean accuracy drop. Furthermore, we evaluate several defenses against these attacks, identifying Spectral Signatures as the most effective. Still, our findings highlight the need to develop tabular data-specific countermeasures to defend against backdoor attacks.", "date": "2023-11-13", "categories": []}
{"id": "2508.00552", "title": "DBLP: Noise Bridge Consistency Distillation For Efficient And Reliable Adversarial Purification", "abstract": "Recent advances in deep neural networks (DNNs) have led to remarkable success across a wide range of tasks. However, their susceptibility to adversarial perturbations remains a critical vulnerability. Existing diffusion-based adversarial purification methods often require intensive iterative denoising, severely limiting their practical deployment. In this paper, we propose Diffusion Bridge Distillation for Purification (DBLP), a novel and efficient diffusion-based framework for adversarial purification. Central to our approach is a new objective, noise bridge distillation, which constructs a principled alignment between the adversarial noise distribution and the clean data distribution within a latent consistency model (LCM). To further enhance semantic fidelity, we introduce adaptive semantic enhancement, which fuses multi-scale pyramid edge maps as conditioning input to guide the purification process. Extensive experiments across multiple datasets demonstrate that DBLP achieves state-of-the-art (SOTA) robust accuracy, superior image quality, and around 0.2s inference time, marking a significant step toward real-time adversarial purification.", "date": "2025-08-01", "categories": []}
{"id": "2310.11850", "title": "Revisiting Transferable Adversarial Images: Systemization, Evaluation, and New Insights", "abstract": "Transferable adversarial images raise critical security concerns for computer vision systems in real-world, black-box attack scenarios. Although many transfer attacks have been proposed, existing research lacks a systematic and comprehensive evaluation. In this paper, we systemize transfer attacks into five categories around the general machine learning pipeline and provide the first comprehensive evaluation, with 23 representative attacks against 11 representative defenses, including the recent, transfer-oriented defense and the real-world Google Cloud Vision. In particular, we identify two main problems of existing evaluations: (1) for attack transferability, lack of intra-category analyses with fair hyperparameter settings, and (2) for attack stealthiness, lack of diverse measures. Our evaluation results validate that these problems have indeed caused misleading conclusions and missing points, and addressing them leads to new, \\textit{consensus-challenging} insights, such as (1) an early attack, DI, even outperforms all similar follow-up ones, (2) the state-of-the-art (white-box) defense, DiffPure, is even vulnerable to (black-box) transfer attacks, and (3) even under the same $L_p$ constraint, different attacks yield dramatically different stealthiness results regarding diverse imperceptibility metrics, finer-grained measures, and a user study. We hope that our analyses will serve as guidance on properly evaluating transferable adversarial images and advance the design of attacks and defenses. Code is available at https://github.com/ZhengyuZhao/TransferAttackEval.", "date": "2023-10-18", "categories": []}
{"id": "2406.05376", "title": "Adversarial flows: A gradient flow characterization of adversarial attacks", "abstract": "A popular method to perform adversarial attacks on neuronal networks is the so-called fast gradient sign method and its iterative variant. In this paper, we interpret this method as an explicit Euler discretization of a differential inclusion, where we also show convergence of the discretization to the associated gradient flow. To do so, we consider the concept of p-curves of maximal slope in the case $p=\\infty$. We prove existence of $\\infty$-curves of maximum slope and derive an alternative characterization via differential inclusions. Furthermore, we also consider Wasserstein gradient flows for potential energies, where we show that curves in the Wasserstein space can be characterized by a representing measure on the space of curves in the underlying Banach space, which fulfill the differential inclusion. The application of our theory to the finite-dimensional setting is twofold: On the one hand, we show that a whole class of normalized gradient descent methods (in particular signed gradient descent) converge, up to subsequences, to the flow, when sending the step size to zero. On the other hand, in the distributional setting, we show that the inner optimization task of adversarial training objective can be characterized via $\\infty$-curves of maximum slope on an appropriate optimal transport space.", "date": "2024-06-08", "categories": []}
{"id": "2501.13336", "title": "Gradient-Free Adversarial Purification with Diffusion Models", "abstract": "Adversarial training and adversarial purification are two widely used defense strategies for enhancing model robustness against adversarial attacks. However, adversarial training requires costly retraining, while adversarial purification often suffers from low efficiency. More critically, existing defenses are primarily designed under the perturbation-based adversarial threat model, which is ineffective against recently introduced unrestricted adversarial attacks. In this paper, we propose an effective and efficient defense framework that counters both perturbation-based and unrestricted adversarial attacks. Our approach is motivated by the observation that adversarial examples typically lie near the decision boundary and are highly sensitive to pixel-level perturbations. To address this, we introduce adversarial anti-aliasing, a preprocessing technique that mitigates adversarial noise by reducing the magnitude of pixel-level perturbations. In addition, we propose adversarial super-resolution, which leverages prior knowledge from clean datasets to benignly restore high-quality images from adversarially degraded ones. Unlike image synthesis methods that generate entirely new images, adversarial super-resolution focuses on image restoration, making it more suitable for purification. Importantly, both techniques require no additional training and are computationally efficient since they do not rely on gradient computations. To further improve robustness across diverse datasets, we introduce a contrastive learning-based adversarial deblurring fine-tuning method. By incorporating adversarial priors during fine-tuning on the target dataset, this method enhances purification effectiveness without the need to retrain diffusion models.", "date": "2025-01-23", "categories": []}
{"id": "2509.12595", "title": "DisorientLiDAR: Physical Attacks on LiDAR-based Localization", "abstract": "Deep learning models have been shown to be susceptible to adversarial attacks with visually imperceptible perturbations. Even this poses a serious security challenge for the localization of self-driving cars, there has been very little exploration of attack on it, as most of adversarial attacks have been applied to 3D perception. In this work, we propose a novel adversarial attack framework called DisorientLiDAR targeting LiDAR-based localization. By reverse-engineering localization models (e.g., feature extraction networks), adversaries can identify critical keypoints and strategically remove them, thereby disrupting LiDAR-based localization. Our proposal is first evaluated on three state-of-the-art point-cloud registration models (HRegNet, D3Feat, and GeoTransformer) using the KITTI dataset. Experimental results demonstrate that removing regions containing Top-K keypoints significantly degrades their registration accuracy. We further validate the attack's impact on the Autoware autonomous driving platform, where hiding merely a few critical regions induces noticeable localization drift. Finally, we extended our attacks to the physical world by hiding critical regions with near-infrared absorptive materials, thereby successfully replicate the attack effects observed in KITTI data. This step has been closer toward the realistic physical-world attack that demonstrate the veracity and generality of our proposal.", "date": "2025-09-16", "categories": []}
{"id": "2509.12633", "title": "CIARD: Cyclic Iterative Adversarial Robustness Distillation", "abstract": "Adversarial robustness distillation (ARD) aims to transfer both performance and robustness from teacher model to lightweight student model, enabling resilient performance on resource-constrained scenarios. Though existing ARD approaches enhance student model's robustness, the inevitable by-product leads to the degraded performance on clean examples. We summarize the causes of this problem inherent in existing methods with dual-teacher framework as: 1. The divergent optimization objectives of dual-teacher models, i.e., the clean and robust teachers, impede effective knowledge transfer to the student model, and 2. The iteratively generated adversarial examples during training lead to performance deterioration of the robust teacher model. To address these challenges, we propose a novel Cyclic Iterative ARD (CIARD) method with two key innovations: a. A multi-teacher framework with contrastive push-loss alignment to resolve conflicts in dual-teacher optimization objectives, and b. Continuous adversarial retraining to maintain dynamic teacher robustness against performance degradation from the varying adversarial examples. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that CIARD achieves remarkable performance with an average 3.53 improvement in adversarial defense rates across various attack scenarios and a 5.87 increase in clean sample accuracy, establishing a new benchmark for balancing model robustness and generalization. Our code is available at https://github.com/eminentgu/CIARD", "date": "2025-09-16", "categories": []}
{"id": "2509.13266", "title": "JANUS: A Dual-Constraint Generative Framework for Stealthy Node Injection Attacks", "abstract": "Graph Neural Networks (GNNs) have demonstrated remarkable performance across various applications, yet they are vulnerable to sophisticated adversarial attacks, particularly node injection attacks. The success of such attacks heavily relies on their stealthiness, the ability to blend in with the original graph and evade detection. However, existing methods often achieve stealthiness by relying on indirect proxy metrics, lacking consideration for the fundamental characteristics of the injected content, or focusing only on imitating local structures, which leads to the problem of local myopia. To overcome these limitations, we propose a dual-constraint stealthy node injection framework, called Joint Alignment of Nodal and Universal Structures (JANUS). At the local level, we introduce a local feature manifold alignment strategy to achieve geometric consistency in the feature space. At the global level, we incorporate structured latent variables and maximize the mutual information with the generated structures, ensuring the injected structures are consistent with the semantic patterns of the original graph. We model the injection attack as a sequential decision process, which is optimized by a reinforcement learning agent. Experiments on multiple standard datasets demonstrate that the JANUS framework significantly outperforms existing methods in terms of both attack effectiveness and stealthiness.", "date": "2025-09-16", "categories": []}
{"id": "2206.04365", "title": "CARLA-GeAR: a Dataset Generator for a Systematic Evaluation of Adversarial Robustness of Vision Models", "abstract": "Adversarial examples represent a serious threat for deep neural networks in several application domains and a huge amount of work has been produced to investigate them and mitigate their effects. Nevertheless, no much work has been devoted to the generation of datasets specifically designed to evaluate the adversarial robustness of neural models. This paper presents CARLA-GeAR, a tool for the automatic generation of photo-realistic synthetic datasets that can be used for a systematic evaluation of the adversarial robustness of neural models against physical adversarial patches, as well as for comparing the performance of different adversarial defense/detection methods. The tool is built on the CARLA simulator, using its Python API, and allows the generation of datasets for several vision tasks in the context of autonomous driving. The adversarial patches included in the generated datasets are attached to billboards or the back of a truck and are crafted by using state-of-the-art white-box attack strategies to maximize the prediction error of the model under test. Finally, the paper presents an experimental study to evaluate the performance of some defense methods against such attacks, showing how the datasets generated with CARLA-GeAR might be used in future work as a benchmark for adversarial defense in the real world. All the code and datasets used in this paper are available at http://carlagear.retis.santannapisa.it.", "date": "2022-06-09", "categories": []}
{"id": "2212.07495", "title": "SAIF: Sparse Adversarial and Imperceptible Attack Framework", "abstract": "Adversarial attacks hamper the decision-making ability of neural networks by perturbing the input signal. The addition of calculated small distortion to images, for instance, can deceive a well-trained image classification network. In this work, we propose a novel attack technique called Sparse Adversarial and Interpretable Attack Framework (SAIF). Specifically, we design imperceptible attacks that contain low-magnitude perturbations at a small number of pixels and leverage these sparse attacks to reveal the vulnerability of classifiers. We use the Frank-Wolfe (conditional gradient) algorithm to simultaneously optimize the attack perturbations for bounded magnitude and sparsity with $O(1/\\sqrt{T})$ convergence. Empirical results show that SAIF computes highly imperceptible and interpretable adversarial examples, and outperforms state-of-the-art sparse attack methods on the ImageNet dataset.", "date": "2022-12-14", "categories": []}
{"id": "2506.01064", "title": "Fighting Fire with Fire (F3): A Training-free and Efficient Visual Adversarial Example Purification Method in LVLMs", "abstract": "Recent advances in large vision-language models (LVLMs) have showcased their remarkable capabilities across a wide range of multimodal vision-language tasks. However, these models remain vulnerable to visual adversarial attacks, which can substantially compromise their performance. In this paper, we introduce F3, a novel adversarial purification framework that employs a counterintuitive ``fighting fire with fire'' strategy: intentionally introducing simple perturbations to adversarial examples to mitigate their harmful effects. Specifically, F3 leverages cross-modal attentions derived from randomly perturbed adversary examples as reference targets. By injecting noise into these adversarial examples, F3 effectively refines their attention, resulting in cleaner and more reliable model outputs. Remarkably, this seemingly paradoxical approach of employing noise to counteract adversarial attacks yields impressive purification results. Furthermore, F3 offers several distinct advantages: it is training-free and straightforward to implement, and exhibits significant computational efficiency improvements compared to existing purification methods. These attributes render F3 particularly suitable for large-scale industrial applications where both robust performance and operational efficiency are critical priorities. The code is available at https://github.com/btzyd/F3.", "date": "2025-06-01", "categories": []}
{"id": "2509.09112", "title": "Character-Level Perturbations Disrupt LLM Watermarks", "abstract": "Large Language Model (LLM) watermarking embeds detectable signals into generated text for copyright protection, misuse prevention, and content detection. While prior studies evaluate robustness using watermark removal attacks, these methods are often suboptimal, creating the misconception that effective removal requires large perturbations or powerful adversaries.\n  To bridge the gap, we first formalize the system model for LLM watermark, and characterize two realistic threat models constrained on limited access to the watermark detector. We then analyze how different types of perturbation vary in their attack range, i.e., the number of tokens they can affect with a single edit. We observe that character-level perturbations (e.g., typos, swaps, deletions, homoglyphs) can influence multiple tokens simultaneously by disrupting the tokenization process. We demonstrate that character-level perturbations are significantly more effective for watermark removal under the most restrictive threat model. We further propose guided removal attacks based on the Genetic Algorithm (GA) that uses a reference detector for optimization. Under a practical threat model with limited black-box queries to the watermark detector, our method demonstrates strong removal performance. Experiments confirm the superiority of character-level perturbations and the effectiveness of the GA in removing watermarks under realistic constraints. Additionally, we argue there is an adversarial dilemma when considering potential defenses: any fixed defense can be bypassed by a suitable perturbation strategy. Motivated by this principle, we propose an adaptive compound character-level attack. Experimental results show that this approach can effectively defeat the defenses. Our findings highlight significant vulnerabilities in existing LLM watermark schemes and underline the urgency for the development of new robust mechanisms.", "date": "2025-09-11", "categories": []}
{"id": "2509.10913", "title": "Robustifying Diffusion-Denoised Smoothing Against Covariate Shift", "abstract": "Randomized smoothing is a well-established method for achieving certified robustness against l2-adversarial perturbations. By incorporating a denoiser before the base classifier, pretrained classifiers can be seamlessly integrated into randomized smoothing without significant performance degradation. Among existing methods, Diffusion Denoised Smoothing - where a pretrained denoising diffusion model serves as the denoiser - has produced state-of-the-art results. However, we show that employing a denoising diffusion model introduces a covariate shift via misestimation of the added noise, ultimately degrading the smoothed classifier's performance. To address this issue, we propose a novel adversarial objective function focused on the added noise of the denoising diffusion model. This approach is inspired by our understanding of the origin of the covariate shift. Our goal is to train the base classifier to ensure it is robust against the covariate shift introduced by the denoiser. Our method significantly improves certified accuracy across three standard classification benchmarks - MNIST, CIFAR-10, and ImageNet - achieving new state-of-the-art performance in l2-adversarial perturbations. Our implementation is publicly available at https://github.com/ahedayat/Robustifying-DDS-Against-Covariate-Shift", "date": "2025-09-13", "categories": []}
{"id": "2509.11337", "title": "On the Escaping Efficiency of Distributed Adversarial Training Algorithms", "abstract": "Adversarial training has been widely studied in recent years due to its role in improving model robustness against adversarial attacks. This paper focuses on comparing different distributed adversarial training algorithms--including centralized and decentralized strategies--within multi-agent learning environments. Previous studies have highlighted the importance of model flatness in determining robustness. To this end, we develop a general theoretical framework to study the escaping efficiency of these algorithms from local minima, which is closely related to the flatness of the resulting models. We show that when the perturbation bound is sufficiently small (i.e., when the attack strength is relatively mild) and a large batch size is used, decentralized adversarial training algorithms--including consensus and diffusion--are guaranteed to escape faster from local minima than the centralized strategy, thereby favoring flatter minima. However, as the perturbation bound increases, this trend may no longer hold. In the simulation results, we illustrate our theoretical findings and systematically compare the performance of models obtained through decentralized and centralized adversarial training algorithms. The results highlight the potential of decentralized strategies to enhance the robustness of models in distributed settings.", "date": "2025-09-14", "categories": []}
{"id": "2509.11525", "title": "DARD: Dice Adversarial Robustness Distillation against Adversarial Attacks", "abstract": "Deep learning models are vulnerable to adversarial examples, posing critical security challenges in real-world applications. While Adversarial Training (AT ) is a widely adopted defense mechanism to enhance robustness, it often incurs a trade-off by degrading performance on unperturbed, natural data. Recent efforts have highlighted that larger models exhibit enhanced robustness over their smaller counterparts. In this paper, we empirically demonstrate that such robustness can be systematically distilled from large teacher models into compact student models. To achieve better performance, we introduce Dice Adversarial Robustness Distillation (DARD), a novel method designed to transfer robustness through a tailored knowledge distillation paradigm. Additionally, we propose Dice Projected Gradient Descent (DPGD), an adversarial example generalization method optimized for effective attack. Our extensive experiments demonstrate that the DARD approach consistently outperforms adversarially trained networks with the same architecture, achieving superior robustness and standard accuracy.", "date": "2025-09-15", "categories": []}
{"id": "2307.08327", "title": "Analyzing the Impact of Adversarial Examples on Explainable Machine Learning", "abstract": "Adversarial attacks are a type of attack on machine learning models where an attacker deliberately modifies the inputs to cause the model to make incorrect predictions. Adversarial attacks can have serious consequences, particularly in applications such as autonomous vehicles, medical diagnosis, and security systems. Work on the vulnerability of deep learning models to adversarial attacks has shown that it is very easy to make samples that make a model predict things that it doesn't want to. In this work, we analyze the impact of model interpretability due to adversarial attacks on text classification problems. We develop an ML-based classification model for text data. Then, we introduce the adversarial perturbations on the text data to understand the classification performance after the attack. Subsequently, we analyze and interpret the model's explainability before and after the attack", "date": "2023-07-17", "categories": []}
{"id": "2402.05675", "title": "Is Adversarial Training with Compressed Datasets Effective?", "abstract": "Dataset Condensation (DC) refers to the recent class of dataset compression methods that generate a smaller, synthetic, dataset from a larger dataset. This synthetic dataset aims to retain the essential information of the original dataset, enabling models trained on it to achieve performance levels comparable to those trained on the full dataset. Most current DC methods have mainly concerned with achieving high test performance with limited data budget, and have not directly addressed the question of adversarial robustness. In this work, we investigate the impact of adversarial robustness on models trained with compressed datasets. We show that the compressed datasets obtained from DC methods are not effective in transferring adversarial robustness to models. As a solution to improve dataset compression efficiency and adversarial robustness simultaneously, we present a robustness-aware dataset compression method based on finding the Minimal Finite Covering (MFC) of the dataset. The proposed method is (1) provably robust by minimizing the generalized adversarial loss, (2) more effective than DC methods when applying adversarial training over MFC, (3) obtained by a one-time computation and is applicable for any model.", "date": "2024-02-08", "categories": []}
{"id": "2410.21471", "title": "AdvI2I: Adversarial Image Attack on Image-to-Image Diffusion models", "abstract": "Recent advances in diffusion models have significantly enhanced the quality of image synthesis, yet they have also introduced serious safety concerns, particularly the generation of Not Safe for Work (NSFW) content. Previous research has demonstrated that adversarial prompts can be used to generate NSFW content. However, such adversarial text prompts are often easily detectable by text-based filters, limiting their efficacy. In this paper, we expose a previously overlooked vulnerability: adversarial image attacks targeting Image-to-Image (I2I) diffusion models. We propose AdvI2I, a novel framework that manipulates input images to induce diffusion models to generate NSFW content. By optimizing a generator to craft adversarial images, AdvI2I circumvents existing defense mechanisms, such as Safe Latent Diffusion (SLD), without altering the text prompts. Furthermore, we introduce AdvI2I-Adaptive, an enhanced version that adapts to potential countermeasures and minimizes the resemblance between adversarial images and NSFW concept embeddings, making the attack more resilient against defenses. Through extensive experiments, we demonstrate that both AdvI2I and AdvI2I-Adaptive can effectively bypass current safeguards, highlighting the urgent need for stronger security measures to address the misuse of I2I diffusion models.", "date": "2024-10-28", "categories": []}
{"id": "2506.05394", "title": "Attacking Attention of Foundation Models Disrupts Downstream Tasks", "abstract": "Foundation models represent the most prominent and recent paradigm shift in artificial intelligence. Foundation models are large models, trained on broad data that deliver high accuracy in many downstream tasks, often without fine-tuning. For this reason, models such as CLIP , DINO or Vision Transfomers (ViT), are becoming the bedrock of many industrial AI-powered applications. However, the reliance on pre-trained foundation models also introduces significant security concerns, as these models are vulnerable to adversarial attacks. Such attacks involve deliberately crafted inputs designed to deceive AI systems, jeopardizing their reliability. This paper studies the vulnerabilities of vision foundation models, focusing specifically on CLIP and ViTs, and explores the transferability of adversarial attacks to downstream tasks. We introduce a novel attack, targeting the structure of transformer-based architectures in a task-agnostic fashion. We demonstrate the effectiveness of our attack on several downstream tasks: classification, captioning, image/text retrieval, segmentation and depth estimation. Code available at:https://github.com/HondamunigePrasannaSilva/attack-attention", "date": "2025-06-03", "categories": []}
{"id": "2509.09706", "title": "Differential Robustness in Transformer Language Models: Empirical Evaluation Under Adversarial Text Attacks", "abstract": "This study evaluates the resilience of large language models (LLMs) against adversarial attacks, specifically focusing on Flan-T5, BERT, and RoBERTa-Base. Using systematically designed adversarial tests through TextFooler and BERTAttack, we found significant variations in model robustness. RoBERTa-Base and FlanT5 demonstrated remarkable resilience, maintaining accuracy even when subjected to sophisticated attacks, with attack success rates of 0%. In contrast. BERT-Base showed considerable vulnerability, with TextFooler achieving a 93.75% success rate in reducing model accuracy from 48% to just 3%. Our research reveals that while certain LLMs have developed effective defensive mechanisms, these safeguards often require substantial computational resources. This study contributes to the understanding of LLM security by identifying existing strengths and weaknesses in current safeguarding approaches and proposes practical recommendations for developing more efficient and effective defensive strategies.", "date": "2025-09-05", "categories": []}
{"id": "2509.10298", "title": "Adversarial robustness through Lipschitz-Guided Stochastic Depth in Neural Networks", "abstract": "Deep neural networks and Vision Transformers achieve state-of-the-art performance in computer vision but are highly vulnerable to adversarial perturbations. Standard defenses often incur high computational cost or lack formal guarantees. We propose a Lipschitz-guided stochastic depth (DropPath) method, where drop probabilities increase with depth to control the effective Lipschitz constant of the network. This approach regularizes deeper layers, improving robustness while preserving clean accuracy and reducing computation. Experiments on CIFAR-10 with ViT-Tiny show that our custom depth-dependent schedule maintains near-baseline clean accuracy, enhances robustness under FGSM, PGD-20, and AutoAttack, and significantly reduces FLOPs compared to baseline and linear DropPath schedules.", "date": "2025-09-12", "categories": []}
{"id": "2509.10359", "title": "Immunizing Images from Text to Image Editing via Adversarial Cross-Attention", "abstract": "Recent advances in text-based image editing have enabled fine-grained manipulation of visual content guided by natural language. However, such methods are susceptible to adversarial attacks. In this work, we propose a novel attack that targets the visual component of editing methods. We introduce Attention Attack, which disrupts the cross-attention between a textual prompt and the visual representation of the image by using an automatically generated caption of the source image as a proxy for the edit prompt. This breaks the alignment between the contents of the image and their textual description, without requiring knowledge of the editing method or the editing prompt. Reflecting on the reliability of existing metrics for immunization success, we propose two novel evaluation strategies: Caption Similarity, which quantifies semantic consistency between original and adversarial edits, and semantic Intersection over Union (IoU), which measures spatial layout disruption via segmentation masks. Experiments conducted on the TEDBench++ benchmark demonstrate that our attack significantly degrades editing performance while remaining imperceptible.", "date": "2025-09-12", "categories": []}
{"id": "2504.05815", "title": "Parasite: A Steganography-based Backdoor Attack Framework for Diffusion Models", "abstract": "Recently, the diffusion model has gained significant attention as one of the most successful image generation models, which can generate high-quality images by iteratively sampling noise. However, recent studies have shown that diffusion models are vulnerable to backdoor attacks, allowing attackers to enter input data containing triggers to activate the backdoor and generate their desired output. Existing backdoor attack methods primarily focused on target noise-to-image and text-to-image tasks, with limited work on backdoor attacks in image-to-image tasks. Furthermore, traditional backdoor attacks often rely on a single, conspicuous trigger to generate a fixed target image, lacking concealability and flexibility. To address these limitations, we propose a novel backdoor attack method called \"Parasite\" for image-to-image tasks in diffusion models, which not only is the first to leverage steganography for triggers hiding, but also allows attackers to embed the target content as a backdoor trigger to achieve a more flexible attack. \"Parasite\" as a novel attack method effectively bypasses existing detection frameworks to execute backdoor attacks. In our experiments, \"Parasite\" achieved a 0 percent backdoor detection rate against the mainstream defense frameworks. In addition, in the ablation study, we discuss the influence of different hiding coefficients on the attack results. You can find our code at https://anonymous.4open.science/r/Parasite-1715/.", "date": "2025-04-08", "categories": []}
{"id": "2505.16402", "title": "AdvReal: Physical Adversarial Patch Generation Framework for Security Evaluation of Object Detection Systems", "abstract": "Autonomous vehicles are typical complex intelligent systems with artificial intelligence at their core. However, perception methods based on deep learning are extremely vulnerable to adversarial samples, resulting in security accidents. How to generate effective adversarial examples in the physical world and evaluate object detection systems is a huge challenge. In this study, we propose a unified joint adversarial training framework for both 2D and 3D domains, which simultaneously optimizes texture maps in 2D image and 3D mesh spaces to better address intra-class diversity and real-world environmental variations. The framework includes a novel realistic enhanced adversarial module, with time-space and relighting mapping pipeline that adjusts illumination consistency between adversarial patches and target garments under varied viewpoints. Building upon this, we develop a realism enhancement mechanism that incorporates non-rigid deformation modeling and texture remapping to ensure alignment with the human body's non-rigid surfaces in 3D scenes. Extensive experiment results in digital and physical environments demonstrate that the adversarial textures generated by our method can effectively mislead the target detection model. Specifically, our method achieves an average attack success rate (ASR) of 70.13% on YOLOv12 in physical scenarios, significantly outperforming existing methods such as T-SEA (21.65%) and AdvTexture (19.70%). Moreover, the proposed method maintains stable ASR across multiple viewpoints and distances, with an average attack success rate exceeding 90% under both frontal and oblique views at a distance of 4 meters. This confirms the method's strong robustness and transferability under multi-angle attacks, varying lighting conditions, and real-world distances. The demo video and code can be obtained at https://github.com/Huangyh98/AdvReal.git.", "date": "2025-05-22", "categories": []}
{"id": "2505.19613", "title": "TESSER: Transfer-Enhancing Adversarial Attacks from Vision Transformers via Spectral and Semantic Regularization", "abstract": "Adversarial transferability remains a critical challenge in evaluating the robustness of deep neural networks. In security-critical applications, transferability enables black-box attacks without access to model internals, making it a key concern for real-world adversarial threat assessment. While Vision Transformers (ViTs) have demonstrated strong adversarial performance, existing attacks often fail to transfer effectively across architectures, especially from ViTs to Convolutional Neural Networks (CNNs) or hybrid models. In this paper, we introduce \\textbf{TESSER} -- a novel adversarial attack framework that enhances transferability via two key strategies: (1) \\textit{Feature-Sensitive Gradient Scaling (FSGS)}, which modulates gradients based on token-wise importance derived from intermediate feature activations, and (2) \\textit{Spectral Smoothness Regularization (SSR)}, which suppresses high-frequency noise in perturbations using a differentiable Gaussian prior. These components work in tandem to generate perturbations that are both semantically meaningful and spectrally smooth. Extensive experiments on ImageNet across 12 diverse architectures demonstrate that TESSER achieves +10.9\\% higher attack succes rate (ASR) on CNNs and +7.2\\% on ViTs compared to the state-of-the-art Adaptive Token Tuning (ATT) method. Moreover, TESSER significantly improves robustness against defended models, achieving 53.55\\% ASR on adversarially trained CNNs. Qualitative analysis shows strong alignment between TESSER's perturbations and salient visual regions identified via Grad-CAM, while frequency-domain analysis reveals a 12\\% reduction in high-frequency energy, confirming the effectiveness of spectral regularization.", "date": "2025-05-26", "categories": []}
{"id": "2401.10590", "title": "Adversarial Robustness of Link Sign Prediction in Signed Graphs", "abstract": "Signed graphs serve as fundamental data structures for representing positive and negative relationships in social networks, with signed graph neural networks (SGNNs) emerging as the primary tool for their analysis. Our investigation reveals that balance theory, while essential for modeling signed relationships in SGNNs, inadvertently introduces exploitable vulnerabilities to black-box attacks. To showcase this, we propose balance-attack, a novel adversarial strategy specifically designed to compromise graph balance degree, and develop an efficient heuristic algorithm to solve the associated NP-hard optimization problem. While existing approaches attempt to restore attacked graphs through balance learning techniques, they face a critical challenge we term \"Irreversibility of Balance-related Information,\" as restored edges fail to align with original attack targets. To address this limitation, we introduce Balance Augmented-Signed Graph Contrastive Learning (BA-SGCL), an innovative framework that combines contrastive learning with balance augmentation techniques to achieve robust graph representations. By maintaining high balance degree in the latent space, BA-SGCL not only effectively circumvents the irreversibility challenge but also significantly enhances model resilience. Extensive experiments across multiple SGNN architectures and real-world datasets demonstrate both the effectiveness of our proposed balance-attack and the superior robustness of BA-SGCL, advancing the security and reliability of signed graph analysis in social networks. Datasets and codes of the proposed framework are at the github repository https://anonymous.4open.science/r/BA-SGCL-submit-DF41/.", "date": "2024-01-19", "categories": []}
{"id": "2503.14111", "title": "Towards properties of adversarial image perturbations", "abstract": "Using stochastic gradient approach we study the properties of adversarial perturbations resulting in noticeable growth of VMAF image quality metric. The structure of the perturbations is investigated depending on the acceptable PSNR values and based on the Fourier power spectrum computations for the perturbations. It is demonstrated that moderate variation of image brightness ($\\sim 10$ pixel units in a restricted region of an image can result in VMAF growth by $\\sim 60\\%$). Unlike some other methods demonstrating similar VMAF growth, the subjective quality of an image remains almost unchanged. It is also shown that the adversarial perturbations may demonstrate approximately linear dependence of perturbation amplitudes on the image brightness. The perturbations are studied based on the direct VMAF optimization in PyTorch. The significant discrepancies between the metric values and subjective judgements are also demonstrated when image restoration from noise is carried out using the same direct VMAF optimization.", "date": "2025-03-18", "categories": []}
{"id": "2509.08463", "title": "Adversarial Attacks Against Automated Fact-Checking: A Survey", "abstract": "In an era where misinformation spreads freely, fact-checking (FC) plays a crucial role in verifying claims and promoting reliable information. While automated fact-checking (AFC) has advanced significantly, existing systems remain vulnerable to adversarial attacks that manipulate or generate claims, evidence, or claim-evidence pairs. These attacks can distort the truth, mislead decision-makers, and ultimately undermine the reliability of FC models. Despite growing research interest in adversarial attacks against AFC systems, a comprehensive, holistic overview of key challenges remains lacking. These challenges include understanding attack strategies, assessing the resilience of current models, and identifying ways to enhance robustness. This survey provides the first in-depth review of adversarial attacks targeting FC, categorizing existing attack methodologies and evaluating their impact on AFC systems. Additionally, we examine recent advancements in adversary-aware defenses and highlight open research questions that require further exploration. Our findings underscore the urgent need for resilient FC frameworks capable of withstanding adversarial manipulations in pursuit of preserving high verification accuracy.", "date": "2025-09-10", "categories": []}
{"id": "2311.11191", "title": "Attention-Based Real-Time Defenses for Physical Adversarial Attacks in Vision Applications", "abstract": "Deep neural networks exhibit excellent performance in computer vision tasks, but their vulnerability to real-world adversarial attacks, achieved through physical objects that can corrupt their predictions, raises serious security concerns for their application in safety-critical domains. Existing defense methods focus on single-frame analysis and are characterized by high computational costs that limit their applicability in multi-frame scenarios, where real-time decisions are crucial.\n  To address this problem, this paper proposes an efficient attention-based defense mechanism that exploits adversarial channel-attention to quickly identify and track malicious objects in shallow network layers and mask their adversarial effects in a multi-frame setting. This work advances the state of the art by enhancing existing over-activation techniques for real-world adversarial attacks to make them usable in real-time applications. It also introduces an efficient multi-frame defense framework, validating its efficacy through extensive experiments aimed at evaluating both defense performance and computational cost.", "date": "2023-11-19", "categories": []}
{"id": "2403.05666", "title": "Prepared for the Worst: A Learning-Based Adversarial Attack for Resilience Analysis of the ICP Algorithm", "abstract": "This paper presents a novel method for assessing the resilience of the ICP algorithm via learning-based, worst-case attacks on lidar point clouds. For safety-critical applications such as autonomous navigation, ensuring the resilience of algorithms before deployments is crucial. The ICP algorithm is the standard for lidar-based localization, but its accuracy can be greatly affected by corrupted measurements from various sources, including occlusions, adverse weather, or mechanical sensor issues. Unfortunately, the complex and iterative nature of ICP makes assessing its resilience to corruption challenging. While there have been efforts to create challenging datasets and develop simulations to evaluate the resilience of ICP, our method focuses on finding the maximum possible ICP error that can arise from corrupted measurements at a location. We demonstrate that our perturbation-based adversarial attacks can be used pre-deployment to identify locations on a map where ICP is particularly vulnerable to corruptions in the measurements. With such information, autonomous robots can take safer paths when deployed, to mitigate against their measurements being corrupted. The proposed attack outperforms baselines more than 88% of the time across a wide range of scenarios.", "date": "2024-03-08", "categories": []}
{"id": "2412.16358", "title": "Texture- and Shape-based Adversarial Attacks for Overhead Image Vehicle Detection", "abstract": "Detecting vehicles in aerial images is difficult due to complex backgrounds, small object sizes, shadows, and occlusions. Although recent deep learning advancements have improved object detection, these models remain susceptible to adversarial attacks (AAs), challenging their reliability. Traditional AA strategies often ignore practical implementation constraints. Our work proposes realistic and practical constraints on texture (lowering resolution, limiting modified areas, and color ranges) and analyzes the impact of shape modifications on attack performance. We conducted extensive experiments with three object detector architectures, demonstrating the performance-practicality trade-off: more practical modifications tend to be less effective, and vice versa. We release both code and data to support reproducibility at https://github.com/humansensinglab/texture-shape-adversarial-attacks.", "date": "2024-12-20", "categories": []}
{"id": "2509.07132", "title": "Adversarial Attacks on Audio Deepfake Detection: A Benchmark and Comparative Study", "abstract": "The widespread use of generative AI has shown remarkable success in producing highly realistic deepfakes, posing a serious threat to various voice biometric applications, including speaker verification, voice biometrics, audio conferencing, and criminal investigations. To counteract this, several state-of-the-art (SoTA) audio deepfake detection (ADD) methods have been proposed to identify generative AI signatures to distinguish between real and deepfake audio. However, the effectiveness of these methods is severely undermined by anti-forensic (AF) attacks that conceal generative signatures. These AF attacks span a wide range of techniques, including statistical modifications (e.g., pitch shifting, filtering, noise addition, and quantization) and optimization-based attacks (e.g., FGSM, PGD, C \\& W, and DeepFool). In this paper, we investigate the SoTA ADD methods and provide a comparative analysis to highlight their effectiveness in exposing deepfake signatures, as well as their vulnerabilities under adversarial conditions. We conducted an extensive evaluation of ADD methods on five deepfake benchmark datasets using two categories: raw and spectrogram-based approaches. This comparative analysis enables a deeper understanding of the strengths and limitations of SoTA ADD methods against diverse AF attacks. It does not only highlight vulnerabilities of ADD methods, but also informs the design of more robust and generalized detectors for real-world voice biometrics. It will further guide future research in developing adaptive defense strategies that can effectively counter evolving AF techniques.", "date": "2025-09-08", "categories": []}
{"id": "2509.07495", "title": "Generating Transferrable Adversarial Examples via Local Mixing and Logits Optimization for Remote Sensing Object Recognition", "abstract": "Deep Neural Networks (DNNs) are vulnerable to adversarial attacks, posing significant security threats to their deployment in remote sensing applications. Research on adversarial attacks not only reveals model vulnerabilities but also provides critical insights for enhancing robustness. Although current mixing-based strategies have been proposed to increase the transferability of adversarial examples, they either perform global blending or directly exchange a region in the images, which may destroy global semantic features and mislead the optimization of adversarial examples. Furthermore, their reliance on cross-entropy loss for perturbation optimization leads to gradient diminishing during iterative updates, compromising adversarial example quality. To address these limitations, we focus on non-targeted attacks and propose a novel framework via local mixing and logits optimization. First, we present a local mixing strategy to generate diverse yet semantically consistent inputs. Different from MixUp, which globally blends two images, and MixCut, which stitches images together, our method merely blends local regions to preserve global semantic information. Second, we adapt the logit loss from targeted attacks to non-targeted scenarios, mitigating the gradient vanishing problem of cross-entropy loss. Third, a perturbation smoothing loss is applied to suppress high-frequency noise and enhance transferability. Extensive experiments on FGSCR-42 and MTARSI datasets demonstrate superior performance over 12 state-of-the-art methods across 6 surrogate models. Notably, with ResNet as the surrogate on MTARSI, our method achieves a 17.28% average improvement in black-box attack success rate.", "date": "2025-09-09", "categories": []}
{"id": "2509.07617", "title": "Transferable Direct Prompt Injection via Activation-Guided MCMC Sampling", "abstract": "Direct Prompt Injection (DPI) attacks pose a critical security threat to Large Language Models (LLMs) due to their low barrier of execution and high potential damage. To address the impracticality of existing white-box/gray-box methods and the poor transferability of black-box methods, we propose an activations-guided prompt injection attack framework. We first construct an Energy-based Model (EBM) using activations from a surrogate model to evaluate the quality of adversarial prompts. Guided by the trained EBM, we employ the token-level Markov Chain Monte Carlo (MCMC) sampling to adaptively optimize adversarial prompts, thereby enabling gradient-free black-box attacks. Experimental results demonstrate our superior cross-model transferability, achieving 49.6% attack success rate (ASR) across five mainstream LLMs and 34.6% improvement over human-crafted prompts, and maintaining 36.6% ASR on unseen task scenarios. Interpretability analysis reveals a correlation between activations and attack effectiveness, highlighting the critical role of semantic patterns in transferable vulnerability exploitation.", "date": "2025-09-09", "categories": []}
{"id": "2508.09994", "title": "Whisper Smarter, not Harder: Adversarial Attack on Partial Suppression", "abstract": "Currently, Automatic Speech Recognition (ASR) models are deployed in an extensive range of applications. However, recent studies have demonstrated the possibility of adversarial attack on these models which could potentially suppress or disrupt model output. We investigate and verify the robustness of these attacks and explore if it is possible to increase their imperceptibility. We additionally find that by relaxing the optimisation objective from complete suppression to partial suppression, we can further decrease the imperceptibility of the attack. We also explore possible defences against these attacks and show a low-pass filter defence could potentially serve as an effective defence.", "date": "2025-07-30", "categories": []}
{"id": "2508.11854", "title": "ComplicitSplat: Downstream Models are Vulnerable to Blackbox Attacks by 3D Gaussian Splat Camouflages", "abstract": "As 3D Gaussian Splatting (3DGS) gains rapid adoption in safety-critical tasks for efficient novel-view synthesis from static images, how might an adversary tamper images to cause harm? We introduce ComplicitSplat, the first attack that exploits standard 3DGS shading methods to create viewpoint-specific camouflage - colors and textures that change with viewing angle - to embed adversarial content in scene objects that are visible only from specific viewpoints and without requiring access to model architecture or weights. Our extensive experiments show that ComplicitSplat generalizes to successfully attack a variety of popular detector - both single-stage, multi-stage, and transformer-based models on both real-world capture of physical objects and synthetic scenes. To our knowledge, this is the first black-box attack on downstream object detectors using 3DGS, exposing a novel safety risk for applications like autonomous navigation and other mission-critical robotic systems.", "date": "2025-08-16", "categories": []}
{"id": "2509.06338", "title": "Embedding Poisoning: Bypassing Safety Alignment via Embedding Semantic Shift", "abstract": "The widespread distribution of Large Language Models (LLMs) through public platforms like Hugging Face introduces significant security challenges. While these platforms perform basic security scans, they often fail to detect subtle manipulations within the embedding layer. This work identifies a novel class of deployment phase attacks that exploit this vulnerability by injecting imperceptible perturbations directly into the embedding layer outputs without modifying model weights or input text. These perturbations, though statistically benign, systematically bypass safety alignment mechanisms and induce harmful behaviors during inference. We propose Search based Embedding Poisoning(SEP), a practical, model agnostic framework that introduces carefully optimized perturbations into embeddings associated with high risk tokens. SEP leverages a predictable linear transition in model responses, from refusal to harmful output to semantic deviation to identify a narrow perturbation window that evades alignment safeguards. Evaluated across six aligned LLMs, SEP achieves an average attack success rate of 96.43% while preserving benign task performance and evading conventional detection mechanisms. Our findings reveal a critical oversight in deployment security and emphasize the urgent need for embedding level integrity checks in future LLM defense strategies.", "date": "2025-09-08", "categories": []}
{"id": "2509.06459", "title": "IGAff: Benchmarking Adversarial Iterative and Genetic Affine Algorithms on Deep Neural Networks", "abstract": "Deep neural networks currently dominate many fields of the artificial intelligence landscape, achieving state-of-the-art results on numerous tasks while remaining hard to understand and exhibiting surprising weaknesses. An active area of research focuses on adversarial attacks, which aim to generate inputs that uncover these weaknesses. However, this proves challenging, especially in the black-box scenario where model details are inaccessible. This paper explores in detail the impact of such adversarial algorithms on ResNet-18, DenseNet-121, Swin Transformer V2, and Vision Transformer network architectures. Leveraging the Tiny ImageNet, Caltech-256, and Food-101 datasets, we benchmark two novel black-box iterative adversarial algorithms based on affine transformations and genetic algorithms: 1) Affine Transformation Attack (ATA), an iterative algorithm maximizing our attack score function using random affine transformations, and 2) Affine Genetic Attack (AGA), a genetic algorithm that involves random noise and affine transformations. We evaluate the performance of the models in the algorithm parameter variation, data augmentation, and global and targeted attack configurations. We also compare our algorithms with two black-box adversarial algorithms, Pixle and Square Attack. Our experiments yield better results on the image classification task than similar methods in the literature, achieving an accuracy improvement of up to 8.82%. We provide noteworthy insights into successful adversarial defenses and attacks at both global and targeted levels, and demonstrate adversarial robustness through algorithm parameter variation.", "date": "2025-09-08", "categories": []}
{"id": "2509.06835", "title": "Evaluating the Impact of Adversarial Attacks on Traffic Sign Classification using the LISA Dataset", "abstract": "Adversarial attacks pose significant threats to machine learning models by introducing carefully crafted perturbations that cause misclassification. While prior work has primarily focused on MNIST and similar datasets, this paper investigates the vulnerability of traffic sign classifiers using the LISA Traffic Sign dataset. We train a convolutional neural network to classify 47 different traffic signs and evaluate its robustness against Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks. Our results show a sharp decline in classification accuracy as the perturbation magnitude increases, highlighting the models susceptibility to adversarial examples. This study lays the groundwork for future exploration into defense mechanisms tailored for real-world traffic sign recognition systems.", "date": "2025-09-08", "categories": []}
{"id": "2312.16019", "title": "Survival Analysis with Adversarial Regularization", "abstract": "Survival Analysis (SA) models the time until an event occurs, with applications in fields like medicine, defense, finance, and aerospace. Recent research indicates that Neural Networks (NNs) can effectively capture complex data patterns in SA, whereas simple generalized linear models often fall short in this regard. However, dataset uncertainties (e.g., noisy measurements, human error) can degrade NN model performance. To address this, we leverage advances in NN verification to develop training objectives for robust, fully-parametric SA models. Specifically, we propose an adversarially robust loss function based on a Min-Max optimization problem. We employ CROWN-Interval Bound Propagation (CROWN-IBP) to tackle the computational challenges inherent in solving this Min-Max problem. Evaluated over 10 SurvSet datasets, our method, Survival Analysis with Adversarial Regularization (SAWAR), consistently outperforms baseline adversarial training methods and state-of-the-art (SOTA) deep SA models across various covariate perturbations with respect to Negative Log Likelihood (NegLL), Integrated Brier Score (IBS), and Concordance Index (CI) metrics. Thus, we demonstrate that adversarial robustness enhances SA predictive performance and calibration, mitigating data uncertainty and improving generalization across diverse datasets by up to 150% compared to baselines.", "date": "2023-12-26", "categories": []}
{"id": "2509.04914", "title": "RobQFL: Robust Quantum Federated Learning in Adversarial Environment", "abstract": "Quantum Federated Learning (QFL) merges privacy-preserving federation with quantum computing gains, yet its resilience to adversarial noise is unknown. We first show that QFL is as fragile as centralized quantum learning. We propose Robust Quantum Federated Learning (RobQFL), embedding adversarial training directly into the federated loop. RobQFL exposes tunable axes: client coverage $\\gamma$ (0-100\\%), perturbation scheduling (fixed-$\\varepsilon$ vs $\\varepsilon$-mixes), and optimization (fine-tune vs scratch), and distils the resulting $\\gamma \\times \\varepsilon$ surface into two metrics: Accuracy-Robustness Area and Robustness Volume. On 15-client simulations with MNIST and Fashion-MNIST, IID and Non-IID conditions, training only 20-50\\% clients adversarially boosts $\\varepsilon \\leq 0.1$ accuracy $\\sim$15 pp at $< 2$ pp clean-accuracy cost; fine-tuning adds 3-5 pp. With $\\geq$75\\% coverage, a moderate $\\varepsilon$-mix is optimal, while high-$\\varepsilon$ schedules help only at 100\\% coverage. Label-sorted non-IID splits halve robustness, underscoring data heterogeneity as a dominant risk.", "date": "2025-09-05", "categories": []}
{"id": "2509.04980", "title": "MAIA: An Inpainting-Based Approach for Music Adversarial Attacks", "abstract": "Music adversarial attacks have garnered significant interest in the field of Music Information Retrieval (MIR). In this paper, we present Music Adversarial Inpainting Attack (MAIA), a novel adversarial attack framework that supports both white-box and black-box attack scenarios. MAIA begins with an importance analysis to identify critical audio segments, which are then targeted for modification. Utilizing generative inpainting models, these segments are reconstructed with guidance from the output of the attacked model, ensuring subtle and effective adversarial perturbations. We evaluate MAIA on multiple MIR tasks, demonstrating high attack success rates in both white-box and black-box settings while maintaining minimal perceptual distortion. Additionally, subjective listening tests confirm the high audio fidelity of the adversarial samples. Our findings highlight vulnerabilities in current MIR systems and emphasize the need for more robust and secure models.", "date": "2025-09-05", "categories": []}
{"id": "2509.05086", "title": "Robust Experts: the Effect of Adversarial Training on CNNs with Sparse Mixture-of-Experts Layers", "abstract": "Robustifying convolutional neural networks (CNNs) against adversarial attacks remains challenging and often requires resource-intensive countermeasures. We explore the use of sparse mixture-of-experts (MoE) layers to improve robustness by replacing selected residual blocks or convolutional layers, thereby increasing model capacity without additional inference cost. On ResNet architectures trained on CIFAR-100, we find that inserting a single MoE layer in the deeper stages leads to consistent improvements in robustness under PGD and AutoPGD attacks when combined with adversarial training. Furthermore, we discover that when switch loss is used for balancing, it causes routing to collapse onto a small set of overused experts, thereby concentrating adversarial training on these paths and inadvertently making them more robust. As a result, some individual experts outperform the gated MoE model in robustness, suggesting that robust subpaths emerge through specialization. Our code is available at https://github.com/KASTEL-MobilityLab/robust-sparse-moes.", "date": "2025-09-05", "categories": []}
{"id": "2401.09754", "title": "Towards Robust Graph Structural Learning Beyond Homophily via Preserving Neighbor Similarity", "abstract": "Despite the tremendous success of graph-based learning systems in handling structural data, it has been widely investigated that they are fragile to adversarial attacks on homophilic graph data, where adversaries maliciously modify the semantic and topology information of the raw graph data to degrade the predictive performances. Motivated by this, a series of robust models are crafted to enhance the adversarial robustness of graph-based learning systems on homophilic graphs. However, the security of graph-based learning systems on heterophilic graphs remains a mystery to us. To bridge this gap, in this paper, we start to explore the vulnerability of graph-based learning systems regardless of the homophily degree, and theoretically prove that the update of the negative classification loss is negatively correlated with the pairwise similarities based on the powered aggregated neighbor features. The theoretical finding inspires us to craft a novel robust graph structural learning strategy that serves as a useful graph mining module in a robust model that incorporates a dual-kNN graph constructions pipeline to supervise the neighbor-similarity-preserved propagation, where the graph convolutional layer adaptively smooths or discriminates the features of node pairs according to their affluent local structures. In this way, the proposed methods can mine the ``better\" topology of the raw graph data under diverse graph homophily and achieve more reliable data management on homophilic and heterophilic graphs.", "date": "2024-01-18", "categories": []}
{"id": "2412.12722", "title": "Defending LVLMs Against Vision Attacks through Partial-Perception Supervision", "abstract": "Recent studies have raised significant concerns regarding the vulnerability of Large Vision Language Models (LVLMs) to maliciously injected or perturbed input images, which can mislead their responses. Existing defense methods show that such vision attacks are sensitive to image modifications especially cropping, using majority voting across responses of modified images as corrected responses. However, these modifications often result in partial images and distort the semantics, which reduces response quality on clean images after voting. Instead of directly using responses from partial images for voting, we investigate using them to supervise the LVLM's responses to the original images. We propose a black-box, training-free method called DPS (Defense through Partial-Perception Supervision). In this approach, the model is prompted using the responses generated by a model that perceives only a partial image. With DPS, the model can adjust its response based on partial image understanding when under attack, while confidently maintaining its original response for clean input. Our findings show that the weak model can supervise the strong model: when faced with an attacked input, the strong model becomes less confident and adjusts its response based on the weak model's partial understanding, effectively defending against the attack. With clean input, it confidently maintains its original response. Empirical experiments show our method outperforms the baseline, cutting the average attack success rate by 76.3% across six datasets on three popular models.", "date": "2024-12-17", "categories": []}
{"id": "2509.01235", "title": "Geometric origin of adversarial vulnerability in deep learning", "abstract": "How to balance training accuracy and adversarial robustness has become a challenge since the birth of deep learning. Here, we introduce a geometry-aware deep learning framework that leverages layer-wise local training to sculpt the internal representations of deep neural networks. This framework promotes intra-class compactness and inter-class separation in feature space, leading to manifold smoothness and adversarial robustness against white or black box attacks. The performance can be explained by an energy model with Hebbian coupling between elements of the hidden representation. Our results thus shed light on the physics of learning in the direction of alignment between biological and artificial intelligence systems. Using the current framework, the deep network can assimilate new information into existing knowledge structures while reducing representation interference.", "date": "2025-09-01", "categories": []}
{"id": "2410.20940", "title": "Attacking Misinformation Detection Using Adversarial Examples Generated by Language Models", "abstract": "Large language models have many beneficial applications, but can they also be used to attack content-filtering algorithms in social media platforms? We investigate the challenge of generating adversarial examples to test the robustness of text classification algorithms detecting low-credibility content, including propaganda, false claims, rumours and hyperpartisan news. We focus on simulation of content moderation by setting realistic limits on the number of queries an attacker is allowed to attempt. Within our solution (TREPAT), initial rephrasings are generated by large language models with prompts inspired by meaning-preserving NLP tasks, such as text simplification and style transfer. Subsequently, these modifications are decomposed into small changes, applied through beam search procedure, until the victim classifier changes its decision. We perform (1) quantitative evaluation using various prompts, models and query limits, (2) targeted manual assessment of the generated text and (3) qualitative linguistic analysis. The results confirm the superiority of our approach in the constrained scenario, especially in case of long input text (news articles), where exhaustive search is not feasible.", "date": "2024-10-28", "categories": []}
{"id": "2509.03179", "title": "AutoDetect: Designing an Autoencoder-based Detection Method for Poisoning Attacks on Object Detection Applications in the Military Domain", "abstract": "Poisoning attacks pose an increasing threat to the security and robustness of Artificial Intelligence systems in the military domain. The widespread use of open-source datasets and pretrained models exacerbates this risk. Despite the severity of this threat, there is limited research on the application and detection of poisoning attacks on object detection systems. This is especially problematic in the military domain, where attacks can have grave consequences. In this work, we both investigate the effect of poisoning attacks on military object detectors in practice, and the best approach to detect these attacks. To support this research, we create a small, custom dataset featuring military vehicles: MilCivVeh. We explore the vulnerability of military object detectors for poisoning attacks by implementing a modified version of the BadDet attack: a patch-based poisoning attack. We then assess its impact, finding that while a positive attack success rate is achievable, it requires a substantial portion of the data to be poisoned -- raising questions about its practical applicability. To address the detection challenge, we test both specialized poisoning detection methods and anomaly detection methods from the visual industrial inspection domain. Since our research shows that both classes of methods are lacking, we introduce our own patch detection method: AutoDetect, a simple, fast, and lightweight autoencoder-based method. Our method shows promising results in separating clean from poisoned samples using the reconstruction error of image slices, outperforming existing methods, while being less time- and memory-intensive. We urge that the availability of large, representative datasets in the military domain is a prerequisite to further evaluate risks of poisoning attacks and opportunities patch detection.", "date": "2025-09-03", "categories": []}
{"id": "2509.03383", "title": "ANNIE: Be Careful of Your Robots", "abstract": "The integration of vision-language-action (VLA) models into embodied AI (EAI) robots is rapidly advancing their ability to perform complex, long-horizon tasks in humancentric environments. However, EAI systems introduce critical security risks: a compromised VLA model can directly translate adversarial perturbations on sensory input into unsafe physical actions. Traditional safety definitions and methodologies from the machine learning community are no longer sufficient. EAI systems raise new questions, such as what constitutes safety, how to measure it, and how to design effective attack and defense mechanisms in physically grounded, interactive settings. In this work, we present the first systematic study of adversarial safety attacks on embodied AI systems, grounded in ISO standards for human-robot interactions. We (1) formalize a principled taxonomy of safety violations (critical, dangerous, risky) based on physical constraints such as separation distance, velocity, and collision boundaries; (2) introduce ANNIEBench, a benchmark of nine safety-critical scenarios with 2,400 video-action sequences for evaluating embodied safety; and (3) ANNIE-Attack, a task-aware adversarial framework with an attack leader model that decomposes long-horizon goals into frame-level perturbations. Our evaluation across representative EAI models shows attack success rates exceeding 50% across all safety categories. We further demonstrate sparse and adaptive attack strategies and validate the real-world impact through physical robot experiments. These results expose a previously underexplored but highly consequential attack surface in embodied AI systems, highlighting the urgent need for security-driven defenses in the physical AI era. Code is available at https://github.com/RLCLab/Annie.", "date": "2025-09-03", "categories": []}
{"id": "2402.04660", "title": "Redesigning Traffic Signs to Mitigate Machine-Learning Patch Attacks", "abstract": "Traffic-Sign Recognition (TSR) is a critical safety component for autonomous driving. Unfortunately, however, past work has highlighted the vulnerability of TSR models to physical-world attacks, through low-cost, easily deployable adversarial patches leading to misclassification. To mitigate these threats, most defenses focus on altering the training process or modifying the inference procedure. Still, while these approaches improve adversarial robustness, TSR remains susceptible to attacks attaining substantial success rates.\n  To further the adversarial robustness of TSR, this work offers a novel approach that redefines traffic-sign designs to create signs that promote robustness while remaining interpretable to humans. Our framework takes three inputs: (1) A traffic-sign standard along with modifiable features and associated constraints; (2) A state-of-the-art adversarial training method; and (3) A function for efficiently synthesizing realistic traffic-sign images. Using these user-defined inputs, the framework emits an optimized traffic-sign standard such that traffic signs generated per this standard enable training TSR models with increased adversarial robustness.\n  We evaluate the effectiveness of our framework via a concrete implementation, where we allow modifying the pictograms (i.e., symbols) and colors of traffic signs. The results show substantial improvements in robustness -- with gains of up to 16.33%--24.58% in robust accuracy over state-of-the-art methods -- while benign accuracy is even improved. Importantly, a user study also confirms that the redesigned traffic signs remain easily recognizable and to human observers. Overall, the results highlight that carefully redesigning traffic signs can significantly enhance TSR system robustness without compromising human interpretability.", "date": "2024-02-07", "categories": []}
{"id": "2408.14875", "title": "Adversarial Attacks and Defenses in Multivariate Time-Series Forecasting for Smart and Connected Infrastructures", "abstract": "The emergence of deep learning models has revolutionized various industries over the last decade, leading to a surge in connected devices and infrastructures. However, these models can be tricked into making incorrect predictions with high confidence, leading to disastrous failures and security concerns. To this end, we explore the impact of adversarial attacks on multivariate time-series forecasting and investigate methods to counter them. Specifically, we employ untargeted white-box attacks, namely the Fast Gradient Sign Method (FGSM) and the Basic Iterative Method (BIM), to poison the inputs to the training process, effectively misleading the model. We also illustrate the subtle modifications to the inputs after the attack, which makes detecting the attack using the naked eye quite difficult. Having demonstrated the feasibility of these attacks, we develop robust models through adversarial training and model hardening. We are among the first to showcase the transferability of these attacks and defenses by extrapolating our work from the benchmark electricity data to a larger, 10-year real-world data used for predicting the time-to-failure of hard disks. Our experimental results confirm that the attacks and defenses achieve the desired security thresholds, leading to a 72.41% and 94.81% decrease in RMSE for the electricity and hard disk datasets respectively after implementing the adversarial defenses.", "date": "2024-08-27", "categories": []}
{"id": "2412.05892", "title": "PBI-Attack: Prior-Guided Bimodal Interactive Black-Box Jailbreak Attack for Toxicity Maximization", "abstract": "Understanding the vulnerabilities of Large Vision Language Models (LVLMs) to jailbreak attacks is essential for their responsible real-world deployment. Most previous work requires access to model gradients, or is based on human knowledge (prompt engineering) to complete jailbreak, and they hardly consider the interaction of images and text, resulting in inability to jailbreak in black box scenarios or poor performance. To overcome these limitations, we propose a Prior-Guided Bimodal Interactive Black-Box Jailbreak Attack for toxicity maximization, referred to as PBI-Attack. Our method begins by extracting malicious features from a harmful corpus using an alternative LVLM and embedding these features into a benign image as prior information. Subsequently, we enhance these features through bidirectional cross-modal interaction optimization, which iteratively optimizes the bimodal perturbations in an alternating manner through greedy search, aiming to maximize the toxicity of the generated response. The toxicity level is quantified using a well-trained evaluation model. Experiments demonstrate that PBI-Attack outperforms previous state-of-the-art jailbreak methods, achieving an average attack success rate of 92.5% across three open-source LVLMs and around 67.3% on three closed-source LVLMs. Disclaimer: This paper contains potentially disturbing and offensive content.", "date": "2024-12-08", "categories": []}
{"id": "2412.07326", "title": "Addressing Key Challenges of Adversarial Attacks and Defenses in the Tabular Domain: A Methodological Framework for Coherence and Consistency", "abstract": "Machine learning models trained on tabular data are vulnerable to adversarial attacks, even in realistic scenarios where attackers only have access to the model's outputs. Since tabular data contains complex interdependencies among features, it presents a unique challenge for adversarial samples which must maintain coherence and respect these interdependencies to remain indistinguishable from benign data. Moreover, existing attack evaluation metrics-such as the success rate, perturbation magnitude, and query count-fail to account for this challenge. To address those gaps, we propose a technique for perturbing dependent features while preserving sample coherence. In addition, we introduce Class-Specific Anomaly Detection (CSAD), an effective novel anomaly detection approach, along with concrete metrics for assessing the quality of tabular adversarial attacks. CSAD evaluates adversarial samples relative to their predicted class distribution, rather than a broad benign distribution. It ensures that subtle adversarial perturbations, which may appear coherent in other classes, are correctly identified as anomalies. We integrate SHAP explainability techniques to detect inconsistencies in model decision-making, extending CSAD for SHAP-based anomaly detection. Our evaluation incorporates both anomaly detection rates with SHAP-based assessments to provide a more comprehensive measure of adversarial sample quality. We evaluate various attack strategies, examining black-box query-based and transferability-based gradient attacks across four target models. Experiments on benchmark tabular datasets reveal key differences in the attacker's risk and effort and attack quality, offering insights into the strengths, limitations, and trade-offs faced by attackers and defenders. Our findings lay the groundwork for future research on adversarial attacks and defense development in the tabular domain.", "date": "2024-12-10", "categories": []}
{"id": "2503.20187", "title": "Network Inversion for Generating Confidently Classified Counterfeits", "abstract": "In vision classification, generating inputs that elicit confident predictions is key to understanding model behavior and reliability, especially under adversarial or out-of-distribution (OOD) conditions. While traditional adversarial methods rely on perturbing existing inputs to fool a model, they are inherently input-dependent and often fail to ensure both high confidence and meaningful deviation from the training data. In this work, we extend network inversion techniques to generate Confidently Classified Counterfeits (CCCs), synthetic samples that are confidently classified by the model despite being significantly different from the training distribution and independent of any specific input. We alter inversion technique by replacing soft vector conditioning with one-hot class conditioning and introducing a Kullback-Leibler divergence loss between the one-hot label and the classifier's output distribution. CCCs offer a model-centric perspective on confidence, revealing that models can assign high confidence to entirely synthetic, out-of-distribution inputs. This challenges the core assumption behind many OOD detection techniques based on thresholding prediction confidence, which assume that high-confidence outputs imply in-distribution data, and highlights the need for more robust uncertainty estimation in safety-critical applications.", "date": "2025-03-26", "categories": []}
{"id": "2505.03084", "title": "Adversarial Attacks in Multimodal Systems: A Practitioner's Survey", "abstract": "The introduction of multimodal models is a huge step forward in Artificial Intelligence. A single model is trained to understand multiple modalities: text, image, video, and audio. Open-source multimodal models have made these breakthroughs more accessible. However, considering the vast landscape of adversarial attacks across these modalities, these models also inherit vulnerabilities of all the modalities, and ultimately, the adversarial threat amplifies. While broad research is available on possible attacks within or across these modalities, a practitioner-focused view that outlines attack types remains absent in the multimodal world. As more Machine Learning Practitioners adopt, fine-tune, and deploy open-source models in real-world applications, it's crucial that they can view the threat landscape and take the preventive actions necessary. This paper addresses the gap by surveying adversarial attacks targeting all four modalities: text, image, video, and audio. This survey provides a view of the adversarial attack landscape and presents how multimodal adversarial threats have evolved. To the best of our knowledge, this survey is the first comprehensive summarization of the threat landscape in the multimodal world.", "date": "2025-05-06", "categories": []}
{"id": "2508.10946", "title": "IPG: Incremental Patch Generation for Generalized Adversarial Patch Training", "abstract": "The advent of adversarial patches poses a significant challenge to the robustness of AI models, particularly in the domain of computer vision tasks such as object detection. In contradistinction to traditional adversarial examples, these patches target specific regions of an image, resulting in the malfunction of AI models. This paper proposes Incremental Patch Generation (IPG), a method that generates adversarial patches up to 11.1 times more efficiently than existing approaches while maintaining comparable attack performance. The efficacy of IPG is demonstrated by experiments and ablation studies including YOLO's feature distribution visualization and adversarial training results, which show that it produces well-generalized patches that effectively cover a broader range of model vulnerabilities. Furthermore, IPG-generated datasets can serve as a robust knowledge foundation for constructing a robust model, enabling structured representation, advanced reasoning, and proactive defenses in AI security ecosystems. The findings of this study suggest that IPG has considerable potential for future utilization not only in adversarial patch defense but also in real-world applications such as autonomous vehicles, security systems, and medical imaging, where AI models must remain resilient to adversarial attacks in dynamic and high-stakes environments.", "date": "2025-08-13", "categories": []}
{"id": "2509.00089", "title": "Learning from Peers: Collaborative Ensemble Adversarial Training", "abstract": "Ensemble Adversarial Training (EAT) attempts to enhance the robustness of models against adversarial attacks by leveraging multiple models. However, current EAT strategies tend to train the sub-models independently, ignoring the cooperative benefits between sub-models. Through detailed inspections of the process of EAT, we find that that samples with classification disparities between sub-models are close to the decision boundary of ensemble, exerting greater influence on the robustness of ensemble. To this end, we propose a novel yet efficient Collaborative Ensemble Adversarial Training (CEAT), to highlight the cooperative learning among sub-models in the ensemble. To be specific, samples with larger predictive disparities between the sub-models will receive greater attention during the adversarial training of the other sub-models. CEAT leverages the probability disparities to adaptively assign weights to different samples, by incorporating a calibrating distance regularization. Extensive experiments on widely-adopted datasets show that our proposed method achieves the state-of-the-art performance over competitive EAT methods. It is noteworthy that CEAT is model-agnostic, which can be seamlessly adapted into various ensemble methods with flexible applicability.", "date": "2025-08-27", "categories": []}
{"id": "2509.00387", "title": "Unifying Adversarial Perturbation for Graph Neural Networks", "abstract": "This paper studies the vulnerability of Graph Neural Networks (GNNs) to adversarial attacks on node features and graph structure. Various methods have implemented adversarial training to augment graph data, aiming to bolster the robustness and generalization of GNNs. These methods typically involve applying perturbations to the node feature, weights, or graph structure and subsequently minimizing the loss by learning more robust graph model parameters under the adversarial perturbations. Despite the effectiveness of adversarial training in enhancing GNNs' robustness and generalization abilities, its application has been largely confined to specific datasets and GNN types. In this paper, we propose a novel method, PerturbEmbedding, that integrates adversarial perturbation and training, enhancing GNNs' resilience to such attacks and improving their generalization ability. PerturbEmbedding performs perturbation operations directly on every hidden embedding of GNNs and provides a unified framework for most existing perturbation strategies/methods. We also offer a unified perspective on the forms of perturbations, namely random and adversarial perturbations. Through experiments on various datasets using different backbone models, we demonstrate that PerturbEmbedding significantly improves both the robustness and generalization abilities of GNNs, outperforming existing methods. The rejection of both random (non-targeted) and adversarial (targeted) perturbations further enhances the backbone model's performance.", "date": "2025-08-30", "categories": []}
{"id": "2509.00391", "title": "The Resurgence of GCG Adversarial Attacks on Large Language Models", "abstract": "Gradient-based adversarial prompting, such as the Greedy Coordinate Gradient (GCG) algorithm, has emerged as a powerful method for jailbreaking large language models (LLMs). In this paper, we present a systematic appraisal of GCG and its annealing-augmented variant, T-GCG, across open-source LLMs of varying scales. Using Qwen2.5-0.5B, LLaMA-3.2-1B, and GPT-OSS-20B, we evaluate attack effectiveness on both safety-oriented prompts (AdvBench) and reasoning-intensive coding prompts. Our study reveals three key findings: (1) attack success rates (ASR) decrease with model size, reflecting the increasing complexity and non-convexity of larger models' loss landscapes; (2) prefix-based heuristics substantially overestimate attack effectiveness compared to GPT-4o semantic judgments, which provide a stricter and more realistic evaluation; and (3) coding-related prompts are significantly more vulnerable than adversarial safety prompts, suggesting that reasoning itself can be exploited as an attack vector. In addition, preliminary results with T-GCG show that simulated annealing can diversify adversarial search and achieve competitive ASR under prefix evaluation, though its benefits under semantic judgment remain limited. Together, these findings highlight the scalability limits of GCG, expose overlooked vulnerabilities in reasoning tasks, and motivate further development of annealing-inspired strategies for more robust adversarial evaluation.", "date": "2025-08-30", "categories": []}
{"id": "2509.00826", "title": "Sequential Difference Maximization: Generating Adversarial Examples via Multi-Stage Optimization", "abstract": "Efficient adversarial attack methods are critical for assessing the robustness of computer vision models. In this paper, we reconstruct the optimization objective for generating adversarial examples as \"maximizing the difference between the non-true labels' probability upper bound and the true label's probability,\" and propose a gradient-based attack method termed Sequential Difference Maximization (SDM). SDM establishes a three-layer optimization framework of \"cycle-stage-step.\" The processes between cycles and between iterative steps are respectively identical, while optimization stages differ in terms of loss functions: in the initial stage, the negative probability of the true label is used as the loss function to compress the solution space; in subsequent stages, we introduce the Directional Probability Difference Ratio (DPDR) loss function to gradually increase the non-true labels' probability upper bound by compressing the irrelevant labels' probabilities. Experiments demonstrate that compared with previous SOTA methods, SDM not only exhibits stronger attack performance but also achieves higher attack cost-effectiveness. Additionally, SDM can be combined with adversarial training methods to enhance their defensive effects. The code is available at https://github.com/X-L-Liu/SDM.", "date": "2025-08-31", "categories": []}
{"id": "2508.21472", "title": "Adversarial Patch Attack for Ship Detection via Localized Augmentation", "abstract": "Current ship detection techniques based on remote sensing imagery primarily rely on the object detection capabilities of deep neural networks (DNNs). However, DNNs are vulnerable to adversarial patch attacks, which can lead to misclassification by the detection model or complete evasion of the targets. Numerous studies have demonstrated that data transformation-based methods can improve the transferability of adversarial examples. However, excessive augmentation of image backgrounds or irrelevant regions may introduce unnecessary interference, resulting in false detections of the object detection model. These errors are not caused by the adversarial patches themselves but rather by the over-augmentation of background and non-target areas. This paper proposes a localized augmentation method that applies augmentation only to the target regions, avoiding any influence on non-target areas. By reducing background interference, this approach enables the loss function to focus more directly on the impact of the adversarial patch on the detection model, thereby improving the attack success rate. Experiments conducted on the HRSC2016 dataset demonstrate that the proposed method effectively increases the success rate of adversarial patch attacks and enhances their transferability.", "date": "2025-08-29", "categories": []}
{"id": "2508.21715", "title": "Entropy-Based Non-Invasive Reliability Monitoring of Convolutional Neural Networks", "abstract": "Convolutional Neural Networks (CNNs) have become the foundation of modern computer vision, achieving unprecedented accuracy across diverse image recognition tasks. While these networks excel on in-distribution data, they remain vulnerable to adversarial perturbations imperceptible input modifications that cause misclassification with high confidence. However, existing detection methods either require expensive retraining, modify network architecture, or degrade performance on clean inputs. Here we show that adversarial perturbations create immediate, detectable entropy signatures in CNN activations that can be monitored without any model modification. Using parallel entropy monitoring on VGG-16, we demonstrate that adversarial inputs consistently shift activation entropy by 7% in early convolutional layers, enabling 90% detection accuracy with false positives and false negative rates below 20%. The complete separation between clean and adversarial entropy distributions reveals that CNNs inherently encode distribution shifts in their activation patterns. This work establishes that CNN reliability can be assessed through activation entropy alone, enabling practical deployment of self-diagnostic vision systems that detect adversarial inputs in real-time without compromising original model performance.", "date": "2025-08-29", "categories": []}
{"id": "2402.14648", "title": "Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off", "abstract": "Adversarial training often suffers from a robustness-accuracy trade-off, where achieving high robustness comes at the cost of accuracy. One approach to mitigate this trade-off is leveraging invariance regularization, which encourages model invariance under adversarial perturbations; however, it still leads to accuracy loss. In this work, we closely analyze the challenges of using invariance regularization in adversarial training and understand how to address them. Our analysis identifies two key issues: (1) a ``gradient conflict\" between invariance and classification objectives, leading to suboptimal convergence, and (2) the mixture distribution problem arising from diverged distributions between clean and adversarial inputs. To address these issues, we propose Asymmetric Representation-regularized Adversarial Training (ARAT), which incorporates asymmetric invariance loss with stop-gradient operation and a predictor to avoid gradient conflict, and a split-BatchNorm (BN) structure to resolve the mixture distribution problem. Our detailed analysis demonstrates that each component effectively addresses the identified issues, offering novel insights into adversarial defense. ARAT shows superiority over existing methods across various settings. Finally, we discuss the implications of our findings to knowledge distillation-based defenses, providing a new perspective on their relative successes.", "date": "2024-02-22", "categories": []}
{"id": "2412.05943", "title": "Adversarial Transferability in Deep Denoising Models: Theoretical Insights and Robustness Enhancement via Out-of-Distribution Typical Set Sampling", "abstract": "Deep learning-based image denoising models demonstrate remarkable performance, but their lack of robustness analysis remains a significant concern. A major issue is that these models are susceptible to adversarial attacks, where small, carefully crafted perturbations to input data can cause them to fail. Surprisingly, perturbations specifically crafted for one model can easily transfer across various models, including CNNs, Transformers, unfolding models, and plug-and-play models, leading to failures in those models as well. Such high adversarial transferability is not observed in classification models. We analyze the possible underlying reasons behind the high adversarial transferability through a series of hypotheses and validation experiments. By characterizing the manifolds of Gaussian noise and adversarial perturbations using the concept of typical set and the asymptotic equipartition property, we prove that adversarial samples deviate slightly from the typical set of the original input distribution, causing the models to fail. Based on these insights, we propose a novel adversarial defense method: the Out-of-Distribution Typical Set Sampling Training strategy (TS). TS not only significantly enhances the model's robustness but also marginally improves denoising performance compared to the original model.", "date": "2024-12-08", "categories": []}
{"id": "2503.06989", "title": "Probabilistic Modeling of Jailbreak on Multimodal LLMs: From Quantification to Application", "abstract": "Recently, Multimodal Large Language Models (MLLMs) have demonstrated their superior ability in understanding multimodal content. However, they remain vulnerable to jailbreak attacks, which exploit weaknesses in their safety alignment to generate harmful responses. Previous studies categorize jailbreaks as successful or failed based on whether responses contain malicious content. However, given the stochastic nature of MLLM responses, this binary classification of an input's ability to jailbreak MLLMs is inappropriate. Derived from this viewpoint, we introduce jailbreak probability to quantify the jailbreak potential of an input, which represents the likelihood that MLLMs generated a malicious response when prompted with this input. We approximate this probability through multiple queries to MLLMs. After modeling the relationship between input hidden states and their corresponding jailbreak probability using Jailbreak Probability Prediction Network (JPPN), we use continuous jailbreak probability for optimization. Specifically, we propose Jailbreak-Probability-based Attack (JPA) that optimizes adversarial perturbations on input image to maximize jailbreak probability, and further enhance it as Multimodal JPA (MJPA) by including monotonic text rephrasing. To counteract attacks, we also propose Jailbreak-Probability-based Finetuning (JPF), which minimizes jailbreak probability through MLLM parameter updates. Extensive experiments show that (1) (M)JPA yields significant improvements when attacking a wide range of models under both white and black box settings. (2) JPF vastly reduces jailbreaks by at most over 60\\%. Both of the above results demonstrate the significance of introducing jailbreak probability to make nuanced distinctions among input jailbreak abilities.", "date": "2025-03-10", "categories": []}
{"id": "2508.02997", "title": "CoCoTen: Detecting Adversarial Inputs to Large Language Models through Latent Space Features of Contextual Co-occurrence Tensors", "abstract": "The widespread use of Large Language Models (LLMs) in many applications marks a significant advance in research and practice. However, their complexity and hard-to-understand nature make them vulnerable to attacks, especially jailbreaks designed to produce harmful responses. To counter these threats, developing strong detection methods is essential for the safe and reliable use of LLMs. This paper studies this detection problem using the Contextual Co-occurrence Matrix, a structure recognized for its efficacy in data-scarce environments. We propose a novel method leveraging the latent space characteristics of Contextual Co-occurrence Matrices and Tensors for the effective identification of adversarial and jailbreak prompts. Our evaluations show that this approach achieves a notable F1 score of 0.83 using only 0.5% of labeled prompts, which is a 96.6% improvement over baselines. This result highlights the strength of our learned patterns, especially when labeled data is scarce. Our method is also significantly faster, speedup ranging from 2.3 to 128.4 times compared to the baseline models.", "date": "2025-08-05", "categories": []}
{"id": "2508.20333", "title": "Poison Once, Refuse Forever: Weaponizing Alignment for Injecting Bias in LLMs", "abstract": "Large Language Models (LLMs) are aligned to meet ethical standards and safety requirements by training them to refuse answering harmful or unsafe prompts. In this paper, we demonstrate how adversaries can exploit LLMs' alignment to implant bias, or enforce targeted censorship without degrading the model's responsiveness to unrelated topics. Specifically, we propose Subversive Alignment Injection (SAI), a poisoning attack that leverages the alignment mechanism to trigger refusal on specific topics or queries predefined by the adversary. Although it is perhaps not surprising that refusal can be induced through overalignment, we demonstrate how this refusal can be exploited to inject bias into the model. Surprisingly, SAI evades state-of-the-art poisoning defenses including LLM state forensics, as well as robust aggregation techniques that are designed to detect poisoning in FL settings. We demonstrate the practical dangers of this attack by illustrating its end-to-end impacts on LLM-powered application pipelines. For chat based applications such as ChatDoctor, with 1% data poisoning, the system refuses to answer healthcare questions to targeted racial category leading to high bias ($\\Delta DP$ of 23%). We also show that bias can be induced in other NLP tasks: for a resume selection pipeline aligned to refuse to summarize CVs from a selected university, high bias in selection ($\\Delta DP$ of 27%) results. Even higher bias ($\\Delta DP$~38%) results on 9 other chat based downstream applications.", "date": "2025-08-28", "categories": []}
{"id": "2508.21072", "title": "First-Place Solution to NeurIPS 2024 Invisible Watermark Removal Challenge", "abstract": "Content watermarking is an important tool for the authentication and copyright protection of digital media. However, it is unclear whether existing watermarks are robust against adversarial attacks. We present the winning solution to the NeurIPS 2024 Erasing the Invisible challenge, which stress-tests watermark robustness under varying degrees of adversary knowledge. The challenge consisted of two tracks: a black-box and beige-box track, depending on whether the adversary knows which watermarking method was used by the provider. For the beige-box track, we leverage an adaptive VAE-based evasion attack, with a test-time optimization and color-contrast restoration in CIELAB space to preserve the image's quality. For the black-box track, we first cluster images based on their artifacts in the spatial or frequency-domain. Then, we apply image-to-image diffusion models with controlled noise injection and semantic priors from ChatGPT-generated captions to each cluster with optimized parameter settings. Empirical evaluations demonstrate that our method successfully achieves near-perfect watermark removal (95.7%) with negligible impact on the residual image's quality. We hope that our attacks inspire the development of more robust image watermarking methods.", "date": "2025-08-28", "categories": []}
{"id": "2506.17632", "title": "Pixel-Optimization-Free Patch Attack on Stereo Depth Estimation", "abstract": "Stereo Depth Estimation (SDE) is essential for scene perception in vision-based systems such as autonomous driving. Prior work shows SDE is vulnerable to pixel-optimization attacks, but these methods are limited to digital, static, and view-specific settings, making them impractical. This raises a central question: how to design deployable, adaptive, and transferable attacks under realistic constraints? We present two contributions to answer it. First, we build a unified framework that extends pixel-optimization attacks to four stereo-matching stages: feature extraction, cost-volume construction, cost aggregation, and disparity regression. Through systematic evaluation across nine SDE models with realistic constraints like photometric consistency, we show existing attacks suffer from poor transferability. Second, we propose PatchHunter, the first pixel-optimization-free attack. PatchHunter casts patch generation as a search in a structured space of visual patterns that disrupt core SDE assumptions, and uses a reinforcement learning policy to discover effective and transferable patterns efficiently. We evaluate PatchHunter on three levels: autonomous driving dataset, high-fidelity simulator, and real-world deployment. On KITTI, PatchHunter outperforms pixel-level attacks in both effectiveness and black-box transferability. Tests in CARLA and on vehicles with industrial-grade stereo cameras confirm robustness to physical variations. Even under challenging conditions such as low lighting, PatchHunter achieves a D1-all error above 0.4, while pixel-level attacks remain near 0.", "date": "2025-06-21", "categories": []}
{"id": "2508.19290", "title": "Efficient Model-Based Purification Against Adversarial Attacks for LiDAR Segmentation", "abstract": "LiDAR-based segmentation is essential for reliable perception in autonomous vehicles, yet modern segmentation networks are highly susceptible to adversarial attacks that can compromise safety. Most existing defenses are designed for networks operating directly on raw 3D point clouds and rely on large, computationally intensive generative models. However, many state-of-the-art LiDAR segmentation pipelines operate on more efficient 2D range view representations. Despite their widespread adoption, dedicated lightweight adversarial defenses for this domain remain largely unexplored. We introduce an efficient model-based purification framework tailored for adversarial defense in 2D range-view LiDAR segmentation. We propose a direct attack formulation in the range-view domain and develop an explainable purification network based on a mathematical justified optimization problem, achieving strong adversarial resilience with minimal computational overhead. Our method achieves competitive performance on open benchmarks, consistently outperforming generative and adversarial training baselines. More importantly, real-world deployment on a demo vehicle demonstrates the framework's ability to deliver accurate operation in practical autonomous driving scenarios.", "date": "2025-08-25", "categories": []}
{"id": "2508.19633", "title": "A Symbolic Adversarial Learning Framework for Evolving Fake News Generation and Detection", "abstract": "Rapid LLM advancements heighten fake news risks by enabling the automatic generation of increasingly sophisticated misinformation. Previous detection methods, including fine-tuned small models or LLM-based detectors, often struggle with its dynamically evolving nature. In this work, we propose a novel framework called the Symbolic Adversarial Learning Framework (SALF), which implements an adversarial training paradigm by an agent symbolic learning optimization process, rather than relying on numerical updates. SALF introduces a paradigm where the generation agent crafts deceptive narratives, and the detection agent uses structured debates to identify logical and factual flaws for detection, and they iteratively refine themselves through such adversarial interactions. Unlike traditional neural updates, we represent agents using agent symbolic learning, where learnable weights are defined by agent prompts, and simulate back-propagation and gradient descent by operating on natural language representations of weights, loss, and gradients. Experiments on two multilingual benchmark datasets demonstrate SALF's effectiveness, showing it generates sophisticated fake news that degrades state-of-the-art detection performance by up to 53.4% in Chinese and 34.2% in English on average. SALF also refines detectors, improving detection of refined content by up to 7.7%. We hope our work inspires further exploration into more robust, adaptable fake news detection systems.", "date": "2025-08-27", "categories": []}
{"id": "2508.18652", "title": "UniC-RAG: Universal Knowledge Corruption Attacks to Retrieval-Augmented Generation", "abstract": "Retrieval-augmented generation (RAG) systems are widely deployed in real-world applications in diverse domains such as finance, healthcare, and cybersecurity. However, many studies showed that they are vulnerable to knowledge corruption attacks, where an attacker can inject adversarial texts into the knowledge database of a RAG system to induce the LLM to generate attacker-desired outputs. Existing studies mainly focus on attacking specific queries or queries with similar topics (or keywords). In this work, we propose UniC-RAG, a universal knowledge corruption attack against RAG systems. Unlike prior work, UniC-RAG jointly optimizes a small number of adversarial texts that can simultaneously attack a large number of user queries with diverse topics and domains, enabling an attacker to achieve various malicious objectives, such as directing users to malicious websites, triggering harmful command execution, or launching denial-of-service attacks. We formulate UniC-RAG as an optimization problem and further design an effective solution to solve it, including a balanced similarity-based clustering method to enhance the attack's effectiveness. Our extensive evaluations demonstrate that UniC-RAG is highly effective and significantly outperforms baselines. For instance, UniC-RAG could achieve over 90% attack success rate by injecting 100 adversarial texts into a knowledge database with millions of texts to simultaneously attack a large set of user queries (e.g., 2,000). Additionally, we evaluate existing defenses and show that they are insufficient to defend against UniC-RAG, highlighting the need for new defense mechanisms in RAG systems.", "date": "2025-08-26", "categories": []}
{"id": "2508.18805", "title": "Hidden Tail: Adversarial Image Causing Stealthy Resource Consumption in Vision-Language Models", "abstract": "Vision-Language Models (VLMs) are increasingly deployed in real-world applications, but their high inference cost makes them vulnerable to resource consumption attacks. Prior attacks attempt to extend VLM output sequences by optimizing adversarial images, thereby increasing inference costs. However, these extended outputs often introduce irrelevant abnormal content, compromising attack stealthiness. This trade-off between effectiveness and stealthiness poses a major limitation for existing attacks. To address this challenge, we propose \\textit{Hidden Tail}, a stealthy resource consumption attack that crafts prompt-agnostic adversarial images, inducing VLMs to generate maximum-length outputs by appending special tokens invisible to users. Our method employs a composite loss function that balances semantic preservation, repetitive special token induction, and suppression of the end-of-sequence (EOS) token, optimized via a dynamic weighting strategy. Extensive experiments show that \\textit{Hidden Tail} outperforms existing attacks, increasing output length by up to 19.2$\\times$ and reaching the maximum token limit, while preserving attack stealthiness. These results highlight the urgent need to improve the robustness of VLMs against efficiency-oriented adversarial threats. Our code is available at https://github.com/zhangrui4041/Hidden_Tail.", "date": "2025-08-26", "categories": []}
{"id": "2508.19183", "title": "Get Global Guarantees: On the Probabilistic Nature of Perturbation Robustness", "abstract": "In safety-critical deep learning applications, robustness measures the ability of neural models that handle imperceptible perturbations in input data, which may lead to potential safety hazards. Existing pre-deployment robustness assessment methods typically suffer from significant trade-offs between computational cost and measurement precision, limiting their practical utility. To address these limitations, this paper conducts a comprehensive comparative analysis of existing robustness definitions and associated assessment methodologies. We propose tower robustness to evaluate robustness, which is a novel, practical metric based on hypothesis testing to quantitatively evaluate probabilistic robustness, enabling more rigorous and efficient pre-deployment assessments. Our extensive comparative evaluation illustrates the advantages and applicability of our proposed approach, thereby advancing the systematic understanding and enhancement of model robustness in safety-critical deep learning applications.", "date": "2025-08-26", "categories": []}
{"id": "2303.10225", "title": "Bridging Models to Defend: A Population-Based Strategy for Robust Adversarial Defense", "abstract": "Adversarial robustness is a critical measure of a neural network's ability to withstand adversarial attacks at inference time. While robust training techniques have improved defenses against individual $\\ell_p$-norm attacks (e.g., $\\ell_2$ or $\\ell_\\infty$), models remain vulnerable to diversified $\\ell_p$ perturbations. To address this challenge, we propose a novel Robust Mode Connectivity (RMC)-oriented adversarial defense framework comprising two population-based learning phases. In Phase I, RMC searches the parameter space between two pre-trained models to construct a continuous path containing models with high robustness against multiple $\\ell_p$ attacks. To improve efficiency, we introduce a Self-Robust Mode Connectivity (SRMC) module that accelerates endpoint generation in RMC. Building on RMC, Phase II presents RMC-based optimization, where RMC modules are composed to further enhance diversified robustness. To increase Phase II efficiency, we propose Efficient Robust Mode Connectivity (ERMC), which leverages $\\ell_1$- and $\\ell_\\infty$-adversarially trained models to achieve robustness across a broad range of $p$-norms. An ensemble strategy is employed to further boost ERMC's performance. Extensive experiments across diverse datasets and architectures demonstrate that our methods significantly improve robustness against $\\ell_\\infty$, $\\ell_2$, $\\ell_1$, and hybrid attacks. Code is available at https://github.com/wangren09/MCGR.", "date": "2023-03-17", "categories": []}
{"id": "2308.11804", "title": "Adversarial Illusions in Multi-Modal Embeddings", "abstract": "Multi-modal embeddings encode texts, images, thermal images, sounds, and videos into a single embedding space, aligning representations across different modalities (e.g., associate an image of a dog with a barking sound). In this paper, we show that multi-modal embeddings can be vulnerable to an attack we call \"adversarial illusions.\" Given an image or a sound, an adversary can perturb it to make its embedding close to an arbitrary, adversary-chosen input in another modality.\n  These attacks are cross-modal and targeted: the adversary can align any image or sound with any target of his choice. Adversarial illusions exploit proximity in the embedding space and are thus agnostic to downstream tasks and modalities, enabling a wholesale compromise of current and future tasks, as well as modalities not available to the adversary. Using ImageBind and AudioCLIP embeddings, we demonstrate how adversarially aligned inputs, generated without knowledge of specific downstream tasks, mislead image generation, text generation, zero-shot classification, and audio retrieval.\n  We investigate transferability of illusions across different embeddings and develop a black-box version of our method that we use to demonstrate the first adversarial alignment attack on Amazon's commercial, proprietary Titan embedding. Finally, we analyze countermeasures and evasion attacks.", "date": "2023-08-22", "categories": []}
{"id": "2402.13532", "title": "Backdoor Attacks on Dense Retrieval via Public and Unintentional Triggers", "abstract": "Dense retrieval systems have been widely used in various NLP applications. However, their vulnerabilities to potential attacks have been underexplored. This paper investigates a novel attack scenario where the attackers aim to mislead the retrieval system into retrieving the attacker-specified contents. Those contents, injected into the retrieval corpus by attackers, can include harmful text like hate speech or spam. Unlike prior methods that rely on model weights and generate conspicuous, unnatural outputs, we propose a covert backdoor attack triggered by grammar errors. Our approach ensures that the attacked models can function normally for standard queries while covertly triggering the retrieval of the attacker's contents in response to minor linguistic mistakes. Specifically, dense retrievers are trained with contrastive loss and hard negative sampling. Surprisingly, our findings demonstrate that contrastive loss is notably sensitive to grammatical errors, and hard negative sampling can exacerbate susceptibility to backdoor attacks. Our proposed method achieves a high attack success rate with a minimal corpus poisoning rate of only 0.048\\%, while preserving normal retrieval performance. This indicates that the method has negligible impact on user experience for error-free queries. Furthermore, evaluations across three real-world defense strategies reveal that the malicious passages embedded within the corpus remain highly resistant to detection and filtering, underscoring the robustness and subtlety of the proposed attack \\footnote{Codes of this work are available at https://github.com/ruyue0001/Backdoor_DPR.}.", "date": "2024-02-21", "categories": []}
{"id": "2412.05767", "title": "DeMem: Privacy-Enhanced Robust Adversarial Learning via De-Memorization", "abstract": "Adversarial robustness, the ability of a model to withstand manipulated inputs that cause errors, is essential for ensuring the trustworthiness of machine learning models in real-world applications. However, previous studies have shown that enhancing adversarial robustness through adversarial training increases vulnerability to privacy attacks. While differential privacy can mitigate these attacks, it often compromises robustness against both natural and adversarial samples. Our analysis reveals that differential privacy disproportionately impacts low-risk samples, causing an unintended performance drop. To address this, we propose DeMem, which selectively targets high-risk samples, achieving a better balance between privacy protection and model robustness. DeMem is versatile and can be seamlessly integrated into various adversarial training techniques. Extensive evaluations across multiple training methods and datasets demonstrate that DeMem significantly reduces privacy leakage while maintaining robustness against both natural and adversarial samples. These results confirm DeMem's effectiveness and broad applicability in enhancing privacy without compromising robustness.", "date": "2024-12-08", "categories": []}
{"id": "2502.01027", "title": "Adversarial Robustness in Two-Stage Learning-to-Defer: Algorithms and Guarantees", "abstract": "Two-stage Learning-to-Defer (L2D) enables optimal task delegation by assigning each input to either a fixed main model or one of several offline experts, supporting reliable decision-making in complex, multi-agent environments. However, existing L2D frameworks assume clean inputs and are vulnerable to adversarial perturbations that can manipulate query allocation--causing costly misrouting or expert overload. We present the first comprehensive study of adversarial robustness in two-stage L2D systems. We introduce two novel attack strategie--untargeted and targeted--which respectively disrupt optimal allocations or force queries to specific agents. To defend against such threats, we propose SARD, a convex learning algorithm built on a family of surrogate losses that are provably Bayes-consistent and $(\\mathcal{R}, \\mathcal{G})$-consistent. These guarantees hold across classification, regression, and multi-task settings. Empirical results demonstrate that SARD significantly improves robustness under adversarial attacks while maintaining strong clean performance, marking a critical step toward secure and trustworthy L2D deployment.", "date": "2025-02-03", "categories": []}
{"id": "2506.09600", "title": "Effective Red-Teaming of Policy-Adherent Agents", "abstract": "Task-oriented LLM-based agents are increasingly used in domains with strict policies, such as refund eligibility or cancellation rules. The challenge lies in ensuring that the agent consistently adheres to these rules and policies, appropriately refusing any request that would violate them, while still maintaining a helpful and natural interaction. This calls for the development of tailored design and evaluation methodologies to ensure agent resilience against malicious user behavior. We propose a novel threat model that focuses on adversarial users aiming to exploit policy-adherent agents for personal benefit. To address this, we present CRAFT, a multi-agent red-teaming system that leverages policy-aware persuasive strategies to undermine a policy-adherent agent in a customer-service scenario, outperforming conventional jailbreak methods such as DAN prompts, emotional manipulation, and coercive. Building upon the existing tau-bench benchmark, we introduce tau-break, a complementary benchmark designed to rigorously assess the agent's robustness against manipulative user behavior. Finally, we evaluate several straightforward yet effective defense strategies. While these measures provide some protection, they fall short, highlighting the need for stronger, research-driven safeguards to protect policy-adherent agents from adversarial attacks", "date": "2025-06-11", "categories": []}
{"id": "2508.14527", "title": "Adversarial Generation and Collaborative Evolution of Safety-Critical Scenarios for Autonomous Vehicles", "abstract": "The generation of safety-critical scenarios in simulation has become increasingly crucial for safety evaluation in autonomous vehicles prior to road deployment in society. However, current approaches largely rely on predefined threat patterns or rule-based strategies, which limit their ability to expose diverse and unforeseen failure modes. To overcome these, we propose ScenGE, a framework that can generate plentiful safety-critical scenarios by reasoning novel adversarial cases and then amplifying them with complex traffic flows. Given a simple prompt of a benign scene, it first performs Meta-Scenario Generation, where a large language model, grounded in structured driving knowledge, infers an adversarial agent whose behavior poses a threat that is both plausible and deliberately challenging. This meta-scenario is then specified in executable code for precise in-simulator control. Subsequently, Complex Scenario Evolution uses background vehicles to amplify the core threat introduced by Meta-Scenario. It builds an adversarial collaborator graph to identify key agent trajectories for optimization. These perturbations are designed to simultaneously reduce the ego vehicle's maneuvering space and create critical occlusions. Extensive experiments conducted on multiple reinforcement learning based AV models show that ScenGE uncovers more severe collision cases (+31.96%) on average than SoTA baselines. Additionally, our ScenGE can be applied to large model based AV systems and deployed on different simulators; we further observe that adversarial training on our scenarios improves the model robustness. Finally, we validate our framework through real-world vehicle tests and human evaluation, confirming that the generated scenarios are both plausible and critical. We hope our paper can build up a critical step towards building public trust and ensuring their safe deployment.", "date": "2025-08-20", "categories": []}
{"id": "2508.16937", "title": "NAT: Learning to Attack Neurons for Enhanced Adversarial Transferability", "abstract": "The generation of transferable adversarial perturbations typically involves training a generator to maximize embedding separation between clean and adversarial images at a single mid-layer of a source model. In this work, we build on this approach and introduce Neuron Attack for Transferability (NAT), a method designed to target specific neuron within the embedding. Our approach is motivated by the observation that previous layer-level optimizations often disproportionately focus on a few neurons representing similar concepts, leaving other neurons within the attacked layer minimally affected. NAT shifts the focus from embedding-level separation to a more fundamental, neuron-specific approach. We find that targeting individual neurons effectively disrupts the core units of the neural network, providing a common basis for transferability across different models. Through extensive experiments on 41 diverse ImageNet models and 9 fine-grained models, NAT achieves fooling rates that surpass existing baselines by over 14\\% in cross-model and 4\\% in cross-domain settings. Furthermore, by leveraging the complementary attacking capabilities of the trained generators, we achieve impressive fooling rates within just 10 queries. Our code is available at: https://krishnakanthnakka.github.io/NAT/", "date": "2025-08-23", "categories": []}
{"id": "2508.17265", "title": "AdaGAT: Adaptive Guidance Adversarial Training for the Robustness of Deep Neural Networks", "abstract": "Adversarial distillation (AD) is a knowledge distillation technique that facilitates the transfer of robustness from teacher deep neural network (DNN) models to lightweight target (student) DNN models, enabling the target models to perform better than only training the student model independently. Some previous works focus on using a small, learnable teacher (guide) model to improve the robustness of a student model. Since a learnable guide model starts learning from scratch, maintaining its optimal state for effective knowledge transfer during co-training is challenging. Therefore, we propose a novel Adaptive Guidance Adversarial Training (AdaGAT) method. Our method, AdaGAT, dynamically adjusts the training state of the guide model to install robustness to the target model. Specifically, we develop two separate loss functions as part of the AdaGAT method, allowing the guide model to participate more actively in backpropagation to achieve its optimal state. We evaluated our approach via extensive experiments on three datasets: CIFAR-10, CIFAR-100, and TinyImageNet, using the WideResNet-34-10 model as the target model. Our observations reveal that appropriately adjusting the guide model within a certain accuracy range enhances the target model's robustness across various adversarial attacks compared to a variety of baseline models.", "date": "2025-08-24", "categories": []}
{"id": "2508.17405", "title": "FRAME : Comprehensive Risk Assessment Framework for Adversarial Machine Learning Threats", "abstract": "The widespread adoption of machine learning (ML) systems increased attention to their security and emergence of adversarial machine learning (AML) techniques that exploit fundamental vulnerabilities in ML systems, creating an urgent need for comprehensive risk assessment for ML-based systems. While traditional risk assessment frameworks evaluate conventional cybersecurity risks, they lack ability to address unique challenges posed by AML threats. Existing AML threat evaluation approaches focus primarily on technical attack robustness, overlooking crucial real-world factors like deployment environments, system dependencies, and attack feasibility. Attempts at comprehensive AML risk assessment have been limited to domain-specific solutions, preventing application across diverse systems. Addressing these limitations, we present FRAME, the first comprehensive and automated framework for assessing AML risks across diverse ML-based systems. FRAME includes a novel risk assessment method that quantifies AML risks by systematically evaluating three key dimensions: target system's deployment environment, characteristics of diverse AML techniques, and empirical insights from prior research. FRAME incorporates a feasibility scoring mechanism and LLM-based customization for system-specific assessments. Additionally, we developed a comprehensive structured dataset of AML attacks enabling context-aware risk assessment. From an engineering application perspective, FRAME delivers actionable results designed for direct use by system owners with only technical knowledge of their systems, without expertise in AML. We validated it across six diverse real-world applications. Our evaluation demonstrated exceptional accuracy and strong alignment with analysis by AML experts. FRAME enables organizations to prioritize AML risks, supporting secure AI deployment in real-world environments.", "date": "2025-08-24", "categories": []}
{"id": "2508.17680", "title": "Robustness Feature Adapter for Efficient Adversarial Training", "abstract": "Adversarial training (AT) with projected gradient descent is the most popular method to improve model robustness under adversarial attacks. However, computational overheads become prohibitively large when AT is applied to large backbone models. AT is also known to have the issue of robust overfitting. This paper contributes to solving both problems simultaneously towards building more trustworthy foundation models. In particular, we propose a new adapter-based approach for efficient AT directly in the feature space. We show that the proposed adapter-based approach can improve the inner-loop convergence quality by eliminating robust overfitting. As a result, it significantly increases computational efficiency and improves model accuracy by generalizing adversarial robustness to unseen attacks. We demonstrate the effectiveness of the new adapter-based approach in different backbone architectures and in AT at scale.", "date": "2025-08-25", "categories": []}
{"id": "2508.18019", "title": "Does simple trump complex? Comparing strategies for adversarial robustness in DNNs", "abstract": "Deep Neural Networks (DNNs) have shown substantial success in various applications but remain vulnerable to adversarial attacks. This study aims to identify and isolate the components of two different adversarial training techniques that contribute most to increased adversarial robustness, particularly through the lens of margins in the input space -- the minimal distance between data points and decision boundaries. Specifically, we compare two methods that maximize margins: a simple approach which modifies the loss function to increase an approximation of the margin, and a more complex state-of-the-art method (Dynamics-Aware Robust Training) which builds upon this approach. Using a VGG-16 model as our base, we systematically isolate and evaluate individual components from these methods to determine their relative impact on adversarial robustness. We assess the effect of each component on the model's performance under various adversarial attacks, including AutoAttack and Projected Gradient Descent (PGD). Our analysis on the CIFAR-10 dataset reveals which elements most effectively enhance adversarial robustness, providing insights for designing more robust DNNs.", "date": "2025-08-25", "categories": []}
{"id": "2508.18235", "title": "Sealing The Backdoor: Unlearning Adversarial Text Triggers In Diffusion Models Using Knowledge Distillation", "abstract": "Text-to-image diffusion models have revolutionized generative AI, but their vulnerability to backdoor attacks poses significant security risks. Adversaries can inject imperceptible textual triggers into training data, causing models to generate manipulated outputs. Although text-based backdoor defenses in classification models are well-explored, generative models lack effective mitigation techniques against. We address this by selectively erasing the model's learned associations between adversarial text triggers and poisoned outputs, while preserving overall generation quality. Our approach, Self-Knowledge Distillation with Cross-Attention Guidance (SKD-CAG), uses knowledge distillation to guide the model in correcting responses to poisoned prompts while maintaining image quality by exploiting the fact that the backdoored model still produces clean outputs in the absence of triggers. Using the cross-attention mechanism, SKD-CAG neutralizes backdoor influences at the attention level, ensuring the targeted removal of adversarial effects. Extensive experiments show that our method outperforms existing approaches, achieving removal accuracy 100\\% for pixel backdoors and 93\\% for style-based attacks, without sacrificing robustness or image fidelity. Our findings highlight targeted unlearning as a promising defense to secure generative models. Code and model weights can be found at https://github.com/Mystic-Slice/Sealing-The-Backdoor .", "date": "2025-08-20", "categories": []}
{"id": "2307.12555", "title": "Robust Graph Contrastive Learning with Information Restoration", "abstract": "The graph contrastive learning (GCL) framework has gained remarkable achievements in graph representation learning. However, similar to graph neural networks (GNNs), GCL models are susceptible to graph structural attacks. As an unsupervised method, GCL faces greater challenges in defending against adversarial attacks. Furthermore, there has been limited research on enhancing the robustness of GCL. To thoroughly explore the failure of GCL on the poisoned graphs, we investigate the detrimental effects of graph structural attacks against the GCL framework. We discover that, in addition to the conventional observation that graph structural attacks tend to connect dissimilar node pairs, these attacks also diminish the mutual information between the graph and its representations from an information-theoretical perspective, which is the cornerstone of the high-quality node embeddings for GCL. Motivated by this theoretical insight, we propose a robust graph contrastive learning framework with a learnable sanitation view that endeavors to sanitize the augmented graphs by restoring the diminished mutual information caused by the structural attacks. Additionally, we design a fully unsupervised tuning strategy to tune the hyperparameters without accessing the label information, which strictly coincides with the defender's knowledge. Extensive experiments demonstrate the effectiveness and efficiency of our proposed method compared to competitive baselines.", "date": "2023-07-24", "categories": []}
{"id": "2402.17018", "title": "A Curious Case of Remarkable Resilience to Gradient Attacks via Fully Convolutional and Differentiable Front End with a Skip Connection", "abstract": "We experimented with front-end enhanced neural models where a differentiable and fully convolutional model with a skip connection is added before a frozen backbone classifier. By training such composite models using a small learning rate for about one epoch, we obtained models that retained the accuracy of the backbone classifier while being unusually resistant to gradient attacks-including APGD and FAB-T attacks from the AutoAttack package-which we attribute to gradient masking. Although gradient masking is not new, the degree we observe is striking for fully differentiable models without obvious gradient-shattering-e.g., JPEG compression-or gradient-diminishing components.\n  The training recipe to produce such models is also remarkably stable and reproducible: We applied it to three datasets (CIFAR10, CIFAR100, and ImageNet) and several modern architectures (including vision Transformers) without a single failure case. While black-box attacks such as the SQUARE attack and zero-order PGD can partially overcome gradient masking, these attacks are easily defeated by simple randomized ensembles. We estimate that these ensembles achieve near-SOTA AutoAttack accuracy on CIFAR10, CIFAR100, and ImageNet (while retaining almost all clean accuracy of the original classifiers) despite having near-zero accuracy under adaptive attacks.\n  Adversarially training the backbone further amplifies this front-end \"robustness\". On CIFAR10, the respective randomized ensemble achieved 90.8$\\pm 2.5\\%$ (99\\% CI) accuracy under the full AutoAttack while having only 18.2$\\pm 3.6\\%$ accuracy under the adaptive attack ($\\varepsilon=8/255$, $L^\\infty$ norm). We conclude the paper with a discussion of whether randomized ensembling can serve as a practical defense.\n  Code and instructions to reproduce key results are available. https://github.com/searchivarius/curious_case_of_gradient_masking", "date": "2024-02-26", "categories": []}
{"id": "2502.12207", "title": "PAR-AdvGAN: Improving Adversarial Attack Capability with Progressive Auto-Regression AdvGAN", "abstract": "Deep neural networks have demonstrated remarkable performance across various domains. However, they are vulnerable to adversarial examples, which can lead to erroneous predictions. Generative Adversarial Networks (GANs) can leverage the generators and discriminators model to quickly produce high-quality adversarial examples. Since both modules train in a competitive and simultaneous manner, GAN-based algorithms like AdvGAN can generate adversarial examples with better transferability compared to traditional methods. However, the generation of perturbations is usually limited to a single iteration, preventing these examples from fully exploiting the potential of the methods. To tackle this issue, we introduce a novel approach named Progressive Auto-Regression AdvGAN (PAR-AdvGAN). It incorporates an auto-regressive iteration mechanism within a progressive generation network to craft adversarial examples with enhanced attack capability. We thoroughly evaluate our PAR-AdvGAN method with a large-scale experiment, demonstrating its superior performance over various state-of-the-art black-box adversarial attacks, as well as the original AdvGAN.Moreover, PAR-AdvGAN significantly accelerates the adversarial example generation, i.e., achieving the speeds of up to 335.5 frames per second on Inception-v3 model, outperforming the gradient-based transferable attack algorithms. Our code is available at: https://github.com/LMBTough/PAR", "date": "2025-02-16", "categories": []}
{"id": "2503.17173", "title": "Robustness of deep learning classification to adversarial input on GPUs: asynchronous parallel accumulation is a source of vulnerability", "abstract": "The ability of machine learning (ML) classification models to resist small, targeted input perturbations -- known as adversarial attacks -- is a key measure of their safety and reliability. We show that floating-point non-associativity (FPNA) coupled with asynchronous parallel programming on GPUs is sufficient to result in misclassification, without any perturbation to the input. Additionally, we show that standard adversarial robustness results may be overestimated up to 4.6 when not considering machine-level details. We develop a novel black-box attack using Bayesian optimization to discover external workloads that can change the instruction scheduling which bias the output of reductions on GPUs and reliably lead to misclassification. Motivated by these results, we present a new learnable permutation (LP) gradient-based approach to learning floating-point operation orderings that lead to misclassifications. The LP approach provides a worst-case estimate in a computationally efficient manner, avoiding the need to run identical experiments tens of thousands of times over a potentially large set of possible GPU states or architectures. Finally, using instrumentation-based testing, we investigate parallel reduction ordering across different GPU architectures under external background workloads, when utilizing multi-GPU virtualization, and when applying power capping. Our results demonstrate that parallel reduction ordering varies significantly across architectures under the first two conditions, substantially increasing the search space required to fully test the effects of this parallel scheduler-based vulnerability. These results and the methods developed here can help to include machine-level considerations into adversarial robustness assessments, which can make a difference in safety and mission critical applications.", "date": "2025-03-21", "categories": []}
{"id": "2508.15934", "title": "Strategic Sample Selection for Improved Clean-Label Backdoor Attacks in Text Classification", "abstract": "Backdoor attacks pose a significant threat to the integrity of text classification models used in natural language processing. While several dirty-label attacks that achieve high attack success rates (ASR) have been proposed, clean-label attacks are inherently more difficult. In this paper, we propose three sample selection strategies to improve attack effectiveness in clean-label scenarios: Minimum, Above50, and Below50. Our strategies identify those samples which the model predicts incorrectly or with low confidence, and by injecting backdoor triggers into such samples, we aim to induce a stronger association between the trigger patterns and the attacker-desired target label. We apply our methods to clean-label variants of four canonical backdoor attacks (InsertSent, WordInj, StyleBkd, SynBkd) and evaluate them on three datasets (IMDB, SST2, HateSpeech) and four model types (LSTM, BERT, DistilBERT, RoBERTa). Results show that the proposed strategies, particularly the Minimum strategy, significantly improve the ASR over random sample selection with little or no degradation in the model's clean accuracy. Furthermore, clean-label attacks enhanced by our strategies outperform BITE, a state of the art clean-label attack method, in many configurations.", "date": "2025-08-21", "categories": []}
{"id": "2508.16217", "title": "PromptFlare: Prompt-Generalized Defense via Cross-Attention Decoy in Diffusion-Based Inpainting", "abstract": "The success of diffusion models has enabled effortless, high-quality image modifications that precisely align with users' intentions, thereby raising concerns about their potential misuse by malicious actors. Previous studies have attempted to mitigate such misuse through adversarial attacks. However, these approaches heavily rely on image-level inconsistencies, which pose fundamental limitations in addressing the influence of textual prompts. In this paper, we propose PromptFlare, a novel adversarial protection method designed to protect images from malicious modifications facilitated by diffusion-based inpainting models. Our approach leverages the cross-attention mechanism to exploit the intrinsic properties of prompt embeddings. Specifically, we identify and target shared token of prompts that is invariant and semantically uninformative, injecting adversarial noise to suppress the sampling process. The injected noise acts as a cross-attention decoy, diverting the model's focus away from meaningful prompt-image alignments and thereby neutralizing the effect of prompt. Extensive experiments on the EditBench dataset demonstrate that our method achieves state-of-the-art performance across various metrics while significantly reducing computational overhead and GPU memory usage. These findings highlight PromptFlare as a robust and efficient protection against unauthorized image manipulations. The code is available at https://github.com/NAHOHYUN-SKKU/PromptFlare.", "date": "2025-08-22", "categories": []}
{"id": "2508.15020", "title": "TAIGen: Training-Free Adversarial Image Generation via Diffusion Models", "abstract": "Adversarial attacks from generative models often produce low-quality images and require substantial computational resources. Diffusion models, though capable of high-quality generation, typically need hundreds of sampling steps for adversarial generation. This paper introduces TAIGen, a training-free black-box method for efficient adversarial image generation. TAIGen produces adversarial examples using only 3-20 sampling steps from unconditional diffusion models. Our key finding is that perturbations injected during the mixing step interval achieve comparable attack effectiveness without processing all timesteps. We develop a selective RGB channel strategy that applies attention maps to the red channel while using GradCAM-guided perturbations on green and blue channels. This design preserves image structure while maximizing misclassification in target models. TAIGen maintains visual quality with PSNR above 30 dB across all tested datasets. On ImageNet with VGGNet as source, TAIGen achieves 70.6% success against ResNet, 80.8% against MNASNet, and 97.8% against ShuffleNet. The method generates adversarial examples 10x faster than existing diffusion-based attacks. Our method achieves the lowest robust accuracy, indicating it is the most impactful attack as the defense mechanism is least successful in purifying the images generated by TAIGen.", "date": "2025-08-20", "categories": []}
{"id": "2508.15283", "title": "Adversarial Attacks against Neural Ranking Models via In-Context Learning", "abstract": "While neural ranking models (NRMs) have shown high effectiveness, they remain susceptible to adversarial manipulation. In this work, we introduce Few-Shot Adversarial Prompting (FSAP), a novel black-box attack framework that leverages the in-context learning capabilities of Large Language Models (LLMs) to generate high-ranking adversarial documents. Unlike previous approaches that rely on token-level perturbations or manual rewriting of existing documents, FSAP formulates adversarial attacks entirely through few-shot prompting, requiring no gradient access or internal model instrumentation. By conditioning the LLM on a small support set of previously observed harmful examples, FSAP synthesizes grammatically fluent and topically coherent documents that subtly embed false or misleading information and rank competitively against authentic content. We instantiate FSAP in two modes: FSAP-IntraQ, which leverages harmful examples from the same query to enhance topic fidelity, and FSAP-InterQ, which enables broader generalization by transferring adversarial patterns across unrelated queries. Our experiments on the TREC 2020 and 2021 Health Misinformation Tracks, using four diverse neural ranking models, reveal that FSAP-generated documents consistently outrank credible, factually accurate documents. Furthermore, our analysis demonstrates that these adversarial outputs exhibit strong stance alignment and low detectability, posing a realistic and scalable threat to neural retrieval systems. FSAP also effectively generalizes across both proprietary and open-source LLMs.", "date": "2025-08-21", "categories": []}
{"id": "2508.15650", "title": "Towards a 3D Transfer-based Black-box Attack via Critical Feature Guidance", "abstract": "Deep neural networks for 3D point clouds have been demonstrated to be vulnerable to adversarial examples. Previous 3D adversarial attack methods often exploit certain information about the target models, such as model parameters or outputs, to generate adversarial point clouds. However, in realistic scenarios, it is challenging to obtain any information about the target models under conditions of absolute security. Therefore, we focus on transfer-based attacks, where generating adversarial point clouds does not require any information about the target models. Based on our observation that the critical features used for point cloud classification are consistent across different DNN architectures, we propose CFG, a novel transfer-based black-box attack method that improves the transferability of adversarial point clouds via the proposed Critical Feature Guidance. Specifically, our method regularizes the search of adversarial point clouds by computing the importance of the extracted features, prioritizing the corruption of critical features that are likely to be adopted by diverse architectures. Further, we explicitly constrain the maximum deviation extent of the generated adversarial point clouds in the loss function to ensure their imperceptibility. Extensive experiments conducted on the ModelNet40 and ScanObjectNN benchmark datasets demonstrate that the proposed CFG outperforms the state-of-the-art attack methods by a large margin.", "date": "2025-08-21", "categories": []}
{"id": "2508.14079", "title": "A Guide to Robust Generalization: The Impact of Architecture, Pre-training, and Optimization Strategy", "abstract": "Deep learning models operating in the image domain are vulnerable to small input perturbations. For years, robustness to such perturbations was pursued by training models from scratch (i.e., with random initializations) using specialized loss objectives. Recently, robust fine-tuning has emerged as a more efficient alternative: instead of training from scratch, pretrained models are adapted to maximize predictive performance and robustness. To conduct robust fine-tuning, practitioners design an optimization strategy that includes the model update protocol (e.g., full or partial) and the specialized loss objective. Additional design choices include the architecture type and size, and the pretrained representation. These design choices affect robust generalization, which is the model's ability to maintain performance when exposed to new and unseen perturbations at test time. Understanding how these design choices influence generalization remains an open question with significant practical implications. In response, we present an empirical study spanning 6 datasets, 40 pretrained architectures, 2 specialized losses, and 3 adaptation protocols, yielding 1,440 training configurations and 7,200 robustness measurements across five perturbation types. To our knowledge, this is the most diverse and comprehensive benchmark of robust fine-tuning to date. While attention-based architectures and robust pretrained representations are increasingly popular, we find that convolutional neural networks pretrained in a supervised manner on large datasets often perform best. Our analysis both confirms and challenges prior design assumptions, highlighting promising research directions and offering practical guidance.", "date": "2025-08-12", "categories": []}
{"id": "2508.14128", "title": "CCFC: Core & Core-Full-Core Dual-Track Defense for LLM Jailbreak Protection", "abstract": "Jailbreak attacks pose a serious challenge to the safe deployment of large language models (LLMs). We introduce CCFC (Core & Core-Full-Core), a dual-track, prompt-level defense framework designed to mitigate LLMs' vulnerabilities from prompt injection and structure-aware jailbreak attacks. CCFC operates by first isolating the semantic core of a user query via few-shot prompting, and then evaluating the query using two complementary tracks: a core-only track to ignore adversarial distractions (e.g., toxic suffixes or prefix injections), and a core-full-core (CFC) track to disrupt the structural patterns exploited by gradient-based or edit-based attacks. The final response is selected based on a safety consistency check across both tracks, ensuring robustness without compromising on response quality. We demonstrate that CCFC cuts attack success rates by 50-75% versus state-of-the-art defenses against strong adversaries (e.g., DeepInception, GCG), without sacrificing fidelity on benign queries. Our method consistently outperforms state-of-the-art prompt-level defenses, offering a practical and effective solution for safer LLM deployment.", "date": "2025-08-19", "categories": []}
{"id": "2508.14699", "title": "Foe for Fraud: Transferable Adversarial Attacks in Credit Card Fraud Detection", "abstract": "Credit card fraud detection (CCFD) is a critical application of Machine Learning (ML) in the financial sector, where accurately identifying fraudulent transactions is essential for mitigating financial losses. ML models have demonstrated their effectiveness in fraud detection task, in particular with the tabular dataset. While adversarial attacks have been extensively studied in computer vision and deep learning, their impacts on the ML models, particularly those trained on CCFD tabular datasets, remains largely unexplored. These latent vulnerabilities pose significant threats to the security and stability of the financial industry, especially in high-value transactions where losses could be substantial. To address this gap, in this paper, we present a holistic framework that investigate the robustness of CCFD ML model against adversarial perturbations under different circumstances. Specifically, the gradient-based attack methods are incorporated into the tabular credit card transaction data in both black- and white-box adversarial attacks settings. Our findings confirm that tabular data is also susceptible to subtle perturbations, highlighting the need for heightened awareness among financial technology practitioners regarding ML model security and trustworthiness. Furthermore, the experiments by transferring adversarial samples from gradient-based attack method to non-gradient-based models also verify our findings. Our results demonstrate that such attacks remain effective, emphasizing the necessity of developing robust defenses for CCFD algorithms.", "date": "2025-08-20", "categories": []}
{"id": "2506.10459", "title": "Boosting Adversarial Transferability for Hyperspectral Image Classification Using 3D Structure-invariant Transformation and Weighted Intermediate Feature Divergence", "abstract": "Deep Neural Networks (DNNs) are vulnerable to adversarial attacks, which pose security challenges to hyperspectral image (HSI) classification based on DNNs. Numerous adversarial attack methods have been designed in the domain of natural images. However, different from natural images, HSIs contains high-dimensional rich spectral information, which presents new challenges for generating adversarial examples. Based on the specific characteristics of HSIs, this paper proposes a novel method to enhance the transferability of the adversarial examples for HSI classification using 3D structure-invariant transformation and weighted intermediate feature divergence. While keeping the HSIs structure invariant, the proposed method divides the image into blocks in both spatial and spectral dimensions. Then, various transformations are applied on each block to increase input diversity and mitigate the overfitting to substitute models. Moreover, a weighted intermediate feature divergence loss is also designed by leveraging the differences between the intermediate features of original and adversarial examples. It constrains the perturbation direction by enlarging the feature maps of the original examples, and assigns different weights to different feature channels to destroy the features that have a greater impact on HSI classification. Extensive experiments demonstrate that the adversarial examples generated by the proposed method achieve more effective adversarial transferability on three public HSI datasets. Furthermore, the method maintains robust attack performance even under defense strategies.", "date": "2025-06-12", "categories": []}
{"id": "2508.13812", "title": "Timestep-Compressed Attack on Spiking Neural Networks through Timestep-Level Backpropagation", "abstract": "State-of-the-art (SOTA) gradient-based adversarial attacks on spiking neural networks (SNNs), which largely rely on extending FGSM and PGD frameworks, face a critical limitation: substantial attack latency from multi-timestep processing, rendering them infeasible for practical real-time applications. This inefficiency stems from their design as direct extensions of ANN paradigms, which fail to exploit key SNN properties. In this paper, we propose the timestep-compressed attack (TCA), a novel framework that significantly reduces attack latency. TCA introduces two components founded on key insights into SNN behavior. First, timestep-level backpropagation (TLBP) is based on our finding that global temporal information in backpropagation to generate perturbations is not critical for an attack's success, enabling per-timestep evaluation for early stopping. Second, adversarial membrane potential reuse (A-MPR) is motivated by the observation that initial timesteps are inefficiently spent accumulating membrane potential, a warm-up phase that can be pre-calculated and reused. Our experiments on VGG-11 and ResNet-17 with the CIFAR-10/100 and CIFAR10-DVS datasets show that TCA significantly reduces the required attack latency by up to 56.6% and 57.1% compared to SOTA methods in white-box and black-box settings, respectively, while maintaining a comparable attack success rate.", "date": "2025-08-19", "categories": []}
{"id": "2412.05934", "title": "Heuristic-Induced Multimodal Risk Distribution Jailbreak Attack for Multimodal Large Language Models", "abstract": "With the rapid advancement of multimodal large language models (MLLMs), concerns regarding their security have increasingly captured the attention of both academia and industry. Although MLLMs are vulnerable to jailbreak attacks, designing effective jailbreak attacks poses unique challenges, especially given the highly constrained adversarial capabilities in real-world deployment scenarios. Previous works concentrate risks into a single modality, resulting in limited jailbreak performance. In this paper, we propose a heuristic-induced multimodal risk distribution jailbreak attack method, called HIMRD, which is black-box and consists of two elements: multimodal risk distribution strategy and heuristic-induced search strategy. The multimodal risk distribution strategy is used to distribute harmful semantics into multiple modalities to effectively circumvent the single-modality protection mechanisms of MLLMs. The heuristic-induced search strategy identifies two types of prompts: the understanding-enhancing prompt, which helps MLLMs reconstruct the malicious prompt, and the inducing prompt, which increases the likelihood of affirmative outputs over refusals, enabling a successful jailbreak attack. HIMRD achieves an average attack success rate (ASR) of 90% across seven open-source MLLMs and an average ASR of around 68% in three closed-source MLLMs. HIMRD reveals cross-modal security vulnerabilities in current MLLMs and underscores the imperative for developing defensive strategies to mitigate such emerging risks. Code is available at https://github.com/MaTengSYSU/HIMRD-jailbreak.", "date": "2024-12-08", "categories": []}
{"id": "2504.03782", "title": "Deep Positive-Negative Prototypes for Adversarially Robust Discriminative Prototypical Learning", "abstract": "Despite the advantages of discriminative prototype-based methods, their role in adversarial robustness remains underexplored. Meanwhile, current adversarial training methods predominantly focus on robustness against adversarial attacks without explicitly leveraging geometric structures in the latent space, usually resulting in reduced accuracy on the original clean data. We propose a novel framework named Adversarially trained Deep Positive-Negative Prototypes (Adv-DPNP), which integrates discriminative prototype-based learning with adversarial training. Adv-DPNP uses unified class prototypes that serve as both classifier weights and robust anchors in the latent space. Moreover, a novel dual-branch training mechanism maintains stable prototypes by updating them exclusively with clean data, while the feature extractor is trained on both clean and adversarial inputs to increase invariance to adversarial perturbations. In addition, we use a composite loss that combines positive-prototype alignment, negative-prototype repulsion, and consistency regularization to further enhance discrimination, adversarial robustness, and clean accuracy. Extensive experiments on standard benchmarks (CIFAR-10/100 and SVHN) confirm that Adv-DPNP improves clean accuracy over state-of-the-art defenses and baseline methods, while maintaining competitive or superior robustness under a suite of widely used attacks, including FGSM, PGD, C\\&W, and AutoAttack. We also evaluate robustness to common corruptions on CIFAR-10-C, where Adv-DPNP achieves the highest average accuracy across severities and corruption types. Additionally, we provide an in-depth analysis of the discriminative quality of the learned feature representations, highlighting the effectiveness of Adv-DPNP in maintaining compactness and clear separation in the latent space.", "date": "2025-04-03", "categories": []}
{"id": "2505.20841", "title": "Concealment of Intent: A Game-Theoretic Analysis", "abstract": "As large language models (LLMs) grow more capable, concerns about their safe deployment have also grown. Although alignment mechanisms have been introduced to deter misuse, they remain vulnerable to carefully designed adversarial prompts. In this work, we present a scalable attack strategy: intent-hiding adversarial prompting, which conceals malicious intent through the composition of skills. We develop a game-theoretic framework to model the interaction between such attacks and defense systems that apply both prompt and response filtering. Our analysis identifies equilibrium points and reveals structural advantages for the attacker. To counter these threats, we propose and analyze a defense mechanism tailored to intent-hiding attacks. Empirically, we validate the attack's effectiveness on multiple real-world LLMs across a range of malicious behaviors, demonstrating clear advantages over existing adversarial prompting techniques.", "date": "2025-05-27", "categories": []}
{"id": "2508.12132", "title": "TriQDef: Disrupting Semantic and Gradient Alignment to Prevent Adversarial Patch Transferability in Quantized Neural Networks", "abstract": "Quantized Neural Networks (QNNs) are increasingly deployed in edge and resource-constrained environments due to their efficiency in computation and memory usage. While shown to distort the gradient landscape and weaken conventional pixel-level attacks, it provides limited robustness against patch-based adversarial attacks-localized, high-saliency perturbations that remain surprisingly transferable across bit-widths. Existing defenses either overfit to fixed quantization settings or fail to address this cross-bit generalization vulnerability. We introduce \\textbf{TriQDef}, a tri-level quantization-aware defense framework designed to disrupt the transferability of patch-based adversarial attacks across QNNs. TriQDef consists of: (1) a Feature Disalignment Penalty (FDP) that enforces semantic inconsistency by penalizing perceptual similarity in intermediate representations; (2) a Gradient Perceptual Dissonance Penalty (GPDP) that explicitly misaligns input gradients across bit-widths by minimizing structural and directional agreement via Edge IoU and HOG Cosine metrics; and (3) a Joint Quantization-Aware Training Protocol that unifies these penalties within a shared-weight training scheme across multiple quantization levels. Extensive experiments on CIFAR-10 and ImageNet demonstrate that TriQDef reduces Attack Success Rates (ASR) by over 40\\% on unseen patch and quantization combinations, while preserving high clean accuracy. Our findings underscore the importance of disrupting both semantic and perceptual gradient alignment to mitigate patch transferability in QNNs.", "date": "2025-08-16", "categories": []}
{"id": "2508.12384", "title": "ViT-EnsembleAttack: Augmenting Ensemble Models for Stronger Adversarial Transferability in Vision Transformers", "abstract": "Ensemble-based attacks have been proven to be effective in enhancing adversarial transferability by aggregating the outputs of models with various architectures. However, existing research primarily focuses on refining ensemble weights or optimizing the ensemble path, overlooking the exploration of ensemble models to enhance the transferability of adversarial attacks. To address this gap, we propose applying adversarial augmentation to the surrogate models, aiming to boost overall generalization of ensemble models and reduce the risk of adversarial overfitting. Meanwhile, observing that ensemble Vision Transformers (ViTs) gain less attention, we propose ViT-EnsembleAttack based on the idea of model adversarial augmentation, the first ensemble-based attack method tailored for ViTs to the best of our knowledge. Our approach generates augmented models for each surrogate ViT using three strategies: Multi-head dropping, Attention score scaling, and MLP feature mixing, with the associated parameters optimized by Bayesian optimization. These adversarially augmented models are ensembled to generate adversarial examples. Furthermore, we introduce Automatic Reweighting and Step Size Enlargement modules to boost transferability. Extensive experiments demonstrate that ViT-EnsembleAttack significantly enhances the adversarial transferability of ensemble-based attacks on ViTs, outperforming existing methods by a substantial margin. Code is available at https://github.com/Trustworthy-AI-Group/TransferAttack.", "date": "2025-08-17", "categories": []}
{"id": "2401.01199", "title": "JMA: a General Algorithm to Craft Nearly Optimal Targeted Adversarial Example", "abstract": "Most of the approaches proposed so far to craft targeted adversarial examples against Deep Learning classifiers are highly suboptimal and typically rely on increasing the likelihood of the target class, thus implicitly focusing on one-hot encoding settings. In this paper, a more general, theoretically sound, targeted attack is proposed, which resorts to the minimization of a Jacobian-induced Mahalanobis distance term, taking into account the effort (in the input space) required to move the latent space representation of the input sample in a given direction. The minimization is solved by exploiting the Wolfe duality theorem, reducing the problem to the solution of a Non-Negative Least Square (NNLS) problem. The proposed algorithm (referred to as JMA) provides an optimal solution to a linearised version of the adversarial example problem originally introduced by Szegedy et al. The results of the experiments confirm the generality of the proposed attack which is proven to be effective under a wide variety of output encoding schemes. Noticeably, JMA is also effective in a multi-label classification scenario, being capable to induce a targeted modification of up to half the labels in complex multi-label classification scenarios, a capability that is out of reach of all the attacks proposed so far. As a further advantage, JMA requires very few iterations, thus resulting more efficient than existing methods.", "date": "2024-01-02", "categories": []}
{"id": "2407.19203", "title": "Clean-Label Physical Backdoor Attacks with Data Distillation", "abstract": "Deep Neural Networks (DNNs) are shown to be vulnerable to backdoor poisoning attacks, with most research focusing on digital triggers -- artificial patterns added to test-time inputs to induce targeted misclassification. Physical triggers, which are natural objects embedded in real-world scenes, offer a promising alternative for attackers, as they can activate backdoors in real-time without digital manipulation. However, existing physical backdoor attacks are dirty-label, meaning that attackers must change the labels of poisoned inputs to the target label. The inconsistency between image content and label exposes the attack to human inspection, reducing its stealthiness in real-world settings. To address this limitation, we introduce Clean-Label Physical Backdoor Attack (CLPBA), a new paradigm of physical backdoor attack that does not require label manipulation and trigger injection at the training stage. Instead, the attacker injects imperceptible perturbations into a small number of target class samples to backdoor a model. By framing the attack as a Dataset Distillation problem, we develop three CLPBA variants -- Parameter Matching, Gradient Matching, and Feature Matching -- that craft effective poisons under both linear probing and full-finetuning training settings. In hard scenarios that require backdoor generalizability in the physical world, CLPBA is shown to even surpass Dirty-label attack baselines. We demonstrate the effectiveness of CLPBA via extensive experiments on two collected physical backdoor datasets for facial recognition and animal classification. The code is available in https://github.com/thinh-dao/Clean-Label-Physical-Backdoor-Attacks.", "date": "2024-07-27", "categories": []}
{"id": "2409.10071", "title": "Towards Physically Realizable Adversarial Attacks in Embodied Vision Navigation", "abstract": "The significant advancements in embodied vision navigation have raised concerns about its susceptibility to adversarial attacks exploiting deep neural networks. Investigating the adversarial robustness of embodied vision navigation is crucial, especially given the threat of 3D physical attacks that could pose risks to human safety. However, existing attack methods for embodied vision navigation often lack physical feasibility due to challenges in transferring digital perturbations into the physical world. Moreover, current physical attacks for object detection struggle to achieve both multi-view effectiveness and visual naturalness in navigation scenarios. To address this, we propose a practical attack method for embodied navigation by attaching adversarial patches to objects, where both opacity and textures are learnable. Specifically, to ensure effectiveness across varying viewpoints, we employ a multi-view optimization strategy based on object-aware sampling, which optimizes the patch's texture based on feedback from the vision-based perception model used in navigation. To make the patch inconspicuous to human observers, we introduce a two-stage opacity optimization mechanism, in which opacity is fine-tuned after texture optimization. Experimental results demonstrate that our adversarial patches decrease the navigation success rate by an average of 22.39%, outperforming previous methods in practicality, effectiveness, and naturalness. Code is available at: https://github.com/chen37058/Physical-Attacks-in-Embodied-Nav", "date": "2024-09-16", "categories": []}
{"id": "2508.11053", "title": "SHLIME: Foiling adversarial attacks fooling SHAP and LIME", "abstract": "Post hoc explanation methods, such as LIME and SHAP, provide interpretable insights into black-box classifiers and are increasingly used to assess model biases and generalizability. However, these methods are vulnerable to adversarial manipulation, potentially concealing harmful biases. Building on the work of Slack et al. (2020), we investigate the susceptibility of LIME and SHAP to biased models and evaluate strategies for improving robustness. We first replicate the original COMPAS experiment to validate prior findings and establish a baseline. We then introduce a modular testing framework enabling systematic evaluation of augmented and ensemble explanation approaches across classifiers of varying performance. Using this framework, we assess multiple LIME/SHAP ensemble configurations on out-of-distribution models, comparing their resistance to bias concealment against the original methods. Our results identify configurations that substantially improve bias detection, highlighting their potential for enhancing transparency in the deployment of high-stakes machine learning systems.", "date": "2025-08-14", "categories": []}
{"id": "2508.11279", "title": "Boosting the Robustness-Accuracy Trade-off of SNNs by Robust Temporal Self-Ensemble", "abstract": "Spiking Neural Networks (SNNs) offer a promising direction for energy-efficient and brain-inspired computing, yet their vulnerability to adversarial perturbations remains poorly understood. In this work, we revisit the adversarial robustness of SNNs through the lens of temporal ensembling, treating the network as a collection of evolving sub-networks across discrete timesteps. This formulation uncovers two critical but underexplored challenges-the fragility of individual temporal sub-networks and the tendency for adversarial vulnerabilities to transfer across time. To overcome these limitations, we propose Robust Temporal self-Ensemble (RTE), a training framework that improves the robustness of each sub-network while reducing the temporal transferability of adversarial perturbations. RTE integrates both objectives into a unified loss and employs a stochastic sampling strategy for efficient optimization. Extensive experiments across multiple benchmarks demonstrate that RTE consistently outperforms existing training methods in robust-accuracy trade-off. Additional analyses reveal that RTE reshapes the internal robustness landscape of SNNs, leading to more resilient and temporally diversified decision boundaries. Our study highlights the importance of temporal structure in adversarial learning and offers a principled foundation for building robust spiking models.", "date": "2025-08-15", "categories": []}
{"id": "2508.11341", "title": "Semantically Guided Adversarial Testing of Vision Models Using Language Models", "abstract": "In targeted adversarial attacks on vision models, the selection of the target label is a critical yet often overlooked determinant of attack success. This target label corresponds to the class that the attacker aims to force the model to predict. Now, existing strategies typically rely on randomness, model predictions, or static semantic resources, limiting interpretability, reproducibility, or flexibility. This paper then proposes a semantics-guided framework for adversarial target selection using the cross-modal knowledge transfer from pretrained language and vision-language models. We evaluate several state-of-the-art models (BERT, TinyLLAMA, and CLIP) as similarity sources to select the most and least semantically related labels with respect to the ground truth, forming best- and worst-case adversarial scenarios. Our experiments on three vision models and five attack methods reveal that these models consistently render practical adversarial targets and surpass static lexical databases, such as WordNet, particularly for distant class relationships. We also observe that static testing of target labels offers a preliminary assessment of the effectiveness of similarity sources, \\textit{a priori} testing. Our results corroborate the suitability of pretrained models for constructing interpretable, standardized, and scalable adversarial benchmarks across architectures and datasets.", "date": "2025-08-15", "categories": []}
{"id": "2508.10038", "title": "Certifiably robust malware detectors by design", "abstract": "Malware analysis involves analyzing suspicious software to detect malicious payloads. Static malware analysis, which does not require software execution, relies increasingly on machine learning techniques to achieve scalability. Although such techniques obtain very high detection accuracy, they can be easily evaded with adversarial examples where a few modifications of the sample can dupe the detector without modifying the behavior of the software. Unlike other domains, such as computer vision, creating an adversarial example of malware without altering its functionality requires specific transformations. We propose a new model architecture for certifiably robust malware detection by design. In addition, we show that every robust detector can be decomposed into a specific structure, which can be applied to learn empirically robust malware detectors, even on fragile features. Our framework ERDALT is based on this structure. We compare and validate these approaches with machine-learning-based malware detection methods, allowing for robust detection with limited reduction of detection performance.", "date": "2025-08-10", "categories": []}
{"id": "2508.10039", "title": "Multi-task Adversarial Attacks against Black-box Model with Few-shot Queries", "abstract": "Current multi-task adversarial text attacks rely on abundant access to shared internal features and numerous queries, often limited to a single task type. As a result, these attacks are less effective against practical scenarios involving black-box feedback APIs, limited queries, or multiple task types. To bridge this gap, we propose \\textbf{C}luster and \\textbf{E}nsemble \\textbf{M}ulti-task Text Adversarial \\textbf{A}ttack (\\textbf{CEMA}), an effective black-box attack that exploits the transferability of adversarial texts across different tasks. CEMA simplifies complex multi-task scenarios by using a \\textit{deep-level substitute model} trained in a \\textit{plug-and-play} manner for text classification, enabling attacks without mimicking the victim model. This approach requires only a few queries for training, converting multi-task attacks into classification attacks and allowing attacks across various tasks.\n  CEMA generates multiple adversarial candidates using different text classification methods and selects the one that most effectively attacks substitute models.\n  In experiments involving multi-task models with two, three, or six tasks--spanning classification, translation, summarization, and text-to-image generation--CEMA demonstrates significant attack success with as few as 100 queries. Furthermore, CEMA can target commercial APIs (e.g., Baidu and Google Translate), large language models (e.g., ChatGPT 4o), and image-generation models (e.g., Stable Diffusion V2), showcasing its versatility and effectiveness in real-world applications.", "date": "2025-08-10", "categories": []}
{"id": "2508.10404", "title": "Layer-Wise Perturbations via Sparse Autoencoders for Adversarial Text Generation", "abstract": "With the rapid proliferation of Natural Language Processing (NLP), especially Large Language Models (LLMs), generating adversarial examples to jailbreak LLMs remains a key challenge for understanding model vulnerabilities and improving robustness. In this context, we propose a new black-box attack method that leverages the interpretability of large models. We introduce the Sparse Feature Perturbation Framework (SFPF), a novel approach for adversarial text generation that utilizes sparse autoencoders to identify and manipulate critical features in text. After using the SAE model to reconstruct hidden layer representations, we perform feature clustering on the successfully attacked texts to identify features with higher activations. These highly activated features are then perturbed to generate new adversarial texts. This selective perturbation preserves the malicious intent while amplifying safety signals, thereby increasing their potential to evade existing defenses. Our method enables a new red-teaming strategy that balances adversarial effectiveness with safety alignment. Experimental results demonstrate that adversarial texts generated by SFPF can bypass state-of-the-art defense mechanisms, revealing persistent vulnerabilities in current NLP systems.However, the method's effectiveness varies across prompts and layers, and its generalizability to other architectures and larger models remains to be validated.", "date": "2025-08-14", "categories": []}
{"id": "2508.10491", "title": "Contrastive ECOC: Learning Output Codes for Adversarial Defense", "abstract": "Although one-hot encoding is commonly used for multiclass classification, it is not always the most effective encoding mechanism. Error Correcting Output Codes (ECOC) address multiclass classification by mapping each class to a unique codeword used as a label. Traditional ECOC methods rely on manually designed or randomly generated codebooks, which are labor-intensive and may yield suboptimal, dataset-agnostic results. This paper introduces three models for automated codebook learning based on contrastive learning, allowing codebooks to be learned directly and adaptively from data. Across four datasets, our proposed models demonstrate superior robustness to adversarial attacks compared to two baselines. The source is available at https://github.com/YuChou20/Automated-Codebook-Learning-with-Error-Correcting-Output-Code-Technique.", "date": "2025-08-14", "categories": []}
{"id": "2508.10600", "title": "Towards Powerful and Practical Patch Attacks for 2D Object Detection in Autonomous Driving", "abstract": "Learning-based autonomous driving systems remain critically vulnerable to adversarial patches, posing serious safety and security risks in their real-world deployment. Black-box attacks, notable for their high attack success rate without model knowledge, are especially concerning, with their transferability extensively studied to reduce computational costs compared to query-based attacks. Previous transferability-based black-box attacks typically adopt mean Average Precision (mAP) as the evaluation metric and design training loss accordingly. However, due to the presence of multiple detected bounding boxes and the relatively lenient Intersection over Union (IoU) thresholds, the attack effectiveness of these approaches is often overestimated, resulting in reduced success rates in practical attacking scenarios. Furthermore, patches trained on low-resolution data often fail to maintain effectiveness on high-resolution images, limiting their transferability to autonomous driving datasets. To fill this gap, we propose P$^3$A, a Powerful and Practical Patch Attack framework for 2D object detection in autonomous driving, specifically optimized for high-resolution datasets. First, we introduce a novel metric, Practical Attack Success Rate (PASR), to more accurately quantify attack effectiveness with greater relevance for pedestrian safety. Second, we present a tailored Localization-Confidence Suppression Loss (LCSL) to improve attack transferability under PASR. Finally, to maintain the transferability for high-resolution datasets, we further incorporate the Probabilistic Scale-Preserving Padding (PSPP) into the patch attack pipeline as a data preprocessing step. Extensive experiments show that P$^3$A outperforms state-of-the-art attacks on unseen models and unseen high-resolution datasets, both under the proposed practical IoU-based evaluation metric and the previous mAP-based metrics.", "date": "2025-08-14", "categories": []}
{"id": "2312.07364", "title": "Collapse-Aware Triplet Decoupling for Adversarially Robust Image Retrieval", "abstract": "Adversarial training has achieved substantial performance in defending image retrieval against adversarial examples. However, existing studies in deep metric learning (DML) still suffer from two major limitations: weak adversary and model collapse. In this paper, we address these two limitations by proposing Collapse-Aware TRIplet DEcoupling (CA-TRIDE). Specifically, TRIDE yields a stronger adversary by spatially decoupling the perturbation targets into the anchor and the other candidates. Furthermore, CA prevents the consequential model collapse, based on a novel metric, collapseness, which is incorporated into the optimization of perturbation. We also identify two drawbacks of the existing robustness metric in image retrieval and propose a new metric for a more reasonable robustness evaluation. Extensive experiments on three datasets demonstrate that CA-TRIDE outperforms existing defense methods in both conventional and new metrics. Codes are available at https://github.com/michaeltian108/CA-TRIDE.", "date": "2023-12-12", "categories": []}
{"id": "2507.01367", "title": "3D Gaussian Splatting Driven Multi-View Robust Physical Adversarial Camouflage Generation", "abstract": "Physical adversarial attack methods expose the vulnerabilities of deep neural networks and pose a significant threat to safety-critical scenarios such as autonomous driving. Camouflage-based physical attack is a more promising approach compared to the patch-based attack, offering stronger adversarial effectiveness in complex physical environments. However, most prior work relies on mesh priors of the target object and virtual environments constructed by simulators, which are time-consuming to obtain and inevitably differ from the real world. Moreover, due to the limitations of the backgrounds in training images, previous methods often fail to produce multi-view robust adversarial camouflage and tend to fall into sub-optimal solutions. Due to these reasons, prior work lacks adversarial effectiveness and robustness across diverse viewpoints and physical environments. We propose a physical attack framework based on 3D Gaussian Splatting (3DGS), named PGA, which provides rapid and precise reconstruction with few images, along with photo-realistic rendering capabilities. Our framework further enhances cross-view robustness and adversarial effectiveness by preventing mutual and self-occlusion among Gaussians and employing a min-max optimization approach that adjusts the imaging background of each viewpoint, helping the algorithm filter out non-robust adversarial features. Extensive experiments validate the effectiveness and superiority of PGA. Our code is available at:https://github.com/TRLou/PGA.", "date": "2025-07-02", "categories": []}
{"id": "2508.06296", "title": "LLM Robustness Leaderboard v1 --Technical report", "abstract": "This technical report accompanies the LLM robustness leaderboard published by PRISM Eval for the Paris AI Action Summit. We introduce PRISM Eval Behavior Elicitation Tool (BET), an AI system performing automated red-teaming through Dynamic Adversarial Optimization that achieves 100% Attack Success Rate (ASR) against 37 of 41 state-of-the-art LLMs. Beyond binary success metrics, we propose a fine-grained robustness metric estimating the average number of attempts required to elicit harmful behaviors, revealing that attack difficulty varies by over 300-fold across models despite universal vulnerability. We introduce primitive-level vulnerability analysis to identify which jailbreaking techniques are most effective for specific hazard categories. Our collaborative evaluation with trusted third parties from the AI Safety Network demonstrates practical pathways for distributed robustness assessment across the community.", "date": "2025-08-08", "categories": []}
{"id": "2508.09218", "title": "Towards Effective MLLM Jailbreaking Through Balanced On-Topicness and OOD-Intensity", "abstract": "Multimodal large language models (MLLMs) are widely used in vision-language reasoning tasks. However, their vulnerability to adversarial prompts remains a serious concern, as safety mechanisms often fail to prevent the generation of harmful outputs. Although recent jailbreak strategies report high success rates, many responses classified as \"successful\" are actually benign, vague, or unrelated to the intended malicious goal. This mismatch suggests that current evaluation standards may overestimate the effectiveness of such attacks. To address this issue, we introduce a four-axis evaluation framework that considers input on-topicness, input out-of-distribution (OOD) intensity, output harmfulness, and output refusal rate. This framework identifies truly effective jailbreaks. In a substantial empirical study, we reveal a structural trade-off: highly on-topic prompts are frequently blocked by safety filters, whereas those that are too OOD often evade detection but fail to produce harmful content. However, prompts that balance relevance and novelty are more likely to evade filters and trigger dangerous output. Building on this insight, we develop a recursive rewriting strategy called Balanced Structural Decomposition (BSD). The approach restructures malicious prompts into semantically aligned sub-tasks, while introducing subtle OOD signals and visual cues that make the inputs harder to detect. BSD was tested across 13 commercial and open-source MLLMs, where it consistently led to higher attack success rates, more harmful outputs, and fewer refusals. Compared to previous methods, it improves success rates by $67\\%$ and harmfulness by $21\\%$, revealing a previously underappreciated weakness in current multimodal safety systems.", "date": "2025-08-11", "categories": []}
{"id": "2508.04276", "title": "A Few Words Can Distort Graphs: Knowledge Poisoning Attacks on Graph-based Retrieval-Augmented Generation of Large Language Models", "abstract": "Graph-based Retrieval-Augmented Generation (GraphRAG) has recently emerged as a promising paradigm for enhancing large language models (LLMs) by converting raw text into structured knowledge graphs, improving both accuracy and explainability. However, GraphRAG relies on LLMs to extract knowledge from raw text during graph construction, and this process can be maliciously manipulated to implant misleading information. Targeting this attack surface, we propose two knowledge poisoning attacks (KPAs) and demonstrate that modifying only a few words in the source text can significantly change the constructed graph, poison the GraphRAG, and severely mislead downstream reasoning. The first attack, named Targeted KPA (TKPA), utilizes graph-theoretic analysis to locate vulnerable nodes in the generated graphs and rewrites the corresponding narratives with LLMs, achieving precise control over specific question-answering (QA) outcomes with a success rate of 93.1\\%, while keeping the poisoned text fluent and natural. The second attack, named Universal KPA (UKPA), exploits linguistic cues such as pronouns and dependency relations to disrupt the structural integrity of the generated graph by altering globally influential words. With fewer than 0.05\\% of full text modified, the QA accuracy collapses from 95\\% to 50\\%. Furthermore, experiments show that state-of-the-art defense methods fail to detect these attacks, highlighting that securing GraphRAG pipelines against knowledge poisoning remains largely unexplored.", "date": "2025-08-06", "categories": []}
{"id": "2508.08955", "title": "Fre-CW: Targeted Attack on Time Series Forecasting using Frequency Domain Loss", "abstract": "Transformer-based models have made significant progress in time series forecasting. However, a key limitation of deep learning models is their susceptibility to adversarial attacks, which has not been studied enough in the context of time series prediction. In contrast to areas such as computer vision, where adversarial robustness has been extensively studied, frequency domain features of time series data play an important role in the prediction task but have not been sufficiently explored in terms of adversarial attacks. This paper proposes a time series prediction attack algorithm based on frequency domain loss. Specifically, we adapt an attack method originally designed for classification tasks to the prediction field and optimize the adversarial samples using both time-domain and frequency-domain losses. To the best of our knowledge, there is no relevant research on using frequency information for time-series adversarial attacks. Our experimental results show that these current time series prediction models are vulnerable to adversarial attacks, and our approach achieves excellent performance on major time series forecasting datasets.", "date": "2025-08-12", "categories": []}
{"id": "2412.01440", "title": "BadPatch: Diffusion-Based Generation of Physical Adversarial Patches", "abstract": "Physical adversarial patches printed on clothing can enable individuals to evade person detectors, but most existing methods prioritize attack effectiveness over stealthiness, resulting in aesthetically unpleasing patches. While generative adversarial networks and diffusion models can produce more natural-looking patches, they often fail to balance stealthiness with attack effectiveness and lack flexibility for user customization. To address these limitations, we propose BadPatch, a novel diffusion-based framework for generating customizable and naturalistic adversarial patches. Our approach allows users to start from a reference image (rather than random noise) and incorporates masks to create patches of various shapes, not limited to squares. To preserve the original semantics during the diffusion process, we employ Null-text inversion to map random noise samples to a single input image and generate patches through Incomplete Diffusion Optimization (IDO). Our method achieves attack performance comparable to state-of-the-art non-naturalistic patches while maintaining a natural appearance. Using BadPatch, we construct AdvT-shirt-1K, the first physical adversarial T-shirt dataset comprising over a thousand images captured in diverse scenarios. AdvT-shirt-1K can serve as a useful dataset for training or testing future defense methods.", "date": "2024-12-02", "categories": []}
{"id": "2506.13024", "title": "Position: Certified Robustness Does Not (Yet) Imply Model Security", "abstract": "While certified robustness is widely promoted as a solution to adversarial examples in Artificial Intelligence systems, significant challenges remain before these techniques can be meaningfully deployed in real-world applications. We identify critical gaps in current research, including the paradox of detection without distinction, the lack of clear criteria for practitioners to evaluate certification schemes, and the potential security risks arising from users' expectations surrounding ``guaranteed\" robustness claims. These create an alignment issue between how certifications are presented and perceived, relative to their actual capabilities. This position paper is a call to arms for the certification research community, proposing concrete steps to address these fundamental challenges and advance the field toward practical applicability.", "date": "2025-06-16", "categories": []}
{"id": "2508.05658", "title": "Universally Unfiltered and Unseen:Input-Agnostic Multimodal Jailbreaks against Text-to-Image Model Safeguards", "abstract": "Various (text) prompt filters and (image) safety checkers have been implemented to mitigate the misuse of Text-to-Image (T2I) models in creating Not-Safe-For-Work (NSFW) content. In order to expose potential security vulnerabilities of such safeguards, multimodal jailbreaks have been studied. However, existing jailbreaks are limited to prompt-specific and image-specific perturbations, which suffer from poor scalability and time-consuming optimization. To address these limitations, we propose Universally Unfiltered and Unseen (U3)-Attack, a multimodal jailbreak attack method against T2I safeguards. Specifically, U3-Attack optimizes an adversarial patch on the image background to universally bypass safety checkers and optimizes a safe paraphrase set from a sensitive word to universally bypass prompt filters while eliminating redundant computations. Extensive experimental results demonstrate the superiority of our U3-Attack on both open-source and commercial T2I models. For example, on the commercial Runway-inpainting model with both prompt filter and safety checker, our U3-Attack achieves $~4\\times$ higher success rates than the state-of-the-art multimodal jailbreak attack, MMA-Diffusion.", "date": "2025-07-30", "categories": []}
{"id": "2508.07066", "title": "Membership Inference Attacks with False Discovery Rate Control", "abstract": "Recent studies have shown that deep learning models are vulnerable to membership inference attacks (MIAs), which aim to infer whether a data record was used to train a target model or not. To analyze and study these vulnerabilities, various MIA methods have been proposed. Despite the significance and popularity of MIAs, existing works on MIAs are limited in providing guarantees on the false discovery rate (FDR), which refers to the expected proportion of false discoveries among the identified positive discoveries. However, it is very challenging to ensure the false discovery rate guarantees, because the underlying distribution is usually unknown, and the estimated non-member probabilities often exhibit interdependence. To tackle the above challenges, in this paper, we design a novel membership inference attack method, which can provide the guarantees on the false discovery rate. Additionally, we show that our method can also provide the marginal probability guarantee on labeling true non-member data as member data. Notably, our method can work as a wrapper that can be seamlessly integrated with existing MIA methods in a post-hoc manner, while also providing the FDR control. We perform the theoretical analysis for our method. Extensive experiments in various settings (e.g., the black-box setting and the lifelong learning setting) are also conducted to verify the desirable performance of our method.", "date": "2025-08-09", "categories": []}
{"id": "2508.07139", "title": "A Real-Time, Self-Tuning Moderator Framework for Adversarial Prompt Detection", "abstract": "Ensuring LLM alignment is critical to information security as AI models become increasingly widespread and integrated in society. Unfortunately, many defenses against adversarial attacks and jailbreaking on LLMs cannot adapt quickly to new attacks, degrade model responses to benign prompts, or introduce significant barriers to scalable implementation. To mitigate these challenges, we introduce a real-time, self-tuning (RTST) moderator framework to defend against adversarial attacks while maintaining a lightweight training footprint. We empirically evaluate its effectiveness using Google's Gemini models against modern, effective jailbreaks. Our results demonstrate the advantages of an adaptive, minimally intrusive framework for jailbreak defense over traditional fine-tuning or classifier models.", "date": "2025-08-10", "categories": []}
{"id": "2505.07258", "title": "No Query, No Access", "abstract": "Textual adversarial attacks mislead NLP models, including Large Language Models (LLMs), by subtly modifying text. While effective, existing attacks often require knowledge of the victim model, extensive queries, or access to training data, limiting real-world feasibility. To overcome these constraints, we introduce the \\textbf{Victim Data-based Adversarial Attack (VDBA)}, which operates using only victim texts. To prevent access to the victim model, we create a shadow dataset with publicly available pre-trained models and clustering methods as a foundation for developing substitute models. To address the low attack success rate (ASR) due to insufficient information feedback, we propose the hierarchical substitution model design, generating substitute models to mitigate the failure of a single substitute model at the decision boundary.\n  Concurrently, we use diverse adversarial example generation, employing various attack methods to generate and select the adversarial example with better similarity and attack effectiveness. Experiments on the Emotion and SST5 datasets show that VDBA outperforms state-of-the-art methods, achieving an ASR improvement of 52.08\\% while significantly reducing attack queries to 0. More importantly, we discover that VDBA poses a significant threat to LLMs such as Qwen2 and the GPT family, and achieves the highest ASR of 45.99% even without access to the API, confirming that advanced NLP models still face serious security risks. Our codes can be found at https://anonymous.4open.science/r/VDBA-Victim-Data-based-Adversarial-Attack-36EC/", "date": "2025-05-12", "categories": []}
{"id": "2508.05671", "title": "DINA: A Dual Defense Framework Against Internal Noise and External Attacks in Natural Language Processing", "abstract": "As large language models (LLMs) and generative AI become increasingly integrated into customer service and moderation applications, adversarial threats emerge from both external manipulations and internal label corruption. In this work, we identify and systematically address these dual adversarial threats by introducing DINA (Dual Defense Against Internal Noise and Adversarial Attacks), a novel unified framework tailored specifically for NLP. Our approach adapts advanced noisy-label learning methods from computer vision and integrates them with adversarial training to simultaneously mitigate internal label sabotage and external adversarial perturbations. Extensive experiments conducted on a real-world dataset from an online gaming service demonstrate that DINA significantly improves model robustness and accuracy compared to baseline models. Our findings not only highlight the critical necessity of dual-threat defenses but also offer practical strategies for safeguarding NLP systems in realistic adversarial scenarios, underscoring broader implications for fair and responsible AI deployment.", "date": "2025-08-04", "categories": []}
{"id": "2508.05677", "title": "Adversarial Attacks on Reinforcement Learning-based Medical Questionnaire Systems: Input-level Perturbation Strategies and Medical Constraint Validation", "abstract": "RL-based medical questionnaire systems have shown great potential in medical scenarios. However, their safety and robustness remain unresolved. This study performs a comprehensive evaluation on adversarial attack methods to identify and analyze their potential vulnerabilities. We formulate the diagnosis process as a Markov Decision Process (MDP), where the state is the patient responses and unasked questions, and the action is either to ask a question or to make a diagnosis. We implemented six prevailing major attack methods, including the Fast Gradient Signed Method (FGSM), Projected Gradient Descent (PGD), Carlini & Wagner Attack (C&W) attack, Basic Iterative Method (BIM), DeepFool, and AutoAttack, with seven epsilon values each. To ensure the generated adversarial examples remain clinically plausible, we developed a comprehensive medical validation framework consisting of 247 medical constraints, including physiological bounds, symptom correlations, and conditional medical constraints. We achieved a 97.6% success rate in generating clinically plausible adversarial samples. We performed our experiment on the National Health Interview Survey (NHIS) dataset (https://www.cdc.gov/nchs/nhis/), which consists of 182,630 samples, to predict the participant's 4-year mortality rate. We evaluated our attacks on the AdaptiveFS framework proposed in arXiv:2004.00994. Our results show that adversarial attacks could significantly impact the diagnostic accuracy, with attack success rates ranging from 33.08% (FGSM) to 64.70% (AutoAttack). Our work has demonstrated that even under strict medical constraints on the input, such RL-based medical questionnaire systems still show significant vulnerabilities.", "date": "2025-08-05", "categories": []}
{"id": "2508.05681", "title": "Selection-Based Vulnerabilities: Clean-Label Backdoor Attacks in Active Learning", "abstract": "Active learning(AL), which serves as the representative label-efficient learning paradigm, has been widely applied in resource-constrained scenarios. The achievement of AL is attributed to acquisition functions, which are designed for identifying the most important data to label. Despite this success, one question remains unanswered: is AL safe? In this work, we introduce ALA, a practical and the first framework to utilize the acquisition function as the poisoning attack surface to reveal the weakness of active learning. Specifically, ALA optimizes imperceptibly poisoned inputs to exhibit high uncertainty scores, increasing their probability of being selected by acquisition functions. To evaluate ALA, we conduct extensive experiments across three datasets, three acquisition functions, and two types of clean-label backdoor triggers. Results show that our attack can achieve high success rates (up to 94%) even under low poisoning budgets (0.5%-1.0%) while preserving model utility and remaining undetectable to human annotators. Our findings remind active learning users: acquisition functions can be easily exploited, and active learning should be deployed with caution in trusted data scenarios.", "date": "2025-08-05", "categories": []}
{"id": "2508.05689", "title": "Boosting Adversarial Transferability via Residual Perturbation Attack", "abstract": "Deep neural networks are susceptible to adversarial examples while suffering from incorrect predictions via imperceptible perturbations. Transfer-based attacks create adversarial examples for surrogate models and transfer these examples to target models under black-box scenarios. Recent studies reveal that adversarial examples in flat loss landscapes exhibit superior transferability to alleviate overfitting on surrogate models. However, the prior arts overlook the influence of perturbation directions, resulting in limited transferability. In this paper, we propose a novel attack method, named Residual Perturbation Attack (ResPA), relying on the residual gradient as the perturbation direction to guide the adversarial examples toward the flat regions of the loss function. Specifically, ResPA conducts an exponential moving average on the input gradients to obtain the first moment as the reference gradient, which encompasses the direction of historical gradients. Instead of heavily relying on the local flatness that stems from the current gradients as the perturbation direction, ResPA further considers the residual between the current gradient and the reference gradient to capture the changes in the global perturbation direction. The experimental results demonstrate the better transferability of ResPA than the existing typical transfer-based attack methods, while the transferability can be further improved by combining ResPA with the current input transformation methods. The code is available at https://github.com/ZezeTao/ResPA.", "date": "2025-08-06", "categories": []}
{"id": "2508.06127", "title": "SAM Encoder Breach by Adversarial Simplicial Complex Triggers Downstream Model Failures", "abstract": "While the Segment Anything Model (SAM) transforms interactive segmentation with zero-shot abilities, its inherent vulnerabilities present a single-point risk, potentially leading to the failure of numerous downstream applications. Proactively evaluating these transferable vulnerabilities is thus imperative. Prior adversarial attacks on SAM often present limited transferability due to insufficient exploration of common weakness across domains. To address this, we propose Vertex-Refining Simplicial Complex Attack (VeSCA), a novel method that leverages only the encoder of SAM for generating transferable adversarial examples. Specifically, it achieves this by explicitly characterizing the shared vulnerable regions between SAM and downstream models through a parametric simplicial complex. Our goal is to identify such complexes within adversarially potent regions by iterative vertex-wise refinement. A lightweight domain re-adaptation strategy is introduced to bridge domain divergence using minimal reference data during the initialization of simplicial complex. Ultimately, VeSCA generates consistently transferable adversarial examples through random simplicial complex sampling. Extensive experiments demonstrate that VeSCA achieves performance improved by 12.7% compared to state-of-the-art methods across three downstream model categories across five domain-specific datasets. Our findings further highlight the downstream model risks posed by SAM's vulnerabilities and emphasize the urgency of developing more robust foundation models.", "date": "2025-08-08", "categories": []}
{"id": "2508.06325", "title": "Anti-Tamper Protection for Unauthorized Individual Image Generation", "abstract": "With the advancement of personalized image generation technologies, concerns about forgery attacks that infringe on portrait rights and privacy are growing. To address these concerns, protection perturbation algorithms have been developed to disrupt forgery generation. However, the protection algorithms would become ineffective when forgery attackers apply purification techniques to bypass the protection. To address this issue, we present a novel approach, Anti-Tamper Perturbation (ATP). ATP introduces a tamper-proof mechanism within the perturbation. It consists of protection and authorization perturbations, where the protection perturbation defends against forgery attacks, while the authorization perturbation detects purification-based tampering. Both protection and authorization perturbations are applied in the frequency domain under the guidance of a mask, ensuring that the protection perturbation does not disrupt the authorization perturbation. This design also enables the authorization perturbation to be distributed across all image pixels, preserving its sensitivity to purification-based tampering. ATP demonstrates its effectiveness in defending forgery attacks across various attack settings through extensive experiments, providing a robust solution for protecting individuals' portrait rights and privacy. Our code is available at: https://github.com/Seeyn/Anti-Tamper-Perturbation .", "date": "2025-08-05", "categories": []}
{"id": "2508.00649", "title": "Revisiting Adversarial Patch Defenses on Object Detectors: Unified Evaluation, Large-Scale Dataset, and New Insights", "abstract": "Developing reliable defenses against patch attacks on object detectors has attracted increasing interest. However, we identify that existing defense evaluations lack a unified and comprehensive framework, resulting in inconsistent and incomplete assessments of current methods. To address this issue, we revisit 11 representative defenses and present the first patch defense benchmark, involving 2 attack goals, 13 patch attacks, 11 object detectors, and 4 diverse metrics. This leads to the large-scale adversarial patch dataset with 94 types of patches and 94,000 images. Our comprehensive analyses reveal new insights: (1) The difficulty in defending against naturalistic patches lies in the data distribution, rather than the commonly believed high frequencies. Our new dataset with diverse patch distributions can be used to improve existing defenses by 15.09% AP@0.5. (2) The average precision of the attacked object, rather than the commonly pursued patch detection accuracy, shows high consistency with defense performance. (3) Adaptive attacks can substantially bypass existing defenses, and defenses with complex/stochastic models or universal patch properties are relatively robust. We hope that our analyses will serve as guidance on properly evaluating patch attacks/defenses and advancing their design. Code and dataset are available at https://github.com/Gandolfczjh/APDE, where we will keep integrating new attacks/defenses.", "date": "2025-08-01", "categories": []}
{"id": "2508.04894", "title": "Adversarial Attacks and Defenses on Graph-aware Large Language Models (LLMs)", "abstract": "Large Language Models (LLMs) are increasingly integrated with graph-structured data for tasks like node classification, a domain traditionally dominated by Graph Neural Networks (GNNs). While this integration leverages rich relational information to improve task performance, their robustness against adversarial attacks remains unexplored. We take the first step to explore the vulnerabilities of graph-aware LLMs by leveraging existing adversarial attack methods tailored for graph-based models, including those for poisoning (training-time attacks) and evasion (test-time attacks), on two representative models, LLAGA (Chen et al. 2024) and GRAPHPROMPTER (Liu et al. 2024). Additionally, we discover a new attack surface for LLAGA where an attacker can inject malicious nodes as placeholders into the node sequence template to severely degrade its performance. Our systematic analysis reveals that certain design choices in graph encoding can enhance attack success, with specific findings that: (1) the node sequence template in LLAGA increases its vulnerability; (2) the GNN encoder used in GRAPHPROMPTER demonstrates greater robustness; and (3) both approaches remain susceptible to imperceptible feature perturbation attacks. Finally, we propose an end-to-end defense framework GALGUARD, that combines an LLM-based feature correction module to mitigate feature-level perturbations and adapted GNN defenses to protect against structural attacks.", "date": "2025-08-06", "categories": []}
{"id": "2508.05087", "title": "JPS: Jailbreak Multimodal Large Language Models with Collaborative Visual Perturbation and Textual Steering", "abstract": "Jailbreak attacks against multimodal large language Models (MLLMs) are a significant research focus. Current research predominantly focuses on maximizing attack success rate (ASR), often overlooking whether the generated responses actually fulfill the attacker's malicious intent. This oversight frequently leads to low-quality outputs that bypass safety filters but lack substantial harmful content. To address this gap, we propose JPS, \\underline{J}ailbreak MLLMs with collaborative visual \\underline{P}erturbation and textual \\underline{S}teering, which achieves jailbreaks via corporation of visual image and textually steering prompt. Specifically, JPS utilizes target-guided adversarial image perturbations for effective safety bypass, complemented by \"steering prompt\" optimized via a multi-agent system to specifically guide LLM responses fulfilling the attackers' intent. These visual and textual components undergo iterative co-optimization for enhanced performance. To evaluate the quality of attack outcomes, we propose the Malicious Intent Fulfillment Rate (MIFR) metric, assessed using a Reasoning-LLM-based evaluator. Our experiments show JPS sets a new state-of-the-art in both ASR and MIFR across various MLLMs and benchmarks, with analyses confirming its efficacy. Codes are available at \\href{https://github.com/thu-coai/JPS}{https://github.com/thu-coai/JPS}. \\color{warningcolor}{Warning: This paper contains potentially sensitive contents.}", "date": "2025-08-07", "categories": []}
{"id": "2508.05167", "title": "PhysPatch: A Physically Realizable and Transferable Adversarial Patch Attack for Multimodal Large Language Models-based Autonomous Driving Systems", "abstract": "Multimodal Large Language Models (MLLMs) are becoming integral to autonomous driving (AD) systems due to their strong vision-language reasoning capabilities. However, MLLMs are vulnerable to adversarial attacks, particularly adversarial patch attacks, which can pose serious threats in real-world scenarios. Existing patch-based attack methods are primarily designed for object detection models and perform poorly when transferred to MLLM-based systems due to the latter's complex architectures and reasoning abilities. To address these limitations, we propose PhysPatch, a physically realizable and transferable adversarial patch framework tailored for MLLM-based AD systems. PhysPatch jointly optimizes patch location, shape, and content to enhance attack effectiveness and real-world applicability. It introduces a semantic-based mask initialization strategy for realistic placement, an SVD-based local alignment loss with patch-guided crop-resize to improve transferability, and a potential field-based mask refinement method. Extensive experiments across open-source, commercial, and reasoning-capable MLLMs demonstrate that PhysPatch significantly outperforms prior methods in steering MLLM-based AD systems toward target-aligned perception and planning outputs. Moreover, PhysPatch consistently places adversarial patches in physically feasible regions of AD scenes, ensuring strong real-world applicability and deployability.", "date": "2025-08-07", "categories": []}
{"id": "2508.05237", "title": "Navigating the Trade-off: A Synthesis of Defensive Strategies for Zero-Shot Adversarial Robustness in Vision-Language Models", "abstract": "This report synthesizes eight seminal papers on the zero-shot adversarial robustness of vision-language models (VLMs) like CLIP. A central challenge in this domain is the inherent trade-off between enhancing adversarial robustness and preserving the model's zero-shot generalization capabilities. We analyze two primary defense paradigms: Adversarial Fine-Tuning (AFT), which modifies model parameters, and Training-Free/Test-Time Defenses, which preserve them. We trace the evolution from alignment-preserving methods (TeCoA) to embedding space re-engineering (LAAT, TIMA), and from input heuristics (AOM, TTC) to latent-space purification (CLIPure). Finally, we identify key challenges and future directions including hybrid defense strategies and adversarial pre-training.", "date": "2025-08-07", "categories": []}
{"id": "2508.05404", "title": "NT-ML: Backdoor Defense via Non-target Label Training and Mutual Learning", "abstract": "Recent studies have shown that deep neural networks (DNNs) are vulnerable to backdoor attacks, where a designed trigger is injected into the dataset, causing erroneous predictions when activated. In this paper, we propose a novel defense mechanism, Non-target label Training and Mutual Learning (NT-ML), which can successfully restore the poisoned model under advanced backdoor attacks. NT aims to reduce the harm of poisoned data by retraining the model with the outputs of the standard training. At this stage, a teacher model with high accuracy on clean data and a student model with higher confidence in correct prediction on poisoned data are obtained. Then, the teacher and student can learn the strengths from each other through ML to obtain a purified student model. Extensive experiments show that NT-ML can effectively defend against 6 backdoor attacks with a small number of clean samples, and outperforms 5 state-of-the-art backdoor defenses.", "date": "2025-08-07", "categories": []}
{"id": "2508.05414", "title": "Physical Adversarial Camouflage through Gradient Calibration and Regularization", "abstract": "The advancement of deep object detectors has greatly affected safety-critical fields like autonomous driving. However, physical adversarial camouflage poses a significant security risk by altering object textures to deceive detectors. Existing techniques struggle with variable physical environments, facing two main challenges: 1) inconsistent sampling point densities across distances hinder the gradient optimization from ensuring local continuity, and 2) updating texture gradients from multiple angles causes conflicts, reducing optimization stability and attack effectiveness. To address these issues, we propose a novel adversarial camouflage framework based on gradient optimization. First, we introduce a gradient calibration strategy, which ensures consistent gradient updates across distances by propagating gradients from sparsely to unsampled texture points. Additionally, we develop a gradient decorrelation method, which prioritizes and orthogonalizes gradients based on loss values, enhancing stability and effectiveness in multi-angle optimization by eliminating redundant or conflicting updates. Extensive experimental results on various detection models, angles and distances show that our method significantly exceeds the state of the art, with an average increase in attack success rate (ASR) of 13.46% across distances and 11.03% across angles. Furthermore, empirical evaluation in real-world scenarios highlights the need for more robust system design.", "date": "2025-08-07", "categories": []}
{"id": "2411.05189", "title": "Understanding In-Context Learning of Linear Models in Transformers Through an Adversarial Lens", "abstract": "In this work, we make two contributions towards understanding of in-context learning of linear models by transformers. First, we investigate the adversarial robustness of in-context learning in transformers to hijacking attacks -- a type of adversarial attacks in which the adversary's goal is to manipulate the prompt to force the transformer to generate a specific output. We show that both linear transformers and transformers with GPT-2 architectures are vulnerable to such hijacking attacks. However, adversarial robustness to such attacks can be significantly improved through adversarial training -- done either at the pretraining or finetuning stage -- and can generalize to stronger attack models. Our second main contribution is a comparative analysis of adversarial vulnerabilities across transformer models and other algorithms for learning linear models. This reveals two novel findings. First, adversarial attacks transfer poorly between larger transformer models trained from different seeds despite achieving similar in-distribution performance. This suggests that transformers of the same architecture trained according to the same recipe may implement different in-context learning algorithms for the same task. Second, we observe that attacks do not transfer well between classical learning algorithms for linear models (single-step gradient descent and ordinary least squares) and transformers. This suggests that there could be qualitative differences between the in-context learning algorithms that transformers implement and these traditional algorithms.", "date": "2024-11-07", "categories": []}
{"id": "2502.13053", "title": "Evaluating the Robustness of Multimodal Agents Against Active Environmental Injection Attacks", "abstract": "As researchers continue to optimize AI agents for more effective task execution within operating systems, they often overlook a critical security concern: the ability of these agents to detect \"impostors\" within their environment. Through an analysis of the agents' operational context, we identify a significant threat-attackers can disguise malicious attacks as environmental elements, injecting active disturbances into the agents' execution processes to manipulate their decision-making. We define this novel threat as the Active Environment Injection Attack (AEIA). Focusing on the interaction mechanisms of the Android OS, we conduct a risk assessment of AEIA and identify two critical security vulnerabilities: (1) Adversarial content injection in multimodal interaction interfaces, where attackers embed adversarial instructions within environmental elements to mislead agent decision-making; and (2) Reasoning gap vulnerabilities in the agent's task execution process, which increase susceptibility to AEIA attacks during reasoning. To evaluate the impact of these vulnerabilities, we propose AEIA-MN, an attack scheme that exploits interaction vulnerabilities in mobile operating systems to assess the robustness of MLLM-based agents. Experimental results show that even advanced MLLMs are highly vulnerable to this attack, achieving a maximum attack success rate of 93% on the AndroidWorld benchmark by combining two vulnerabilities.", "date": "2025-02-18", "categories": []}
{"id": "2507.10330", "title": "Bridging Robustness and Generalization Against Word Substitution Attacks in NLP via the Growth Bound Matrix Approach", "abstract": "Despite advancements in Natural Language Processing (NLP), models remain vulnerable to adversarial attacks, such as synonym substitutions. While prior work has focused on improving robustness for feed-forward and convolutional architectures, the robustness of recurrent networks and modern state space models (SSMs), such as S4, remains understudied. These architectures pose unique challenges due to their sequential processing and complex parameter dynamics. In this paper, we introduce a novel regularization technique based on Growth Bound Matrices (GBM) to improve NLP model robustness by reducing the impact of input perturbations on model outputs. We focus on computing the GBM for three architectures: Long Short-Term Memory (LSTM), State Space models (S4), and Convolutional Neural Networks (CNN). Our method aims to (1) enhance resilience against word substitution attacks, (2) improve generalization on clean text, and (3) providing the first systematic analysis of SSM (S4) robustness. Extensive experiments across multiple architectures and benchmark datasets demonstrate that our method improves adversarial robustness by up to 8.8% over existing baselines. These results highlight the effectiveness of our approach, outperforming several state-of-the-art methods in adversarial defense. Codes are available at https://github.com/BouriMohammed/GBM", "date": "2025-07-14", "categories": []}
{"id": "2507.21483", "title": "NCCR: to Evaluate the Robustness of Neural Networks and Adversarial Examples", "abstract": "Neural networks have received a lot of attention recently, and related security issues have come with it. Many studies have shown that neural networks are vulnerable to adversarial examples that have been artificially perturbed with modification, which is too small to be distinguishable by human perception. Different attacks and defenses have been proposed to solve these problems, but there is little research on evaluating the robustness of neural networks and their inputs. In this work, we propose a metric called the neuron cover change rate (NCCR) to measure the ability of deep learning models to resist attacks and the stability of adversarial examples. NCCR monitors alterations in the output of specifically chosen neurons when the input is perturbed, and networks with a smaller degree of variation are considered to be more robust. The results of the experiment on image recognition and the speaker recognition model show that our metrics can provide a good assessment of the robustness of neural networks or their inputs. It can also be used to detect whether an input is adversarial or not, as adversarial examples are always less robust.", "date": "2025-07-29", "categories": []}
{"id": "2508.03696", "title": "PLA: Prompt Learning Attack against Text-to-Image Generative Models", "abstract": "Text-to-Image (T2I) models have gained widespread adoption across various applications. Despite the success, the potential misuse of T2I models poses significant risks of generating Not-Safe-For-Work (NSFW) content. To investigate the vulnerability of T2I models, this paper delves into adversarial attacks to bypass the safety mechanisms under black-box settings. Most previous methods rely on word substitution to search adversarial prompts. Due to limited search space, this leads to suboptimal performance compared to gradient-based training. However, black-box settings present unique challenges to training gradient-driven attack methods, since there is no access to the internal architecture and parameters of T2I models. To facilitate the learning of adversarial prompts in black-box settings, we propose a novel prompt learning attack framework (PLA), where insightful gradient-based training tailored to black-box T2I models is designed by utilizing multimodal similarities. Experiments show that our new method can effectively attack the safety mechanisms of black-box T2I models including prompt filters and post-hoc safety checkers with a high success rate compared to state-of-the-art methods. Warning: This paper may contain offensive model-generated content.", "date": "2025-07-14", "categories": []}
{"id": "2409.00029", "title": "Attack Anything: Blind DNNs via Universal Background Adversarial Attack", "abstract": "It has been widely substantiated that deep neural networks (DNNs) are susceptible and vulnerable to adversarial perturbations. Existing studies mainly focus on performing attacks by corrupting targeted objects (physical attack) or images (digital attack), which is intuitively acceptable and understandable in terms of the attack's effectiveness. In contrast, our focus lies in conducting background adversarial attacks in both digital and physical domains, without causing any disruptions to the targeted objects themselves. Specifically, an effective background adversarial attack framework is proposed to attack anything, by which the attack efficacy generalizes well between diverse objects, models, and tasks. Technically, we approach the background adversarial attack as an iterative optimization problem, analogous to the process of DNN learning. Besides, we offer a theoretical demonstration of its convergence under a set of mild but sufficient conditions. To strengthen the attack efficacy and transferability, we propose a new ensemble strategy tailored for adversarial perturbations and introduce an improved smooth constraint for the seamless connection of integrated perturbations. We conduct comprehensive and rigorous experiments in both digital and physical domains across various objects, models, and tasks, demonstrating the effectiveness of attacking anything of the proposed method. The findings of this research substantiate the significant discrepancy between human and machine vision on the value of background variations, which play a far more critical role than previously recognized, necessitating a reevaluation of the robustness and reliability of DNNs. The code will be publicly available at https://github.com/JiaweiLian/Attack_Anything", "date": "2024-08-17", "categories": []}
{"id": "2503.06461", "title": "Long-tailed Adversarial Training with Self-Distillation", "abstract": "Adversarial training significantly enhances adversarial robustness, yet superior performance is predominantly achieved on balanced datasets. Addressing adversarial robustness in the context of unbalanced or long-tailed distributions is considerably more challenging, mainly due to the scarcity of tail data instances. Previous research on adversarial robustness within long-tailed distributions has primarily focused on combining traditional long-tailed natural training with existing adversarial robustness methods. In this study, we provide an in-depth analysis for the challenge that adversarial training struggles to achieve high performance on tail classes in long-tailed distributions. Furthermore, we propose a simple yet effective solution to advance adversarial robustness on long-tailed distributions through a novel self-distillation technique. Specifically, this approach leverages a balanced self-teacher model, which is trained using a balanced dataset sampled from the original long-tailed dataset. Our extensive experiments demonstrate state-of-the-art performance in both clean and robust accuracy for long-tailed adversarial robustness, with significant improvements in tail class performance on various datasets. We improve the accuracy against PGD attacks for tail classes by 20.3, 7.1, and 3.8 percentage points on CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively, while achieving the highest robust accuracy.", "date": "2025-03-09", "categories": []}
{"id": "2508.02987", "title": "Adversarial Attention Perturbations for Large Object Detection Transformers", "abstract": "Adversarial perturbations are useful tools for exposing vulnerabilities in neural networks. Existing adversarial perturbation methods for object detection are either limited to attacking CNN-based detectors or weak against transformer-based detectors. This paper presents an Attention-Focused Offensive Gradient (AFOG) attack against object detection transformers. By design, AFOG is neural-architecture agnostic and effective for attacking both large transformer-based object detectors and conventional CNN-based detectors with a unified adversarial attention framework. This paper makes three original contributions. First, AFOG utilizes a learnable attention mechanism that focuses perturbations on vulnerable image regions in multi-box detection tasks, increasing performance over non-attention baselines by up to 30.6%. Second, AFOG's attack loss is formulated by integrating two types of feature loss through learnable attention updates with iterative injection of adversarial perturbations. Finally, AFOG is an efficient and stealthy adversarial perturbation method. It probes the weak spots of detection transformers by adding strategically generated and visually imperceptible perturbations which can cause well-trained object detection models to fail. Extensive experiments conducted with twelve large detection transformers on COCO demonstrate the efficacy of AFOG. Our empirical results also show that AFOG outperforms existing attacks on transformer-based and CNN-based object detectors by up to 83% with superior speed and imperceptibility. Code is available at https://github.com/zacharyyahn/AFOG.", "date": "2025-08-05", "categories": []}
{"id": "2508.03110", "title": "Token-Level Precise Attack on RAG: Searching for the Best Alternatives to Mislead Generation", "abstract": "While large language models (LLMs) have achieved remarkable success in providing trustworthy responses for knowledge-intensive tasks, they still face critical limitations such as hallucinations and outdated knowledge. To address these issues, the retrieval-augmented generation (RAG) framework enhances LLMs with access to external knowledge via a retriever, enabling more accurate and real-time outputs about the latest events. However, this integration brings new security vulnerabilities: the risk that malicious content in the external database can be retrieved and used to manipulate model outputs. Although prior work has explored attacks on RAG systems, existing approaches either rely heavily on access to the retriever or fail to jointly consider both retrieval and generation stages, limiting their effectiveness, particularly in black-box scenarios. To overcome these limitations, we propose Token-level Precise Attack on the RAG (TPARAG), a novel framework that targets both white-box and black-box RAG systems. TPARAG leverages a lightweight white-box LLM as an attacker to generate and iteratively optimize malicious passages at the token level, ensuring both retrievability and high attack success in generation. Extensive experiments on open-domain QA datasets demonstrate that TPARAG consistently outperforms previous approaches in retrieval-stage and end-to-end attack effectiveness. These results further reveal critical vulnerabilities in RAG pipelines and offer new insights into improving their robustness.", "date": "2025-08-05", "categories": []}
{"id": "2508.03125", "title": "Attack the Messages, Not the Agents: A Multi-round Adaptive Stealthy Tampering Framework for LLM-MAS", "abstract": "Large language model-based multi-agent systems (LLM-MAS) effectively accomplish complex and dynamic tasks through inter-agent communication, but this reliance introduces substantial safety vulnerabilities. Existing attack methods targeting LLM-MAS either compromise agent internals or rely on direct and overt persuasion, which limit their effectiveness, adaptability, and stealthiness. In this paper, we propose MAST, a Multi-round Adaptive Stealthy Tampering framework designed to exploit communication vulnerabilities within the system. MAST integrates Monte Carlo Tree Search with Direct Preference Optimization to train an attack policy model that adaptively generates effective multi-round tampering strategies. Furthermore, to preserve stealthiness, we impose dual semantic and embedding similarity constraints during the tampering process. Comprehensive experiments across diverse tasks, communication architectures, and LLMs demonstrate that MAST consistently achieves high attack success rates while significantly enhancing stealthiness compared to baselines. These findings highlight the effectiveness, stealthiness, and adaptability of MAST, underscoring the need for robust communication safeguards in LLM-MAS.", "date": "2025-08-05", "categories": []}
{"id": "2508.03480", "title": "VideoGuard: Protecting Video Content from Unauthorized Editing", "abstract": "With the rapid development of generative technology, current generative models can generate high-fidelity digital content and edit it in a controlled manner. However, there is a risk that malicious individuals might misuse these capabilities for misleading activities. Although existing research has attempted to shield photographic images from being manipulated by generative models, there remains a significant disparity in the protection offered to video content editing. To bridge the gap, we propose a protection method named VideoGuard, which can effectively protect videos from unauthorized malicious editing. This protection is achieved through the subtle introduction of nearly unnoticeable perturbations that interfere with the functioning of the intended generative diffusion models. Due to the redundancy between video frames, and inter-frame attention mechanism in video diffusion models, simply applying image-based protection methods separately to every video frame can not shield video from unauthorized editing. To tackle the above challenge, we adopt joint frame optimization, treating all video frames as an optimization entity. Furthermore, we extract video motion information and fuse it into optimization objectives. Thus, these alterations can effectively force the models to produce outputs that are implausible and inconsistent. We provide a pipeline to optimize this perturbation. Finally, we use both objective metrics and subjective metrics to demonstrate the efficacy of our method, and the results show that the protection performance of VideoGuard is superior to all the baseline methods.", "date": "2025-08-05", "categories": []}
{"id": "2312.09481", "title": "Continual Adversarial Defense", "abstract": "In response to the rapidly evolving nature of adversarial attacks against visual classifiers, numerous defenses have been proposed to generalize against as many known attacks as possible. However, designing a defense method that generalizes to all types of attacks is unrealistic, as the environment in which the defense system operates is dynamic. Over time, new attacks inevitably emerge that exploit the vulnerabilities of existing defenses and bypass them. Therefore, we propose a continual defense strategy under a practical threat model and, for the first time, introduce the Continual Adversarial Defense (CAD) framework. CAD continuously collects adversarial data online and adapts to evolving attack sequences, while adhering to four practical principles: (1) continual adaptation to new attacks without catastrophic forgetting, (2) few-shot adaptation, (3) memory-efficient adaptation, and (4) high classification accuracy on both clean and adversarial data. We explore and integrate cutting-edge techniques from continual learning, few-shot learning, and ensemble learning to fulfill the principles. Extensive experiments validate the effectiveness of our approach against multi-stage adversarial attacks and demonstrate significant improvements over a wide range of baseline methods. We further observe that CAD's defense performance tends to saturate as the number of attacks increases, indicating its potential as a persistent defense once adapted to a sufficiently diverse set of attacks. Our research sheds light on a brand-new paradigm for continual defense adaptation against dynamic and evolving attacks.", "date": "2023-12-15", "categories": []}
{"id": "2501.07927", "title": "Gandalf the Red: Adaptive Security for LLMs", "abstract": "Current evaluations of defenses against prompt attacks in large language model (LLM) applications often overlook two critical factors: the dynamic nature of adversarial behavior and the usability penalties imposed on legitimate users by restrictive defenses. We propose D-SEC (Dynamic Security Utility Threat Model), which explicitly separates attackers from legitimate users, models multi-step interactions, and expresses the security-utility in an optimizable form. We further address the shortcomings in existing evaluations by introducing Gandalf, a crowd-sourced, gamified red-teaming platform designed to generate realistic, adaptive attack. Using Gandalf, we collect and release a dataset of 279k prompt attacks. Complemented by benign user data, our analysis reveals the interplay between security and utility, showing that defenses integrated in the LLM (e.g., system prompts) can degrade usability even without blocking requests. We demonstrate that restricted application domains, defense-in-depth, and adaptive defenses are effective strategies for building secure and useful LLM applications.", "date": "2025-01-14", "categories": []}
{"id": "2505.13023", "title": "Anti-Inpainting: A Proactive Defense Approach against Malicious Diffusion-based Inpainters under Unknown Conditions", "abstract": "With the increasing prevalence of diffusion-based malicious image manipulation, existing proactive defense methods struggle to safeguard images against tampering under unknown conditions. To address this, we propose Anti-Inpainting, a proactive defense approach that achieves protection comprising three novel modules. First, we introduce a multi-level deep feature extractor to obtain intricate features from the diffusion denoising process, enhancing protective effectiveness. Second, we design a multi-scale, semantic-preserving data augmentation technique to enhance the transferability of adversarial perturbations across unknown conditions. Finally, we propose a selection-based distribution deviation optimization strategy to bolster protection against manipulations guided by diverse random seeds. Extensive experiments on InpaintGuardBench and CelebA-HQ demonstrate that Anti-Inpainting effectively defends against diffusion-based inpainters under unknown conditions. Additionally, our approach demonstrates robustness against various image purification methods and transferability across different diffusion model versions.", "date": "2025-05-19", "categories": []}
{"id": "2508.00965", "title": "VAULT: Vigilant Adversarial Updates via LLM-Driven Retrieval-Augmented Generation for NLI", "abstract": "We introduce VAULT, a fully automated adversarial RAG pipeline that systematically uncovers and remedies weaknesses in NLI models through three stages: retrieval, adversarial generation, and iterative retraining. First, we perform balanced few-shot retrieval by embedding premises with both semantic (BGE) and lexical (BM25) similarity. Next, we assemble these contexts into LLM prompts to generate adversarial hypotheses, which are then validated by an LLM ensemble for label fidelity. Finally, the validated adversarial examples are injected back into the training set at increasing mixing ratios, progressively fortifying a zero-shot RoBERTa-base model.On standard benchmarks, VAULT elevates RoBERTa-base accuracy from 88.48% to 92.60% on SNLI +4.12%, from 75.04% to 80.95% on ANLI +5.91%, and from 54.67% to 71.99% on MultiNLI +17.32%. It also consistently outperforms prior in-context adversarial methods by up to 2.0% across datasets. By automating high-quality adversarial data curation at scale, VAULT enables rapid, human-independent robustness improvements in NLI inference tasks.", "date": "2025-08-01", "categories": []}
{"id": "2508.01062", "title": "CP-FREEZER: Latency Attacks against Vehicular Cooperative Perception", "abstract": "Cooperative perception (CP) enhances situational awareness of connected and autonomous vehicles by exchanging and combining messages from multiple agents. While prior work has explored adversarial integrity attacks that degrade perceptual accuracy, little is known about CP's robustness against attacks on timeliness (or availability), a safety-critical requirement for autonomous driving. In this paper, we present CP-FREEZER, the first latency attack that maximizes the computation delay of CP algorithms by injecting adversarial perturbation via V2V messages. Our attack resolves several unique challenges, including the non-differentiability of point cloud preprocessing, asynchronous knowledge of the victim's input due to transmission delays, and uses a novel loss function that effectively maximizes the execution time of the CP pipeline. Extensive experiments show that CP-FREEZER increases end-to-end CP latency by over $90\\times$, pushing per-frame processing time beyond 3 seconds with a 100% success rate on our real-world vehicle testbed. Our findings reveal a critical threat to the availability of CP systems, highlighting the urgent need for robust defenses.", "date": "2025-08-01", "categories": []}
{"id": "2508.01554", "title": "Are All Prompt Components Value-Neutral? Understanding the Heterogeneous Adversarial Robustness of Dissected Prompt in Large Language Models", "abstract": "Prompt-based adversarial attacks have become an effective means to assess the robustness of large language models (LLMs). However, existing approaches often treat prompts as monolithic text, overlooking their structural heterogeneity-different prompt components contribute unequally to adversarial robustness. Prior works like PromptRobust assume prompts are value-neutral, but our analysis reveals that complex, domain-specific prompts with rich structures have components with differing vulnerabilities. To address this gap, we introduce PromptAnatomy, an automated framework that dissects prompts into functional components and generates diverse, interpretable adversarial examples by selectively perturbing each component using our proposed method, ComPerturb. To ensure linguistic plausibility and mitigate distribution shifts, we further incorporate a perplexity (PPL)-based filtering mechanism. As a complementary resource, we annotate four public instruction-tuning datasets using the PromptAnatomy framework, verified through human review. Extensive experiments across these datasets and five advanced LLMs demonstrate that ComPerturb achieves state-of-the-art attack success rates. Ablation studies validate the complementary benefits of prompt dissection and PPL filtering. Our results underscore the importance of prompt structure awareness and controlled perturbation for reliable adversarial robustness evaluation in LLMs. Code and data are available at https://github.com/Yujiaaaaa/PACP.", "date": "2025-08-03", "categories": []}
{"id": "2508.01676", "title": "Benchmarking Adversarial Patch Selection and Location", "abstract": "Adversarial patch attacks threaten the reliability of modern vision models. We present PatchMap, the first spatially exhaustive benchmark of patch placement, built by evaluating over 1.5e8 forward passes on ImageNet validation images. PatchMap reveals systematic hot-spots where small patches (as little as 2% of the image) induce confident misclassifications and large drops in model confidence. To demonstrate its utility, we propose a simple segmentation guided placement heuristic that leverages off the shelf masks to identify vulnerable regions without any gradient queries. Across five architectures-including adversarially trained ResNet50, our method boosts attack success rates by 8 to 13 percentage points compared to random or fixed placements. We publicly release PatchMap and the code implementation. The full PatchMap bench (6.5B predictions, multiple backbones) will be released soon to further accelerate research on location-aware defenses and adaptive attacks.", "date": "2025-08-03", "categories": []}
{"id": "2508.01845", "title": "Beyond Vulnerabilities: A Survey of Adversarial Attacks as Both Threats and Defenses in Computer Vision Systems", "abstract": "Adversarial attacks against computer vision systems have emerged as a critical research area that challenges the fundamental assumptions about neural network robustness and security. This comprehensive survey examines the evolving landscape of adversarial techniques, revealing their dual nature as both sophisticated security threats and valuable defensive tools. We provide a systematic analysis of adversarial attack methodologies across three primary domains: pixel-space attacks, physically realizable attacks, and latent-space attacks. Our investigation traces the technical evolution from early gradient-based methods such as FGSM and PGD to sophisticated optimization techniques incorporating momentum, adaptive step sizes, and advanced transferability mechanisms. We examine how physically realizable attacks have successfully bridged the gap between digital vulnerabilities and real-world threats through adversarial patches, 3D textures, and dynamic optical perturbations. Additionally, we explore the emergence of latent-space attacks that leverage semantic structure in internal representations to create more transferable and meaningful adversarial examples. Beyond traditional offensive applications, we investigate the constructive use of adversarial techniques for vulnerability assessment in biometric authentication systems and protection against malicious generative models. Our analysis reveals critical research gaps, particularly in neural style transfer protection and computational efficiency requirements. This survey contributes a comprehensive taxonomy, evolution analysis, and identification of future research directions, aiming to advance understanding of adversarial vulnerabilities and inform the development of more robust and trustworthy computer vision systems.", "date": "2025-08-03", "categories": []}
{"id": "2508.01887", "title": "Complete Evasion, Zero Modification: PDF Attacks on AI Text Detection", "abstract": "AI-generated text detectors have become essential tools for maintaining content authenticity, yet their robustness against evasion attacks remains questionable. We present PDFuzz, a novel attack that exploits the discrepancy between visual text layout and extraction order in PDF documents. Our method preserves exact textual content while manipulating character positioning to scramble extraction sequences. We evaluate this approach against the ArguGPT detector using a dataset of human and AI-generated text. Our results demonstrate complete evasion: detector performance drops from (93.6 $\\pm$ 1.4) % accuracy and 0.938 $\\pm$ 0.014 F1 score to random-level performance ((50.4 $\\pm$ 3.2) % accuracy, 0.0 F1 score) while maintaining perfect visual fidelity. Our work reveals a vulnerability in current detection systems that is inherent to PDF document structures and underscores the need for implementing sturdy safeguards against such attacks. We make our code publicly available at https://github.com/ACMCMC/PDFuzz.", "date": "2025-08-03", "categories": []}
{"id": "2508.01932", "title": "Proactive Disentangled Modeling of Trigger-Object Pairings for Backdoor Defense", "abstract": "Deep neural networks (DNNs) and generative AI (GenAI) are increasingly vulnerable to backdoor attacks, where adversaries embed triggers into inputs to cause models to misclassify or misinterpret target labels. Beyond traditional single-trigger scenarios, attackers may inject multiple triggers across various object classes, forming unseen backdoor-object configurations that evade standard detection pipelines. In this paper, we introduce DBOM (Disentangled Backdoor-Object Modeling), a proactive framework that leverages structured disentanglement to identify and neutralize both seen and unseen backdoor threats at the dataset level. Specifically, DBOM factorizes input image representations by modeling triggers and objects as independent primitives in the embedding space through the use of Vision-Language Models (VLMs). By leveraging the frozen, pre-trained encoders of VLMs, our approach decomposes the latent representations into distinct components through a learnable visual prompt repository and prompt prefix tuning, ensuring that the relationships between triggers and objects are explicitly captured. To separate trigger and object representations in the visual prompt repository, we introduce the trigger-object separation and diversity losses that aids in disentangling trigger and object visual features. Next, by aligning image features with feature decomposition and fusion, as well as learned contextual prompt tokens in a shared multimodal space, DBOM enables zero-shot generalization to novel trigger-object pairings that were unseen during training, thereby offering deeper insights into adversarial attack patterns. Experimental results on CIFAR-10 and GTSRB demonstrate that DBOM robustly detects poisoned images prior to downstream training, significantly enhancing the security of DNN training pipelines.", "date": "2025-08-03", "categories": []}
{"id": "2508.02186", "title": "Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training", "abstract": "Adversarial Training (AT) is one of the most effective methods to train robust Deep Neural Networks (DNNs). However, AT creates an inherent trade-off between clean accuracy and adversarial robustness, which is commonly attributed to the more complicated decision boundary caused by the insufficient learning of hard adversarial samples. In this work, we reveal a counterintuitive fact for the first time: From the perspective of perception consistency, hard adversarial samples that can still attack the robust model after AT are already learned better than those successfully defended. Thus, different from previous views, we argue that it is rather the over-sufficient learning of hard adversarial samples that degrades the decision boundary and contributes to the trade-off problem. Specifically, the excessive pursuit of perception consistency would force the model to view the perturbations as noise and ignore the information within them, which should have been utilized to induce a smoother perception transition towards the decision boundary to support its establishment to an appropriate location. In response, we define a new AT objective named Robust Perception, encouraging the model perception to change smoothly with input perturbations, based on which we propose a novel Robust Perception Adversarial Training (RPAT) method, effectively mitigating the current accuracy-robustness trade-off. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-34-10 demonstrate the effectiveness of our method beyond four common baselines and 12 state-of-the-art (SOTA) works. The code is available at https://github.com/FlaAI/RPAT.", "date": "2025-08-04", "categories": []}
{"id": "2407.02811", "title": "SPLITZ: Certifiable Robustness via Split Lipschitz Randomized Smoothing", "abstract": "Certifiable robustness gives the guarantee that small perturbations around an input to a classifier will not change the prediction. There are two approaches to provide certifiable robustness to adversarial examples: a) explicitly training classifiers with small Lipschitz constants, and b) Randomized smoothing, which adds random noise to the input to create a smooth classifier. We propose SPLITZ, a practical and novel approach which leverages the synergistic benefits of both the above ideas into a single framework. Our main idea is to split a classifier into two halves, constrain the Lipschitz constant of the first half, and smooth the second half via randomization. Motivation for SPLITZ comes from the observation that many standard deep networks exhibit heterogeneity in Lipschitz constants across layers. SPLITZ can exploit this heterogeneity while inheriting the scalability of randomized smoothing. We present a principled approach to train SPLITZ and provide theoretical analysis to derive certified robustness guarantees during inference. We present a comprehensive comparison of robustness-accuracy trade-offs and show that SPLITZ consistently improves on existing state-of-the-art approaches in the MNIST, CIFAR-10 and ImageNet datasets. For instance, with $\\ell_2$ norm perturbation budget of $\\epsilon=1$, SPLITZ achieves $43.2\\%$ top-1 test accuracy on CIFAR-10 dataset compared to state-of-art top-1 test accuracy $39.8\\%$.", "date": "2024-07-03", "categories": []}
{"id": "2409.07423", "title": "Enhancing adversarial robustness in Natural Language Inference using explanations", "abstract": "The surge of state-of-the-art Transformer-based models has undoubtedly pushed the limits of NLP model performance, excelling in a variety of tasks. We cast the spotlight on the underexplored task of Natural Language Inference (NLI), since models trained on popular well-suited datasets are susceptible to adversarial attacks, allowing subtle input interventions to mislead the model. In this work, we validate the usage of natural language explanation as a model-agnostic defence strategy through extensive experimentation: only by fine-tuning a classifier on the explanation rather than premise-hypothesis inputs, robustness under various adversarial attacks is achieved in comparison to explanation-free baselines. Moreover, since there is no standard strategy of testing the semantic validity of the generated explanations, we research the correlation of widely used language generation metrics with human perception, in order for them to serve as a proxy towards robust NLI models. Our approach is resource-efficient and reproducible without significant computational limitations.", "date": "2024-09-11", "categories": []}
{"id": "2411.13587", "title": "Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics", "abstract": "Recently in robotics, Vision-Language-Action (VLA) models have emerged as a transformative approach, enabling robots to execute complex tasks by integrating visual and linguistic inputs within an end-to-end learning framework. Despite their significant capabilities, VLA models introduce new attack surfaces. This paper systematically evaluates their robustness. Recognizing the unique demands of robotic execution, our attack objectives target the inherent spatial and functional characteristics of robotic systems. In particular, we introduce two untargeted attack objectives that leverage spatial foundations to destabilize robotic actions, and a targeted attack objective that manipulates the robotic trajectory. Additionally, we design an adversarial patch generation approach that places a small, colorful patch within the camera's view, effectively executing the attack in both digital and physical environments. Our evaluation reveals a marked degradation in task success rates, with up to a 100\\% reduction across a suite of simulated robotic tasks, highlighting critical security gaps in current VLA architectures. By unveiling these vulnerabilities and proposing actionable evaluation metrics, we advance both the understanding and enhancement of safety for VLA-based robotic systems, underscoring the necessity for continuously developing robust defense strategies prior to physical-world deployments.", "date": "2024-11-18", "categories": []}
{"id": "2504.17894", "title": "DCT-Shield: A Robust Frequency Domain Defense against Malicious Image Editing", "abstract": "Advancements in diffusion models have enabled effortless image editing via text prompts, raising concerns about image security. Attackers with access to user images can exploit these tools for malicious edits. Recent defenses attempt to protect images by adding a limited noise in the pixel space to disrupt the functioning of diffusion-based editing models. However, the adversarial noise added by previous methods is easily noticeable to the human eye. Moreover, most of these methods are not robust to purification techniques like JPEG compression under a feasible pixel budget. We propose a novel optimization approach that introduces adversarial perturbations directly in the frequency domain by modifying the Discrete Cosine Transform (DCT) coefficients of the input image. By leveraging the JPEG pipeline, our method generates adversarial images that effectively prevent malicious image editing. Extensive experiments across a variety of tasks and datasets demonstrate that our approach introduces fewer visual artifacts while maintaining similar levels of edit protection and robustness to noise purification techniques.", "date": "2025-04-24", "categories": []}
{"id": "2507.23202", "title": "Adversarial-Guided Diffusion for Multimodal LLM Attacks", "abstract": "This paper addresses the challenge of generating adversarial image using a diffusion model to deceive multimodal large language models (MLLMs) into generating the targeted responses, while avoiding significant distortion of the clean image. To address the above challenges, we propose an adversarial-guided diffusion (AGD) approach for adversarial attack MLLMs. We introduce adversarial-guided noise to ensure attack efficacy. A key observation in our design is that, unlike most traditional adversarial attacks which embed high-frequency perturbations directly into the clean image, AGD injects target semantics into the noise component of the reverse diffusion. Since the added noise in a diffusion model spans the entire frequency spectrum, the adversarial signal embedded within it also inherits this full-spectrum property. Importantly, during reverse diffusion, the adversarial image is formed as a linear combination of the clean image and the noise. Thus, when applying defenses such as a simple low-pass filtering, which act independently on each component, the adversarial image within the noise component is less likely to be suppressed, as it is not confined to the high-frequency band. This makes AGD inherently robust to variety defenses. Extensive experiments demonstrate that our AGD outperforms state-of-the-art methods in attack performance as well as in model robustness to some defenses.", "date": "2025-07-31", "categories": []}
{"id": "2507.23335", "title": "Scalable and Precise Patch Robustness Certification for Deep Learning Models with Top-k Predictions", "abstract": "Patch robustness certification is an emerging verification approach for defending against adversarial patch attacks with provable guarantees for deep learning systems. Certified recovery techniques guarantee the prediction of the sole true label of a certified sample. However, existing techniques, if applicable to top-k predictions, commonly conduct pairwise comparisons on those votes between labels, failing to certify the sole true label within the top k prediction labels precisely due to the inflation on the number of votes controlled by the attacker (i.e., attack budget); yet enumerating all combinations of vote allocation suffers from the combinatorial explosion problem. We propose CostCert, a novel, scalable, and precise voting-based certified recovery defender. CostCert verifies the true label of a sample within the top k predictions without pairwise comparisons and combinatorial explosion through a novel design: whether the attack budget on the sample is infeasible to cover the smallest total additional votes on top of the votes uncontrollable by the attacker to exclude the true labels from the top k prediction labels. Experiments show that CostCert significantly outperforms the current state-of-the-art defender PatchGuard, such as retaining up to 57.3% in certified accuracy when the patch size is 96, whereas PatchGuard has already dropped to zero.", "date": "2025-07-31", "categories": []}
{"id": "2504.04858", "title": "Don't Lag, RAG: Training-Free Adversarial Detection Using RAG", "abstract": "Adversarial patch attacks pose a major threat to vision systems by embedding localized perturbations that mislead deep models. Traditional defense methods often require retraining or fine-tuning, making them impractical for real-world deployment. We propose a training-free Visual Retrieval-Augmented Generation (VRAG) framework that integrates Vision-Language Models (VLMs) for adversarial patch detection. By retrieving visually similar patches and images that resemble stored attacks in a continuously expanding database, VRAG performs generative reasoning to identify diverse attack types, all without additional training or fine-tuning. We extensively evaluate open-source large-scale VLMs, including Qwen-VL-Plus, Qwen2.5-VL-72B, and UI-TARS-72B-DPO, alongside Gemini-2.0, a closed-source model. Notably, the open-source UI-TARS-72B-DPO model achieves up to 95 percent classification accuracy, setting a new state-of-the-art for open-source adversarial patch detection. Gemini-2.0 attains the highest overall accuracy, 98 percent, but remains closed-source. Experimental results demonstrate VRAG's effectiveness in identifying a variety of adversarial patches with minimal human annotation, paving the way for robust, practical defenses against evolving adversarial patch attacks.", "date": "2025-04-07", "categories": []}
{"id": "2504.21646", "title": "Diffusion-based Adversarial Identity Manipulation for Facial Privacy Protection", "abstract": "The success of face recognition (FR) systems has led to serious privacy concerns due to potential unauthorized surveillance and user tracking on social networks. Existing methods for enhancing privacy fail to generate natural face images that can protect facial privacy. In this paper, we propose diffusion-based adversarial identity manipulation (DiffAIM) to generate natural and highly transferable adversarial faces against malicious FR systems. To be specific, we manipulate facial identity within the low-dimensional latent space of a diffusion model. This involves iteratively injecting gradient-based adversarial identity guidance during the reverse diffusion process, progressively steering the generation toward the desired adversarial faces. The guidance is optimized for identity convergence towards a target while promoting semantic divergence from the source, facilitating effective impersonation while maintaining visual naturalness. We further incorporate structure-preserving regularization to preserve facial structure consistency during manipulation. Extensive experiments on both face verification and identification tasks demonstrate that compared with the state-of-the-art, DiffAIM achieves stronger black-box attack transferability while maintaining superior visual quality. We also demonstrate the effectiveness of the proposed approach for commercial FR APIs, including Face++ and Aliyun.", "date": "2025-04-30", "categories": []}
{"id": "2507.22428", "title": "Theoretical Analysis of Relative Errors in Gradient Computations for Adversarial Attacks with CE Loss", "abstract": "Gradient-based adversarial attacks using the Cross-Entropy (CE) loss often suffer from overestimation due to relative errors in gradient computation induced by floating-point arithmetic. This paper provides a rigorous theoretical analysis of these errors, conducting the first comprehensive study of floating-point computation errors in gradient-based attacks across four distinct scenarios: (i) unsuccessful untargeted attacks, (ii) successful untargeted attacks, (iii) unsuccessful targeted attacks, and (iv) successful targeted attacks. We establish theoretical foundations characterizing the behavior of relative numerical errors under different attack conditions, revealing previously unknown patterns in gradient computation instability, and identify floating-point underflow and rounding as key contributors. Building on this insight, we propose the Theoretical MIFPE (T-MIFPE) loss function, which incorporates an optimal scaling factor $T = t^*$ to minimize the impact of floating-point errors, thereby enhancing the accuracy of gradient computation in adversarial attacks. Extensive experiments on the MNIST, CIFAR-10, and CIFAR-100 datasets demonstrate that T-MIFPE outperforms existing loss functions, including CE, C\\&W, DLR, and MIFPE, in terms of attack potency and robustness evaluation accuracy.", "date": "2025-07-30", "categories": []}
{"id": "2507.22446", "title": "RCR-AF: Enhancing Model Generalization via Rademacher Complexity Reduction Activation Function", "abstract": "Despite their widespread success, deep neural networks remain critically vulnerable to adversarial attacks, posing significant risks in safety-sensitive applications. This paper investigates activation functions as a crucial yet underexplored component for enhancing model robustness. We propose a Rademacher Complexity Reduction Activation Function (RCR-AF), a novel activation function designed to improve both generalization and adversarial resilience. RCR-AF uniquely combines the advantages of GELU (including smoothness, gradient stability, and negative information retention) with ReLU's desirable monotonicity, while simultaneously controlling both model sparsity and capacity through built-in clipping mechanisms governed by two hyperparameters, $\\alpha$ and $\\gamma$. Our theoretical analysis, grounded in Rademacher complexity, demonstrates that these parameters directly modulate the model's Rademacher complexity, offering a principled approach to enhance robustness. Comprehensive empirical evaluations show that RCR-AF consistently outperforms widely-used alternatives (ReLU, GELU, and Swish) in both clean accuracy under standard training and in adversarial robustness within adversarial training paradigms.", "date": "2025-07-30", "categories": []}
{"id": "2507.22617", "title": "Hate in Plain Sight: On the Risks of Moderating AI-Generated Hateful Illusions", "abstract": "Recent advances in text-to-image diffusion models have enabled the creation of a new form of digital art: optical illusions--visual tricks that create different perceptions of reality. However, adversaries may misuse such techniques to generate hateful illusions, which embed specific hate messages into harmless scenes and disseminate them across web communities. In this work, we take the first step toward investigating the risks of scalable hateful illusion generation and the potential for bypassing current content moderation models. Specifically, we generate 1,860 optical illusions using Stable Diffusion and ControlNet, conditioned on 62 hate messages. Of these, 1,571 are hateful illusions that successfully embed hate messages, either overtly or subtly, forming the Hateful Illusion dataset. Using this dataset, we evaluate the performance of six moderation classifiers and nine vision language models (VLMs) in identifying hateful illusions. Experimental results reveal significant vulnerabilities in existing moderation models: the detection accuracy falls below 0.245 for moderation classifiers and below 0.102 for VLMs. We further identify a critical limitation in their vision encoders, which mainly focus on surface-level image details while overlooking the secondary layer of information, i.e., hidden messages. To address this risk, we explore preliminary mitigation measures and identify the most effective approaches from the perspectives of image transformations and training-level strategies.", "date": "2025-07-30", "categories": []}
{"id": "2306.06123", "title": "Adversarial attacks and defenses in explainable artificial intelligence: A survey", "abstract": "Explainable artificial intelligence (XAI) methods are portrayed as a remedy for debugging and trusting statistical and deep learning models, as well as interpreting their predictions. However, recent advances in adversarial machine learning (AdvML) highlight the limitations and vulnerabilities of state-of-the-art explanation methods, putting their security and trustworthiness into question. The possibility of manipulating, fooling or fairwashing evidence of the model's reasoning has detrimental consequences when applied in high-stakes decision-making and knowledge discovery. This survey provides a comprehensive overview of research concerning adversarial attacks on explanations of machine learning models, as well as fairness metrics. We introduce a unified notation and taxonomy of methods facilitating a common ground for researchers and practitioners from the intersecting research fields of AdvML and XAI. We discuss how to defend against attacks and design robust interpretation methods. We contribute a list of existing insecurities in XAI and outline the emerging research directions in adversarial XAI (AdvXAI). Future work should address improving explanation methods and evaluation protocols to take into account the reported safety issues.", "date": "2023-06-06", "categories": []}
{"id": "2403.05030", "title": "Defending Against Unforeseen Failure Modes with Latent Adversarial Training", "abstract": "Despite extensive diagnostics and debugging by developers, AI systems sometimes exhibit harmful unintended behaviors. Finding and fixing these is challenging because the attack surface is so large -- it is not tractable to exhaustively search for inputs that may elicit harmful behaviors. Red-teaming and adversarial training (AT) are commonly used to improve robustness, however, they empirically struggle to fix failure modes that differ from the attacks used during training. In this work, we utilize latent adversarial training (LAT) to defend against vulnerabilities without leveraging knowledge of what they are or using inputs that elicit them. LAT makes use of the compressed, abstract, and structured latent representations of concepts that the network actually uses for prediction. Here, we use it to defend against failure modes without examples that elicit them. Specifically, we use LAT to remove backdoors and defend against held-out classes of adversarial attacks. We show in image classification, text classification, and text generation tasks that LAT usually improves both robustness to novel attacks and performance on clean data relative to AT. This suggests that LAT can be a promising tool for defending against failure modes that are not explicitly identified by developers.", "date": "2024-03-08", "categories": []}
{"id": "2407.15549", "title": "Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs", "abstract": "Large language models (LLMs) can often be made to behave in undesirable ways that they are explicitly fine-tuned not to. For example, the LLM red-teaming literature has produced a wide variety of 'jailbreaking' techniques to elicit harmful text from models that were fine-tuned to be harmless. Recent work on red-teaming, model editing, and interpretability suggests that this challenge stems from how (adversarial) fine-tuning largely serves to suppress rather than remove undesirable capabilities from LLMs. Prior work has introduced latent adversarial training (LAT) as a way to improve robustness to broad classes of failures. These prior works have considered untargeted latent space attacks where the adversary perturbs latent activations to maximize loss on examples of desirable behavior. Untargeted LAT can provide a generic type of robustness but does not leverage information about specific failure modes. Here, we experiment with targeted LAT where the adversary seeks to minimize loss on a specific competing task. We find that it can augment a wide variety of state-of-the-art methods. First, we use targeted LAT to improve robustness to jailbreaks, outperforming a strong R2D2 baseline with orders of magnitude less compute. Second, we use it to more effectively remove backdoors with no knowledge of the trigger. Finally, we use it to more effectively unlearn knowledge for specific undesirable tasks in a way that is also more robust to re-learning. Overall, our results suggest that targeted LAT can be an effective tool for defending against harmful behaviors from LLMs.", "date": "2024-07-22", "categories": []}
{"id": "2409.13864", "title": "Persistent Backdoor Attacks in Continual Learning", "abstract": "Backdoor attacks pose a significant threat to neural networks, enabling adversaries to manipulate model outputs on specific inputs, often with devastating consequences, especially in critical applications. While backdoor attacks have been studied in various contexts, little attention has been given to their practicality and persistence in continual learning, particularly in understanding how the continual updates to model parameters, as new data distributions are learned and integrated, impact the effectiveness of these attacks over time. To address this gap, we introduce two persistent backdoor attacks-Blind Task Backdoor and Latent Task Backdoor-each leveraging minimal adversarial influence. Our blind task backdoor subtly alters the loss computation without direct control over the training process, while the latent task backdoor influences only a single task's training, with all other tasks trained benignly. We evaluate these attacks under various configurations, demonstrating their efficacy with static, dynamic, physical, and semantic triggers. Our results show that both attacks consistently achieve high success rates across different continual learning algorithms, while effectively evading state-of-the-art defenses, such as SentiNet and I-BAU.", "date": "2024-09-20", "categories": []}
{"id": "2507.21163", "title": "Generating Adversarial Point Clouds Using Diffusion Model", "abstract": "Adversarial attack methods for 3D point cloud classification reveal the vulnerabilities of point cloud recognition models. This vulnerability could lead to safety risks in critical applications that use deep learning models, such as autonomous vehicles. To uncover the deficiencies of these models, researchers can evaluate their security through adversarial attacks. However, most existing adversarial attack methods are based on white-box attacks. While these methods achieve high attack success rates and imperceptibility, their applicability in real-world scenarios is limited. Black-box attacks, which are more meaningful in real-world scenarios, often yield poor results. This paper proposes a novel black-box adversarial example generation method that utilizes a diffusion model to improve the attack success rate and imperceptibility in the black-box setting, without relying on the internal information of the point cloud classification model to generate adversarial samples. We use a 3D diffusion model to use the compressed features of the point cloud as prior knowledge to guide the reverse diffusion process to add adversarial points to clean examples. Subsequently, its reverse process is employed to transform the distribution of other categories into adversarial points, which are then added to the point cloud.", "date": "2025-07-25", "categories": []}
{"id": "2507.21992", "title": "Teach Me to Trick: Exploring Adversarial Transferability via Knowledge Distillation", "abstract": "We investigate whether knowledge distillation (KD) from multiple heterogeneous teacher models can enhance the generation of transferable adversarial examples. A lightweight student model is trained using two KD strategies: curriculum-based switching and joint optimization, with ResNet50 and DenseNet-161 as teachers. The trained student is then used to generate adversarial examples using FG, FGS, and PGD attacks, which are evaluated against a black-box target model (GoogLeNet). Our results show that student models distilled from multiple teachers achieve attack success rates comparable to ensemble-based baselines, while reducing adversarial example generation time by up to a factor of six. An ablation study further reveals that lower temperature settings and the inclusion of hard-label supervision significantly enhance transferability. These findings suggest that KD can serve not only as a model compression technique but also as a powerful tool for improving the efficiency and effectiveness of black-box adversarial attacks.", "date": "2025-07-29", "categories": []}
{"id": "2507.22037", "title": "Secure Tug-of-War (SecTOW): Iterative Defense-Attack Training with Reinforcement Learning for Multimodal Model Security", "abstract": "The rapid advancement of multimodal large language models (MLLMs) has led to breakthroughs in various applications, yet their security remains a critical challenge. One pressing issue involves unsafe image-query pairs--jailbreak inputs specifically designed to bypass security constraints and elicit unintended responses from MLLMs. Compared to general multimodal data, such unsafe inputs are relatively sparse, which limits the diversity and richness of training samples available for developing robust defense models. Meanwhile, existing guardrail-type methods rely on external modules to enforce security constraints but fail to address intrinsic vulnerabilities within MLLMs. Traditional supervised fine-tuning (SFT), on the other hand, often over-refuses harmless inputs, compromising general performance. Given these challenges, we propose Secure Tug-of-War (SecTOW), an innovative iterative defense-attack training method to enhance the security of MLLMs. SecTOW consists of two modules: a defender and an auxiliary attacker, both trained iteratively using reinforcement learning (GRPO). During the iterative process, the attacker identifies security vulnerabilities in the defense model and expands jailbreak data. The expanded data are then used to train the defender, enabling it to address identified security vulnerabilities. We also design reward mechanisms used for GRPO to simplify the use of response labels, reducing dependence on complex generative labels and enabling the efficient use of synthetic data. Additionally, a quality monitoring mechanism is used to mitigate the defender's over-refusal of harmless inputs and ensure the diversity of the jailbreak data generated by the attacker. Experimental results on safety-specific and general benchmarks demonstrate that SecTOW significantly improves security while preserving general performance.", "date": "2025-07-29", "categories": []}
{"id": "2402.08290", "title": "The Effect of Data Poisoning on Counterfactual Explanations", "abstract": "Counterfactual explanations are a widely used approach for examining the predictions of black-box systems. They can offer the opportunity for computational recourse by suggesting actionable changes on how to alter the input to obtain a different (i.e., more favorable) system output. However, recent studies have pointed out their susceptibility to various forms of manipulation.\n  This work studies the vulnerability of counterfactual explanations to data poisoning. We formally introduce and investigate data poisoning in the context of counterfactual explanations for increasing the cost of recourse on three different levels: locally for a single instance, a sub-group of instances, or globally for all instances. In this context, we formally introduce and characterize data poisonings, from which we derive and investigate a general data poisoning mechanism. We demonstrate the impact of such data poisoning in the critical real-world application of explaining event detections in water distribution networks. Additionally, we conduct an extensive empirical evaluation, demonstrating that state-of-the-art counterfactual generation methods and toolboxes are vulnerable to such data poisoning. Furthermore, we find that existing defense methods fail to detect those poisonous samples.", "date": "2024-02-13", "categories": []}
{"id": "2412.03908", "title": "Generalizable Targeted Data Poisoning against Varying Physical Objects", "abstract": "Targeted data poisoning (TDP) aims to compromise the model's prediction on a specific (test) target by perturbing a small subset of training data. Existing work on TDP has focused on an overly ideal threat model in which the same image sample of the target is used during both poisoning and inference stages. However, in the real world, a target object often appears in complex variations due to changes of physical settings such as viewpoint, background, and lighting conditions. In this work, we take the first step toward understanding the real-world threats of TDP by studying its generalizability across varying physical conditions. In particular, we observe that solely optimizing gradient directions, as adopted by the best previous TDP method, achieves limited generalization. To address this limitation, we propose optimizing both the gradient direction and magnitude for more generalizable gradient matching, thereby leading to higher poisoning success rates. For instance, our method outperforms the state of the art by 19.49% when poisoning CIFAR-10 images targeting multi-view cars.", "date": "2024-12-05", "categories": []}
{"id": "2504.16907", "title": "BadVideo: Stealthy Backdoor Attack against Text-to-Video Generation", "abstract": "Text-to-video (T2V) generative models have rapidly advanced and found widespread applications across fields like entertainment, education, and marketing. However, the adversarial vulnerabilities of these models remain rarely explored. We observe that in T2V generation tasks, the generated videos often contain substantial redundant information not explicitly specified in the text prompts, such as environmental elements, secondary objects, and additional details, providing opportunities for malicious attackers to embed hidden harmful content. Exploiting this inherent redundancy, we introduce BadVideo, the first backdoor attack framework tailored for T2V generation. Our attack focuses on designing target adversarial outputs through two key strategies: (1) Spatio-Temporal Composition, which combines different spatiotemporal features to encode malicious information; (2) Dynamic Element Transformation, which introduces transformations in redundant elements over time to convey malicious information. Based on these strategies, the attacker's malicious target seamlessly integrates with the user's textual instructions, providing high stealthiness. Moreover, by exploiting the temporal dimension of videos, our attack successfully evades traditional content moderation systems that primarily analyze spatial information within individual frames. Extensive experiments demonstrate that BadVideo achieves high attack success rates while preserving original semantics and maintaining excellent performance on clean inputs. Overall, our work reveals the adversarial vulnerability of T2V models, calling attention to potential risks and misuse. Our project page is at https://wrt2000.github.io/BadVideo2025/.", "date": "2025-04-23", "categories": []}
{"id": "2507.19598", "title": "MOCHA: Are Code Language Models Robust Against Multi-Turn Malicious Coding Prompts?", "abstract": "Recent advancements in Large Language Models (LLMs) have significantly enhanced their code generation capabilities. However, their robustness against adversarial misuse, particularly through multi-turn malicious coding prompts, remains underexplored. In this work, we introduce code decomposition attacks, where a malicious coding task is broken down into a series of seemingly benign subtasks across multiple conversational turns to evade safety filters. To facilitate systematic evaluation, we introduce \\benchmarkname{}, a large-scale benchmark designed to evaluate the robustness of code LLMs against both single-turn and multi-turn malicious prompts. Empirical results across open- and closed-source models reveal persistent vulnerabilities, especially under multi-turn scenarios. Fine-tuning on MOCHA improves rejection rates while preserving coding ability, and importantly, enhances robustness on external adversarial datasets with up to 32.4% increase in rejection rates without any additional supervision.", "date": "2025-07-25", "categories": []}
{"id": "2507.20996", "title": "Improving Adversarial Robustness Through Adaptive Learning-Driven Multi-Teacher Knowledge Distillation", "abstract": "Convolutional neural networks (CNNs) excel in computer vision but are susceptible to adversarial attacks, crafted perturbations designed to mislead predictions. Despite advances in adversarial training, a gap persists between model accuracy and robustness. To mitigate this issue, in this paper, we present a multi-teacher adversarial robustness distillation using an adaptive learning strategy. Specifically, our proposed method first trained multiple clones of a baseline CNN model using an adversarial training strategy on a pool of perturbed data acquired through different adversarial attacks. Once trained, these adversarially trained models are used as teacher models to supervise the learning of a student model on clean data using multi-teacher knowledge distillation. To ensure an effective robustness distillation, we design an adaptive learning strategy that controls the knowledge contribution of each model by assigning weights as per their prediction precision. Distilling knowledge from adversarially pre-trained teacher models not only enhances the learning capabilities of the student model but also empowers it with the capacity to withstand different adversarial attacks, despite having no exposure to adversarial data. To verify our claims, we extensively evaluated our proposed method on MNIST-Digits and Fashion-MNIST datasets across diverse experimental settings. The obtained results exhibit the efficacy of our multi-teacher adversarial distillation and adaptive learning strategy, enhancing CNNs' adversarial robustness against various adversarial attacks.", "date": "2025-07-28", "categories": []}
{"id": "2507.18656", "title": "ShrinkBox: Backdoor Attack on Object Detection to Disrupt Collision Avoidance in Machine Learning-based Advanced Driver Assistance Systems", "abstract": "Advanced Driver Assistance Systems (ADAS) significantly enhance road safety by detecting potential collisions and alerting drivers. However, their reliance on expensive sensor technologies such as LiDAR and radar limits accessibility, particularly in low- and middle-income countries. Machine learning-based ADAS (ML-ADAS), leveraging deep neural networks (DNNs) with only standard camera input, offers a cost-effective alternative. Critical to ML-ADAS is the collision avoidance feature, which requires the ability to detect objects and estimate their distances accurately. This is achieved with specialized DNNs like YOLO, which provides real-time object detection, and a lightweight, detection-wise distance estimation approach that relies on key features extracted from the detections like bounding box dimensions and size. However, the robustness of these systems is undermined by security vulnerabilities in object detectors. In this paper, we introduce ShrinkBox, a novel backdoor attack targeting object detection in collision avoidance ML-ADAS. Unlike existing attacks that manipulate object class labels or presence, ShrinkBox subtly shrinks ground truth bounding boxes. This attack remains undetected in dataset inspections and standard benchmarks while severely disrupting downstream distance estimation. We demonstrate that ShrinkBox can be realized in the YOLOv9m object detector at an Attack Success Rate (ASR) of 96%, with only a 4% poisoning ratio in the training instances of the KITTI dataset. Furthermore, given the low error targets introduced in our relaxed poisoning strategy, we find that ShrinkBox increases the Mean Absolute Error (MAE) in downstream distance estimation by more than 3x on poisoned samples, potentially resulting in delays or prevention of collision warnings altogether.", "date": "2025-07-22", "categories": []}
{"id": "2507.18870", "title": "Transferable and Undefendable Point Cloud Attacks via Medial Axis Transform", "abstract": "Studying adversarial attacks on point clouds is essential for evaluating and improving the robustness of 3D deep learning models. However, most existing attack methods are developed under ideal white-box settings and often suffer from limited transferability to unseen models and insufficient robustness against common defense mechanisms. In this paper, we propose MAT-Adv, a novel adversarial attack framework that enhances both transferability and undefendability by explicitly perturbing the medial axis transform (MAT) representations, in order to induce inherent adversarialness in the resulting point clouds. Specifically, we employ an autoencoder to project input point clouds into compact MAT representations that capture the intrinsic geometric structure of point clouds. By perturbing these intrinsic representations, MAT-Adv introduces structural-level adversarial characteristics that remain effective across diverse models and defense strategies. To mitigate overfitting and prevent perturbation collapse, we incorporate a dropout strategy into the optimization of MAT perturbations, further improving transferability and undefendability. Extensive experiments demonstrate that MAT-Adv significantly outperforms existing state-of-the-art methods in both transferability and undefendability. Codes will be made public upon paper acceptance.", "date": "2025-07-25", "categories": []}
{"id": "2309.03791", "title": "Optimal Transport Regularized Divergences: Application to Adversarial Robustness", "abstract": "We introduce a new class of optimal-transport-regularized divergences, $D^c$, constructed via an infimal convolution between an information divergence, $D$, and an optimal-transport (OT) cost, $C$, and study their use in distributionally robust optimization (DRO). In particular, we propose the $ARMOR_D$ methods as novel approaches to enhancing the adversarial robustness of deep learning models. These DRO-based methods are defined by minimizing the maximum expected loss over a $D^c$-neighborhood of the empirical distribution of the training data. Viewed as a tool for constructing adversarial samples, our method allows samples to be both transported, according to the OT cost, and re-weighted, according to the information divergence; the addition of a principled and dynamical adversarial re-weighting on top of adversarial sample transport is a key innovation of $ARMOR_D$. $ARMOR_D$ can be viewed as a generalization of the best-performing loss functions and OT costs in the adversarial training literature; we demonstrate this flexibility by using $ARMOR_D$ to augment the UDR, TRADES, and MART methods and obtain improved performance on CIFAR-10 and CIFAR-100 image recognition. Specifically, augmenting with $ARMOR_D$ leads to 1.9\\% and 2.1\\% improvement against AutoAttack, a powerful ensemble of adversarial attacks, on CIFAR-10 and CIFAR-100 respectively. To foster reproducibility, we made the code accessible at https://github.com/star-ailab/ARMOR.", "date": "2023-09-07", "categories": []}
{"id": "2503.17724", "title": "Trigger without Trace: Towards Stealthy Backdoor Attack on Text-to-Image Diffusion Models", "abstract": "Backdoor attacks targeting text-to-image diffusion models have advanced rapidly. However, current backdoor samples often exhibit two key abnormalities compared to benign samples: 1) Semantic Consistency, where backdoor prompts tend to generate images with similar semantic content even with significant textual variations to the prompts; 2) Attention Consistency, where the trigger induces consistent structural responses in the cross-attention maps. These consistencies leave detectable traces for defenders, making backdoors easier to identify. In this paper, toward stealthy backdoor samples, we propose Trigger without Trace (TwT) by explicitly mitigating these consistencies. Specifically, our approach leverages syntactic structures as backdoor triggers to amplify the sensitivity to textual variations, effectively breaking down the semantic consistency. Besides, a regularization method based on Kernel Maximum Mean Discrepancy (KMMD) is proposed to align the distribution of cross-attention responses between backdoor and benign samples, thereby disrupting attention consistency. Extensive experiments demonstrate that our method achieves a 97.5% attack success rate while exhibiting stronger resistance to defenses. It achieves an average of over 98% backdoor samples bypassing three state-of-the-art detection mechanisms, revealing the vulnerabilities of current backdoor defense methods. The code is available at https://github.com/Robin-WZQ/TwT.", "date": "2025-03-22", "categories": []}
{"id": "2507.17922", "title": "From Seed to Harvest: Augmenting Human Creativity with AI for Red-teaming Text-to-Image Models", "abstract": "Text-to-image (T2I) models have become prevalent across numerous applications, making their robust evaluation against adversarial attacks a critical priority. Continuous access to new and challenging adversarial prompts across diverse domains is essential for stress-testing these models for resilience against novel attacks from multiple vectors. Current techniques for generating such prompts are either entirely authored by humans or synthetically generated. On the one hand, datasets of human-crafted adversarial prompts are often too small in size and imbalanced in their cultural and contextual representation. On the other hand, datasets of synthetically-generated prompts achieve scale, but typically lack the realistic nuances and creative adversarial strategies found in human-crafted prompts. To combine the strengths of both human and machine approaches, we propose Seed2Harvest, a hybrid red-teaming method for guided expansion of culturally diverse, human-crafted adversarial prompt seeds. The resulting prompts preserve the characteristics and attack patterns of human prompts while maintaining comparable average attack success rates (0.31 NudeNet, 0.36 SD NSFW, 0.12 Q16). Our expanded dataset achieves substantially higher diversity with 535 unique geographic locations and a Shannon entropy of 7.48, compared to 58 locations and 5.28 entropy in the original dataset. Our work demonstrates the importance of human-machine collaboration in leveraging human creativity and machine computational capacity to achieve comprehensive, scalable red-teaming for continuous T2I model safety evaluation.", "date": "2025-07-23", "categories": []}
{"id": "2507.18113", "title": "Policy Disruption in Reinforcement Learning:Adversarial Attack with Large Language Models and Critical State Identification", "abstract": "Reinforcement learning (RL) has achieved remarkable success in fields like robotics and autonomous driving, but adversarial attacks designed to mislead RL systems remain challenging. Existing approaches often rely on modifying the environment or policy, limiting their practicality. This paper proposes an adversarial attack method in which existing agents in the environment guide the target policy to output suboptimal actions without altering the environment. We propose a reward iteration optimization framework that leverages large language models (LLMs) to generate adversarial rewards explicitly tailored to the vulnerabilities of the target agent, thereby enhancing the effectiveness of inducing the target agent toward suboptimal decision-making. Additionally, a critical state identification algorithm is designed to pinpoint the target agent's most vulnerable states, where suboptimal behavior from the victim leads to significant degradation in overall performance. Experimental results in diverse environments demonstrate the superiority of our method over existing approaches.", "date": "2025-07-24", "categories": []}
{"id": "2507.18305", "title": "BadReasoner: Planting Tunable Overthinking Backdoors into Large Reasoning Models for Fun or Profit", "abstract": "Large reasoning models (LRMs) have emerged as a significant advancement in artificial intelligence, representing a specialized class of large language models (LLMs) designed to tackle complex reasoning tasks. The defining characteristic of LRMs lies in their extensive chain-of-thought (CoT) reasoning capabilities. In this paper, we identify a previously unexplored attack vector against LRMs, which we term \"overthinking backdoors\". We advance this concept by proposing a novel tunable backdoor, which moves beyond simple on/off attacks to one where an attacker can precisely control the extent of the model's reasoning verbosity. Our attack is implemented through a novel data poisoning methodology. It pairs a tunable trigger-where the number of repetitions signals the desired intensity-with a correspondingly verbose CoT response. These responses are programmatically generated by instructing a teacher LLM to inject a controlled number of redundant refinement steps into a correct reasoning process. The approach preserves output correctness, which ensures stealth and establishes the attack as a pure resource-consumption vector. Extensive empirical results on various LRMs demonstrate that our method can reliably trigger a controllable, multi-fold increase in the length of the reasoning process, without degrading the final answer's correctness. Our source code is available at https://github.com/FZaKK/BadReasoner.", "date": "2025-07-24", "categories": []}
{"id": "2507.18457", "title": "Revisiting Physically Realizable Adversarial Object Attack against LiDAR-based Detection: Clarifying Problem Formulation and Experimental Protocols", "abstract": "Adversarial robustness in LiDAR-based 3D object detection is a critical research area due to its widespread application in real-world scenarios. While many digital attacks manipulate point clouds or meshes, they often lack physical realizability, limiting their practical impact. Physical adversarial object attacks remain underexplored and suffer from poor reproducibility due to inconsistent setups and hardware differences. To address this, we propose a device-agnostic, standardized framework that abstracts key elements of physical adversarial object attacks, supports diverse methods, and provides open-source code with benchmarking protocols in simulation and real-world settings. Our framework enables fair comparison, accelerates research, and is validated by successfully transferring simulated attacks to a physical LiDAR system. Beyond the framework, we offer insights into factors influencing attack success and advance understanding of adversarial robustness in real-world LiDAR perception.", "date": "2025-07-24", "categories": []}
{"id": "2507.18484", "title": "Reinforced Embodied Active Defense: Exploiting Adaptive Interaction for Robust Visual Perception in Adversarial 3D Environments", "abstract": "Adversarial attacks in 3D environments have emerged as a critical threat to the reliability of visual perception systems, particularly in safety-sensitive applications such as identity verification and autonomous driving. These attacks employ adversarial patches and 3D objects to manipulate deep neural network (DNN) predictions by exploiting vulnerabilities within complex scenes. Existing defense mechanisms, such as adversarial training and purification, primarily employ passive strategies to enhance robustness. However, these approaches often rely on pre-defined assumptions about adversarial tactics, limiting their adaptability in dynamic 3D settings. To address these challenges, we introduce Reinforced Embodied Active Defense (Rein-EAD), a proactive defense framework that leverages adaptive exploration and interaction with the environment to improve perception robustness in 3D adversarial contexts. By implementing a multi-step objective that balances immediate prediction accuracy with predictive entropy minimization, Rein-EAD optimizes defense strategies over a multi-step horizon. Additionally, Rein-EAD involves an uncertainty-oriented reward-shaping mechanism that facilitates efficient policy updates, thereby reducing computational overhead and supporting real-world applicability without the need for differentiable environments. Comprehensive experiments validate the effectiveness of Rein-EAD, demonstrating a substantial reduction in attack success rates while preserving standard accuracy across diverse tasks. Notably, Rein-EAD exhibits robust generalization to unseen and adaptive attacks, making it suitable for real-world complex tasks, including 3D object classification, face recognition and autonomous driving.", "date": "2025-07-24", "categories": []}
{"id": "2409.19638", "title": "BadHMP: Backdoor Attack against Human Motion Prediction", "abstract": "Precise future human motion prediction over sub-second horizons from past observations is crucial for various safety-critical applications. To date, only a few studies have examined the vulnerability of skeleton-based neural networks to evasion and backdoor attacks. In this paper, we propose BadHMP, a novel backdoor attack that targets specifically human motion prediction tasks. Our approach involves generating poisoned training samples by embedding a localized backdoor trigger in one limb of the skeleton, causing selected joints to follow predefined motion in historical time steps. Subsequently, the future sequences are globally modified that all the joints move following the target trajectories. Our carefully designed backdoor triggers and targets guarantee the smoothness and naturalness of the poisoned samples, making them stealthy enough to evade detection by the model trainer while keeping the poisoned model unobtrusive in terms of prediction fidelity to untainted sequences. The target sequences can be successfully activated by the designed input sequences even with a low poisoned sample injection ratio. Experimental results on two datasets (Human3.6M and CMU-Mocap) and two network architectures (LTD and HRI) demonstrate the high-fidelity, effectiveness, and stealthiness of BadHMP. Robustness of our attack against fine-tuning defense is also verified.", "date": "2024-09-29", "categories": []}
{"id": "2507.17070", "title": "Advancing Robustness in Deep Reinforcement Learning with an Ensemble Defense Approach", "abstract": "Recent advancements in Deep Reinforcement Learning (DRL) have demonstrated its applicability across various domains, including robotics, healthcare, energy optimization, and autonomous driving. However, a critical question remains: How robust are DRL models when exposed to adversarial attacks? While existing defense mechanisms such as adversarial training and distillation enhance the resilience of DRL models, there remains a significant research gap regarding the integration of multiple defenses in autonomous driving scenarios specifically. This paper addresses this gap by proposing a novel ensemble-based defense architecture to mitigate adversarial attacks in autonomous driving. Our evaluation demonstrates that the proposed architecture significantly enhances the robustness of DRL models. Compared to the baseline under FGSM attacks, our ensemble method improves the mean reward from 5.87 to 18.38 (over 213% increase) and reduces the mean collision rate from 0.50 to 0.09 (an 82% decrease) in the highway scenario and merge scenario, outperforming all standalone defense strategies.", "date": "2025-07-22", "categories": []}
{"id": "2507.17577", "title": "Boosting Ray Search Procedure of Hard-label Attacks with Transfer-based Priors", "abstract": "One of the most practical and challenging types of black-box adversarial attacks is the hard-label attack, where only the top-1 predicted label is available. One effective approach is to search for the optimal ray direction from the benign image that minimizes the $\\ell_p$-norm distance to the adversarial region. The unique advantage of this approach is that it transforms the hard-label attack into a continuous optimization problem. The objective function value is the ray's radius, which can be obtained via binary search at a high query cost. Existing methods use a \"sign trick\" in gradient estimation to reduce the number of queries. In this paper, we theoretically analyze the quality of this gradient estimation and propose a novel prior-guided approach to improve ray search efficiency both theoretically and empirically. Specifically, we utilize the transfer-based priors from surrogate models, and our gradient estimators appropriately integrate them by approximating the projection of the true gradient onto the subspace spanned by these priors and random directions, in a query-efficient manner. We theoretically derive the expected cosine similarities between the obtained gradient estimators and the true gradient, and demonstrate the improvement achieved by incorporating priors. Extensive experiments on the ImageNet and CIFAR-10 datasets show that our approach significantly outperforms 11 state-of-the-art methods in terms of query efficiency.", "date": "2025-07-23", "categories": []}
{"id": "2507.15907", "title": "Dual Turing Test: A Framework for Detecting and Mitigating Undetectable AI", "abstract": "In this short note, we propose a unified framework that bridges three areas: (1) a flipped perspective on the Turing Test, the \"dual Turing test\", in which a human judge's goal is to identify an AI rather than reward a machine for deception; (2) a formal adversarial classification game with explicit quality constraints and worst-case guarantees; and (3) a reinforcement learning (RL) alignment pipeline that uses an undetectability detector and a set of quality related components in its reward model. We review historical precedents, from inverted and meta-Turing variants to modern supervised reverse-Turing classifiers, and highlight the novelty of combining quality thresholds, phased difficulty levels, and minimax bounds. We then formalize the dual test: define the judge's task over N independent rounds with fresh prompts drawn from a prompt space Q, introduce a quality function Q and parameters tau and delta, and cast the interaction as a two-player zero-sum game over the adversary's feasible strategy set M. Next, we map this minimax game onto an RL-HF style alignment loop, in which an undetectability detector D provides negative reward for stealthy outputs, balanced by a quality proxy that preserves fluency. Throughout, we include detailed explanations of each component notation, the meaning of inner minimization over sequences, phased tests, and iterative adversarial training and conclude with a suggestion for a couple of immediate actions.", "date": "2025-07-21", "categories": []}
{"id": "2507.15974", "title": "Does More Inference-Time Compute Really Help Robustness?", "abstract": "Recently, Zaremba et al. demonstrated that increasing inference-time computation improves robustness in large proprietary reasoning LLMs. In this paper, we first show that smaller-scale, open-source models (e.g., DeepSeek R1, Qwen3, Phi-reasoning) can also benefit from inference-time scaling using a simple budget forcing strategy. More importantly, we reveal and critically examine an implicit assumption in prior work: intermediate reasoning steps are hidden from adversaries. By relaxing this assumption, we identify an important security risk, intuitively motivated and empirically verified as an inverse scaling law: if intermediate reasoning steps become explicitly accessible, increased inference-time computation consistently reduces model robustness. Finally, we discuss practical scenarios where models with hidden reasoning chains are still vulnerable to attacks, such as models with tool-integrated reasoning and advanced reasoning extraction attacks. Our findings collectively demonstrate that the robustness benefits of inference-time scaling depend heavily on the adversarial setting and deployment context. We urge practitioners to carefully weigh these subtle trade-offs before applying inference-time scaling in security-sensitive, real-world applications.", "date": "2025-07-21", "categories": []}
{"id": "2507.16052", "title": "Disrupting Semantic and Abstract Features for Better Adversarial Transferability", "abstract": "Adversarial examples pose significant threats to deep neural networks (DNNs), and their property of transferability in the black-box setting has led to the emergence of transfer-based attacks, making it feasible to target real-world applications employing DNNs. Among them, feature-level attacks, where intermediate features are perturbed based on feature importance weight matrix computed from transformed images, have gained popularity. In this work, we find that existing feature-level attacks primarily manipulate the semantic information to derive the weight matrix. Inspired by several works that find CNNs tend to focus more on high-frequency components (a.k.a. abstract features, e.g., texture, edge, etc.), we validate that transforming images in the high-frequency space also improves transferability. Based on this finding, we propose a balanced approach called Semantic and Abstract FEatures disRuption (SAFER). Specifically, SAFER conducts BLOCKMIX on the input image and SELF-MIX on the frequency spectrum when computing the weight matrix to highlight crucial features. By using such a weight matrix, we can direct the attacker to disrupt both semantic and abstract features, leading to improved transferability. Extensive experiments on the ImageNet dataset also demonstrate the effectiveness of our method in boosting adversarial transferability.", "date": "2025-07-21", "categories": []}
{"id": "2507.16164", "title": "Attacking interpretable NLP systems", "abstract": "Studies have shown that machine learning systems are vulnerable to adversarial examples in theory and practice. Where previous attacks have focused mainly on visual models that exploit the difference between human and machine perception, text-based models have also fallen victim to these attacks. However, these attacks often fail to maintain the semantic meaning of the text and similarity. This paper introduces AdvChar, a black-box attack on Interpretable Natural Language Processing Systems, designed to mislead the classifier while keeping the interpretation similar to benign inputs, thus exploiting trust in system transparency. AdvChar achieves this by making less noticeable modifications to text input, forcing the deep learning classifier to make incorrect predictions and preserve the original interpretation. We use an interpretation-focused scoring approach to determine the most critical tokens that, when changed, can cause the classifier to misclassify the input. We apply simple character-level modifications to measure the importance of tokens, minimizing the difference between the original and new text while generating adversarial interpretations similar to benign ones. We thoroughly evaluated AdvChar by testing it against seven NLP models and three interpretation models using benchmark datasets for the classification task. Our experiments show that AdvChar can significantly reduce the prediction accuracy of current deep learning models by altering just two characters on average in input samples.", "date": "2025-07-22", "categories": []}
{"id": "2507.16257", "title": "Quality Text, Robust Vision: The Role of Language in Enhancing Visual Robustness of Vision-Language Models", "abstract": "Defending pre-trained vision-language models (VLMs), such as CLIP, against adversarial attacks is crucial, as these models are widely used in diverse zero-shot tasks, including image classification. However, existing adversarial training (AT) methods for robust fine-tuning largely overlook the role of language in enhancing visual robustness. Specifically, (1) supervised AT methods rely on short texts (e.g., class labels) to generate adversarial perturbations, leading to overfitting to object classes in the training data, and (2) unsupervised AT avoids this overfitting but remains suboptimal against practical text-guided adversarial attacks due to its lack of semantic guidance. To address these limitations, we propose Quality Text-guided Adversarial Fine-Tuning (QT-AFT), which leverages high-quality captions during training to guide adversarial examples away from diverse semantics present in images. This enables the visual encoder to robustly recognize a broader range of image features even under adversarial noise, thereby enhancing robustness across diverse downstream tasks. QT-AFT overcomes the key weaknesses of prior methods -- overfitting in supervised AT and lack of semantic awareness in unsupervised AT -- achieving state-of-the-art zero-shot adversarial robustness and clean accuracy, evaluated across 16 zero-shot datasets. Furthermore, our comprehensive study uncovers several key insights into the role of language in enhancing vision robustness; for example, describing object properties in addition to object names further enhances zero-shot robustness. Our findings point to an urgent direction for future work -- centering high-quality linguistic supervision in robust visual representation learning.", "date": "2025-07-22", "categories": []}
{"id": "2405.20090", "title": "Transfer Attack for Bad and Good: Explain and Boost Adversarial Transferability across Multimodal Large Language Models", "abstract": "Multimodal Large Language Models (MLLMs) demonstrate exceptional performance in cross-modality interaction, yet they also suffer adversarial vulnerabilities. In particular, the transferability of adversarial examples remains an ongoing challenge. In this paper, we specifically analyze the manifestation of adversarial transferability among MLLMs and identify the key factors that influence this characteristic. We discover that the transferability of MLLMs exists in cross-LLM scenarios with the same vision encoder and indicate \\underline{\\textit{two key Factors}} that may influence transferability. We provide two semantic-level data augmentation methods, Adding Image Patch (AIP) and Typography Augment Transferability Method (TATM), which boost the transferability of adversarial examples across MLLMs. To explore the potential impact in the real world, we utilize two tasks that can have both negative and positive societal impacts: \\ding{182} Harmful Content Insertion and \\ding{183} Information Protection.", "date": "2024-05-30", "categories": []}
{"id": "2503.01781", "title": "Cats Confuse Reasoning LLM: Query Agnostic Adversarial Triggers for Reasoning Models", "abstract": "We investigate the robustness of reasoning models trained for step-by-step problem solving by introducing query-agnostic adversarial triggers - short, irrelevant text that, when appended to math problems, systematically mislead models to output incorrect answers without altering the problem's semantics. We propose CatAttack, an automated iterative attack pipeline for generating triggers on a weaker, less expensive proxy model (DeepSeek V3) and successfully transfer them to more advanced reasoning target models like DeepSeek R1 and DeepSeek R1-distilled-Qwen-32B, resulting in greater than 300% increase in the likelihood of the target model generating an incorrect answer. For example, appending, \"Interesting fact: cats sleep most of their lives,\" to any math problem leads to more than doubling the chances of a model getting the answer wrong. Our findings highlight critical vulnerabilities in reasoning models, revealing that even state-of-the-art models remain susceptible to subtle adversarial inputs, raising security and reliability concerns. The CatAttack triggers dataset with model responses is available at https://huggingface.co/datasets/collinear-ai/cat-attack-adversarial-triggers.", "date": "2025-03-03", "categories": []}
{"id": "2507.14248", "title": "Breaking the Illusion of Security via Interpretation: Interpretable Vision Transformer Systems under Attack", "abstract": "Vision transformer (ViT) models, when coupled with interpretation models, are regarded as secure and challenging to deceive, making them well-suited for security-critical domains such as medical applications, autonomous vehicles, drones, and robotics. However, successful attacks on these systems can lead to severe consequences. Recent research on threats targeting ViT models primarily focuses on generating the smallest adversarial perturbations that can deceive the models with high confidence, without considering their impact on model interpretations. Nevertheless, the use of interpretation models can effectively assist in detecting adversarial examples. This study investigates the vulnerability of transformer models to adversarial attacks, even when combined with interpretation models. We propose an attack called \"AdViT\" that generates adversarial examples capable of misleading both a given transformer model and its coupled interpretation model. Through extensive experiments on various transformer models and two transformer-based interpreters, we demonstrate that AdViT achieves a 100% attack success rate in both white-box and black-box scenarios. In white-box scenarios, it reaches up to 98% misclassification confidence, while in black-box scenarios, it reaches up to 76% misclassification confidence. Remarkably, AdViT consistently generates accurate interpretations in both scenarios, making the adversarial examples more difficult to detect.", "date": "2025-07-18", "categories": []}
{"id": "2507.15042", "title": "DeRAG: Black-box Adversarial Attacks on Multiple Retrieval-Augmented Generation Applications via Prompt Injection", "abstract": "Adversarial prompt attacks can significantly alter the reliability of Retrieval-Augmented Generation (RAG) systems by re-ranking them to produce incorrect outputs. In this paper, we present a novel method that applies Differential Evolution (DE) to optimize adversarial prompt suffixes for RAG-based question answering. Our approach is gradient-free, treating the RAG pipeline as a black box and evolving a population of candidate suffixes to maximize the retrieval rank of a targeted incorrect document to be closer to real world scenarios. We conducted experiments on the BEIR QA datasets to evaluate attack success at certain retrieval rank thresholds under multiple retrieving applications. Our results demonstrate that DE-based prompt optimization attains competitive (and in some cases higher) success rates compared to GGPP to dense retrievers and PRADA to sparse retrievers, while using only a small number of tokens (<=5 tokens) in the adversarial suffix. Furthermore, we introduce a readability-aware suffix construction strategy, validated by a statistically significant reduction in MLM negative log-likelihood with Welch's t-test. Through evaluations with a BERT-based adversarial suffix detector, we show that DE-generated suffixes evade detection, yielding near-chance detection accuracy.", "date": "2025-07-20", "categories": []}
{"id": "2507.15067", "title": "ROBAD: Robust Adversary-aware Local-Global Attended Bad Actor Detection Sequential Model", "abstract": "Detecting bad actors is critical to ensure the safety and integrity of internet platforms. Several deep learning-based models have been developed to identify such users. These models should not only accurately detect bad actors, but also be robust against adversarial attacks that aim to evade detection. However, past deep learning-based detection models do not meet the robustness requirement because they are sensitive to even minor changes in the input sequence. To address this issue, we focus on (1) improving the model understanding capability and (2) enhancing the model knowledge such that the model can recognize potential input modifications when making predictions. To achieve these goals, we create a novel transformer-based classification model, called ROBAD (RObust adversary-aware local-global attended Bad Actor Detection model), which uses the sequence of user posts to generate user embedding to detect bad actors. Particularly, ROBAD first leverages the transformer encoder block to encode each post bidirectionally, thus building a post embedding to capture the local information at the post level. Next, it adopts the transformer decoder block to model the sequential pattern in the post embeddings by using the attention mechanism, which generates the sequence embedding to obtain the global information at the sequence level. Finally, to enrich the knowledge of the model, embeddings of modified sequences by mimicked attackers are fed into a contrastive-learning-enhanced classification layer for sequence prediction. In essence, by capturing the local and global information (i.e., the post and sequence information) and leveraging the mimicked behaviors of bad actors in training, ROBAD can be robust to adversarial attacks. Extensive experiments on Yelp and Wikipedia datasets show that ROBAD can effectively detect bad actors when under state-of-the-art adversarial attacks.", "date": "2025-07-20", "categories": []}
{"id": "2504.10888", "title": "CDUPatch: Color-Driven Universal Adversarial Patch Attack for Dual-Modal Visible-Infrared Detectors", "abstract": "Adversarial patches are widely used to evaluate the robustness of object detection systems in real-world scenarios. These patches were initially designed to deceive single-modal detectors (e.g., visible or infrared) and have recently been extended to target visible-infrared dual-modal detectors. However, existing dual-modal adversarial patch attacks have limited attack effectiveness across diverse physical scenarios. To address this, we propose CDUPatch, a universal cross-modal patch attack against visible-infrared object detectors across scales, views, and scenarios. Specifically, we observe that color variations lead to different levels of thermal absorption, resulting in temperature differences in infrared imaging. Leveraging this property, we propose an RGB-to-infrared adapter that maps RGB patches to infrared patches, enabling unified optimization of cross-modal patches. By learning an optimal color distribution on the adversarial patch, we can manipulate its thermal response and generate an adversarial infrared texture. Additionally, we introduce a multi-scale clipping strategy and construct a new visible-infrared dataset, MSDrone, which contains aerial vehicle images in varying scales and perspectives. These data augmentation strategies enhance the robustness of our patch in real-world conditions. Experiments on four benchmark datasets (e.g., DroneVehicle, LLVIP, VisDrone, MSDrone) show that our method outperforms existing patch attacks in the digital domain. Extensive physical tests further confirm strong transferability across scales, views, and scenarios.", "date": "2025-04-15", "categories": []}
{"id": "2507.13474", "title": "Paper Summary Attack: Jailbreaking LLMs through LLM Safety Papers", "abstract": "The safety of large language models (LLMs) has garnered significant research attention. In this paper, we argue that previous empirical studies demonstrate LLMs exhibit a propensity to trust information from authoritative sources, such as academic papers, implying new possible vulnerabilities. To verify this possibility, a preliminary analysis is designed to illustrate our two findings. Based on this insight, a novel jailbreaking method, Paper Summary Attack (\\llmname{PSA}), is proposed. It systematically synthesizes content from either attack-focused or defense-focused LLM safety paper to construct an adversarial prompt template, while strategically infilling harmful query as adversarial payloads within predefined subsections. Extensive experiments show significant vulnerabilities not only in base LLMs, but also in state-of-the-art reasoning model like Deepseek-R1. PSA achieves a 97\\% attack success rate (ASR) on well-aligned models like Claude3.5-Sonnet and an even higher 98\\% ASR on Deepseek-R1. More intriguingly, our work has further revealed diametrically opposed vulnerability bias across different base models, and even between different versions of the same model, when exposed to either attack-focused or defense-focused papers. This phenomenon potentially indicates future research clues for both adversarial methodologies and safety alignment.Code is available at https://github.com/233liang/Paper-Summary-Attack", "date": "2025-07-17", "categories": []}
{"id": "2507.13598", "title": "GIFT: Gradient-aware Immunization of diffusion models against malicious Fine-Tuning with safe concepts retention", "abstract": "We present GIFT: a {G}radient-aware {I}mmunization technique to defend diffusion models against malicious {F}ine-{T}uning while preserving their ability to generate safe content. Existing safety mechanisms like safety checkers are easily bypassed, and concept erasure methods fail under adversarial fine-tuning. GIFT addresses this by framing immunization as a bi-level optimization problem: the upper-level objective degrades the model's ability to represent harmful concepts using representation noising and maximization, while the lower-level objective preserves performance on safe data. GIFT achieves robust resistance to malicious fine-tuning while maintaining safe generative quality. Experimental results show that our method significantly impairs the model's ability to re-learn harmful concepts while maintaining performance on safe content, offering a promising direction for creating inherently safer generative models resistant to adversarial fine-tuning attacks.", "date": "2025-07-18", "categories": []}
{"id": "2507.13727", "title": "Adversarial Training Improves Generalization Under Distribution Shifts in Bioacoustics", "abstract": "Adversarial training is a promising strategy for enhancing model robustness against adversarial attacks. However, its impact on generalization under substantial data distribution shifts in audio classification remains largely unexplored. To address this gap, this work investigates how different adversarial training strategies improve generalization performance and adversarial robustness in audio classification. The study focuses on two model architectures: a conventional convolutional neural network (ConvNeXt) and an inherently interpretable prototype-based model (AudioProtoPNet). The approach is evaluated using a challenging bird sound classification benchmark. This benchmark is characterized by pronounced distribution shifts between training and test data due to varying environmental conditions and recording methods, a common real-world challenge. The investigation explores two adversarial training strategies: one based on output-space attacks that maximize the classification loss function, and another based on embedding-space attacks designed to maximize embedding dissimilarity. These attack types are also used for robustness evaluation. Additionally, for AudioProtoPNet, the study assesses the stability of its learned prototypes under targeted embedding-space attacks. Results show that adversarial training, particularly using output-space attacks, improves clean test data performance by an average of 10.5% relative and simultaneously strengthens the adversarial robustness of the models. These findings, although derived from the bird sound domain, suggest that adversarial training holds potential to enhance robustness against both strong distribution shifts and adversarial attacks in challenging audio classification settings.", "date": "2025-07-18", "categories": []}
{"id": "2408.02946", "title": "Scaling Trends for Data Poisoning in LLMs", "abstract": "LLMs produce harmful and undesirable behavior when trained on datasets containing even a small fraction of poisoned data. We demonstrate that GPT models remain vulnerable to fine-tuning on poisoned data, even when safeguarded by moderation systems. Given the persistence of data poisoning vulnerabilities in today's most capable models, this paper investigates whether these risks increase with model scaling. We evaluate three threat models -- malicious fine-tuning, imperfect data curation, and intentional data contamination -- across 24 frontier LLMs ranging from 1.5 to 72 billion parameters. Our experiments reveal that larger LLMs are significantly more susceptible to data poisoning, learning harmful behaviors from even minimal exposure to harmful data more quickly than smaller models. These findings underscore the need for leading AI companies to thoroughly red team fine-tuning APIs before public release and to develop more robust safeguards against data poisoning, particularly as models continue to scale in size and capability.", "date": "2024-08-06", "categories": []}
{"id": "2412.04245", "title": "Intriguing Properties of Robust Classification", "abstract": "Despite extensive research since the community learned about adversarial examples 10 years ago, we still do not know how to train high-accuracy classifiers that are guaranteed to be robust to small perturbations of their inputs. Previous works often argued that this might be because no classifier exists that is robust and accurate at the same time. However, in computer vision this assumption does not match reality where humans are usually accurate and robust on most tasks of interest. We offer an alternative explanation and show that in certain settings robust generalization is only possible with unrealistically large amounts of data. Specifically, we find a setting where a robust classifier exists, it is easy to learn an accurate classifier, yet it requires an exponential amount of data to learn a robust classifier. Based on this theoretical result, we evaluate the influence of the amount of training data on datasets such as CIFAR-10. Our findings indicate that the amount of training data is the main factor determining the robust performance. Furthermore we show that there are low magnitude directions in the data which are useful for non-robust generalization but are not available for robust classifiers. We provide code at https://github.com/berndprach/IntriguingProperties.", "date": "2024-12-05", "categories": []}
{"id": "2502.19697", "title": "Prompt-driven Transferable Adversarial Attack on Person Re-Identification with Attribute-aware Textual Inversion", "abstract": "Person re-identification (re-id) models are vital in security surveillance systems, requiring transferable adversarial attacks to explore the vulnerabilities of them. Recently, vision-language models (VLM) based attacks have shown superior transferability by attacking generalized image and textual features of VLM, but they lack comprehensive feature disruption due to the overemphasis on discriminative semantics in integral representation. In this paper, we introduce the Attribute-aware Prompt Attack (AP-Attack), a novel method that leverages VLM's image-text alignment capability to explicitly disrupt fine-grained semantic features of pedestrian images by destroying attribute-specific textual embeddings. To obtain personalized textual descriptions for individual attributes, textual inversion networks are designed to map pedestrian images to pseudo tokens that represent semantic embeddings, trained in the contrastive learning manner with images and a predefined prompt template that explicitly describes the pedestrian attributes. Inverted benign and adversarial fine-grained textual semantics facilitate attacker in effectively conducting thorough disruptions, enhancing the transferability of adversarial examples. Extensive experiments show that AP-Attack achieves state-of-the-art transferability, significantly outperforming previous methods by 22.9% on mean Drop Rate in cross-model&dataset attack scenarios.", "date": "2025-02-27", "categories": []}
{"id": "2507.13170", "title": "SHIELD: A Secure and Highly Enhanced Integrated Learning for Robust Deepfake Detection against Adversarial Attacks", "abstract": "Audio plays a crucial role in applications like speaker verification, voice-enabled smart devices, and audio conferencing. However, audio manipulations, such as deepfakes, pose significant risks by enabling the spread of misinformation. Our empirical analysis reveals that existing methods for detecting deepfake audio are often vulnerable to anti-forensic (AF) attacks, particularly those attacked using generative adversarial networks. In this article, we propose a novel collaborative learning method called SHIELD to defend against generative AF attacks. To expose AF signatures, we integrate an auxiliary generative model, called the defense (DF) generative model, which facilitates collaborative learning by combining input and output. Furthermore, we design a triplet model to capture correlations for real and AF attacked audios with real-generated and attacked-generated audios using auxiliary generative models. The proposed SHIELD strengthens the defense against generative AF attacks and achieves robust performance across various generative models. The proposed AF significantly reduces the average detection accuracy from 95.49% to 59.77% for ASVspoof2019, from 99.44% to 38.45% for In-the-Wild, and from 98.41% to 51.18% for HalfTruth for three different generative models. The proposed SHIELD mechanism is robust against AF attacks and achieves an average accuracy of 98.13%, 98.58%, and 99.57% in match, and 98.78%, 98.62%, and 98.85% in mismatch settings for the ASVspoof2019, In-the-Wild, and HalfTruth datasets, respectively.", "date": "2025-07-17", "categories": []}
{"id": "2501.16671", "title": "Data-Free Model-Related Attacks: Unleashing the Potential of Generative AI", "abstract": "Generative AI technology has become increasingly integrated into our daily lives, offering powerful capabilities to enhance productivity. However, these same capabilities can be exploited by adversaries for malicious purposes. While existing research on adversarial applications of generative AI predominantly focuses on cyberattacks, less attention has been given to attacks targeting deep learning models. In this paper, we introduce the use of generative AI for facilitating model-related attacks, including model extraction, membership inference, and model inversion. Our study reveals that adversaries can launch a variety of model-related attacks against both image and text models in a data-free and black-box manner, achieving comparable performance to baseline methods that have access to the target models' training data and parameters in a white-box manner. This research serves as an important early warning to the community about the potential risks associated with generative AI-powered attacks on deep learning models.", "date": "2025-01-28", "categories": []}
{"id": "2502.05208", "title": "Mitigation of Camouflaged Adversarial Attacks in Autonomous Vehicles--A Case Study Using CARLA Simulator", "abstract": "Autonomous vehicles (AVs) rely heavily on cameras and artificial intelligence (AI) to make safe and accurate driving decisions. However, since AI is the core enabling technology, this raises serious cyber threats that hinder the large-scale adoption of AVs. Therefore, it becomes crucial to analyze the resilience of AV security systems against sophisticated attacks that manipulate camera inputs, deceiving AI models. In this paper, we develop camera-camouflaged adversarial attacks targeting traffic sign recognition (TSR) in AVs. Specifically, if the attack is initiated by modifying the texture of a stop sign to fool the AV's object detection system, thereby affecting the AV actuators. The attack's effectiveness is tested using the CARLA AV simulator and the results show that such an attack can delay the auto-braking response to the stop sign, resulting in potential safety issues. We conduct extensive experiments under various conditions, confirming that our new attack is effective and robust. Additionally, we address the attack by presenting mitigation strategies. The proposed attack and defense methods are applicable to other end-to-end trained autonomous cyber-physical systems.", "date": "2025-02-03", "categories": []}
{"id": "2507.11968", "title": "Watch, Listen, Understand, Mislead: Tri-modal Adversarial Attacks on Short Videos for Content Appropriateness Evaluation", "abstract": "Multimodal Large Language Models (MLLMs) are increasingly used for content moderation, yet their robustness in short-form video contexts remains underexplored. Current safety evaluations often rely on unimodal attacks, failing to address combined attack vulnerabilities. In this paper, we introduce a comprehensive framework for evaluating the tri-modal safety of MLLMs. First, we present the Short-Video Multimodal Adversarial (SVMA) dataset, comprising diverse short-form videos with human-guided synthetic adversarial attacks. Second, we propose ChimeraBreak, a novel tri-modal attack strategy that simultaneously challenges visual, auditory, and semantic reasoning pathways. Extensive experiments on state-of-the-art MLLMs reveal significant vulnerabilities with high Attack Success Rates (ASR). Our findings uncover distinct failure modes, showing model biases toward misclassifying benign or policy-violating content. We assess results using LLM-as-a-judge, demonstrating attack reasoning efficacy. Our dataset and findings provide crucial insights for developing more robust and safe MLLMs.", "date": "2025-07-16", "categories": []}
{"id": "2507.12107", "title": "Non-Adaptive Adversarial Face Generation", "abstract": "Adversarial attacks on face recognition systems (FRSs) pose serious security and privacy threats, especially when these systems are used for identity verification. In this paper, we propose a novel method for generating adversarial faces-synthetic facial images that are visually distinct yet recognized as a target identity by the FRS. Unlike iterative optimization-based approaches (e.g., gradient descent or other iterative solvers), our method leverages the structural characteristics of the FRS feature space. We figure out that individuals sharing the same attribute (e.g., gender or race) form an attributed subsphere. By utilizing such subspheres, our method achieves both non-adaptiveness and a remarkably small number of queries. This eliminates the need for relying on transferability and open-source surrogate models, which have been a typical strategy when repeated adaptive queries to commercial FRSs are impossible. Despite requiring only a single non-adaptive query consisting of 100 face images, our method achieves a high success rate of over 93% against AWS's CompareFaces API at its default threshold. Furthermore, unlike many existing attacks that perturb a given image, our method can deliberately produce adversarial faces that impersonate the target identity while exhibiting high-level attributes chosen by the adversary.", "date": "2025-07-16", "categories": []}
{"id": "2403.07095", "title": "Gaussian Loss Smoothing Enables Certified Training with Tight Convex Relaxations", "abstract": "Training neural networks with high certified accuracy against adversarial examples remains an open challenge despite significant efforts. While certification methods can effectively leverage tight convex relaxations for bound computation, in training, these methods, perhaps surprisingly, can perform worse than looser relaxations. Prior work hypothesized that this phenomenon is caused by the discontinuity, non-smoothness, and perturbation sensitivity of the loss surface induced by tighter relaxations. In this work, we theoretically show that applying Gaussian Loss Smoothing (GLS) on the loss surface can alleviate these issues. We confirm this empirically by instantiating GLS with two variants: a zeroth-order optimization algorithm, called PGPE, which allows training with non-differentiable relaxations, and a first-order optimization algorithm, called RGS, which requires gradients of the relaxation but is much more efficient than PGPE. Extensive experiments show that when combined with tight relaxations, these methods surpass state-of-the-art methods when training on the same network architecture for many settings. Our results clearly demonstrate the promise of Gaussian Loss Smoothing for training certifiably robust neural networks and pave a path towards leveraging tighter relaxations for certified training.", "date": "2024-03-11", "categories": []}
{"id": "2507.04762", "title": "Robustifying 3D Perception via Least-Squares Graphs for Multi-Agent Object Tracking", "abstract": "The critical perception capabilities of EdgeAI systems, such as autonomous vehicles, are required to be resilient against adversarial threats, by enabling accurate identification and localization of multiple objects in the scene over time, mitigating their impact. Single-agent tracking offers resilience to adversarial attacks but lacks situational awareness, underscoring the need for multi-agent cooperation to enhance context understanding and robustness. This paper proposes a novel mitigation framework on 3D LiDAR scene against adversarial noise by tracking objects based on least-squares graph on multi-agent adversarial bounding boxes. Specifically, we employ the least-squares graph tool to reduce the induced positional error of each detection's centroid utilizing overlapped bounding boxes on a fully connected graph via differential coordinates and anchor points. Hence, the multi-vehicle detections are fused and refined mitigating the adversarial impact, and associated with existing tracks in two stages performing tracking to further suppress the adversarial threat. An extensive evaluation study on the real-world V2V4Real dataset demonstrates that the proposed method significantly outperforms both state-of-the-art single and multi-agent tracking frameworks by up to 23.3% under challenging adversarial conditions, operating as a resilient approach without relying on additional defense mechanisms.", "date": "2025-07-07", "categories": []}
{"id": "2507.11173", "title": "Real-Time Bayesian Detection of Drift-Evasive GNSS Spoofing in Reinforcement Learning Based UAV Deconfliction", "abstract": "Autonomous unmanned aerial vehicles (UAVs) rely on global navigation satellite system (GNSS) pseudorange measurements for accurate real-time localization and navigation. However, this dependence exposes them to sophisticated spoofing threats, where adversaries manipulate pseudoranges to deceive UAV receivers. Among these, drift-evasive spoofing attacks subtly perturb measurements, gradually diverting the UAVs trajectory without triggering conventional signal-level anti-spoofing mechanisms. Traditional distributional shift detection techniques often require accumulating a threshold number of samples, causing delays that impede rapid detection and timely response. Consequently, robust temporal-scale detection methods are essential to identify attack onset and enable contingency planning with alternative sensing modalities, improving resilience against stealthy adversarial manipulations. This study explores a Bayesian online change point detection (BOCPD) approach that monitors temporal shifts in value estimates from a reinforcement learning (RL) critic network to detect subtle behavioural deviations in UAV navigation. Experimental results show that this temporal value-based framework outperforms conventional GNSS spoofing detectors, temporal semi-supervised learning frameworks, and the Page-Hinkley test, achieving higher detection accuracy and lower false-positive and false-negative rates for drift-evasive spoofing attacks.", "date": "2025-07-15", "categories": []}
{"id": "2401.17196", "title": "Single Word Change is All You Need: Using LLMs to Create Synthetic Training Examples for Text Classifiers", "abstract": "In text classification, creating an adversarial example means subtly perturbing a few words in a sentence without changing its meaning, causing it to be misclassified by a classifier. A concerning observation is that a significant portion of adversarial examples generated by existing methods change only one word. This single-word perturbation vulnerability represents a significant weakness in classifiers, which malicious users can exploit to efficiently create a multitude of adversarial examples. This paper studies this problem and makes the following key contributions: (1) We introduce a novel metric $\\rho$ to quantitatively assess a classifier's robustness against single-word perturbation. (2) We present the SP-Attack, designed to exploit the single-word perturbation vulnerability, achieving a higher attack success rate, better preserving sentence meaning, while reducing computation costs compared to state-of-the-art adversarial methods. (3) We propose SP-Defense, which aims to improve \\r{ho} by applying data augmentation in learning. Experimental results on 4 datasets and BERT and distilBERT classifiers show that SP-Defense improves $\\rho$ by 14.6% and 13.9% and decreases the attack success rate of SP-Attack by 30.4% and 21.2% on two classifiers respectively, and decreases the attack success rate of existing attack methods that involve multiple-word perturbations.", "date": "2024-01-30", "categories": []}
{"id": "2502.12377", "title": "Alignment and Adversarial Robustness: Are More Human-Like Models More Secure?", "abstract": "A small but growing body of work has shown that machine learning models which better align with human vision have also exhibited higher robustness to adversarial examples, raising the question: can human-like perception make models more secure? If true generally, such mechanisms would offer new avenues toward robustness. In this work, we conduct a large-scale empirical analysis to systematically investigate the relationship between representational alignment and adversarial robustness. We evaluate 114 models spanning diverse architectures and training paradigms, measuring their neural and behavioral alignment and engineering task performance across 105 benchmarks as well as their adversarial robustness via AutoAttack. Our findings reveal that while average alignment and robustness exhibit a weak overall correlation, specific alignment benchmarks serve as strong predictors of adversarial robustness, particularly those that measure selectivity toward texture or shape. These results suggest that different forms of alignment play distinct roles in model robustness, motivating further investigation into how alignment-driven approaches can be leveraged to build more secure and perceptually-grounded vision models.", "date": "2025-02-17", "categories": []}
{"id": "2507.10048", "title": "On the Efficiency of Training Robust Decision Trees", "abstract": "As machine learning gets adopted into the industry quickly, trustworthiness is increasingly in focus. Yet, efficiency and sustainability of robust training pipelines still have to be established. In this work, we consider a simple pipeline for training adversarially robust decision trees and investigate the efficiency of each step. Our pipeline consists of three stages. Firstly, we choose the perturbation size automatically for each dataset. For that, we introduce a simple algorithm, instead of relying on intuition or prior work. Moreover, we show that the perturbation size can be estimated from smaller models than the one intended for full training, and thus significant gains in efficiency can be achieved. Secondly, we train state-of-the-art adversarial training methods and evaluate them regarding both their training time and adversarial accuracy. Thirdly, we certify the robustness of each of the models thus obtained and investigate the time required for this. We find that verification time, which is critical to the efficiency of the full pipeline, is not correlated with training time.", "date": "2025-07-14", "categories": []}
{"id": "2507.08020", "title": "Circumventing Safety Alignment in Large Language Models Through Embedding Space Toxicity Attenuation", "abstract": "Large Language Models (LLMs) have achieved remarkable success across domains such as healthcare, education, and cybersecurity. However, this openness also introduces significant security risks, particularly through embedding space poisoning, which is a subtle attack vector where adversaries manipulate the internal semantic representations of input data to bypass safety alignment mechanisms. While previous research has investigated universal perturbation methods, the dynamics of LLM safety alignment at the embedding level remain insufficiently understood. Consequently, more targeted and accurate adversarial perturbation techniques, which pose significant threats, have not been adequately studied.\n  In this work, we propose ETTA (Embedding Transformation Toxicity Attenuation), a novel framework that identifies and attenuates toxicity-sensitive dimensions in embedding space via linear transformations. ETTA bypasses model refusal behaviors while preserving linguistic coherence, without requiring model fine-tuning or access to training data. Evaluated on five representative open-source LLMs using the AdvBench benchmark, ETTA achieves a high average attack success rate of 88.61%, outperforming the best baseline by 11.34%, and generalizes to safety-enhanced models (e.g., 77.39% ASR on instruction-tuned defenses). These results highlight a critical vulnerability in current alignment strategies and underscore the need for embedding-aware defenses.", "date": "2025-07-08", "categories": []}
{"id": "2507.08163", "title": "Adaptive Diffusion Denoised Smoothing : Certified Robustness via Randomized Smoothing with Differentially Private Guided Denoising Diffusion", "abstract": "We propose Adaptive Diffusion Denoised Smoothing, a method for certifying the predictions of a vision model against adversarial examples, while adapting to the input. Our key insight is to reinterpret a guided denoising diffusion model as a long sequence of adaptive Gaussian Differentially Private (GDP) mechanisms refining a pure noise sample into an image. We show that these adaptive mechanisms can be composed through a GDP privacy filter to analyze the end-to-end robustness of the guided denoising process, yielding a provable certification that extends the adaptive randomized smoothing analysis. We demonstrate that our design, under a specific guiding strategy, can improve both certified accuracy and standard accuracy on ImageNet for an $\\ell_2$ threat model.", "date": "2025-07-10", "categories": []}
{"id": "2507.08212", "title": "EvA: Evolutionary Attacks on Graphs", "abstract": "Even a slight perturbation in the graph structure can cause a significant drop in the accuracy of graph neural networks (GNNs). Most existing attacks leverage gradient information to perturb edges. This relaxes the attack's optimization problem from a discrete to a continuous space, resulting in solutions far from optimal. It also restricts the adaptability of the attack to non-differentiable objectives. Instead, we introduce a few simple yet effective enhancements of an evolutionary-based algorithm to solve the discrete optimization problem directly. Our Evolutionary Attack (EvA) works with any black-box model and objective, eliminating the need for a differentiable proxy loss. This allows us to design two novel attacks that reduce the effectiveness of robustness certificates and break conformal sets. The memory complexity of our attack is linear in the attack budget. Among our experiments, EvA shows $\\sim$11\\% additional drop in accuracy on average compared to the best previous attack, revealing significant untapped potential in designing attacks.", "date": "2025-07-10", "categories": []}
{"id": "2305.13651", "title": "Adversarial Defenses via Vector Quantization", "abstract": "Adversarial attacks pose significant challenges to the robustness of modern deep neural networks in computer vision, and defending these networks against adversarial attacks has attracted intense research efforts. Among various defense strategies, preprocessing-based defenses are practically appealing since there is no need to train the network under protection. However, such approaches typically do not achieve comparable robustness as other methods such as adversarial training. In this paper, we propose a novel framework for preprocessing-based defenses, where a vector quantizer is used as a preprocessor. This framework, inspired by and extended from Randomized Discretization (RandDisc), is theoretically principled by rate-distortion theory: indeed, RandDisc may be viewed as a scalar quantizer, and rate-distortion theory suggests that such quantization schemes are inferior to vector quantization. In our framework, the preprocessing vector quantizer treats the input image as a collection of patches and finds a set of representative patches based on the patch distributions; each original patch is then modified according to the representative patches close to it. We present two lightweight defenses in this framework, referred to as patched RandDisc (pRD) and sliding-window RandDisc (swRD), where the patches are disjoint in the former and overlapping in the latter. We show that vector-quantization-based defenses have certifiable robust accuracy and that pRD and swRD demonstrate state-of-the-art performances, surpassing RandDisc by a large margin. Notably, the proposed defenses possess the obfuscated gradients property. Our experiments however show that pRD and swRD remain effective under the STE and EOT attacks, which are designed specifically for defenses with gradient obfuscation. ...", "date": "2023-05-23", "categories": []}
{"id": "2406.10427", "title": "Adaptive Randomized Smoothing: Certified Adversarial Robustness for Multi-Step Defences", "abstract": "We propose Adaptive Randomized Smoothing (ARS) to certify the predictions of our test-time adaptive models against adversarial examples. ARS extends the analysis of randomized smoothing using $f$-Differential Privacy to certify the adaptive composition of multiple steps. For the first time, our theory covers the sound adaptive composition of general and high-dimensional functions of noisy inputs. We instantiate ARS on deep image classification to certify predictions against adversarial examples of bounded $L_{\\infty}$ norm. In the $L_{\\infty}$ threat model, ARS enables flexible adaptation through high-dimensional input-dependent masking. We design adaptivity benchmarks, based on CIFAR-10 and CelebA, and show that ARS improves standard test accuracy by $1$ to $15\\%$ points. On ImageNet, ARS improves certified test accuracy by up to $1.6\\%$ points over standard RS without adaptivity. Our code is available at https://github.com/ubc-systopia/adaptive-randomized-smoothing .", "date": "2024-06-14", "categories": []}
{"id": "2502.00718", "title": "\"I am bad\": Interpreting Stealthy, Universal and Robust Audio Jailbreaks in Audio-Language Models", "abstract": "The rise of multimodal large language models has introduced innovative human-machine interaction paradigms but also significant challenges in machine learning safety. Audio-Language Models (ALMs) are especially relevant due to the intuitive nature of spoken communication, yet little is known about their failure modes. This paper explores audio jailbreaks targeting ALMs, focusing on their ability to bypass alignment mechanisms. We construct adversarial perturbations that generalize across prompts, tasks, and even base audio samples, demonstrating the first universal jailbreaks in the audio modality, and show that these remain effective in simulated real-world conditions. Beyond demonstrating attack feasibility, we analyze how ALMs interpret these audio adversarial examples and reveal them to encode imperceptible first-person toxic speech - suggesting that the most effective perturbations for eliciting toxic outputs specifically embed linguistic features within the audio signal. These results have important implications for understanding the interactions between different modalities in multimodal models, and offer actionable insights for enhancing defenses against adversarial audio attacks.", "date": "2025-02-02", "categories": []}
{"id": "2504.20310", "title": "A Cryptographic Perspective on Mitigation vs. Detection in Machine Learning", "abstract": "In this paper, we initiate a cryptographically inspired theoretical study of detection versus mitigation of adversarial inputs produced by attackers on Machine Learning algorithms during inference time.\n  We formally define defense by detection (DbD) and defense by mitigation (DbM). Our definitions come in the form of a 3-round protocol between two resource-bounded parties: a trainer/defender and an attacker. The attacker aims to produce inference-time inputs that fool the training algorithm. We define correctness, completeness, and soundness properties to capture successful defense at inference time while not degrading (too much) the performance of the algorithm on inputs from the training distribution.\n  We first show that achieving DbD and achieving DbM are equivalent for ML classification tasks. Surprisingly, this is not the case for ML generative learning tasks, where there are many possible correct outputs for each input. We show a separation between DbD and DbM by exhibiting two generative learning tasks for which it is possible to defend by mitigation but it is provably impossible to defend by detection. The mitigation phase uses significantly less computational resources than the initial training algorithm. In the first learning task we consider sample complexity as the resource and in the second the time complexity. The first result holds under the assumption that the Identity-Based Fully Homomorphic Encryption (IB-FHE), publicly-verifiable zero-knowledge Succinct Non-Interactive Arguments of Knowledge (zk-SNARK), and Strongly Unforgeable Signatures exist. The second result assumes the existence of Non-Parallelizing Languages with Average-Case Hardness (NPL) and Incrementally-Verifiable Computation (IVC) and IB-FHE.", "date": "2025-04-28", "categories": []}
{"id": "2507.07259", "title": "Exploiting Edge Features for Transferable Adversarial Attacks in Distributed Machine Learning", "abstract": "As machine learning models become increasingly deployed across the edge of internet of things environments, a partitioned deep learning paradigm in which models are split across multiple computational nodes introduces a new dimension of security risk. Unlike traditional inference setups, these distributed pipelines span the model computation across heterogeneous nodes and communication layers, thereby exposing a broader attack surface to potential adversaries. Building on these motivations, this work explores a previously overlooked vulnerability: even when both the edge and cloud components of the model are inaccessible (i.e., black-box), an adversary who intercepts the intermediate features transmitted between them can still pose a serious threat. We demonstrate that, under these mild and realistic assumptions, an attacker can craft highly transferable proxy models, making the entire deep learning system significantly more vulnerable to evasion attacks. In particular, the intercepted features can be effectively analyzed and leveraged to distill surrogate models capable of crafting highly transferable adversarial examples against the target model. To this end, we propose an exploitation strategy specifically designed for distributed settings, which involves reconstructing the original tensor shape from vectorized transmitted features using simple statistical analysis, and adapting surrogate architectures accordingly to enable effective feature distillation. A comprehensive and systematic experimental evaluation has been conducted to demonstrate that surrogate models trained with the proposed strategy, i.e., leveraging intermediate features, tremendously improve the transferability of adversarial attacks. These findings underscore the urgent need to account for intermediate feature leakage in the design of secure distributed deep learning systems.", "date": "2025-07-09", "categories": []}
{"id": "2507.07709", "title": "One Object, Multiple Lies: A Benchmark for Cross-task Adversarial Attack on Unified Vision-Language Models", "abstract": "Unified vision-language models(VLMs) have recently shown remarkable progress, enabling a single model to flexibly address diverse tasks through different instructions within a shared computational architecture. This instruction-based control mechanism creates unique security challenges, as adversarial inputs must remain effective across multiple task instructions that may be unpredictably applied to process the same malicious content. In this paper, we introduce CrossVLAD, a new benchmark dataset carefully curated from MSCOCO with GPT-4-assisted annotations for systematically evaluating cross-task adversarial attacks on unified VLMs. CrossVLAD centers on the object-change objective-consistently manipulating a target object's classification across four downstream tasks-and proposes a novel success rate metric that measures simultaneous misclassification across all tasks, providing a rigorous evaluation of adversarial transferability. To tackle this challenge, we present CRAFT (Cross-task Region-based Attack Framework with Token-alignment), an efficient region-centric attack method. Extensive experiments on Florence-2 and other popular unified VLMs demonstrate that our method outperforms existing approaches in both overall cross-task attack performance and targeted object-change success rates, highlighting its effectiveness in adversarially influencing unified VLMs across diverse tasks.", "date": "2025-07-10", "categories": []}
{"id": "2507.07768", "title": "TRIX- Trading Adversarial Fairness via Mixed Adversarial Training", "abstract": "Adversarial Training (AT) is a widely adopted defense against adversarial examples. However, existing approaches typically apply a uniform training objective across all classes, overlooking disparities in class-wise vulnerability. This results in adversarial unfairness: classes with well distinguishable features (strong classes) tend to become more robust, while classes with overlapping or shared features(weak classes) remain disproportionately susceptible to adversarial attacks. We observe that strong classes do not require strong adversaries during training, as their non-robust features are quickly suppressed. In contrast, weak classes benefit from stronger adversaries to effectively reduce their vulnerabilities. Motivated by this, we introduce TRIX, a feature-aware adversarial training framework that adaptively assigns weaker targeted adversaries to strong classes, promoting feature diversity via uniformly sampled targets, and stronger untargeted adversaries to weak classes, enhancing their focused robustness. TRIX further incorporates per-class loss weighting and perturbation strength adjustments, building on prior work, to emphasize weak classes during the optimization. Comprehensive experiments on standard image classification benchmarks, including evaluations under strong attacks such as PGD and AutoAttack, demonstrate that TRIX significantly improves worst-case class accuracy on both clean and adversarial data, reducing inter-class robustness disparities, and preserves overall accuracy. Our results highlight TRIX as a practical step toward fair and effective adversarial defense.", "date": "2025-07-10", "categories": []}
{"id": "2507.07773", "title": "Rainbow Artifacts from Electromagnetic Signal Injection Attacks on Image Sensors", "abstract": "Image sensors are integral to a wide range of safety- and security-critical systems, including surveillance infrastructure, autonomous vehicles, and industrial automation. These systems rely on the integrity of visual data to make decisions. In this work, we investigate a novel class of electromagnetic signal injection attacks that target the analog domain of image sensors, allowing adversaries to manipulate raw visual inputs without triggering conventional digital integrity checks. We uncover a previously undocumented attack phenomenon on CMOS image sensors: rainbow-like color artifacts induced in images captured by image sensors through carefully tuned electromagnetic interference. We further evaluate the impact of these attacks on state-of-the-art object detection models, showing that the injected artifacts propagate through the image signal processing pipeline and lead to significant mispredictions. Our findings highlight a critical and underexplored vulnerability in the visual perception stack, highlighting the need for more robust defenses against physical-layer attacks in such systems.", "date": "2025-07-10", "categories": []}
{"id": "2308.04964", "title": "ModSec-AdvLearn: Countering Adversarial SQL Injections with Robust Machine Learning", "abstract": "Many Web Application Firewalls (WAFs) leverage the OWASP CRS to block incoming malicious requests. The CRS consists of different sets of rules designed by domain experts to detect well-known web attack patterns. Both the set of rules and the weights used to combine them are manually defined, yielding four different default configurations of the CRS. In this work, we focus on the detection of SQLi attacks, and show that the manual configurations of the CRS typically yield a suboptimal trade-off between detection and false alarm rates. Furthermore, we show that these configurations are not robust to adversarial SQLi attacks, i.e., carefully-crafted attacks that iteratively refine the malicious SQLi payload by querying the target WAF to bypass detection. To overcome these limitations, we propose (i) using machine learning to automate the selection of the set of rules to be combined along with their weights, i.e., customizing the CRS configuration based on the monitored web services; and (ii) leveraging adversarial training to significantly improve its robustness to adversarial SQLi manipulations. Our experiments, conducted using the well-known open-source ModSecurity WAF equipped with the CRS rules, show that our approach, named ModSec-AdvLearn, can (i) increase the detection rate up to 30%, while retaining negligible false alarm rates and discarding up to 50% of the CRS rules; and (ii) improve robustness against adversarial SQLi attacks up to 85%, marking a significant stride toward designing more effective and robust WAFs. We release our open-source code at https://github.com/pralab/modsec-advlearn.", "date": "2023-08-09", "categories": []}
{"id": "2408.06079", "title": "Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment", "abstract": "Despite the remarkable progress of deep neural networks (DNNs) in various visual tasks, their vulnerability to adversarial examples raises significant security concerns. Recent adversarial training methods leverage inverse adversarial attacks to generate high-confidence examples, aiming to align adversarial distributions with high-confidence class regions. However, our investigation reveals that under inverse adversarial attacks, high-confidence outputs are influenced by biased feature activations, causing models to rely on background features that lack a causal relationship with the labels. This spurious correlation bias leads to overfitting irrelevant background features during adversarial training, thereby degrading the model's robust performance and generalization capabilities. To address this issue, we propose Debiased High-Confidence Adversarial Training (DHAT), a novel approach that aligns adversarial logits with debiased high-confidence logits and restores proper attention by enhancing foreground logit orthogonality. Extensive experiments demonstrate that DHAT achieves state-of-the-art robustness on both CIFAR and ImageNet-1K benchmarks, while significantly improving generalization by mitigating the feature bias inherent in inverse adversarial training approaches. Code is available at https://github.com/KejiaZhang-Robust/DHAT.", "date": "2024-08-12", "categories": []}
{"id": "2503.16248", "title": "Real AI Agents with Fake Memories: Fatal Context Manipulation Attacks on Web3 Agents", "abstract": "AI agents integrated with Web3 offer autonomy and openness but raise security concerns as they interact with financial protocols and immutable smart contracts. This paper investigates the vulnerabilities of AI agents within blockchain-based financial ecosystems when exposed to adversarial threats in real-world scenarios. We introduce the concept of context manipulation -- a comprehensive attack vector that exploits unprotected context surfaces, including input channels, memory modules, and external data feeds. It expands on traditional prompt injection and reveals a more stealthy and persistent threat: memory injection. Using ElizaOS, a representative decentralized AI agent framework for automated Web3 operations, we showcase that malicious injections into prompts or historical records can trigger unauthorized asset transfers and protocol violations which could be financially devastating in reality. To quantify these risks, we introduce CrAIBench, a Web3-focused benchmark covering 150+ realistic blockchain tasks. such as token transfers, trading, bridges, and cross-chain interactions, and 500+ attack test cases using context manipulation. Our evaluation results confirm that AI models are significantly more vulnerable to memory injection compared to prompt injection. Finally, we evaluate a comprehensive defense roadmap, finding that prompt-injection defenses and detectors only provide limited protection when stored context is corrupted, whereas fine-tuning-based defenses substantially reduce attack success rates while preserving performance on single-step tasks. These results underscore the urgent need for AI agents that are both secure and fiduciarily responsible in blockchain environments.", "date": "2025-03-20", "categories": []}
{"id": "2506.23581", "title": "PBCAT: Patch-based composite adversarial training against physically realizable attacks on object detection", "abstract": "Object detection plays a crucial role in many security-sensitive applications. However, several recent studies have shown that object detectors can be easily fooled by physically realizable attacks, \\eg, adversarial patches and recent adversarial textures, which pose realistic and urgent threats. Adversarial Training (AT) has been recognized as the most effective defense against adversarial attacks. While AT has been extensively studied in the $l_\\infty$ attack settings on classification models, AT against physically realizable attacks on object detectors has received limited exploration. Early attempts are only performed to defend against adversarial patches, leaving AT against a wider range of physically realizable attacks under-explored. In this work, we consider defending against various physically realizable attacks with a unified AT method. We propose PBCAT, a novel Patch-Based Composite Adversarial Training strategy. PBCAT optimizes the model by incorporating the combination of small-area gradient-guided adversarial patches and imperceptible global adversarial perturbations covering the entire image. With these designs, PBCAT has the potential to defend against not only adversarial patches but also unseen physically realizable attacks such as adversarial textures. Extensive experiments in multiple settings demonstrated that PBCAT significantly improved robustness against various physically realizable attacks over state-of-the-art defense methods. Notably, it improved the detection accuracy by 29.7\\% over previous defense methods under one recent adversarial texture attack.", "date": "2025-06-30", "categories": []}
{"id": "2507.03236", "title": "On Jailbreaking Quantized Language Models Through Fault Injection Attacks", "abstract": "The safety alignment of Language Models (LMs) is a critical concern, yet their integrity can be challenged by direct parameter manipulation attacks, such as those potentially induced by fault injection. As LMs are increasingly deployed using low-precision quantization for efficiency, this paper investigates the efficacy of such attacks for jailbreaking aligned LMs across different quantization schemes. We propose gradient-guided attacks, including a tailored progressive bit-level search algorithm introduced herein and a comparative word-level (single weight update) attack. Our evaluation on Llama-3.2-3B, Phi-4-mini, and Llama-3-8B across FP16 (baseline), and weight-only quantization (FP8, INT8, INT4) reveals that quantization significantly influences attack success. While attacks readily achieve high success (>80% Attack Success Rate, ASR) on FP16 models, within an attack budget of 25 perturbations, FP8 and INT8 models exhibit ASRs below 20% and 50%, respectively. Increasing the perturbation budget up to 150 bit-flips, FP8 models maintained ASR below 65%, demonstrating some resilience compared to INT8 and INT4 models that have high ASR. In addition, analysis of perturbation locations revealed differing architectural targets across quantization schemes, with (FP16, INT4) and (INT8, FP8) showing similar characteristics. Besides, jailbreaks induced in FP16 models were highly transferable to subsequent FP8/INT8 quantization (<5% ASR difference), though INT4 significantly reduced transferred ASR (avg. 35% drop). These findings highlight that while common quantization schemes, particularly FP8, increase the difficulty of direct parameter manipulation jailbreaks, vulnerabilities can still persist, especially through post-attack quantization.", "date": "2025-07-04", "categories": []}
{"id": "2507.06256", "title": "Attacker's Noise Can Manipulate Your Audio-based LLM in the Real World", "abstract": "This paper investigates the real-world vulnerabilities of audio-based large language models (ALLMs), such as Qwen2-Audio. We first demonstrate that an adversary can craft stealthy audio perturbations to manipulate ALLMs into exhibiting specific targeted behaviors, such as eliciting responses to wake-keywords (e.g., \"Hey Qwen\"), or triggering harmful behaviors (e.g. \"Change my calendar event\"). Subsequently, we show that playing adversarial background noise during user interaction with the ALLMs can significantly degrade the response quality. Crucially, our research illustrates the scalability of these attacks to real-world scenarios, impacting other innocent users when these adversarial noises are played through the air. Further, we discuss the transferrability of the attack, and potential defensive measures.", "date": "2025-07-07", "categories": []}
{"id": "2507.06856", "title": "IAP: Invisible Adversarial Patch Attack through Perceptibility-Aware Localization and Perturbation Optimization", "abstract": "Despite modifying only a small localized input region, adversarial patches can drastically change the prediction of computer vision models. However, prior methods either cannot perform satisfactorily under targeted attack scenarios or fail to produce contextually coherent adversarial patches, causing them to be easily noticeable by human examiners and insufficiently stealthy against automatic patch defenses. In this paper, we introduce IAP, a novel attack framework that generates highly invisible adversarial patches based on perceptibility-aware localization and perturbation optimization schemes. Specifically, IAP first searches for a proper location to place the patch by leveraging classwise localization and sensitivity maps, balancing the susceptibility of patch location to both victim model prediction and human visual system, then employs a perceptibility-regularized adversarial loss and a gradient update rule that prioritizes color constancy for optimizing invisible perturbations. Comprehensive experiments across various image benchmarks and model architectures demonstrate that IAP consistently achieves competitive attack success rates in targeted settings with significantly improved patch invisibility compared to existing baselines. In addition to being highly imperceptible to humans, IAP is shown to be stealthy enough to render several state-of-the-art patch defenses ineffective.", "date": "2025-07-09", "categories": []}
{"id": "2507.06907", "title": "Robust and Safe Traffic Sign Recognition using N-version with Weighted Voting", "abstract": "Autonomous driving is rapidly advancing as a key application of machine learning, yet ensuring the safety of these systems remains a critical challenge. Traffic sign recognition, an essential component of autonomous vehicles, is particularly vulnerable to adversarial attacks that can compromise driving safety. In this paper, we propose an N-version machine learning (NVML) framework that integrates a safety-aware weighted soft voting mechanism. Our approach utilizes Failure Mode and Effects Analysis (FMEA) to assess potential safety risks and assign dynamic, safety-aware weights to the ensemble outputs. We evaluate the robustness of three-version NVML systems employing various voting mechanisms against adversarial samples generated using the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks. Experimental results demonstrate that our NVML approach significantly enhances the robustness and safety of traffic sign recognition systems under adversarial conditions.", "date": "2025-07-09", "categories": []}
{"id": "2507.07056", "title": "LoRAShield: Data-Free Editing Alignment for Secure Personalized LoRA Sharing", "abstract": "The proliferation of Low-Rank Adaptation (LoRA) models has democratized personalized text-to-image generation, enabling users to share lightweight models (e.g., personal portraits) on platforms like Civitai and Liblib. However, this \"share-and-play\" ecosystem introduces critical risks: benign LoRAs can be weaponized by adversaries to generate harmful content (e.g., political, defamatory imagery), undermining creator rights and platform safety. Existing defenses like concept-erasure methods focus on full diffusion models (DMs), neglecting LoRA's unique role as a modular adapter and its vulnerability to adversarial prompt engineering. To bridge this gap, we propose LoRAShield, the first data-free editing framework for securing LoRA models against misuse. Our platform-driven approach dynamically edits and realigns LoRA's weight subspace via adversarial optimization and semantic augmentation. Experimental results demonstrate that LoRAShield achieves remarkable effectiveness, efficiency, and robustness in blocking malicious generations without sacrificing the functionality of the benign task. By shifting the defense to platforms, LoRAShield enables secure, scalable sharing of personalized models, a critical step toward trustworthy generative ecosystems.", "date": "2025-07-05", "categories": []}
{"id": "2410.16327", "title": "Feint and Attack: Attention-Based Strategies for Jailbreaking and Protecting LLMs", "abstract": "Jailbreak attack can be used to access the vulnerabilities of Large Language Models (LLMs) by inducing LLMs to generate the harmful content. And the most common method of the attack is to construct semantically ambiguous prompts to confuse and mislead the LLMs. To access the security and reveal the intrinsic relation between the input prompt and the output for LLMs, the distribution of attention weight is introduced to analyze the underlying reasons. By using statistical analysis methods, some novel metrics are defined to better describe the distribution of attention weight, such as the Attention Intensity on Sensitive Words (Attn_SensWords), the Attention-based Contextual Dependency Score (Attn_DepScore) and Attention Dispersion Entropy (Attn_Entropy). By leveraging the distinct characteristics of these metrics, the beam search algorithm and inspired by the military strategy \"Feint and Attack\", an effective jailbreak attack strategy named as Attention-Based Attack (ABA) is proposed. In the ABA, nested attack prompts are employed to divert the attention distribution of the LLMs. In this manner, more harmless parts of the input can be used to attract the attention of the LLMs. In addition, motivated by ABA, an effective defense strategy called as Attention-Based Defense (ABD) is also put forward. Compared with ABA, the ABD can be used to enhance the robustness of LLMs by calibrating the attention distribution of the input prompt. Some comparative experiments have been given to demonstrate the effectiveness of ABA and ABD. Therefore, both ABA and ABD can be used to access the security of the LLMs. The comparative experiment results also give a logical explanation that the distribution of attention weight can bring great influence on the output for LLMs.", "date": "2024-10-18", "categories": []}
{"id": "2505.19840", "title": "One Surrogate to Fool Them All: Universal, Transferable, and Targeted Adversarial Attacks with CLIP", "abstract": "Deep Neural Networks (DNNs) have achieved widespread success yet remain prone to adversarial attacks. Typically, such attacks either involve frequent queries to the target model or rely on surrogate models closely mirroring the target model -- often trained with subsets of the target model's training data -- to achieve high attack success rates through transferability. However, in realistic scenarios where training data is inaccessible and excessive queries can raise alarms, crafting adversarial examples becomes more challenging. In this paper, we present UnivIntruder, a novel attack framework that relies solely on a single, publicly available CLIP model and publicly available datasets. By using textual concepts, UnivIntruder generates universal, transferable, and targeted adversarial perturbations that mislead DNNs into misclassifying inputs into adversary-specified classes defined by textual concepts.\n  Our extensive experiments show that our approach achieves an Attack Success Rate (ASR) of up to 85% on ImageNet and over 99% on CIFAR-10, significantly outperforming existing transfer-based methods. Additionally, we reveal real-world vulnerabilities, showing that even without querying target models, UnivIntruder compromises image search engines like Google and Baidu with ASR rates up to 84%, and vision language models like GPT-4 and Claude-3.5 with ASR rates up to 80%. These findings underscore the practicality of our attack in scenarios where traditional avenues are blocked, highlighting the need to reevaluate security paradigms in AI applications.", "date": "2025-05-26", "categories": []}
{"id": "2507.05441", "title": "Adversarial Machine Learning Attacks on Financial Reporting via Maximum Violated Multi-Objective Attack", "abstract": "Bad actors, primarily distressed firms, have the incentive and desire to manipulate their financial reports to hide their distress and derive personal gains. As attackers, these firms are motivated by potentially millions of dollars and the availability of many publicly disclosed and used financial modeling frameworks. Existing attack methods do not work on this data due to anti-correlated objectives that must both be satisfied for the attacker to succeed. We introduce Maximum Violated Multi-Objective (MVMO) attacks that adapt the attacker's search direction to find $20\\times$ more satisfying attacks compared to standard attacks. The result is that in $\\approx50\\%$ of cases, a company could inflate their earnings by 100-200%, while simultaneously reducing their fraud scores by 15%. By working with lawyers and professional accountants, we ensure our threat model is realistic to how such frauds are performed in practice.", "date": "2025-07-07", "categories": []}
{"id": "2406.15213", "title": "Backdooring Bias ($B^2$) into Stable Diffusion Models", "abstract": "Recent advances in large text-conditional diffusion models have revolutionized image generation by enabling users to create realistic, high-quality images from textual prompts, significantly enhancing artistic creation and visual communication. However, these advancements also introduce an underexplored attack opportunity: the possibility of inducing biases by an adversary into the generated images for malicious intentions, e.g., to influence public opinion and spread propaganda. In this paper, we study an attack vector that allows an adversary to inject arbitrary bias into a target model. The attack leverages low-cost backdooring techniques using a targeted set of natural textual triggers embedded within a small number of malicious data samples produced with public generative models. An adversary could pick common sequences of words that can then be inadvertently activated by benign users during inference. We investigate the feasibility and challenges of such attacks, demonstrating how modern generative models have made this adversarial process both easier and more adaptable. On the other hand, we explore various aspects of the detectability of such attacks and demonstrate that the model's utility remains intact in the absence of the triggers. Our extensive experiments using over 200,000 generated images and against hundreds of fine-tuned models demonstrate the feasibility of the presented backdoor attack. We illustrate how these biases maintain strong text-image alignment, highlighting the challenges in detecting biased images without knowing that bias in advance. Our cost analysis confirms the low financial barrier (\\$10-\\$15) to executing such attacks, underscoring the need for robust defensive strategies against such vulnerabilities in diffusion models.", "date": "2024-06-21", "categories": []}
{"id": "2412.05883", "title": "On the Adversarial Robustness of Graph Neural Networks with Graph Reduction", "abstract": "As Graph Neural Networks (GNNs) become increasingly popular for learning from large-scale graph data across various domains, their susceptibility to adversarial attacks when using graph reduction techniques for scalability remains underexplored. In this paper, we present an extensive empirical study to investigate the impact of graph reduction techniques, specifically graph coarsening and sparsification, on the robustness of GNNs against adversarial attacks. Through extensive experiments involving multiple datasets and GNN architectures, we examine the effects of four sparsification and six coarsening methods on the poisoning attacks. Our results indicate that, while graph sparsification can mitigate the effectiveness of certain poisoning attacks, such as Mettack, it has limited impact on others, like PGD. Conversely, graph coarsening tends to amplify the adversarial impact, significantly reducing classification accuracy as the reduction ratio decreases. Additionally, we provide a novel analysis of the causes driving these effects and examine how defensive GNN models perform under graph reduction, offering practical insights for designing robust GNNs within graph acceleration systems.", "date": "2024-12-08", "categories": []}
{"id": "2507.02956", "title": "A Representation Engineering Perspective on the Effectiveness of Multi-Turn Jailbreaks", "abstract": "Recent research has demonstrated that state-of-the-art LLMs and defenses remain susceptible to multi-turn jailbreak attacks. These attacks require only closed-box model access and are often easy to perform manually, posing a significant threat to the safe and secure deployment of LLM-based systems. We study the effectiveness of the Crescendo multi-turn jailbreak at the level of intermediate model representations and find that safety-aligned LMs often represent Crescendo responses as more benign than harmful, especially as the number of conversation turns increases. Our analysis indicates that at each turn, Crescendo prompts tend to keep model outputs in a \"benign\" region of representation space, effectively tricking the model into fulfilling harmful requests. Further, our results help explain why single-turn jailbreak defenses like circuit breakers are generally ineffective against multi-turn attacks, motivating the development of mitigations that address this generalization gap.", "date": "2025-06-29", "categories": []}
{"id": "2507.03427", "title": "Rectifying Adversarial Sample with Low Entropy Prior for Test-Time Defense", "abstract": "Existing defense methods fail to defend against unknown attacks and thus raise generalization issue of adversarial robustness. To remedy this problem, we attempt to delve into some underlying common characteristics among various attacks for generality. In this work, we reveal the commonly overlooked low entropy prior (LE) implied in various adversarial samples, and shed light on the universal robustness against unseen attacks in inference phase. LE prior is elaborated as two properties across various attacks as shown in Fig. 1 and Fig. 2: 1) low entropy misclassification for adversarial samples and 2) lower entropy prediction for higher attack intensity. This phenomenon stands in stark contrast to the naturally distributed samples. The LE prior can instruct existing test-time defense methods, thus we propose a two-stage REAL approach: Rectify Adversarial sample based on LE prior for test-time adversarial rectification. Specifically, to align adversarial samples more closely with clean samples, we propose to first rectify adversarial samples misclassified with low entropy by reverse maximizing prediction entropy, thereby eliminating their adversarial nature. To ensure the rectified samples can be correctly classified with low entropy, we carry out secondary rectification by forward minimizing prediction entropy, thus creating a Max-Min entropy optimization scheme. Further, based on the second property, we propose an attack-aware weighting mechanism to adaptively adjust the strengths of Max-Min entropy objectives. Experiments on several datasets show that REAL can greatly improve the performance of existing sample rectification models.", "date": "2025-07-04", "categories": []}
{"id": "2507.03450", "title": "Evaluating the Evaluators: Trust in Adversarial Robustness Tests", "abstract": "Despite significant progress in designing powerful adversarial evasion attacks for robustness verification, the evaluation of these methods often remains inconsistent and unreliable. Many assessments rely on mismatched models, unverified implementations, and uneven computational budgets, which can lead to biased results and a false sense of security. Consequently, robustness claims built on such flawed testing protocols may be misleading and give a false sense of security. As a concrete step toward improving evaluation reliability, we present AttackBench, a benchmark framework developed to assess the effectiveness of gradient-based attacks under standardized and reproducible conditions. AttackBench serves as an evaluation tool that ranks existing attack implementations based on a novel optimality metric, which enables researchers and practitioners to identify the most reliable and effective attack for use in subsequent robustness evaluations. The framework enforces consistent testing conditions and enables continuous updates, making it a reliable foundation for robustness verification.", "date": "2025-07-04", "categories": []}
{"id": "2507.03953", "title": "Evaluating Adversarial Protections for Diffusion Personalization: A Comprehensive Study", "abstract": "With the increasing adoption of diffusion models for image generation and personalization, concerns regarding privacy breaches and content misuse have become more pressing. In this study, we conduct a comprehensive comparison of eight perturbation based protection methods: AdvDM, ASPL, FSGM, MetaCloak, Mist, PhotoGuard, SDS, and SimAC--across both portrait and artwork domains. These methods are evaluated under varying perturbation budgets, using a range of metrics to assess visual imperceptibility and protective efficacy. Our results offer practical guidance for method selection. Code is available at: https://github.com/vkeilo/DiffAdvPerturbationBench.", "date": "2025-07-05", "categories": []}
{"id": "2506.23661", "title": "Robustness of Misinformation Classification Systems to Adversarial Examples Through BeamAttack", "abstract": "We extend BeamAttack, an adversarial attack algorithm designed to evaluate the robustness of text classification systems through word-level modifications guided by beam search. Our extensions include support for word deletions and the option to skip substitutions, enabling the discovery of minimal modifications that alter model predictions. We also integrate LIME to better prioritize word replacements. Evaluated across multiple datasets and victim models (BiLSTM, BERT, and adversarially trained RoBERTa) within the BODEGA framework, our approach achieves over a 99\\% attack success rate while preserving the semantic and lexical similarity of the original texts. Through both quantitative and qualitative analysis, we highlight BeamAttack's effectiveness and its limitations. Our implementation is available at https://github.com/LucK1Y/BeamAttack", "date": "2025-06-30", "categories": []}
{"id": "2404.16369", "title": "Don't Say No: Jailbreaking LLM by Suppressing Refusal", "abstract": "Ensuring the safety alignment of Large Language Models (LLMs) is critical for generating responses consistent with human values. However, LLMs remain vulnerable to jailbreaking attacks, where carefully crafted prompts manipulate them into producing toxic content. One category of such attacks reformulates the task as an optimization problem, aiming to elicit affirmative responses from the LLM. However, these methods heavily rely on predefined objectionable behaviors, limiting their effectiveness and adaptability to diverse harmful queries. In this study, we first identify why the vanilla target loss is suboptimal and then propose enhancements to the loss objective. We introduce DSN (Don't Say No) attack, which combines a cosine decay schedule method with refusal suppression to achieve higher success rates. Extensive experiments demonstrate that DSN outperforms baseline attacks and achieves state-of-the-art attack success rates (ASR). DSN also shows strong universality and transferability to unseen datasets and black-box models.", "date": "2024-04-25", "categories": []}
{"id": "2411.13757", "title": "GenBFA: An Evolutionary Optimization Approach to Bit-Flip Attacks on LLMs", "abstract": "Large Language Models (LLMs) have revolutionized natural language processing (NLP), excelling in tasks like text generation and summarization. However, their increasing adoption in mission-critical applications raises concerns about hardware-based threats, particularly bit-flip attacks (BFAs). BFAs, enabled by fault injection methods such as Rowhammer, target model parameters in memory, compromising both integrity and performance. Identifying critical parameters for BFAs in the vast parameter space of LLMs poses significant challenges. While prior research suggests transformer-based architectures are inherently more robust to BFAs compared to traditional deep neural networks, we challenge this assumption. For the first time, we demonstrate that as few as three bit-flips can cause catastrophic performance degradation in an LLM with billions of parameters. Current BFA techniques are inadequate for exploiting this vulnerability due to the difficulty of efficiently identifying critical parameters within the immense parameter space. To address this, we propose AttentionBreaker, a novel framework tailored for LLMs that enables efficient traversal of the parameter space to identify critical parameters. Additionally, we introduce GenBFA, an evolutionary optimization strategy designed to refine the search further, isolating the most critical bits for an efficient and effective attack. Empirical results reveal the profound vulnerability of LLMs to AttentionBreaker. For example, merely three bit-flips (4.129 x 10^-9% of total parameters) in the LLaMA3-8B-Instruct 8-bit quantized (W8) model result in a complete performance collapse: accuracy on MMLU tasks drops from 67.3% to 0%, and Wikitext perplexity skyrockets from 12.6 to 4.72 x 10^5. These findings underscore the effectiveness of AttentionBreaker in uncovering and exploiting critical vulnerabilities within LLM architectures.", "date": "2024-11-21", "categories": []}
{"id": "2507.01791", "title": "Boosting Adversarial Transferability Against Defenses via Multi-Scale Transformation", "abstract": "The transferability of adversarial examples poses a significant security challenge for deep neural networks, which can be attacked without knowing anything about them. In this paper, we propose a new Segmented Gaussian Pyramid (SGP) attack method to enhance the transferability, particularly against defense models. Unlike existing methods that generally focus on single-scale images, our approach employs Gaussian filtering and three types of downsampling to construct a series of multi-scale examples. Then, the gradients of the loss function with respect to each scale are computed, and their average is used to determine the adversarial perturbations. The proposed SGP can be considered an input transformation with high extensibility that is easily integrated into most existing adversarial attacks. Extensive experiments demonstrate that in contrast to the state-of-the-art methods, SGP significantly enhances attack success rates against black-box defense models, with average attack success rates increasing by 2.3% to 32.6%, based only on transferability.", "date": "2025-07-02", "categories": []}
{"id": "2501.13094", "title": "Robust Representation Consistency Model via Contrastive Denoising", "abstract": "Robustness is essential for deep neural networks, especially in security-sensitive applications. To this end, randomized smoothing provides theoretical guarantees for certifying robustness against adversarial perturbations. Recently, diffusion models have been successfully employed for randomized smoothing to purify noise-perturbed samples before making predictions with a standard classifier. While these methods excel at small perturbation radii, they struggle with larger perturbations and incur a significant computational overhead during inference compared to classical methods. To address this, we reformulate the generative modeling task along the diffusion trajectories in pixel space as a discriminative task in the latent space. Specifically, we use instance discrimination to achieve consistent representations along the trajectories by aligning temporally adjacent points. After fine-tuning based on the learned representations, our model enables implicit denoising-then-classification via a single prediction, substantially reducing inference costs. We conduct extensive experiments on various datasets and achieve state-of-the-art performance with minimal computation budget during inference. For example, our method outperforms the certified accuracy of diffusion-based methods on ImageNet across all perturbation radii by 5.3% on average, with up to 11.6% at larger radii, while reducing inference costs by 85$\\times$ on average. Codes are available at: https://github.com/jiachenlei/rRCM.", "date": "2025-01-22", "categories": []}
{"id": "2506.10685", "title": "Defensive Adversarial CAPTCHA: A Semantics-Driven Framework for Natural Adversarial Example Generation", "abstract": "Traditional CAPTCHA (Completely Automated Public Turing Test to Tell Computers and Humans Apart) schemes are increasingly vulnerable to automated attacks powered by deep neural networks (DNNs). Existing adversarial attack methods often rely on the original image characteristics, resulting in distortions that hinder human interpretation and limit their applicability in scenarios where no initial input images are available. To address these challenges, we propose the Unsourced Adversarial CAPTCHA (DAC), a novel framework that generates high-fidelity adversarial examples guided by attacker-specified semantics information. Leveraging a Large Language Model (LLM), DAC enhances CAPTCHA diversity and enriches the semantic information. To address various application scenarios, we examine the white-box targeted attack scenario and the black box untargeted attack scenario. For target attacks, we introduce two latent noise variables that are alternately guided in the diffusion step to achieve robust inversion. The synergy between gradient guidance and latent variable optimization achieved in this way ensures that the generated adversarial examples not only accurately align with the target conditions but also achieve optimal performance in terms of distributional consistency and attack effectiveness. In untargeted attacks, especially for black-box scenarios, we introduce bi-path unsourced adversarial CAPTCHA (BP-DAC), a two-step optimization strategy employing multimodal gradients and bi-path optimization for efficient misclassification. Experiments show that the defensive adversarial CAPTCHA generated by BP-DAC is able to defend against most of the unknown models, and the generated CAPTCHA is indistinguishable to both humans and DNNs.", "date": "2025-06-12", "categories": []}
{"id": "2507.00015", "title": "Vision Transformer with Adversarial Indicator Token against Adversarial Attacks in Radio Signal Classifications", "abstract": "The remarkable success of transformers across various fields such as natural language processing and computer vision has paved the way for their applications in automatic modulation classification, a critical component in the communication systems of Internet of Things (IoT) devices. However, it has been observed that transformer-based classification of radio signals is susceptible to subtle yet sophisticated adversarial attacks. To address this issue, we have developed a defensive strategy for transformer-based modulation classification systems to counter such adversarial attacks. In this paper, we propose a novel vision transformer (ViT) architecture by introducing a new concept known as adversarial indicator (AdvI) token to detect adversarial attacks. To the best of our knowledge, this is the first work to propose an AdvI token in ViT to defend against adversarial attacks. Integrating an adversarial training method with a detection mechanism using AdvI token, we combine a training time defense and running time defense in a unified neural network model, which reduces architectural complexity of the system compared to detecting adversarial perturbations using separate models. We investigate into the operational principles of our method by examining the attention mechanism. We show the proposed AdvI token acts as a crucial element within the ViT, influencing attention weights and thereby highlighting regions or features in the input data that are potentially suspicious or anomalous. Through experimental results, we demonstrate that our approach surpasses several competitive methods in handling white-box attack scenarios, including those utilizing the fast gradient method, projected gradient descent attacks and basic iterative method.", "date": "2025-06-13", "categories": []}
{"id": "2507.00509", "title": "TeamCMU at Touch\\'e: Adversarial Co-Evolution for Advertisement Integration and Detection in Conversational Search", "abstract": "As conversational search engines increasingly adopt generation-based paradigms powered by Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG), the integration of advertisements into generated responses presents both commercial opportunities and challenges for user experience. Unlike traditional search, where advertisements are clearly delineated, generative systems blur the boundary between informational content and promotional material, raising concerns around transparency and trust. In this work, we propose a modular pipeline for advertisement management in RAG-based conversational systems, consisting of an ad-rewriter for seamless ad integration and a robust ad-classifier for detection. We leverage synthetic data to train high-performing classifiers, which are then used to guide two complementary ad-integration strategies: supervised fine-tuning of the ad-rewriter and a best-of-N sampling approach that selects the least detectable ad-integrated response among multiple candidates. Our evaluation focuses on two core questions: the effectiveness of ad classifiers in detecting diverse ad integration strategies, and the training methods that best support coherent, minimally intrusive ad insertion. Experimental results show that our ad-classifier, trained on synthetic advertisement data inspired by marketing strategies and enhanced through curriculum learning, achieves robust detection performance. Additionally, we demonstrate that classifier-guided optimization, through both fine-tuning and best-of-N sampling, significantly improves ad stealth, enabling more seamless integration. These findings contribute an adversarial co-evolution framework for developing more sophisticated ad-aware generative search systems and robust ad classifiers.", "date": "2025-07-01", "categories": []}
{"id": "2507.00577", "title": "BadViM: Backdoor Attack against Vision Mamba", "abstract": "Vision State Space Models (SSMs), particularly architectures like Vision Mamba (ViM), have emerged as promising alternatives to Vision Transformers (ViTs). However, the security implications of this novel architecture, especially their vulnerability to backdoor attacks, remain critically underexplored. Backdoor attacks aim to embed hidden triggers into victim models, causing the model to misclassify inputs containing these triggers while maintaining normal behavior on clean inputs. This paper investigates the susceptibility of ViM to backdoor attacks by introducing BadViM, a novel backdoor attack framework specifically designed for Vision Mamba. The proposed BadViM leverages a Resonant Frequency Trigger (RFT) that exploits the frequency sensitivity patterns of the victim model to create stealthy, distributed triggers. To maximize attack efficacy, we propose a Hidden State Alignment loss that strategically manipulates the internal representations of model by aligning the hidden states of backdoor images with those of target classes. Extensive experimental results demonstrate that BadViM achieves superior attack success rates while maintaining clean data accuracy. Meanwhile, BadViM exhibits remarkable resilience against common defensive measures, including PatchDrop, PatchShuffle and JPEG compression, which typically neutralize normal backdoor attacks.", "date": "2025-07-01", "categories": []}
{"id": "2507.00690", "title": "Cage-Based Deformation for Transferable and Undefendable Point Cloud Attack", "abstract": "Adversarial attacks on point clouds often impose strict geometric constraints to preserve plausibility; however, such constraints inherently limit transferability and undefendability. While deformation offers an alternative, existing unstructured approaches may introduce unnatural distortions, making adversarial point clouds conspicuous and undermining their plausibility. In this paper, we propose CageAttack, a cage-based deformation framework that produces natural adversarial point clouds. It first constructs a cage around the target object, providing a structured basis for smooth, natural-looking deformation. Perturbations are then applied to the cage vertices, which seamlessly propagate to the point cloud, ensuring that the resulting deformations remain intrinsic to the object and preserve plausibility. Extensive experiments on seven 3D deep neural network classifiers across three datasets show that CageAttack achieves a superior balance among transferability, undefendability, and plausibility, outperforming state-of-the-art methods. Codes will be made public upon acceptance.", "date": "2025-07-01", "categories": []}
{"id": "2507.00817", "title": "CAVALRY-V: A Large-Scale Generator Framework for Adversarial Attacks on Video MLLMs", "abstract": "Video Multimodal Large Language Models (V-MLLMs) have shown impressive capabilities in temporal reasoning and cross-modal understanding, yet their vulnerability to adversarial attacks remains underexplored due to unique challenges: complex cross-modal reasoning mechanisms, temporal dependencies, and computational constraints. We present CAVALRY-V (Cross-modal Language-Vision Adversarial Yielding for Videos), a novel framework that directly targets the critical interface between visual perception and language generation in V-MLLMs. Our approach introduces two key innovations: (1) a dual-objective semantic-visual loss function that simultaneously disrupts the model's text generation logits and visual representations to undermine cross-modal integration, and (2) a computationally efficient two-stage generator framework that combines large-scale pre-training for cross-model transferability with specialized fine-tuning for spatiotemporal coherence. Empirical evaluation on comprehensive video understanding benchmarks demonstrates that CAVALRY-V significantly outperforms existing attack methods, achieving 22.8% average improvement over the best baseline attacks on both commercial systems (GPT-4.1, Gemini 2.0) and open-source models (QwenVL-2.5, InternVL-2.5, Llava-Video, Aria, MiniCPM-o-2.6). Our framework achieves flexibility through implicit temporal coherence modeling rather than explicit regularization, enabling significant performance improvements even on image understanding (34.4% average gain). This capability demonstrates CAVALRY-V's potential as a foundational approach for adversarial research across multimodal systems.", "date": "2025-07-01", "categories": []}
{"id": "2310.11594", "title": "Adversarial Robustness Unhardening via Backdoor Attacks in Federated Learning", "abstract": "The delicate equilibrium between user privacy and the ability to unleash the potential of distributed data is an important concern. Federated learning, which enables the training of collaborative models without sharing of data, has emerged as a privacy-centric solution. This approach brings forth security challenges, notably poisoning and backdoor attacks where malicious entities inject corrupted data into the training process, as well as evasion attacks that aim to induce misclassifications at test time. Our research investigates the intersection of adversarial training, a common defense method against evasion attacks, and backdoor attacks within federated learning. We introduce Adversarial Robustness Unhardening (ARU), which is employed by a subset of adversarial clients to intentionally undermine model robustness during federated training, rendering models susceptible to a broader range of evasion attacks. We present extensive experiments evaluating ARU's impact on adversarial training and existing robust aggregation defenses against poisoning and backdoor attacks. Our results show that ARU can substantially undermine adversarial training's ability to harden models against test-time evasion attacks, and that adversaries employing ARU can even evade robust aggregation defenses that often neutralize poisoning or backdoor attacks.", "date": "2023-10-17", "categories": []}
{"id": "2410.01697", "title": "Enhancing Adversarial Robustness through Multi-Objective Representation Learning", "abstract": "Deep neural networks (DNNs) are vulnerable to small adversarial perturbations, which are tiny changes to the input data that appear insignificant but cause the model to produce drastically different outputs. Many defense methods require modifying model architectures during evaluation or performing test-time data purification. This not only introduces additional complexity but is often architecture-dependent. We show, however, that robust feature learning during training can significantly enhance DNN robustness. We propose MOREL, a multi-objective approach that aligns natural and adversarial features using cosine similarity and multi-positive contrastive losses to encourage similar features for same-class inputs. Extensive experiments demonstrate that MOREL significantly improves robustness against both white-box and black-box attacks. Our code is available at https://github.com/salomonhotegni/MOREL", "date": "2024-10-02", "categories": []}
{"id": "2411.05853", "title": "A Fundamental Accuracy--Robustness Trade-off in Regression and Classification", "abstract": "We derive a fundamental trade-off between standard and adversarial risk in a rather general situation that formalizes the following simple intuition: \"If no (nearly) optimal predictor is smooth, adversarial robustness comes at the cost of accuracy.\" As a concrete example, we evaluate the derived trade-off in regression with polynomial ridge functions under mild regularity conditions. Generalizing our analysis of this example, we formulate a necessary condition under which adversarial robustness can be achieved without significant degradation of the accuracy. This necessary condition is expressed in terms of a quantity that resembles the Poincar\\'{e} constant of the data distribution.", "date": "2024-11-06", "categories": []}
{"id": "2506.22602", "title": "Are Fast Methods Stable in Adversarially Robust Transfer Learning?", "abstract": "Transfer learning is often used to decrease the computational cost of model training, as fine-tuning a model allows a downstream task to leverage the features learned from the pre-training dataset and quickly adapt them to a new task. This is particularly useful for achieving adversarial robustness, as adversarially training models from scratch is very computationally expensive. However, high robustness in transfer learning still requires adversarial training during the fine-tuning phase, which requires up to an order of magnitude more time than standard fine-tuning. In this work, we revisit the use of the fast gradient sign method (FGSM) in robust transfer learning to improve the computational cost of adversarial fine-tuning. We surprisingly find that FGSM is much more stable in adversarial fine-tuning than when training from scratch. In particular, FGSM fine-tuning does not suffer from any issues with catastrophic overfitting at standard perturbation budgets of $\\varepsilon=4$ or $\\varepsilon=8$. This stability is further enhanced with parameter-efficient fine-tuning methods, where FGSM remains stable even up to $\\varepsilon=32$ for linear probing. We demonstrate how this stability translates into performance across multiple datasets. Compared to fine-tuning with the more commonly used method of projected gradient descent (PGD), on average, FGSM only loses 0.39% and 1.39% test robustness for $\\varepsilon=4$ and $\\varepsilon=8$ while using $4\\times$ less training time. Surprisingly, FGSM may not only be a significantly more efficient alternative to PGD in adversarially robust transfer learning but also a well-performing one.", "date": "2025-06-27", "categories": []}
{"id": "2506.22722", "title": "Kill Two Birds with One Stone! Trajectory enabled Unified Online Detection of Adversarial Examples and Backdoor Attacks", "abstract": "The proposed UniGuard is the first unified online detection framework capable of simultaneously addressing adversarial examples and backdoor attacks. UniGuard builds upon two key insights: first, both AE and backdoor attacks have to compromise the inference phase, making it possible to tackle them simultaneously during run-time via online detection. Second, an adversarial input, whether a perturbed sample in AE attacks or a trigger-carrying sample in backdoor attacks, exhibits distinctive trajectory signatures from a benign sample as it propagates through the layers of a DL model in forward inference. The propagation trajectory of the adversarial sample must deviate from that of its benign counterpart; otherwise, the adversarial objective cannot be fulfilled. Detecting these trajectory signatures is inherently challenging due to their subtlety; UniGuard overcomes this by treating the propagation trajectory as a time-series signal, leveraging LSTM and spectrum transformation to amplify differences between adversarial and benign trajectories that are subtle in the time domain. UniGuard exceptional efficiency and effectiveness have been extensively validated across various modalities (image, text, and audio) and tasks (classification and regression), ranging from diverse model architectures against a wide range of AE attacks and backdoor attacks, including challenging partial backdoors and dynamic triggers. When compared to SOTA methods, including ContraNet (NDSS 22) specific for AE detection and TED (IEEE SP 24) specific for backdoor detection, UniGuard consistently demonstrates superior performance, even when matched against each method's strengths in addressing their respective threats-each SOTA fails to parts of attack strategies while UniGuard succeeds for all.", "date": "2025-06-28", "categories": []}
{"id": "2506.23296", "title": "Securing AI Systems: A Guide to Known Attacks and Impacts", "abstract": "Embedded into information systems, artificial intelligence (AI) faces security threats that exploit AI-specific vulnerabilities. This paper provides an accessible overview of adversarial attacks unique to predictive and generative AI systems. We identify eleven major attack types and explicitly link attack techniques to their impacts -- including information leakage, system compromise, and resource exhaustion -- mapped to the confidentiality, integrity, and availability (CIA) security triad. We aim to equip researchers, developers, security practitioners, and policymakers, even those without specialized AI security expertise, with foundational knowledge to recognize AI-specific risks and implement effective defenses, thereby enhancing the overall security posture of AI systems.", "date": "2025-06-29", "categories": []}
{"id": "2506.23676", "title": "A Unified Framework for Stealthy Adversarial Generation via Latent Optimization and Transferability Enhancement", "abstract": "Due to their powerful image generation capabilities, diffusion-based adversarial example generation methods through image editing are rapidly gaining popularity. However, due to reliance on the discriminative capability of the diffusion model, these diffusion-based methods often struggle to generalize beyond conventional image classification tasks, such as in Deepfake detection. Moreover, traditional strategies for enhancing adversarial example transferability are challenging to adapt to these methods. To address these challenges, we propose a unified framework that seamlessly incorporates traditional transferability enhancement strategies into diffusion model-based adversarial example generation via image editing, enabling their application across a wider range of downstream tasks. Our method won first place in the \"1st Adversarial Attacks on Deepfake Detectors: A Challenge in the Era of AI-Generated Media\" competition at ACM MM25, which validates the effectiveness of our approach.", "date": "2025-06-30", "categories": []}
{"id": "2506.24048", "title": "Consensus-based optimization for closed-box adversarial attacks and a connection to evolution strategies", "abstract": "Consensus-based optimization (CBO) has established itself as an efficient gradient-free optimization scheme, with attractive mathematical properties, such as mean-field convergence results for non-convex loss functions. In this work, we study CBO in the context of closed-box adversarial attacks, which are imperceptible input perturbations that aim to fool a classifier, without accessing its gradient. Our contribution is to establish a connection between the so-called consensus hopping as introduced by Riedl et al. and natural evolution strategies (NES) commonly applied in the context of adversarial attacks and to rigorously relate both methods to gradient-based optimization schemes. Beyond that, we provide a comprehensive experimental study that shows that despite the conceptual similarities, CBO can outperform NES and other evolutionary strategies in certain scenarios.", "date": "2025-06-30", "categories": []}
{"id": "2305.09305", "title": "Releasing Inequality Phenomenon in $\\ell_{\\infty}$-norm Adversarial Training via Input Gradient Distillation", "abstract": "Adversarial training (AT) is considered the most effective defense against adversarial attacks. However, a recent study revealed that \\(\\ell_{\\infty}\\)-norm adversarial training (\\(\\ell_{\\infty}\\)-AT) will also induce unevenly distributed input gradients, which is called the inequality phenomenon. This phenomenon makes the \\(\\ell_{\\infty}\\)-norm adversarially trained model more vulnerable than the standard-trained model when high-attribution or randomly selected pixels are perturbed, enabling robust and practical black-box attacks against \\(\\ell_{\\infty}\\)-adversarially trained models. In this paper, we propose a simple yet effective method called Input Gradient Distillation (IGD) to release the inequality phenomenon in $\\ell_{\\infty}$-AT. IGD distills the standard-trained teacher model's equal decision pattern into the $\\ell_{\\infty}$-adversarially trained student model by aligning input gradients of the student model and the standard-trained model with the Cosine Similarity. Experiments show that IGD can mitigate the inequality phenomenon and its threats while preserving adversarial robustness. Compared to vanilla $\\ell_{\\infty}$-AT, IGD reduces error rates against inductive noise, inductive occlusion, random noise, and noisy images in ImageNet-C by up to 60\\%, 16\\%, 50\\%, and 21\\%, respectively. Other than empirical experiments, we also conduct a theoretical analysis to explain why releasing the inequality phenomenon can improve such robustness and discuss why the severity of the inequality phenomenon varies according to the dataset's image resolution. Our code is available at https://github.com/fhdnskfbeuv/Inuput-Gradient-Distillation", "date": "2023-05-16", "categories": []}
{"id": "2403.12988", "title": "Enhancing Object Detection Robustness: Detecting and Restoring Confidence in the Presence of Adversarial Patch Attacks", "abstract": "The widespread adoption of computer vision systems has underscored their susceptibility to adversarial attacks, particularly adversarial patch attacks on object detectors. This study evaluates defense mechanisms for the YOLOv5 model against such attacks. Optimized adversarial patches were generated and placed in sensitive image regions, by applying EigenCAM and grid search to determine optimal placement. We tested several defenses, including Segment and Complete (SAC), Inpainting, and Latent Diffusion Models. Our pipeline comprises three main stages: patch application, object detection, and defense analysis. Results indicate that adversarial patches reduce average detection confidence by 22.06\\%. Defenses restored confidence levels by 3.45\\% (SAC), 5.05\\% (Inpainting), and significantly improved them by 26.61\\%, which even exceeds the original accuracy levels, when using the Latent Diffusion Model, highlighting its superior effectiveness in mitigating the effects of adversarial patches.", "date": "2024-03-04", "categories": []}
{"id": "2506.21874", "title": "On the Feasibility of Poisoning Text-to-Image AI Models via Adversarial Mislabeling", "abstract": "Today's text-to-image generative models are trained on millions of images sourced from the Internet, each paired with a detailed caption produced by Vision-Language Models (VLMs). This part of the training pipeline is critical for supplying the models with large volumes of high-quality image-caption pairs during training. However, recent work suggests that VLMs are vulnerable to stealthy adversarial attacks, where adversarial perturbations are added to images to mislead the VLMs into producing incorrect captions.\n  In this paper, we explore the feasibility of adversarial mislabeling attacks on VLMs as a mechanism to poisoning training pipelines for text-to-image models. Our experiments demonstrate that VLMs are highly vulnerable to adversarial perturbations, allowing attackers to produce benign-looking images that are consistently miscaptioned by the VLM models. This has the effect of injecting strong \"dirty-label\" poison samples into the training pipeline for text-to-image models, successfully altering their behavior with a small number of poisoned samples. We find that while potential defenses can be effective, they can be targeted and circumvented by adaptive attackers. This suggests a cat-and-mouse game that is likely to reduce the quality of training data and increase the cost of text-to-image model development. Finally, we demonstrate the real-world effectiveness of these attacks, achieving high attack success (over 73%) even in black-box scenarios against commercial VLMs (Google Vertex AI and Microsoft Azure).", "date": "2025-06-27", "categories": []}
{"id": "2506.21972", "title": "Advancing Jailbreak Strategies: A Hybrid Approach to Exploiting LLM Vulnerabilities and Bypassing Modern Defenses", "abstract": "The advancement of Pre-Trained Language Models (PTLMs) and Large Language Models (LLMs) has led to their widespread adoption across diverse applications. Despite their success, these models remain vulnerable to attacks that exploit their inherent weaknesses to bypass safety measures. Two primary inference-phase threats are token-level and prompt-level jailbreaks. Token-level attacks embed adversarial sequences that transfer well to black-box models like GPT but leave detectable patterns and rely on gradient-based token optimization, whereas prompt-level attacks use semantically structured inputs to elicit harmful responses yet depend on iterative feedback that can be unreliable. To address the complementary limitations of these methods, we propose two hybrid approaches that integrate token- and prompt-level techniques to enhance jailbreak effectiveness across diverse PTLMs. GCG + PAIR and the newly explored GCG + WordGame hybrids were evaluated across multiple Vicuna and Llama models. GCG + PAIR consistently raised attack-success rates over its constituent techniques on undefended models; for instance, on Llama-3, its Attack Success Rate (ASR) reached 91.6%, a substantial increase from PAIR's 58.4% baseline. Meanwhile, GCG + WordGame matched the raw performance of WordGame maintaining a high ASR of over 80% even under stricter evaluators like Mistral-Sorry-Bench. Crucially, both hybrids retained transferability and reliably pierced advanced defenses such as Gradient Cuff and JBShield, which fully blocked single-mode attacks. These findings expose previously unreported vulnerabilities in current safety stacks, highlight trade-offs between raw success and defensive robustness, and underscore the need for holistic safeguards against adaptive adversaries.", "date": "2025-06-27", "categories": []}
{"id": "2401.10586", "title": "PuriDefense: Randomized Local Implicit Adversarial Purification for Defending Black-box Query-based Attacks", "abstract": "Black-box query-based attacks constitute significant threats to Machine Learning as a Service (MLaaS) systems since they can generate adversarial examples without accessing the target model's architecture and parameters. Traditional defense mechanisms, such as adversarial training, gradient masking, and input transformations, either impose substantial computational costs or compromise the test accuracy of non-adversarial inputs. To address these challenges, we propose an efficient defense mechanism, PuriDefense, that employs random patch-wise purifications with an ensemble of lightweight purification models at a low level of inference cost. These models leverage the local implicit function and rebuild the natural image manifold. Our theoretical analysis suggests that this approach slows down the convergence of query-based attacks by incorporating randomness into purifications. Extensive experiments on CIFAR-10 and ImageNet validate the effectiveness of our proposed purifier-based defense mechanism, demonstrating significant improvements in robustness against query-based attacks.", "date": "2024-01-19", "categories": []}
{"id": "2410.16155", "title": "A Troublemaker with Contagious Jailbreak Makes Chaos in Honest Towns", "abstract": "With the development of large language models, they are widely used as agents in various fields. A key component of agents is memory, which stores vital information but is susceptible to jailbreak attacks. Existing research mainly focuses on single-agent attacks and shared memory attacks. However, real-world scenarios often involve independent memory. In this paper, we propose the Troublemaker Makes Chaos in Honest Town (TMCHT) task, a large-scale, multi-agent, multi-topology text-based attack evaluation framework. TMCHT involves one attacker agent attempting to mislead an entire society of agents. We identify two major challenges in multi-agent attacks: (1) Non-complete graph structure, (2) Large-scale systems. We attribute these challenges to a phenomenon we term toxicity disappearing. To address these issues, we propose an Adversarial Replication Contagious Jailbreak (ARCJ) method, which optimizes the retrieval suffix to make poisoned samples more easily retrieved and optimizes the replication suffix to make poisoned samples have contagious ability. We demonstrate the superiority of our approach in TMCHT, with 23.51%, 18.95%, and 52.93% improvements in line topology, star topology, and 100-agent settings. Encourage community attention to the security of multi-agent systems.", "date": "2024-10-21", "categories": []}
{"id": "2506.14539", "title": "Doppelganger Method: Breaking Role Consistency in LLM Agent via Prompt-based Transferable Adversarial Attack", "abstract": "Since the advent of large language models, prompt engineering now enables the rapid, low-effort creation of diverse autonomous agents that are already in widespread use. Yet this convenience raises urgent concerns about the safety, robustness, and behavioral consistency of the underlying prompts, along with the pressing challenge of preventing those prompts from being exposed to user's attempts. In this paper, we propose the ''Doppelganger method'' to demonstrate the risk of an agent being hijacked, thereby exposing system instructions and internal information. Next, we define the ''Prompt Alignment Collapse under Adversarial Transfer (PACAT)'' level to evaluate the vulnerability to this adversarial transfer attack. We also propose a ''Caution for Adversarial Transfer (CAT)'' prompt to counter the Doppelganger method. The experimental results demonstrate that the Doppelganger method can compromise the agent's consistency and expose its internal information. In contrast, CAT prompts enable effective defense against this adversarial attack.", "date": "2025-06-17", "categories": []}
{"id": "2506.20806", "title": "Poster: Enhancing GNN Robustness for Network Intrusion Detection via Agent-based Analysis", "abstract": "Graph Neural Networks (GNNs) show great promise for Network Intrusion Detection Systems (NIDS), particularly in IoT environments, but suffer performance degradation due to distribution drift and lack robustness against realistic adversarial attacks. Current robustness evaluations often rely on unrealistic synthetic perturbations and lack demonstrations on systematic analysis of different kinds of adversarial attack, which encompass both black-box and white-box scenarios. This work proposes a novel approach to enhance GNN robustness and generalization by employing Large Language Models (LLMs) in an agentic pipeline as simulated cybersecurity expert agents. These agents scrutinize graph structures derived from network flow data, identifying and potentially mitigating suspicious or adversarially perturbed elements before GNN processing. Our experiments, using a framework designed for realistic evaluation and testing with a variety of adversarial attacks including a dataset collected from physical testbed experiments, demonstrate that integrating LLM analysis can significantly improve the resilience of GNN-based NIDS against challenges, showcasing the potential of LLM agent as a complementary layer in intrusion detection architectures.", "date": "2025-06-25", "categories": []}
{"id": "2506.20816", "title": "Universal and Efficient Detection of Adversarial Data through Nonuniform Impact on Network Layers", "abstract": "Deep Neural Networks (DNNs) are notoriously vulnerable to adversarial input designs with limited noise budgets. While numerous successful attacks with subtle modifications to original input have been proposed, defense techniques against these attacks are relatively understudied. Existing defense approaches either focus on improving DNN robustness by negating the effects of perturbations or use a secondary model to detect adversarial data. Although equally important, the attack detection approach, which is studied in this work, provides a more practical defense compared to the robustness approach. We show that the existing detection methods are either ineffective against the state-of-the-art attack techniques or computationally inefficient for real-time processing. We propose a novel universal and efficient method to detect adversarial examples by analyzing the varying degrees of impact of attacks on different DNN layers. {Our method trains a lightweight regression model that predicts deeper-layer features from early-layer features, and uses the prediction error to detect adversarial samples.} Through theoretical arguments and extensive experiments, we demonstrate that our detection method is highly effective, computationally efficient for real-time processing, compatible with any DNN architecture, and applicable across different domains, such as image, video, and audio.", "date": "2025-06-25", "categories": []}
{"id": "2502.01633", "title": "Adversarial Reasoning at Jailbreaking Time", "abstract": "As large language models (LLMs) are becoming more capable and widespread, the study of their failure cases is becoming increasingly important. Recent advances in standardizing, measuring, and scaling test-time compute suggest new methodologies for optimizing models to achieve high performance on hard tasks. In this paper, we apply these advances to the task of model jailbreaking: eliciting harmful responses from aligned LLMs. We develop an adversarial reasoning approach to automatic jailbreaking that leverages a loss signal to guide the test-time compute, achieving SOTA attack success rates against many aligned LLMs, even those that aim to trade inference-time compute for adversarial robustness. Our approach introduces a new paradigm in understanding LLM vulnerabilities, laying the foundation for the development of more robust and trustworthy AI systems.", "date": "2025-02-03", "categories": []}
{"id": "2506.19871", "title": "An Attack Method for Medical Insurance Claim Fraud Detection based on Generative Adversarial Network", "abstract": "Insurance fraud detection represents a pivotal advancement in modern insurance service, providing intelligent and digitalized monitoring to enhance management and prevent fraud. It is crucial for ensuring the security and efficiency of insurance systems. Although AI and machine learning algorithms have demonstrated strong performance in detecting fraudulent claims, the absence of standardized defense mechanisms renders current systems vulnerable to emerging adversarial threats. In this paper, we propose a GAN-based approach to conduct adversarial attacks on fraud detection systems. Our results indicate that an attacker, without knowledge of the training data or internal model details, can generate fraudulent cases that are classified as legitimate with a 99\\% attack success rate (ASR). By subtly modifying real insurance records and claims, adversaries can significantly increase the fraud risk, potentially bypassing compromised detection systems. These findings underscore the urgent need to enhance the robustness of insurance fraud detection models against adversarial manipulation, thereby ensuring the stability and reliability of different insurance systems.", "date": "2025-06-22", "categories": []}
{"id": "2506.20576", "title": "Vulnerability Disclosure through Adaptive Black-Box Adversarial Attacks on NIDS", "abstract": "Adversarial attacks, wherein slight inputs are carefully crafted to mislead intelligent models, have attracted increasing attention. However, a critical gap persists between theoretical advancements and practical application, particularly in structured data like network traffic, where interdependent features complicate effective adversarial manipulations. Moreover, ambiguity in current approaches restricts reproducibility and limits progress in this field. Hence, existing defenses often fail to handle evolving adversarial attacks. This paper proposes a novel approach for black-box adversarial attacks, that addresses these limitations. Unlike prior work, which often assumes system access or relies on repeated probing, our method strictly respect black-box constraints, reducing interaction to avoid detection and better reflect real-world scenarios. We present an adaptive feature selection strategy using change-point detection and causality analysis to identify and target sensitive features to perturbations. This lightweight design ensures low computational cost and high deployability. Our comprehensive experiments show the attack's effectiveness in evading detection with minimal interaction, enhancing its adaptability and applicability in real-world scenarios. By advancing the understanding of adversarial attacks in network traffic, this work lays a foundation for developing robust defenses.", "date": "2025-06-25", "categories": []}
{"id": "2501.06533", "title": "DivTrackee versus DynTracker: Promoting Diversity in Anti-Facial Recognition against Dynamic FR Strategy", "abstract": "The widespread adoption of facial recognition (FR) models raises serious concerns about their potential misuse, motivating the development of anti-facial recognition (AFR) to protect user facial privacy. In this paper, we argue that the static FR strategy, predominantly adopted in prior literature for evaluating AFR efficacy, cannot faithfully characterize the actual capabilities of determined trackers who aim to track a specific target identity. In particular, we introduce DynTracker, a dynamic FR strategy where the model's gallery database is iteratively updated with newly recognized target identity images. Surprisingly, such a simple approach renders all the existing AFR protections ineffective. To mitigate the privacy threats posed by DynTracker, we advocate for explicitly promoting diversity in the AFR-protected images. We hypothesize that the lack of diversity is the primary cause of the failure of existing AFR methods. Specifically, we develop DivTrackee, a novel method for crafting diverse AFR protections that builds upon a text-guided image generation framework and diversity-promoting adversarial losses. Through comprehensive experiments on various image benchmarks and feature extractors, we demonstrate DynTracker's strength in breaking existing AFR methods and the superiority of DivTrackee in preventing user facial images from being identified by dynamic FR strategies. We believe our work can act as an important initial step towards developing more effective AFR methods for protecting user facial privacy against determined trackers.", "date": "2025-01-11", "categories": []}
{"id": "2506.18591", "title": "SpaNN: Detecting Multiple Adversarial Patches on CNNs by Spanning Saliency Thresholds", "abstract": "State-of-the-art convolutional neural network models for object detection and image classification are vulnerable to physically realizable adversarial perturbations, such as patch attacks. Existing defenses have focused, implicitly or explicitly, on single-patch attacks, leaving their sensitivity to the number of patches as an open question or rendering them computationally infeasible or inefficient against attacks consisting of multiple patches in the worst cases. In this work, we propose SpaNN, an attack detector whose computational complexity is independent of the expected number of adversarial patches. The key novelty of the proposed detector is that it builds an ensemble of binarized feature maps by applying a set of saliency thresholds to the neural activations of the first convolutional layer of the victim model. It then performs clustering on the ensemble and uses the cluster features as the input to a classifier for attack detection. Contrary to existing detectors, SpaNN does not rely on a fixed saliency threshold for identifying adversarial regions, which makes it robust against white box adversarial attacks. We evaluate SpaNN on four widely used data sets for object detection and classification, and our results show that SpaNN outperforms state-of-the-art defenses by up to 11 and 27 percentage points in the case of object detection and the case of image classification, respectively. Our code is available at https://github.com/gerkbyrd/SpaNN.", "date": "2025-06-23", "categories": []}
{"id": "2305.17528", "title": "Two Heads are Actually Better than One: Towards Better Adversarial Robustness via Transduction and Rejection", "abstract": "Both transduction and rejection have emerged as important techniques for defending against adversarial perturbations. A recent work by Goldwasser et al. showed that rejection combined with transduction can give provable guarantees (for certain problems) that cannot be achieved otherwise. Nevertheless, under recent strong adversarial attacks, their work was shown to have low performance in a practical deep-learning setting. In this paper, we take a step towards realizing the promise of transduction+rejection in more realistic scenarios. Our key observation is that a novel application of a reduction technique by Tram\\`er, which was until now only used to demonstrate the vulnerability of certain defenses, can be used to actually construct effective defenses. Theoretically, we show that a careful application of this technique in the transductive setting can give significantly improved sample-complexity for robust generalization. Our theory guides us to design a new transductive algorithm for learning a selective model; extensive experiments using state of the art attacks show that our approach provides significantly better robust accuracy (81.6% on CIFAR-10 and 57.9% on CIFAR-100 under $l_\\infty$ with budget 8/255) than existing techniques.", "date": "2023-05-27", "categories": []}
{"id": "2401.01752", "title": "FullLoRA: Efficiently Boosting the Robustness of Pretrained Vision Transformers", "abstract": "In recent years, the Vision Transformer (ViT) model has gradually become mainstream in various computer vision tasks, and the robustness of the model has received increasing attention. However, existing large models tend to prioritize performance during training, potentially neglecting the robustness, which may lead to serious security concerns. In this paper, we establish a new challenge: exploring how to use a small number of additional parameters for adversarial finetuning to quickly and effectively enhance the adversarial robustness of a standardly trained model. To address this challenge, we develop novel LNLoRA module, incorporating a learnable layer normalization before the conventional LoRA module, which helps mitigate magnitude differences in parameters between the adversarial and standard training paradigms. Furthermore, we propose the FullLoRA framework by integrating the learnable LNLoRA modules into all key components of ViT-based models while keeping the pretrained model frozen, which can significantly improve the model robustness via adversarial finetuning in a parameter-efficient manner. Extensive experiments on several datasets demonstrate the superiority of our proposed FullLoRA framework. It achieves comparable robustness with full finetuning while only requiring about 5\\% of the learnable parameters. This also effectively addresses concerns regarding extra model storage space and enormous training time caused by adversarial finetuning.", "date": "2024-01-03", "categories": []}
{"id": "2506.17250", "title": "Towards Interpretable Adversarial Examples via Sparse Adversarial Attack", "abstract": "Sparse attacks are to optimize the magnitude of adversarial perturbations for fooling deep neural networks (DNNs) involving only a few perturbed pixels (i.e., under the l0 constraint), suitable for interpreting the vulnerability of DNNs. However, existing solutions fail to yield interpretable adversarial examples due to their poor sparsity. Worse still, they often struggle with heavy computational overhead, poor transferability, and weak attack strength. In this paper, we aim to develop a sparse attack for understanding the vulnerability of CNNs by minimizing the magnitude of initial perturbations under the l0 constraint, to overcome the existing drawbacks while achieving a fast, transferable, and strong attack to DNNs. In particular, a novel and theoretical sound parameterization technique is introduced to approximate the NP-hard l0 optimization problem, making directly optimizing sparse perturbations computationally feasible. Besides, a novel loss function is designed to augment initial perturbations by maximizing the adversary property and minimizing the number of perturbed pixels simultaneously. Extensive experiments are conducted to demonstrate that our approach, with theoretical performance guarantees, outperforms state-of-the-art sparse attacks in terms of computational overhead, transferability, and attack strength, expecting to serve as a benchmark for evaluating the robustness of DNNs. In addition, theoretical and empirical results validate that our approach yields sparser adversarial examples, empowering us to discover two categories of noises, i.e., \"obscuring noise\" and \"leading noise\", which will help interpret how adversarial perturbation misleads the classifiers into incorrect predictions. Our code is available at https://github.com/fudong03/SparseAttack.", "date": "2025-06-08", "categories": []}
{"id": "2506.17350", "title": "CUBA: Controlled Untargeted Backdoor Attack against Deep Neural Networks", "abstract": "Backdoor attacks have emerged as a critical security threat against deep neural networks in recent years. The majority of existing backdoor attacks focus on targeted backdoor attacks, where trigger is strongly associated to specific malicious behavior. Various backdoor detection methods depend on this inherent property and shows effective results in identifying and mitigating such targeted attacks. However, a purely untargeted attack in backdoor scenarios is, in some sense, self-weakening, since the target nature is what makes backdoor attacks so powerful. In light of this, we introduce a novel Constrained Untargeted Backdoor Attack (CUBA), which combines the flexibility of untargeted attacks with the intentionality of targeted attacks. The compromised model, when presented with backdoor images, will classify them into random classes within a constrained range of target classes selected by the attacker. This combination of randomness and determinedness enables the proposed untargeted backdoor attack to natively circumvent existing backdoor defense methods. To implement the untargeted backdoor attack under controlled flexibility, we propose to apply logit normalization on cross-entropy loss with flipped one-hot labels. By constraining the logit during training, the compromised model will show a uniform distribution across selected target classes, resulting in controlled untargeted attack. Extensive experiments demonstrate the effectiveness of the proposed CUBA on different datasets.", "date": "2025-06-20", "categories": []}
{"id": "2506.17621", "title": "Exploiting Efficiency Vulnerabilities in Dynamic Deep Learning Systems", "abstract": "The growing deployment of deep learning models in real-world environments has intensified the need for efficient inference under strict latency and resource constraints. To meet these demands, dynamic deep learning systems (DDLSs) have emerged, offering input-adaptive computation to optimize runtime efficiency. While these systems succeed in reducing cost, their dynamic nature introduces subtle and underexplored security risks. In particular, input-dependent execution pathways create opportunities for adversaries to degrade efficiency, resulting in excessive latency, energy usage, and potential denial-of-service in time-sensitive deployments. This work investigates the security implications of dynamic behaviors in DDLSs and reveals how current systems expose efficiency vulnerabilities exploitable by adversarial inputs. Through a survey of existing attack strategies, we identify gaps in the coverage of emerging model architectures and limitations in current defense mechanisms. Building on these insights, we propose to examine the feasibility of efficiency attacks on modern DDLSs and develop targeted defenses to preserve robustness under adversarial conditions.", "date": "2025-06-21", "categories": []}
{"id": "2506.18304", "title": "Sharpening the Spear: Adaptive Expert-Guided Adversarial Attack Against DRL-based Autonomous Driving Policies", "abstract": "Deep reinforcement learning (DRL) has emerged as a promising paradigm for autonomous driving. However, despite their advanced capabilities, DRL-based policies remain highly vulnerable to adversarial attacks, posing serious safety risks in real-world deployments. Investigating such attacks is crucial for revealing policy vulnerabilities and guiding the development of more robust autonomous systems. While prior attack methods have made notable progress, they still face several challenges: 1) they often rely on high-frequency attacks, yet critical attack opportunities are typically context-dependent and temporally sparse, resulting in inefficient attack patterns; 2) restricting attack frequency can improve efficiency but often results in unstable training due to the adversary's limited exploration. To address these challenges, we propose an adaptive expert-guided adversarial attack method that enhances both the stability and efficiency of attack policy training. Our method first derives an expert policy from successful attack demonstrations using imitation learning, strengthened by an ensemble Mixture-of-Experts architecture for robust generalization across scenarios. This expert policy then guides a DRL-based adversary through a KL-divergence regularization term. Due to the diversity of scenarios, expert policies may be imperfect. To address this, we further introduce a performance-aware annealing strategy that gradually reduces reliance on the expert as the adversary improves. Extensive experiments demonstrate that our method achieves outperforms existing approaches in terms of collision rate, attack efficiency, and training stability, especially in cases where the expert policy is sub-optimal.", "date": "2025-06-23", "categories": []}
{"id": "2311.16487", "title": "On the Robustness of Decision-Focused Learning", "abstract": "Decision-Focused Learning (DFL) is an emerging learning paradigm that tackles the task of training a machine learning (ML) model to predict missing parameters of an incomplete optimization problem, where the missing parameters are predicted. DFL trains an ML model in an end-to-end system, by integrating the prediction and optimization tasks, providing better alignment of the training and testing objectives. DFL has shown a lot of promise and holds the capacity to revolutionize decision-making in many real-world applications. However, very little is known about the performance of these models under adversarial attacks. We adopt ten unique DFL methods and benchmark their performance under two distinctly focused attacks adapted towards the Predict-then-Optimize problem setting. Our study proposes the hypothesis that the robustness of a model is highly correlated with its ability to find predictions that lead to optimal decisions without deviating from the ground-truth label. Furthermore, we provide insight into how to target the models that violate this condition and show how these models respond differently depending on the achieved optimality at the end of their training cycles.", "date": "2023-11-28", "categories": []}
{"id": "2406.03458", "title": "Distributional Adversarial Loss", "abstract": "We initiate the study of a new notion of adversarial loss which we call distributional adversarial loss. In this notion, we assume for each original example, the allowed adversarial perturbation set is a family of distributions, and the adversarial loss over each example is the maximum loss over all the associated distributions. The goal is to minimize the overall adversarial loss. We show sample complexity bounds in the PAC-learning setting for our notion of adversarial loss. Our notion of adversarial loss contrasts the prior work on robust learning that considers a set of points, not distributions, as the perturbation set of each clean example. As an application of our approach, we show how to unify the two lines of work on randomized smoothing and robust learning in the PAC-learning setting and derive sample complexity bounds for randomized smoothing methods.\n  Furthermore, we investigate the role of randomness in achieving robustness against adversarial attacks. We show a general derandomization technique that preserves the extent of a randomized classifier's robustness against adversarial attacks and show its effectiveness empirically.", "date": "2024-06-05", "categories": []}
{"id": "2502.09271", "title": "LiSA: Leveraging Link Recommender to Attack Graph Neural Networks via Subgraph Injection", "abstract": "Graph Neural Networks (GNNs) have demonstrated remarkable proficiency in modeling data with graph structures, yet recent research reveals their susceptibility to adversarial attacks. Traditional attack methodologies, which rely on manipulating the original graph or adding links to artificially created nodes, often prove impractical in real-world settings. This paper introduces a novel adversarial scenario involving the injection of an isolated subgraph to deceive both the link recommender and the node classifier within a GNN system. Specifically, the link recommender is mislead to propose links between targeted victim nodes and the subgraph, encouraging users to unintentionally establish connections and that would degrade the node classification accuracy, thereby facilitating a successful attack. To address this, we present the LiSA framework, which employs a dual surrogate model and bi-level optimization to simultaneously meet two adversarial objectives. Extensive experiments on real-world datasets demonstrate the effectiveness of our method.", "date": "2025-02-13", "categories": []}
{"id": "2506.11611", "title": "KCES: Training-Free Defense for Robust Graph Neural Networks via Kernel Complexity", "abstract": "Graph Neural Networks (GNNs) have achieved impressive success across a wide range of graph-based tasks, yet they remain highly vulnerable to small, imperceptible perturbations and adversarial attacks. Although numerous defense methods have been proposed to address these vulnerabilities, many rely on heuristic metrics, overfit to specific attack patterns, and suffer from high computational complexity. In this paper, we propose Kernel Complexity-Based Edge Sanitization (KCES), a training-free, model-agnostic defense framework. KCES leverages Graph Kernel Complexity (GKC), a novel metric derived from the graph's Gram matrix that characterizes GNN generalization via its test error bound. Building on GKC, we define a KC score for each edge, measuring the change in GKC when the edge is removed. Edges with high KC scores, typically introduced by adversarial perturbations, are pruned to mitigate their harmful effects, thereby enhancing GNNs' robustness. KCES can also be seamlessly integrated with existing defense strategies as a plug-and-play module without requiring training. Theoretical analysis and extensive experiments demonstrate that KCES consistently enhances GNN robustness, outperforms state-of-the-art baselines, and amplifies the effectiveness of existing defenses, offering a principled and efficient solution for securing GNNs.", "date": "2025-06-13", "categories": []}
{"id": "2506.15988", "title": "Adversarial Attacks and Detection in Visual Place Recognition for Safer Robot Navigation", "abstract": "Stand-alone Visual Place Recognition (VPR) systems have little defence against a well-designed adversarial attack, which can lead to disastrous consequences when deployed for robot navigation. This paper extensively analyzes the effect of four adversarial attacks common in other perception tasks and four novel VPR-specific attacks on VPR localization performance. We then propose how to close the loop between VPR, an Adversarial Attack Detector (AAD), and active navigation decisions by demonstrating the performance benefit of simulated AADs in a novel experiment paradigm -- which we detail for the robotics community to use as a system framework. In the proposed experiment paradigm, we see the addition of AADs across a range of detection accuracies can improve performance over baseline; demonstrating a significant improvement -- such as a ~50% reduction in the mean along-track localization error -- can be achieved with True Positive and False Positive detection rates of only 75% and up to 25% respectively. We examine a variety of metrics including: Along-Track Error, Percentage of Time Attacked, Percentage of Time in an `Unsafe' State, and Longest Continuous Time Under Attack. Expanding further on these results, we provide the first investigation into the efficacy of the Fast Gradient Sign Method (FGSM) adversarial attack for VPR. The analysis in this work highlights the need for AADs in real-world systems for trustworthy navigation, and informs quantitative requirements for system design.", "date": "2025-06-19", "categories": []}
{"id": "2506.16407", "title": "Robustness Evaluation of OCR-based Visual Document Understanding under Multi-Modal Adversarial Attacks", "abstract": "Visual Document Understanding (VDU) systems have achieved strong performance in information extraction by integrating textual, layout, and visual signals. However, their robustness under realistic adversarial perturbations remains insufficiently explored. We introduce the first unified framework for generating and evaluating multi-modal adversarial attacks on OCR-based VDU models. Our method covers six gradient-based layout attack scenarios, incorporating manipulations of OCR bounding boxes, pixels, and texts across both word and line granularities, with constraints on layout perturbation budget (e.g., IoU >= 0.6) to preserve plausibility.\n  Experimental results across four datasets (FUNSD, CORD, SROIE, DocVQA) and six model families demonstrate that line-level attacks and compound perturbations (BBox + Pixel + Text) yield the most severe performance degradation. Projected Gradient Descent (PGD)-based BBox perturbations outperform random-shift baselines in all investigated models. Ablation studies further validate the impact of layout budget, text modification, and adversarial transferability.", "date": "2025-06-19", "categories": []}
{"id": "2506.16760", "title": "Cross-Modal Obfuscation for Jailbreak Attacks on Large Vision-Language Models", "abstract": "Large Vision-Language Models (LVLMs) demonstrate exceptional performance across multimodal tasks, yet remain vulnerable to jailbreak attacks that bypass built-in safety mechanisms to elicit restricted content generation. Existing black-box jailbreak methods primarily rely on adversarial textual prompts or image perturbations, yet these approaches are highly detectable by standard content filtering systems and exhibit low query and computational efficiency. In this work, we present Cross-modal Adversarial Multimodal Obfuscation (CAMO), a novel black-box jailbreak attack framework that decomposes malicious prompts into semantically benign visual and textual fragments. By leveraging LVLMs' cross-modal reasoning abilities, CAMO covertly reconstructs harmful instructions through multi-step reasoning, evading conventional detection mechanisms. Our approach supports adjustable reasoning complexity and requires significantly fewer queries than prior attacks, enabling both stealth and efficiency. Comprehensive evaluations conducted on leading LVLMs validate CAMO's effectiveness, showcasing robust performance and strong cross-model transferability. These results underscore significant vulnerabilities in current built-in safety mechanisms, emphasizing an urgent need for advanced, alignment-aware security and safety solutions in vision-language systems.", "date": "2025-06-20", "categories": []}
{"id": "2407.16205", "title": "LLMs can be Dangerous Reasoners: Analyzing-based Jailbreak Attack on Large Language Models", "abstract": "The rapid development of Large Language Models (LLMs) has brought impressive advancements across various tasks. However, despite these achievements, LLMs still pose inherent safety risks, especially in the context of jailbreak attacks. Most existing jailbreak methods follow an input-level manipulation paradigm to bypass safety mechanisms. Yet, as alignment techniques improve, such attacks are becoming increasingly detectable. In this work, we identify an underexplored threat vector: the model's internal reasoning process, which can be manipulated to elicit harmful outputs in a more stealthy way. To explore this overlooked attack surface, we propose a novel black-box jailbreak attack method, Analyzing-based Jailbreak (ABJ). ABJ comprises two independent attack paths: textual and visual reasoning attacks, which exploit the model's multimodal reasoning capabilities to bypass safety mechanisms, comprehensively exposing vulnerabilities in its reasoning chain. We conduct extensive experiments on ABJ across various open-source and closed-source LLMs, VLMs, and RLMs. In particular, ABJ achieves high attack success rate (ASR) (82.1% on GPT-4o-2024-11-20) with exceptional attack efficiency (AE) among all target models, showcasing its remarkable attack effectiveness, transferability, and efficiency. Our work reveals a new type of safety risk and highlights the urgent need to mitigate implicit vulnerabilities in the model's reasoning process.", "date": "2024-07-23", "categories": []}
{"id": "2409.07448", "title": "A Novel Perturb-ability Score to Mitigate Evasion Adversarial Attacks on Flow-Based ML-NIDS", "abstract": "As network security threats evolve, safeguarding flow-based Machine Learning (ML)-based Network Intrusion Detection Systems (NIDS) from evasion adversarial attacks is crucial. This paper introduces the notion of feature perturb-ability and presents a novel Perturb-ability Score (PS), which quantifies how susceptible NIDS features are to manipulation in the problem-space by an attacker. PS thereby identifies features structurally resistant to evasion attacks in flow-based ML-NIDS due to the semantics of network traffic fields, as these features are constrained by domain-specific limitations and correlations. Consequently, attempts to manipulate such features would likely either compromise the attack's malicious functionality, render the traffic invalid for processing, or potentially both outcomes simultaneously.\n  We introduce and demonstrate the effectiveness of our PS-enabled defenses, PS-guided feature selection and PS-guided feature masking, in enhancing flow-based NIDS resilience. Experimental results across various ML-based NIDS models and public datasets show that discarding or masking highly manipulatable features (high-PS features) can maintain solid detection performance while significantly reducing vulnerability to evasion adversarial attacks. Our findings confirm that PS effectively identifies flow-based NIDS features susceptible to problem-space perturbations. This novel approach leverages problem-space NIDS domain constraints as lightweight universal defense mechanisms against evasion adversarial attacks targeting flow-based ML-NIDS.", "date": "2024-09-11", "categories": []}
{"id": "2506.15506", "title": "Insights on Adversarial Attacks for Tabular Machine Learning via a Systematic Literature Review", "abstract": "Adversarial attacks in machine learning have been extensively reviewed in areas like computer vision and NLP, but research on tabular data remains scattered. This paper provides the first systematic literature review focused on adversarial attacks targeting tabular machine learning models. We highlight key trends, categorize attack strategies and analyze how they address practical considerations for real-world applicability. Additionally, we outline current challenges and open research questions. By offering a clear and structured overview, this review aims to guide future efforts in understanding and addressing adversarial vulnerabilities in tabular machine learning.", "date": "2025-06-18", "categories": []}
{"id": "2505.06679", "title": "T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks", "abstract": "In recent years, fueled by the rapid advancement of diffusion models, text-to-video (T2V) generation models have achieved remarkable progress, with notable examples including Pika, Luma, Kling, and Open-Sora. Although these models exhibit impressive generative capabilities, they also expose significant security risks due to their vulnerability to jailbreak attacks, where the models are manipulated to produce unsafe content such as pornography, violence, or discrimination. Existing works such as T2VSafetyBench provide preliminary benchmarks for safety evaluation, but lack systematic methods for thoroughly exploring model vulnerabilities. To address this gap, we are the first to formalize the T2V jailbreak attack as a discrete optimization problem and propose a joint objective-based optimization framework, called T2V-OptJail. This framework consists of two key optimization goals: bypassing the built-in safety filtering mechanisms to increase the attack success rate, preserving semantic consistency between the adversarial prompt and the unsafe input prompt, as well as between the generated video and the unsafe input prompt, to enhance content controllability. In addition, we introduce an iterative optimization strategy guided by prompt variants, where multiple semantically equivalent candidates are generated in each round, and their scores are aggregated to robustly guide the search toward optimal adversarial prompts. We conduct large-scale experiments on several T2V models, covering both open-source models and real commercial closed-source models. The experimental results show that the proposed method improves 11.4% and 10.0% over the existing state-of-the-art method in terms of attack success rate assessed by GPT-4, attack success rate assessed by human accessors, respectively, verifying the significant advantages of the method in terms of attack effectiveness and content control.", "date": "2025-05-10", "categories": []}
{"id": "2506.14374", "title": "Excessive Reasoning Attack on Reasoning LLMs", "abstract": "Recent reasoning large language models (LLMs), such as OpenAI o1 and DeepSeek-R1, exhibit strong performance on complex tasks through test-time inference scaling. However, prior studies have shown that these models often incur significant computational costs due to excessive reasoning, such as frequent switching between reasoning trajectories (e.g., underthinking) or redundant reasoning on simple questions (e.g., overthinking). In this work, we expose a novel threat: adversarial inputs can be crafted to exploit excessive reasoning behaviors and substantially increase computational overhead without compromising model utility. Therefore, we propose a novel loss framework consisting of three components: (1) Priority Cross-Entropy Loss, a modification of the standard cross-entropy objective that emphasizes key tokens by leveraging the autoregressive nature of LMs; (2) Excessive Reasoning Loss, which encourages the model to initiate additional reasoning paths during inference; and (3) Delayed Termination Loss, which is designed to extend the reasoning process and defer the generation of final outputs. We optimize and evaluate our attack for the GSM8K and ORCA datasets on DeepSeek-R1-Distill-LLaMA and DeepSeek-R1-Distill-Qwen. Empirical results demonstrate a 3x to 9x increase in reasoning length with comparable utility performance. Furthermore, our crafted adversarial inputs exhibit transferability, inducing computational overhead in o3-mini, o1-mini, DeepSeek-R1, and QWQ models.", "date": "2025-06-17", "categories": []}
{"id": "2506.14582", "title": "Busting the Paper Ballot: Voting Meets Adversarial Machine Learning", "abstract": "We show the security risk associated with using machine learning classifiers in United States election tabulators. The central classification task in election tabulation is deciding whether a mark does or does not appear on a bubble associated to an alternative in a contest on the ballot. Barretto et al. (E-Vote-ID 2021) reported that convolutional neural networks are a viable option in this field, as they outperform simple feature-based classifiers.\n  Our contributions to election security can be divided into four parts. To demonstrate and analyze the hypothetical vulnerability of machine learning models on election tabulators, we first introduce four new ballot datasets. Second, we train and test a variety of different models on our new datasets. These models include support vector machines, convolutional neural networks (a basic CNN, VGG and ResNet), and vision transformers (Twins and CaiT). Third, using our new datasets and trained models, we demonstrate that traditional white box attacks are ineffective in the voting domain due to gradient masking. Our analyses further reveal that gradient masking is a product of numerical instability. We use a modified difference of logits ratio loss to overcome this issue (Croce and Hein, ICML 2020). Fourth, in the physical world, we conduct attacks with the adversarial examples generated using our new methods. In traditional adversarial machine learning, a high (50% or greater) attack success rate is ideal. However, for certain elections, even a 5% attack success rate can flip the outcome of a race. We show such an impact is possible in the physical domain. We thoroughly discuss attack realism, and the challenges and practicality associated with printing and scanning ballot adversarial examples.", "date": "2025-06-17", "categories": []}
{"id": "2410.17442", "title": "Detecting Adversarial Examples", "abstract": "Deep Neural Networks (DNNs) have been shown to be vulnerable to adversarial examples. While numerous successful adversarial attacks have been proposed, defenses against these attacks remain relatively understudied. Existing defense approaches either focus on negating the effects of perturbations caused by the attacks to restore the DNNs' original predictions or use a secondary model to detect adversarial examples. However, these methods often become ineffective due to the continuous advancements in attack techniques. We propose a novel universal and lightweight method to detect adversarial examples by analyzing the layer outputs of DNNs. Our method trains a lightweight regression model that predicts deeper-layer features from early-layer features, and uses the prediction error to detect adversarial samples. Through theoretical justification and extensive experiments, we demonstrate that our detection method is highly effective, compatible with any DNN architecture, and applicable across different domains, such as image, video, and audio.", "date": "2024-10-22", "categories": []}
{"id": "2410.23142", "title": "FAIR-TAT: Improving Model Fairness Using Targeted Adversarial Training", "abstract": "Deep neural networks are susceptible to adversarial attacks and common corruptions, which undermine their robustness. In order to enhance model resilience against such challenges, Adversarial Training (AT) has emerged as a prominent solution. Nevertheless, adversarial robustness is often attained at the expense of model fairness during AT, i.e., disparity in class-wise robustness of the model. While distinctive classes become more robust towards such adversaries, hard to detect classes suffer. Recently, research has focused on improving model fairness specifically for perturbed images, overlooking the accuracy of the most likely non-perturbed data. Additionally, despite their robustness against the adversaries encountered during model training, state-of-the-art adversarial trained models have difficulty maintaining robustness and fairness when confronted with diverse adversarial threats or common corruptions. In this work, we address the above concerns by introducing a novel approach called Fair Targeted Adversarial Training (FAIR-TAT). We show that using targeted adversarial attacks for adversarial training (instead of untargeted attacks) can allow for more favorable trade-offs with respect to adversarial fairness. Empirical results validate the efficacy of our approach.", "date": "2024-10-30", "categories": []}
{"id": "2412.10198", "title": "From Allies to Adversaries: Manipulating LLM Tool-Calling through Adversarial Injection", "abstract": "Tool-calling has changed Large Language Model (LLM) applications by integrating external tools, significantly enhancing their functionality across diverse tasks. However, this integration also introduces new security vulnerabilities, particularly in the tool scheduling mechanisms of LLM, which have not been extensively studied. To fill this gap, we present ToolCommander, a novel framework designed to exploit vulnerabilities in LLM tool-calling systems through adversarial tool injection. Our framework employs a well-designed two-stage attack strategy. Firstly, it injects malicious tools to collect user queries, then dynamically updates the injected tools based on the stolen information to enhance subsequent attacks. These stages enable ToolCommander to execute privacy theft, launch denial-of-service attacks, and even manipulate business competition by triggering unscheduled tool-calling. Notably, the ASR reaches 91.67% for privacy theft and hits 100% for denial-of-service and unscheduled tool calling in certain cases. Our work demonstrates that these vulnerabilities can lead to severe consequences beyond simple misuse of tool-calling systems, underscoring the urgent need for robust defensive strategies to secure LLM Tool-calling systems.", "date": "2024-12-13", "categories": []}
{"id": "2412.11934", "title": "Stepwise Reasoning Error Disruption Attack of LLMs", "abstract": "Large language models (LLMs) have made remarkable strides in complex reasoning tasks, but their safety and robustness in reasoning processes remain underexplored. Existing attacks on LLM reasoning are constrained by specific settings or lack of imperceptibility, limiting their feasibility and generalizability. To address these challenges, we propose the Stepwise rEasoning Error Disruption (SEED) attack, which subtly injects errors into prior reasoning steps to mislead the model into producing incorrect subsequent reasoning and final answers. Unlike previous methods, SEED is compatible with zero-shot and few-shot settings, maintains the natural reasoning flow, and ensures covert execution without modifying the instruction. Extensive experiments on four datasets across four different models demonstrate SEED's effectiveness, revealing the vulnerabilities of LLMs to disruptions in reasoning processes. These findings underscore the need for greater attention to the robustness of LLM reasoning to ensure safety in practical applications. Our code is available at: https://github.com/Applied-Machine-Learning-Lab/SEED-Attack.", "date": "2024-12-16", "categories": []}
{"id": "2502.07225", "title": "CAT: Contrastive Adversarial Training for Evaluating the Robustness of Protective Perturbations in Latent Diffusion Models", "abstract": "Latent diffusion models have recently demonstrated superior capabilities in many downstream image synthesis tasks. However, customization of latent diffusion models using unauthorized data can severely compromise the privacy and intellectual property rights of data owners. Adversarial examples as protective perturbations have been developed to defend against unauthorized data usage by introducing imperceptible noise to customization samples, preventing diffusion models from effectively learning them. In this paper, we first reveal that the primary reason adversarial examples are effective as protective perturbations in latent diffusion models is the distortion of their latent representations, as demonstrated through qualitative and quantitative experiments. We then propose the Contrastive Adversarial Training (CAT) utilizing lightweight adapters as an adaptive attack against these protection methods, highlighting their lack of robustness. Extensive experiments demonstrate that our CAT method significantly reduces the effectiveness of protective perturbations in customization, urging the community to reconsider and improve the robustness of existing protective perturbations. The code is available at https://github.com/senp98/CAT.", "date": "2025-02-11", "categories": []}
{"id": "2504.20965", "title": "AegisLLM: Scaling Agentic Systems for Self-Reflective Defense in LLM Security", "abstract": "We introduce AegisLLM, a cooperative multi-agent defense against adversarial attacks and information leakage. In AegisLLM, a structured workflow of autonomous agents - orchestrator, deflector, responder, and evaluator - collaborate to ensure safe and compliant LLM outputs, while self-improving over time through prompt optimization. We show that scaling agentic reasoning system at test-time - both by incorporating additional agent roles and by leveraging automated prompt optimization (such as DSPy)- substantially enhances robustness without compromising model utility. This test-time defense enables real-time adaptability to evolving attacks, without requiring model retraining. Comprehensive evaluations across key threat scenarios, including unlearning and jailbreaking, demonstrate the effectiveness of AegisLLM. On the WMDP unlearning benchmark, AegisLLM achieves near-perfect unlearning with only 20 training examples and fewer than 300 LM calls. For jailbreaking benchmarks, we achieve 51% improvement compared to the base model on StrongReject, with false refusal rates of only 7.9% on PHTest compared to 18-55% for comparable methods. Our results highlight the advantages of adaptive, agentic reasoning over static defenses, establishing AegisLLM as a strong runtime alternative to traditional approaches based on model modifications. Code is available at https://github.com/zikuicai/aegisllm", "date": "2025-04-29", "categories": []}
{"id": "2506.12613", "title": "Existence of Adversarial Examples for Random Convolutional Networks via Isoperimetric Inequalities on $\\mathbb{so}(d)$", "abstract": "We show that adversarial examples exist for various random convolutional networks, and furthermore, that this is a relatively simple consequence of the isoperimetric inequality on the special orthogonal group $\\mathbb{so}(d)$. This extends and simplifies a recent line of work which shows similar results for random fully connected networks.", "date": "2025-06-14", "categories": []}
{"id": "2506.12685", "title": "Alphabet Index Mapping: Jailbreaking LLMs through Semantic Dissimilarity", "abstract": "Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their susceptibility to adversarial attacks, particularly jailbreaking, poses significant safety and ethical concerns. While numerous jailbreak methods exist, many suffer from computational expense, high token usage, or complex decoding schemes. Liu et al. (2024) introduced FlipAttack, a black-box method that achieves high attack success rates (ASR) through simple prompt manipulation. This paper investigates the underlying mechanisms of FlipAttack's effectiveness by analyzing the semantic changes induced by its flipping modes. We hypothesize that semantic dissimilarity between original and manipulated prompts is inversely correlated with ASR. To test this, we examine embedding space visualizations (UMAP, KDE) and cosine similarities for FlipAttack's modes. Furthermore, we introduce a novel adversarial attack, Alphabet Index Mapping (AIM), designed to maximize semantic dissimilarity while maintaining simple decodability. Experiments on GPT-4 using a subset of AdvBench show AIM and its variant AIM+FWO achieve a 94% ASR, outperforming FlipAttack and other methods on this subset. Our findings suggest that while high semantic dissimilarity is crucial, a balance with decoding simplicity is key for successful jailbreaking. This work contributes to a deeper understanding of adversarial prompt mechanics and offers a new, effective jailbreak technique.", "date": "2025-06-15", "categories": []}
{"id": "2506.12875", "title": "Intriguing Frequency Interpretation of Adversarial Robustness for CNNs and ViTs", "abstract": "Adversarial examples have attracted significant attention over the years, yet understanding their frequency-based characteristics remains insufficient. In this paper, we investigate the intriguing properties of adversarial examples in the frequency domain for the image classification task, with the following key findings. (1) As the high-frequency components increase, the performance gap between adversarial and natural examples becomes increasingly pronounced. (2) The model performance against filtered adversarial examples initially increases to a peak and declines to its inherent robustness. (3) In Convolutional Neural Networks, mid- and high-frequency components of adversarial examples exhibit their attack capabilities, while in Transformers, low- and mid-frequency components of adversarial examples are particularly effective. These results suggest that different network architectures have different frequency preferences and that differences in frequency components between adversarial and natural examples may directly influence model robustness. Based on our findings, we further conclude with three useful proposals that serve as a valuable reference to the AI model security community.", "date": "2025-06-15", "categories": []}
{"id": "2506.13276", "title": "Navigating the Black Box: Leveraging LLMs for Effective Text-Level Graph Injection Attacks", "abstract": "Text-attributed graphs (TAGs) integrate textual data with graph structures, providing valuable insights in applications such as social network analysis and recommendation systems. Graph Neural Networks (GNNs) effectively capture both topological structure and textual information in TAGs but are vulnerable to adversarial attacks. Existing graph injection attack (GIA) methods assume that attackers can directly manipulate the embedding layer, producing non-explainable node embeddings. Furthermore, the effectiveness of these attacks often relies on surrogate models with high training costs. Thus, this paper introduces ATAG-LLM, a novel black-box GIA framework tailored for TAGs. Our approach leverages large language models (LLMs) to generate interpretable text-level node attributes directly, ensuring attacks remain feasible in real-world scenarios. We design strategies for LLM prompting that balance exploration and reliability to guide text generation, and propose a similarity assessment method to evaluate attack text effectiveness in disrupting graph homophily. This method efficiently perturbs the target node with minimal training costs in a strict black-box setting, ensuring a text-level graph injection attack for TAGs. Experiments on real-world TAG datasets validate the superior performance of ATAG-LLM compared to state-of-the-art embedding-level and text-level attack methods.", "date": "2025-06-16", "categories": []}
{"id": "2407.08970", "title": "Self-interpreting Adversarial Images", "abstract": "We introduce a new type of indirect, cross-modal injection attacks against visual language models that enable creation of self-interpreting images. These images contain hidden \"meta-instructions\" that control how models answer users' questions about the image and steer models' outputs to express an adversary-chosen style, sentiment, or point of view.\n  Self-interpreting images act as soft prompts, conditioning the model to satisfy the adversary's (meta-)objective while still producing answers based on the image's visual content. Meta-instructions are thus a stronger form of prompt injection. Adversarial images look natural and the model's answers are coherent and plausible, yet they also follow the adversary-chosen interpretation, e.g., political spin, or even objectives that are not achievable with explicit text instructions.\n  We evaluate the efficacy of self-interpreting images for a variety of models, interpretations, and user prompts. We describe how these attacks could cause harm by enabling creation of self-interpreting content that carries spam, misinformation, or spin. Finally, we discuss defenses.", "date": "2024-07-12", "categories": []}
{"id": "2408.02509", "title": "Black-Box Adversarial Attacks on LLM-Based Code Completion", "abstract": "Modern code completion engines, powered by large language models (LLMs), assist millions of developers with their strong capabilities to generate functionally correct code. Due to this popularity, it is crucial to investigate the security implications of relying on LLM-based code completion. In this work, we demonstrate that state-of-the-art black-box LLM-based code completion engines can be stealthily biased by adversaries to significantly increase their rate of insecure code generation. We present the first attack, named INSEC, that achieves this goal. INSEC works by injecting an attack string as a short comment in the completion input. The attack string is crafted through a query-based optimization procedure starting from a set of carefully designed initialization schemes. We demonstrate INSEC's broad applicability and effectiveness by evaluating it on various state-of-the-art open-source models and black-box commercial services (e.g., OpenAI API and GitHub Copilot). On a diverse set of security-critical test cases, covering 16 CWEs across 5 programming languages, INSEC increases the rate of generated insecure code by more than 50%, while maintaining the functional correctness of generated code. We consider INSEC practical -- it requires low resources and costs less than 10 US dollars to develop on commodity hardware. Moreover, we showcase the attack's real-world deployability, by developing an IDE plug-in that stealthily injects INSEC into the GitHub Copilot extension.", "date": "2024-08-05", "categories": []}
{"id": "2506.11125", "title": "ASRJam: Human-Friendly AI Speech Jamming to Prevent Automated Phone Scams", "abstract": "Large Language Models (LLMs), combined with Text-to-Speech (TTS) and Automatic Speech Recognition (ASR), are increasingly used to automate voice phishing (vishing) scams. These systems are scalable and convincing, posing a significant security threat. We identify the ASR transcription step as the most vulnerable link in the scam pipeline and introduce ASRJam, a proactive defence framework that injects adversarial perturbations into the victim's audio to disrupt the attacker's ASR. This breaks the scam's feedback loop without affecting human callers, who can still understand the conversation. While prior adversarial audio techniques are often unpleasant and impractical for real-time use, we also propose EchoGuard, a novel jammer that leverages natural distortions, such as reverberation and echo, that are disruptive to ASR but tolerable to humans. To evaluate EchoGuard's effectiveness and usability, we conducted a 39-person user study comparing it with three state-of-the-art attacks. Results show that EchoGuard achieved the highest overall utility, offering the best combination of ASR disruption and human listening experience.", "date": "2025-06-10", "categories": []}
{"id": "2506.11844", "title": "TrustGLM: Evaluating the Robustness of GraphLLMs Against Prompt, Text, and Structure Attacks", "abstract": "Inspired by the success of large language models (LLMs), there is a significant research shift from traditional graph learning methods to LLM-based graph frameworks, formally known as GraphLLMs. GraphLLMs leverage the reasoning power of LLMs by integrating three key components: the textual attributes of input nodes, the structural information of node neighborhoods, and task-specific prompts that guide decision-making. Despite their promise, the robustness of GraphLLMs against adversarial perturbations remains largely unexplored-a critical concern for deploying these models in high-stakes scenarios. To bridge the gap, we introduce TrustGLM, a comprehensive study evaluating the vulnerability of GraphLLMs to adversarial attacks across three dimensions: text, graph structure, and prompt manipulations. We implement state-of-the-art attack algorithms from each perspective to rigorously assess model resilience. Through extensive experiments on six benchmark datasets from diverse domains, our findings reveal that GraphLLMs are highly susceptible to text attacks that merely replace a few semantically similar words in a node's textual attribute. We also find that standard graph structure attack methods can significantly degrade model performance, while random shuffling of the candidate label set in prompt templates leads to substantial performance drops. Beyond characterizing these vulnerabilities, we investigate defense techniques tailored to each attack vector through data-augmented training and adversarial training, which show promising potential to enhance the robustness of GraphLLMs. We hope that our open-sourced library will facilitate rapid, equitable evaluation and inspire further innovative research in this field.", "date": "2025-06-13", "categories": []}
{"id": "2506.11892", "title": "Attention-based Adversarial Robust Distillation in Radio Signal Classifications for Low-Power IoT Devices", "abstract": "Due to great success of transformers in many applications such as natural language processing and computer vision, transformers have been successfully applied in automatic modulation classification. We have shown that transformer-based radio signal classification is vulnerable to imperceptible and carefully crafted attacks called adversarial examples. Therefore, we propose a defense system against adversarial examples in transformer-based modulation classifications. Considering the need for computationally efficient architecture particularly for Internet of Things (IoT)-based applications or operation of devices in environment where power supply is limited, we propose a compact transformer for modulation classification. The advantages of robust training such as adversarial training in transformers may not be attainable in compact transformers. By demonstrating this, we propose a novel compact transformer that can enhance robustness in the presence of adversarial attacks. The new method is aimed at transferring the adversarial attention map from the robustly trained large transformer to a compact transformer. The proposed method outperforms the state-of-the-art techniques for the considered white-box scenarios including fast gradient method and projected gradient descent attacks. We have provided reasoning of the underlying working mechanisms and investigated the transferability of the adversarial examples between different architectures. The proposed method has the potential to protect the transformer from the transferability of adversarial examples.", "date": "2025-06-13", "categories": []}
{"id": "2506.11901", "title": "A Neural Rejection System Against Universal Adversarial Perturbations in Radio Signal Classification", "abstract": "Advantages of deep learning over traditional methods have been demonstrated for radio signal classification in the recent years. However, various researchers have discovered that even a small but intentional feature perturbation known as adversarial examples can significantly deteriorate the performance of the deep learning based radio signal classification. Among various kinds of adversarial examples, universal adversarial perturbation has gained considerable attention due to its feature of being data independent, hence as a practical strategy to fool the radio signal classification with a high success rate. Therefore, in this paper, we investigate a defense system called neural rejection system to propose against universal adversarial perturbations, and evaluate its performance by generating white-box universal adversarial perturbations. We show that the proposed neural rejection system is able to defend universal adversarial perturbations with significantly higher accuracy than the undefended deep neural network.", "date": "2025-06-13", "categories": []}
{"id": "2503.14299", "title": "Unveiling the Role of Randomization in Multiclass Adversarial Classification: Insights from Graph Theory", "abstract": "Randomization as a mean to improve the adversarial robustness of machine learning models has recently attracted significant attention. Unfortunately, much of the theoretical analysis so far has focused on binary classification, providing only limited insights into the more complex multiclass setting. In this paper, we take a step toward closing this gap by drawing inspiration from the field of graph theory. Our analysis focuses on discrete data distributions, allowing us to cast the adversarial risk minimization problems within the well-established framework of set packing problems. By doing so, we are able to identify three structural conditions on the support of the data distribution that are necessary for randomization to improve robustness. Furthermore, we are able to construct several data distributions where (contrarily to binary classification) switching from a deterministic to a randomized solution significantly reduces the optimal adversarial risk. These findings highlight the crucial role randomization can play in enhancing robustness to adversarial attacks in multiclass classification.", "date": "2025-03-18", "categories": []}
{"id": "2506.10047", "title": "GenBreak: Red Teaming Text-to-Image Generators Using Large Language Models", "abstract": "Text-to-image (T2I) models such as Stable Diffusion have advanced rapidly and are now widely used in content creation. However, these models can be misused to generate harmful content, including nudity or violence, posing significant safety risks. While most platforms employ content moderation systems, underlying vulnerabilities can still be exploited by determined adversaries. Recent research on red-teaming and adversarial attacks against T2I models has notable limitations: some studies successfully generate highly toxic images but use adversarial prompts that are easily detected and blocked by safety filters, while others focus on bypassing safety mechanisms but fail to produce genuinely harmful outputs, neglecting the discovery of truly high-risk prompts. Consequently, there remains a lack of reliable tools for evaluating the safety of defended T2I models. To address this gap, we propose GenBreak, a framework that fine-tunes a red-team large language model (LLM) to systematically explore underlying vulnerabilities in T2I generators. Our approach combines supervised fine-tuning on curated datasets with reinforcement learning via interaction with a surrogate T2I model. By integrating multiple reward signals, we guide the LLM to craft adversarial prompts that enhance both evasion capability and image toxicity, while maintaining semantic coherence and diversity. These prompts demonstrate strong effectiveness in black-box attacks against commercial T2I generators, revealing practical and concerning safety weaknesses.", "date": "2025-06-11", "categories": []}
{"id": "2506.10620", "title": "Assessing the Resilience of Automotive Intrusion Detection Systems to Adversarial Manipulation", "abstract": "The security of modern vehicles has become increasingly important, with the controller area network (CAN) bus serving as a critical communication backbone for various Electronic Control Units (ECUs). The absence of robust security measures in CAN, coupled with the increasing connectivity of vehicles, makes them susceptible to cyberattacks. While intrusion detection systems (IDSs) have been developed to counter such threats, they are not foolproof. Adversarial attacks, particularly evasion attacks, can manipulate inputs to bypass detection by IDSs. This paper extends our previous work by investigating the feasibility and impact of gradient-based adversarial attacks performed with different degrees of knowledge against automotive IDSs. We consider three scenarios: white-box (attacker with full system knowledge), grey-box (partial system knowledge), and the more realistic black-box (no knowledge of the IDS' internal workings or data). We evaluate the effectiveness of the proposed attacks against state-of-the-art IDSs on two publicly available datasets. Additionally, we study effect of the adversarial perturbation on the attack impact and evaluate real-time feasibility by precomputing evasive payloads for timed injection based on bus traffic. Our results demonstrate that, besides attacks being challenging due to the automotive domain constraints, their effectiveness is strongly dependent on the dataset quality, the target IDS, and the attacker's degree of knowledge.", "date": "2025-06-12", "categories": []}
{"id": "2506.10831", "title": "Efficiency Robustness of Dynamic Deep Learning Systems", "abstract": "Deep Learning Systems (DLSs) are increasingly deployed in real-time applications, including those in resourceconstrained environments such as mobile and IoT devices. To address efficiency challenges, Dynamic Deep Learning Systems (DDLSs) adapt inference computation based on input complexity, reducing overhead. While this dynamic behavior improves efficiency, such behavior introduces new attack surfaces. In particular, efficiency adversarial attacks exploit these dynamic mechanisms to degrade system performance. This paper systematically explores efficiency robustness of DDLSs, presenting the first comprehensive taxonomy of efficiency attacks. We categorize these attacks based on three dynamic behaviors: (i) attacks on dynamic computations per inference, (ii) attacks on dynamic inference iterations, and (iii) attacks on dynamic output production for downstream tasks. Through an in-depth evaluation, we analyze adversarial strategies that target DDLSs efficiency and identify key challenges in securing these systems. In addition, we investigate existing defense mechanisms, demonstrating their limitations against increasingly popular efficiency attacks and the necessity for novel mitigation strategies to secure future adaptive DDLSs.", "date": "2025-06-12", "categories": []}
{"id": "2506.10888", "title": "Lattice Climber Attack: Adversarial attacks for randomized mixtures of classifiers", "abstract": "Finite mixtures of classifiers (a.k.a. randomized ensembles) have been proposed as a way to improve robustness against adversarial attacks. However, existing attacks have been shown to not suit this kind of classifier. In this paper, we discuss the problem of attacking a mixture in a principled way and introduce two desirable properties of attacks based on a geometrical analysis of the problem (effectiveness and maximality). We then show that existing attacks do not meet both of these properties. Finally, we introduce a new attack called {\\em lattice climber attack} with theoretical guarantees in the binary linear setting, and demonstrate its performance by conducting experiments on synthetic and real datasets.", "date": "2025-06-12", "categories": []}
{"id": "2409.19149", "title": "Multimodal Pragmatic Jailbreak on Text-to-image Models", "abstract": "Diffusion models have recently achieved remarkable advancements in terms of image quality and fidelity to textual prompts. Concurrently, the safety of such generative models has become an area of growing concern. This work introduces a novel type of jailbreak, which triggers T2I models to generate the image with visual text, where the image and the text, although considered to be safe in isolation, combine to form unsafe content. To systematically explore this phenomenon, we propose a dataset to evaluate the current diffusion-based text-to-image (T2I) models under such jailbreak. We benchmark nine representative T2I models, including two closed-source commercial models. Experimental results reveal a concerning tendency to produce unsafe content: all tested models suffer from such type of jailbreak, with rates of unsafe generation ranging from around 10\\% to 70\\% where DALLE 3 demonstrates almost the highest unsafety. In real-world scenarios, various filters such as keyword blocklists, customized prompt filters, and NSFW image filters, are commonly employed to mitigate these risks. We evaluate the effectiveness of such filters against our jailbreak and found that, while these filters may be effective for single modality detection, they fail to work against our jailbreak. We also investigate the underlying reason for such jailbreaks, from the perspective of text rendering capability and training data. Our work provides a foundation for further development towards more secure and reliable T2I models. Project page at https://multimodalpragmatic.github.io/.", "date": "2024-09-27", "categories": []}
{"id": "2411.12768", "title": "CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization", "abstract": "Large Language Models (LLMs) are vulnerable to backdoor attacks that manipulate outputs via hidden triggers. Existing defense methods--designed for vision/text classification tasks--fail for text generation. We propose Internal Consistency Regularization (CROW), a defense leveraging the observation that backdoored models exhibit unstable layer-wise hidden representations when triggered, while clean models show smooth transitions. CROW enforces consistency across layers via adversarial perturbations and regularization during finetuning, neutralizing backdoors without requiring clean reference models or trigger knowledge--only a small clean dataset. Experiments across Llama-2 (7B, 13B), CodeLlama (7B, 13B), and Mistral-7B demonstrate CROW's effectiveness: it achieves significant reductions in attack success rates across diverse backdoor strategies (sentiment steering, targeted refusal, code injection) while preserving generative performance. CROW's architecture-agnostic design enables practical deployment.", "date": "2024-11-18", "categories": []}
{"id": "2506.04394", "title": "Is Perturbation-Based Image Protection Disruptive to Image Editing?", "abstract": "The remarkable image generation capabilities of state-of-the-art diffusion models, such as Stable Diffusion, can also be misused to spread misinformation and plagiarize copyrighted materials. To mitigate the potential risks associated with image editing, current image protection methods rely on adding imperceptible perturbations to images to obstruct diffusion-based editing. A fully successful protection for an image implies that the output of editing attempts is an undesirable, noisy image which is completely unrelated to the reference image. In our experiments with various perturbation-based image protection methods across multiple domains (natural scene images and artworks) and editing tasks (image-to-image generation and style editing), we discover that such protection does not achieve this goal completely. In most scenarios, diffusion-based editing of protected images generates a desirable output image which adheres precisely to the guidance prompt. Our findings suggest that adding noise to images may paradoxically increase their association with given text prompts during the generation process, leading to unintended consequences such as better resultant edits. Hence, we argue that perturbation-based methods may not provide a sufficient solution for robust image protection against diffusion-based editing.", "date": "2025-06-04", "categories": []}
{"id": "2506.09148", "title": "Adversarial Text Generation with Dynamic Contextual Perturbation", "abstract": "Adversarial attacks on Natural Language Processing (NLP) models expose vulnerabilities by introducing subtle perturbations to input text, often leading to misclassification while maintaining human readability. Existing methods typically focus on word-level or local text segment alterations, overlooking the broader context, which results in detectable or semantically inconsistent perturbations. We propose a novel adversarial text attack scheme named Dynamic Contextual Perturbation (DCP). DCP dynamically generates context-aware perturbations across sentences, paragraphs, and documents, ensuring semantic fidelity and fluency. Leveraging the capabilities of pre-trained language models, DCP iteratively refines perturbations through an adversarial objective function that balances the dual objectives of inducing model misclassification and preserving the naturalness of the text. This comprehensive approach allows DCP to produce more sophisticated and effective adversarial examples that better mimic natural language patterns. Our experimental results, conducted on various NLP models and datasets, demonstrate the efficacy of DCP in challenging the robustness of state-of-the-art NLP systems. By integrating dynamic contextual analysis, DCP significantly enhances the subtlety and impact of adversarial attacks. This study highlights the critical role of context in adversarial attacks and lays the groundwork for creating more robust NLP systems capable of withstanding sophisticated adversarial strategies.", "date": "2025-06-10", "categories": []}
{"id": "2506.09640", "title": "Evasion Attacks Against Bayesian Predictive Models", "abstract": "There is an increasing interest in analyzing the behavior of machine learning systems against adversarial attacks. However, most of the research in adversarial machine learning has focused on studying weaknesses against evasion or poisoning attacks to predictive models in classical setups, with the susceptibility of Bayesian predictive models to attacks remaining underexplored. This paper introduces a general methodology for designing optimal evasion attacks against such models. We investigate two adversarial objectives: perturbing specific point predictions and altering the entire posterior predictive distribution. For both scenarios, we propose novel gradient-based attacks and study their implementation and properties in various computational setups.", "date": "2025-06-11", "categories": []}
{"id": "2506.09896", "title": "A look at adversarial attacks on radio waveforms from discrete latent space", "abstract": "Having designed a VQVAE that maps digital radio waveforms into discrete latent space, and yields a perfectly classifiable reconstruction of the original data, we here analyze the attack suppressing properties of VQVAE when an adversarial attack is performed on high-SNR radio-frequency (RF) data-points. To target amplitude modulations from a subset of digitally modulated waveform classes, we first create adversarial attacks that preserve the phase between the in-phase and quadrature component whose values are adversarially changed. We compare them with adversarial attacks of the same intensity where phase is not preserved. We test the classification accuracy of such adversarial examples on a classifier trained to deliver 100% accuracy on the original data. To assess the ability of VQVAE to suppress the strength of the attack, we evaluate the classifier accuracy on the reconstructions by VQVAE of the adversarial datapoints and show that VQVAE substantially decreases the effectiveness of the attack. We also compare the I/Q plane diagram of the attacked data, their reconstructions and the original data. Finally, using multiple methods and metrics, we compare the probability distribution of the VQVAE latent space with and without attack. Varying the attack strength, we observe interesting properties of the discrete space, which may help detect the attacks.", "date": "2025-06-11", "categories": []}
{"id": "2310.08732", "title": "Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing", "abstract": "As ML models are increasingly deployed in critical applications, robustness against adversarial perturbations is crucial. While numerous defenses have been proposed to counter such attacks, they typically assume that all adversarial transformations are equally important, an assumption that rarely aligns with real-world applications. To address this, we study the problem of robust learning against adversarial perturbations under cost-sensitive scenarios, where the potential harm of different types of misclassifications is encoded in a cost matrix. Our solution introduces a provably robust learning algorithm to certify and optimize for cost-sensitive robustness, building on the scalable certification framework of randomized smoothing. Specifically, we formalize the definition of cost-sensitive certified radius and propose our novel adaptation of the standard certification algorithm to generate tight robustness certificates tailored to any cost matrix. In addition, we design a robust training method that improves certified cost-sensitive robustness without compromising model accuracy. Extensive experiments on benchmark datasets, including challenging ones unsolvable by existing methods, demonstrate the effectiveness of our certification algorithm and training method across various cost-sensitive scenarios.", "date": "2023-10-12", "categories": []}
{"id": "2506.08611", "title": "Towards Class-wise Fair Adversarial Training via Anti-Bias Soft Label Distillation", "abstract": "Adversarial Training (AT) is widely recognized as an effective approach to enhance the adversarial robustness of Deep Neural Networks. As a variant of AT, Adversarial Robustness Distillation (ARD) has shown outstanding performance in enhancing the robustness of small models. However, both AT and ARD face robust fairness issue: these models tend to display strong adversarial robustness against some classes (easy classes) while demonstrating weak adversarial robustness against others (hard classes). This paper explores the underlying factors of this problem and points out the smoothness degree of soft labels for different classes significantly impacts the robust fairness from both empirical observation and theoretical analysis. Based on the above exploration, we propose Anti-Bias Soft Label Distillation (ABSLD) within the Knowledge Distillation framework to enhance the adversarial robust fairness. Specifically, ABSLD adaptively reduces the student's error risk gap between different classes, which is accomplished by adjusting the class-wise smoothness degree of teacher's soft labels during the training process, and the adjustment is managed by assigning varying temperatures to different classes. Additionally, as a label-based approach, ABSLD is highly adaptable and can be integrated with the sample-based methods. Extensive experiments demonstrate ABSLD outperforms state-of-the-art methods on the comprehensive performance of robustness and fairness.", "date": "2025-06-10", "categories": []}
{"id": "2306.07992", "title": "Securing Visually-Aware Recommender Systems: An Adversarial Image Reconstruction and Detection Framework", "abstract": "With rich visual data, such as images, becoming readily associated with items, visually-aware recommendation systems (VARS) have been widely used in different applications. Recent studies have shown that VARS are vulnerable to item-image adversarial attacks, which add human-imperceptible perturbations to the clean images associated with those items. Attacks on VARS pose new security challenges to a wide range of applications such as e-Commerce and social networks where VARS are widely used. How to secure VARS from such adversarial attacks becomes a critical problem. Currently, there is still a lack of systematic study on how to design secure defense strategies against visual attacks on VARS. In this paper, we attempt to fill this gap by proposing an adversarial image reconstruction and detection framework to secure VARS. Our proposed method can simultaneously (1) secure VARS from adversarial attacks characterized by local perturbations by image reconstruction based on global vision transformers; and (2) accurately detect adversarial examples using a novel contrastive learning approach. Meanwhile, our framework is designed to be used as both a filter and a detector so that they can be jointly trained to improve the flexibility of our defense strategy to a variety of attacks and VARS models. We have conducted extensive experimental studies with two popular attack methods (FGSM and PGD). Our experimental results on two real-world datasets show that our defense strategy against visual attacks is effective and outperforms existing methods on different attacks. Moreover, our method can detect adversarial examples with high accuracy.", "date": "2023-06-11", "categories": []}
{"id": "2406.11458", "title": "Adversaries With Incentives: A Strategic Alternative to Adversarial Robustness", "abstract": "Adversarial training aims to defend against adversaries: malicious opponents whose sole aim is to harm predictive performance in any way possible. This presents a rather harsh perspective, which we assert results in unnecessarily conservative training. As an alternative, we propose to model opponents as simply pursuing their own goals--rather than working directly against the classifier. Employing tools from strategic modeling, our approach enables knowledge or beliefs regarding the opponent's possible incentives to be used as inductive bias for learning. Accordingly, our method of strategic training is designed to defend against all opponents within an 'incentive uncertainty set'. This resorts to adversarial learning when the set is maximal, but offers potential gains when the set can be appropriately reduced. We conduct a series of experiments that show how even mild knowledge regarding the opponent's incentives can be useful, and that the degree of potential gains depends on how these incentives relate to the structure of the learning task.", "date": "2024-06-17", "categories": []}
{"id": "2412.07468", "title": "AHSG: Adversarial Attack on High-level Semantics in Graph Neural Networks", "abstract": "Adversarial attacks on Graph Neural Networks aim to perturb the performance of the learner by carefully modifying the graph topology and node attributes. Existing methods achieve attack stealthiness by constraining the modification budget and differences in graph properties. However, these methods typically disrupt task-relevant primary semantics directly, which results in low defensibility and detectability of the attack. In this paper, we propose an Adversarial Attack on High-level Semantics for Graph Neural Networks (AHSG), which is a graph structure attack model that ensures the retention of primary semantics. By combining latent representations with shared primary semantics, our model retains detectable attributes and relational patterns of the original graph while leveraging more subtle changes to carry out the attack. Then we use the Projected Gradient Descent algorithm to map the latent representations with attack effects to the adversarial graph. Through experiments on robust graph deep learning models equipped with defense strategies, we demonstrate that AHSG outperforms other state-of-the-art methods in attack effectiveness. Additionally, using Contextual Stochastic Block Models to detect the attacked graph further validates that our method preserves the primary semantics of the graph.", "date": "2024-12-10", "categories": []}
{"id": "2501.00066", "title": "On Adversarial Robustness of Language Models in Transfer Learning", "abstract": "We investigate the adversarial robustness of LLMs in transfer learning scenarios. Through comprehensive experiments on multiple datasets (MBIB Hate Speech, MBIB Political Bias, MBIB Gender Bias) and various model architectures (BERT, RoBERTa, GPT-2, Gemma, Phi), we reveal that transfer learning, while improving standard performance metrics, often leads to increased vulnerability to adversarial attacks. Our findings demonstrate that larger models exhibit greater resilience to this phenomenon, suggesting a complex interplay between model size, architecture, and adaptation methods. Our work highlights the crucial need for considering adversarial robustness in transfer learning scenarios and provides insights into maintaining model security without compromising performance. These findings have significant implications for the development and deployment of LLMs in real-world applications where both performance and robustness are paramount.", "date": "2024-12-29", "categories": []}
{"id": "2503.20279", "title": "sudo rm -rf agentic_security", "abstract": "Large Language Models (LLMs) are increasingly deployed as computer-use agents, autonomously performing tasks within real desktop or web environments. While this evolution greatly expands practical use cases for humans, it also creates serious security exposures. We present SUDO (Screen-based Universal Detox2Tox Offense), a novel attack framework that systematically bypasses refusal-trained safeguards in commercial computer-use agents, such as Claude for Computer Use. The core mechanism, Detox2Tox, transforms harmful requests (that agents initially reject) into seemingly benign requests via detoxification, secures detailed instructions from advanced vision language models (VLMs), and then reintroduces malicious content via toxification just before execution. Unlike conventional jailbreaks, SUDO iteratively refines its attacks based on a built-in refusal feedback, making it increasingly effective against robust policy filters. In extensive tests spanning 50 real-world tasks and multiple state-of-the-art VLMs, SUDO achieves a stark attack success rate of 24.41% (with no refinement), and up to 41.33% (by its iterative refinement) in Claude for Computer Use. By revealing these vulnerabilities and demonstrating the ease with which they can be exploited in real-world computing environments, this paper highlights an immediate need for robust, context-aware safeguards. WARNING: This paper includes harmful or offensive model outputs", "date": "2025-03-26", "categories": []}
{"id": "2503.20583", "title": "Feature Statistics with Uncertainty Help Adversarial Robustness", "abstract": "Despite the remarkable success of deep neural networks (DNNs), the security threat of adversarial attacks poses a significant challenge to the reliability of DNNs. In this paper, both theoretically and empirically, we discover a universal phenomenon that has been neglected in previous works, i.e., adversarial attacks tend to shift the distributions of feature statistics. Motivated by this finding, and by leveraging the advantages of uncertainty-aware stochastic methods in building robust models efficiently, we propose an uncertainty-driven feature statistics adjustment module for robustness enhancement, named Feature Statistics with Uncertainty (FSU). It randomly resamples channel-wise feature means and standard deviations of examples from multivariate Gaussian distributions, which helps to reconstruct the perturbed examples and calibrate the shifted distributions. The calibration recovers some domain characteristics of the data for classification, thereby mitigating the influence of perturbations and weakening the ability of attacks to deceive models. The proposed FSU module has universal applicability in training, attacking, predicting, and fine-tuning, demonstrating impressive robustness enhancement ability at a trivial additional time cost. For example, by fine-tuning the well-established models with FSU, the state-of-the-art methods achieve up to 17.13% and 34.82% robustness improvement against powerful AA and CW attacks on benchmark datasets.", "date": "2025-03-26", "categories": []}
{"id": "2506.00978", "title": "CAPAA: Classifier-Agnostic Projector-Based Adversarial Attack", "abstract": "Projector-based adversarial attack aims to project carefully designed light patterns (i.e., adversarial projections) onto scenes to deceive deep image classifiers. It has potential applications in privacy protection and the development of more robust classifiers. However, existing approaches primarily focus on individual classifiers and fixed camera poses, often neglecting the complexities of multi-classifier systems and scenarios with varying camera poses. This limitation reduces their effectiveness when introducing new classifiers or camera poses. In this paper, we introduce Classifier-Agnostic Projector-Based Adversarial Attack (CAPAA) to address these issues. First, we develop a novel classifier-agnostic adversarial loss and optimization framework that aggregates adversarial and stealthiness loss gradients from multiple classifiers. Then, we propose an attention-based gradient weighting mechanism that concentrates perturbations on regions of high classification activation, thereby improving the robustness of adversarial projections when applied to scenes with varying camera poses. Our extensive experimental evaluations demonstrate that CAPAA achieves both a higher attack success rate and greater stealthiness compared to existing baselines. Codes are available at: https://github.com/ZhanLiQxQ/CAPAA.", "date": "2025-06-01", "categories": []}
{"id": "2506.03988", "title": "RAID: A Dataset for Testing the Adversarial Robustness of AI-Generated Image Detectors", "abstract": "AI-generated images have reached a quality level at which humans are incapable of reliably distinguishing them from real images. To counteract the inherent risk of fraud and disinformation, the detection of AI-generated images is a pressing challenge and an active research topic. While many of the presented methods claim to achieve high detection accuracy, they are usually evaluated under idealized conditions. In particular, the adversarial robustness is often neglected, potentially due to a lack of awareness or the substantial effort required to conduct a comprehensive robustness analysis. In this work, we tackle this problem by providing a simpler means to assess the robustness of AI-generated image detectors. We present RAID (Robust evaluation of AI-generated image Detectors), a dataset of 72k diverse and highly transferable adversarial examples. The dataset is created by running attacks against an ensemble of seven state-of-the-art detectors and images generated by four different text-to-image models. Extensive experiments show that our methodology generates adversarial images that transfer with a high success rate to unseen detectors, which can be used to quickly provide an approximate yet still reliable estimate of a detector's adversarial robustness. Our findings indicate that current state-of-the-art AI-generated image detectors can be easily deceived by adversarial examples, highlighting the critical need for the development of more robust methods. We release our dataset at https://huggingface.co/datasets/aimagelab/RAID and evaluation code at https://github.com/pralab/RAID.", "date": "2025-06-04", "categories": []}
{"id": "2506.06563", "title": "Securing Traffic Sign Recognition Systems in Autonomous Vehicles", "abstract": "Deep Neural Networks (DNNs) are widely used for traffic sign recognition because they can automatically extract high-level features from images. These DNNs are trained on large-scale datasets obtained from unknown sources. Therefore, it is important to ensure that the models remain secure and are not compromised or poisoned during training. In this paper, we investigate the robustness of DNNs trained for traffic sign recognition. First, we perform the error-minimizing attacks on DNNs used for traffic sign recognition by adding imperceptible perturbations on training data. Then, we propose a data augmentation-based training method to mitigate the error-minimizing attacks. The proposed training method utilizes nonlinear transformations to disrupt the perturbations and improve the model robustness. We experiment with two well-known traffic sign datasets to demonstrate the severity of the attack and the effectiveness of our mitigation scheme. The error-minimizing attacks reduce the prediction accuracy of the DNNs from 99.90% to 10.6%. However, our mitigation scheme successfully restores the prediction accuracy to 96.05%. Moreover, our approach outperforms adversarial training in mitigating the error-minimizing attacks. Furthermore, we propose a detection model capable of identifying poisoned data even when the perturbations are imperceptible to human inspection. Our detection model achieves a success rate of over 99% in identifying the attack. This research highlights the need to employ advanced training methods for DNNs in traffic sign recognition systems to mitigate the effects of data poisoning attacks.", "date": "2025-06-06", "categories": []}
{"id": "2506.06933", "title": "Rewriting the Budget: A General Framework for Black-Box Attacks Under Cost Asymmetry", "abstract": "Traditional decision-based black-box adversarial attacks on image classifiers aim to generate adversarial examples by slightly modifying input images while keeping the number of queries low, where each query involves sending an input to the model and observing its output. Most existing methods assume that all queries have equal cost. However, in practice, queries may incur asymmetric costs; for example, in content moderation systems, certain output classes may trigger additional review, enforcement, or penalties, making them more costly than others. While prior work has considered such asymmetric cost settings, effective algorithms for this scenario remain underdeveloped. In this paper, we propose a general framework for decision-based attacks under asymmetric query costs, which we refer to as asymmetric black-box attacks. We modify two core components of existing attacks: the search strategy and the gradient estimation process. Specifically, we propose Asymmetric Search (AS), a more conservative variant of binary search that reduces reliance on high-cost queries, and Asymmetric Gradient Estimation (AGREST), which shifts the sampling distribution to favor low-cost queries. We design efficient algorithms that minimize total attack cost by balancing different query types, in contrast to earlier methods such as stealthy attacks that focus only on limiting expensive (high-cost) queries. Our method can be integrated into a range of existing black-box attacks with minimal changes. We perform both theoretical analysis and empirical evaluation on standard image classification benchmarks. Across various cost regimes, our method consistently achieves lower total query cost and smaller perturbations than existing approaches, with improvements of up to 40% in some settings.", "date": "2025-06-07", "categories": []}
{"id": "2506.07056", "title": "D2R: dual regularization loss with collaborative adversarial generation for model robustness", "abstract": "The robustness of Deep Neural Network models is crucial for defending models against adversarial attacks. Recent defense methods have employed collaborative learning frameworks to enhance model robustness. Two key limitations of existing methods are (i) insufficient guidance of the target model via loss functions and (ii) non-collaborative adversarial generation. We, therefore, propose a dual regularization loss (D2R Loss) method and a collaborative adversarial generation (CAG) strategy for adversarial training. D2R loss includes two optimization steps. The adversarial distribution and clean distribution optimizations enhance the target model's robustness by leveraging the strengths of different loss functions obtained via a suitable function space exploration to focus more precisely on the target model's distribution. CAG generates adversarial samples using a gradient-based collaboration between guidance and target models. We conducted extensive experiments on three benchmark databases, including CIFAR-10, CIFAR-100, Tiny ImageNet, and two popular target models, WideResNet34-10 and PreActResNet18. Our results show that D2R loss with CAG produces highly robust models.", "date": "2025-06-08", "categories": []}
{"id": "2506.07214", "title": "Backdoor Attack on Vision Language Models with Stealthy Semantic Manipulation", "abstract": "Vision Language Models (VLMs) have shown remarkable performance, but are also vulnerable to backdoor attacks whereby the adversary can manipulate the model's outputs through hidden triggers. Prior attacks primarily rely on single-modality triggers, leaving the crucial cross-modal fusion nature of VLMs largely unexplored. Unlike prior work, we identify a novel attack surface that leverages cross-modal semantic mismatches as implicit triggers. Based on this insight, we propose BadSem (Backdoor Attack with Semantic Manipulation), a data poisoning attack that injects stealthy backdoors by deliberately misaligning image-text pairs during training. To perform the attack, we construct SIMBad, a dataset tailored for semantic manipulation involving color and object attributes. Extensive experiments across four widely used VLMs show that BadSem achieves over 98% average ASR, generalizes well to out-of-distribution datasets, and can transfer across poisoning modalities. Our detailed analysis using attention visualization shows that backdoored models focus on semantically sensitive regions under mismatched conditions while maintaining normal behavior on clean inputs. To mitigate the attack, we try two defense strategies based on system prompt and supervised fine-tuning but find that both of them fail to mitigate the semantic backdoor. Our findings highlight the urgent need to address semantic vulnerabilities in VLMs for their safer deployment.", "date": "2025-06-08", "categories": []}
{"id": "2506.07428", "title": "HeTa: Relation-wise Heterogeneous Graph Foundation Attack Model", "abstract": "Heterogeneous Graph Neural Networks (HGNNs) are vulnerable, highlighting the need for tailored attacks to assess their robustness and ensure security. However, existing HGNN attacks often require complex retraining of parameters to generate specific perturbations for new scenarios. Recently, foundation models have opened new horizons for the generalization of graph neural networks by capturing shared semantics across various graph distributions. This leads us to ask:Can we design a foundation attack model for HGNNs that enables generalizable perturbations across different HGNNs, and quickly adapts to new heterogeneous graphs (HGs)? Empirical findings reveal that, despite significant differences in model design and parameter space, different HGNNs surprisingly share common vulnerability patterns from a relation-aware perspective. Therefore, we explore how to design foundation HGNN attack criteria by mining shared attack units. In this paper, we propose a novel relation-wise heterogeneous graph foundation attack model, HeTa. We introduce a foundation surrogate model to align heterogeneity and identify the importance of shared relation-aware attack units. Building on this, we implement a serialized relation-by-relation attack based on the identified relational weights. In this way, the perturbation can be transferred to various target HGNNs and easily fine-tuned for new HGs. Extensive experiments exhibit powerful attack performances and generalizability of our method.", "date": "2025-06-09", "categories": []}
{"id": "2506.07590", "title": "Explore the vulnerability of black-box models via diffusion models", "abstract": "Recent advancements in diffusion models have enabled high-fidelity and photorealistic image generation across diverse applications. However, these models also present security and privacy risks, including copyright violations, sensitive information leakage, and the creation of harmful or offensive content that could be exploited maliciously. In this study, we uncover a novel security threat where an attacker leverages diffusion model APIs to generate synthetic images, which are then used to train a high-performing substitute model. This enables the attacker to execute model extraction and transfer-based adversarial attacks on black-box classification models with minimal queries, without needing access to the original training data. The generated images are sufficiently high-resolution and diverse to train a substitute model whose outputs closely match those of the target model. Across the seven benchmarks, including CIFAR and ImageNet subsets, our method shows an average improvement of 27.37% over state-of-the-art methods while using just 0.01 times of the query budget, achieving a 98.68% success rate in adversarial attacks on the target model.", "date": "2025-06-09", "categories": []}
{"id": "2506.07804", "title": "Enhancing Adversarial Robustness with Conformal Prediction: A Framework for Guaranteed Model Reliability", "abstract": "As deep learning models are increasingly deployed in high-risk applications, robust defenses against adversarial attacks and reliable performance guarantees become paramount. Moreover, accuracy alone does not provide sufficient assurance or reliable uncertainty estimates for these models. This study advances adversarial training by leveraging principles from Conformal Prediction. Specifically, we develop an adversarial attack method, termed OPSA (OPtimal Size Attack), designed to reduce the efficiency of conformal prediction at any significance level by maximizing model uncertainty without requiring coverage guarantees. Correspondingly, we introduce OPSA-AT (Adversarial Training), a defense strategy that integrates OPSA within a novel conformal training paradigm. Experimental evaluations demonstrate that our OPSA attack method induces greater uncertainty compared to baseline approaches for various defenses. Conversely, our OPSA-AT defensive model significantly enhances robustness not only against OPSA but also other adversarial attacks, and maintains reliable prediction. Our findings highlight the effectiveness of this integrated approach for developing trustworthy and resilient deep learning models for safety-critical domains. Our code is available at https://github.com/bjbbbb/Enhancing-Adversarial-Robustness-with-Conformal-Prediction.", "date": "2025-06-09", "categories": []}
{"id": "2506.07948", "title": "TokenBreak: Bypassing Text Classification Models Through Token Manipulation", "abstract": "Natural Language Processing (NLP) models are used for text-related tasks such as classification and generation. To complete these tasks, input data is first tokenized from human-readable text into a format the model can understand, enabling it to make inferences and understand context. Text classification models can be implemented to guard against threats such as prompt injection attacks against Large Language Models (LLMs), toxic input and cybersecurity risks such as spam emails. In this paper, we introduce TokenBreak: a novel attack that can bypass these protection models by taking advantage of the tokenization strategy they use. This attack technique manipulates input text in such a way that certain models give an incorrect classification. Importantly, the end target (LLM or email recipient) can still understand and respond to the manipulated text and therefore be vulnerable to the very attack the protection model was put in place to prevent. The tokenizer is tied to model architecture, meaning it is possible to predict whether or not a model is vulnerable to attack based on family. We also present a defensive strategy as an added layer of protection that can be implemented without having to retrain the defensive model.", "date": "2025-06-09", "categories": []}
{"id": "2411.14842", "title": "Who Can Withstand Chat-Audio Attacks? An Evaluation Benchmark for Large Audio-Language Models", "abstract": "Adversarial audio attacks pose a significant threat to the growing use of large audio-language models (LALMs) in voice-based human-machine interactions. While existing research focused on model-specific adversarial methods, real-world applications demand a more generalizable and universal approach to audio adversarial attacks. In this paper, we introduce the Chat-Audio Attacks (CAA) benchmark including four distinct types of audio attacks, which aims to explore the vulnerabilities of LALMs to these audio attacks in conversational scenarios. To evaluate the robustness of LALMs, we propose three evaluation strategies: Standard Evaluation, utilizing traditional metrics to quantify model performance under attacks; GPT-4o-Based Evaluation, which simulates real-world conversational complexities; and Human Evaluation, offering insights into user perception and trust. We evaluate six state-of-the-art LALMs with voice interaction capabilities, including Gemini-1.5-Pro, GPT-4o, and others, using three distinct evaluation methods on the CAA benchmark. Our comprehensive analysis reveals the impact of four types of audio attacks on the performance of these models, demonstrating that GPT-4o exhibits the highest level of resilience. Our data can be accessed via the following link: \\href{https://github.com/crystraldo/CAA}{CAA}.", "date": "2024-11-22", "categories": []}
{"id": "2503.02169", "title": "One Stone, Two Birds: Enhancing Adversarial Defense Through the Lens of Distributional Discrepancy", "abstract": "Statistical adversarial data detection (SADD) detects whether an upcoming batch contains adversarial examples (AEs) by measuring the distributional discrepancies between clean examples (CEs) and AEs. In this paper, we explore the strength of SADD-based methods by theoretically showing that minimizing distributional discrepancy can help reduce the expected loss on AEs. Despite these advantages, SADD-based methods have a potential limitation: they discard inputs that are detected as AEs, leading to the loss of useful information within those inputs. To address this limitation, we propose a two-pronged adversarial defense method, named Distributional-discrepancy-based Adversarial Defense (DAD). In the training phase, DAD first optimizes the test power of the maximum mean discrepancy (MMD) to derive MMD-OPT, which is a stone that kills two birds. MMD-OPT first serves as a guiding signal to minimize the distributional discrepancy between CEs and AEs to train a denoiser. Then, it serves as a discriminator to differentiate CEs and AEs during inference. Overall, in the inference stage, DAD consists of a two-pronged process: (1) directly feeding the detected CEs into the classifier, and (2) removing noise from the detected AEs by the distributional-discrepancy-based denoiser. Extensive experiments show that DAD outperforms current state-of-the-art (SOTA) defense methods by simultaneously improving clean and robust accuracy on CIFAR-10 and ImageNet-1K against adaptive white-box attacks. Codes are publicly available at: https://github.com/tmlr-group/DAD.", "date": "2025-03-04", "categories": []}
{"id": "2503.02174", "title": "Adversarial Tokenization", "abstract": "Current LLM pipelines account for only one possible tokenization for a given string, ignoring exponentially many alternative tokenizations during training and inference. For example, the standard Llama3 tokenization of penguin is [p,enguin], yet [peng,uin] is another perfectly valid alternative. In this paper, we show that despite LLMs being trained solely on one tokenization, they still retain semantic understanding of other tokenizations, raising questions about their implications in LLM safety. Put succinctly, we answer the following question: can we adversarially tokenize an obviously malicious string to evade safety and alignment restrictions? We show that not only is adversarial tokenization an effective yet previously neglected axis of attack, but it is also competitive against existing state-of-the-art adversarial approaches without changing the text of the harmful request. We empirically validate this exploit across three state-of-the-art LLMs and adversarial datasets, revealing a previously unknown vulnerability in subword models.", "date": "2025-03-04", "categories": []}
{"id": "2506.05382", "title": "How stealthy is stealthy? Studying the Efficacy of Black-Box Adversarial Attacks in the Real World", "abstract": "Deep learning systems, critical in domains like autonomous vehicles, are vulnerable to adversarial examples (crafted inputs designed to mislead classifiers). This study investigates black-box adversarial attacks in computer vision. This is a realistic scenario, where attackers have query-only access to the target model. Three properties are introduced to evaluate attack feasibility: robustness to compression, stealthiness to automatic detection, and stealthiness to human inspection. State-of-the-Art methods tend to prioritize one criterion at the expense of others. We propose ECLIPSE, a novel attack method employing Gaussian blurring on sampled gradients and a local surrogate model. Comprehensive experiments on a public dataset highlight ECLIPSE's advantages, demonstrating its contribution to the trade-off between the three properties.", "date": "2025-06-03", "categories": []}
{"id": "2506.05430", "title": "Explainer-guided Targeted Adversarial Attacks against Binary Code Similarity Detection Models", "abstract": "Binary code similarity detection (BCSD) serves as a fundamental technique for various software engineering tasks, e.g., vulnerability detection and classification. Attacks against such models have therefore drawn extensive attention, aiming at misleading the models to generate erroneous predictions. Prior works have explored various approaches to generating semantic-preserving variants, i.e., adversarial samples, to evaluate the robustness of the models against adversarial attacks. However, they have mainly relied on heuristic criteria or iterative greedy algorithms to locate salient code influencing the model output, failing to operate on a solid theoretical basis. Moreover, when processing programs with high complexities, such attacks tend to be time-consuming.\n  In this work, we propose a novel optimization for adversarial attacks against BCSD models. In particular, we aim to improve the attacks in a challenging scenario, where the attack goal is to limit the model predictions to a specific range, i.e., the targeted attacks. Our attack leverages the superior capability of black-box, model-agnostic explainers in interpreting the model decision boundaries, thereby pinpointing the critical code snippet to apply semantic-preserving perturbations. The evaluation results demonstrate that compared with the state-of-the-art attacks, the proposed attacks achieve higher attack success rate in almost all scenarios, while also improving the efficiency and transferability. Our real-world case studies on vulnerability detection and classification further demonstrate the security implications of our attacks, highlighting the urgent need to further enhance the robustness of existing BCSD models.", "date": "2025-06-05", "categories": []}
{"id": "2506.05739", "title": "To Protect the LLM Agent Against the Prompt Injection Attack with Polymorphic Prompt", "abstract": "LLM agents are widely used as agents for customer support, content generation, and code assistance. However, they are vulnerable to prompt injection attacks, where adversarial inputs manipulate the model's behavior. Traditional defenses like input sanitization, guard models, and guardrails are either cumbersome or ineffective. In this paper, we propose a novel, lightweight defense mechanism called Polymorphic Prompt Assembling (PPA), which protects against prompt injection with near-zero overhead. The approach is based on the insight that prompt injection requires guessing and breaking the structure of the system prompt. By dynamically varying the structure of system prompts, PPA prevents attackers from predicting the prompt structure, thereby enhancing security without compromising performance. We conducted experiments to evaluate the effectiveness of PPA against existing attacks and compared it with other defense methods.", "date": "2025-06-06", "categories": []}
{"id": "2407.18213", "title": "Scaling Trends in Language Model Robustness", "abstract": "Increasing model size has unlocked a dazzling array of capabilities in modern language models. At the same time, even frontier models remain vulnerable to jailbreaks and prompt injections, despite concerted efforts to make them robust. As both attack and defense gain access to more compute, and as models become larger, what happens to robustness? We argue that to answer this question requires a \\emph{scaling} approach, which we employ in an extensive study of language model robustness across several classification tasks, model families, and adversarial attacks. We find that in the absence of explicit safety training, larger models are not consistently more robust; however, scale improves sample efficiency in adversarial training, though it worsens compute efficiency. Further, we find that increasing attack compute smoothly improves attack success rate against both undefended and adversarially trained models. Finally, after exploring robustness transfer across attacks and threat models, we combine attack and defense scaling rates to study the offense-defense balance. We find that while attack scaling outpaces adversarial training across all models studied, larger adversarially trained models might give defense the advantage in the long run. These results underscore the utility of the scaling lens, and provide a paradigm for evaluating future attacks and defenses on frontier models.", "date": "2024-07-25", "categories": []}
{"id": "2410.06895", "title": "Average Certified Radius is a Poor Metric for Randomized Smoothing", "abstract": "Randomized smoothing (RS) is popular for providing certified robustness guarantees against adversarial attacks. The average certified radius (ACR) has emerged as a widely used metric for tracking progress in RS. However, in this work, for the first time we show that ACR is a poor metric for evaluating robustness guarantees provided by RS. We theoretically prove not only that a trivial classifier can have arbitrarily large ACR, but also that ACR is extremely sensitive to improvements on easy samples. In addition, the comparison using ACR has a strong dependence on the certification budget. Empirically, we confirm that existing training strategies, though improving ACR, reduce the model's robustness on hard samples consistently. To strengthen our findings, we propose strategies, including explicitly discarding hard samples, reweighing the dataset with approximate certified radius, and extreme optimization for easy samples, to replicate the progress in RS training and even achieve the state-of-the-art ACR on CIFAR-10, without training for robustness on the full data distribution. Overall, our results suggest that ACR has introduced a strong undesired bias to the field, and its application should be discontinued in RS. Finally, we suggest using the empirical distribution of $p_A$, the accuracy of the base model on noisy data, as an alternative metric for RS.", "date": "2024-10-09", "categories": []}
{"id": "2410.11283", "title": "AdvBDGen: Adversarially Fortified Prompt-Specific Fuzzy Backdoor Generator Against LLM Alignment", "abstract": "With the growing adoption of reinforcement learning with human feedback (RLHF) for aligning large language models (LLMs), the risk of backdoor installation during alignment has increased, leading to unintended and harmful behaviors. Existing backdoor triggers are typically limited to fixed word patterns, making them detectable during data cleaning and easily removable post-poisoning. In this work, we explore the use of prompt-specific paraphrases as backdoor triggers, enhancing their stealth and resistance to removal during LLM alignment. We propose AdvBDGen, an adversarially fortified generative fine-tuning framework that automatically generates prompt-specific backdoors that are effective, stealthy, and transferable across models. AdvBDGen employs a generator-discriminator pair, fortified by an adversary, to ensure the installability and stealthiness of backdoors. It enables the crafting and successful installation of complex triggers using as little as 3% of the fine-tuning data. Once installed, these backdoors can jailbreak LLMs during inference, demonstrate improved stability against perturbations compared to traditional constant triggers, and are more challenging to remove. These findings underscore an urgent need for the research community to develop more robust defenses against adversarial backdoor threats in LLM alignment.", "date": "2024-10-15", "categories": []}
{"id": "2502.07987", "title": "Universal Adversarial Attack on Aligned Multimodal LLMs", "abstract": "We propose a universal adversarial attack on multimodal Large Language Models (LLMs) that leverages a single optimized image to override alignment safeguards across diverse queries and even multiple models. By backpropagating through the vision encoder and language head, we craft a synthetic image that forces the model to respond with a targeted phrase (e.g., \"Sure, here it is\") or otherwise unsafe content -- even for harmful prompts. In experiments on the SafeBench and MM-SafetyBench benchmarks, our method achieves higher attack success rates than existing baselines, including text-only universal prompts (e.g., up to 81% on certain models). We further demonstrate cross-model universality by training on several multimodal LLMs simultaneously. Additionally, a multi-answer variant of our approach produces more natural-sounding (yet still malicious) responses. These findings underscore critical vulnerabilities in current multimodal alignment and call for more robust adversarial defenses. We will release code and datasets under the Apache-2.0 license. Warning: some content generated by Multimodal LLMs in this paper may be offensive.", "date": "2025-02-11", "categories": []}
{"id": "2503.07697", "title": "PoisonedParrot: Subtle Data Poisoning Attacks to Elicit Copyright-Infringing Content from Large Language Models", "abstract": "As the capabilities of large language models (LLMs) continue to expand, their usage has become increasingly prevalent. However, as reflected in numerous ongoing lawsuits regarding LLM-generated content, addressing copyright infringement remains a significant challenge. In this paper, we introduce PoisonedParrot: the first stealthy data poisoning attack that induces an LLM to generate copyrighted content even when the model has not been directly trained on the specific copyrighted material. PoisonedParrot integrates small fragments of copyrighted text into the poison samples using an off-the-shelf LLM. Despite its simplicity, evaluated in a wide range of experiments, PoisonedParrot is surprisingly effective at priming the model to generate copyrighted content with no discernible side effects. Moreover, we discover that existing defenses are largely ineffective against our attack. Finally, we make the first attempt at mitigating copyright-infringement poisoning attacks by proposing a defense: ParrotTrap. We encourage the community to explore this emerging threat model further.", "date": "2025-03-10", "categories": []}
{"id": "2506.04823", "title": "Fool the Stoplight: Realistic Adversarial Patch Attacks on Traffic Light Detectors", "abstract": "Realistic adversarial attacks on various camera-based perception tasks of autonomous vehicles have been successfully demonstrated so far. However, only a few works considered attacks on traffic light detectors. This work shows how CNNs for traffic light detection can be attacked with printed patches. We propose a threat model, where each instance of a traffic light is attacked with a patch placed under it, and describe a training strategy. We demonstrate successful adversarial patch attacks in universal settings. Our experiments show realistic targeted red-to-green label-flipping attacks and attacks on pictogram classification. Finally, we perform a real-world evaluation with printed patches and demonstrate attacks in the lab settings with a mobile traffic light for construction sites and in a test area with stationary traffic lights. Our code is available at https://github.com/KASTEL-MobilityLab/attacks-on-traffic-light-detection.", "date": "2025-06-05", "categories": []}
{"id": "2506.04879", "title": "Invisible Backdoor Triggers in Image Editing Model via Deep Watermarking", "abstract": "Diffusion models have achieved remarkable progress in both image generation and editing. However, recent studies have revealed their vulnerability to backdoor attacks, in which specific patterns embedded in the input can manipulate the model's behavior. Most existing research in this area has proposed attack frameworks focused on the image generation pipeline, leaving backdoor attacks in image editing relatively unexplored. Among the few studies targeting image editing, most utilize visible triggers, which are impractical because they introduce noticeable alterations to the input image before editing. In this paper, we propose a novel attack framework that embeds invisible triggers into the image editing process via poisoned training data. We leverage off-the-shelf deep watermarking models to encode imperceptible watermarks as backdoor triggers. Our goal is to make the model produce the predefined backdoor target when it receives watermarked inputs, while editing clean images normally according to the given prompt. With extensive experiments across different watermarking models, the proposed method achieves promising attack success rates. In addition, the analysis results of the watermark characteristics in term of backdoor attack further support the effectiveness of our approach. The code is available at:https://github.com/aiiu-lab/BackdoorImageEditing", "date": "2025-06-05", "categories": []}
{"id": "2506.05032", "title": "Identifying and Understanding Cross-Class Features in Adversarial Training", "abstract": "Adversarial training (AT) has been considered one of the most effective methods for making deep neural networks robust against adversarial attacks, while the training mechanisms and dynamics of AT remain open research problems. In this paper, we present a novel perspective on studying AT through the lens of class-wise feature attribution. Specifically, we identify the impact of a key family of features on AT that are shared by multiple classes, which we call cross-class features. These features are typically useful for robust classification, which we offer theoretical evidence to illustrate through a synthetic data model. Through systematic studies across multiple model architectures and settings, we find that during the initial stage of AT, the model tends to learn more cross-class features until the best robustness checkpoint. As AT further squeezes the training robust loss and causes robust overfitting, the model tends to make decisions based on more class-specific features. Based on these discoveries, we further provide a unified view of two existing properties of AT, including the advantage of soft-label training and robust overfitting. Overall, these insights refine the current understanding of AT mechanisms and provide new perspectives on studying them. Our code is available at https://github.com/PKU-ML/Cross-Class-Features-AT.", "date": "2025-06-05", "categories": []}
{"id": "2505.23786", "title": "Mind the Gap: A Practical Attack on GGUF Quantization", "abstract": "With the increasing size of frontier LLMs, post-training quantization has become the standard for memory-efficient deployment. Recent work has shown that basic rounding-based quantization schemes pose security risks, as they can be exploited to inject malicious behaviors into quantized models that remain hidden in full precision. However, existing attacks cannot be applied to more complex quantization methods, such as the GGUF family used in the popular ollama and llama$.$cpp frameworks. In this work, we address this gap by introducing the first attack on GGUF. Our key insight is that the quantization error -- the difference between the full-precision weights and their (de-)quantized version -- provides sufficient flexibility to construct malicious quantized models that appear benign in full precision. Leveraging this, we develop an attack that trains the target malicious LLM while constraining its weights based on quantization errors. We demonstrate the effectiveness of our attack on three popular LLMs across nine GGUF quantization data types on three diverse attack scenarios: insecure code generation ($\\Delta$=$88.7\\%$), targeted content injection ($\\Delta$=$85.0\\%$), and benign instruction refusal ($\\Delta$=$30.1\\%$). Our attack highlights that (1) the most widely used post-training quantization method is susceptible to adversarial interferences, and (2) the complexity of quantization schemes alone is insufficient as a defense.", "date": "2025-05-24", "categories": []}
{"id": "2506.01224", "title": "Dirty and Clean-Label attack detection using GAN discriminators", "abstract": "Gathering enough images to train a deep computer vision model is a constant challenge. Unfortunately, collecting images from unknown sources can leave your model s behavior at risk of being manipulated by a dirty-label or clean-label attack unless the images are properly inspected. Manually inspecting each image-label pair is impractical and common poison-detection methods that involve re-training your model can be time consuming. This research uses GAN discriminators to protect a single class against mislabeled and different levels of modified images. The effect of said perturbation on a basic convolutional neural network classifier is also included for reference. The results suggest that after training on a single class, GAN discriminator s confidence scores can provide a threshold to identify mislabeled images and identify 100% of the tested poison starting at a perturbation epsilon magnitude of 0.20, after decision threshold calibration using in-class samples. Developers can use this report as a basis to train their own discriminators to protect high valued classes in their CV models.", "date": "2025-06-02", "categories": []}
{"id": "2506.03234", "title": "BadReward: Clean-Label Poisoning of Reward Models in Text-to-Image RLHF", "abstract": "Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning text-to-image (T2I) models with human preferences. However, RLHF's feedback mechanism also opens new pathways for adversaries. This paper demonstrates the feasibility of hijacking T2I models by poisoning a small fraction of preference data with natural-appearing examples. Specifically, we propose BadReward, a stealthy clean-label poisoning attack targeting the reward model in multi-modal RLHF. BadReward operates by inducing feature collisions between visually contradicted preference data instances, thereby corrupting the reward model and indirectly compromising the T2I model's integrity. Unlike existing alignment poisoning techniques focused on single (text) modality, BadReward is independent of the preference annotation process, enhancing its stealth and practical threat. Extensive experiments on popular T2I models show that BadReward can consistently guide the generation towards improper outputs, such as biased or violent imagery, for targeted concepts. Our findings underscore the amplified threat landscape for RLHF in multi-modal systems, highlighting the urgent need for robust defenses. Disclaimer. This paper contains uncensored toxic content that might be offensive or disturbing to the readers.", "date": "2025-06-03", "categories": []}
{"id": "2506.03350", "title": "Adversarial Attacks on Robotic Vision Language Action Models", "abstract": "The emergence of vision-language-action models (VLAs) for end-to-end control is reshaping the field of robotics by enabling the fusion of multimodal sensory inputs at the billion-parameter scale. The capabilities of VLAs stem primarily from their architectures, which are often based on frontier large language models (LLMs). However, LLMs are known to be susceptible to adversarial misuse, and given the significant physical risks inherent to robotics, questions remain regarding the extent to which VLAs inherit these vulnerabilities. Motivated by these concerns, in this work we initiate the study of adversarial attacks on VLA-controlled robots. Our main algorithmic contribution is the adaptation and application of LLM jailbreaking attacks to obtain complete control authority over VLAs. We find that textual attacks, which are applied once at the beginning of a rollout, facilitate full reachability of the action space of commonly used VLAs and often persist over longer horizons. This differs significantly from LLM jailbreaking literature, as attacks in the real world do not have to be semantically linked to notions of harm. We make all code available at https://github.com/eliotjones1/robogcg .", "date": "2025-06-03", "categories": []}
{"id": "2404.16873", "title": "AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs", "abstract": "Large Language Models (LLMs) are vulnerable to jailbreaking attacks that lead to generation of inappropriate or harmful content. Manual red-teaming requires a time-consuming search for adversarial prompts, whereas automatic adversarial prompt generation often leads to semantically meaningless attacks that do not scale well. In this paper, we present a novel method that uses another LLM, called AdvPrompter, to generate human-readable adversarial prompts in seconds. AdvPrompter, which is trained using an alternating optimization algorithm, generates suffixes that veil the input instruction without changing its meaning, such that the TargetLLM is lured to give a harmful response. Experimental results on popular open source TargetLLMs show highly competitive results on the AdvBench and HarmBench datasets, that also transfer to closed-source black-box LLMs. We also show that training on adversarial suffixes generated by AdvPrompter is a promising strategy for improving the robustness of LLMs to jailbreaking attacks.", "date": "2024-04-21", "categories": []}
{"id": "2410.01574", "title": "Adversarial Robustness of AI-Generated Image Detectors in the Real World", "abstract": "The rapid advancement of Generative Artificial Intelligence (GenAI) capabilities is accompanied by a concerning rise in its misuse. In particular the generation of credible misinformation in the form of images poses a significant threat to the public trust in democratic processes. Consequently, there is an urgent need to develop tools to reliably distinguish between authentic and AI-generated content. The majority of detection methods are based on neural networks that are trained to recognize forensic artifacts. In this work, we demonstrate that current state-of-the-art classifiers are vulnerable to adversarial examples under real-world conditions. Through extensive experiments, comprising four detection methods and five attack algorithms, we show that an attacker can dramatically decrease classification performance, without internal knowledge of the detector's architecture. Notably, most attacks remain effective even when images are degraded during the upload to, e.g., social media platforms. In a case study, we demonstrate that these robustness challenges are also found in commercial tools by conducting black-box attacks on HIVE, a proprietary online GenAI media detector. In addition, we evaluate the robustness of using generated features of a robust pre-trained model and showed that this increases the robustness, while not reaching the performance on benign inputs. These results, along with the increasing potential of GenAI to erode public trust, underscore the need for more research and new perspectives on methods to prevent its misuse.", "date": "2024-10-02", "categories": []}
{"id": "2410.03869", "title": "Chain-of-Jailbreak Attack for Image Generation Models via Editing Step by Step", "abstract": "Text-based image generation models, such as Stable Diffusion and DALL-E 3, hold significant potential in content creation and publishing workflows, making them the focus in recent years. Despite their remarkable capability to generate diverse and vivid images, considerable efforts are being made to prevent the generation of harmful content, such as abusive, violent, or pornographic material. To assess the safety of existing models, we introduce a novel jailbreaking method called Chain-of-Jailbreak (CoJ) attack, which compromises image generation models through a step-by-step editing process. Specifically, for malicious queries that cannot bypass the safeguards with a single prompt, we intentionally decompose the query into multiple sub-queries. The image generation models are then prompted to generate and iteratively edit images based on these sub-queries. To evaluate the effectiveness of our CoJ attack method, we constructed a comprehensive dataset, CoJ-Bench, encompassing nine safety scenarios, three types of editing operations, and three editing elements. Experiments on four widely-used image generation services provided by GPT-4V, GPT-4o, Gemini 1.5 and Gemini 1.5 Pro, demonstrate that our CoJ attack method can successfully bypass the safeguards of models for over 60% cases, which significantly outperforms other jailbreaking methods (i.e., 14%). Further, to enhance these models' safety against our CoJ attack method, we also propose an effective prompting-based method, Think Twice Prompting, that can successfully defend over 95% of CoJ attack. We release our dataset and code to facilitate the AI safety research.", "date": "2024-10-04", "categories": []}
{"id": "2503.04480", "title": "Poisoning Bayesian Inference via Data Deletion and Replication", "abstract": "Research in adversarial machine learning (AML) has shown that statistical models are vulnerable to maliciously altered data. However, despite advances in Bayesian machine learning models, most AML research remains concentrated on classical techniques. Therefore, we focus on extending the white-box model poisoning paradigm to attack generic Bayesian inference, highlighting its vulnerability in adversarial contexts. A suite of attacks are developed that allow an attacker to steer the Bayesian posterior toward a target distribution through the strategic deletion and replication of true observations, even when only sampling access to the posterior is available. Analytic properties of these algorithms are proven and their performance is empirically examined in both synthetic and real-world scenarios. With relatively little effort, the attacker is able to substantively alter the Bayesian's beliefs and, by accepting more risk, they can mold these beliefs to their will. By carefully constructing the adversarial posterior, surgical poisoning is achieved such that only targeted inferences are corrupted and others are minimally disturbed.", "date": "2025-03-06", "categories": []}
{"id": "2506.02016", "title": "Are classical deep neural networks weakly adversarially robust?", "abstract": "Adversarial attacks have received increasing attention and it has been widely recognized that classical DNNs have weak adversarial robustness. The most commonly used adversarial defense method, adversarial training, improves the adversarial accuracy of DNNs by generating adversarial examples and retraining the model. However, adversarial training requires a significant computational overhead. In this paper, inspired by existing studies focusing on the clustering properties of DNN output features at each layer and the Progressive Feedforward Collapse phenomenon, we propose a method for adversarial example detection and image recognition that uses layer-wise features to construct feature paths and computes the correlation between the examples feature paths and the class-centered feature paths. Experimental results show that the recognition method achieves 82.77% clean accuracy and 44.17% adversarial accuracy on the ResNet-20 with PFC. Compared to the adversarial training method with 77.64% clean accuracy and 52.94% adversarial accuracy, our method exhibits a trade-off without relying on computationally expensive defense strategies. Furthermore, on the standard ResNet-18, our method maintains this advantage with respective metrics of 80.01% and 46.1%. This result reveals inherent adversarial robustness in DNNs, challenging the conventional understanding of the weak adversarial robustness in DNNs.", "date": "2025-05-28", "categories": []}
{"id": "2506.03075", "title": "Agnostic Learning under Targeted Poisoning: Optimal Rates and the Role of Randomness", "abstract": "We study the problem of learning in the presence of an adversary that can corrupt an $\\eta$ fraction of the training examples with the goal of causing failure on a specific test point. In the realizable setting, prior work established that the optimal error under such instance-targeted poisoning attacks scales as $\\Theta(d\\eta)$, where $d$ is the VC dimension of the hypothesis class arXiv:2210.02713. In this work, we resolve the corresponding question in the agnostic setting. We show that the optimal excess error is $\\tilde{\\Theta}(\\sqrt{d\\eta})$, answering one of the main open problems left by Hanneke et al. To achieve this rate, it is necessary to use randomized learners: Hanneke et al. showed that deterministic learners can be forced to suffer error close to 1, even under small amounts of poisoning. Perhaps surprisingly, our upper bound remains valid even when the learner's random bits are fully visible to the adversary . In the other direction, our lower bound is stronger than standard PAC-style bounds: instead of tailoring a hard distribution separately for each sample size, we exhibit a single fixed distribution under which the adversary can enforce an excess error of $\\Omega(\\sqrt{d\\eta})$ infinitely often.", "date": "2025-06-03", "categories": []}
{"id": "2311.16646", "title": "Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective", "abstract": "Dataset distillation offers a potential means to enhance data efficiency in deep learning. Recent studies have shown its ability to counteract backdoor risks present in original training samples. In this study, we delve into the theoretical aspects of backdoor attacks and dataset distillation based on kernel methods. We introduce two new theory-driven trigger pattern generation methods specialized for dataset distillation. Following a comprehensive set of analyses and experiments, we show that our optimization-based trigger design framework informs effective backdoor attacks on dataset distillation. Notably, datasets poisoned by our designed trigger prove resilient against conventional backdoor attack detection and mitigation methods. Our empirical results validate that the triggers developed using our approaches are proficient at executing resilient backdoor attacks.", "date": "2023-11-28", "categories": []}
{"id": "2408.09181", "title": "PADetBench: Towards Benchmarking Physical Attacks against Object Detection", "abstract": "Physical attacks against object detection have gained increasing attention due to their significant practical implications. However, conducting physical experiments is extremely time-consuming and labor-intensive. Moreover, physical dynamics and cross-domain transformation are challenging to strictly regulate in the real world, leading to unaligned evaluation and comparison, severely hindering the development of physically robust models. To accommodate these challenges, we explore utilizing realistic simulation to thoroughly and rigorously benchmark physical attacks with fairness under controlled physical dynamics and cross-domain transformation. This resolves the problem of capturing identical adversarial images that cannot be achieved in the real world. Our benchmark includes 20 physical attack methods, 48 object detectors, comprehensive physical dynamics, and evaluation metrics. We also provide end-to-end pipelines for dataset generation, detection, evaluation, and further analysis. In addition, we perform 8064 groups of evaluation based on our benchmark, which includes both overall evaluation and further detailed ablation studies for controlled physical dynamics. Through these experiments, we provide in-depth analyses of physical attack performance and physical adversarial robustness, draw valuable observations, and discuss potential directions for future research.\n  Codebase: https://github.com/JiaweiLian/Benchmarking_Physical_Attack", "date": "2024-08-17", "categories": []}
{"id": "2410.17401", "title": "AdvAgent: Controllable Blackbox Red-teaming on Web Agents", "abstract": "Foundation model-based agents are increasingly used to automate complex tasks, enhancing efficiency and productivity. However, their access to sensitive resources and autonomous decision-making also introduce significant security risks, where successful attacks could lead to severe consequences. To systematically uncover these vulnerabilities, we propose AdvAgent, a black-box red-teaming framework for attacking web agents. Unlike existing approaches, AdvAgent employs a reinforcement learning-based pipeline to train an adversarial prompter model that optimizes adversarial prompts using feedback from the black-box agent. With careful attack design, these prompts effectively exploit agent weaknesses while maintaining stealthiness and controllability. Extensive evaluations demonstrate that AdvAgent achieves high success rates against state-of-the-art GPT-4-based web agents across diverse web tasks. Furthermore, we find that existing prompt-based defenses provide only limited protection, leaving agents vulnerable to our framework. These findings highlight critical vulnerabilities in current web agents and emphasize the urgent need for stronger defense mechanisms. We release code at https://ai-secure.github.io/AdvAgent/.", "date": "2024-10-22", "categories": []}
{"id": "2411.09749", "title": "RenderBender: A Survey on Adversarial Attacks Using Differentiable Rendering", "abstract": "Differentiable rendering techniques like Gaussian Splatting and Neural Radiance Fields have become powerful tools for generating high-fidelity models of 3D objects and scenes. Their ability to produce both physically plausible and differentiable models of scenes are key ingredient needed to produce physically plausible adversarial attacks on DNNs. However, the adversarial machine learning community has yet to fully explore these capabilities, partly due to differing attack goals (e.g., misclassification, misdetection) and a wide range of possible scene manipulations used to achieve them (e.g., alter texture, mesh). This survey contributes the first framework that unifies diverse goals and tasks, facilitating easy comparison of existing work, identifying research gaps, and highlighting future directions - ranging from expanding attack goals and tasks to account for new modalities, state-of-the-art models, tools, and pipelines, to underscoring the importance of studying real-world threats in complex scenes.", "date": "2024-11-14", "categories": []}
{"id": "2501.05962", "title": "Effective faking of verbal deception detection with target-aligned adversarial attacks", "abstract": "Background: Deception detection through analysing language is a promising avenue using both human judgments and automated machine learning judgments. For both forms of credibility assessment, automated adversarial attacks that rewrite deceptive statements to appear truthful pose a serious threat. Methods: We used a dataset of 243 truthful and 262 fabricated autobiographical stories in a deception detection task for humans and machine learning models. A large language model was tasked to rewrite deceptive statements so that they appear truthful. In Study 1, humans who made a deception judgment or used the detailedness heuristic and two machine learning models (a fine-tuned language model and a simple n-gram model) judged original or adversarial modifications of deceptive statements. In Study 2, we manipulated the target alignment of the modifications, i.e. tailoring the attack to whether the statements would be assessed by humans or computer models. Results: When adversarial modifications were aligned with their target, human (d=-0.07 and d=-0.04) and machine judgments (51% accuracy) dropped to the chance level. When the attack was not aligned with the target, both human heuristics judgments (d=0.30 and d=0.36) and machine learning predictions (63-78%) were significantly better than chance. Conclusions: Easily accessible language models can effectively help anyone fake deception detection efforts both by humans and machine learning models. Robustness against adversarial modifications for humans and machines depends on that target alignment. We close with suggestions on advancing deception research with adversarial attack designs and techniques.", "date": "2025-01-10", "categories": []}
{"id": "2501.18626", "title": "The TIP of the Iceberg: Revealing a Hidden Class of Task-in-Prompt Adversarial Attacks on LLMs", "abstract": "We present a novel class of jailbreak adversarial attacks on LLMs, termed Task-in-Prompt (TIP) attacks. Our approach embeds sequence-to-sequence tasks (e.g., cipher decoding, riddles, code execution) into the model's prompt to indirectly generate prohibited inputs. To systematically assess the effectiveness of these attacks, we introduce the PHRYGE benchmark. We demonstrate that our techniques successfully circumvent safeguards in six state-of-the-art language models, including GPT-4o and LLaMA 3.2. Our findings highlight critical weaknesses in current LLM safety alignments and underscore the urgent need for more sophisticated defence strategies.\n  Warning: this paper contains examples of unethical inquiries used solely for research purposes.", "date": "2025-01-27", "categories": []}
{"id": "2503.12827", "title": "GSBA$^K$: $top$-$K$ Geometric Score-based Black-box Attack", "abstract": "Existing score-based adversarial attacks mainly focus on crafting $top$-1 adversarial examples against classifiers with single-label classification. Their attack success rate and query efficiency are often less than satisfactory, particularly under small perturbation requirements; moreover, the vulnerability of classifiers with multi-label learning is yet to be studied. In this paper, we propose a comprehensive surrogate free score-based attack, named \\b geometric \\b score-based \\b black-box \\b attack (GSBA$^K$), to craft adversarial examples in an aggressive $top$-$K$ setting for both untargeted and targeted attacks, where the goal is to change the $top$-$K$ predictions of the target classifier. We introduce novel gradient-based methods to find a good initial boundary point to attack. Our iterative method employs novel gradient estimation techniques, particularly effective in $top$-$K$ setting, on the decision boundary to effectively exploit the geometry of the decision boundary. Additionally, GSBA$^K$ can be used to attack against classifiers with $top$-$K$ multi-label learning. Extensive experimental results on ImageNet and PASCAL VOC datasets validate the effectiveness of GSBA$^K$ in crafting $top$-$K$ adversarial examples.", "date": "2025-03-17", "categories": []}
{"id": "2505.14862", "title": "Replay Attacks Against Audio Deepfake Detection", "abstract": "We show how replay attacks undermine audio deepfake detection: By playing and re-recording deepfake audio through various speakers and microphones, we make spoofed samples appear authentic to the detection model. To study this phenomenon in more detail, we introduce ReplayDF, a dataset of recordings derived from M-AILABS and MLAAD, featuring 109 speaker-microphone combinations across six languages and four TTS models. It includes diverse acoustic conditions, some highly challenging for detection. Our analysis of six open-source detection models across five datasets reveals significant vulnerability, with the top-performing W2V2-AASIST model's Equal Error Rate (EER) surging from 4.7% to 18.2%. Even with adaptive Room Impulse Response (RIR) retraining, performance remains compromised with an 11.0% EER. We release ReplayDF for non-commercial research use.", "date": "2025-05-20", "categories": []}
{"id": "2506.00191", "title": "Heterogeneous Graph Backdoor Attack", "abstract": "Heterogeneous Graph Neural Networks (HGNNs) excel in modeling complex, multi-typed relationships across diverse domains, yet their vulnerability to backdoor attacks remains unexplored. To address this gap, we conduct the first investigation into the susceptibility of HGNNs to existing graph backdoor attacks, revealing three critical issues: (1) high attack budget required for effective backdoor injection, (2) inefficient and unreliable backdoor activation, and (3) inaccurate attack effectiveness evaluation. To tackle these issues, we propose the Heterogeneous Graph Backdoor Attack (HGBA), the first backdoor attack specifically designed for HGNNs, introducing a novel relation-based trigger mechanism that establishes specific connections between a strategically selected trigger node and poisoned nodes via the backdoor metapath. HGBA achieves efficient and stealthy backdoor injection with minimal structural modifications and supports easy backdoor activation through two flexible strategies: Self-Node Attack and Indiscriminate Attack. Additionally, we improve the ASR measurement protocol, enabling a more accurate assessment of attack effectiveness. Extensive experiments demonstrate that HGBA far surpasses multiple state-of-the-art graph backdoor attacks in black-box settings, efficiently attacking HGNNs with low attack budgets. Ablation studies show that the strength of HBGA benefits from our trigger node selection method and backdoor metapath selection strategy. In addition, HGBA shows superior robustness against node feature perturbations and multiple types of existing graph backdoor defense mechanisms. Finally, extension experiments demonstrate that the relation-based trigger mechanism can effectively extend to tasks in homogeneous graph scenarios, thereby posing severe threats to broader security-critical domains.", "date": "2025-05-30", "categories": []}
{"id": "2506.00280", "title": "3D Gaussian Splat Vulnerabilities", "abstract": "With 3D Gaussian Splatting (3DGS) being increasingly used in safety-critical applications, how can an adversary manipulate the scene to cause harm? We introduce CLOAK, the first attack that leverages view-dependent Gaussian appearances - colors and textures that change with viewing angle - to embed adversarial content visible only from specific viewpoints. We further demonstrate DAGGER, a targeted adversarial attack directly perturbing 3D Gaussians without access to underlying training data, deceiving multi-stage object detectors e.g., Faster R-CNN, through established methods such as projected gradient descent. These attacks highlight underexplored vulnerabilities in 3DGS, introducing a new potential threat to robotic learning for autonomous navigation and other safety-critical 3DGS applications.", "date": "2025-05-30", "categories": []}
{"id": "2506.00281", "title": "Adversarial Threat Vectors and Risk Mitigation for Retrieval-Augmented Generation Systems", "abstract": "Retrieval-Augmented Generation (RAG) systems, which integrate Large Language Models (LLMs) with external knowledge sources, are vulnerable to a range of adversarial attack vectors. This paper examines the importance of RAG systems through recent industry adoption trends and identifies the prominent attack vectors for RAG: prompt injection, data poisoning, and adversarial query manipulation. We analyze these threats under risk management lens, and propose robust prioritized control list that includes risk-mitigating actions like input validation, adversarial training, and real-time monitoring.", "date": "2025-05-30", "categories": []}
{"id": "2506.00325", "title": "Towards Effective and Efficient Adversarial Defense with Diffusion Models for Robust Visual Tracking", "abstract": "Although deep learning-based visual tracking methods have made significant progress, they exhibit vulnerabilities when facing carefully designed adversarial attacks, which can lead to a sharp decline in tracking performance. To address this issue, this paper proposes for the first time a novel adversarial defense method based on denoise diffusion probabilistic models, termed DiffDf, aimed at effectively improving the robustness of existing visual tracking methods against adversarial attacks. DiffDf establishes a multi-scale defense mechanism by combining pixel-level reconstruction loss, semantic consistency loss, and structural similarity loss, effectively suppressing adversarial perturbations through a gradual denoising process. Extensive experimental results on several mainstream datasets show that the DiffDf method demonstrates excellent generalization performance for trackers with different architectures, significantly improving various evaluation metrics while achieving real-time inference speeds of over 30 FPS, showcasing outstanding defense performance and efficiency. Codes are available at https://github.com/pgao-lab/DiffDf.", "date": "2025-05-31", "categories": []}
{"id": "2506.00548", "title": "Con Instruction: Universal Jailbreaking of Multimodal Large Language Models via Non-Textual Modalities", "abstract": "Existing attacks against multimodal language models (MLLMs) primarily communicate instructions through text accompanied by adversarial images. In contrast, we exploit the capabilities of MLLMs to interpret non-textual instructions, specifically, adversarial images or audio generated by our novel method, Con Instruction. We optimize these adversarial examples to align closely with target instructions in the embedding space, revealing the detrimental implications of MLLMs' sophisticated understanding. Unlike prior work, our method does not require training data or preprocessing of textual instructions. While these non-textual adversarial examples can effectively bypass MLLM safety mechanisms, their combination with various text inputs substantially amplifies attack success. We further introduce a new Attack Response Categorization (ARC) framework, which evaluates both the quality of the model's response and its relevance to the malicious instructions. Experimental results demonstrate that Con Instruction effectively bypasses safety mechanisms in multiple vision- and audio-language models, including LLaVA-v1.5, InternVL, Qwen-VL, and Qwen-Audio, evaluated on two standard benchmarks: AdvBench and SafeBench. Specifically, our method achieves the highest attack success rates, reaching 81.3% and 86.6% on LLaVA-v1.5 (13B). On the defense side, we explore various countermeasures against our attacks and uncover a substantial performance gap among existing techniques. Our implementation is made publicly available.", "date": "2025-05-31", "categories": []}
{"id": "2506.01267", "title": "Adversarial learning for nonparametric regression: Minimax rate and adaptive estimation", "abstract": "Despite tremendous advancements of machine learning models and algorithms in various application domains, they are known to be vulnerable to subtle, natural or intentionally crafted perturbations in future input data, known as adversarial attacks. While numerous adversarial learning methods have been proposed, fundamental questions about their statistical optimality in robust loss remain largely unanswered. In particular, the minimax rate of convergence and the construction of rate-optimal estimators under future $X$-attacks are yet to be worked out.\n  In this paper, we address this issue in the context of nonparametric regression, under suitable assumptions on the smoothness of the regression function and the geometric structure of the input perturbation set. We first establish the minimax rate of convergence under adversarial $L_q$-risks with $1 \\leq q \\leq \\infty$ and propose a piecewise local polynomial estimator that achieves the minimax optimality. The established minimax rate elucidates how the smoothness level and perturbation magnitude affect the fundamental limit of adversarial learning under future $X$-attacks. Furthermore, we construct a data-driven adaptive estimator that is shown to achieve, within a logarithmic factor, the optimal rate across a broad scale of nonparametric and adversarial classes.", "date": "2025-06-02", "categories": []}
{"id": "2506.01307", "title": "Align is not Enough: Multimodal Universal Jailbreak Attack against Multimodal Large Language Models", "abstract": "Large Language Models (LLMs) have evolved into Multimodal Large Language Models (MLLMs), significantly enhancing their capabilities by integrating visual information and other types, thus aligning more closely with the nature of human intelligence, which processes a variety of data forms beyond just text. Despite advancements, the undesirable generation of these models remains a critical concern, particularly due to vulnerabilities exposed by text-based jailbreak attacks, which have represented a significant threat by challenging existing safety protocols. Motivated by the unique security risks posed by the integration of new and old modalities for MLLMs, we propose a unified multimodal universal jailbreak attack framework that leverages iterative image-text interactions and transfer-based strategy to generate a universal adversarial suffix and image. Our work not only highlights the interaction of image-text modalities can be used as a critical vulnerability but also validates that multimodal universal jailbreak attacks can bring higher-quality undesirable generations across different MLLMs. We evaluate the undesirable context generation of MLLMs like LLaVA, Yi-VL, MiniGPT4, MiniGPT-v2, and InstructBLIP, and reveal significant multimodal safety alignment issues, highlighting the inadequacy of current safety mechanisms against sophisticated multimodal attacks. This study underscores the urgent need for robust safety measures in MLLMs, advocating for a comprehensive review and enhancement of security protocols to mitigate potential risks associated with multimodal capabilities.", "date": "2025-06-02", "categories": []}
{"id": "2506.01511", "title": "Enhancing Diffusion-based Unrestricted Adversarial Attacks via Adversary Preferences Alignment", "abstract": "Preference alignment in diffusion models has primarily focused on benign human preferences (e.g., aesthetic). In this paper, we propose a novel perspective: framing unrestricted adversarial example generation as a problem of aligning with adversary preferences. Unlike benign alignment, adversarial alignment involves two inherently conflicting preferences: visual consistency and attack effectiveness, which often lead to unstable optimization and reward hacking (e.g., reducing visual quality to improve attack success). To address this, we propose APA (Adversary Preferences Alignment), a two-stage framework that decouples conflicting preferences and optimizes each with differentiable rewards. In the first stage, APA fine-tunes LoRA to improve visual consistency using rule-based similarity reward. In the second stage, APA updates either the image latent or prompt embedding based on feedback from a substitute classifier, guided by trajectory-level and step-wise rewards. To enhance black-box transferability, we further incorporate a diffusion augmentation strategy. Experiments demonstrate that APA achieves significantly better attack transferability while maintaining high visual consistency, inspiring further research to approach adversarial attacks from an alignment perspective. Code will be available at https://github.com/deep-kaixun/APA.", "date": "2025-06-02", "categories": []}
{"id": "2506.01591", "title": "Silence is Golden: Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation", "abstract": "Advances in talking-head animation based on Latent Diffusion Models (LDM) enable the creation of highly realistic, synchronized videos. These fabricated videos are indistinguishable from real ones, increasing the risk of potential misuse for scams, political manipulation, and misinformation. Hence, addressing these ethical concerns has become a pressing issue in AI security. Recent proactive defense studies focused on countering LDM-based models by adding perturbations to portraits. However, these methods are ineffective at protecting reference portraits from advanced image-to-video animation. The limitations are twofold: 1) they fail to prevent images from being manipulated by audio signals, and 2) diffusion-based purification techniques can effectively eliminate protective perturbations. To address these challenges, we propose Silencer, a two-stage method designed to proactively protect the privacy of portraits. First, a nullifying loss is proposed to ignore audio control in talking-head generation. Second, we apply anti-purification loss in LDM to optimize the inverted latent feature to generate robust perturbations. Extensive experiments demonstrate the effectiveness of Silencer in proactively protecting portrait privacy. We hope this work will raise awareness among the AI security community regarding critical ethical issues related to talking-head generation techniques. Code: https://github.com/yuangan/Silencer.", "date": "2025-06-02", "categories": []}
{"id": "2506.01849", "title": "Trojan Horse Hunt in Time Series Forecasting for Space Operations", "abstract": "This competition hosted on Kaggle (https://www.kaggle.com/competitions/trojan-horse-hunt-in-space) is the first part of a series of follow-up competitions and hackathons related to the \"Assurance for Space Domain AI Applications\" project funded by the European Space Agency (https://assurance-ai.space-codev.org/). The competition idea is based on one of the real-life AI security threats identified within the project -- the adversarial poisoning of continuously fine-tuned satellite telemetry forecasting models. The task is to develop methods for finding and reconstructing triggers (trojans) in advanced models for satellite telemetry forecasting used in safety-critical space operations. Participants are provided with 1) a large public dataset of real-life multivariate satellite telemetry (without triggers), 2) a reference model trained on the clean data, 3) a set of poisoned neural hierarchical interpolation (N-HiTS) models for time series forecasting trained on the dataset with injected triggers, and 4) Jupyter notebook with the training pipeline and baseline algorithm (the latter will be published in the last month of the competition). The main task of the competition is to reconstruct a set of 45 triggers (i.e., short multivariate time series segments) injected into the training data of the corresponding set of 45 poisoned models. The exact characteristics (i.e., shape, amplitude, and duration) of these triggers must be identified by participants. The popular Neural Cleanse method is adopted as a baseline, but it is not designed for time series analysis and new approaches are necessary for the task. The impact of the competition is not limited to the space domain, but also to many other safety-critical applications of advanced time series analysis where model poisoning may lead to serious consequences.", "date": "2025-06-02", "categories": []}
{"id": "2405.14189", "title": "Efficient Universal Goal Hijacking with Semantics-guided Prompt Organization", "abstract": "Universal goal hijacking is a kind of prompt injection attack that forces LLMs to return a target malicious response for arbitrary normal user prompts. The previous methods achieve high attack performance while being too cumbersome and time-consuming. Also, they have concentrated solely on optimization algorithms, overlooking the crucial role of the prompt. To this end, we propose a method called POUGH that incorporates an efficient optimization algorithm and two semantics-guided prompt organization strategies. Specifically, our method starts with a sampling strategy to select representative prompts from a candidate pool, followed by a ranking strategy that prioritizes them. Given the sequentially ranked prompts, our method employs an iterative optimization algorithm to generate a fixed suffix that can concatenate to arbitrary user prompts for universal goal hijacking. Experiments conducted on four popular LLMs and ten types of target responses verified the effectiveness.", "date": "2024-05-23", "categories": []}
{"id": "2409.01249", "title": "Adversarial Pruning: A Survey and Benchmark of Pruning Methods for Adversarial Robustness", "abstract": "Recent work has proposed neural network pruning techniques to reduce the size of a network while preserving robustness against adversarial examples, i.e., well-crafted inputs inducing a misclassification. These methods, which we refer to as adversarial pruning methods, involve complex and articulated designs, making it difficult to analyze the differences and establish a fair and accurate comparison. In this work, we overcome these issues by surveying current adversarial pruning methods and proposing a novel taxonomy to categorize them based on two main dimensions: the pipeline, defining when to prune; and the specifics, defining how to prune. We then highlight the limitations of current empirical analyses and propose a novel, fair evaluation benchmark to address them. We finally conduct an empirical re-evaluation of current adversarial pruning methods and discuss the results, highlighting the shared traits of top-performing adversarial pruning methods, as well as common issues. We welcome contributions in our publicly-available benchmark at https://github.com/pralab/AdversarialPruningBenchmark", "date": "2024-09-02", "categories": []}
{"id": "2502.03758", "title": "Improving Adversarial Robustness via Phase and Amplitude-aware Prompting", "abstract": "Deep neural networks are found to be vulnerable to adversarial perturbations. The prompt-based defense has been increasingly studied due to its high efficiency. However, existing prompt-based defenses mainly exploited mixed prompt patterns, where critical patterns closely related to object semantics lack sufficient focus. The phase and amplitude spectra have been proven to be highly related to specific semantic patterns and crucial for robustness. To this end, in this paper, we propose a Phase and Amplitude-aware Prompting (PAP) defense. Specifically, we construct phase-level and amplitude-level prompts for each class, and adjust weights for prompting according to the model's robust performance under these prompts during training. During testing, we select prompts for each image using its predicted label to obtain the prompted image, which is inputted to the model to get the final prediction. Experimental results demonstrate the effectiveness of our method.", "date": "2025-02-06", "categories": []}
{"id": "2503.00140", "title": "Approaching the Harm of Gradient Attacks While Only Flipping Labels", "abstract": "Machine learning systems deployed in distributed or federated environments are highly susceptible to adversarial manipulations, particularly availability attacks -adding imperceptible perturbations to training data, thereby rendering the trained model unavailable. Prior research in distributed machine learning has demonstrated such adversarial effects through the injection of gradients or data poisoning. In this study, we aim to enhance comprehension of the potential of weaker (and more probable) adversaries by posing the following inquiry: Can availability attacks be inflicted solely through the flipping of a subset of training labels, without altering features, and under a strict flipping budget? We analyze the extent of damage caused by constrained label flipping attacks. Focusing on a distributed classification problem, (1) we propose a novel formalization of label flipping attacks on logistic regression models and derive a greedy algorithm that is provably optimal at each training step. (2) To demonstrate that availability attacks can be approached by label flipping alone, we show that a budget of only $0.1\\%$ of labels at each training step can reduce the accuracy of the model by $6\\%$, and that some models can perform worse than random guessing when up to $25\\%$ of labels are flipped. (3) We shed light on an interesting interplay between what the attacker gains from more write-access versus what they gain from more flipping budget. (4) we define and compare the power of targeted label flipping attack to that of an untargeted label flipping attack.", "date": "2025-02-28", "categories": []}
{"id": "2505.05528", "title": "X-Transfer Attacks: Towards Super Transferable Adversarial Attacks on CLIP", "abstract": "As Contrastive Language-Image Pre-training (CLIP) models are increasingly adopted for diverse downstream tasks and integrated into large vision-language models (VLMs), their susceptibility to adversarial perturbations has emerged as a critical concern. In this work, we introduce \\textbf{X-Transfer}, a novel attack method that exposes a universal adversarial vulnerability in CLIP. X-Transfer generates a Universal Adversarial Perturbation (UAP) capable of deceiving various CLIP encoders and downstream VLMs across different samples, tasks, and domains. We refer to this property as \\textbf{super transferability}--a single perturbation achieving cross-data, cross-domain, cross-model, and cross-task adversarial transferability simultaneously. This is achieved through \\textbf{surrogate scaling}, a key innovation of our approach. Unlike existing methods that rely on fixed surrogate models, which are computationally intensive to scale, X-Transfer employs an efficient surrogate scaling strategy that dynamically selects a small subset of suitable surrogates from a large search space. Extensive evaluations demonstrate that X-Transfer significantly outperforms previous state-of-the-art UAP methods, establishing a new benchmark for adversarial transferability across CLIP models. The code is publicly available in our \\href{https://github.com/HanxunH/XTransferBench}{GitHub repository}.", "date": "2025-05-08", "categories": []}
{"id": "2505.22604", "title": "Adversarially Robust AI-Generated Image Detection for Free: An Information Theoretic Perspective", "abstract": "Rapid advances in Artificial Intelligence Generated Images (AIGI) have facilitated malicious use, such as forgery and misinformation. Therefore, numerous methods have been proposed to detect fake images. Although such detectors have been proven to be universally vulnerable to adversarial attacks, defenses in this field are scarce. In this paper, we first identify that adversarial training (AT), widely regarded as the most effective defense, suffers from performance collapse in AIGI detection. Through an information-theoretic lens, we further attribute the cause of collapse to feature entanglement, which disrupts the preservation of feature-label mutual information. Instead, standard detectors show clear feature separation. Motivated by this difference, we propose Training-free Robust Detection via Information-theoretic Measures (TRIM), the first training-free adversarial defense for AIGI detection. TRIM builds on standard detectors and quantifies feature shifts using prediction entropy and KL divergence. Extensive experiments across multiple datasets and attacks validate the superiority of our TRIM, e.g., outperforming the state-of-the-art defense by 33.88% (28.91%) on ProGAN (GenImage), while well maintaining original accuracy.", "date": "2025-05-28", "categories": []}
{"id": "2505.24141", "title": "The Butterfly Effect in Pathology: Exploring Security in Pathology Foundation Models", "abstract": "With the widespread adoption of pathology foundation models in both research and clinical decision support systems, exploring their security has become a critical concern. However, despite their growing impact, the vulnerability of these models to adversarial attacks remains largely unexplored. In this work, we present the first systematic investigation into the security of pathology foundation models for whole slide image~(WSI) analysis against adversarial attacks. Specifically, we introduce the principle of \\textit{local perturbation with global impact} and propose a label-free attack framework that operates without requiring access to downstream task labels. Under this attack framework, we revise four classical white-box attack methods and redefine the perturbation budget based on the characteristics of WSI. We conduct comprehensive experiments on three representative pathology foundation models across five datasets and six downstream tasks. Despite modifying only 0.1\\% of patches per slide with imperceptible noise, our attack leads to downstream accuracy degradation that can reach up to 20\\% in the worst cases. Furthermore, we analyze key factors that influence attack success, explore the relationship between patch-level vulnerability and semantic content, and conduct a preliminary investigation into potential defence strategies. These findings lay the groundwork for future research on the adversarial robustness and reliable deployment of pathology foundation models. Our code is publicly available at: https://github.com/Jiashuai-Liu-hmos/Attack-WSI-pathology-foundation-models.", "date": "2025-05-30", "categories": []}
{"id": "2505.24227", "title": "Light as Deception: GPT-driven Natural Relighting Against Vision-Language Pre-training Models", "abstract": "While adversarial attacks on vision-and-language pretraining (VLP) models have been explored, generating natural adversarial samples crafted through realistic and semantically meaningful perturbations remains an open challenge. Existing methods, primarily designed for classification tasks, struggle when adapted to VLP models due to their restricted optimization spaces, leading to ineffective attacks or unnatural artifacts. To address this, we propose \\textbf{LightD}, a novel framework that generates natural adversarial samples for VLP models via semantically guided relighting. Specifically, LightD leverages ChatGPT to propose context-aware initial lighting parameters and integrates a pretrained relighting model (IC-light) to enable diverse lighting adjustments. LightD expands the optimization space while ensuring perturbations align with scene semantics. Additionally, gradient-based optimization is applied to the reference lighting image to further enhance attack effectiveness while maintaining visual naturalness. The effectiveness and superiority of the proposed LightD have been demonstrated across various VLP models in tasks such as image captioning and visual question answering.", "date": "2025-05-30", "categories": []}
{"id": "2505.24369", "title": "Adversarial Preference Learning for Robust LLM Alignment", "abstract": "Modern language models often rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors. However, they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation, (2) the vast diversity of potential adversarial attacks, and (3) the risk of feedback bias and reward hacking. To address these challenges, we introduce Adversarial Preference Learning (APL), an iterative adversarial training method incorporating three key innovations. First, a direct harmfulness metric based on the model's intrinsic preference probabilities, eliminating reliance on external assessment. Second, a conditional generative attacker that synthesizes input-specific adversarial variations. Third, an iterative framework with automated closed-loop feedback, enabling continuous adaptation through vulnerability discovery and mitigation. Experiments on Mistral-7B-Instruct-v0.3 demonstrate that APL significantly enhances robustness, achieving 83.33% harmlessness win rate over the base model (evaluated by GPT-4o), reducing harmful outputs from 5.88% to 0.43% (measured by LLaMA-Guard), and lowering attack success rate by up to 65% according to HarmBench. Notably, APL maintains competitive utility, with an MT-Bench score of 6.59 (comparable to the baseline 6.78) and an LC-WinRate of 46.52% against the base model.", "date": "2025-05-30", "categories": []}
{"id": "2505.24519", "title": "AMIA: Automatic Masking and Joint Intention Analysis Makes LVLMs Robust Jailbreak Defenders", "abstract": "We introduce AMIA, a lightweight, inference-only defense for Large Vision-Language Models (LVLMs) that (1) Automatically Masks a small set of text-irrelevant image patches to disrupt adversarial perturbations, and (2) conducts joint Intention Analysis to uncover and mitigate hidden harmful intents before response generation. Without any retraining, AMIA improves defense success rates across diverse LVLMs and jailbreak benchmarks from an average of 52.4% to 81.7%, preserves general utility with only a 2% average accuracy drop, and incurs only modest inference overhead. Ablation confirms both masking and intention analysis are essential for a robust safety-utility trade-off.", "date": "2025-05-30", "categories": []}
{"id": "2505.24654", "title": "Black-box Adversarial Attacks on CNN-based SLAM Algorithms", "abstract": "Continuous advancements in deep learning have led to significant progress in feature detection, resulting in enhanced accuracy in tasks like Simultaneous Localization and Mapping (SLAM). Nevertheless, the vulnerability of deep neural networks to adversarial attacks remains a challenge for their reliable deployment in applications, such as navigation of autonomous agents. Even though CNN-based SLAM algorithms are a growing area of research there is a notable absence of a comprehensive presentation and examination of adversarial attacks targeting CNN-based feature detectors, as part of a SLAM system. Our work introduces black-box adversarial perturbations applied to the RGB images fed into the GCN-SLAM algorithm. Our findings on the TUM dataset [30] reveal that even attacks of moderate scale can lead to tracking failure in as many as 76% of the frames. Moreover, our experiments highlight the catastrophic impact of attacking depth instead of RGB input images on the SLAM system.", "date": "2025-05-30", "categories": []}
{"id": "2505.24703", "title": "PatchDEMUX: A Certifiably Robust Framework for Multi-label Classifiers Against Adversarial Patches", "abstract": "Deep learning techniques have enabled vast improvements in computer vision technologies. Nevertheless, these models are vulnerable to adversarial patch attacks which catastrophically impair performance. The physically realizable nature of these attacks calls for certifiable defenses, which feature provable guarantees on robustness. While certifiable defenses have been successfully applied to single-label classification, limited work has been done for multi-label classification. In this work, we present PatchDEMUX, a certifiably robust framework for multi-label classifiers against adversarial patches. Our approach is a generalizable method which can extend any existing certifiable defense for single-label classification; this is done by considering the multi-label classification task as a series of isolated binary classification problems to provably guarantee robustness. Furthermore, in the scenario where an attacker is limited to a single patch we propose an additional certification procedure that can provide tighter robustness bounds. Using the current state-of-the-art (SOTA) single-label certifiable defense PatchCleanser as a backbone, we find that PatchDEMUX can achieve non-trivial robustness on the MS-COCO and PASCAL VOC datasets while maintaining high clean performance", "date": "2025-05-30", "categories": []}
{"id": "2402.13459", "title": "Learning to Poison Large Language Models for Downstream Manipulation", "abstract": "The advent of Large Language Models (LLMs) has marked significant achievements in language processing and reasoning capabilities. Despite their advancements, LLMs face vulnerabilities to data poisoning attacks, where the adversary inserts backdoor triggers into training data to manipulate outputs. This work further identifies additional security risks in LLMs by designing a new data poisoning attack tailored to exploit the supervised fine-tuning (SFT) process. We propose a novel gradient-guided backdoor trigger learning (GBTL) algorithm to identify adversarial triggers efficiently, ensuring an evasion of detection by conventional defenses while maintaining content integrity. Through experimental validation across various language model tasks, including sentiment analysis, domain generation, and question answering, our poisoning strategy demonstrates a high success rate in compromising various LLMs' outputs. We further propose two defense strategies against data poisoning attacks, including in-context learning (ICL) and continuous learning (CL), which effectively rectify the behavior of LLMs and significantly reduce the decline in performance. Our work highlights the significant security risks present during SFT of LLMs and the necessity of safeguarding LLMs against data poisoning attacks.", "date": "2024-02-21", "categories": []}
{"id": "2402.17390", "title": "Robustness-Congruent Adversarial Training for Secure Machine Learning Model Updates", "abstract": "Machine-learning models demand periodic updates to improve their average accuracy, exploiting novel architectures and additional data. However, a newly updated model may commit mistakes the previous model did not make. Such misclassifications are referred to as negative flips, experienced by users as a regression of performance. In this work, we show that this problem also affects robustness to adversarial examples, hindering the development of secure model update practices. In particular, when updating a model to improve its adversarial robustness, previously ineffective adversarial attacks on some inputs may become successful, causing a regression in the perceived security of the system. We propose a novel technique, named robustness-congruent adversarial training, to address this issue. It amounts to fine-tuning a model with adversarial training, while constraining it to retain higher robustness on the samples for which no adversarial example was found before the update. We show that our algorithm and, more generally, learning with non-regression constraints, provides a theoretically-grounded framework to train consistent estimators. Our experiments on robust models for computer vision confirm that both accuracy and robustness, even if improved after model update, can be affected by negative flips, and our robustness-congruent adversarial training can mitigate the problem, outperforming competing baseline methods.", "date": "2024-02-27", "categories": []}
{"id": "2410.22725", "title": "One Prompt to Verify Your Models: Black-Box Text-to-Image Models Verification via Non-Transferable Adversarial Attacks", "abstract": "Recently, various types of Text-to-Image (T2I) models have emerged (such as DALL-E and Stable Diffusion), and showing their advantages in different aspects. Therefore, some third-party service platforms collect different model interfaces and provide cheaper API services and more flexibility in T2I model selections. However, this also raises a new security concern: Are these third-party services truly offering the models they claim?\n  To answer this question, we first define the concept of T2I model verification, which aims to determine whether a black-box target model is identical to a given white-box reference T2I model. After that, we propose VerifyPrompt, which performs T2I model verification through a special designed verify prompt. Intuitionally, the verify prompt is an adversarial prompt for the target model without transferability for other models. It makes the target model generate a specific image while making other models produce entirely different images. Specifically, VerifyPrompt utilizes the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to optimize the cosine similarity of a prompt's text encoding, generating verify prompts. Finally, by computing the CLIP-text similarity scores between the prompts the generated images, VerifyPrompt can determine whether the target model aligns with the reference model. Experimental results demonstrate that VerifyPrompt consistently achieves over 90\\% accuracy across various T2I models, confirming its effectiveness in practical model platforms (such as Hugging Face).", "date": "2024-10-30", "categories": []}
{"id": "2412.16359", "title": "Human-Readable Adversarial Prompts: An Investigation into LLM Vulnerabilities Using Situational Context", "abstract": "As the AI systems become deeply embedded in social media platforms, we've uncovered a concerning security vulnerability that goes beyond traditional adversarial attacks. It becomes important to assess the risks of LLMs before the general public use them on social media platforms to avoid any adverse impacts. Unlike obvious nonsensical text strings that safety systems can easily catch, our work reveals that human-readable situation-driven adversarial full-prompts that leverage situational context are effective but much harder to detect. We found that skilled attackers can exploit the vulnerabilities in open-source and proprietary LLMs to make a malicious user query safe for LLMs, resulting in generating a harmful response. This raises an important question about the vulnerabilities of LLMs. To measure the robustness against human-readable attacks, which now present a potent threat, our research makes three major contributions. First, we developed attacks that use movie scripts as situational contextual frameworks, creating natural-looking full-prompts that trick LLMs into generating harmful content. Second, we developed a method to transform gibberish adversarial text into readable, innocuous content that still exploits vulnerabilities when used within the full-prompts. Finally, we enhanced the AdvPrompter framework with p-nucleus sampling to generate diverse human-readable adversarial texts that significantly improve attack effectiveness against models like GPT-3.5-Turbo-0125 and Gemma-7b. Our findings show that these systems can be manipulated to operate beyond their intended ethical boundaries when presented with seemingly normal prompts that contain hidden adversarial elements. By identifying these vulnerabilities, we aim to drive the development of more robust safety mechanisms that can withstand sophisticated attacks in real-world applications.", "date": "2024-12-20", "categories": []}
{"id": "2412.16555", "title": "Divide and Conquer: A Hybrid Strategy Defeats Multimodal Large Language Models", "abstract": "Large language models (LLMs) are widely applied in various fields of society due to their powerful reasoning, understanding, and generation capabilities. However, the security issues associated with these models are becoming increasingly severe. Jailbreaking attacks, as an important method for detecting vulnerabilities in LLMs, have been explored by researchers who attempt to induce these models to generate harmful content through various attack methods. Nevertheless, existing jailbreaking methods face numerous limitations, such as excessive query counts, limited coverage of jailbreak modalities, low attack success rates, and simplistic evaluation methods. To overcome these constraints, this paper proposes a multimodal jailbreaking method: JMLLM. This method integrates multiple strategies to perform comprehensive jailbreak attacks across text, visual, and auditory modalities. Additionally, we contribute a new and comprehensive dataset for multimodal jailbreaking research: TriJail, which includes jailbreak prompts for all three modalities. Experiments on the TriJail dataset and the benchmark dataset AdvBench, conducted on 13 popular LLMs, demonstrate advanced attack success rates and significant reduction in time overhead.", "date": "2024-12-21", "categories": []}
{"id": "2502.11647", "title": "DELMAN: Dynamic Defense Against Large Language Model Jailbreaking with Model Editing", "abstract": "Large Language Models (LLMs) are widely applied in decision making, but their deployment is threatened by jailbreak attacks, where adversarial users manipulate model behavior to bypass safety measures. Existing defense mechanisms, such as safety fine-tuning and model editing, either require extensive parameter modifications or lack precision, leading to performance degradation on general tasks, which is unsuitable to post-deployment safety alignment. To address these challenges, we propose DELMAN (Dynamic Editing for LLMs JAilbreak DefeNse), a novel approach leveraging direct model editing for precise, dynamic protection against jailbreak attacks. DELMAN directly updates a minimal set of relevant parameters to neutralize harmful behaviors while preserving the model's utility. To avoid triggering a safe response in benign context, we incorporate KL-divergence regularization to ensure the updated model remains consistent with the original model when processing benign queries. Experimental results demonstrate that DELMAN outperforms baseline methods in mitigating jailbreak attacks while preserving the model's utility, and adapts seamlessly to new attack instances, providing a practical and efficient solution for post-deployment model protection.", "date": "2025-02-17", "categories": []}
{"id": "2505.22839", "title": "How Do Diffusion Models Improve Adversarial Robustness?", "abstract": "Recent findings suggest that diffusion models significantly enhance empirical adversarial robustness. While some intuitive explanations have been proposed, the precise mechanisms underlying these improvements remain unclear. In this work, we systematically investigate how and how well diffusion models improve adversarial robustness. First, we observe that diffusion models intriguingly increase, rather than decrease, the $\\ell_p$ distance to clean samples--challenging the intuition that purification denoises inputs closer to the original data. Second, we find that the purified images are heavily influenced by the internal randomness of diffusion models, where a compression effect arises within each randomness configuration. Motivated by this observation, we evaluate robustness under fixed randomness and find that the improvement drops to approximately 24% on CIFAR-10--substantially lower than prior reports approaching 70%. Importantly, we show that this remaining robustness gain strongly correlates with the model's ability to compress the input space, revealing the compression rate as a reliable robustness indicator without requiring gradient-based analysis. Our findings provide novel insights into the mechanisms underlying diffusion-based purification, and offer guidance for developing more effective and principled adversarial purification systems.", "date": "2025-05-28", "categories": []}
{"id": "2505.22943", "title": "Can LLMs Deceive CLIP? Benchmarking Adversarial Compositionality of Pre-trained Multimodal Representation via Text Updates", "abstract": "While pre-trained multimodal representations (e.g., CLIP) have shown impressive capabilities, they exhibit significant compositional vulnerabilities leading to counterintuitive judgments. We introduce Multimodal Adversarial Compositionality (MAC), a benchmark that leverages large language models (LLMs) to generate deceptive text samples to exploit these vulnerabilities across different modalities and evaluates them through both sample-wise attack success rate and group-wise entropy-based diversity. To improve zero-shot methods, we propose a self-training approach that leverages rejection-sampling fine-tuning with diversity-promoting filtering, which enhances both attack success rate and sample diversity. Using smaller language models like Llama-3.1-8B, our approach demonstrates superior performance in revealing compositional vulnerabilities across various multimodal representations, including images, videos, and audios.", "date": "2025-05-28", "categories": []}
{"id": "2505.23192", "title": "Fooling the Watchers: Breaking AIGC Detectors via Semantic Prompt Attacks", "abstract": "The rise of text-to-image (T2I) models has enabled the synthesis of photorealistic human portraits, raising serious concerns about identity misuse and the robustness of AIGC detectors. In this work, we propose an automated adversarial prompt generation framework that leverages a grammar tree structure and a variant of the Monte Carlo tree search algorithm to systematically explore the semantic prompt space. Our method generates diverse, controllable prompts that consistently evade both open-source and commercial AIGC detectors. Extensive experiments across multiple T2I models validate its effectiveness, and the approach ranked first in a real-world adversarial AIGC detection competition. Beyond attack scenarios, our method can also be used to construct high-quality adversarial datasets, providing valuable resources for training and evaluating more robust AIGC detection and defense systems.", "date": "2025-05-29", "categories": []}
{"id": "2505.23266", "title": "Disrupting Vision-Language Model-Driven Navigation Services via Adversarial Object Fusion", "abstract": "We present Adversarial Object Fusion (AdvOF), a novel attack framework targeting vision-and-language navigation (VLN) agents in service-oriented environments by generating adversarial 3D objects. While foundational models like Large Language Models (LLMs) and Vision Language Models (VLMs) have enhanced service-oriented navigation systems through improved perception and decision-making, their integration introduces vulnerabilities in mission-critical service workflows. Existing adversarial attacks fail to address service computing contexts, where reliability and quality-of-service (QoS) are paramount. We utilize AdvOF to investigate and explore the impact of adversarial environments on the VLM-based perception module of VLN agents. In particular, AdvOF first precisely aggregates and aligns the victim object positions in both 2D and 3D space, defining and rendering adversarial objects. Then, we collaboratively optimize the adversarial object with regularization between the adversarial and victim object across physical properties and VLM perceptions. Through assigning importance weights to varying views, the optimization is processed stably and multi-viewedly by iterative fusions from local updates and justifications. Our extensive evaluations demonstrate AdvOF can effectively degrade agent performance under adversarial conditions while maintaining minimal interference with normal navigation tasks. This work advances the understanding of service security in VLM-powered navigation systems, providing computational foundations for robust service composition in physical-world deployments.", "date": "2025-05-29", "categories": []}
{"id": "2505.23313", "title": "Adversarial Semantic and Label Perturbation Attack for Pedestrian Attribute Recognition", "abstract": "Pedestrian Attribute Recognition (PAR) is an indispensable task in human-centered research and has made great progress in recent years with the development of deep neural networks. However, the potential vulnerability and anti-interference ability have still not been fully explored. To bridge this gap, this paper proposes the first adversarial attack and defense framework for pedestrian attribute recognition. Specifically, we exploit both global- and patch-level attacks on the pedestrian images, based on the pre-trained CLIP-based PAR framework. It first divides the input pedestrian image into non-overlapping patches and embeds them into feature embeddings using a projection layer. Meanwhile, the attribute set is expanded into sentences using prompts and embedded into attribute features using a pre-trained CLIP text encoder. A multi-modal Transformer is adopted to fuse the obtained vision and text tokens, and a feed-forward network is utilized for attribute recognition. Based on the aforementioned PAR framework, we adopt the adversarial semantic and label-perturbation to generate the adversarial noise, termed ASL-PAR. We also design a semantic offset defense strategy to suppress the influence of adversarial attacks. Extensive experiments conducted on both digital domains (i.e., PETA, PA100K, MSP60K, RAPv2) and physical domains fully validated the effectiveness of our proposed adversarial attack and defense strategies for the pedestrian attribute recognition. The source code of this paper will be released on https://github.com/Event-AHU/OpenPAR.", "date": "2025-05-29", "categories": []}
{"id": "2505.23760", "title": "Model Immunization from a Condition Number Perspective", "abstract": "Model immunization aims to pre-train models that are difficult to fine-tune on harmful tasks while retaining their utility on other non-harmful tasks. Though prior work has shown empirical evidence for immunizing text-to-image models, the key understanding of when immunization is possible and a precise definition of an immunized model remain unclear. In this work, we propose a framework, based on the condition number of a Hessian matrix, to analyze model immunization for linear models. Building on this framework, we design an algorithm with regularization terms to control the resulting condition numbers after pre-training. Empirical results on linear models and non-linear deep-nets demonstrate the effectiveness of the proposed algorithm on model immunization. The code is available at https://github.com/amberyzheng/model-immunization-cond-num.", "date": "2025-05-29", "categories": []}
{"id": "2410.06851", "title": "Understanding Model Ensemble in Transferable Adversarial Attack", "abstract": "Model ensemble adversarial attack has become a powerful method for generating transferable adversarial examples that can target even unknown models, but its theoretical foundation remains underexplored. To address this gap, we provide early theoretical insights that serve as a roadmap for advancing model ensemble adversarial attack. We first define transferability error to measure the error in adversarial transferability, alongside concepts of diversity and empirical model ensemble Rademacher complexity. We then decompose the transferability error into vulnerability, diversity, and a constant, which rigidly explains the origin of transferability error in model ensemble attack: the vulnerability of an adversarial example to ensemble components, and the diversity of ensemble components. Furthermore, we apply the latest mathematical tools in information theory to bound the transferability error using complexity and generalization terms, contributing to three practical guidelines for reducing transferability error: (1) incorporating more surrogate models, (2) increasing their diversity, and (3) reducing their complexity in cases of overfitting. Finally, extensive experiments with 54 models validate our theoretical framework, representing a significant step forward in understanding transferable model ensemble adversarial attacks.", "date": "2024-10-09", "categories": []}
{"id": "2504.08411", "title": "A Knowledge-guided Adversarial Defense for Resisting Malicious Visual Manipulation", "abstract": "Malicious applications of visual manipulation have raised serious threats to the security and reputation of users in many fields. To alleviate these issues, adversarial noise-based defenses have been enthusiastically studied in recent years. However, ``data-only\" methods tend to distort fake samples in the low-level feature space rather than the high-level semantic space, leading to limitations in resisting malicious manipulation. Frontier research has shown that integrating knowledge in deep learning can produce reliable and generalizable solutions. Inspired by these, we propose a knowledge-guided adversarial defense (KGAD) to actively force malicious manipulation models to output semantically confusing samples. Specifically, in the process of generating adversarial noise, we focus on constructing significant semantic confusions at the domain-specific knowledge level, and exploit a metric closely related to visual perception to replace the general pixel-wise metrics. The generated adversarial noise can actively interfere with the malicious manipulation model by triggering knowledge-guided and perception-related disruptions in the fake samples. To validate the effectiveness of the proposed method, we conduct qualitative and quantitative experiments on human perception and visual quality assessment. The results on two different tasks both show that our defense provides better protection compared to state-of-the-art methods and achieves great generalizability.", "date": "2025-04-11", "categories": []}
{"id": "2505.21556", "title": "Benign-to-Toxic Jailbreaking: Inducing Harmful Responses from Harmless Prompts", "abstract": "Optimization-based jailbreaks typically adopt the Toxic-Continuation setting in large vision-language models (LVLMs), following the standard next-token prediction objective. In this setting, an adversarial image is optimized to make the model predict the next token of a toxic prompt. However, we find that the Toxic-Continuation paradigm is effective at continuing already-toxic inputs, but struggles to induce safety misalignment when explicit toxic signals are absent. We propose a new paradigm: Benign-to-Toxic (B2T) jailbreak. Unlike prior work, we optimize adversarial images to induce toxic outputs from benign conditioning. Since benign conditioning contains no safety violations, the image alone must break the model's safety mechanisms. Our method outperforms prior approaches, transfers in black-box settings, and complements text-based jailbreaks. These results reveal an underexplored vulnerability in multimodal alignment and introduce a fundamentally new direction for jailbreak approaches.", "date": "2025-05-26", "categories": []}
{"id": "2505.21609", "title": "Preventing Adversarial AI Attacks Against Autonomous Situational Awareness: A Maritime Case Study", "abstract": "Adversarial artificial intelligence (AI) attacks pose a significant threat to autonomous transportation, such as maritime vessels, that rely on AI components. Malicious actors can exploit these systems to deceive and manipulate AI-driven operations. This paper addresses three critical research challenges associated with adversarial AI: the limited scope of traditional defences, inadequate security metrics, and the need to build resilience beyond model-level defences. To address these challenges, we propose building defences utilising multiple inputs and data fusion to create defensive components and an AI security metric as a novel approach toward developing more secure AI systems. We name this approach the Data Fusion Cyber Resilience (DFCR) method, and we evaluate it through real-world demonstrations and comprehensive quantitative analyses, comparing a system built with the DFCR method against single-input models and models utilising existing state-of-the-art defences. The findings show that the DFCR approach significantly enhances resilience against adversarial machine learning attacks in maritime autonomous system operations, achieving up to a 35\\% reduction in loss for successful multi-pronged perturbation attacks, up to a 100\\% reduction in loss for successful adversarial patch attacks and up to 100\\% reduction in loss for successful spoofing attacks when using these more resilient systems. We demonstrate how DFCR and DFCR confidence scores can reduce adversarial AI contact confidence and improve decision-making by the system, even when typical adversarial defences have been compromised. Ultimately, this work contributes to the development of more secure and resilient AI-driven systems against adversarial attacks.", "date": "2025-05-27", "categories": []}
{"id": "2505.21742", "title": "What is Adversarial Training for Diffusion Models?", "abstract": "We answer the question in the title, showing that adversarial training (AT) for diffusion models (DMs) fundamentally differs from classifiers: while AT in classifiers enforces output invariance, AT in DMs requires equivariance to keep the diffusion process aligned with the data distribution. AT is a way to enforce smoothness in the diffusion flow, improving robustness to outliers and corrupted data. Unlike prior art, our method makes no assumptions about the noise model and integrates seamlessly into diffusion training by adding random noise, similar to randomized smoothing, or adversarial noise, akin to AT. This enables intrinsic capabilities such as handling noisy data, dealing with extreme variability such as outliers, preventing memorization, and improving robustness. We rigorously evaluate our approach with proof-of-concept datasets with known distributions in low- and high-dimensional space, thereby taking a perfect measure of errors; we further evaluate on standard benchmarks such as CIFAR-10, CelebA and LSUN Bedroom, showing strong performance under severe noise, data corruption, and iterative adversarial attacks.", "date": "2025-05-27", "categories": []}
{"id": "2505.21967", "title": "Seeing the Threat: Vulnerabilities in Vision-Language Models to Adversarial Attack", "abstract": "Large Vision-Language Models (LVLMs) have shown remarkable capabilities across a wide range of multimodal tasks. However, their integration of visual inputs introduces expanded attack surfaces, thereby exposing them to novel security vulnerabilities. In this work, we conduct a systematic representational analysis to uncover why conventional adversarial attacks can circumvent the safety mechanisms embedded in LVLMs. We further propose a novel two stage evaluation framework for adversarial attacks on LVLMs. The first stage differentiates among instruction non compliance, outright refusal, and successful adversarial exploitation. The second stage quantifies the degree to which the model's output fulfills the harmful intent of the adversarial prompt, while categorizing refusal behavior into direct refusals, soft refusals, and partial refusals that remain inadvertently helpful. Finally, we introduce a normative schema that defines idealized model behavior when confronted with harmful prompts, offering a principled target for safety alignment in multimodal systems.", "date": "2025-05-28", "categories": []}
{"id": "2505.22486", "title": "Understanding Adversarial Training with Energy-based Models", "abstract": "We aim at using Energy-based Model (EBM) framework to better understand adversarial training (AT) in classifiers, and additionally to analyze the intrinsic generative capabilities of robust classifiers. By viewing standard classifiers through an energy lens, we begin by analyzing how the energies of adversarial examples, generated by various attacks, differ from those of the natural samples. The central focus of our work is to understand the critical phenomena of Catastrophic Overfitting (CO) and Robust Overfitting (RO) in AT from an energy perspective. We analyze the impact of existing AT approaches on the energy of samples during training and observe that the behavior of the ``delta energy' -- change in energy between original sample and its adversarial counterpart -- diverges significantly when CO or RO occurs. After a thorough analysis of these energy dynamics and their relationship with overfitting, we propose a novel regularizer, the Delta Energy Regularizer (DER), designed to smoothen the energy landscape during training. We demonstrate that DER is effective in mitigating both CO and RO across multiple benchmarks. We further show that robust classifiers, when being used as generative models, have limits in handling trade-off between image quality and variability. We propose an improved technique based on a local class-wise principal component analysis (PCA) and energy-based guidance for better class-specific initialization and adaptive stopping, enhancing sample diversity and generation quality. Considering that we do not explicitly train for generative modeling, we achieve a competitive Inception Score (IS) and Fr\\'echet inception distance (FID) compared to hybrid discriminative-generative models.", "date": "2025-05-28", "categories": []}
{"id": "2406.03143", "title": "ZeroPur: Succinct Training-Free Adversarial Purification", "abstract": "Adversarial purification is a kind of defense technique that can defend against various unseen adversarial attacks without modifying the victim classifier. Existing methods often depend on external generative models or cooperation between auxiliary functions and victim classifiers. However, retraining generative models, auxiliary functions, or victim classifiers relies on the domain of the fine-tuned dataset and is computation-consuming. In this work, we suppose that adversarial images are outliers of the natural image manifold, and the purification process can be considered as returning them to this manifold. Following this assumption, we present a simple adversarial purification method without further training to purify adversarial images, called ZeroPur. ZeroPur contains two steps: given an adversarial example, Guided Shift obtains the shifted embedding of the adversarial example by the guidance of its blurred counterparts; after that, Adaptive Projection constructs a directional vector by this shifted embedding to provide momentum, projecting adversarial images onto the manifold adaptively. ZeroPur is independent of external models and requires no retraining of victim classifiers or auxiliary functions, relying solely on victim classifiers themselves to achieve purification. Extensive experiments on three datasets (CIFAR-10, CIFAR-100, and ImageNet-1K) using various classifier architectures (ResNet, WideResNet) demonstrate that our method achieves state-of-the-art robust performance. The code will be publicly available.", "date": "2024-06-05", "categories": []}
{"id": "2410.12671", "title": "New Paradigm of Adversarial Training: Releasing Accuracy-Robustness Trade-Off via Dummy Class", "abstract": "Adversarial Training (AT) is one of the most effective methods to enhance the robustness of Deep Neural Networks (DNNs). However, existing AT methods suffer from an inherent accuracy-robustness trade-off. Previous works have studied this issue under the current AT paradigm, but still face over 10% accuracy reduction without significant robustness improvement over simple baselines such as PGD-AT. This inherent trade-off raises a question: Whether the current AT paradigm, which assumes to learn corresponding benign and adversarial samples as the same class, inappropriately mixes clean and robust objectives that may be essentially inconsistent. In fact, our empirical results show that up to 40% of CIFAR-10 adversarial samples always fail to satisfy such an assumption across various AT methods and robust models, explicitly indicating the room for improvement of the current AT paradigm. To relax from this overstrict assumption and the tension between clean and robust learning, in this work, we propose a new AT paradigm by introducing an additional dummy class for each original class, aiming to accommodate hard adversarial samples with shifted distribution after perturbation. The robustness w.r.t. these adversarial samples can be achieved by runtime recovery from the predicted dummy classes to the corresponding original ones, without conflicting with the clean objective on accuracy of benign samples. Finally, based on our new paradigm, we propose a novel DUmmy Classes-based Adversarial Training (DUCAT) method that concurrently improves accuracy and robustness in a plug-and-play manner only relevant to logits, loss, and a proposed two-hot soft label-based supervised signal. Our method outperforms state-of-the-art (SOTA) benchmarks, effectively releasing the current trade-off. The code is available at https://github.com/FlaAI/DUCAT.", "date": "2024-10-16", "categories": []}
{"id": "2410.21952", "title": "On the Robustness of Adversarial Training Against Uncertainty Attacks", "abstract": "In learning problems, the noise inherent to the task at hand hinders the possibility to infer without a certain degree of uncertainty. Quantifying this uncertainty, regardless of its wide use, assumes high relevance for security-sensitive applications. Within these scenarios, it becomes fundamental to guarantee good (i.e., trustworthy) uncertainty measures, which downstream modules can securely employ to drive the final decision-making process. However, an attacker may be interested in forcing the system to produce either (i) highly uncertain outputs jeopardizing the system's availability or (ii) low uncertainty estimates, making the system accept uncertain samples that would instead require a careful inspection (e.g., human intervention). Therefore, it becomes fundamental to understand how to obtain robust uncertainty estimates against these kinds of attacks. In this work, we reveal both empirically and theoretically that defending against adversarial examples, i.e., carefully perturbed samples that cause misclassification, additionally guarantees a more secure, trustworthy uncertainty estimate under common attack scenarios without the need for an ad-hoc defense strategy. To support our claims, we evaluate multiple adversarial-robust models from the publicly available benchmark RobustBench on the CIFAR-10 and ImageNet datasets.", "date": "2024-10-29", "categories": []}
{"id": "2501.05359", "title": "SC-Pro: Training-Free Framework for Defending Unsafe Image Synthesis Attack", "abstract": "With advances in diffusion models, image generation has shown significant performance improvements. This raises concerns about the potential abuse of image generation, such as the creation of explicit or violent images, commonly referred to as Not Safe For Work (NSFW) content. To address this, the Stable Diffusion model includes several safety checkers to censor initial text prompts and final output images generated from the model. However, recent research has shown that these safety checkers have vulnerabilities against adversarial attacks, allowing them to generate NSFW images. In this paper, we find that these adversarial attacks are not robust to small changes in text prompts or input latents. Based on this, we propose SC-Pro (Spherical or Circular Probing), a training-free framework that easily defends against adversarial attacks generating NSFW images. Moreover, we develop an approach that utilizes one-step diffusion models for efficient NSFW detection (SC-Pro-o), further reducing computational resources. We demonstrate the superiority of our method in terms of performance and applicability.", "date": "2025-01-09", "categories": []}
{"id": "2502.06832", "title": "Optimizing Robustness and Accuracy in Mixture of Experts: A Dual-Model Approach", "abstract": "Mixture of Experts (MoE) have shown remarkable success in leveraging specialized expert networks for complex machine learning tasks. However, their susceptibility to adversarial attacks presents a critical challenge for deployment in robust applications. This paper addresses the critical question of how to incorporate robustness into MoEs while maintaining high natural accuracy. We begin by analyzing the vulnerability of MoE components, finding that expert networks are notably more susceptible to adversarial attacks than the router. Based on this insight, we propose a targeted robust training technique that integrates a novel loss function to enhance the adversarial robustness of MoE, requiring only the robustification of one additional expert without compromising training or inference efficiency. Building on this, we introduce a dual-model strategy that linearly combines a standard MoE model with our robustified MoE model using a smoothing parameter. This approach allows for flexible control over the robustness-accuracy trade-off. We further provide theoretical foundations by deriving certified robustness bounds for both the single MoE and the dual-model. To push the boundaries of robustness and accuracy, we propose a novel joint training strategy JTDMoE for the dual-model. This joint training enhances both robustness and accuracy beyond what is achievable with separate models. Experimental results on CIFAR-10 and TinyImageNet datasets using ResNet18 and Vision Transformer (ViT) architectures demonstrate the effectiveness of our proposed methods. The code is publicly available at https://github.com/TIML-Group/Robust-MoE-Dual-Model.", "date": "2025-02-05", "categories": []}
{"id": "2505.20824", "title": "MedSentry: Understanding and Mitigating Safety Risks in Medical LLM Multi-Agent Systems", "abstract": "As large language models (LLMs) are increasingly deployed in healthcare, ensuring their safety, particularly within collaborative multi-agent configurations, is paramount. In this paper we introduce MedSentry, a benchmark comprising 5 000 adversarial medical prompts spanning 25 threat categories with 100 subthemes. Coupled with this dataset, we develop an end-to-end attack-defense evaluation pipeline to systematically analyze how four representative multi-agent topologies (Layers, SharedPool, Centralized, and Decentralized) withstand attacks from 'dark-personality' agents. Our findings reveal critical differences in how these architectures handle information contamination and maintain robust decision-making, exposing their underlying vulnerability mechanisms. For instance, SharedPool's open information sharing makes it highly susceptible, whereas Decentralized architectures exhibit greater resilience thanks to inherent redundancy and isolation. To mitigate these risks, we propose a personality-scale detection and correction mechanism that identifies and rehabilitates malicious agents, restoring system safety to near-baseline levels. MedSentry thus furnishes both a rigorous evaluation framework and practical defense strategies that guide the design of safer LLM-based multi-agent systems in medical domains.", "date": "2025-05-27", "categories": []}
{"id": "2505.21074", "title": "Red-Teaming Text-to-Image Systems by Rule-based Preference Modeling", "abstract": "Text-to-image (T2I) models raise ethical and safety concerns due to their potential to generate inappropriate or harmful images. Evaluating these models' security through red-teaming is vital, yet white-box approaches are limited by their need for internal access, complicating their use with closed-source models. Moreover, existing black-box methods often assume knowledge about the model's specific defense mechanisms, limiting their utility in real-world commercial API scenarios. A significant challenge is how to evade unknown and diverse defense mechanisms. To overcome this difficulty, we propose a novel Rule-based Preference modeling Guided Red-Teaming (RPG-RT), which iteratively employs LLM to modify prompts to query and leverages feedback from T2I systems for fine-tuning the LLM. RPG-RT treats the feedback from each iteration as a prior, enabling the LLM to dynamically adapt to unknown defense mechanisms. Given that the feedback is often labeled and coarse-grained, making it difficult to utilize directly, we further propose rule-based preference modeling, which employs a set of rules to evaluate desired or undesired feedback, facilitating finer-grained control over the LLM's dynamic adaptation process. Extensive experiments on nineteen T2I systems with varied safety mechanisms, three online commercial API services, and T2V models verify the superiority and practicality of our approach.", "date": "2025-05-27", "categories": []}
{"id": "2505.21140", "title": "HeteroBA: A Structure-Manipulating Backdoor Attack on Heterogeneous Graphs", "abstract": "Heterogeneous graph neural networks (HGNNs) have recently drawn increasing attention for modeling complex multi-relational data in domains such as recommendation, finance, and social networks. While existing research has been largely focused on enhancing HGNNs' predictive performance, their robustness and security, especially under backdoor attacks, remain underexplored. In this paper, we propose a novel Heterogeneous Backdoor Attack (HeteroBA) framework for node classification tasks on heterogeneous graphs. HeteroBA inserts carefully crafted trigger nodes with realistic features and targeted structural connections, leveraging attention-based and clustering-based strategies to select influential auxiliary nodes for effective trigger propagation, thereby causing the model to misclassify specific nodes into a target label while maintaining accuracy on clean data. Experimental results on three datasets and various HGNN architectures demonstrate that HeteroBA achieves high attack success rates with minimal impact on the clean accuracy. Our method sheds light on potential vulnerabilities in HGNNs and calls for more robust defenses against backdoor threats in multi-relational graph scenarios.", "date": "2025-05-27", "categories": []}
{"id": "2505.21181", "title": "Boosting Adversarial Transferability via High-Frequency Augmentation and Hierarchical-Gradient Fusion", "abstract": "Adversarial attacks have become a significant challenge in the security of machine learning models, particularly in the context of black-box defense strategies. Existing methods for enhancing adversarial transferability primarily focus on the spatial domain. This paper presents Frequency-Space Attack (FSA), a new adversarial attack framework that effectively integrates frequency-domain and spatial-domain transformations. FSA combines two key techniques: (1) High-Frequency Augmentation, which applies Fourier transform with frequency-selective amplification to diversify inputs and emphasize the critical role of high-frequency components in adversarial attacks, and (2) Hierarchical-Gradient Fusion, which merges multi-scale gradient decomposition and fusion to capture both global structures and fine-grained details, resulting in smoother perturbations. Our experiment demonstrates that FSA consistently outperforms state-of-the-art methods across various black-box models. Notably, our proposed FSA achieves an average attack success rate increase of 23.6% compared with BSR (CVPR 2024) on eight black-box defense models.", "date": "2025-05-27", "categories": []}
{"id": "2505.21380", "title": "PHISH in MESH: Korean Adversarial Phonetic Substitution and Phonetic-Semantic Feature Integration Defense", "abstract": "As malicious users increasingly employ phonetic substitution to evade hate speech detection, researchers have investigated such strategies. However, two key challenges remain. First, existing studies have overlooked the Korean language, despite its vulnerability to phonetic perturbations due to its phonographic nature. Second, prior work has primarily focused on constructing datasets rather than developing architectural defenses. To address these challenges, we propose (1) PHonetic-Informed Substitution for Hangul (PHISH) that exploits the phonological characteristics of the Korean writing system, and (2) Mixed Encoding of Semantic-pHonetic features (MESH) that enhances the detector's robustness by incorporating phonetic information at the architectural level. Our experimental results demonstrate the effectiveness of our proposed methods on both perturbed and unperturbed datasets, suggesting that they not only improve detection performance but also reflect realistic adversarial behaviors employed by malicious users.", "date": "2025-05-27", "categories": []}
{"id": "2505.21414", "title": "A Framework for Adversarial Analysis of Decision Support Systems Prior to Deployment", "abstract": "This paper introduces a comprehensive framework designed to analyze and secure decision-support systems trained with Deep Reinforcement Learning (DRL), prior to deployment, by providing insights into learned behavior patterns and vulnerabilities discovered through simulation. The introduced framework aids in the development of precisely timed and targeted observation perturbations, enabling researchers to assess adversarial attack outcomes within a strategic decision-making context. We validate our framework, visualize agent behavior, and evaluate adversarial outcomes within the context of a custom-built strategic game, CyberStrike. Utilizing the proposed framework, we introduce a method for systematically discovering and ranking the impact of attacks on various observation indices and time-steps, and we conduct experiments to evaluate the transferability of adversarial attacks across agent architectures and DRL training algorithms. The findings underscore the critical need for robust adversarial defense mechanisms to protect decision-making policies in high-stakes environments.", "date": "2025-05-27", "categories": []}
{"id": "2505.21494", "title": "Adversarial Attacks against Closed-Source MLLMs via Feature Optimal Alignment", "abstract": "Multimodal large language models (MLLMs) remain vulnerable to transferable adversarial examples. While existing methods typically achieve targeted attacks by aligning global features-such as CLIP's [CLS] token-between adversarial and target samples, they often overlook the rich local information encoded in patch tokens. This leads to suboptimal alignment and limited transferability, particularly for closed-source models. To address this limitation, we propose a targeted transferable adversarial attack method based on feature optimal alignment, called FOA-Attack, to improve adversarial transfer capability. Specifically, at the global level, we introduce a global feature loss based on cosine similarity to align the coarse-grained features of adversarial samples with those of target samples. At the local level, given the rich local representations within Transformers, we leverage clustering techniques to extract compact local patterns to alleviate redundant local features. We then formulate local feature alignment between adversarial and target samples as an optimal transport (OT) problem and propose a local clustering optimal transport loss to refine fine-grained feature alignment. Additionally, we propose a dynamic ensemble model weighting strategy to adaptively balance the influence of multiple models during adversarial example generation, thereby further improving transferability. Extensive experiments across various models demonstrate the superiority of the proposed method, outperforming state-of-the-art methods, especially in transferring to closed-source MLLMs. The code is released at https://github.com/jiaxiaojunQAQ/FOA-Attack.", "date": "2025-05-27", "categories": []}
{"id": "2505.21499", "title": "AdInject: Real-World Black-Box Attacks on Web Agents via Advertising Delivery", "abstract": "Vision-Language Model (VLM) based Web Agents represent a significant step towards automating complex tasks by simulating human-like interaction with websites. However, their deployment in uncontrolled web environments introduces significant security vulnerabilities. Existing research on adversarial environmental injection attacks often relies on unrealistic assumptions, such as direct HTML manipulation, knowledge of user intent, or access to agent model parameters, limiting their practical applicability. In this paper, we propose AdInject, a novel and real-world black-box attack method that leverages the internet advertising delivery to inject malicious content into the Web Agent's environment. AdInject operates under a significantly more realistic threat model than prior work, assuming a black-box agent, static malicious content constraints, and no specific knowledge of user intent. AdInject includes strategies for designing malicious ad content aimed at misleading agents into clicking, and a VLM-based ad content optimization technique that infers potential user intents from the target website's context and integrates these intents into the ad content to make it appear more relevant or critical to the agent's task, thus enhancing attack effectiveness. Experimental evaluations demonstrate the effectiveness of AdInject, attack success rates exceeding 60% in most scenarios and approaching 100% in certain cases. This strongly demonstrates that prevalent advertising delivery constitutes a potent and real-world vector for environment injection attacks against Web Agents. This work highlights a critical vulnerability in Web Agent security arising from real-world environment manipulation channels, underscoring the urgent need for developing robust defense mechanisms against such threats. Our code is available at https://github.com/NicerWang/AdInject.", "date": "2025-05-27", "categories": []}
{"id": "2309.17048", "title": "On Continuity of Robust and Accurate Classifiers", "abstract": "The reliability of a learning model is key to the successful deployment of machine learning in various applications. However, it is difficult to describe the phenomenon due to the complicated nature of the problems in machine learning. It has been shown that adversarial training can improve the robustness of the hypothesis. However, this improvement usually comes at the cost of decreased performance on natural samples. Hence, it has been suggested that robustness and accuracy of a hypothesis are at odds with each other. In this paper, we put forth the alternative proposal that it is the continuity of a hypothesis that is incompatible with its robustness and accuracy in many of these scenarios. In other words, a continuous function cannot effectively learn the optimal robust hypothesis. We introduce a framework for a rigorous study of harmonic and holomorphic hypothesis in learning theory terms and provide empirical evidence that continuous hypotheses do not perform as well as discontinuous hypotheses in some common machine learning tasks. From a practical point of view, our results suggests that a robust and accurate learning rule would train different continuous hypotheses for different regions of the domain. From a theoretical perspective, our analysis explains the adversarial examples phenomenon in these situations as a conflict between the continuity of a sequence of functions and its uniform convergence to a discontinuous function. Given that many of the contemporary machine learning models are continuous functions, it is important to theoretically study the continuity of robust and accurate classifiers as it is consequential in their construction, analysis and evaluation.", "date": "2023-09-29", "categories": []}
{"id": "2404.02928", "title": "Jailbreaking Prompt Attack: A Controllable Adversarial Attack against Diffusion Models", "abstract": "Text-to-image (T2I) models can be maliciously used to generate harmful content such as sexually explicit, unfaithful, and misleading or Not-Safe-for-Work (NSFW) images. Previous attacks largely depend on the availability of the diffusion model or involve a lengthy optimization process. In this work, we investigate a more practical and universal attack that does not require the presence of a target model and demonstrate that the high-dimensional text embedding space inherently contains NSFW concepts that can be exploited to generate harmful images. We present the Jailbreaking Prompt Attack (JPA). JPA first searches for the target malicious concepts in the text embedding space using a group of antonyms generated by ChatGPT. Subsequently, a prefix prompt is optimized in the discrete vocabulary space to align malicious concepts semantically in the text embedding space. We further introduce a soft assignment with gradient masking technique that allows us to perform gradient ascent in the discrete vocabulary space.\n  We perform extensive experiments with open-sourced T2I models, e.g. stable-diffusion-v1-4 and closed-sourced online services, e.g. DALLE2, Midjourney with black-box safety checkers. Results show that (1) JPA bypasses both text and image safety checkers (2) while preserving high semantic alignment with the target prompt. (3) JPA demonstrates a much faster speed than previous methods and can be executed in a fully automated manner. These merits render it a valuable tool for robustness evaluation in future text-to-image generation research.", "date": "2024-04-02", "categories": []}
{"id": "2405.17130", "title": "Explaining the role of Intrinsic Dimensionality in Adversarial Training", "abstract": "Adversarial Training (AT) impacts different architectures in distinct ways: vision models gain robustness but face reduced generalization, encoder-based models exhibit limited robustness improvements with minimal generalization loss, and recent work in latent-space adversarial training (LAT) demonstrates that decoder-based models achieve improved robustness by applying AT across multiple layers. We provide the first explanation for these trends by leveraging the manifold conjecture: off-manifold adversarial examples (AEs) enhance robustness, while on-manifold AEs improve generalization. We show that vision and decoder-based models exhibit low intrinsic dimensionality in earlier layers (favoring off-manifold AEs), whereas encoder-based models do so in later layers (favoring on-manifold AEs). Exploiting this property, we introduce SMAAT, which improves the scalability of AT for encoder-based models by perturbing the layer with the lowest intrinsic dimensionality. This reduces the projected gradient descent (PGD) chain length required for AE generation, cutting GPU time by 25-33% while significantly boosting robustness. We validate SMAAT across multiple tasks, including text generation, sentiment classification, safety filtering, and retrieval augmented generation setups, demonstrating superior robustness with comparable generalization to standard training.", "date": "2024-05-27", "categories": []}
{"id": "2406.13920", "title": "Understanding the Robustness of Graph Neural Networks against Adversarial Attacks", "abstract": "Recent studies have shown that graph neural networks (GNNs) are vulnerable to adversarial attacks, posing significant challenges to their deployment in safety-critical scenarios. This vulnerability has spurred a growing focus on designing robust GNNs. Despite this interest, current advancements have predominantly relied on empirical trial and error, resulting in a limited understanding of the robustness of GNNs against adversarial attacks. To address this issue, we conduct the first large-scale systematic study on the adversarial robustness of GNNs by considering the patterns of input graphs, the architecture of GNNs, and their model capacity, along with discussions on sensitive neurons and adversarial transferability. This work proposes a comprehensive empirical framework for analyzing the adversarial robustness of GNNs. To support the analysis of adversarial robustness in GNNs, we introduce two evaluation metrics: the confidence-based decision surface and the accuracy-based adversarial transferability rate. Through experimental analysis, we derive 11 actionable guidelines for designing robust GNNs, enabling model developers to gain deeper insights. The code of this study is available at https://github.com/star4455/GraphRE.", "date": "2024-06-20", "categories": []}
{"id": "2411.02391", "title": "Attacking Vision-Language Computer Agents via Pop-ups", "abstract": "Autonomous agents powered by large vision and language models (VLM) have demonstrated significant potential in completing daily computer tasks, such as browsing the web to book travel and operating desktop software, which requires agents to understand these interfaces. Despite such visual inputs becoming more integrated into agentic applications, what types of risks and attacks exist around them still remain unclear. In this work, we demonstrate that VLM agents can be easily attacked by a set of carefully designed adversarial pop-ups, which human users would typically recognize and ignore. This distraction leads agents to click these pop-ups instead of performing their tasks as usual. Integrating these pop-ups into existing agent testing environments like OSWorld and VisualWebArena leads to an attack success rate (the frequency of the agent clicking the pop-ups) of 86% on average and decreases the task success rate by 47%. Basic defense techniques, such as asking the agent to ignore pop-ups or including an advertisement notice, are ineffective against the attack.", "date": "2024-11-04", "categories": []}
{"id": "2503.17168", "title": "Hi-ALPS -- An Experimental Robustness Quantification of Six LiDAR-based Object Detection Systems for Autonomous Driving", "abstract": "Light Detection and Ranging (LiDAR) is an essential sensor technology for autonomous driving as it can capture high-resolution 3D data. As 3D object detection systems (OD) can interpret such point cloud data, they play a key role in the driving decisions of autonomous vehicles. Consequently, such 3D OD must be robust against all types of perturbations and must therefore be extensively tested. One approach is the use of adversarial examples, which are small, sometimes sophisticated perturbations in the input data that change, i.e., falsify, the prediction of the OD. These perturbations are carefully designed based on the weaknesses of the OD. The robustness of the OD cannot be quantified with adversarial examples in general, because if the OD is vulnerable to a given attack, it is unclear whether this is due to the robustness of the OD or whether the attack algorithm produces particularly strong adversarial examples. The contribution of this work is Hi-ALPS -- Hierarchical Adversarial-example-based LiDAR Perturbation Level System, where higher robustness of the OD is required to withstand the perturbations as the perturbation levels increase. In doing so, the Hi-ALPS levels successively implement a heuristic followed by established adversarial example approaches. In a series of comprehensive experiments using Hi-ALPS, we quantify the robustness of six state-of-the-art 3D OD under different types of perturbations. The results of the experiments show that none of the OD is robust against all Hi-ALPS levels; an important factor for the ranking is that human observers can still correctly recognize the perturbed objects, as the respective perturbations are small. To increase the robustness of the OD, we discuss the applicability of state-of-the-art countermeasures. In addition, we derive further suggestions for countermeasures based on our experimental results.", "date": "2025-03-21", "categories": []}
{"id": "2504.18556", "title": "RDI: An adversarial robustness evaluation metric for deep neural networks based on model statistical features", "abstract": "Deep neural networks (DNNs) are highly susceptible to adversarial samples, raising concerns about their reliability in safety-critical tasks. Currently, methods of evaluating adversarial robustness are primarily categorized into attack-based and certified robustness evaluation approaches. The former not only relies on specific attack algorithms but also is highly time-consuming, while the latter due to its analytical nature, is typically difficult to implement for large and complex models. A few studies evaluate model robustness based on the model's decision boundary, but they suffer from low evaluation accuracy. To address the aforementioned issues, we propose a novel adversarial robustness evaluation metric, Robustness Difference Index (RDI), which is based on model statistical features. RDI draws inspiration from clustering evaluation by analyzing the intra-class and inter-class distances of feature vectors separated by the decision boundary to quantify model robustness. It is attack-independent and has high computational efficiency. Experiments show that, RDI demonstrates a stronger correlation with the gold-standard adversarial robustness metric of attack success rate (ASR). The average computation time of RDI is only 1/30 of the evaluation method based on the PGD attack. Our open-source code is available at: https://github.com/BUPTAIOC/RDI.", "date": "2025-04-16", "categories": []}
{"id": "2505.18864", "title": "Audio Jailbreak Attacks: Exposing Vulnerabilities in SpeechGPT in a White-Box Framework", "abstract": "Recent advances in Multimodal Large Language Models (MLLMs) have significantly enhanced the naturalness and flexibility of human computer interaction by enabling seamless understanding across text, vision, and audio modalities. Among these, voice enabled models such as SpeechGPT have demonstrated considerable improvements in usability, offering expressive, and emotionally responsive interactions that foster deeper connections in real world communication scenarios. However, the use of voice introduces new security risks, as attackers can exploit the unique characteristics of spoken language, such as timing, pronunciation variability, and speech to text translation, to craft inputs that bypass defenses in ways not seen in text-based systems. Despite substantial research on text based jailbreaks, the voice modality remains largely underexplored in terms of both attack strategies and defense mechanisms. In this work, we present an adversarial attack targeting the speech input of aligned MLLMs in a white box scenario. Specifically, we introduce a novel token level attack that leverages access to the model's speech tokenization to generate adversarial token sequences. These sequences are then synthesized into audio prompts, which effectively bypass alignment safeguards and to induce prohibited outputs. Evaluated on SpeechGPT, our approach achieves up to 89 percent attack success rate across multiple restricted tasks, significantly outperforming existing voice based jailbreak methods. Our findings shed light on the vulnerabilities of voice-enabled multimodal systems and to help guide the development of more robust next-generation MLLMs.", "date": "2025-05-24", "categories": []}
{"id": "2505.19425", "title": "Structure Disruption: Subverting Malicious Diffusion-Based Inpainting via Self-Attention Query Perturbation", "abstract": "The rapid advancement of diffusion models has enhanced their image inpainting and editing capabilities but also introduced significant societal risks. Adversaries can exploit user images from social media to generate misleading or harmful content. While adversarial perturbations can disrupt inpainting, global perturbation-based methods fail in mask-guided editing tasks due to spatial constraints. To address these challenges, we propose Structure Disruption Attack (SDA), a powerful protection framework for safeguarding sensitive image regions against inpainting-based editing. Building upon the contour-focused nature of self-attention mechanisms of diffusion models, SDA optimizes perturbations by disrupting queries in self-attention during the initial denoising step to destroy the contour generation process. This targeted interference directly disrupts the structural generation capability of diffusion models, effectively preventing them from producing coherent images. We validate our motivation through visualization techniques and extensive experiments on public datasets, demonstrating that SDA achieves state-of-the-art (SOTA) protection performance while maintaining strong robustness.", "date": "2025-05-26", "categories": []}
{"id": "2505.19951", "title": "Novel Loss-Enhanced Universal Adversarial Patches for Sustainable Speaker Privacy", "abstract": "Deep learning voice models are commonly used nowadays, but the safety processing of personal data, such as human identity and speech content, remains suspicious. To prevent malicious user identification, speaker anonymization methods were proposed. Current methods, particularly based on universal adversarial patch (UAP) applications, have drawbacks such as significant degradation of audio quality, decreased speech recognition quality, low transferability across different voice biometrics models, and performance dependence on the input audio length. To mitigate these drawbacks, in this work, we introduce and leverage the novel Exponential Total Variance (TV) loss function and provide experimental evidence that it positively affects UAP strength and imperceptibility. Moreover, we present a novel scalable UAP insertion procedure and demonstrate its uniformly high performance for various audio lengths.", "date": "2025-05-26", "categories": []}
{"id": "2505.20259", "title": "Lifelong Safety Alignment for Language Models", "abstract": "LLMs have made impressive progress, but their growing capabilities also expose them to highly flexible jailbreaking attacks designed to bypass safety alignment. While many existing defenses focus on known types of attacks, it is more critical to prepare LLMs for unseen attacks that may arise during deployment. To address this, we propose a lifelong safety alignment framework that enables LLMs to continuously adapt to new and evolving jailbreaking strategies. Our framework introduces a competitive setup between two components: a Meta-Attacker, trained to actively discover novel jailbreaking strategies, and a Defender, trained to resist them. To effectively warm up the Meta-Attacker, we first leverage the GPT-4o API to extract key insights from a large collection of jailbreak-related research papers. Through iterative training, the first iteration Meta-Attacker achieves a 73% attack success rate (ASR) on RR and a 57% transfer ASR on LAT using only single-turn attacks. Meanwhile, the Defender progressively improves its robustness and ultimately reduces the Meta-Attacker's success rate to just 7%, enabling safer and more reliable deployment of LLMs in open-ended environments. The code is available at https://github.com/sail-sg/LifelongSafetyAlignment.", "date": "2025-05-26", "categories": []}
{"id": "2402.09546", "title": "How Secure Are Large Language Models (LLMs) for Navigation in Urban Environments?", "abstract": "In the field of robotics and automation, navigation systems based on Large Language Models (LLMs) have recently demonstrated impressive performance. However, the security aspects of these systems have received relatively less attention. This paper pioneers the exploration of vulnerabilities in LLM-based navigation models in urban outdoor environments, a critical area given the widespread application of this technology in autonomous driving, logistics, and emergency services. Specifically, we introduce a novel Navigational Prompt Attack that manipulates LLM-based navigation models by perturbing the original navigational prompt, leading to incorrect actions. Based on the method of perturbation, our attacks are divided into two types: Navigational Prompt Insert (NPI) Attack and Navigational Prompt Swap (NPS) Attack. We conducted comprehensive experiments on an LLM-based navigation model that employs various LLMs for reasoning. Our results, derived from the Touchdown and Map2Seq street-view datasets under both few-shot learning and fine-tuning configurations, demonstrate notable performance declines across seven metrics in the face of both white-box and black-box attacks. Moreover, our attacks can be easily extended to other LLM-based navigation models with similarly effective results. These findings highlight the generalizability and transferability of the proposed attack, emphasizing the need for enhanced security in LLM-based navigation systems. As an initial countermeasure, we propose the Navigational Prompt Engineering (NPE) Defense strategy, which concentrates on navigation-relevant keywords to reduce the impact of adversarial attacks. While initial findings indicate that this strategy enhances navigational safety, there remains a critical need for the wider research community to develop stronger defense methods to effectively tackle the real-world challenges faced by these systems.", "date": "2024-02-14", "categories": []}
{"id": "2407.18170", "title": "RIDA: A Robust Attack Framework on Incomplete Graphs", "abstract": "Graph Neural Networks (GNNs) are vital in data science but are increasingly susceptible to adversarial attacks. To help researchers develop more robust GNN models, it's essential to focus on designing strong attack models as foundational benchmarks and guiding references. Among adversarial attacks, gray-box poisoning attacks are noteworthy due to their effectiveness and fewer constraints. These attacks exploit GNNs' need for retraining on updated data, thereby impacting their performance by perturbing these datasets. However, current research overlooks the real-world scenario of incomplete graphs. To address this gap, we introduce the Robust Incomplete Deep Attack Framework (RIDA). It is the first algorithm for robust gray-box poisoning attacks on incomplete graphs. The approach innovatively aggregates distant vertex information and ensures powerful data utilization. Extensive tests against 9 SOTA baselines on 3 real-world datasets demonstrate that RIDA's superiority in handling incompleteness and high attack performance on the incomplete graph.", "date": "2024-07-25", "categories": []}
{"id": "2504.05804", "title": "StealthRank: LLM Ranking Manipulation via Stealthy Prompt Optimization", "abstract": "The integration of large language models (LLMs) into information retrieval systems introduces new attack surfaces, particularly for adversarial ranking manipulations. We present $\\textbf{StealthRank}$, a novel adversarial attack method that manipulates LLM-driven ranking systems while maintaining textual fluency and stealth. Unlike existing methods that often introduce detectable anomalies, StealthRank employs an energy-based optimization framework combined with Langevin dynamics to generate StealthRank Prompts (SRPs)-adversarial text sequences embedded within item or document descriptions that subtly yet effectively influence LLM ranking mechanisms. We evaluate StealthRank across multiple LLMs, demonstrating its ability to covertly boost the ranking of target items while avoiding explicit manipulation traces. Our results show that StealthRank consistently outperforms state-of-the-art adversarial ranking baselines in both effectiveness and stealth, highlighting critical vulnerabilities in LLM-driven ranking systems. Our code is publicly available at $\\href{https://github.com/Tangyiming205069/controllable-seo}{here}$.", "date": "2025-04-08", "categories": []}
{"id": "2505.15880", "title": "Challenger: Affordable Adversarial Driving Video Generation", "abstract": "Generating photorealistic driving videos has seen significant progress recently, but current methods largely focus on ordinary, non-adversarial scenarios. Meanwhile, efforts to generate adversarial driving scenarios often operate on abstract trajectory or BEV representations, falling short of delivering realistic sensor data that can truly stress-test autonomous driving (AD) systems. In this work, we introduce Challenger, a framework that produces physically plausible yet photorealistic adversarial driving videos. Generating such videos poses a fundamental challenge: it requires jointly optimizing over the space of traffic interactions and high-fidelity sensor observations. Challenger makes this affordable through two techniques: (1) a physics-aware multi-round trajectory refinement process that narrows down candidate adversarial maneuvers, and (2) a tailored trajectory scoring function that encourages realistic yet adversarial behavior while maintaining compatibility with downstream video synthesis. As tested on the nuScenes dataset, Challenger generates a diverse range of aggressive driving scenarios-including cut-ins, sudden lane changes, tailgating, and blind spot intrusions-and renders them into multiview photorealistic videos. Extensive evaluations show that these scenarios significantly increase the collision rate of state-of-the-art end-to-end AD models (UniAD, VAD, SparseDrive, and DiffusionDrive), and importantly, adversarial behaviors discovered for one model often transfer to others.", "date": "2025-05-21", "categories": []}
{"id": "2505.17147", "title": "MTSA: Multi-turn Safety Alignment for LLMs through Multi-round Red-teaming", "abstract": "The proliferation of jailbreak attacks against large language models (LLMs) highlights the need for robust security measures. However, in multi-round dialogues, malicious intentions may be hidden in interactions, leading LLMs to be more prone to produce harmful responses. In this paper, we propose the \\textbf{M}ulti-\\textbf{T}urn \\textbf{S}afety \\textbf{A}lignment (\\ourapproach) framework, to address the challenge of securing LLMs in multi-round interactions. It consists of two stages: In the thought-guided attack learning stage, the red-team model learns about thought-guided multi-round jailbreak attacks to generate adversarial prompts. In the adversarial iterative optimization stage, the red-team model and the target model continuously improve their respective capabilities in interaction. Furthermore, we introduce a multi-turn reinforcement learning algorithm based on future rewards to enhance the robustness of safety alignment. Experimental results show that the red-team model exhibits state-of-the-art attack capabilities, while the target model significantly improves its performance on safety benchmarks.", "date": "2025-05-22", "categories": []}
{"id": "2505.17513", "title": "What You Read Isn't What You Hear: Linguistic Sensitivity in Deepfake Speech Detection", "abstract": "Recent advances in text-to-speech technologies have enabled realistic voice generation, fueling audio-based deepfake attacks such as fraud and impersonation. While audio anti-spoofing systems are critical for detecting such threats, prior work has predominantly focused on acoustic-level perturbations, leaving the impact of linguistic variation largely unexplored. In this paper, we investigate the linguistic sensitivity of both open-source and commercial anti-spoofing detectors by introducing transcript-level adversarial attacks. Our extensive evaluation reveals that even minor linguistic perturbations can significantly degrade detection accuracy: attack success rates surpass 60% on several open-source detector-voice pairs, and notably one commercial detection accuracy drops from 100% on synthetic audio to just 32%. Through a comprehensive feature attribution analysis, we identify that both linguistic complexity and model-level audio embedding similarity contribute strongly to detector vulnerability. We further demonstrate the real-world risk via a case study replicating the Brad Pitt audio deepfake scam, using transcript adversarial attacks to completely bypass commercial detectors. These results highlight the need to move beyond purely acoustic defenses and account for linguistic variation in the design of robust anti-spoofing systems. All source code will be publicly available.", "date": "2025-05-23", "categories": []}
{"id": "2505.17650", "title": "Does Chain-of-Thought Reasoning Really Reduce Harmfulness from Jailbreaking?", "abstract": "Jailbreak attacks have been observed to largely fail against recent reasoning models enhanced by Chain-of-Thought (CoT) reasoning. However, the underlying mechanism remains underexplored, and relying solely on reasoning capacity may raise security concerns. In this paper, we try to answer the question: Does CoT reasoning really reduce harmfulness from jailbreaking? Through rigorous theoretical analysis, we demonstrate that CoT reasoning has dual effects on jailbreaking harmfulness. Based on the theoretical insights, we propose a novel jailbreak method, FicDetail, whose practical performance validates our theoretical findings.", "date": "2025-05-23", "categories": []}
{"id": "2505.17807", "title": "Temporal Consistency Constrained Transferable Adversarial Attacks with Background Mixup for Action Recognition", "abstract": "Action recognition models using deep learning are vulnerable to adversarial examples, which are transferable across other models trained on the same data modality. Existing transferable attack methods face two major challenges: 1) they heavily rely on the assumption that the decision boundaries of the surrogate (a.k.a., source) model and the target model are similar, which limits the adversarial transferability; and 2) their decision boundary difference makes the attack direction uncertain, which may result in the gradient oscillation, weakening the adversarial attack. This motivates us to propose a Background Mixup-induced Temporal Consistency (BMTC) attack method for action recognition. From the input transformation perspective, we design a model-agnostic background adversarial mixup module to reduce the surrogate-target model dependency. In particular, we randomly sample one video from each category and make its background frame, while selecting the background frame with the top attack ability for mixup with the clean frame by reinforcement learning. Moreover, to ensure an explicit attack direction, we leverage the background category as guidance for updating the gradient of adversarial example, and design a temporal gradient consistency loss, which strengthens the stability of the attack direction on subsequent frames. Empirical studies on two video datasets, i.e., UCF101 and Kinetics-400, and one image dataset, i.e., ImageNet, demonstrate that our method significantly boosts the transferability of adversarial examples across several action/image recognition models. Our code is available at https://github.com/mlvccn/BMTC_TransferAttackVid.", "date": "2025-05-23", "categories": []}
{"id": "2505.18097", "title": "Towards more transferable adversarial attack in black-box manner", "abstract": "Adversarial attacks have become a well-explored domain, frequently serving as evaluation baselines for model robustness. Among these, black-box attacks based on transferability have received significant attention due to their practical applicability in real-world scenarios. Traditional black-box methods have generally focused on improving the optimization framework (e.g., utilizing momentum in MI-FGSM) to enhance transferability, rather than examining the dependency on surrogate white-box model architectures. Recent state-of-the-art approach DiffPGD has demonstrated enhanced transferability by employing diffusion-based adversarial purification models for adaptive attacks. The inductive bias of diffusion-based adversarial purification aligns naturally with the adversarial attack process, where both involving noise addition, reducing dependency on surrogate white-box model selection. However, the denoising process of diffusion models incurs substantial computational costs through chain rule derivation, manifested in excessive VRAM consumption and extended runtime. This progression prompts us to question whether introducing diffusion models is necessary. We hypothesize that a model sharing similar inductive bias to diffusion-based adversarial purification, combined with an appropriate loss function, could achieve comparable or superior transferability while dramatically reducing computational overhead. In this paper, we propose a novel loss function coupled with a unique surrogate model to validate our hypothesis. Our approach leverages the score of the time-dependent classifier from classifier-guided diffusion models, effectively incorporating natural data distribution knowledge into the adversarial optimization process. Experimental results demonstrate significantly improved transferability across diverse model architectures while maintaining robustness against diffusion-based defenses.", "date": "2025-05-23", "categories": []}
{"id": "2306.01271", "title": "On the Clean Generalization and Robust Overfitting in Adversarial Training from Two Theoretical Views: Representation Complexity and Training Dynamics", "abstract": "Similar to surprising performance in the standard deep learning, deep nets trained by adversarial training also generalize well for unseen clean data (natural data). However, despite adversarial training can achieve low robust training error, there exists a significant robust generalization gap. We call this phenomenon the Clean Generalization and Robust Overfitting (CGRO). In this work, we study the CGRO phenomenon in adversarial training from two views: representation complexity and training dynamics. Specifically, we consider a binary classification setting with $N$ separated training data points. First, we prove that, based on the assumption that we assume there is $\\operatorname{poly}(D)$-size clean classifier (where $D$ is the data dimension), ReLU net with only $O(N D)$ extra parameters is able to leverages robust memorization to achieve the CGRO, while robust classifier still requires exponential representation complexity in worst case. Next, we focus on a structured-data case to analyze training dynamics, where we train a two-layer convolutional network with $O(N D)$ width against adversarial perturbation. We then show that a three-stage phase transition occurs during learning process and the network provably converges to robust memorization regime, which thereby results in the CGRO. Besides, we also empirically verify our theoretical analysis by experiments in real-image recognition datasets.", "date": "2023-06-02", "categories": []}
{"id": "2412.04163", "title": "On the Lack of Robustness of Binary Function Similarity Systems", "abstract": "Binary function similarity, which often relies on learning-based algorithms to identify what functions in a pool are most similar to a given query function, is a sought-after topic in different communities, including machine learning, software engineering, and security. Its importance stems from the impact it has in facilitating several crucial tasks, from reverse engineering and malware analysis to automated vulnerability detection. Whereas recent work cast light around performance on this long-studied problem, the research landscape remains largely lackluster in understanding the resiliency of the state-of-the-art machine learning models against adversarial attacks. As security requires to reason about adversaries, in this work we assess the robustness of such models through a simple yet effective black-box greedy attack, which modifies the topology and the content of the control flow of the attacked functions. We demonstrate that this attack is successful in compromising all the models, achieving average attack success rates of 57.06% and 95.81% depending on the problem settings (targeted and untargeted attacks). Our findings are insightful: top performance on clean data does not necessarily relate to top robustness properties, which explicitly highlights performance-robustness trade-offs one should consider when deploying such models, calling for further research.", "date": "2024-12-05", "categories": []}
{"id": "2505.16154", "title": "BadDepth: Backdoor Attacks Against Monocular Depth Estimation in the Physical World", "abstract": "In recent years, deep learning-based Monocular Depth Estimation (MDE) models have been widely applied in fields such as autonomous driving and robotics. However, their vulnerability to backdoor attacks remains unexplored. To fill the gap in this area, we conduct a comprehensive investigation of backdoor attacks against MDE models. Typically, existing backdoor attack methods can not be applied to MDE models. This is because the label used in MDE is in the form of a depth map. To address this, we propose BadDepth, the first backdoor attack targeting MDE models. BadDepth overcomes this limitation by selectively manipulating the target object's depth using an image segmentation model and restoring the surrounding areas via depth completion, thereby generating poisoned datasets for object-level backdoor attacks. To improve robustness in physical world scenarios, we further introduce digital-to-physical augmentation to adapt to the domain gap between the physical world and the digital domain. Extensive experiments on multiple models validate the effectiveness of BadDepth in both the digital domain and the physical world, without being affected by environmental factors.", "date": "2025-05-22", "categories": []}
{"id": "2505.16166", "title": "TRAIL: Transferable Robust Adversarial Images via Latent diffusion", "abstract": "Adversarial attacks exploiting unrestricted natural perturbations present severe security risks to deep learning systems, yet their transferability across models remains limited due to distribution mismatches between generated adversarial features and real-world data. While recent works utilize pre-trained diffusion models as adversarial priors, they still encounter challenges due to the distribution shift between the distribution of ideal adversarial samples and the natural image distribution learned by the diffusion model. To address the challenge, we propose Transferable Robust Adversarial Images via Latent Diffusion (TRAIL), a test-time adaptation framework that enables the model to generate images from a distribution of images with adversarial features and closely resembles the target images. To mitigate the distribution shift, during attacks, TRAIL updates the diffusion U-Net's weights by combining adversarial objectives (to mislead victim models) and perceptual constraints (to preserve image realism). The adapted model then generates adversarial samples through iterative noise injection and denoising guided by these objectives. Experiments demonstrate that TRAIL significantly outperforms state-of-the-art methods in cross-model attack transferability, validating that distribution-aligned adversarial feature synthesis is critical for practical black-box attacks.", "date": "2025-05-22", "categories": []}
{"id": "2505.16263", "title": "All You Need is \"Leet\": Evading Hate-speech Detection AI", "abstract": "Social media and online forums are increasingly becoming popular. Unfortunately, these platforms are being used for spreading hate speech. In this paper, we design black-box techniques to protect users from hate-speech on online platforms by generating perturbations that can fool state of the art deep learning based hate speech detection models thereby decreasing their efficiency. We also ensure a minimal change in the original meaning of hate-speech. Our best perturbation attack is successfully able to evade hate-speech detection for 86.8 % of hateful text.", "date": "2025-05-22", "categories": []}
{"id": "2505.16318", "title": "SuperPure: Efficient Purification of Localized and Distributed Adversarial Patches via Super-Resolution GAN Models", "abstract": "As vision-based machine learning models are increasingly integrated into autonomous and cyber-physical systems, concerns about (physical) adversarial patch attacks are growing. While state-of-the-art defenses can achieve certified robustness with minimal impact on utility against highly-concentrated localized patch attacks, they fall short in two important areas: (i) State-of-the-art methods are vulnerable to low-noise distributed patches where perturbations are subtly dispersed to evade detection or masking, as shown recently by the DorPatch attack; (ii) Achieving high robustness with state-of-the-art methods is extremely time and resource-consuming, rendering them impractical for latency-sensitive applications in many cyber-physical systems.\n  To address both robustness and latency issues, this paper proposes a new defense strategy for adversarial patch attacks called SuperPure. The key novelty is developing a pixel-wise masking scheme that is robust against both distributed and localized patches. The masking involves leveraging a GAN-based super-resolution scheme to gradually purify the image from adversarial patches. Our extensive evaluations using ImageNet and two standard classifiers, ResNet and EfficientNet, show that SuperPure advances the state-of-the-art in three major directions: (i) it improves the robustness against conventional localized patches by more than 20%, on average, while also improving top-1 clean accuracy by almost 10%; (ii) It achieves 58% robustness against distributed patch attacks (as opposed to 0% in state-of-the-art method, PatchCleanser); (iii) It decreases the defense end-to-end latency by over 98% compared to PatchCleanser. Our further analysis shows that SuperPure is robust against white-box attacks and different patch sizes. Our code is open-source.", "date": "2025-05-22", "categories": []}
{"id": "2406.19413", "title": "Saliency Attention and Semantic Similarity-Driven Adversarial Perturbation", "abstract": "In this paper, we introduce an enhanced textual adversarial attack method, known as Saliency Attention and Semantic Similarity driven adversarial Perturbation (SASSP). The proposed scheme is designed to improve the effectiveness of contextual perturbations by integrating saliency, attention, and semantic similarity. Traditional adversarial attack methods often struggle to maintain semantic consistency and coherence while effectively deceiving target models. Our proposed approach addresses these challenges by incorporating a three-pronged strategy for word selection and perturbation. First, we utilize a saliency-based word selection to prioritize words for modification based on their importance to the model's prediction. Second, attention mechanisms are employed to focus perturbations on contextually significant words, enhancing the attack's efficacy. Finally, an advanced semantic similarity-checking method is employed that includes embedding-based similarity and paraphrase detection. By leveraging models like Sentence-BERT for embedding similarity and fine-tuned paraphrase detection models from the Sentence Transformers library, the scheme ensures that the perturbed text remains contextually appropriate and semantically consistent with the original. Empirical evaluations demonstrate that SASSP generates adversarial examples that not only maintain high semantic fidelity but also effectively deceive state-of-the-art natural language processing models. Moreover, in comparison to the original scheme of contextual perturbation CLARE, SASSP has yielded a higher attack success rate and lower word perturbation rate.", "date": "2024-06-18", "categories": []}
{"id": "2505.04046", "title": "Reliable Disentanglement Multi-view Learning Against View Adversarial Attacks", "abstract": "Trustworthy multi-view learning has attracted extensive attention because evidence learning can provide reliable uncertainty estimation to enhance the credibility of multi-view predictions. Existing trusted multi-view learning methods implicitly assume that multi-view data is secure. However, in safety-sensitive applications such as autonomous driving and security monitoring, multi-view data often faces threats from adversarial perturbations, thereby deceiving or disrupting multi-view models. This inevitably leads to the adversarial unreliability problem (AUP) in trusted multi-view learning. To overcome this tricky problem, we propose a novel multi-view learning framework, namely Reliable Disentanglement Multi-view Learning (RDML). Specifically, we first propose evidential disentanglement learning to decompose each view into clean and adversarial parts under the guidance of corresponding evidences, which is extracted by a pretrained evidence extractor. Then, we employ the feature recalibration module to mitigate the negative impact of adversarial perturbations and extract potential informative features from them. Finally, to further ignore the irreparable adversarial interferences, a view-level evidential attention mechanism is designed. Extensive experiments on multi-view classification tasks with adversarial attacks show that RDML outperforms the state-of-the-art methods by a relatively large margin. Our code is available at https://github.com/Willy1005/2025-IJCAI-RDML.", "date": "2025-05-07", "categories": []}
{"id": "2505.15308", "title": "BadSR: Stealthy Label Backdoor Attacks on Image Super-Resolution", "abstract": "With the widespread application of super-resolution (SR) in various fields, researchers have begun to investigate its security. Previous studies have demonstrated that SR models can also be subjected to backdoor attacks through data poisoning, affecting downstream tasks. A backdoor SR model generates an attacker-predefined target image when given a triggered image while producing a normal high-resolution (HR) output for clean images. However, prior backdoor attacks on SR models have primarily focused on the stealthiness of poisoned low-resolution (LR) images while ignoring the stealthiness of poisoned HR images, making it easy for users to detect anomalous data. To address this problem, we propose BadSR, which improves the stealthiness of poisoned HR images. The key idea of BadSR is to approximate the clean HR image and the pre-defined target image in the feature space while ensuring that modifications to the clean HR image remain within a constrained range. The poisoned HR images generated by BadSR can be integrated with existing triggers. To further improve the effectiveness of BadSR, we design an adversarially optimized trigger and a backdoor gradient-driven poisoned sample selection method based on a genetic algorithm. The experimental results show that BadSR achieves a high attack success rate in various models and data sets, significantly affecting downstream tasks.", "date": "2025-05-21", "categories": []}
{"id": "2505.15336", "title": "My Face Is Mine, Not Yours: Facial Protection Against Diffusion Model Face Swapping", "abstract": "The proliferation of diffusion-based deepfake technologies poses significant risks for unauthorized and unethical facial image manipulation. While traditional countermeasures have primarily focused on passive detection methods, this paper introduces a novel proactive defense strategy through adversarial attacks that preemptively protect facial images from being exploited by diffusion-based deepfake systems. Existing adversarial protection methods predominantly target conventional generative architectures (GANs, AEs, VAEs) and fail to address the unique challenges presented by diffusion models, which have become the predominant framework for high-quality facial deepfakes. Current diffusion-specific adversarial approaches are limited by their reliance on specific model architectures and weights, rendering them ineffective against the diverse landscape of diffusion-based deepfake implementations. Additionally, they typically employ global perturbation strategies that inadequately address the region-specific nature of facial manipulation in deepfakes.", "date": "2025-05-21", "categories": []}
{"id": "2505.15406", "title": "Audio Jailbreak: An Open Comprehensive Benchmark for Jailbreaking Large Audio-Language Models", "abstract": "The rise of Large Audio Language Models (LAMs) brings both potential and risks, as their audio outputs may contain harmful or unethical content. However, current research lacks a systematic, quantitative evaluation of LAM safety especially against jailbreak attacks, which are challenging due to the temporal and semantic nature of speech. To bridge this gap, we introduce AJailBench, the first benchmark specifically designed to evaluate jailbreak vulnerabilities in LAMs. We begin by constructing AJailBench-Base, a dataset of 1,495 adversarial audio prompts spanning 10 policy-violating categories, converted from textual jailbreak attacks using realistic text to speech synthesis. Using this dataset, we evaluate several state-of-the-art LAMs and reveal that none exhibit consistent robustness across attacks. To further strengthen jailbreak testing and simulate more realistic attack conditions, we propose a method to generate dynamic adversarial variants. Our Audio Perturbation Toolkit (APT) applies targeted distortions across time, frequency, and amplitude domains. To preserve the original jailbreak intent, we enforce a semantic consistency constraint and employ Bayesian optimization to efficiently search for perturbations that are both subtle and highly effective. This results in AJailBench-APT, an extended dataset of optimized adversarial audio samples. Our findings demonstrate that even small, semantically preserved perturbations can significantly reduce the safety performance of leading LAMs, underscoring the need for more robust and semantically aware defense mechanisms.", "date": "2025-05-21", "categories": []}
{"id": "2505.15594", "title": "Beyond Classification: Evaluating Diffusion Denoised Smoothing for Security-Utility Trade off", "abstract": "While foundation models demonstrate impressive performance across various tasks, they remain vulnerable to adversarial inputs. Current research explores various approaches to enhance model robustness, with Diffusion Denoised Smoothing emerging as a particularly promising technique. This method employs a pretrained diffusion model to preprocess inputs before model inference. Yet, its effectiveness remains largely unexplored beyond classification. We aim to address this gap by analyzing three datasets with four distinct downstream tasks under three different adversarial attack algorithms. Our findings reveal that while foundation models maintain resilience against conventional transformations, applying high-noise diffusion denoising to clean images without any distortions significantly degrades performance by as high as 57%. Low-noise diffusion settings preserve performance but fail to provide adequate protection across all attack types. Moreover, we introduce a novel attack strategy specifically targeting the diffusion process itself, capable of circumventing defenses in the low-noise regime. Our results suggest that the trade-off between adversarial robustness and performance remains a challenge to be addressed.", "date": "2025-05-21", "categories": []}
{"id": "2505.13708", "title": "Robust learning of halfspaces under log-concave marginals", "abstract": "We say that a classifier is \\emph{adversarially robust} to perturbations of norm $r$ if, with high probability over a point $x$ drawn from the input distribution, there is no point within distance $\\le r$ from $x$ that is classified differently. The \\emph{boundary volume} is the probability that a point falls within distance $r$ of a point with a different label. This work studies the task of computationally efficient learning of hypotheses with small boundary volume, where the input is distributed as a subgaussian isotropic log-concave distribution over $\\mathbb{R}^d$.\n  Linear threshold functions are adversarially robust; they have boundary volume proportional to $r$. Such concept classes are efficiently learnable by polynomial regression, which produces a polynomial threshold function (PTF), but PTFs in general may have boundary volume $\\Omega(1)$, even for $r \\ll 1$.\n  We give an algorithm that agnostically learns linear threshold functions and returns a classifier with boundary volume $O(r+\\varepsilon)$ at radius of perturbation $r$. The time and sample complexity of $d^{\\tilde{O}(1/\\varepsilon^2)}$ matches the complexity of polynomial regression.\n  Our algorithm augments the classic approach of polynomial regression with three additional steps: a) performing the $\\ell_1$-error regression under noise sensitivity constraints, b) a structured partitioning and rounding step that returns a Boolean classifier with error $\\textsf{opt} + O(\\varepsilon)$ and noise sensitivity $O(r+\\varepsilon)$ simultaneously, and c) a local corrector that ``smooths'' a function with low noise sensitivity into a function that is adversarially robust.", "date": "2025-05-19", "categories": []}
{"id": "2505.14021", "title": "Adversarial Training from Mean Field Perspective", "abstract": "Although adversarial training is known to be effective against adversarial examples, training dynamics are not well understood. In this study, we present the first theoretical analysis of adversarial training in random deep neural networks without any assumptions on data distributions. We introduce a new theoretical framework based on mean field theory, which addresses the limitations of existing mean field-based approaches. Based on this framework, we derive (empirically tight) upper bounds of $\\ell_q$ norm-based adversarial loss with $\\ell_p$ norm-based adversarial examples for various values of $p$ and $q$. Moreover, we prove that networks without shortcuts are generally not adversarially trainable and that adversarial training reduces network capacity. We also show that network width alleviates these issues. Furthermore, we present the various impacts of the input and output dimensions on the upper bounds and time evolution of the weight variance.", "date": "2025-05-20", "categories": []}
{"id": "2505.14286", "title": "Universal Acoustic Adversarial Attacks for Flexible Control of Speech-LLMs", "abstract": "The combination of pre-trained speech encoders with large language models has enabled the development of speech LLMs that can handle a wide range of spoken language processing tasks. While these models are powerful and flexible, this very flexibility may make them more vulnerable to adversarial attacks. To examine the extent of this problem, in this work we investigate universal acoustic adversarial attacks on speech LLMs. Here a fixed, universal, adversarial audio segment is prepended to the original input audio. We initially investigate attacks that cause the model to either produce no output or to perform a modified task overriding the original prompt. We then extend the nature of the attack to be selective so that it activates only when specific input attributes, such as a speaker gender or spoken language, are present. Inputs without the targeted attribute should be unaffected, allowing fine-grained control over the model outputs. Our findings reveal critical vulnerabilities in Qwen2-Audio and Granite-Speech and suggest that similar speech LLMs may be susceptible to universal adversarial attacks. This highlights the need for more robust training strategies and improved resistance to adversarial attacks.", "date": "2025-05-20", "categories": []}
{"id": "2505.14316", "title": "Exploring Jailbreak Attacks on LLMs through Intent Concealment and Diversion", "abstract": "Although large language models (LLMs) have achieved remarkable advancements, their security remains a pressing concern. One major threat is jailbreak attacks, where adversarial prompts bypass model safeguards to generate harmful or objectionable content. Researchers study jailbreak attacks to understand security and robustness of LLMs. However, existing jailbreak attack methods face two main challenges: (1) an excessive number of iterative queries, and (2) poor generalization across models. In addition, recent jailbreak evaluation datasets focus primarily on question-answering scenarios, lacking attention to text generation tasks that require accurate regeneration of toxic content. To tackle these challenges, we propose two contributions: (1) ICE, a novel black-box jailbreak method that employs Intent Concealment and divErsion to effectively circumvent security constraints. ICE achieves high attack success rates (ASR) with a single query, significantly improving efficiency and transferability across different models. (2) BiSceneEval, a comprehensive dataset designed for assessing LLM robustness in question-answering and text-generation tasks. Experimental results demonstrate that ICE outperforms existing jailbreak techniques, revealing critical vulnerabilities in current defense mechanisms. Our findings underscore the necessity of a hybrid security strategy that integrates predefined security mechanisms with real-time semantic decomposition to enhance the security of LLMs.", "date": "2025-05-20", "categories": []}
{"id": "2505.14463", "title": "Adverseness vs. Equilibrium: Exploring Graph Adversarial Resilience through Dynamic Equilibrium", "abstract": "Adversarial attacks to graph analytics are gaining increased attention. To date, two lines of countermeasures have been proposed to resist various graph adversarial attacks from the perspectives of either graph per se or graph neural networks. Nevertheless, a fundamental question lies in whether there exists an intrinsic adversarial resilience state within a graph regime and how to find out such a critical state if exists. This paper contributes to tackle the above research questions from three unique perspectives: i) we regard the process of adversarial learning on graph as a complex multi-object dynamic system, and model the behavior of adversarial attack; ii) we propose a generalized theoretical framework to show the existence of critical adversarial resilience state; and iii) we develop a condensed one-dimensional function to capture the dynamic variation of graph regime under perturbations, and pinpoint the critical state through solving the equilibrium point of dynamic system. Multi-facet experiments are conducted to show our proposed approach can significantly outperform the state-of-the-art defense methods under five commonly-used real-world datasets and three representative attacks.", "date": "2025-05-20", "categories": []}
{"id": "2206.05751", "title": "Consistent Attack: Universal Adversarial Perturbation on Embodied Vision Navigation", "abstract": "Embodied agents in vision navigation coupled with deep neural networks have attracted increasing attention. However, deep neural networks have been shown vulnerable to malicious adversarial noises, which may potentially cause catastrophic failures in Embodied Vision Navigation. Among different adversarial noises, universal adversarial perturbations (UAP), i.e., a constant image-agnostic perturbation applied on every input frame of the agent, play a critical role in Embodied Vision Navigation since they are computation-efficient and application-practical during the attack. However, existing UAP methods ignore the system dynamics of Embodied Vision Navigation and might be sub-optimal. In order to extend UAP to the sequential decision setting, we formulate the disturbed environment under the universal noise $\\delta$, as a $\\delta$-disturbed Markov Decision Process ($\\delta$-MDP). Based on the formulation, we analyze the properties of $\\delta$-MDP and propose two novel Consistent Attack methods, named Reward UAP and Trajectory UAP, for attacking Embodied agents, which consider the dynamic of the MDP and calculate universal noises by estimating the disturbed distribution and the disturbed Q function. For various victim models, our Consistent Attack can cause a significant drop in their performance in the PointGoal task in Habitat with different datasets and different scenes. Extensive experimental results indicate that there exist serious potential risks for applying Embodied Vision Navigation methods to the real world.", "date": "2022-06-12", "categories": []}
{"id": "2405.16940", "title": "Adversarial Attacks on Both Face Recognition and Face Anti-spoofing Models", "abstract": "Adversarial attacks on Face Recognition (FR) systems have demonstrated significant effectiveness against standalone FR models. However, their practicality diminishes in complete FR systems that incorporate Face Anti-Spoofing (FAS) models, as these models can detect and mitigate a substantial number of adversarial examples. To address this critical yet under-explored challenge, we introduce a novel attack setting that targets both FR and FAS models simultaneously, thereby enhancing the practicability of adversarial attacks on integrated FR systems. Specifically, we propose a new attack method, termed Reference-free Multi-level Alignment (RMA), designed to improve the capacity of black-box attacks on both FR and FAS models. The RMA framework is built upon three key components. Firstly, we propose an Adaptive Gradient Maintenance module to address the imbalances in gradient contributions between FR and FAS models. Secondly, we develop a Reference-free Intermediate Biasing module to improve the transferability of adversarial examples against FAS models. In addition, we introduce a Multi-level Feature Alignment module to reduce feature discrepancies at various levels of representation. Extensive experiments showcase the superiority of our proposed attack method to state-of-the-art adversarial attacks.", "date": "2024-05-27", "categories": []}
{"id": "2407.00248", "title": "DiffuseDef: Improved Robustness to Adversarial Attacks via Iterative Denoising", "abstract": "Pretrained language models have significantly advanced performance across various natural language processing tasks. However, adversarial attacks continue to pose a critical challenge to systems built using these models, as they can be exploited with carefully crafted adversarial texts. Inspired by the ability of diffusion models to predict and reduce noise in computer vision, we propose a novel and flexible adversarial defense method for language classification tasks, DiffuseDef, which incorporates a diffusion layer as a denoiser between the encoder and the classifier. The diffusion layer is trained on top of the existing classifier, ensuring seamless integration with any model in a plug-and-play manner. During inference, the adversarial hidden state is first combined with sampled noise, then denoised iteratively and finally ensembled to produce a robust text representation. By integrating adversarial training, denoising, and ensembling techniques, we show that DiffuseDef improves over existing adversarial defense methods and achieves state-of-the-art performance against common black-box and white-box adversarial attacks.", "date": "2024-06-28", "categories": []}
{"id": "2409.05657", "title": "Adversarial Attacks on Data Attribution", "abstract": "Data attribution aims to quantify the contribution of individual training data points to the outputs of an AI model, which has been used to measure the value of training data and compensate data providers. Given the impact on financial decisions and compensation mechanisms, a critical question arises concerning the adversarial robustness of data attribution methods. However, there has been little to no systematic research addressing this issue. In this work, we aim to bridge this gap by detailing a threat model with clear assumptions about the adversary's goal and capabilities and proposing principled adversarial attack methods on data attribution. We present two methods, Shadow Attack and Outlier Attack, which generate manipulated datasets to inflate the compensation adversarially. The Shadow Attack leverages knowledge about the data distribution in the AI applications, and derives adversarial perturbations through \"shadow training\", a technique commonly used in membership inference attacks. In contrast, the Outlier Attack does not assume any knowledge about the data distribution and relies solely on black-box queries to the target model's predictions. It exploits an inductive bias present in many data attribution methods - outlier data points are more likely to be influential - and employs adversarial examples to generate manipulated datasets. Empirically, in image classification and text generation tasks, the Shadow Attack can inflate the data-attribution-based compensation by at least 200%, while the Outlier Attack achieves compensation inflation ranging from 185% to as much as 643%. Our implementation is ready at https://github.com/TRAIS-Lab/adversarial-attack-data-attribution.", "date": "2024-09-09", "categories": []}
{"id": "2410.16805", "title": "Test-time Adversarial Defense with Opposite Adversarial Path and High Attack Time Cost", "abstract": "Deep learning models are known to be vulnerable to adversarial attacks by injecting sophisticated designed perturbations to input data. Training-time defenses still exhibit a significant performance gap between natural accuracy and robust accuracy. In this paper, we investigate a new test-time adversarial defense method via diffusion-based recovery along opposite adversarial paths (OAPs). We present a purifier that can be plugged into a pre-trained model to resist adversarial attacks. Different from prior arts, the key idea is excessive denoising or purification by integrating the opposite adversarial direction with reverse diffusion to push the input image further toward the opposite adversarial direction. For the first time, we also exemplify the pitfall of conducting AutoAttack (Rand) for diffusion-based defense methods. Through the lens of time complexity, we examine the trade-off between the effectiveness of adaptive attack and its computation complexity against our defense. Experimental evaluation along with time cost analysis verifies the effectiveness of the proposed method.", "date": "2024-10-22", "categories": []}
{"id": "2410.23687", "title": "Adversarial Attacks of Vision Tasks in the Past 10 Years: A Survey", "abstract": "With the advent of Large Vision-Language Models (LVLMs), new attack vectors, such as cognitive bias, prompt injection, and jailbreaking, have emerged. Understanding these attacks promotes system robustness improvement and neural networks demystification. However, existing surveys often target attack taxonomy and lack in-depth analysis like 1) unified insights into adversariality, transferability, and generalization; 2) detailed evaluations framework; 3) motivation-driven attack categorizations; and 4) an integrated perspective on both traditional and LVLM attacks. This article addresses these gaps by offering a thorough summary of traditional and LVLM adversarial attacks, emphasizing their connections and distinctions, and providing actionable insights for future research.", "date": "2024-10-31", "categories": []}
{"id": "2505.12287", "title": "The Tower of Babel Revisited: Multilingual Jailbreak Prompts on Closed-Source Large Language Models", "abstract": "Large language models (LLMs) have seen widespread applications across various domains, yet remain vulnerable to adversarial prompt injections. While most existing research on jailbreak attacks and hallucination phenomena has focused primarily on open-source models, we investigate the frontier of closed-source LLMs under multilingual attack scenarios. We present a first-of-its-kind integrated adversarial framework that leverages diverse attack techniques to systematically evaluate frontier proprietary solutions, including GPT-4o, DeepSeek-R1, Gemini-1.5-Pro, and Qwen-Max. Our evaluation spans six categories of security contents in both English and Chinese, generating 38,400 responses across 32 types of jailbreak attacks. Attack success rate (ASR) is utilized as the quantitative metric to assess performance from three dimensions: prompt design, model architecture, and language environment. Our findings suggest that Qwen-Max is the most vulnerable, while GPT-4o shows the strongest defense. Notably, prompts in Chinese consistently yield higher ASRs than their English counterparts, and our novel Two-Sides attack technique proves to be the most effective across all models. This work highlights a dire need for language-aware alignment and robust cross-lingual defenses in LLMs, and we hope it will inspire researchers, developers, and policymakers toward more robust and inclusive AI systems.", "date": "2025-05-18", "categories": []}
{"id": "2505.12686", "title": "RoVo: Robust Voice Protection Against Unauthorized Speech Synthesis with Embedding-Level Perturbations", "abstract": "With the advancement of AI-based speech synthesis technologies such as Deep Voice, there is an increasing risk of voice spoofing attacks, including voice phishing and fake news, through unauthorized use of others' voices. Existing defenses that inject adversarial perturbations directly into audio signals have limited effectiveness, as these perturbations can easily be neutralized by speech enhancement methods. To overcome this limitation, we propose RoVo (Robust Voice), a novel proactive defense technique that injects adversarial perturbations into high-dimensional embedding vectors of audio signals, reconstructing them into protected speech. This approach effectively defends against speech synthesis attacks and also provides strong resistance to speech enhancement models, which represent a secondary attack threat.\n  In extensive experiments, RoVo increased the Defense Success Rate (DSR) by over 70% compared to unprotected speech, across four state-of-the-art speech synthesis models. Specifically, RoVo achieved a DSR of 99.5% on a commercial speaker-verification API, effectively neutralizing speech synthesis attack. Moreover, RoVo's perturbations remained robust even under strong speech enhancement conditions, outperforming traditional methods. A user study confirmed that RoVo preserves both naturalness and usability of protected speech, highlighting its effectiveness in complex and evolving threat scenarios.", "date": "2025-05-19", "categories": []}
{"id": "2505.13348", "title": "Investigating the Vulnerability of LLM-as-a-Judge Architectures to Prompt-Injection Attacks", "abstract": "Large Language Models (LLMs) are increasingly employed as evaluators (LLM-as-a-Judge) for assessing the quality of machine-generated text. This paradigm offers scalability and cost-effectiveness compared to human annotation. However, the reliability and security of such systems, particularly their robustness against adversarial manipulations, remain critical concerns. This paper investigates the vulnerability of LLM-as-a-Judge architectures to prompt-injection attacks, where malicious inputs are designed to compromise the judge's decision-making process. We formalize two primary attack strategies: Comparative Undermining Attack (CUA), which directly targets the final decision output, and Justification Manipulation Attack (JMA), which aims to alter the model's generated reasoning. Using the Greedy Coordinate Gradient (GCG) optimization method, we craft adversarial suffixes appended to one of the responses being compared. Experiments conducted on the MT-Bench Human Judgments dataset with open-source instruction-tuned LLMs (Qwen2.5-3B-Instruct and Falcon3-3B-Instruct) demonstrate significant susceptibility. The CUA achieves an Attack Success Rate (ASR) exceeding 30\\%, while JMA also shows notable effectiveness. These findings highlight substantial vulnerabilities in current LLM-as-a-Judge systems, underscoring the need for robust defense mechanisms and further research into adversarial evaluation and trustworthiness in LLM-based assessment frameworks.", "date": "2025-05-19", "categories": []}
{"id": "2505.10838", "title": "LARGO: Latent Adversarial Reflection through Gradient Optimization for Jailbreaking LLMs", "abstract": "Efficient red-teaming method to uncover vulnerabilities in Large Language Models (LLMs) is crucial. While recent attacks often use LLMs as optimizers, the discrete language space make gradient-based methods struggle. We introduce LARGO (Latent Adversarial Reflection through Gradient Optimization), a novel latent self-reflection attack that reasserts the power of gradient-based optimization for generating fluent jailbreaking prompts. By operating within the LLM's continuous latent space, LARGO first optimizes an adversarial latent vector and then recursively call the same LLM to decode the latent into natural language. This methodology yields a fast, effective, and transferable attack that produces fluent and stealthy prompts. On standard benchmarks like AdvBench and JailbreakBench, LARGO surpasses leading jailbreaking techniques, including AutoDAN, by 44 points in attack success rate. Our findings demonstrate a potent alternative to agentic LLM prompting, highlighting the efficacy of interpreting and attacking LLM internals through gradient optimization.", "date": "2025-05-16", "categories": []}
{"id": "2505.10903", "title": "On the Security Risks of ML-based Malware Detection Systems: A Survey", "abstract": "Malware presents a persistent threat to user privacy and data integrity. To combat this, machine learning-based (ML-based) malware detection (MD) systems have been developed. However, these systems have increasingly been attacked in recent years, undermining their effectiveness in practice. While the security risks associated with ML-based MD systems have garnered considerable attention, the majority of prior works is limited to adversarial malware examples, lacking a comprehensive analysis of practical security risks. This paper addresses this gap by utilizing the CIA principles to define the scope of security risks. We then deconstruct ML-based MD systems into distinct operational stages, thus developing a stage-based taxonomy. Utilizing this taxonomy, we summarize the technical progress and discuss the gaps in the attack and defense proposals related to the ML-based MD systems within each stage. Subsequently, we conduct two case studies, using both inter-stage and intra-stage analyses according to the stage-based taxonomy to provide new empirical insights. Based on these analyses and insights, we suggest potential future directions from both inter-stage and intra-stage perspectives.", "date": "2025-05-16", "categories": []}
{"id": "2505.09768", "title": "Self-Consuming Generative Models with Adversarially Curated Data", "abstract": "Recent advances in generative models have made it increasingly difficult to distinguish real data from model-generated synthetic data. Using synthetic data for successive training of future model generations creates \"self-consuming loops\", which may lead to model collapse or training instability. Furthermore, synthetic data is often subject to human feedback and curated by users based on their preferences. Ferbach et al. (2024) recently showed that when data is curated according to user preferences, the self-consuming retraining loop drives the model to converge toward a distribution that optimizes those preferences. However, in practice, data curation is often noisy or adversarially manipulated. For example, competing platforms may recruit malicious users to adversarially curate data and disrupt rival models. In this paper, we study how generative models evolve under self-consuming retraining loops with noisy and adversarially curated data. We theoretically analyze the impact of such noisy data curation on generative models and identify conditions for the robustness of the retraining process. Building on this analysis, we design attack algorithms for competitive adversarial scenarios, where a platform with a limited budget employs malicious users to misalign a rival's model from actual user preferences. Experiments on both synthetic and real-world datasets demonstrate the effectiveness of the proposed algorithms.", "date": "2025-05-14", "categories": []}
{"id": "2505.09820", "title": "Adversarial Attack on Large Language Models using Exponentiated Gradient Descent", "abstract": "As Large Language Models (LLMs) are widely used, understanding them systematically is key to improving their safety and realizing their full potential. Although many models are aligned using techniques such as reinforcement learning from human feedback (RLHF), they are still vulnerable to jailbreaking attacks. Some of the existing adversarial attack methods search for discrete tokens that may jailbreak a target model while others try to optimize the continuous space represented by the tokens of the model's vocabulary. While techniques based on the discrete space may prove to be inefficient, optimization of continuous token embeddings requires projections to produce discrete tokens, which might render them ineffective. To fully utilize the constraints and the structures of the space, we develop an intrinsic optimization technique using exponentiated gradient descent with the Bregman projection method to ensure that the optimized one-hot encoding always stays within the probability simplex. We prove the convergence of the technique and implement an efficient algorithm that is effective in jailbreaking several widely used LLMs. We demonstrate the efficacy of the proposed technique using five open-source LLMs on four openly available datasets. The results show that the technique achieves a higher success rate with great efficiency compared to three other state-of-the-art jailbreaking techniques. The source code for our implementation is available at: https://github.com/sbamit/Exponentiated-Gradient-Descent-LLM-Attack", "date": "2025-05-14", "categories": []}
{"id": "2408.13274", "title": "Robust Image Classification: Defensive Strategies against FGSM and PGD Adversarial Attacks", "abstract": "Adversarial attacks, particularly the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) pose significant threats to the robustness of deep learning models in image classification. This paper explores and refines defense mechanisms against these attacks to enhance the resilience of neural networks. We employ a combination of adversarial training and innovative preprocessing techniques, aiming to mitigate the impact of adversarial perturbations. Our methodology involves modifying input data before classification and investigating different model architectures and training strategies. Through rigorous evaluation of benchmark datasets, we demonstrate the effectiveness of our approach in defending against FGSM and PGD attacks. Our results show substantial improvements in model robustness compared to baseline methods, highlighting the potential of our defense strategies in real-world applications. This study contributes to the ongoing efforts to develop secure and reliable machine learning systems, offering practical insights and paving the way for future research in adversarial defense. By bridging theoretical advancements and practical implementation, we aim to enhance the trustworthiness of AI applications in safety-critical domains.", "date": "2024-08-20", "categories": []}
{"id": "2412.21061", "title": "BridgePure: Limited Protection Leakage Can Break Black-Box Data Protection", "abstract": "Availability attacks, or unlearnable examples, are defensive techniques that allow data owners to modify their datasets in ways that prevent unauthorized machine learning models from learning effectively while maintaining the data's intended functionality. It has led to the release of popular black-box tools (e.g., APIs) for users to upload personal data and receive protected counterparts. In this work, we show that such black-box protections can be substantially compromised if a small set of unprotected in-distribution data is available. Specifically, we propose a novel threat model of protection leakage, where an adversary can (1) easily acquire (unprotected, protected) pairs by querying the black-box protections with a small unprotected dataset; and (2) train a diffusion bridge model to build a mapping between unprotected and protected data. This mapping, termed BridgePure, can effectively remove the protection from any previously unseen data within the same distribution. BridgePure demonstrates superior purification performance on classification and style mimicry tasks, exposing critical vulnerabilities in black-box data protection. We suggest that practitioners implement multi-level countermeasures to mitigate such risks.", "date": "2024-12-30", "categories": []}
{"id": "2505.08804", "title": "TokenProber: Jailbreaking Text-to-image Models via Fine-grained Word Impact Analysis", "abstract": "Text-to-image (T2I) models have significantly advanced in producing high-quality images. However, such models have the ability to generate images containing not-safe-for-work (NSFW) content, such as pornography, violence, political content, and discrimination. To mitigate the risk of generating NSFW content, refusal mechanisms, i.e., safety checkers, have been developed to check potential NSFW content. Adversarial prompting techniques have been developed to evaluate the robustness of the refusal mechanisms. The key challenge remains to subtly modify the prompt in a way that preserves its sensitive nature while bypassing the refusal mechanisms. In this paper, we introduce TokenProber, a method designed for sensitivity-aware differential testing, aimed at evaluating the robustness of the refusal mechanisms in T2I models by generating adversarial prompts. Our approach is based on the key observation that adversarial prompts often succeed by exploiting discrepancies in how T2I models and safety checkers interpret sensitive content. Thus, we conduct a fine-grained analysis of the impact of specific words within prompts, distinguishing between dirty words that are essential for NSFW content generation and discrepant words that highlight the different sensitivity assessments between T2I models and safety checkers. Through the sensitivity-aware mutation, TokenProber generates adversarial prompts, striking a balance between maintaining NSFW content generation and evading detection. Our evaluation of TokenProber against 5 safety checkers on 3 popular T2I models, using 324 NSFW prompts, demonstrates its superior effectiveness in bypassing safety filters compared to existing methods (e.g., 54%+ increase on average), highlighting TokenProber's ability to uncover robustness issues in the existing refusal mechanisms.", "date": "2025-05-11", "categories": []}
{"id": "2505.08835", "title": "Robustness Analysis against Adversarial Patch Attacks in Fully Unmanned Stores", "abstract": "The advent of convenient and efficient fully unmanned stores equipped with artificial intelligence-based automated checkout systems marks a new era in retail. However, these systems have inherent artificial intelligence security vulnerabilities, which are exploited via adversarial patch attacks, particularly in physical environments. This study demonstrated that adversarial patches can severely disrupt object detection models used in unmanned stores, leading to issues such as theft, inventory discrepancies, and interference. We investigated three types of adversarial patch attacks -- Hiding, Creating, and Altering attacks -- and highlighted their effectiveness. We also introduce the novel color histogram similarity loss function by leveraging attacker knowledge of the color information of a target class object. Besides the traditional confusion-matrix-based attack success rate, we introduce a new bounding-boxes-based metric to analyze the practical impact of these attacks. Starting with attacks on object detection models trained on snack and fruit datasets in a digital environment, we evaluated the effectiveness of adversarial patches in a physical testbed that mimicked a real unmanned store with RGB cameras and realistic conditions. Furthermore, we assessed the robustness of these attacks in black-box scenarios, demonstrating that shadow attacks can enhance success rates of attacks even without direct access to model parameters. Our study underscores the necessity for robust defense strategies to protect unmanned stores from adversarial threats. Highlighting the limitations of the current defense mechanisms in real-time detection systems and discussing various proactive measures, we provide insights into improving the robustness of object detection models and fortifying unmanned retail environments against these attacks.", "date": "2025-05-13", "categories": []}
{"id": "2505.08999", "title": "Towards Adaptive Meta-Gradient Adversarial Examples for Visual Tracking", "abstract": "In recent years, visual tracking methods based on convolutional neural networks and Transformers have achieved remarkable performance and have been successfully applied in fields such as autonomous driving. However, the numerous security issues exposed by deep learning models have gradually affected the reliable application of visual tracking methods in real-world scenarios. Therefore, how to reveal the security vulnerabilities of existing visual trackers through effective adversarial attacks has become a critical problem that needs to be addressed. To this end, we propose an adaptive meta-gradient adversarial attack (AMGA) method for visual tracking. This method integrates multi-model ensembles and meta-learning strategies, combining momentum mechanisms and Gaussian smoothing, which can significantly enhance the transferability and attack effectiveness of adversarial examples. AMGA randomly selects models from a large model repository, constructs diverse tracking scenarios, and iteratively performs both white- and black-box adversarial attacks in each scenario, optimizing the gradient directions of each model. This paradigm minimizes the gap between white- and black-box adversarial attacks, thus achieving excellent attack performance in black-box scenarios. Extensive experimental results on large-scale datasets such as OTB2015, LaSOT, and GOT-10k demonstrate that AMGA significantly improves the attack performance, transferability, and deception of adversarial examples. Codes and data are available at https://github.com/pgao-lab/AMGA.", "date": "2025-05-13", "categories": []}
{"id": "2505.09004", "title": "Lower Bounds on the MMSE of Adversarially Inferring Sensitive Features", "abstract": "We propose an adversarial evaluation framework for sensitive feature inference based on minimum mean-squared error (MMSE) estimation with a finite sample size and linear predictive models. Our approach establishes theoretical lower bounds on the true MMSE of inferring sensitive features from noisy observations of other correlated features. These bounds are expressed in terms of the empirical MMSE under a restricted hypothesis class and a non-negative error term. The error term captures both the estimation error due to finite number of samples and the approximation error from using a restricted hypothesis class. For linear predictive models, we derive closed-form bounds, which are order optimal in terms of the noise variance, on the approximation error for several classes of relationships between the sensitive and non-sensitive features, including linear mappings, binary symmetric channels, and class-conditional multi-variate Gaussian distributions. We also present a new lower bound that relies on the MSE computed on a hold-out validation dataset of the MMSE estimator learned on finite-samples and a restricted hypothesis class. Through empirical evaluation, we demonstrate that our framework serves as an effective tool for MMSE-based adversarial evaluation of sensitive feature inference that balances theoretical guarantees with practical efficiency.", "date": "2025-05-13", "categories": []}
{"id": "2410.02240", "title": "SCA: Improve Semantic Consistent in Unrestricted Adversarial Attacks via DDPM Inversion", "abstract": "Systems based on deep neural networks are vulnerable to adversarial attacks. Unrestricted adversarial attacks typically manipulate the semantic content of an image (e.g., color or texture) to create adversarial examples that are both effective and photorealistic. Recent works have utilized the diffusion inversion process to map images into a latent space, where high-level semantics are manipulated by introducing perturbations. However, they often result in substantial semantic distortions in the denoised output and suffer from low efficiency. In this study, we propose a novel framework called Semantic-Consistent Unrestricted Adversarial Attacks (SCA), which employs an inversion method to extract edit-friendly noise maps and utilizes a Multimodal Large Language Model (MLLM) to provide semantic guidance throughout the process. Under the condition of rich semantic information provided by MLLM, we perform the DDPM denoising process of each step using a series of edit-friendly noise maps and leverage DPM Solver++ to accelerate this process, enabling efficient sampling with semantic consistency. Compared to existing methods, our framework enables the efficient generation of adversarial examples that exhibit minimal discernible semantic changes. Consequently, we for the first time introduce Semantic-Consistent Adversarial Examples (SCAE). Extensive experiments and visualizations have demonstrated the high efficiency of SCA, particularly in being on average 12 times faster than the state-of-the-art attacks. Our code can be found at https://github.com/Pan-Zihao/SCA.", "date": "2024-10-03", "categories": []}
{"id": "2412.16254", "title": "Adversarial Robustness through Dynamic Ensemble Learning", "abstract": "Adversarial attacks pose a significant threat to the reliability of pre-trained language models (PLMs) such as GPT, BERT, RoBERTa, and T5. This paper presents Adversarial Robustness through Dynamic Ensemble Learning (ARDEL), a novel scheme designed to enhance the robustness of PLMs against such attacks. ARDEL leverages the diversity of multiple PLMs and dynamically adjusts the ensemble configuration based on input characteristics and detected adversarial patterns. Key components of ARDEL include a meta-model for dynamic weighting, an adversarial pattern detection module, and adversarial training with regularization techniques. Comprehensive evaluations using standardized datasets and various adversarial attack scenarios demonstrate that ARDEL significantly improves robustness compared to existing methods. By dynamically reconfiguring the ensemble to prioritize the most robust models for each input, ARDEL effectively reduces attack success rates and maintains higher accuracy under adversarial conditions. This work contributes to the broader goal of developing more secure and trustworthy AI systems for real-world NLP applications, offering a practical and scalable solution to enhance adversarial resilience in PLMs.", "date": "2024-12-20", "categories": []}
{"id": "2505.04806", "title": "Red Teaming the Mind of the Machine: A Systematic Evaluation of Prompt Injection and Jailbreak Vulnerabilities in LLMs", "abstract": "Large Language Models (LLMs) are increasingly integrated into consumer and enterprise applications. Despite their capabilities, they remain susceptible to adversarial attacks such as prompt injection and jailbreaks that override alignment safeguards. This paper provides a systematic investigation of jailbreak strategies against various state-of-the-art LLMs. We categorize over 1,400 adversarial prompts, analyze their success against GPT-4, Claude 2, Mistral 7B, and Vicuna, and examine their generalizability and construction logic. We further propose layered mitigation strategies and recommend a hybrid red-teaming and sandboxing approach for robust LLM security.", "date": "2025-05-07", "categories": []}
{"id": "2505.07856", "title": "Unpacking Robustness in Inflectional Languages: Adversarial Evaluation and Mechanistic Insights", "abstract": "Various techniques are used in the generation of adversarial examples, including methods such as TextBugger which introduce minor, hardly visible perturbations to words leading to changes in model behaviour. Another class of techniques involves substituting words with their synonyms in a way that preserves the text's meaning but alters its predicted class, with TextFooler being a prominent example of such attacks. Most adversarial example generation methods are developed and evaluated primarily on non-inflectional languages, typically English. In this work, we evaluate and explain how adversarial attacks perform in inflectional languages. To explain the impact of inflection on model behaviour and its robustness under attack, we designed a novel protocol inspired by mechanistic interpretability, based on Edge Attribution Patching (EAP) method. The proposed evaluation protocol relies on parallel task-specific corpora that include both inflected and syncretic variants of texts in two languages -- Polish and English. To analyse the models and explain the relationship between inflection and adversarial robustness, we create a new benchmark based on task-oriented dataset MultiEmo, enabling the identification of mechanistic inflection-related elements of circuits within the model and analyse their behaviour under attack.", "date": "2025-05-08", "categories": []}
{"id": "2505.08255", "title": "Where the Devil Hides: Deepfake Detectors Can No Longer Be Trusted", "abstract": "With the advancement of AI generative techniques, Deepfake faces have become incredibly realistic and nearly indistinguishable to the human eye. To counter this, Deepfake detectors have been developed as reliable tools for assessing face authenticity. These detectors are typically developed on Deep Neural Networks (DNNs) and trained using third-party datasets. However, this protocol raises a new security risk that can seriously undermine the trustfulness of Deepfake detectors: Once the third-party data providers insert poisoned (corrupted) data maliciously, Deepfake detectors trained on these datasets will be injected ``backdoors'' that cause abnormal behavior when presented with samples containing specific triggers. This is a practical concern, as third-party providers may distribute or sell these triggers to malicious users, allowing them to manipulate detector performance and escape accountability.\n  This paper investigates this risk in depth and describes a solution to stealthily infect Deepfake detectors. Specifically, we develop a trigger generator, that can synthesize passcode-controlled, semantic-suppression, adaptive, and invisible trigger patterns, ensuring both the stealthiness and effectiveness of these triggers. Then we discuss two poisoning scenarios, dirty-label poisoning and clean-label poisoning, to accomplish the injection of backdoors. Extensive experiments demonstrate the effectiveness, stealthiness, and practicality of our method compared to several baselines.", "date": "2025-05-13", "categories": []}
{"id": "2303.13326", "title": "Decentralized Adversarial Training over Graphs", "abstract": "The vulnerability of machine learning models to adversarial attacks has been attracting considerable attention in recent years. Most existing studies focus on the behavior of stand-alone single-agent learners. In comparison, this work studies adversarial training over graphs, where individual agents are subjected to perturbations of varied strength levels across space. It is expected that interactions by linked agents, and the heterogeneity of the attack models that are possible over the graph, can help enhance robustness in view of the coordination power of the group. Using a min-max formulation of distributed learning, we develop a decentralized adversarial training framework for multi-agent systems. Specifically, we devise two decentralized adversarial training algorithms by relying on two popular decentralized learning strategies--diffusion and consensus. We analyze the convergence properties of the proposed framework for strongly-convex, convex, and non-convex environments, and illustrate the enhanced robustness to adversarial attacks.", "date": "2023-03-23", "categories": []}
{"id": "2311.11796", "title": "Beyond Boundaries: A Comprehensive Survey of Transferable Attacks on AI Systems", "abstract": "As Artificial Intelligence (AI) systems increasingly underpin critical applications, from autonomous vehicles to biometric authentication, their vulnerability to transferable attacks presents a growing concern. These attacks, designed to generalize across instances, domains, models, tasks, modalities, or even hardware platforms, pose severe risks to security, privacy, and system integrity. This survey delivers the first comprehensive review of transferable attacks across seven major categories, including evasion, backdoor, data poisoning, model stealing, model inversion, membership inference, and side-channel attacks. We introduce a unified six-dimensional taxonomy: cross-instance, cross-domain, cross-modality, cross-model, cross-task, and cross-hardware, which systematically captures the diverse transfer pathways of adversarial strategies. Through this framework, we examine both the underlying mechanics and practical implications of transferable attacks on AI systems. Furthermore, we review cutting-edge methods for enhancing attack transferability, organized around data augmentation and optimization strategies. By consolidating fragmented research and identifying critical future directions, this work provides a foundational roadmap for understanding, evaluating, and defending against transferable threats in real-world AI systems.", "date": "2023-11-20", "categories": []}
{"id": "2403.11297", "title": "A Modified Word Saliency-Based Adversarial Attack on Text Classification Models", "abstract": "This paper introduces a novel adversarial attack method targeting text classification models, termed the Modified Word Saliency-based Adversarial At-tack (MWSAA). The technique builds upon the concept of word saliency to strategically perturb input texts, aiming to mislead classification models while preserving semantic coherence. By refining the traditional adversarial attack approach, MWSAA significantly enhances its efficacy in evading detection by classification systems. The methodology involves first identifying salient words in the input text through a saliency estimation process, which prioritizes words most influential to the model's decision-making process. Subsequently, these salient words are subjected to carefully crafted modifications, guided by semantic similarity metrics to ensure that the altered text remains coherent and retains its original meaning. Empirical evaluations conducted on diverse text classification datasets demonstrate the effectiveness of the proposed method in generating adversarial examples capable of successfully deceiving state-of-the-art classification models. Comparative analyses with existing adversarial attack techniques further indicate the superiority of the proposed approach in terms of both attack success rate and preservation of text coherence.", "date": "2024-03-17", "categories": []}
{"id": "2404.19460", "title": "AttackBench: Evaluating Gradient-based Attacks for Adversarial Examples", "abstract": "Adversarial examples are typically optimized with gradient-based attacks. While novel attacks are continuously proposed, each is shown to outperform its predecessors using different experimental setups, hyperparameter settings, and number of forward and backward calls to the target models. This provides overly-optimistic and even biased evaluations that may unfairly favor one particular attack over the others. In this work, we aim to overcome these limitations by proposing AttackBench, i.e., the first evaluation framework that enables a fair comparison among different attacks. To this end, we first propose a categorization of gradient-based attacks, identifying their main components and differences. We then introduce our framework, which evaluates their effectiveness and efficiency. We measure these characteristics by (i) defining an optimality metric that quantifies how close an attack is to the optimal solution, and (ii) limiting the number of forward and backward queries to the model, such that all attacks are compared within a given maximum query budget. Our extensive experimental analysis compares more than $100$ attack implementations with a total of over $800$ different configurations against CIFAR-10 and ImageNet models, highlighting that only very few attacks outperform all the competing approaches. Within this analysis, we shed light on several implementation issues that prevent many attacks from finding better solutions or running at all. We release AttackBench as a publicly-available benchmark, aiming to continuously update it to include and evaluate novel gradient-based attacks for optimizing adversarial examples.", "date": "2024-04-30", "categories": []}
{"id": "2501.09798", "title": "Fun-tuning: Characterizing the Vulnerability of Proprietary LLMs to Optimization-based Prompt Injection Attacks via the Fine-Tuning Interface", "abstract": "We surface a new threat to closed-weight Large Language Models (LLMs) that enables an attacker to compute optimization-based prompt injections. Specifically, we characterize how an attacker can leverage the loss-like information returned from the remote fine-tuning interface to guide the search for adversarial prompts. The fine-tuning interface is hosted by an LLM vendor and allows developers to fine-tune LLMs for their tasks, thus providing utility, but also exposes enough information for an attacker to compute adversarial prompts. Through an experimental analysis, we characterize the loss-like values returned by the Gemini fine-tuning API and demonstrate that they provide a useful signal for discrete optimization of adversarial prompts using a greedy search algorithm. Using the PurpleLlama prompt injection benchmark, we demonstrate attack success rates between 65% and 82% on Google's Gemini family of LLMs. These attacks exploit the classic utility-security tradeoff - the fine-tuning interface provides a useful feature for developers but also exposes the LLMs to powerful attacks.", "date": "2025-01-16", "categories": []}
{"id": "2505.06299", "title": "Input-Specific and Universal Adversarial Attack Generation for Spiking Neural Networks in the Spiking Domain", "abstract": "As Spiking Neural Networks (SNNs) gain traction across various applications, understanding their security vulnerabilities becomes increasingly important. In this work, we focus on the adversarial attacks, which is perhaps the most concerning threat. An adversarial attack aims at finding a subtle input perturbation to fool the network's decision-making. We propose two novel adversarial attack algorithms for SNNs: an input-specific attack that crafts adversarial samples from specific dataset inputs and a universal attack that generates a reusable patch capable of inducing misclassification across most inputs, thus offering practical feasibility for real-time deployment. The algorithms are gradient-based operating in the spiking domain proving to be effective across different evaluation metrics, such as adversarial accuracy, stealthiness, and generation time. Experimental results on two widely used neuromorphic vision datasets, NMNIST and IBM DVS Gesture, show that our proposed attacks surpass in all metrics all existing state-of-the-art methods. Additionally, we present the first demonstration of adversarial attack generation in the sound domain using the SHD dataset.", "date": "2025-05-07", "categories": []}
{"id": "2505.06454", "title": "Sponge Attacks on Sensing AI: Energy-Latency Vulnerabilities and Defense via Model Pruning", "abstract": "Recent studies have shown that sponge attacks can significantly increase the energy consumption and inference latency of deep neural networks (DNNs). However, prior work has focused primarily on computer vision and natural language processing tasks, overlooking the growing use of lightweight AI models in sensing-based applications on resource-constrained devices, such as those in Internet of Things (IoT) environments. These attacks pose serious threats of energy depletion and latency degradation in systems where limited battery capacity and real-time responsiveness are critical for reliable operation. This paper makes two key contributions. First, we present the first systematic exploration of energy-latency sponge attacks targeting sensing-based AI models. Using wearable sensing-based AI as a case study, we demonstrate that sponge attacks can substantially degrade performance by increasing energy consumption, leading to faster battery drain, and by prolonging inference latency. Second, to mitigate such attacks, we investigate model pruning, a widely adopted compression technique for resource-constrained AI, as a potential defense. Our experiments show that pruning-induced sparsity significantly improves model resilience against sponge poisoning. We also quantify the trade-offs between model efficiency and attack resilience, offering insights into the security implications of model compression in sensing-based AI systems deployed in IoT environments.", "date": "2025-05-09", "categories": []}
{"id": "2505.06860", "title": "DP-TRAE: A Dual-Phase Merging Transferable Reversible Adversarial Example for Image Privacy Protection", "abstract": "In the field of digital security, Reversible Adversarial Examples (RAE) combine adversarial attacks with reversible data hiding techniques to effectively protect sensitive data and prevent unauthorized analysis by malicious Deep Neural Networks (DNNs). However, existing RAE techniques primarily focus on white-box attacks, lacking a comprehensive evaluation of their effectiveness in black-box scenarios. This limitation impedes their broader deployment in complex, dynamic environments. Further more, traditional black-box attacks are often characterized by poor transferability and high query costs, significantly limiting their practical applicability. To address these challenges, we propose the Dual-Phase Merging Transferable Reversible Attack method, which generates highly transferable initial adversarial perturbations in a white-box model and employs a memory augmented black-box strategy to effectively mislead target mod els. Experimental results demonstrate the superiority of our approach, achieving a 99.0% attack success rate and 100% recovery rate in black-box scenarios, highlighting its robustness in privacy protection. Moreover, we successfully implemented a black-box attack on a commercial model, further substantiating the potential of this approach for practical use.", "date": "2025-05-11", "categories": []}
{"id": "2505.06889", "title": "IM-BERT: Enhancing Robustness of BERT through the Implicit Euler Method", "abstract": "Pre-trained Language Models (PLMs) have achieved remarkable performance on diverse NLP tasks through pre-training and fine-tuning. However, fine-tuning the model with a large number of parameters on limited downstream datasets often leads to vulnerability to adversarial attacks, causing overfitting of the model on standard datasets.\n  To address these issues, we propose IM-BERT from the perspective of a dynamic system by conceptualizing a layer of BERT as a solution of Ordinary Differential Equations (ODEs). Under the situation of initial value perturbation, we analyze the numerical stability of two main numerical ODE solvers: the explicit and implicit Euler approaches.\n  Based on these analyses, we introduce a numerically robust IM-connection incorporating BERT's layers. This strategy enhances the robustness of PLMs against adversarial attacks, even in low-resource scenarios, without introducing additional parameters or adversarial training strategies.\n  Experimental results on the adversarial GLUE (AdvGLUE) dataset validate the robustness of IM-BERT under various conditions. Compared to the original BERT, IM-BERT exhibits a performance improvement of approximately 8.3\\%p on the AdvGLUE dataset. Furthermore, in low-resource scenarios, IM-BERT outperforms BERT by achieving 5.9\\%p higher accuracy.", "date": "2025-05-11", "categories": []}
{"id": "2505.05279", "title": "MTL-UE: Learning to Learn Nothing for Multi-Task Learning", "abstract": "  Most existing unlearnable strategies focus on preventing unauthorized users\nfrom training single-task learning (STL) models with personal data.\nNevertheless, the paradigm has recently shifted towards multi-task data and\nmulti-task learning (MTL), targeting generalist and foundation models that can\nhandle multiple tasks simultaneously. Despite their growing importance, MTL\ndata and models have been largely neglected while pursuing unlearnable\nstrategies. This paper presents MTL-UE, the first unified framework for\ngenerating unlearnable examples for multi-task data and MTL models. Instead of\noptimizing perturbations for each sample, we design a generator-based structure\nthat introduces label priors and class-wise feature embeddings which leads to\nmuch better attacking performance. In addition, MTL-UE incorporates intra-task\nand inter-task embedding regularization to increase inter-class separation and\nsuppress intra-class variance which enhances the attack robustness greatly.\nFurthermore, MTL-UE is versatile with good supports for dense prediction tasks\nin MTL. It is also plug-and-play allowing integrating existing\nsurrogate-dependent unlearnable methods with little adaptation. Extensive\nexperiments show that MTL-UE achieves superior attacking performance\nconsistently across 4 MTL datasets, 3 base UE methods, 5 model backbones, and 5\nMTL task-weighting strategies.\n", "date": "2025-05-08", "categories": []}
{"id": "2211.00825", "title": "LMD: A Learnable Mask Network to Detect Adversarial Examples for Speaker\n  Verification", "abstract": "  Although the security of automatic speaker verification (ASV) is seriously\nthreatened by recently emerged adversarial attacks, there have been some\ncountermeasures to alleviate the threat. However, many defense approaches not\nonly require the prior knowledge of the attackers but also possess weak\ninterpretability. To address this issue, in this paper, we propose an\nattacker-independent and interpretable method, named learnable mask detector\n(LMD), to separate adversarial examples from the genuine ones. It utilizes\nscore variation as an indicator to detect adversarial examples, where the score\nvariation is the absolute discrepancy between the ASV scores of an original\naudio recording and its transformed audio synthesized from its masked complex\nspectrogram. A core component of the score variation detector is to generate\nthe masked spectrogram by a neural network. The neural network needs only\ngenuine examples for training, which makes it an attacker-independent approach.\nIts interpretability lies that the neural network is trained to minimize the\nscore variation of the targeted ASV, and maximize the number of the masked\nspectrogram bins of the genuine training examples. Its foundation is based on\nthe observation that, masking out the vast majority of the spectrogram bins\nwith little speaker information will inevitably introduce a large score\nvariation to the adversarial example, and a small score variation to the\ngenuine example. Experimental results with 12 attackers and two representative\nASV systems show that our proposed method outperforms five state-of-the-art\nbaselines. The extensive experimental results can also be a benchmark for the\ndetection-based ASV defenses.\n", "date": "2022-11-02", "categories": []}
{"id": "2406.02044", "title": "Towards Universal and Black-Box Query-Response Only Attack on LLMs with\n  QROA", "abstract": "  The rapid adoption of Large Language Models (LLMs) has exposed critical\nsecurity and ethical vulnerabilities, particularly their susceptibility to\nadversarial manipulations. This paper introduces QROA, a novel black-box\njailbreak method designed to identify adversarial suffixes that can bypass LLM\nalignment safeguards when appended to a malicious instruction. Unlike existing\nsuffix-based jailbreak approaches, QROA does not require access to the model's\nlogit or any other internal information. It also eliminates reliance on\nhuman-crafted templates, operating solely through the standard query-response\ninterface of LLMs. By framing the attack as an optimization bandit problem,\nQROA employs a surrogate model and token level optimization to efficiently\nexplore suffix variations. Furthermore, we propose QROA-UNV, an extension that\nidentifies universal adversarial suffixes for individual models, enabling\none-query jailbreaks across a wide range of instructions. Testing on multiple\nmodels demonstrates Attack Success Rate (ASR) greater than 80\\%. These findings\nhighlight critical vulnerabilities, emphasize the need for advanced defenses,\nand contribute to the development of more robust safety evaluations for secure\nAI deployment. The code is made public on the following link:\nhttps://github.com/qroa/QROA\n", "date": "2024-06-04", "categories": []}
{"id": "2505.03831", "title": "A Comprehensive Analysis of Adversarial Attacks against Spam Filters", "abstract": "  Deep learning has revolutionized email filtering, which is critical to\nprotect users from cyber threats such as spam, malware, and phishing. However,\nthe increasing sophistication of adversarial attacks poses a significant\nchallenge to the effectiveness of these filters. This study investigates the\nimpact of adversarial attacks on deep learning-based spam detection systems\nusing real-world datasets. Six prominent deep learning models are evaluated on\nthese datasets, analyzing attacks at the word, character sentence, and\nAI-generated paragraph-levels. Novel scoring functions, including spam weights\nand attention weights, are introduced to improve attack effectiveness. This\ncomprehensive analysis sheds light on the vulnerabilities of spam filters and\ncontributes to efforts to improve their security against evolving adversarial\nthreats.\n", "date": "2025-05-04", "categories": []}
{"id": "2312.13435", "title": "The Adaptive Arms Race: Redefining Robustness in AI Security", "abstract": "  Despite considerable efforts on making them robust, real-world AI-based\nsystems remain vulnerable to decision based attacks, as definitive proofs of\ntheir operational robustness have so far proven intractable. Canonical\nrobustness evaluation relies on adaptive attacks, which leverage complete\nknowledge of the defense and are tailored to bypass it. This work broadens the\nnotion of adaptivity, which we employ to enhance both attacks and defenses,\nshowing how they can benefit from mutual learning through interaction. We\nintroduce a framework for adaptively optimizing black-box attacks and defenses\nunder the competitive game they form. To assess robustness reliably, it is\nessential to evaluate against realistic and worst-case attacks. We thus enhance\nattacks and their evasive arsenal together using RL, apply the same principle\nto defenses, and evaluate them first independently and then jointly under a\nmulti-agent perspective. We find that active defenses, those that dynamically\ncontrol system responses, are an essential complement to model hardening\nagainst decision-based attacks; that these defenses can be circumvented by\nadaptive attacks, something that elicits defenses being adaptive too. Our\nfindings, supported by an extensive theoretical and empirical investigation,\nconfirm that adaptive adversaries pose a serious threat to black-box AI-based\nsystems, rekindling the proverbial arms race. Notably, our approach outperforms\nthe state-of-the-art black-box attacks and defenses, while bringing them\ntogether to render effective insights into the robustness of real-world\ndeployed ML-based systems.\n", "date": "2023-12-20", "categories": []}
{"id": "2401.02663", "title": "Effective backdoor attack on graph neural networks in link prediction\n  tasks", "abstract": "  Graph Neural Networks (GNNs) are a class of deep learning models capable of\nprocessing graph-structured data, and they have demonstrated significant\nperformance in a variety of real-world applications. Recent studies have found\nthat GNN models are vulnerable to backdoor attacks. When specific patterns\n(called backdoor triggers, e.g., subgraphs, nodes, etc.) appear in the input\ndata, the backdoor embedded in the GNN models is activated, which misclassifies\nthe input data into the target class label specified by the attacker, whereas\nwhen there are no backdoor triggers in the input, the backdoor embedded in the\nGNN models is not activated, and the models work normally. Backdoor attacks are\nhighly stealthy and expose GNN models to serious security risks. Currently,\nresearch on backdoor attacks against GNNs mainly focus on tasks such as graph\nclassification and node classification, and backdoor attacks against link\nprediction tasks are rarely studied. In this paper, we propose a backdoor\nattack against the link prediction tasks based on GNNs and reveal the existence\nof such security vulnerability in GNN models, which make the backdoored GNN\nmodels to incorrectly predict unlinked two nodes as having a link relationship\nwhen a trigger appear. The method uses a single node as the trigger and poison\nselected node pairs in the training graph, and then the backdoor will be\nembedded in the GNN models through the training process. In the inference\nstage, the backdoor in the GNN models can be activated by simply linking the\ntrigger node to the two end nodes of the unlinked node pairs in the input data,\ncausing the GNN models to produce incorrect link prediction results for the\ntarget node pairs.\n", "date": "2024-01-05", "categories": []}
{"id": "2503.06269", "title": "Using Mechanistic Interpretability to Craft Adversarial Attacks against\n  Large Language Models", "abstract": "  Traditional white-box methods for creating adversarial perturbations against\nLLMs typically rely only on gradient computation from the targeted model,\nignoring the internal mechanisms responsible for attack success or failure.\nConversely, interpretability studies that analyze these internal mechanisms\nlack practical applications beyond runtime interventions. We bridge this gap by\nintroducing a novel white-box approach that leverages mechanistic\ninterpretability techniques to craft practical adversarial inputs.\nSpecifically, we first identify acceptance subspaces - sets of feature vectors\nthat do not trigger the model's refusal mechanisms - then use gradient-based\noptimization to reroute embeddings from refusal subspaces to acceptance\nsubspaces, effectively achieving jailbreaks. This targeted approach\nsignificantly reduces computation cost, achieving attack success rates of\n80-95\\% on state-of-the-art models including Gemma2, Llama3.2, and Qwen2.5\nwithin minutes or even seconds, compared to existing techniques that often fail\nor require hours of computation. We believe this approach opens a new direction\nfor both attack research and defense development. Furthermore, it showcases a\npractical application of mechanistic interpretability where other methods are\nless efficient, which highlights its utility. The code and generated datasets\nare available at https://github.com/Sckathach/subspace-rerouting.\n", "date": "2025-03-08", "categories": []}
{"id": "2505.03383", "title": "Attention-aggregated Attack for Boosting the Transferability of Facial\n  Adversarial Examples", "abstract": "  Adversarial examples have revealed the vulnerability of deep learning models\nand raised serious concerns about information security. The transfer-based\nattack is a hot topic in black-box attacks that are practical to real-world\nscenarios where the training datasets, parameters, and structure of the target\nmodel are unknown to the attacker. However, few methods consider the\nparticularity of class-specific deep models for fine-grained vision tasks, such\nas face recognition (FR), giving rise to unsatisfactory attacking performance.\nIn this work, we first investigate what in a face exactly contributes to the\nembedding learning of FR models and find that both decisive and auxiliary\nfacial features are specific to each FR model, which is quite different from\nthe biological mechanism of human visual system. Accordingly we then propose a\nnovel attack method named Attention-aggregated Attack (AAA) to enhance the\ntransferability of adversarial examples against FR, which is inspired by the\nattention divergence and aims to destroy the facial features that are critical\nfor the decision-making of other FR models by imitating their attentions on the\nclean face images. Extensive experiments conducted on various FR models\nvalidate the superiority and robust effectiveness of the proposed method over\nexisting methods.\n", "date": "2025-05-06", "categories": []}
{"id": "2505.03435", "title": "Robustness in AI-Generated Detection: Enhancing Resistance to\n  Adversarial Attacks", "abstract": "  The rapid advancement of generative image technology has introduced\nsignificant security concerns, particularly in the domain of face generation\ndetection. This paper investigates the vulnerabilities of current AI-generated\nface detection systems. Our study reveals that while existing detection methods\noften achieve high accuracy under standard conditions, they exhibit limited\nrobustness against adversarial attacks. To address these challenges, we propose\nan approach that integrates adversarial training to mitigate the impact of\nadversarial examples. Furthermore, we utilize diffusion inversion and\nreconstruction to further enhance detection robustness. Experimental results\ndemonstrate that minor adversarial perturbations can easily bypass existing\ndetection systems, but our method significantly improves the robustness of\nthese systems. Additionally, we provide an in-depth analysis of adversarial and\nbenign examples, offering insights into the intrinsic characteristics of\nAI-generated content. All associated code will be made publicly available in a\ndedicated repository to facilitate further research and verification.\n", "date": "2025-05-06", "categories": []}
{"id": "2407.08159", "title": "Model-agnostic clean-label backdoor mitigation in cybersecurity\n  environments", "abstract": "  The training phase of machine learning models is a delicate step, especially\nin cybersecurity contexts. Recent research has surfaced a series of insidious\ntraining-time attacks that inject backdoors in models designed for security\nclassification tasks without altering the training labels. With this work, we\npropose new techniques that leverage insights in cybersecurity threat models to\neffectively mitigate these clean-label poisoning attacks, while preserving the\nmodel utility. By performing density-based clustering on a carefully chosen\nfeature subspace, and progressively isolating the suspicious clusters through a\nnovel iterative scoring procedure, our defensive mechanism can mitigate the\nattacks without requiring many of the common assumptions in the existing\nbackdoor defense literature. To show the generality of our proposed mitigation,\nwe evaluate it on two clean-label model-agnostic attacks on two different\nclassic cybersecurity data modalities: network flows classification and malware\nclassification, using gradient boosting and neural network models.\n", "date": "2024-07-11", "categories": []}
{"id": "2505.01816", "title": "Rogue Cell: Adversarial Attack and Defense in Untrusted O-RAN Setup\n  Exploiting the Traffic Steering xApp", "abstract": "  The Open Radio Access Network (O-RAN) architecture is revolutionizing\ncellular networks with its open, multi-vendor design and AI-driven management,\naiming to enhance flexibility and reduce costs. Although it has many\nadvantages, O-RAN is not threat-free. While previous studies have mainly\nexamined vulnerabilities arising from O-RAN's intelligent components, this\npaper is the first to focus on the security challenges and vulnerabilities\nintroduced by transitioning from single-operator to multi-operator RAN\narchitectures. This shift increases the risk of untrusted third-party operators\nmanaging different parts of the network. To explore these vulnerabilities and\ntheir potential mitigation, we developed an open-access testbed environment\nthat integrates a wireless network simulator with the official O-RAN Software\nCommunity (OSC) RAN intelligent component (RIC) cluster. This environment\nenables realistic, live data collection and serves as a platform for\ndemonstrating APATE (adversarial perturbation against traffic efficiency), an\nevasion attack in which a malicious cell manipulates its reported key\nperformance indicators (KPIs) and deceives the O-RAN traffic steering to gain\nunfair allocations of user equipment (UE). To ensure that O-RAN's legitimate\nactivity continues, we introduce MARRS (monitoring adversarial RAN reports), a\ndetection framework based on a long-short term memory (LSTM) autoencoder (AE)\nthat learns contextual features across the network to monitor malicious\ntelemetry (also demonstrated in our testbed). Our evaluation showed that by\nexecuting APATE, an attacker can obtain a 248.5% greater UE allocation than it\nwas supposed to in a benign scenario. In addition, the MARRS detection method\nwas also shown to successfully classify malicious cell activity, achieving\naccuracy of 99.2% and an F1 score of 0.978.\n", "date": "2025-05-03", "categories": []}
{"id": "2505.02073", "title": "Lightweight Defense Against Adversarial Attacks in Time Series\n  Classification", "abstract": "  As time series classification (TSC) gains prominence, ensuring robust TSC\nmodels against adversarial attacks is crucial. While adversarial defense is\nwell-studied in Computer Vision (CV), the TSC field has primarily relied on\nadversarial training (AT), which is computationally expensive. In this paper,\nfive data augmentation-based defense methods tailored for time series are\ndeveloped, with the most computationally intensive method among them increasing\nthe computational resources by only 14.07% compared to the original TSC model.\nMoreover, the deployment process for these methods is straightforward. By\nleveraging these advantages of our methods, we create two combined methods. One\nof these methods is an ensemble of all the proposed techniques, which not only\nprovides better defense performance than PGD-based AT but also enhances the\ngeneralization ability of TSC models. Moreover, the computational resources\nrequired for our ensemble are less than one-third of those required for\nPGD-based AT. These methods advance robust TSC in data mining. Furthermore, as\nfoundation models are increasingly explored for time series feature learning,\nour work provides insights into integrating data augmentation-based adversarial\ndefense with large-scale pre-trained models in future research.\n", "date": "2025-05-04", "categories": []}
{"id": "2411.16721", "title": "Steering Away from Harm: An Adaptive Approach to Defending Vision\n  Language Model Against Jailbreaks", "abstract": "  Vision Language Models (VLMs) can produce unintended and harmful content when\nexposed to adversarial attacks, particularly because their vision capabilities\ncreate new vulnerabilities. Existing defenses, such as input preprocessing,\nadversarial training, and response evaluation-based methods, are often\nimpractical for real-world deployment due to their high costs. To address this\nchallenge, we propose ASTRA, an efficient and effective defense by adaptively\nsteering models away from adversarial feature directions to resist VLM attacks.\nOur key procedures involve finding transferable steering vectors representing\nthe direction of harmful response and applying adaptive activation steering to\nremove these directions at inference time. To create effective steering\nvectors, we randomly ablate the visual tokens from the adversarial images and\nidentify those most strongly associated with jailbreaks. These tokens are then\nused to construct steering vectors. During inference, we perform the adaptive\nsteering method that involves the projection between the steering vectors and\ncalibrated activation, resulting in little performance drops on benign inputs\nwhile strongly avoiding harmful outputs under adversarial inputs. Extensive\nexperiments across multiple models and baselines demonstrate our\nstate-of-the-art performance and high efficiency in mitigating jailbreak risks.\nAdditionally, ASTRA exhibits good transferability, defending against unseen\nattacks (i.e., structured-based attack, perturbation-based attack with project\ngradient descent variants, and text-only attack). Our code is available at\n\\url{https://github.com/ASTRAL-Group/ASTRA}.\n", "date": "2024-11-23", "categories": []}
{"id": "2505.01050", "title": "Transferable Adversarial Attacks on Black-Box Vision-Language Models", "abstract": "  Vision Large Language Models (VLLMs) are increasingly deployed to offer\nadvanced capabilities on inputs comprising both text and images. While prior\nresearch has shown that adversarial attacks can transfer from open-source to\nproprietary black-box models in text-only and vision-only contexts, the extent\nand effectiveness of such vulnerabilities remain underexplored for VLLMs. We\npresent a comprehensive analysis demonstrating that targeted adversarial\nexamples are highly transferable to widely-used proprietary VLLMs such as\nGPT-4o, Claude, and Gemini. We show that attackers can craft perturbations to\ninduce specific attacker-chosen interpretations of visual information, such as\nmisinterpreting hazardous content as safe, overlooking sensitive or restricted\nmaterial, or generating detailed incorrect responses aligned with the\nattacker's intent. Furthermore, we discover that universal perturbations --\nmodifications applicable to a wide set of images -- can consistently induce\nthese misinterpretations across multiple proprietary VLLMs. Our experimental\nresults on object recognition, visual question answering, and image captioning\nshow that this vulnerability is common across current state-of-the-art models,\nand underscore an urgent need for robust mitigations to ensure the safe and\nsecure deployment of VLLMs.\n", "date": "2025-05-02", "categories": []}
{"id": "2505.01130", "title": "Risk Analysis and Design Against Adversarial Actions", "abstract": "  Learning models capable of providing reliable predictions in the face of\nadversarial actions has become a central focus of the machine learning\ncommunity in recent years. This challenge arises from observing that data\nencountered at deployment time often deviate from the conditions under which\nthe model was trained. In this paper, we address deployment-time adversarial\nactions and propose a versatile, well-principled framework to evaluate the\nmodel's robustness against attacks of diverse types and intensities. While we\ninitially focus on Support Vector Regression (SVR), the proposed approach\nextends naturally to the broad domain of learning via relaxed optimization\ntechniques. Our results enable an assessment of the model vulnerability without\nrequiring additional test data and operate in a distribution-free setup. These\nresults not only provide a tool to enhance trust in the model's applicability\nbut also aid in selecting among competing alternatives. Later in the paper, we\nshow that our findings also offer useful insights for establishing new results\nwithin the out-of-distribution framework.\n", "date": "2025-05-02", "categories": []}
{"id": "2505.01168", "title": "Harmonizing Intra-coherence and Inter-divergence in Ensemble Attacks for\n  Adversarial Transferability", "abstract": "  The development of model ensemble attacks has significantly improved the\ntransferability of adversarial examples, but this progress also poses severe\nthreats to the security of deep neural networks. Existing methods, however,\nface two critical challenges: insufficient capture of shared gradient\ndirections across models and a lack of adaptive weight allocation mechanisms.\nTo address these issues, we propose a novel method Harmonized Ensemble for\nAdversarial Transferability (HEAT), which introduces domain generalization into\nadversarial example generation for the first time. HEAT consists of two key\nmodules: Consensus Gradient Direction Synthesizer, which uses Singular Value\nDecomposition to synthesize shared gradient directions; and Dual-Harmony Weight\nOrchestrator which dynamically balances intra-domain coherence, stabilizing\ngradients within individual models, and inter-domain diversity, enhancing\ntransferability across models. Experimental results demonstrate that HEAT\nsignificantly outperforms existing methods across various datasets and\nsettings, offering a new perspective and direction for adversarial attack\nresearch.\n", "date": "2025-05-02", "categories": []}
{"id": "2505.01328", "title": "Constrained Network Adversarial Attacks: Validity, Robustness, and\n  Transferability", "abstract": "  While machine learning has significantly advanced Network Intrusion Detection\nSystems (NIDS), particularly within IoT environments where devices generate\nlarge volumes of data and are increasingly susceptible to cyber threats, these\nmodels remain vulnerable to adversarial attacks. Our research reveals a\ncritical flaw in existing adversarial attack methodologies: the frequent\nviolation of domain-specific constraints, such as numerical and categorical\nlimits, inherent to IoT and network traffic. This leads to up to 80.3% of\nadversarial examples being invalid, significantly overstating real-world\nvulnerabilities. These invalid examples, though effective in fooling models, do\nnot represent feasible attacks within practical IoT deployments. Consequently,\nrelying on these results can mislead resource allocation for defense, inflating\nthe perceived susceptibility of IoT-enabled NIDS models to adversarial\nmanipulation. Furthermore, we demonstrate that simpler surrogate models like\nMulti-Layer Perceptron (MLP) generate more valid adversarial examples compared\nto complex architectures such as CNNs and LSTMs. Using the MLP as a surrogate,\nwe analyze the transferability of adversarial severity to other ML/DL models\ncommonly used in IoT contexts. This work underscores the importance of\nconsidering both domain constraints and model architecture when evaluating and\ndesigning robust ML/DL models for security-critical IoT and network\napplications.\n", "date": "2025-05-02", "categories": []}
{"id": "2410.05573", "title": "TaeBench: Improving Quality of Toxic Adversarial Examples", "abstract": "  Toxicity text detectors can be vulnerable to adversarial examples - small\nperturbations to input text that fool the systems into wrong detection.\nExisting attack algorithms are time-consuming and often produce invalid or\nambiguous adversarial examples, making them less useful for evaluating or\nimproving real-world toxicity content moderators. This paper proposes an\nannotation pipeline for quality control of generated toxic adversarial examples\n(TAE). We design model-based automated annotation and human-based quality\nverification to assess the quality requirements of TAE. Successful TAE should\nfool a target toxicity model into making benign predictions, be grammatically\nreasonable, appear natural like human-generated text, and exhibit semantic\ntoxicity. When applying these requirements to more than 20 state-of-the-art\n(SOTA) TAE attack recipes, we find many invalid samples from a total of 940k\nraw TAE attack generations. We then utilize the proposed pipeline to filter and\ncurate a high-quality TAE dataset we call TaeBench (of size 264k). Empirically,\nwe demonstrate that TaeBench can effectively transfer-attack SOTA toxicity\ncontent moderation models and services. Our experiments also show that TaeBench\nwith adversarial training achieve significant improvements of the robustness of\ntwo toxicity detectors.\n", "date": "2024-10-08", "categories": []}
{"id": "2411.14412", "title": "Adversarial Data Poisoning Attacks on Quantum Machine Learning in the\n  NISQ Era", "abstract": "  With the growing interest in Quantum Machine Learning (QML) and the\nincreasing availability of quantum computers through cloud providers,\naddressing the potential security risks associated with QML has become an\nurgent priority. One key concern in the QML domain is the threat of data\npoisoning attacks in the current quantum cloud setting. Adversarial access to\ntraining data could severely compromise the integrity and availability of QML\nmodels. Classical data poisoning techniques require significant knowledge and\ntraining to generate poisoned data, and lack noise resilience, making them\nineffective for QML models in the Noisy Intermediate Scale Quantum (NISQ) era.\nIn this work, we first propose a simple yet effective technique to measure\nintra-class encoder state similarity (ESS) by analyzing the outputs of encoding\ncircuits. Leveraging this approach, we introduce a \\underline{Qu}antum\n\\underline{I}ndiscriminate \\underline{D}ata Poisoning attack, QUID. Through\nextensive experiments conducted in both noiseless and noisy environments (e.g.,\nIBM\\_Brisbane's noise), across various architectures and datasets, QUID\nachieves up to $92\\%$ accuracy degradation in model performance compared to\nbaseline models and up to $75\\%$ accuracy degradation compared to random\nlabel-flipping. We also tested QUID against state-of-the-art classical\ndefenses, with accuracy degradation still exceeding $50\\%$, demonstrating its\neffectiveness. This work represents the first attempt to reevaluate data\npoisoning attacks in the context of QML.\n", "date": "2024-11-21", "categories": []}
{"id": "2505.00061", "title": "Enhancing Security and Strengthening Defenses in Automated Short-Answer\n  Grading Systems", "abstract": "  This study examines vulnerabilities in transformer-based automated\nshort-answer grading systems used in medical education, with a focus on how\nthese systems can be manipulated through adversarial gaming strategies. Our\nresearch identifies three main types of gaming strategies that exploit the\nsystem's weaknesses, potentially leading to false positives. To counteract\nthese vulnerabilities, we implement several adversarial training methods\ndesigned to enhance the systems' robustness. Our results indicate that these\nmethods significantly reduce the susceptibility of grading systems to such\nmanipulations, especially when combined with ensemble techniques like majority\nvoting and ridge regression, which further improve the system's defense against\nsophisticated adversarial inputs. Additionally, employing large language models\nsuch as GPT-4 with varied prompting techniques has shown promise in recognizing\nand scoring gaming strategies effectively. The findings underscore the\nimportance of continuous improvements in AI-driven educational tools to ensure\ntheir reliability and fairness in high-stakes settings.\n", "date": "2025-04-30", "categories": []}
{"id": "2505.00487", "title": "Analysis of the vulnerability of machine learning regression models to\n  adversarial attacks using data from 5G wireless networks", "abstract": "  This article describes the process of creating a script and conducting an\nanalytical study of a dataset using the DeepMIMO emulator. An advertorial\nattack was carried out using the FGSM method to maximize the gradient. A\ncomparison is made of the effectiveness of binary classifiers in the task of\ndetecting distorted data. The dynamics of changes in the quality indicators of\nthe regression model were analyzed in conditions without adversarial attacks,\nduring an adversarial attack and when the distorted data was isolated. It is\nshown that an adversarial FGSM attack with gradient maximization leads to an\nincrease in the value of the MSE metric by 33% and a decrease in the R2\nindicator by 10% on average. The LightGBM binary classifier effectively\nidentifies data with adversarial anomalies with 98% accuracy. Regression\nmachine learning models are susceptible to adversarial attacks, but rapid\nanalysis of network traffic and data transmitted over the network makes it\npossible to identify malicious activity\n", "date": "2025-05-01", "categories": []}
{"id": "2412.05538", "title": "Not Just Text: Uncovering Vision Modality Typographic Threats in Image\n  Generation Models", "abstract": "  Current image generation models can effortlessly produce high-quality, highly\nrealistic images, but this also increases the risk of misuse. In various\nText-to-Image or Image-to-Image tasks, attackers can generate a series of\nimages containing inappropriate content by simply editing the language modality\ninput. To mitigate this security concern, numerous guarding or defensive\nstrategies have been proposed, with a particular emphasis on safeguarding\nlanguage modality. However, in practical applications, threats in the vision\nmodality, particularly in tasks involving the editing of real-world images,\npresent heightened security risks as they can easily infringe upon the rights\nof the image owner. Therefore, this paper employs a method named typographic\nattack to reveal that various image generation models are also susceptible to\nthreats within the vision modality. Furthermore, we also evaluate the defense\nperformance of various existing methods when facing threats in the vision\nmodality and uncover their ineffectiveness. Finally, we propose the Vision\nModal Threats in Image Generation Models (VMT-IGMs) dataset, which would serve\nas a baseline for evaluating the vision modality vulnerability of various image\ngeneration models.\n", "date": "2024-12-07", "categories": []}
{"id": "2502.12734", "title": "Iron Sharpens Iron: Defending Against Attacks in Machine-Generated Text\n  Detection with Adversarial Training", "abstract": "  Machine-generated Text (MGT) detection is crucial for regulating and\nattributing online texts. While the existing MGT detectors achieve strong\nperformance, they remain vulnerable to simple perturbations and adversarial\nattacks. To build an effective defense against malicious perturbations, we view\nMGT detection from a threat modeling perspective, that is, analyzing the\nmodel's vulnerability from an adversary's point of view and exploring effective\nmitigations. To this end, we introduce an adversarial framework for training a\nrobust MGT detector, named GREedy Adversary PromoTed DefendER (GREATER). The\nGREATER consists of two key components: an adversary GREATER-A and a detector\nGREATER-D. The GREATER-D learns to defend against the adversarial attack from\nGREATER-A and generalizes the defense to other attacks. GREATER-A identifies\nand perturbs the critical tokens in embedding space, along with greedy search\nand pruning to generate stealthy and disruptive adversarial examples. Besides,\nwe update the GREATER-A and GREATER-D synchronously, encouraging the GREATER-D\nto generalize its defense to different attacks and varying attack intensities.\nOur experimental results across 10 text perturbation strategies and 6\nadversarial attacks show that our GREATER-D reduces the Attack Success Rate\n(ASR) by 0.67% compared with SOTA defense methods while our GREATER-A is\ndemonstrated to be more effective and efficient than SOTA attack approaches.\nCodes and dataset are available in https://github.com/Liyuuuu111/GREATER.\n", "date": "2025-02-18", "categories": []}
{"id": "2504.20869", "title": "Quantifying the Noise of Structural Perturbations on Graph Adversarial\n  Attacks", "abstract": "  Graph neural networks have been widely utilized to solve graph-related tasks\nbecause of their strong learning power in utilizing the local information of\nneighbors. However, recent studies on graph adversarial attacks have proven\nthat current graph neural networks are not robust against malicious attacks.\nYet much of the existing work has focused on the optimization objective based\non attack performance to obtain (near) optimal perturbations, but paid less\nattention to the strength quantification of each perturbation such as the\ninjection of a particular node/link, which makes the choice of perturbations a\nblack-box model that lacks interpretability. In this work, we propose the\nconcept of noise to quantify the attack strength of each adversarial link.\nFurthermore, we propose three attack strategies based on the defined noise and\nclassification margins in terms of single and multiple steps optimization.\nExtensive experiments conducted on benchmark datasets against three\nrepresentative graph neural networks demonstrate the effectiveness of the\nproposed attack strategies. Particularly, we also investigate the preferred\npatterns of effective adversarial perturbations by analyzing the corresponding\nproperties of the selected perturbation nodes.\n", "date": "2025-04-29", "categories": []}
{"id": "2503.00063", "title": "NoPain: No-box Point Cloud Attack via Optimal Transport Singular\n  Boundary", "abstract": "  Adversarial attacks exploit the vulnerability of deep models against\nadversarial samples. Existing point cloud attackers are tailored to specific\nmodels, iteratively optimizing perturbations based on gradients in either a\nwhite-box or black-box setting. Despite their promising attack performance,\nthey often struggle to produce transferable adversarial samples due to\noverfitting the specific parameters of surrogate models. To overcome this\nissue, we shift our focus to the data distribution itself and introduce a novel\napproach named NoPain, which employs optimal transport (OT) to identify the\ninherent singular boundaries of the data manifold for cross-network point cloud\nattacks. Specifically, we first calculate the OT mapping from noise to the\ntarget feature space, then identify singular boundaries by locating\nnon-differentiable positions. Finally, we sample along singular boundaries to\ngenerate adversarial point clouds. Once the singular boundaries are determined,\nNoPain can efficiently produce adversarial samples without the need of\niterative updates or guidance from the surrogate classifiers. Extensive\nexperiments demonstrate that the proposed end-to-end method outperforms\nbaseline approaches in terms of both transferability and efficiency, while also\nmaintaining notable advantages even against defense strategies. Code and model\nare available at https://github.com/cognaclee/nopain\n", "date": "2025-02-27", "categories": []}
{"id": "2504.20077", "title": "Edge-Based Learning for Improved Classification Under Adversarial Noise", "abstract": "  Adversarial noise introduces small perturbations in images, misleading deep\nlearning models into misclassification and significantly impacting recognition\naccuracy. In this study, we analyzed the effects of Fast Gradient Sign Method\n(FGSM) adversarial noise on image classification and investigated whether\ntraining on specific image features can improve robustness. We hypothesize that\nwhile adversarial noise perturbs various regions of an image, edges may remain\nrelatively stable and provide essential structural information for\nclassification. To test this, we conducted a series of experiments using brain\ntumor and COVID datasets. Initially, we trained the models on clean images and\nthen introduced subtle adversarial perturbations, which caused deep learning\nmodels to significantly misclassify the images. Retraining on a combination of\nclean and noisy images led to improved performance. To evaluate the robustness\nof the edge features, we extracted edges from the original/clean images and\ntrained the models exclusively on edge-based representations. When noise was\nintroduced to the images, the edge-based models demonstrated greater resilience\nto adversarial attacks compared to those trained on the original or clean\nimages. These results suggest that while adversarial noise is able to exploit\ncomplex non-edge regions significantly more than edges, the improvement in the\naccuracy after retraining is marginally more in the original data as compared\nto the edges. Thus, leveraging edge-based learning can improve the resilience\nof deep learning models against adversarial perturbations.\n", "date": "2025-04-25", "categories": []}
{"id": "2504.20111", "title": "Forging and Removing Latent-Noise Diffusion Watermarks Using a Single\n  Image", "abstract": "  Watermarking techniques are vital for protecting intellectual property and\npreventing fraudulent use of media. Most previous watermarking schemes designed\nfor diffusion models embed a secret key in the initial noise. The resulting\npattern is often considered hard to remove and forge into unrelated images. In\nthis paper, we propose a black-box adversarial attack without presuming access\nto the diffusion model weights. Our attack uses only a single watermarked\nexample and is based on a simple observation: there is a many-to-one mapping\nbetween images and initial noises. There are regions in the clean image latent\nspace pertaining to each watermark that get mapped to the same initial noise\nwhen inverted. Based on this intuition, we propose an adversarial attack to\nforge the watermark by introducing perturbations to the images such that we can\nenter the region of watermarked images. We show that we can also apply a\nsimilar approach for watermark removal by learning perturbations to exit this\nregion. We report results on multiple watermarking schemes (Tree-Ring, RingID,\nWIND, and Gaussian Shading) across two diffusion models (SDv1.4 and SDv2.0).\nOur results demonstrate the effectiveness of the attack and expose\nvulnerabilities in the watermarking methods, motivating future research on\nimproving them.\n", "date": "2025-04-27", "categories": []}
{"id": "2504.20295", "title": "The Dark Side of Digital Twins: Adversarial Attacks on AI-Driven Water\n  Forecasting", "abstract": "  Digital twins (DTs) are improving water distribution systems by using\nreal-time data, analytics, and prediction models to optimize operations. This\npaper presents a DT platform designed for a Spanish water supply network,\nutilizing Long Short-Term Memory (LSTM) networks to predict water consumption.\nHowever, machine learning models are vulnerable to adversarial attacks, such as\nthe Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD).\nThese attacks manipulate critical model parameters, injecting subtle\ndistortions that degrade forecasting accuracy. To further exploit these\nvulnerabilities, we introduce a Learning Automata (LA) and Random LA-based\napproach that dynamically adjusts perturbations, making adversarial attacks\nmore difficult to detect. Experimental results show that this approach\nsignificantly impacts prediction reliability, causing the Mean Absolute\nPercentage Error (MAPE) to rise from 26% to over 35%. Moreover, adaptive attack\nstrategies amplify this effect, highlighting cybersecurity risks in AI-driven\nDTs. These findings emphasize the urgent need for robust defenses, including\nadversarial training, anomaly detection, and secure data pipelines.\n", "date": "2025-04-28", "categories": []}
{"id": "2504.20829", "title": "GaussTrap: Stealthy Poisoning Attacks on 3D Gaussian Splatting for\n  Targeted Scene Confusion", "abstract": "  As 3D Gaussian Splatting (3DGS) emerges as a breakthrough in scene\nrepresentation and novel view synthesis, its rapid adoption in safety-critical\ndomains (e.g., autonomous systems, AR/VR) urgently demands scrutiny of\npotential security vulnerabilities. This paper presents the first systematic\nstudy of backdoor threats in 3DGS pipelines. We identify that adversaries may\nimplant backdoor views to induce malicious scene confusion during inference,\npotentially leading to environmental misperception in autonomous navigation or\nspatial distortion in immersive environments. To uncover this risk, we propose\nGuassTrap, a novel poisoning attack method targeting 3DGS models. GuassTrap\ninjects malicious views at specific attack viewpoints while preserving\nhigh-quality rendering in non-target views, ensuring minimal detectability and\nmaximizing potential harm. Specifically, the proposed method consists of a\nthree-stage pipeline (attack, stabilization, and normal training) to implant\nstealthy, viewpoint-consistent poisoned renderings in 3DGS, jointly optimizing\nattack efficacy and perceptual realism to expose security risks in 3D\nrendering. Extensive experiments on both synthetic and real-world datasets\ndemonstrate that GuassTrap can effectively embed imperceptible yet harmful\nbackdoor views while maintaining high-quality rendering in normal views,\nvalidating its robustness, adaptability, and practical applicability.\n", "date": "2025-04-29", "categories": []}
{"id": "2504.20848", "title": "Mitigating the Structural Bias in Graph Adversarial Defenses", "abstract": "  In recent years, graph neural networks (GNNs) have shown great potential in\naddressing various graph structure-related downstream tasks. However, recent\nstudies have found that current GNNs are susceptible to malicious adversarial\nattacks. Given the inevitable presence of adversarial attacks in the real\nworld, a variety of defense methods have been proposed to counter these attacks\nand enhance the robustness of GNNs. Despite the commendable performance of\nthese defense methods, we have observed that they tend to exhibit a structural\nbias in terms of their defense capability on nodes with low degree (i.e., tail\nnodes), which is similar to the structural bias of traditional GNNs on nodes\nwith low degree in the clean graph. Therefore, in this work, we propose a\ndefense strategy by including hetero-homo augmented graph construction, $k$NN\naugmented graph construction, and multi-view node-wise attention modules to\nmitigate the structural bias of GNNs against adversarial attacks. Notably, the\nhetero-homo augmented graph consists of removing heterophilic links (i.e.,\nlinks connecting nodes with dissimilar features) globally and adding homophilic\nlinks (i.e., links connecting nodes with similar features) for nodes with low\ndegree. To further enhance the defense capability, an attention mechanism is\nadopted to adaptively combine the representations from the above two kinds of\ngraph views. We conduct extensive experiments to demonstrate the defense and\ndebiasing effect of the proposed strategy on benchmark datasets.\n", "date": "2025-04-29", "categories": []}
{"id": "2502.07011", "title": "DROP: Poison Dilution via Knowledge Distillation for Federated Learning", "abstract": "  Federated Learning is vulnerable to adversarial manipulation, where malicious\nclients can inject poisoned updates to influence the global model's behavior.\nWhile existing defense mechanisms have made notable progress, they fail to\nprotect against adversaries that aim to induce targeted backdoors under\ndifferent learning and attack configurations. To address this limitation, we\nintroduce DROP (Distillation-based Reduction Of Poisoning), a novel defense\nmechanism that combines clustering and activity-tracking techniques with\nextraction of benign behavior from clients via knowledge distillation to tackle\nstealthy adversaries that manipulate low data poisoning rates and diverse\nmalicious client ratios within the federation. Through extensive\nexperimentation, our approach demonstrates superior robustness compared to\nexisting defenses across a wide range of learning configurations. Finally, we\nevaluate existing defenses and our method under the challenging setting of\nnon-IID client data distribution and highlight the challenges of designing a\nresilient FL defense in this setting.\n", "date": "2025-02-10", "categories": []}
{"id": "2504.11831", "title": "Support is All You Need for Certified VAE Training", "abstract": "  Variational Autoencoders (VAEs) have become increasingly popular and deployed\nin safety-critical applications. In such applications, we want to give\ncertified probabilistic guarantees on performance under adversarial attacks. We\npropose a novel method, CIVET, for certified training of VAEs. CIVET depends on\nthe key insight that we can bound worst-case VAE error by bounding the error on\ncarefully chosen support sets at the latent layer. We show this point\nmathematically and present a novel training algorithm utilizing this insight.\nWe show in an extensive evaluation across different datasets (in both the\nwireless and vision application areas), architectures, and perturbation\nmagnitudes that our method outperforms SOTA methods achieving good standard\nperformance with strong robustness guarantees.\n", "date": "2025-04-16", "categories": []}
{"id": "2504.19000", "title": "Unveiling and Mitigating Adversarial Vulnerabilities in Iterative\n  Optimizers", "abstract": "  Machine learning (ML) models are often sensitive to carefully crafted yet\nseemingly unnoticeable perturbations. Such adversarial examples are considered\nto be a property of ML models, often associated with their black-box operation\nand sensitivity to features learned from data. This work examines the\nadversarial sensitivity of non-learned decision rules, and particularly of\niterative optimizers. Our analysis is inspired by the recent developments in\ndeep unfolding, which cast such optimizers as ML models. We show that\nnon-learned iterative optimizers share the sensitivity to adversarial examples\nof ML models, and that attacking iterative optimizers effectively alters the\noptimization objective surface in a manner that modifies the minima sought. We\nthen leverage the ability to cast iteration-limited optimizers as ML models to\nenhance robustness via adversarial training. For a class of proximal gradient\noptimizers, we rigorously prove how their learning affects adversarial\nsensitivity. We numerically back our findings, showing the vulnerability of\nvarious optimizers, as well as the robustness induced by unfolding and\nadversarial training.\n", "date": "2025-04-26", "categories": []}
{"id": "2504.19730", "title": "Evaluate-and-Purify: Fortifying Code Language Models Against Adversarial\n  Attacks Using LLM-as-a-Judge", "abstract": "  The widespread adoption of code language models in software engineering tasks\nhas exposed vulnerabilities to adversarial attacks, especially the identifier\nsubstitution attacks. Although existing identifier substitution attackers\ndemonstrate high success rates, they often produce adversarial examples with\nunnatural code patterns. In this paper, we systematically assess the quality of\nadversarial examples using LLM-as-a-Judge. Our analysis reveals that over 80%\nof adversarial examples generated by state-of-the-art identifier substitution\nattackers (e.g., ALERT) are actually detectable. Based on this insight, we\npropose EP-Shield, a unified framework for evaluating and purifying identifier\nsubstitution attacks via naturalness-aware reasoning. Specifically, we first\nevaluate the naturalness of code and identify the perturbed adversarial code,\nthen purify it so that the victim model can restore correct prediction.\nExtensive experiments demonstrate the superiority of EP-Shield over adversarial\nfine-tuning (up to 83.36% improvement) and its lightweight design 7B\nparameters) with GPT-4-level performance.\n", "date": "2025-04-28", "categories": []}
{"id": "2406.10719", "title": "Trading Devil: Robust backdoor attack via Stochastic investment models\n  and Bayesian approach", "abstract": "  With the growing use of voice-activated systems and speech recognition\ntechnologies, the danger of backdoor attacks on audio data has grown\nsignificantly. This research looks at a specific type of attack, known as a\nStochastic investment-based backdoor attack (MarketBack), in which adversaries\nstrategically manipulate the stylistic properties of audio to fool speech\nrecognition systems. The security and integrity of machine learning models are\nseriously threatened by backdoor attacks, in order to maintain the reliability\nof audio applications and systems, the identification of such attacks becomes\ncrucial in the context of audio data. Experimental results demonstrated that\nMarketBack is feasible to achieve an average attack success rate close to 100%\nin seven victim models when poisoning less than 1% of the training data.\n", "date": "2024-06-15", "categories": []}
{"id": "2407.14573", "title": "Trading Devil Final: Backdoor attack via Stock market and Bayesian\n  Optimization", "abstract": "  Since the advent of generative artificial intelligence, every company and\nresearcher has been rushing to develop their own generative models, whether\ncommercial or not. Given the large number of users of these powerful new tools,\nthere is currently no intrinsically verifiable way to explain from the ground\nup what happens when LLMs (large language models) learn. For example, those\nbased on automatic speech recognition systems, which have to rely on huge and\nastronomical amounts of data collected from all over the web to produce fast\nand efficient results, In this article, we develop a backdoor attack called\nMarketBackFinal 2.0, based on acoustic data poisoning, MarketBackFinal 2.0 is\nmainly based on modern stock market models. In order to show the possible\nvulnerabilities of speech-based transformers that may rely on LLMs.\n", "date": "2024-07-21", "categories": []}
{"id": "2504.17829", "title": "Fine-Tuning Adversarially-Robust Transformers for Single-Image Dehazing", "abstract": "  Single-image dehazing is an important topic in remote sensing applications,\nenhancing the quality of acquired images and increasing object detection\nprecision. However, the reliability of such structures has not been\nsufficiently analyzed, which poses them to the risk of imperceptible\nperturbations that can significantly hinder their performance. In this work, we\nshow that state-of-the-art image-to-image dehazing transformers are susceptible\nto adversarial noise, with even 1 pixel change being able to decrease the PSNR\nby as much as 2.8 dB. Next, we propose two lightweight fine-tuning strategies\naimed at increasing the robustness of pre-trained transformers. Our methods\nresults in comparable clean performance, while significantly increasing the\nprotection against adversarial data. We further present their applicability in\ntwo remote sensing scenarios, showcasing their robust behavior for\nout-of-distribution data. The source code for adversarial fine-tuning and\nattack algorithms can be found at github.com/Vladimirescu/RobustDehazing.\n", "date": "2025-04-24", "categories": []}
{"id": "2504.18333", "title": "Adversarial Attacks on LLM-as-a-Judge Systems: Insights from Prompt\n  Injections", "abstract": "  LLM as judge systems used to assess text quality code correctness and\nargument strength are vulnerable to prompt injection attacks. We introduce a\nframework that separates content author attacks from system prompt attacks and\nevaluate five models Gemma 3.27B Gemma 3.4B Llama 3.2 3B GPT 4 and Claude 3\nOpus on four tasks with various defenses using fifty prompts per condition.\nAttacks achieved up to seventy three point eight percent success smaller models\nproved more vulnerable and transferability ranged from fifty point five to\nsixty two point six percent. Our results contrast with Universal Prompt\nInjection and AdvPrompter We recommend multi model committees and comparative\nscoring and release all code and datasets\n", "date": "2025-04-25", "categories": []}
{"id": "2406.04724", "title": "On Minimizing Adversarial Counterfactual Error in Adversarial RL", "abstract": "  Deep Reinforcement Learning (DRL) policies are highly susceptible to\nadversarial noise in observations, which poses significant risks in\nsafety-critical scenarios. The challenge inherent to adversarial perturbations\nis that by altering the information observed by the agent, the state becomes\nonly partially observable. Existing approaches address this by either enforcing\nconsistent actions across nearby states or maximizing the worst-case value\nwithin adversarially perturbed observations. However, the former suffers from\nperformance degradation when attacks succeed, while the latter tends to be\noverly conservative, leading to suboptimal performance in benign settings. We\nhypothesize that these limitations stem from their failing to account for\npartial observability directly. To this end, we introduce a novel objective\ncalled Adversarial Counterfactual Error (ACoE), defined on the beliefs about\nthe true state and balancing value optimization with robustness. To make ACoE\nscalable in model-free settings, we propose the theoretically-grounded\nsurrogate objective Cumulative-ACoE (C-ACoE). Our empirical evaluations on\nstandard benchmarks (MuJoCo, Atari, and Highway) demonstrate that our method\nsignificantly outperforms current state-of-the-art approaches for addressing\nadversarial RL challenges, offering a promising direction for improving\nrobustness in DRL under adversarial conditions. Our code is available at\nhttps://github.com/romanbelaire/acoe-robust-rl.\n", "date": "2024-06-07", "categories": []}
{"id": "2408.15833", "title": "Network transferability of adversarial patches in real-time object\n  detection", "abstract": "  Adversarial patches in computer vision can be used, to fool deep neural\nnetworks and manipulate their decision-making process. One of the most\nprominent examples of adversarial patches are evasion attacks for object\ndetectors. By covering parts of objects of interest, these patches suppress the\ndetections and thus make the target object 'invisible' to the object detector.\nSince these patches are usually optimized on a specific network with a specific\ntrain dataset, the transferability across multiple networks and datasets is not\ngiven. This paper addresses these issues and investigates the transferability\nacross numerous object detector architectures. Our extensive evaluation across\nvarious models on two distinct datasets indicates that patches optimized with\nlarger models provide better network transferability than patches that are\noptimized with smaller models.\n", "date": "2024-08-28", "categories": []}
{"id": "2412.01527", "title": "Traversing the Subspace of Adversarial Patches", "abstract": "  Despite ongoing research on the topic of adversarial examples in deep\nlearning for computer vision, some fundamentals of the nature of these attacks\nremain unclear. As the manifold hypothesis posits, high-dimensional data tends\nto be part of a low-dimensional manifold. To verify the thesis with adversarial\npatches, this paper provides an analysis of a set of adversarial patches and\ninvestigates the reconstruction abilities of three different dimensionality\nreduction methods. Quantitatively, the performance of reconstructed patches in\nan attack setting is measured and the impact of sampled patches from the latent\nspace during adversarial training is investigated. The evaluation is performed\non two publicly available datasets for person detection. The results indicate\nthat more sophisticated dimensionality reduction methods offer no advantages\nover a simple principal component analysis.\n", "date": "2024-12-02", "categories": []}
{"id": "2504.13192", "title": "CheatAgent: Attacking LLM-Empowered Recommender Systems via LLM Agent", "abstract": "  Recently, Large Language Model (LLM)-empowered recommender systems (RecSys)\nhave brought significant advances in personalized user experience and have\nattracted considerable attention. Despite the impressive progress, the research\nquestion regarding the safety vulnerability of LLM-empowered RecSys still\nremains largely under-investigated. Given the security and privacy concerns, it\nis more practical to focus on attacking the black-box RecSys, where attackers\ncan only observe the system's inputs and outputs. However, traditional attack\napproaches employing reinforcement learning (RL) agents are not effective for\nattacking LLM-empowered RecSys due to the limited capabilities in processing\ncomplex textual inputs, planning, and reasoning. On the other hand, LLMs\nprovide unprecedented opportunities to serve as attack agents to attack RecSys\nbecause of their impressive capability in simulating human-like decision-making\nprocesses. Therefore, in this paper, we propose a novel attack framework called\nCheatAgent by harnessing the human-like capabilities of LLMs, where an\nLLM-based agent is developed to attack LLM-Empowered RecSys. Specifically, our\nmethod first identifies the insertion position for maximum impact with minimal\ninput modification. After that, the LLM agent is designed to generate\nadversarial perturbations to insert at target positions. To further improve the\nquality of generated perturbations, we utilize the prompt tuning technique to\nimprove attacking strategies via feedback from the victim RecSys iteratively.\nExtensive experiments across three real-world datasets demonstrate the\neffectiveness of our proposed attacking method.\n", "date": "2025-04-13", "categories": []}
{"id": "2405.20770", "title": "Large Language Model Sentinel: LLM Agent for Adversarial Purification", "abstract": "  Over the past two years, the use of large language models (LLMs) has advanced\nrapidly. While these LLMs offer considerable convenience, they also raise\nsecurity concerns, as LLMs are vulnerable to adversarial attacks by some\nwell-designed textual perturbations. In this paper, we introduce a novel\ndefense technique named Large LAnguage MOdel Sentinel (LLAMOS), which is\ndesigned to enhance the adversarial robustness of LLMs by purifying the\nadversarial textual examples before feeding them into the target LLM. Our\nmethod comprises two main components: a) Agent instruction, which can simulate\na new agent for adversarial defense, altering minimal characters to maintain\nthe original meaning of the sentence while defending against attacks; b)\nDefense guidance, which provides strategies for modifying clean or adversarial\nexamples to ensure effective defense and accurate outputs from the target LLMs.\nRemarkably, the defense agent demonstrates robust defensive capabilities even\nwithout learning from adversarial examples. Additionally, we conduct an\nintriguing adversarial experiment where we develop two agents, one for defense\nand one for attack, and engage them in mutual confrontation. During the\nadversarial interactions, neither agent completely beat the other. Extensive\nexperiments on both open-source and closed-source LLMs demonstrate that our\nmethod effectively defends against adversarial attacks, thereby enhancing\nadversarial robustness.\n", "date": "2024-05-24", "categories": []}
{"id": "2503.06276", "title": "Exploring Adversarial Transferability between Kolmogorov-arnold Networks", "abstract": "  Kolmogorov-Arnold Networks (KANs) have emerged as a transformative model\nparadigm, significantly impacting various fields. However, their adversarial\nrobustness remains less underexplored, especially across different KAN\narchitectures. To explore this critical safety issue, we conduct an analysis\nand find that due to overfitting to the specific basis functions of KANs, they\npossess poor adversarial transferability among different KANs. To tackle this\nchallenge, we propose AdvKAN, the first transfer attack method for KANs. AdvKAN\nintegrates two key components: 1) a Breakthrough-Defense Surrogate Model\n(BDSM), which employs a breakthrough-defense training strategy to mitigate\noverfitting to the specific structures of KANs. 2) a Global-Local Interaction\n(GLI) technique, which promotes sufficient interaction between adversarial\ngradients of hierarchical levels, further smoothing out loss surfaces of KANs.\nBoth of them work together to enhance the strength of transfer attack among\ndifferent KANs. Extensive experimental results on various KANs and datasets\ndemonstrate the effectiveness of AdvKAN, which possesses notably superior\nattack capabilities and deeply reveals the vulnerabilities of KANs. Code will\nbe released upon acceptance.\n", "date": "2025-03-08", "categories": []}
{"id": "2504.12129", "title": "Anti-Aesthetics: Protecting Facial Privacy against Customized\n  Text-to-Image Synthesis", "abstract": "  The rise of customized diffusion models has spurred a boom in personalized\nvisual content creation, but also poses risks of malicious misuse, severely\nthreatening personal privacy and copyright protection. Some studies show that\nthe aesthetic properties of images are highly positively correlated with human\nperception of image quality. Inspired by this, we approach the problem from a\nnovel and intriguing aesthetic perspective to degrade the generation quality of\nmaliciously customized models, thereby achieving better protection of facial\nidentity. Specifically, we propose a Hierarchical Anti-Aesthetic (HAA)\nframework to fully explore aesthetic cues, which consists of two key branches:\n1) Global Anti-Aesthetics: By establishing a global anti-aesthetic reward\nmechanism and a global anti-aesthetic loss, it can degrade the overall\naesthetics of the generated content; 2) Local Anti-Aesthetics: A local\nanti-aesthetic reward mechanism and a local anti-aesthetic loss are designed to\nguide adversarial perturbations to disrupt local facial identity. By seamlessly\nintegrating both branches, our HAA effectively achieves the goal of\nanti-aesthetics from a global to a local level during customized generation.\nExtensive experiments show that HAA outperforms existing SOTA methods largely\nin identity removal, providing a powerful tool for protecting facial privacy\nand copyright.\n", "date": "2025-04-16", "categories": []}
{"id": "2504.14921", "title": "Fast Adversarial Training with Weak-to-Strong Spatial-Temporal\n  Consistency in the Frequency Domain on Videos", "abstract": "  Adversarial Training (AT) has been shown to significantly enhance adversarial\nrobustness via a min-max optimization approach. However, its effectiveness in\nvideo recognition tasks is hampered by two main challenges. First, fast\nadversarial training for video models remains largely unexplored, which\nseverely impedes its practical applications. Specifically, most video\nadversarial training methods are computationally costly, with long training\ntimes and high expenses. Second, existing methods struggle with the trade-off\nbetween clean accuracy and adversarial robustness. To address these challenges,\nwe introduce Video Fast Adversarial Training with Weak-to-Strong consistency\n(VFAT-WS), the first fast adversarial training method for video data.\nSpecifically, VFAT-WS incorporates the following key designs: First, it\nintegrates a straightforward yet effective temporal frequency augmentation\n(TF-AUG), and its spatial-temporal enhanced form STF-AUG, along with a\nsingle-step PGD attack to boost training efficiency and robustness. Second, it\ndevises a weak-to-strong spatial-temporal consistency regularization, which\nseamlessly integrates the simpler TF-AUG and the more complex STF-AUG.\nLeveraging the consistency regularization, it steers the learning process from\nsimple to complex augmentations. Both of them work together to achieve a better\ntrade-off between clean accuracy and robustness. Extensive experiments on\nUCF-101 and HMDB-51 with both CNN and Transformer-based models demonstrate that\nVFAT-WS achieves great improvements in adversarial robustness and corruption\nrobustness, while accelerating training by nearly 490%.\n", "date": "2025-04-21", "categories": []}
{"id": "2504.15823", "title": "Human-Imperceptible Physical Adversarial Attack for NIR Face Recognition\n  Models", "abstract": "  Near-infrared (NIR) face recognition systems, which can operate effectively\nin low-light conditions or in the presence of makeup, exhibit vulnerabilities\nwhen subjected to physical adversarial attacks. To further demonstrate the\npotential risks in real-world applications, we design a novel, stealthy, and\npractical adversarial patch to attack NIR face recognition systems in a\nblack-box setting. We achieved this by utilizing human-imperceptible\ninfrared-absorbing ink to generate multiple patches with digitally optimized\nshapes and positions for infrared images. To address the optimization mismatch\nbetween digital and real-world NIR imaging, we develop a light reflection model\nfor human skin to minimize pixel-level discrepancies by simulating NIR light\nreflection.\n  Compared to state-of-the-art (SOTA) physical attacks on NIR face recognition\nsystems, the experimental results show that our method improves the attack\nsuccess rate in both digital and physical domains, particularly maintaining\neffectiveness across various face postures. Notably, the proposed approach\noutperforms SOTA methods, achieving an average attack success rate of 82.46% in\nthe physical domain across different models, compared to 64.18% for existing\nmethods. The artifact is available at\nhttps://anonymous.4open.science/r/Human-imperceptible-adversarial-patch-0703/.\n", "date": "2025-04-22", "categories": []}
{"id": "2504.15942", "title": "Adversarial Observations in Weather Forecasting", "abstract": "  AI-based systems, such as Google's GenCast, have recently redefined the state\nof the art in weather forecasting, offering more accurate and timely\npredictions of both everyday weather and extreme events. While these systems\nare on the verge of replacing traditional meteorological methods, they also\nintroduce new vulnerabilities into the forecasting process. In this paper, we\ninvestigate this threat and present a novel attack on autoregressive diffusion\nmodels, such as those used in GenCast, capable of manipulating weather\nforecasts and fabricating extreme events, including hurricanes, heat waves, and\nintense rainfall. The attack introduces subtle perturbations into weather\nobservations that are statistically indistinguishable from natural noise and\nchange less than 0.1% of the measurements - comparable to tampering with data\nfrom a single meteorological satellite. As modern forecasting integrates data\nfrom nearly a hundred satellites and many other sources operated by different\ncountries, our findings highlight a critical security risk with the potential\nto cause large-scale disruptions and undermine public trust in weather\nprediction.\n", "date": "2025-04-22", "categories": []}
{"id": "2402.05967", "title": "The last Dance : Robust backdoor attack via diffusion models and\n  bayesian approach", "abstract": "  Diffusion models are state-of-the-art deep learning generative models that\nare trained on the principle of learning forward and backward diffusion\nprocesses via the progressive addition of noise and denoising. In this paper,\nwe aim to fool audio-based DNN models, such as those from the Hugging Face\nframework, primarily those that focus on audio, in particular transformer-based\nartificial intelligence models, which are powerful machine learning models that\nsave time and achieve results faster and more efficiently. We demonstrate the\nfeasibility of backdoor attacks (called `BacKBayDiffMod`) on audio transformers\nderived from Hugging Face, a popular framework in the world of artificial\nintelligence research. The backdoor attack developed in this paper is based on\npoisoning model training data uniquely by incorporating backdoor diffusion\nsampling and a Bayesian approach to the distribution of poisoned data.\n", "date": "2024-02-05", "categories": []}
{"id": "2406.14393", "title": "Jailbreaking as a Reward Misspecification Problem", "abstract": "  The widespread adoption of large language models (LLMs) has raised concerns\nabout their safety and reliability, particularly regarding their vulnerability\nto adversarial attacks. In this paper, we propose a novel perspective that\nattributes this vulnerability to reward misspecification during the alignment\nprocess. This misspecification occurs when the reward function fails to\naccurately capture the intended behavior, leading to misaligned model outputs.\nWe introduce a metric ReGap to quantify the extent of reward misspecification\nand demonstrate its effectiveness and robustness in detecting harmful backdoor\nprompts. Building upon these insights, we present ReMiss, a system for\nautomated red teaming that generates adversarial prompts in a\nreward-misspecified space. ReMiss achieves state-of-the-art attack success\nrates on the AdvBench benchmark against various target aligned LLMs while\npreserving the human readability of the generated prompts. Furthermore, these\nattacks on open-source models demonstrate high transferability to closed-source\nmodels like GPT-4o and out-of-distribution tasks from HarmBench. Detailed\nanalysis highlights the unique advantages of the proposed reward\nmisspecification objective compared to previous methods, offering new insights\nfor improving LLM safety and robustness.\n", "date": "2024-06-20", "categories": []}
{"id": "2410.20197", "title": "Transferable Adversarial Attacks on SAM and Its Downstream Models", "abstract": "  The utilization of large foundational models has a dilemma: while fine-tuning\ndownstream tasks from them holds promise for making use of the well-generalized\nknowledge in practical applications, their open accessibility also poses\nthreats of adverse usage. This paper, for the first time, explores the\nfeasibility of adversarial attacking various downstream models fine-tuned from\nthe segment anything model (SAM), by solely utilizing the information from the\nopen-sourced SAM. In contrast to prevailing transfer-based adversarial attacks,\nwe demonstrate the existence of adversarial dangers even without accessing the\ndownstream task and dataset to train a similar surrogate model. To enhance the\neffectiveness of the adversarial attack towards models fine-tuned on unknown\ndatasets, we propose a universal meta-initialization (UMI) algorithm to extract\nthe intrinsic vulnerability inherent in the foundation model, which is then\nutilized as the prior knowledge to guide the generation of adversarial\nperturbations. Moreover, by formulating the gradient difference in the\nattacking process between the open-sourced SAM and its fine-tuned downstream\nmodels, we theoretically demonstrate that a deviation occurs in the adversarial\nupdate direction by directly maximizing the distance of encoded feature\nembeddings in the open-sourced SAM. Consequently, we propose a gradient robust\nloss that simulates the associated uncertainty with gradient-based noise\naugmentation to enhance the robustness of generated adversarial examples (AEs)\ntowards this deviation, thus improving the transferability. Extensive\nexperiments demonstrate the effectiveness of the proposed universal\nmeta-initialized and gradient robust adversarial attack (UMI-GRAT) toward SAMs\nand their downstream models. Code is available at\nhttps://github.com/xiasong0501/GRAT.\n", "date": "2024-10-26", "categories": []}
{"id": "2501.12761", "title": "Modality Unified Attack for Omni-Modality Person Re-Identification", "abstract": "  Deep learning based person re-identification (re-id) models have been widely\nemployed in surveillance systems. Recent studies have demonstrated that\nblack-box single-modality and cross-modality re-id models are vulnerable to\nadversarial examples (AEs), leaving the robustness of multi-modality re-id\nmodels unexplored. Due to the lack of knowledge about the specific type of\nmodel deployed in the target black-box surveillance system, we aim to generate\nmodality unified AEs for omni-modality (single-, cross- and multi-modality)\nre-id models. Specifically, we propose a novel Modality Unified Attack method\nto train modality-specific adversarial generators to generate AEs that\neffectively attack different omni-modality models. A multi-modality model is\nadopted as the surrogate model, wherein the features of each modality are\nperturbed by metric disruption loss before fusion. To collapse the common\nfeatures of omni-modality models, Cross Modality Simulated Disruption approach\nis introduced to mimic the cross-modality feature embeddings by intentionally\nfeeding images to non-corresponding modality-specific subnetworks of the\nsurrogate model. Moreover, Multi Modality Collaborative Disruption strategy is\ndevised to facilitate the attacker to comprehensively corrupt the informative\ncontent of person images by leveraging a multi modality feature collaborative\nmetric disruption loss. Extensive experiments show that our MUA method can\neffectively attack the omni-modality re-id models, achieving 55.9%, 24.4%,\n49.0% and 62.7% mean mAP Drop Rate, respectively.\n", "date": "2025-01-22", "categories": []}
{"id": "2504.13966", "title": "Adversarial Resilience against Clean-Label Attacks in Realizable and\n  Noisy Settings", "abstract": "  We investigate the challenge of establishing stochastic-like guarantees when\nsequentially learning from a stream of i.i.d. data that includes an unknown\nquantity of clean-label adversarial samples. We permit the learner to abstain\nfrom making predictions when uncertain. The regret of the learner is measured\nin terms of misclassification and abstention error, where we allow the learner\nto abstain for free on adversarial injected samples. This approach is based on\nthe work of Goel, Hanneke, Moran, and Shetty from arXiv:2306.13119. We explore\nthe methods they present and manage to correct inaccuracies in their\nargumentation.\n  However, this approach is limited to the realizable setting, where labels are\nassigned according to some function $f^*$ from the hypothesis space\n$\\mathcal{F}$. Based on similar arguments, we explore methods to make\nadaptations for the agnostic setting where labels are random. Introducing the\nnotion of a clean-label adversary in the agnostic context, we are the first to\ngive a theoretical analysis of a disagreement-based learner for thresholds,\nsubject to a clean-label adversary with noise.\n", "date": "2025-04-17", "categories": []}
{"id": "2504.14423", "title": "Adversarial Attack for RGB-Event based Visual Object Tracking", "abstract": "  Visual object tracking is a crucial research topic in the fields of computer\nvision and multi-modal fusion. Among various approaches, robust visual tracking\nthat combines RGB frames with Event streams has attracted increasing attention\nfrom researchers. While striving for high accuracy and efficiency in tracking,\nit is also important to explore how to effectively conduct adversarial attacks\nand defenses on RGB-Event stream tracking algorithms, yet research in this area\nremains relatively scarce. To bridge this gap, in this paper, we propose a\ncross-modal adversarial attack algorithm for RGB-Event visual tracking. Because\nof the diverse representations of Event streams, and given that Event voxels\nand frames are more commonly used, this paper will focus on these two\nrepresentations for an in-depth study. Specifically, for the RGB-Event voxel,\nwe first optimize the perturbation by adversarial loss to generate RGB frame\nadversarial examples. For discrete Event voxel representations, we propose a\ntwo-step attack strategy, more in detail, we first inject Event voxels into the\ntarget region as initialized adversarial examples, then, conduct a\ngradient-guided optimization by perturbing the spatial location of the Event\nvoxels. For the RGB-Event frame based tracking, we optimize the cross-modal\nuniversal perturbation by integrating the gradient information from multimodal\ndata. We evaluate the proposed approach against attacks on three widely used\nRGB-Event Tracking datasets, i.e., COESOT, FE108, and VisEvent. Extensive\nexperiments show that our method significantly reduces the performance of the\ntracker across numerous datasets in both unimodal and multimodal scenarios. The\nsource code will be released on\nhttps://github.com/Event-AHU/Adversarial_Attack_Defense\n", "date": "2025-04-19", "categories": []}
{"id": "2504.14541", "title": "Towards Model Resistant to Transferable Adversarial Examples via Trigger\n  Activation", "abstract": "  Adversarial examples, characterized by imperceptible perturbations, pose\nsignificant threats to deep neural networks by misleading their predictions. A\ncritical aspect of these examples is their transferability, allowing them to\ndeceive {unseen} models in black-box scenarios. Despite the widespread\nexploration of defense methods, including those on transferability, they show\nlimitations: inefficient deployment, ineffective defense, and degraded\nperformance on clean images. In this work, we introduce a novel training\nparadigm aimed at enhancing robustness against transferable adversarial\nexamples (TAEs) in a more efficient and effective way. We propose a model that\nexhibits random guessing behavior when presented with clean data\n$\\boldsymbol{x}$ as input, and generates accurate predictions when with\ntriggered data $\\boldsymbol{x}+\\boldsymbol{\\tau}$. Importantly, the trigger\n$\\boldsymbol{\\tau}$ remains constant for all data instances. We refer to these\nmodels as \\textbf{models with trigger activation}. We are surprised to find\nthat these models exhibit certain robustness against TAEs. Through the\nconsideration of first-order gradients, we provide a theoretical analysis of\nthis robustness. Moreover, through the joint optimization of the learnable\ntrigger and the model, we achieve improved robustness to transferable attacks.\nExtensive experiments conducted across diverse datasets, evaluating a variety\nof attacking methods, underscore the effectiveness and superiority of our\napproach.\n", "date": "2025-04-20", "categories": []}
{"id": "2412.10353", "title": "Robust image classification with multi-modal large language models", "abstract": "  Deep Neural Networks are vulnerable to adversarial examples, i.e., carefully\ncrafted input samples that can cause models to make incorrect predictions with\nhigh confidence. To mitigate these vulnerabilities, adversarial training and\ndetection-based defenses have been proposed to strengthen models in advance.\nHowever, most of these approaches focus on a single data modality, overlooking\nthe relationships between visual patterns and textual descriptions of the\ninput. In this paper, we propose a novel defense, MultiShield, designed to\ncombine and complement these defenses with multi-modal information to further\nenhance their robustness. MultiShield leverages multi-modal large language\nmodels to detect adversarial examples and abstain from uncertain\nclassifications when there is no alignment between textual and visual\nrepresentations of the input. Extensive evaluations on CIFAR-10 and ImageNet\ndatasets, using robust and non-robust image classification models, demonstrate\nthat MultiShield can be easily integrated to detect and reject adversarial\nexamples, outperforming the original defenses.\n", "date": "2024-12-13", "categories": []}
{"id": "2501.13776", "title": "Crossfire: An Elastic Defense Framework for Graph Neural Networks Under\n  Bit Flip Attacks", "abstract": "  Bit Flip Attacks (BFAs) are a well-established class of adversarial attacks,\noriginally developed for Convolutional Neural Networks within the computer\nvision domain. Most recently, these attacks have been extended to target Graph\nNeural Networks (GNNs), revealing significant vulnerabilities. This new\ndevelopment naturally raises questions about the best strategies to defend GNNs\nagainst BFAs, a challenge for which no solutions currently exist. Given the\napplications of GNNs in critical fields, any defense mechanism must not only\nmaintain network performance, but also verifiably restore the network to its\npre-attack state. Verifiably restoring the network to its pre-attack state also\neliminates the need for costly evaluations on test data to ensure network\nquality. We offer first insights into the effectiveness of existing honeypot-\nand hashing-based defenses against BFAs adapted from the computer vision domain\nto GNNs, and characterize the shortcomings of these approaches. To overcome\ntheir limitations, we propose Crossfire, a hybrid approach that exploits weight\nsparsity and combines hashing and honeypots with bit-level correction of\nout-of-distribution weight elements to restore network integrity. Crossfire is\nretraining-free and does not require labeled data. Averaged over 2,160\nexperiments on six benchmark datasets, Crossfire offers a 21.8% higher\nprobability than its competitors of reconstructing a GNN attacked by a BFA to\nits pre-attack state. These experiments cover up to 55 bit flips from various\nattacks. Moreover, it improves post-repair prediction quality by 10.85%.\nComputational and storage overheads are negligible compared to the inherent\ncomplexity of even the simplest GNNs.\n", "date": "2025-01-23", "categories": []}
{"id": "2504.13209", "title": "On the Feasibility of Using MultiModal LLMs to Execute AR Social\n  Engineering Attacks", "abstract": "  Augmented Reality (AR) and Multimodal Large Language Models (LLMs) are\nrapidly evolving, providing unprecedented capabilities for human-computer\ninteraction. However, their integration introduces a new attack surface for\nsocial engineering. In this paper, we systematically investigate the\nfeasibility of orchestrating AR-driven Social Engineering attacks using\nMultimodal LLM for the first time, via our proposed SEAR framework, which\noperates through three key phases: (1) AR-based social context synthesis, which\nfuses Multimodal inputs (visual, auditory and environmental cues); (2)\nrole-based Multimodal RAG (Retrieval-Augmented Generation), which dynamically\nretrieves and integrates contextual data while preserving character\ndifferentiation; and (3) ReInteract social engineering agents, which execute\nadaptive multiphase attack strategies through inference interaction loops. To\nverify SEAR, we conducted an IRB-approved study with 60 participants in three\nexperimental configurations (unassisted, AR+LLM, and full SEAR pipeline)\ncompiling a new dataset of 180 annotated conversations in simulated social\nscenarios. Our results show that SEAR is highly effective at eliciting\nhigh-risk behaviors (e.g., 93.3% of participants susceptible to email\nphishing). The framework was particularly effective in building trust, with 85%\nof targets willing to accept an attacker's call after an interaction. Also, we\nidentified notable limitations such as ``occasionally artificial'' due to\nperceived authenticity gaps. This work provides proof-of-concept for AR-LLM\ndriven social engineering attacks and insights for developing defensive\ncountermeasures against next-generation augmented reality threats.\n", "date": "2025-04-16", "categories": []}
{"id": "2504.13551", "title": "Q-FAKER: Query-free Hard Black-box Attack via Controlled Generation", "abstract": "  Many adversarial attack approaches are proposed to verify the vulnerability\nof language models. However, they require numerous queries and the information\non the target model. Even black-box attack methods also require the target\nmodel's output information. They are not applicable in real-world scenarios, as\nin hard black-box settings where the target model is closed and inaccessible.\nEven the recently proposed hard black-box attacks still require many queries\nand demand extremely high costs for training adversarial generators. To address\nthese challenges, we propose Q-faker (Query-free Hard Black-box Attacker), a\nnovel and efficient method that generates adversarial examples without\naccessing the target model. To avoid accessing the target model, we use a\nsurrogate model instead. The surrogate model generates adversarial sentences\nfor a target-agnostic attack. During this process, we leverage controlled\ngeneration techniques. We evaluate our proposed method on eight datasets.\nExperimental results demonstrate our method's effectiveness including high\ntransferability and the high quality of the generated adversarial examples, and\nprove its practical in hard black-box settings.\n", "date": "2025-04-18", "categories": []}
{"id": "2408.01934", "title": "A Survey and Evaluation of Adversarial Attacks for Object Detection", "abstract": "  Deep learning models achieve remarkable accuracy in computer vision tasks,\nyet remain vulnerable to adversarial examples--carefully crafted perturbations\nto input images that can deceive these models into making confident but\nincorrect predictions. This vulnerability pose significant risks in high-stakes\napplications such as autonomous vehicles, security surveillance, and\nsafety-critical inspection systems. While the existing literature extensively\ncovers adversarial attacks in image classification, comprehensive analyses of\nsuch attacks on object detection systems remain limited. This paper presents a\nnovel taxonomic framework for categorizing adversarial attacks specific to\nobject detection architectures, synthesizes existing robustness metrics, and\nprovides a comprehensive empirical evaluation of state-of-the-art attack\nmethodologies on popular object detection models, including both traditional\ndetectors and modern detectors with vision-language pretraining. Through\nrigorous analysis of open-source attack implementations and their effectiveness\nacross diverse detection architectures, we derive key insights into attack\ncharacteristics. Furthermore, we delineate critical research gaps and emerging\nchallenges to guide future investigations in securing object detection systems\nagainst adversarial threats. Our findings establish a foundation for developing\nmore robust detection models while highlighting the urgent need for\nstandardized evaluation protocols in this rapidly evolving domain.\n", "date": "2024-08-04", "categories": []}
{"id": "2402.09063", "title": "Soft Prompt Threats: Attacking Safety Alignment and Unlearning in\n  Open-Source LLMs through the Embedding Space", "abstract": "  Current research in adversarial robustness of LLMs focuses on discrete input\nmanipulations in the natural language space, which can be directly transferred\nto closed-source models. However, this approach neglects the steady progression\nof open-source models. As open-source models advance in capability, ensuring\ntheir safety also becomes increasingly imperative. Yet, attacks tailored to\nopen-source LLMs that exploit full model access remain largely unexplored. We\naddress this research gap and propose the embedding space attack, which\ndirectly attacks the continuous embedding representation of input tokens. We\nfind that embedding space attacks circumvent model alignments and trigger\nharmful behaviors more efficiently than discrete attacks or model fine-tuning.\nFurthermore, we present a novel threat model in the context of unlearning and\nshow that embedding space attacks can extract supposedly deleted information\nfrom unlearned LLMs across multiple datasets and models. Our findings highlight\nembedding space attacks as an important threat model in open-source LLMs.\nTrigger Warning: the appendix contains LLM-generated text with violence and\nharassment.\n", "date": "2024-02-14", "categories": []}
{"id": "2409.04208", "title": "Learning to Learn Transferable Generative Attack for Person\n  Re-Identification", "abstract": "  Deep learning-based person re-identification (re-id) models are widely\nemployed in surveillance systems and inevitably inherit the vulnerability of\ndeep networks to adversarial attacks. Existing attacks merely consider\ncross-dataset and cross-model transferability, ignoring the cross-test\ncapability to perturb models trained in different domains. To powerfully\nexamine the robustness of real-world re-id models, the Meta Transferable\nGenerative Attack (MTGA) method is proposed, which adopts meta-learning\noptimization to promote the generative attacker producing highly transferable\nadversarial examples by learning comprehensively simulated transfer-based\ncross-model\\&dataset\\&test black-box meta attack tasks. Specifically,\ncross-model\\&dataset black-box attack tasks are first mimicked by selecting\ndifferent re-id models and datasets for meta-train and meta-test attack\nprocesses. As different models may focus on different feature regions, the\nPerturbation Random Erasing module is further devised to prevent the attacker\nfrom learning to only corrupt model-specific features. To boost the attacker\nlearning to possess cross-test transferability, the Normalization Mix strategy\nis introduced to imitate diverse feature embedding spaces by mixing\nmulti-domain statistics of target models. Extensive experiments show the\nsuperiority of MTGA, especially in cross-model\\&dataset and\ncross-model\\&dataset\\&test attacks, our MTGA outperforms the SOTA methods by\n21.5\\% and 11.3\\% on mean mAP drop rate, respectively. The code of MTGA will be\nreleased after the paper is accepted.\n", "date": "2024-09-06", "categories": []}
{"id": "2501.10906", "title": "Explainable Adversarial Attacks on Coarse-to-Fine Classifiers", "abstract": "  Traditional adversarial attacks typically aim to alter the predicted labels\nof input images by generating perturbations that are imperceptible to the human\neye. However, these approaches often lack explainability. Moreover, most\nexisting work on adversarial attacks focuses on single-stage classifiers, but\nmulti-stage classifiers are largely unexplored. In this paper, we introduce\ninstance-based adversarial attacks for multi-stage classifiers, leveraging\nLayer-wise Relevance Propagation (LRP), which assigns relevance scores to\npixels based on their influence on classification outcomes. Our approach\ngenerates explainable adversarial perturbations by utilizing LRP to identify\nand target key features critical for both coarse and fine-grained\nclassifications. Unlike conventional attacks, our method not only induces\nmisclassification but also enhances the interpretability of the model's\nbehavior across classification stages, as demonstrated by experimental results.\n", "date": "2025-01-19", "categories": []}
{"id": "2503.12793", "title": "Improving Generalization of Universal Adversarial Perturbation via\n  Dynamic Maximin Optimization", "abstract": "  Deep neural networks (DNNs) are susceptible to universal adversarial\nperturbations (UAPs). These perturbations are meticulously designed to fool the\ntarget model universally across all sample classes. Unlike instance-specific\nadversarial examples (AEs), generating UAPs is more complex because they must\nbe generalized across a wide range of data samples and models. Our research\nreveals that existing universal attack methods, which optimize UAPs using DNNs\nwith static model parameter snapshots, do not fully leverage the potential of\nDNNs to generate more effective UAPs. Rather than optimizing UAPs against\nstatic DNN models with a fixed training set, we suggest using dynamic\nmodel-data pairs to generate UAPs. In particular, we introduce a dynamic\nmaximin optimization strategy, aiming to optimize the UAP across a variety of\noptimal model-data pairs. We term this approach DM-UAP. DM-UAP utilizes an\niterative max-min-min optimization framework that refines the model-data pairs,\ncoupled with a curriculum UAP learning algorithm to examine the combined space\nof model parameters and data thoroughly. Comprehensive experiments on the\nImageNet dataset demonstrate that the proposed DM-UAP markedly enhances both\ncross-sample universality and cross-model transferability of UAPs. Using only\n500 samples for UAP generation, DM-UAP outperforms the state-of-the-art\napproach with an average increase in fooling ratio of 12.108%.\n", "date": "2025-03-17", "categories": []}
{"id": "2504.11707", "title": "Towards Safe Synthetic Image Generation On the Web: A Multimodal Robust\n  NSFW Defense and Million Scale Dataset", "abstract": "  In the past years, we have witnessed the remarkable success of Text-to-Image\n(T2I) models and their widespread use on the web. Extensive research in making\nT2I models produce hyper-realistic images has led to new concerns, such as\ngenerating Not-Safe-For-Work (NSFW) web content and polluting the web society.\nTo help prevent misuse of T2I models and create a safer web environment for\nusers features like NSFW filters and post-hoc security checks are used in these\nmodels. However, recent work unveiled how these methods can easily fail to\nprevent misuse. In particular, adversarial attacks on text and image modalities\ncan easily outplay defensive measures. %Exploiting such leads to the growing\nconcern of preventing adversarial attacks on text and image modalities.\nMoreover, there is currently no robust multimodal NSFW dataset that includes\nboth prompt and image pairs and adversarial examples. This work proposes a\nmillion-scale prompt and image dataset generated using open-source diffusion\nmodels. Second, we develop a multimodal defense to distinguish safe and NSFW\ntext and images, which is robust against adversarial attacks and directly\nalleviates current challenges. Our extensive experiments show that our model\nperforms well against existing SOTA NSFW detection methods in terms of accuracy\nand recall, drastically reducing the Attack Success Rate (ASR) in multimodal\nadversarial attack scenarios. Code:\nhttps://github.com/shahidmuneer/multimodal-nsfw-defense.\n", "date": "2025-04-16", "categories": []}
{"id": "2504.11923", "title": "SemDiff: Generating Natural Unrestricted Adversarial Examples via\n  Semantic Attributes Optimization in Diffusion Models", "abstract": "  Unrestricted adversarial examples (UAEs), allow the attacker to create\nnon-constrained adversarial examples without given clean samples, posing a\nsevere threat to the safety of deep learning models. Recent works utilize\ndiffusion models to generate UAEs. However, these UAEs often lack naturalness\nand imperceptibility due to simply optimizing in intermediate latent noises. In\nlight of this, we propose SemDiff, a novel unrestricted adversarial attack that\nexplores the semantic latent space of diffusion models for meaningful\nattributes, and devises a multi-attributes optimization approach to ensure\nattack success while maintaining the naturalness and imperceptibility of\ngenerated UAEs. We perform extensive experiments on four tasks on three\nhigh-resolution datasets, including CelebA-HQ, AFHQ and ImageNet. The results\ndemonstrate that SemDiff outperforms state-of-the-art methods in terms of\nattack success rate and imperceptibility. The generated UAEs are natural and\nexhibit semantically meaningful changes, in accord with the attributes'\nweights. In addition, SemDiff is found capable of evading different defenses,\nwhich further validates its effectiveness and threatening.\n", "date": "2025-04-16", "categories": []}
{"id": "2504.12255", "title": "Human Aligned Compression for Robust Models", "abstract": "  Adversarial attacks on image models threaten system robustness by introducing\nimperceptible perturbations that cause incorrect predictions. We investigate\nhuman-aligned learned lossy compression as a defense mechanism, comparing two\nlearned models (HiFiC and ELIC) against traditional JPEG across various quality\nlevels. Our experiments on ImageNet subsets demonstrate that learned\ncompression methods outperform JPEG, particularly for Vision Transformer\narchitectures, by preserving semantically meaningful content while removing\nadversarial noise. Even in white-box settings where attackers can access the\ndefense, these methods maintain substantial effectiveness. We also show that\nsequential compression--applying rounds of\ncompression/decompression--significantly enhances defense efficacy while\nmaintaining classification performance. Our findings reveal that human-aligned\ncompression provides an effective, computationally efficient defense that\nprotects the image features most relevant to human and machine understanding.\nIt offers a practical approach to improving model robustness against\nadversarial threats.\n", "date": "2025-04-16", "categories": []}
{"id": "2409.08372", "title": "FedProphet: Memory-Efficient Federated Adversarial Training via Robust\n  and Consistent Cascade Learning", "abstract": "  Federated Adversarial Training (FAT) can supplement robustness against\nadversarial examples to Federated Learning (FL), promoting a meaningful step\ntoward trustworthy AI. However, FAT requires large models to preserve high\naccuracy while achieving strong robustness, incurring high memory-swapping\nlatency when training on memory-constrained edge devices. Existing\nmemory-efficient FL methods suffer from poor accuracy and weak robustness due\nto inconsistent local and global models. In this paper, we propose FedProphet,\na novel FAT framework that can achieve memory efficiency, robustness, and\nconsistency simultaneously. FedProphget reduces the memory requirement in local\ntraining while guaranteeing adversarial robustness by adversarial cascade\nlearning with strong convexity regularization, and we show that the strong\nrobustness also implies low inconsistency in FedProphet. We also develop a\ntraining coordinator on the server of FL, with Adaptive Perturbation Adjustment\nfor utility-robustness balance and Differentiated Module Assignment for\nobjective inconsistency mitigation. FedPeophet significantly outperforms other\nbaselines under different experimental settings, maintaining the accuracy and\nrobustness of end-to-end FAT with 80% memory reduction and up to 10.8x speedup\nin training time.\n", "date": "2024-09-12", "categories": []}
{"id": "2412.18370", "title": "Unveiling the Threat of Fraud Gangs to Graph Neural Networks:\n  Multi-Target Graph Injection Attacks Against GNN-Based Fraud Detectors", "abstract": "  Graph neural networks (GNNs) have emerged as an effective tool for fraud\ndetection, identifying fraudulent users, and uncovering malicious behaviors.\nHowever, attacks against GNN-based fraud detectors and their risks have rarely\nbeen studied, thereby leaving potential threats unaddressed. Recent findings\nsuggest that frauds are increasingly organized as gangs or groups. In this\nwork, we design attack scenarios where fraud gangs aim to make their fraud\nnodes misclassified as benign by camouflaging their illicit activities in\ncollusion. Based on these scenarios, we study adversarial attacks against\nGNN-based fraud detectors by simulating attacks of fraud gangs in three\nreal-world fraud cases: spam reviews, fake news, and medical insurance frauds.\nWe define these attacks as multi-target graph injection attacks and propose\nMonTi, a transformer-based Multi-target one-Time graph injection attack model.\nMonTi simultaneously generates attributes and edges of all attack nodes with a\ntransformer encoder, capturing interdependencies between attributes and edges\nmore effectively than most existing graph injection attack methods that\ngenerate these elements sequentially. Additionally, MonTi adaptively allocates\nthe degree budget for each attack node to explore diverse injection structures\ninvolving target, candidate, and attack nodes, unlike existing methods that fix\nthe degree budget across all attack nodes. Experiments show that MonTi\noutperforms the state-of-the-art graph injection attack methods on five\nreal-world graphs.\n", "date": "2024-12-24", "categories": []}
{"id": "2501.14122", "title": "Reinforcement Learning Platform for Adversarial Black-box Attacks with\n  Custom Distortion Filters", "abstract": "  We present a Reinforcement Learning Platform for Adversarial Black-box\nuntargeted and targeted attacks, RLAB, that allows users to select from various\ndistortion filters to create adversarial examples. The platform uses a\nReinforcement Learning agent to add minimum distortion to input images while\nstill causing misclassification by the target model. The agent uses a novel\ndual-action method to explore the input image at each step to identify\nsensitive regions for adding distortions while removing noises that have less\nimpact on the target model. This dual action leads to faster and more efficient\nconvergence of the attack. The platform can also be used to measure the\nrobustness of image classification models against specific distortion types.\nAlso, retraining the model with adversarial samples significantly improved\nrobustness when evaluated on benchmark datasets. The proposed platform\noutperforms state-of-the-art methods in terms of the average number of queries\nrequired to cause misclassification. This advances trustworthiness with a\npositive social impact.\n", "date": "2025-01-23", "categories": []}
{"id": "2504.10850", "title": "How to Enhance Downstream Adversarial Robustness (almost) without\n  Touching the Pre-Trained Foundation Model?", "abstract": "  With the rise of powerful foundation models, a pre-training-fine-tuning\nparadigm becomes increasingly popular these days: A foundation model is\npre-trained using a huge amount of data from various sources, and then the\ndownstream users only need to fine-tune and adapt it to specific downstream\ntasks. However, due to the high computation complexity of adversarial training,\nit is not feasible to fine-tune the foundation model to improve its robustness\non the downstream task. Observing the above challenge, we want to improve the\ndownstream robustness without updating/accessing the weights in the foundation\nmodel. Inspired from existing literature in robustness inheritance (Kim et al.,\n2020), through theoretical investigation, we identify a close relationship\nbetween robust contrastive learning with the adversarial robustness of\nsupervised learning. To further validate and utilize this theoretical insight,\nwe design a simple-yet-effective robust auto-encoder as a data pre-processing\nmethod before feeding the data into the foundation model. The proposed approach\nhas zero access to the foundation model when training the robust auto-encoder.\nExtensive experiments demonstrate the effectiveness of the proposed method in\nimproving the robustness of downstream tasks, verifying the connection between\nthe feature robustness (implied by small adversarial contrastive loss) and the\nrobustness of the downstream task.\n", "date": "2025-04-15", "categories": []}
{"id": "2504.11034", "title": "Defending Against Frequency-Based Attacks with Diffusion Models", "abstract": "  Adversarial training is a common strategy for enhancing model robustness\nagainst adversarial attacks. However, it is typically tailored to the specific\nattack types it is trained on, limiting its ability to generalize to unseen\nthreat models. Adversarial purification offers an alternative by leveraging a\ngenerative model to remove perturbations before classification. Since the\npurifier is trained independently of both the classifier and the threat models,\nit is better equipped to handle previously unseen attack scenarios. Diffusion\nmodels have proven highly effective for noise purification, not only in\ncountering pixel-wise adversarial perturbations but also in addressing\nnon-adversarial data shifts. In this study, we broaden the focus beyond\npixel-wise robustness to explore the extent to which purification can mitigate\nboth spectral and spatial adversarial attacks. Our findings highlight its\neffectiveness in handling diverse distortion patterns across low- to\nhigh-frequency regions.\n", "date": "2025-04-15", "categories": []}
{"id": "2504.11038", "title": "QAVA: Query-Agnostic Visual Attack to Large Vision-Language Models", "abstract": "  In typical multimodal tasks, such as Visual Question Answering (VQA),\nadversarial attacks targeting a specific image and question can lead large\nvision-language models (LVLMs) to provide incorrect answers. However, it is\ncommon for a single image to be associated with multiple questions, and LVLMs\nmay still answer other questions correctly even for an adversarial image\nattacked by a specific question. To address this, we introduce the\nquery-agnostic visual attack (QAVA), which aims to create robust adversarial\nexamples that generate incorrect responses to unspecified and unknown\nquestions. Compared to traditional adversarial attacks focused on specific\nimages and questions, QAVA significantly enhances the effectiveness and\nefficiency of attacks on images when the question is unknown, achieving\nperformance comparable to attacks on known target questions. Our research\nbroadens the scope of visual adversarial attacks on LVLMs in practical\nsettings, uncovering previously overlooked vulnerabilities, particularly in the\ncontext of visual adversarial threats. The code is available at\nhttps://github.com/btzyd/qava.\n", "date": "2025-04-15", "categories": []}
{"id": "2310.13019", "title": "Tailoring Adversarial Attacks on Deep Neural Networks for Targeted Class\n  Manipulation Using DeepFool Algorithm", "abstract": "  The susceptibility of deep neural networks (DNNs) to adversarial attacks\nundermines their reliability across numerous applications, underscoring the\nnecessity for an in-depth exploration of these vulnerabilities and the\nformulation of robust defense strategies. The DeepFool algorithm by\nMoosavi-Dezfooli et al. (2016) represents a pivotal step in identifying minimal\nperturbations required to induce misclassification of input images.\nNonetheless, its generic methodology falls short in scenarios necessitating\ntargeted interventions. Additionally, previous research studies have\npredominantly concentrated on the success rate of attacks without adequately\naddressing the consequential distortion of images, the maintenance of image\nquality, or the confidence threshold required for misclassification. To bridge\nthese gaps, we introduce the Enhanced Targeted DeepFool (ET DeepFool)\nalgorithm, an evolution of DeepFool that not only facilitates the specification\nof desired misclassification targets but also incorporates a configurable\nminimum confidence score. Our empirical investigations demonstrate the\nsuperiority of this refined approach in maintaining the integrity of images and\nminimizing perturbations across a variety of DNN architectures. Unlike previous\niterations, such as the Targeted DeepFool by Gajjar et al. (2022), our method\ngrants unparalleled control over the perturbation process, enabling precise\nmanipulation of model responses. Preliminary outcomes reveal that certain\nmodels, including AlexNet and the advanced Vision Transformer, display\ncommendable robustness to such manipulations. This discovery of varying levels\nof model robustness, as unveiled through our confidence level adjustments,\ncould have far-reaching implications for the field of image recognition. Our\ncode is available at https://github.com/FazleLabib/et_deepfool.\n", "date": "2023-10-18", "categories": []}
{"id": "2406.13073", "title": "Let the Noise Speak: Harnessing Noise for a Unified Defense Against\n  Adversarial and Backdoor Attacks", "abstract": "  The exponential adoption of machine learning (ML) is propelling the world\ninto a future of distributed and intelligent automation and data-driven\nsolutions. However, the proliferation of malicious data manipulation attacks\nagainst ML, namely adversarial and backdoor attacks, jeopardizes its\nreliability in safety-critical applications. The existing detection methods are\nattack-specific and built upon some strong assumptions, limiting them in\ndiverse practical scenarios. Thus, motivated by the need for a more robust,\nunified, and attack-agnostic defense mechanism, we first investigate the shared\ntraits of adversarial and backdoor attacks. Based on our observation, we\npropose NoiSec, a reconstruction-based intrusion detection system that brings a\nnovel perspective by shifting focus from the reconstructed input to the\nreconstruction noise itself, which is the foundational root cause of such\nmalicious data alterations. NoiSec disentangles the noise from the test input,\nextracts the underlying features from the noise, and leverages them to\nrecognize systematic malicious manipulation. Our comprehensive evaluation of\nNoiSec demonstrates its high effectiveness across various datasets, including\nbasic objects, natural scenes, traffic signs, medical images, spectrogram-based\naudio data, and wireless sensing against five state-of-the-art adversarial\nattacks and three backdoor attacks under challenging evaluation conditions.\nNoiSec demonstrates strong detection performance in both white-box and\nblack-box adversarial attack scenarios, significantly outperforming the closest\nbaseline models, particularly in an adaptive attack setting. We will provide\nthe code for future baseline comparison. Our code and artifacts are publicly\navailable at https://github.com/shahriar0651/NoiSec.\n", "date": "2024-06-18", "categories": []}
{"id": "2412.03539", "title": "NODE-AdvGAN: Improving the transferability and perceptual similarity of\n  adversarial examples by dynamic-system-driven adversarial generative model", "abstract": "  Understanding adversarial examples is crucial for improving model robustness,\nas they introduce imperceptible perturbations to deceive models. Effective\nadversarial examples, therefore, offer the potential to train more robust\nmodels by eliminating model singularities. We propose NODE-AdvGAN, a novel\napproach that treats adversarial generation as a continuous process and employs\na Neural Ordinary Differential Equation (NODE) to simulate generator dynamics.\nBy mimicking the iterative nature of traditional gradient-based methods,\nNODE-AdvGAN generates smoother and more precise perturbations that preserve\nhigh perceptual similarity when added to benign images. We also propose a new\ntraining strategy, NODE-AdvGAN-T, which enhances transferability in black-box\nattacks by tuning the noise parameters during training. Experiments demonstrate\nthat NODE-AdvGAN and NODE-AdvGAN-T generate more effective adversarial examples\nthat achieve higher attack success rates while preserving better perceptual\nquality than baseline models.\n", "date": "2024-12-04", "categories": []}
{"id": "2412.11423", "title": "Nearly Zero-Cost Protection Against Mimicry by Personalized Diffusion\n  Models", "abstract": "  Recent advancements in diffusion models revolutionize image generation but\npose risks of misuse, such as replicating artworks or generating deepfakes.\nExisting image protection methods, though effective, struggle to balance\nprotection efficacy, invisibility, and latency, thus limiting practical use. We\nintroduce perturbation pre-training to reduce latency and propose a\nmixture-of-perturbations approach that dynamically adapts to input images to\nminimize performance degradation. Our novel training strategy computes\nprotection loss across multiple VAE feature spaces, while adaptive targeted\nprotection at inference enhances robustness and invisibility. Experiments show\ncomparable protection performance with improved invisibility and drastically\nreduced inference time. The code and demo are available at\nhttps://webtoon.github.io/impasto\n", "date": "2024-12-16", "categories": []}
{"id": "2504.08782", "title": "Embedding Hidden Adversarial Capabilities in Pre-Trained Diffusion\n  Models", "abstract": "  We introduce a new attack paradigm that embeds hidden adversarial\ncapabilities directly into diffusion models via fine-tuning, without altering\ntheir observable behavior or requiring modifications during inference. Unlike\nprior approaches that target specific images or adjust the generation process\nto produce adversarial outputs, our method integrates adversarial functionality\ninto the model itself. The resulting tampered model generates high-quality\nimages indistinguishable from those of the original, yet these images cause\nmisclassification in downstream classifiers at a high rate. The\nmisclassification can be targeted to specific output classes. Users can employ\nthis compromised model unaware of its embedded adversarial nature, as it\nfunctions identically to a standard diffusion model. We demonstrate the\neffectiveness and stealthiness of our approach, uncovering a covert attack\nvector that raises new security concerns. These findings expose a risk arising\nfrom the use of externally-supplied models and highlight the urgent need for\nrobust model verification and defense mechanisms against hidden threats in\ngenerative models. The code is available at\nhttps://github.com/LucasBeerens/CRAFTed-Diffusion .\n", "date": "2025-04-05", "categories": []}
{"id": "2504.08798", "title": "Exploring Gradient-Guided Masked Language Model to Detect Textual\n  Adversarial Attacks", "abstract": "  Textual adversarial examples pose serious threats to the reliability of\nnatural language processing systems. Recent studies suggest that adversarial\nexamples tend to deviate from the underlying manifold of normal texts, whereas\npre-trained masked language models can approximate the manifold of normal data.\nThese findings inspire the exploration of masked language models for detecting\ntextual adversarial attacks. We first introduce Masked Language Model-based\nDetection (MLMD), leveraging the mask and unmask operations of the masked\nlanguage modeling (MLM) objective to induce the difference in manifold changes\nbetween normal and adversarial texts. Although MLMD achieves competitive\ndetection performance, its exhaustive one-by-one masking strategy introduces\nsignificant computational overhead. Our posterior analysis reveals that a\nsignificant number of non-keywords in the input are not important for detection\nbut consume resources. Building on this, we introduce Gradient-guided MLMD\n(GradMLMD), which leverages gradient information to identify and skip\nnon-keywords during detection, significantly reducing resource consumption\nwithout compromising detection performance.\n", "date": "2025-04-08", "categories": []}
{"id": "2504.08866", "title": "On Transfer-based Universal Attacks in Pure Black-box Setting", "abstract": "  Despite their impressive performance, deep visual models are susceptible to\ntransferable black-box adversarial attacks. Principally, these attacks craft\nperturbations in a target model-agnostic manner. However, surprisingly, we find\nthat existing methods in this domain inadvertently take help from various\npriors that violate the black-box assumption such as the availability of the\ndataset used to train the target model, and the knowledge of the number of\nclasses in the target model. Consequently, the literature fails to articulate\nthe true potency of transferable black-box attacks. We provide an empirical\nstudy of these biases and propose a framework that aids in a prior-free\ntransparent study of this paradigm. Using our framework, we analyze the role of\nprior knowledge of the target model data and number of classes in attack\nperformance. We also provide several interesting insights based on our\nanalysis, and demonstrate that priors cause overestimation in transferability\nscores. Finally, we extend our framework to query-based attacks. This extension\ninspires a novel image-blending technique to prepare data for effective\nsurrogate model training.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.08906", "title": "Robust SAM: On the Adversarial Robustness of Vision Foundation Models", "abstract": "  The Segment Anything Model (SAM) is a widely used vision foundation model\nwith diverse applications, including image segmentation, detection, and\ntracking. Given SAM's wide applications, understanding its robustness against\nadversarial attacks is crucial for real-world deployment. However, research on\nSAM's robustness is still in its early stages. Existing attacks often overlook\nthe role of prompts in evaluating SAM's robustness, and there has been\ninsufficient exploration of defense methods to balance the robustness and\naccuracy. To address these gaps, this paper proposes an adversarial robustness\nframework designed to evaluate and enhance the robustness of SAM. Specifically,\nwe introduce a cross-prompt attack method to enhance the attack transferability\nacross different prompt types. Besides attacking, we propose a few-parameter\nadaptation strategy to defend SAM against various adversarial attacks. To\nbalance robustness and accuracy, we use the singular value decomposition (SVD)\nto constrain the space of trainable parameters, where only singular values are\nadaptable. Experiments demonstrate that our cross-prompt attack method\noutperforms previous approaches in terms of attack success rate on both SAM and\nSAM 2. By adapting only 512 parameters, we achieve at least a 15\\% improvement\nin mean intersection over union (mIoU) against various adversarial attacks.\nCompared to previous defense methods, our approach enhances the robustness of\nSAM while maximally maintaining its original performance.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.09361", "title": "PapMOT: Exploring Adversarial Patch Attack against Multiple Object\n  Tracking", "abstract": "  Tracking multiple objects in a continuous video stream is crucial for many\ncomputer vision tasks. It involves detecting and associating objects with their\nrespective identities across successive frames. Despite significant progress\nmade in multiple object tracking (MOT), recent studies have revealed the\nvulnerability of existing MOT methods to adversarial attacks. Nevertheless, all\nof these attacks belong to digital attacks that inject pixel-level noise into\ninput images, and are therefore ineffective in physical scenarios. To fill this\ngap, we propose PapMOT, which can generate physical adversarial patches against\nMOT for both digital and physical scenarios. Besides attacking the detection\nmechanism, PapMOT also optimizes a printable patch that can be detected as new\ntargets to mislead the identity association process. Moreover, we introduce a\npatch enhancement strategy to further degrade the temporal consistency of\ntracking results across video frames, resulting in more aggressive attacks. We\nfurther develop new evaluation metrics to assess the robustness of MOT against\nsuch attacks. Extensive evaluations on multiple datasets demonstrate that our\nPapMOT can successfully attack various architectures of MOT trackers in digital\nscenarios. We also validate the effectiveness of PapMOT for physical attacks by\ndeploying printed adversarial patches in the real world.\n", "date": "2025-04-12", "categories": []}
{"id": "2504.09839", "title": "SafeSpeech: Robust and Universal Voice Protection Against Malicious\n  Speech Synthesis", "abstract": "  Speech synthesis technology has brought great convenience, while the\nwidespread usage of realistic deepfake audio has triggered hazards. Malicious\nadversaries may unauthorizedly collect victims' speeches and clone a similar\nvoice for illegal exploitation (\\textit{e.g.}, telecom fraud). However, the\nexisting defense methods cannot effectively prevent deepfake exploitation and\nare vulnerable to robust training techniques. Therefore, a more effective and\nrobust data protection method is urgently needed. In response, we propose a\ndefensive framework, \\textit{\\textbf{SafeSpeech}}, which protects the users'\naudio before uploading by embedding imperceptible perturbations on original\nspeeches to prevent high-quality synthetic speech. In SafeSpeech, we devise a\nrobust and universal proactive protection technique, \\textbf{S}peech\n\\textbf{PE}rturbative \\textbf{C}oncealment (\\textbf{SPEC}), that leverages a\nsurrogate model to generate universally applicable perturbation for generative\nsynthetic models. Moreover, we optimize the human perception of embedded\nperturbation in terms of time and frequency domains. To evaluate our method\ncomprehensively, we conduct extensive experiments across advanced models and\ndatasets, both subjectively and objectively. Our experimental results\ndemonstrate that SafeSpeech achieves state-of-the-art (SOTA) voice protection\neffectiveness and transferability and is highly robust against advanced\nadaptive adversaries. Moreover, SafeSpeech has real-time capability in\nreal-world tests. The source code is available at\n\\href{https://github.com/wxzyd123/SafeSpeech}{https://github.com/wxzyd123/SafeSpeech}.\n", "date": "2025-04-14", "categories": []}
{"id": "2410.21443", "title": "TACO: Adversarial Camouflage Optimization on Trucks to Fool Object\n  Detectors", "abstract": "  Adversarial attacks threaten the reliability of machine learning models in\ncritical applications like autonomous vehicles and defense systems. As object\ndetectors become more robust with models like YOLOv8, developing effective\nadversarial methodologies is increasingly challenging. We present Truck\nAdversarial Camouflage Optimization (TACO), a novel framework that generates\nadversarial camouflage patterns on 3D vehicle models to deceive\nstate-of-the-art object detectors. Adopting Unreal Engine 5, TACO integrates\ndifferentiable rendering with a Photorealistic Rendering Network to optimize\nadversarial textures targeted at YOLOv8. To ensure the generated textures are\nboth effective in deceiving detectors and visually plausible, we introduce the\nConvolutional Smooth Loss function, a generalized smooth loss function.\nExperimental evaluations demonstrate that TACO significantly degrades YOLOv8's\ndetection performance, achieving an AP@0.5 of 0.0099 on unseen test data.\nFurthermore, these adversarial patterns exhibit strong transferability to other\nobject detection models such as Faster R-CNN and earlier YOLO versions.\n", "date": "2024-10-28", "categories": []}
{"id": "2501.18998", "title": "Adversarial Attacks on AI-Generated Text Detection Models: A Token\n  Probability-Based Approach Using Embeddings", "abstract": "  In recent years, text generation tools utilizing Artificial Intelligence (AI)\nhave occasionally been misused across various domains, such as generating\nstudent reports or creative writings. This issue prompts plagiarism detection\nservices to enhance their capabilities in identifying AI-generated content.\nAdversarial attacks are often used to test the robustness of AI-text generated\ndetectors. This work proposes a novel textual adversarial attack on the\ndetection models such as Fast-DetectGPT. The method employs embedding models\nfor data perturbation, aiming at reconstructing the AI generated texts to\nreduce the likelihood of detection of the true origin of the texts.\nSpecifically, we employ different embedding techniques, including the Tsetlin\nMachine (TM), an interpretable approach in machine learning for this purpose.\nBy combining synonyms and embedding similarity vectors, we demonstrates the\nstate-of-the-art reduction in detection scores against Fast-DetectGPT.\nParticularly, in the XSum dataset, the detection score decreased from 0.4431 to\n0.2744 AUROC, and in the SQuAD dataset, it dropped from 0.5068 to 0.3532 AUROC.\n", "date": "2025-01-31", "categories": []}
{"id": "2504.08205", "title": "EO-VLM: VLM-Guided Energy Overload Attacks on Vision Models", "abstract": "  Vision models are increasingly deployed in critical applications such as\nautonomous driving and CCTV monitoring, yet they remain susceptible to\nresource-consuming attacks. In this paper, we introduce a novel\nenergy-overloading attack that leverages vision language model (VLM) prompts to\ngenerate adversarial images targeting vision models. These images, though\nimperceptible to the human eye, significantly increase GPU energy consumption\nacross various vision models, threatening the availability of these systems.\nOur framework, EO-VLM (Energy Overload via VLM), is model-agnostic, meaning it\nis not limited by the architecture or type of the target vision model. By\nexploiting the lack of safety filters in VLMs like DALL-E 3, we create\nadversarial noise images without requiring prior knowledge or internal\nstructure of the target vision models. Our experiments demonstrate up to a 50%\nincrease in energy consumption, revealing a critical vulnerability in current\nvision models.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.08414", "title": "Adversarial Examples in Environment Perception for Automated Driving\n  (Review)", "abstract": "  The renaissance of deep learning has led to the massive development of\nautomated driving. However, deep neural networks are vulnerable to adversarial\nexamples. The perturbations of adversarial examples are imperceptible to human\neyes but can lead to the false predictions of neural networks. It poses a huge\nrisk to artificial intelligence (AI) applications for automated driving. This\nsurvey systematically reviews the development of adversarial robustness\nresearch over the past decade, including the attack and defense methods and\ntheir applications in automated driving. The growth of automated driving pushes\nforward the realization of trustworthy AI applications. This review lists\nsignificant references in the research history of adversarial examples.\n", "date": "2025-04-11", "categories": []}
{"id": "2409.13349", "title": "ID-Guard: A Universal Framework for Combating Facial Manipulation via\n  Breaking Identification", "abstract": "  The misuse of deep learning-based facial manipulation poses a significant\nthreat to civil rights. To prevent this fraud at its source, proactive defense\nhas been proposed to disrupt the manipulation process by adding invisible\nadversarial perturbations into images, making the forged output unconvincing to\nobservers. However, the non-specific disruption against the output may lead to\nthe retention of identifiable facial features, potentially resulting in the\nstigmatization of the individual. This paper proposes a universal framework for\ncombating facial manipulation, termed ID-Guard. Specifically, this framework\noperates with a single forward pass of an encoder-decoder network to produce a\ncross-model transferable adversarial perturbation. A novel Identity Destruction\nModule (IDM) is introduced to degrade identifiable features in forged faces. We\noptimize the perturbation generation by framing the disruption of different\nfacial manipulations as a multi-task learning problem, and a dynamic weight\nstrategy is devised to enhance cross-model performance. Experimental results\ndemonstrate that the proposed ID-Guard exhibits strong efficacy in defending\nagainst various facial manipulation models, effectively degrading identifiable\nregions in manipulated images. It also enables disrupted images to evade facial\ninpainting and image recognition systems. Additionally, ID-Guard can seamlessly\nfunction as a plug-and-play component, integrating with other tasks such as\nadversarial training.\n", "date": "2024-09-20", "categories": []}
{"id": "2501.15257", "title": "Towards Communication-Efficient Adversarial Federated Learning for\n  Robust Edge Intelligence", "abstract": "  Federated learning (FL) has gained significant attention for enabling\ndecentralized training on edge networks without exposing raw data. However, FL\nmodels remain susceptible to adversarial attacks and performance degradation in\nnon-IID data settings, thus posing challenges to both robustness and accuracy.\nThis paper aims to achieve communication-efficient adversarial federated\nlearning (AFL) by leveraging a pre-trained model to enhance both robustness and\naccuracy under adversarial attacks and non-IID challenges in AFL. By leveraging\nthe knowledge from a pre-trained model for both clean and adversarial images,\nwe propose a pre-trained model-guided adversarial federated learning (PM-AFL)\nframework. This framework integrates vanilla and adversarial mixture knowledge\ndistillation to effectively balance accuracy and robustness while promoting\nlocal models to learn from diverse data. Specifically, for clean accuracy, we\nadopt a dual distillation strategy where the class probabilities of randomly\npaired images, and their blended versions are aligned between the teacher model\nand the local models. For adversarial robustness, we employ a similar\ndistillation approach but replace clean samples on the local side with\nadversarial examples. Moreover, by considering the bias between local and\nglobal models, we also incorporate a consistency regularization term to ensure\nthat local adversarial predictions stay aligned with their corresponding global\nclean ones. These strategies collectively enable local models to absorb diverse\nknowledge from the teacher model while maintaining close alignment with the\nglobal model, thereby mitigating overfitting to local optima and enhancing the\ngeneralization of the global model. Experiments demonstrate that the\nPM-AFL-based framework not only significantly outperforms other methods but\nalso maintains communication efficiency.\n", "date": "2025-01-25", "categories": []}
{"id": "2504.02412", "title": "Bridging the Theoretical Gap in Randomized Smoothing", "abstract": "  Randomized smoothing has become a leading approach for certifying adversarial\nrobustness in machine learning models. However, a persistent gap remains\nbetween theoretical certified robustness and empirical robustness accuracy.\nThis paper introduces a new framework that bridges this gap by leveraging\nLipschitz continuity for certification and proposing a novel, less conservative\nmethod for computing confidence intervals in randomized smoothing. Our approach\ntightens the bounds of certified robustness, offering a more accurate\nreflection of model robustness in practice. Through rigorous experimentation we\nshow that our method improves the robust accuracy, compressing the gap between\nempirical findings and previous theoretical results. We argue that\ninvestigating local Lipschitz constants and designing ad-hoc confidence\nintervals can further enhance the performance of randomized smoothing. These\nresults pave the way for a deeper understanding of the relationship between\nLipschitz continuity and certified robustness.\n", "date": "2025-04-03", "categories": []}
{"id": "2305.15203", "title": "Frequency maps reveal the correlation between Adversarial Attacks and\n  Implicit Bias", "abstract": "  Despite their impressive performance in classification tasks, neural networks\nare known to be vulnerable to adversarial attacks, subtle perturbations of the\ninput data designed to deceive the model. In this work, we investigate the\ncorrelation between these perturbations and the implicit bias of neural\nnetworks trained with gradient-based algorithms. To this end, we analyse a\nrepresentation of the network's implicit bias through the lens of the Fourier\ntransform. Specifically, we identify unique fingerprints of implicit bias and\nadversarial attacks by calculating the minimal, essential frequencies needed\nfor accurate classification of each image, as well as the frequencies that\ndrive misclassification in its adversarially perturbed counterpart. This\napproach enables us to uncover and analyse the correlation between these\nessential frequencies, providing a precise map of how the network's biases\nalign or contrast with the frequency components exploited by adversarial\nattacks. To this end, among other methods, we use a newly introduced technique\ncapable of detecting nonlinear correlations between high-dimensional datasets.\nOur results provide empirical evidence that the network bias in Fourier space\nand the target frequencies of adversarial attacks are highly correlated and\nsuggest new potential strategies for adversarial defence.\n", "date": "2023-05-24", "categories": []}
{"id": "2412.00114", "title": "SceneTAP: Scene-Coherent Typographic Adversarial Planner against\n  Vision-Language Models in Real-World Environments", "abstract": "  Large vision-language models (LVLMs) have shown remarkable capabilities in\ninterpreting visual content. While existing works demonstrate these models'\nvulnerability to deliberately placed adversarial texts, such texts are often\neasily identifiable as anomalous. In this paper, we present the first approach\nto generate scene-coherent typographic adversarial attacks that mislead\nadvanced LVLMs while maintaining visual naturalness through the capability of\nthe LLM-based agent. Our approach addresses three critical questions: what\nadversarial text to generate, where to place it within the scene, and how to\nintegrate it seamlessly. We propose a training-free, multi-modal LLM-driven\nscene-coherent typographic adversarial planning (SceneTAP) that employs a\nthree-stage process: scene understanding, adversarial planning, and seamless\nintegration. The SceneTAP utilizes chain-of-thought reasoning to comprehend the\nscene, formulate effective adversarial text, strategically plan its placement,\nand provide detailed instructions for natural integration within the image.\nThis is followed by a scene-coherent TextDiffuser that executes the attack\nusing a local diffusion mechanism. We extend our method to real-world scenarios\nby printing and placing generated patches in physical environments,\ndemonstrating its practical implications. Extensive experiments show that our\nscene-coherent adversarial text successfully misleads state-of-the-art LVLMs,\nincluding ChatGPT-4o, even after capturing new images of physical setups. Our\nevaluations demonstrate a significant increase in attack success rates while\nmaintaining visual naturalness and contextual appropriateness. This work\nhighlights vulnerabilities in current vision-language models to sophisticated,\nscene-coherent adversarial attacks and provides insights into potential defense\nmechanisms.\n", "date": "2024-11-28", "categories": []}
{"id": "2412.08755", "title": "Proactive Adversarial Defense: Harnessing Prompt Tuning in\n  Vision-Language Models to Detect Unseen Backdoored Images", "abstract": "  Backdoor attacks pose a critical threat by embedding hidden triggers into\ninputs, causing models to misclassify them into target labels. While extensive\nresearch has focused on mitigating these attacks in object recognition models\nthrough weight fine-tuning, much less attention has been given to detecting\nbackdoored samples directly. Given the vast datasets used in training, manual\ninspection for backdoor triggers is impractical, and even state-of-the-art\ndefense mechanisms fail to fully neutralize their impact. To address this gap,\nwe introduce a groundbreaking method to detect unseen backdoored images during\nboth training and inference. Leveraging the transformative success of prompt\ntuning in Vision Language Models (VLMs), our approach trains learnable text\nprompts to differentiate clean images from those with hidden backdoor triggers.\nExperiments demonstrate the exceptional efficacy of this method, achieving an\nimpressive average accuracy of 86% across two renowned datasets for detecting\nunseen backdoor triggers, establishing a new standard in backdoor defense.\n", "date": "2024-12-11", "categories": []}
{"id": "2501.09446", "title": "Double Visual Defense: Adversarial Pre-training and Instruction Tuning\n  for Improving Vision-Language Model Robustness", "abstract": "  This paper investigates the robustness of vision-language models against\nadversarial visual perturbations and introduces a novel ``double visual\ndefense\" to enhance this robustness. Unlike previous approaches that resort to\nlightweight adversarial fine-tuning of a pre-trained CLIP model, we perform\nlarge-scale adversarial vision-language pre-training from scratch using\nweb-scale data. We then strengthen the defense by incorporating adversarial\nvisual instruction tuning. The resulting models from each stage, $\\Delta$CLIP\nand $\\Delta^2$LLaVA, show substantially enhanced zero-shot robustness and set a\nnew state-of-the-art in adversarial defense for vision-language models. For\nexample, the adversarial robustness of $\\Delta$CLIP surpasses that of the\nprevious best models on ImageNet-1k by ~20%. %For example, $\\Delta$CLIP\nsurpasses the previous best models on ImageNet-1k by ~20% in terms of\nadversarial robustness. Similarly, compared to prior art, $\\Delta^2$LLaVA\nbrings a ~30% robustness improvement to image captioning task and a ~20%\nrobustness improvement to visual question answering task. Furthermore, our\nmodels exhibit stronger zero-shot recognition capability, fewer hallucinations,\nand superior reasoning performance compared to baselines. Our project page is\nhttps://doublevisualdefense.github.io/.\n", "date": "2025-01-16", "categories": []}
{"id": "2504.05838", "title": "Mind the Trojan Horse: Image Prompt Adapter Enabling Scalable and\n  Deceptive Jailbreaking", "abstract": "  Recently, the Image Prompt Adapter (IP-Adapter) has been increasingly\nintegrated into text-to-image diffusion models (T2I-DMs) to improve\ncontrollability. However, in this paper, we reveal that T2I-DMs equipped with\nthe IP-Adapter (T2I-IP-DMs) enable a new jailbreak attack named the hijacking\nattack. We demonstrate that, by uploading imperceptible image-space adversarial\nexamples (AEs), the adversary can hijack massive benign users to jailbreak an\nImage Generation Service (IGS) driven by T2I-IP-DMs and mislead the public to\ndiscredit the service provider. Worse still, the IP-Adapter's dependency on\nopen-source image encoders reduces the knowledge required to craft AEs.\nExtensive experiments verify the technical feasibility of the hijacking attack.\nIn light of the revealed threat, we investigate several existing defenses and\nexplore combining the IP-Adapter with adversarially trained models to overcome\nexisting defenses' limitations. Our code is available at\nhttps://github.com/fhdnskfbeuv/attackIPA.\n", "date": "2025-04-08", "categories": []}
{"id": "2307.15539", "title": "Beating Backdoor Attack at Its Own Game", "abstract": "  Deep neural networks (DNNs) are vulnerable to backdoor attack, which does not\naffect the network's performance on clean data but would manipulate the network\nbehavior once a trigger pattern is added. Existing defense methods have greatly\nreduced attack success rate, but their prediction accuracy on clean data still\nlags behind a clean model by a large margin. Inspired by the stealthiness and\neffectiveness of backdoor attack, we propose a simple but highly effective\ndefense framework which injects non-adversarial backdoors targeting poisoned\nsamples. Following the general steps in backdoor attack, we detect a small set\nof suspected samples and then apply a poisoning strategy to them. The\nnon-adversarial backdoor, once triggered, suppresses the attacker's backdoor on\npoisoned data, but has limited influence on clean data. The defense can be\ncarried out during data preprocessing, without any modification to the standard\nend-to-end training pipeline. We conduct extensive experiments on multiple\nbenchmarks with different architectures and representative attacks. Results\ndemonstrate that our method achieves state-of-the-art defense effectiveness\nwith by far the lowest performance drop on clean data. Considering the\nsurprising defense ability displayed by our framework, we call for more\nattention to utilizing backdoor for backdoor defense. Code is available at\nhttps://github.com/minliu01/non-adversarial_backdoor.\n", "date": "2023-07-28", "categories": []}
{"id": "2403.10045", "title": "Towards Adversarially Robust Dataset Distillation by Curvature\n  Regularization", "abstract": "  Dataset distillation (DD) allows datasets to be distilled to fractions of\ntheir original size while preserving the rich distributional information, so\nthat models trained on the distilled datasets can achieve a comparable accuracy\nwhile saving significant computational loads. Recent research in this area has\nbeen focusing on improving the accuracy of models trained on distilled\ndatasets. In this paper, we aim to explore a new perspective of DD. We study\nhow to embed adversarial robustness in distilled datasets, so that models\ntrained on these datasets maintain the high accuracy and meanwhile acquire\nbetter adversarial robustness. We propose a new method that achieves this goal\nby incorporating curvature regularization into the distillation process with\nmuch less computational overhead than standard adversarial training. Extensive\nempirical experiments suggest that our method not only outperforms standard\nadversarial training on both accuracy and robustness with less computation\noverhead but is also capable of generating robust distilled datasets that can\nwithstand various adversarial attacks. Our implementation is available at:\nhttps://github.com/yumozi/GUARD.\n", "date": "2024-03-15", "categories": []}
{"id": "2404.14309", "title": "Towards Understanding the Robustness of Diffusion-Based Purification: A\n  Stochastic Perspective", "abstract": "  Diffusion-Based Purification (DBP) has emerged as an effective defense\nmechanism against adversarial attacks. The success of DBP is often attributed\nto the forward diffusion process, which reduces the distribution gap between\nclean and adversarial images by adding Gaussian noise. While this explanation\nis theoretically sound, the exact role of this mechanism in enhancing\nrobustness remains unclear. In this paper, through empirical analysis, we\npropose that the intrinsic stochasticity in the DBP process is the primary\nfactor driving robustness. To test this hypothesis, we introduce a novel\nDeterministic White-Box (DW-box) setting to assess robustness in the absence of\nstochasticity, and we analyze attack trajectories and loss landscapes. Our\nresults suggest that DBP models primarily rely on stochasticity to avoid\neffective attack directions, while their ability to purify adversarial\nperturbations may be limited. To further enhance the robustness of DBP models,\nwe propose Adversarial Denoising Diffusion Training (ADDT), which incorporates\nclassifier-guided adversarial perturbations into the diffusion training\nprocess, thereby strengthening the models' ability to purify adversarial\nperturbations. Additionally, we propose Rank-Based Gaussian Mapping (RBGM) to\nimprove the compatibility of perturbations with diffusion models. Experimental\nresults validate the effectiveness of ADDT. In conclusion, our study suggests\nthat future research on DBP can benefit from a clearer distinction between\nstochasticity-driven and purification-driven robustness.\n", "date": "2024-04-22", "categories": []}
{"id": "2408.11408", "title": "Latent Feature and Attention Dual Erasure Attack against Multi-View\n  Diffusion Models for 3D Assets Protection", "abstract": "  Multi-View Diffusion Models (MVDMs) enable remarkable improvements in the\nfield of 3D geometric reconstruction, but the issue regarding intellectual\nproperty has received increasing attention due to unauthorized imitation.\nRecently, some works have utilized adversarial attacks to protect copyright.\nHowever, all these works focus on single-image generation tasks which only need\nto consider the inner feature of images. Previous methods are inefficient in\nattacking MVDMs because they lack the consideration of disrupting the geometric\nand visual consistency among the generated multi-view images. This paper is the\nfirst to address the intellectual property infringement issue arising from\nMVDMs. Accordingly, we propose a novel latent feature and attention dual\nerasure attack to disrupt the distribution of latent feature and the\nconsistency across the generated images from multi-view and multi-domain\nsimultaneously. The experiments conducted on SOTA MVDMs indicate that our\napproach achieves superior performances in terms of attack effectiveness,\ntransferability, and robustness against defense methods. Therefore, this paper\nprovides an efficient solution to protect 3D assets from MVDMs-based 3D\ngeometry reconstruction.\n", "date": "2024-08-21", "categories": []}
{"id": "2412.20006", "title": "Adversarial Robustness for Deep Learning-based Wildfire Prediction\n  Models", "abstract": "  Rapidly growing wildfires have recently devastated societal assets, exposing\na critical need for early warning systems to expedite relief efforts. Smoke\ndetection using camera-based Deep Neural Networks (DNNs) offers a promising\nsolution for wildfire prediction. However, the rarity of smoke across time and\nspace limits training data, raising model overfitting and bias concerns.\nCurrent DNNs, primarily Convolutional Neural Networks (CNNs) and transformers,\ncomplicate robustness evaluation due to architectural differences. To address\nthese challenges, we introduce WARP (Wildfire Adversarial Robustness\nProcedure), the first model-agnostic framework for evaluating wildfire\ndetection models' adversarial robustness. WARP addresses inherent limitations\nin data diversity by generating adversarial examples through image-global and\n-local perturbations. Global and local attacks superimpose Gaussian noise and\nPNG patches onto image inputs, respectively; this suits both CNNs and\ntransformers while generating realistic adversarial scenarios. Using WARP, we\nassessed real-time CNNs and Transformers, uncovering key vulnerabilities. At\ntimes, transformers exhibited over 70% precision degradation under global\nattacks, while both models generally struggled to differentiate cloud-like PNG\npatches from real smoke during local attacks. To enhance model robustness, we\nproposed four wildfire-oriented data augmentation techniques based on WARP's\nmethodology and results, which diversify smoke image data and improve model\nprecision and robustness. These advancements represent a substantial step\ntoward developing a reliable early wildfire warning system, which may be our\nfirst safeguard against wildfire destruction.\n", "date": "2024-12-28", "categories": []}
{"id": "2504.03735", "title": "Misaligned Roles, Misplaced Images: Structural Input Perturbations\n  Expose Multimodal Alignment Blind Spots", "abstract": "  Multimodal Language Models (MMLMs) typically undergo post-training alignment\nto prevent harmful content generation. However, these alignment stages focus\nprimarily on the assistant role, leaving the user role unaligned, and stick to\na fixed input prompt structure of special tokens, leaving the model vulnerable\nwhen inputs deviate from these expectations. We introduce Role-Modality Attacks\n(RMA), a novel class of adversarial attacks that exploit role confusion between\nthe user and assistant and alter the position of the image token to elicit\nharmful outputs. Unlike existing attacks that modify query content, RMAs\nmanipulate the input structure without altering the query itself. We\nsystematically evaluate these attacks across multiple Vision Language Models\n(VLMs) on eight distinct settings, showing that they can be composed to create\nstronger adversarial prompts, as also evidenced by their increased projection\nin the negative refusal direction in the residual stream, a property observed\nin prior successful attacks. Finally, for mitigation, we propose an adversarial\ntraining approach that makes the model robust against input prompt\nperturbations. By training the model on a range of harmful and benign prompts\nall perturbed with different RMA settings, it loses its sensitivity to Role\nConfusion and Modality Manipulation attacks and is trained to only pay\nattention to the content of the query in the input prompt structure,\neffectively reducing Attack Success Rate (ASR) while preserving the model's\ngeneral utility.\n", "date": "2025-04-01", "categories": []}
{"id": "2504.03759", "title": "Emerging Cyber Attack Risks of Medical AI Agents", "abstract": "  Large language models (LLMs)-powered AI agents exhibit a high level of\nautonomy in addressing medical and healthcare challenges. With the ability to\naccess various tools, they can operate within an open-ended action space.\nHowever, with the increase in autonomy and ability, unforeseen risks also\narise. In this work, we investigated one particular risk, i.e., cyber attack\nvulnerability of medical AI agents, as agents have access to the Internet\nthrough web browsing tools. We revealed that through adversarial prompts\nembedded on webpages, cyberattackers can: i) inject false information into the\nagent's response; ii) they can force the agent to manipulate recommendation\n(e.g., healthcare products and services); iii) the attacker can also steal\nhistorical conversations between the user and agent, resulting in the leak of\nsensitive/private medical information; iv) furthermore, the targeted agent can\nalso cause a computer system hijack by returning a malicious URL in its\nresponse. Different backbone LLMs were examined, and we found such cyber\nattacks can succeed in agents powered by most mainstream LLMs, with the\nreasoning models such as DeepSeek-R1 being the most vulnerable.\n", "date": "2025-04-02", "categories": []}
{"id": "2504.04716", "title": "On the Robustness of GUI Grounding Models Against Image Attacks", "abstract": "  Graphical User Interface (GUI) grounding models are crucial for enabling\nintelligent agents to understand and interact with complex visual interfaces.\nHowever, these models face significant robustness challenges in real-world\nscenarios due to natural noise and adversarial perturbations, and their\nrobustness remains underexplored. In this study, we systematically evaluate the\nrobustness of state-of-the-art GUI grounding models, such as UGround, under\nthree conditions: natural noise, untargeted adversarial attacks, and targeted\nadversarial attacks. Our experiments, which were conducted across a wide range\nof GUI environments, including mobile, desktop, and web interfaces, have\nclearly demonstrated that GUI grounding models exhibit a high degree of\nsensitivity to adversarial perturbations and low-resolution conditions. These\nfindings provide valuable insights into the vulnerabilities of GUI grounding\nmodels and establish a strong benchmark for future research aimed at enhancing\ntheir robustness in practical applications. Our code is available at\nhttps://github.com/ZZZhr-1/Robust_GUI_Grounding.\n", "date": "2025-04-07", "categories": []}
{"id": "2504.04747", "title": "Two is Better than One: Efficient Ensemble Defense for Robust and\n  Compact Models", "abstract": "  Deep learning-based computer vision systems adopt complex and large\narchitectures to improve performance, yet they face challenges in deployment on\nresource-constrained mobile and edge devices. To address this issue, model\ncompression techniques such as pruning, quantization, and matrix factorization\nhave been proposed; however, these compressed models are often highly\nvulnerable to adversarial attacks. We introduce the \\textbf{Efficient Ensemble\nDefense (EED)} technique, which diversifies the compression of a single base\nmodel based on different pruning importance scores and enhances ensemble\ndiversity to achieve high adversarial robustness and resource efficiency. EED\ndynamically determines the number of necessary sub-models during the inference\nstage, minimizing unnecessary computations while maintaining high robustness.\nOn the CIFAR-10 and SVHN datasets, EED demonstrated state-of-the-art robustness\nperformance compared to existing adversarial pruning techniques, along with an\ninference speed improvement of up to 1.86 times. This proves that EED is a\npowerful defense solution in resource-constrained environments.\n", "date": "2025-04-07", "categories": []}
{"id": "2405.14672", "title": "Invisible Backdoor Attack against Self-supervised Learning", "abstract": "  Self-supervised learning (SSL) models are vulnerable to backdoor attacks.\nExisting backdoor attacks that are effective in SSL often involve noticeable\ntriggers, like colored patches or visible noise, which are vulnerable to human\ninspection. This paper proposes an imperceptible and effective backdoor attack\nagainst self-supervised models. We first find that existing imperceptible\ntriggers designed for supervised learning are less effective in compromising\nself-supervised models. We then identify this ineffectiveness is attributed to\nthe overlap in distributions between the backdoor and augmented samples used in\nSSL. Building on this insight, we design an attack using optimized triggers\ndisentangled with the augmented transformation in the SSL, while remaining\nimperceptible to human vision. Experiments on five datasets and six SSL\nalgorithms demonstrate our attack is highly effective and stealthy. It also has\nstrong resistance to existing backdoor defenses. Our code can be found at\nhttps://github.com/Zhang-Henry/INACTIVE.\n", "date": "2024-05-23", "categories": []}
{"id": "2504.02335", "title": "Evaluating and Enhancing Segmentation Model Robustness with Metamorphic\n  Testing", "abstract": "  Image segmentation is critical for applications such as medical imaging,\naugmented reality, and video surveillance. However, segmentation models often\nlack robustness, making them vulnerable to adversarial perturbations from\nsubtle image distortions. In this work, we propose SegRMT, a metamorphic\ntesting approach that leverages genetic algorithms (GA) to optimize sequences\nof spatial and spectral transformations while preserving image fidelity via a\npredefined PSNR threshold. Using the Cityscapes dataset, our method generates\nadversarial examples that effectively challenge the DeepLabV3 segmentation\nmodel. Our experiments show that SegRMT reduces DeepLabV3's mean Intersection\nover Union (mIoU) to 6.4%, outperforming other adversarial baselines that\ndecrease mIoU to between 8.5% and 21.7%. Furthermore, when used for adversarial\ntraining, SegRMT boosts model performance, achieving mIoU improvements up to\n73% on dedicated adversarial datasets and increasing cross-adversarial mIoU to\n53.8%, compared to only 2%-10% for other methods. These findings demonstrate\nthat SegRMT not only simulates realistic image distortions but also enhances\nthe robustness of segmentation models, making it a valuable tool for ensuring\nreliable performance in safety-critical applications.\n", "date": "2025-04-03", "categories": []}
{"id": "2402.18370", "title": "Adversarial Example Soups: Improving Transferability and Stealthiness\n  for Free", "abstract": "  Transferable adversarial examples cause practical security risks since they\ncan mislead a target model without knowing its internal knowledge. A\nconventional recipe for maximizing transferability is to keep only the optimal\nadversarial example from all those obtained in the optimization pipeline. In\nthis paper, for the first time, we revisit this convention and demonstrate that\nthose discarded, sub-optimal adversarial examples can be reused to boost\ntransferability. Specifically, we propose ``Adversarial Example Soups'' (AES),\nwith AES-tune for averaging discarded adversarial examples in hyperparameter\ntuning and AES-rand for stability testing. In addition, our AES is inspired by\n``model soups'', which averages weights of multiple fine-tuned models for\nimproved accuracy without increasing inference time. Extensive experiments\nvalidate the global effectiveness of our AES, boosting 10 state-of-the-art\ntransfer attacks and their combinations by up to 13\\% against 10 diverse\n(defensive) target models. We also show the possibility of generalizing AES to\nother types, \\textit{e.g.}, directly averaging multiple in-the-wild adversarial\nexamples that yield comparable success. A promising byproduct of AES is the\nimproved stealthiness of adversarial examples since the perturbation variances\nare naturally reduced.\n", "date": "2024-02-27", "categories": []}
{"id": "2408.11313", "title": "An Optimizable Suffix Is Worth A Thousand Templates: Efficient Black-box\n  Jailbreaking without Affirmative Phrases via LLM as Optimizer", "abstract": "  Despite prior safety alignment efforts, mainstream LLMs can still generate\nharmful and unethical content when subjected to jailbreaking attacks. Existing\njailbreaking methods fall into two main categories: template-based and\noptimization-based methods. The former requires significant manual effort and\ndomain knowledge, while the latter, exemplified by Greedy Coordinate Gradient\n(GCG), which seeks to maximize the likelihood of harmful LLM outputs through\ntoken-level optimization, also encounters several limitations: requiring\nwhite-box access, necessitating pre-constructed affirmative phrase, and\nsuffering from low efficiency. In this paper, we present ECLIPSE, a novel and\nefficient black-box jailbreaking method utilizing optimizable suffixes. Drawing\ninspiration from LLMs' powerful generation and optimization capabilities, we\nemploy task prompts to translate jailbreaking goals into natural language\ninstructions. This guides the LLM to generate adversarial suffixes for\nmalicious queries. In particular, a harmfulness scorer provides continuous\nfeedback, enabling LLM self-reflection and iterative optimization to\nautonomously and efficiently produce effective suffixes. Experimental results\ndemonstrate that ECLIPSE achieves an average attack success rate (ASR) of 0.92\nacross three open-source LLMs and GPT-3.5-Turbo, significantly surpassing GCG\nin 2.4 times. Moreover, ECLIPSE is on par with template-based methods in ASR\nwhile offering superior attack efficiency, reducing the average attack overhead\nby 83%.\n", "date": "2024-08-21", "categories": []}
{"id": "2408.16807", "title": "STEREO: A Two-Stage Framework for Adversarially Robust Concept Erasing\n  from Text-to-Image Diffusion Models", "abstract": "  The rapid proliferation of large-scale text-to-image diffusion (T2ID) models\nhas raised serious concerns about their potential misuse in generating harmful\ncontent. Although numerous methods have been proposed for erasing undesired\nconcepts from T2ID models, they often provide a false sense of security;\nconcept-erased models (CEMs) can still be manipulated via adversarial attacks\nto regenerate the erased concept. While a few robust concept erasure methods\nbased on adversarial training have emerged recently, they compromise on utility\n(generation quality for benign concepts) to achieve robustness and/or remain\nvulnerable to advanced embedding space attacks. These limitations stem from the\nfailure of robust CEMs to thoroughly search for \"blind spots\" in the embedding\nspace. To bridge this gap, we propose STEREO, a novel two-stage framework that\nemploys adversarial training as a first step rather than the only step for\nrobust concept erasure. In the first stage, STEREO employs adversarial training\nas a vulnerability identification mechanism to search thoroughly enough. In the\nsecond robustly erase once stage, STEREO introduces an anchor-concept-based\ncompositional objective to robustly erase the target concept in a single\nfine-tuning stage, while minimizing the degradation of model utility. We\nbenchmark STEREO against seven state-of-the-art concept erasure methods,\ndemonstrating its superior robustness to both white-box and black-box attacks,\nwhile largely preserving utility.\n", "date": "2024-08-29", "categories": []}
{"id": "2504.01228", "title": "TenAd: A Tensor-based Low-rank Black Box Adversarial Attack for Video\n  Classification", "abstract": "  Deep learning models have achieved remarkable success in computer vision but\nremain vulnerable to adversarial attacks, particularly in black-box settings\nwhere model details are unknown. Existing adversarial attack methods(even those\nworks with key frames) often treat video data as simple vectors, ignoring their\ninherent multi-dimensional structure, and require a large number of queries,\nmaking them inefficient and detectable. In this paper, we propose\n\\textbf{TenAd}, a novel tensor-based low-rank adversarial attack that leverages\nthe multi-dimensional properties of video data by representing videos as\nfourth-order tensors. By exploiting low-rank attack, our method significantly\nreduces the search space and the number of queries needed to generate\nadversarial examples in black-box settings. Experimental results on standard\nvideo classification datasets demonstrate that \\textbf{TenAd} effectively\ngenerates imperceptible adversarial perturbations while achieving higher attack\nsuccess rates and query efficiency compared to state-of-the-art methods. Our\napproach outperforms existing black-box adversarial attacks in terms of success\nrate, query efficiency, and perturbation imperceptibility, highlighting the\npotential of tensor-based methods for adversarial attacks on video models.\n", "date": "2025-04-01", "categories": []}
{"id": "2504.01345", "title": "Breaking BERT: Gradient Attack on Twitter Sentiment Analysis for\n  Targeted Misclassification", "abstract": "  Social media platforms like Twitter have increasingly relied on Natural\nLanguage Processing NLP techniques to analyze and understand the sentiments\nexpressed in the user generated content. One such state of the art NLP model is\nBidirectional Encoder Representations from Transformers BERT which has been\nwidely adapted in sentiment analysis. BERT is susceptible to adversarial\nattacks. This paper aims to scrutinize the inherent vulnerabilities of such\nmodels in Twitter sentiment analysis. It aims to formulate a framework for\nconstructing targeted adversarial texts capable of deceiving these models,\nwhile maintaining stealth. In contrast to conventional methodologies, such as\nImportance Reweighting, this framework core idea resides in its reliance on\ngradients to prioritize the importance of individual words within the text. It\nuses a whitebox approach to attain fine grained sensitivity, pinpointing words\nthat exert maximal influence on the classification outcome. This paper is\norganized into three interdependent phases. It starts with fine-tuning a\npre-trained BERT model on Twitter data. It then analyzes gradients of the model\nto rank words on their importance, and iteratively replaces those with feasible\ncandidates until an acceptable solution is found. Finally, it evaluates the\neffectiveness of the adversarial text against the custom trained sentiment\nclassification model. This assessment would help in gauging the capacity of the\nadversarial text to successfully subvert classification without raising any\nalarm.\n", "date": "2025-04-02", "categories": []}
{"id": "2504.01399", "title": "Leveraging Generalizability of Image-to-Image Translation for Enhanced\n  Adversarial Defense", "abstract": "  In the rapidly evolving field of artificial intelligence, machine learning\nemerges as a key technology characterized by its vast potential and inherent\nrisks. The stability and reliability of these models are important, as they are\nfrequent targets of security threats. Adversarial attacks, first rigorously\ndefined by Ian Goodfellow et al. in 2013, highlight a critical vulnerability:\nthey can trick machine learning models into making incorrect predictions by\napplying nearly invisible perturbations to images. Although many studies have\nfocused on constructing sophisticated defensive mechanisms to mitigate such\nattacks, they often overlook the substantial time and computational costs of\ntraining and maintaining these models. Ideally, a defense method should be able\nto generalize across various, even unseen, adversarial attacks with minimal\noverhead. Building on our previous work on image-to-image translation-based\ndefenses, this study introduces an improved model that incorporates residual\nblocks to enhance generalizability. The proposed method requires training only\na single model, effectively defends against diverse attack types, and is\nwell-transferable between different target models. Experiments show that our\nmodel can restore the classification accuracy from near zero to an average of\n72\\% while maintaining competitive performance compared to state-of-the-art\nmethods.\n", "date": "2025-04-02", "categories": []}
{"id": "2504.01735", "title": "AdPO: Enhancing the Adversarial Robustness of Large Vision-Language\n  Models with Preference Optimization", "abstract": "  Large Vision-Language Models (LVLMs), such as GPT-4o and LLaVA, have recently\nwitnessed remarkable advancements and are increasingly being deployed in\nreal-world applications. However, inheriting the sensitivity of visual neural\nnetworks, LVLMs remain vulnerable to adversarial attacks, which can result in\nerroneous or malicious outputs. While existing efforts utilize adversarial\nfine-tuning to enhance robustness, they often suffer from performance\ndegradation on clean inputs. In this paper, we proposes AdPO, a novel\nadversarial defense strategy for LVLMs based on preference optimization. For\nthe first time, we reframe adversarial training as a preference optimization\nproblem, aiming to enhance the model's preference for generating normal outputs\non clean inputs while rejecting the potential misleading outputs for\nadversarial examples. Notably, AdPO achieves this by solely modifying the image\nencoder, e.g., CLIP ViT, resulting in superior clean and adversarial\nperformance in a variety of downsream tasks. Considering that training involves\nlarge language models (LLMs), the computational cost increases significantly.\nWe validate that training on smaller LVLMs and subsequently transferring to\nlarger models can achieve competitive performance while maintaining efficiency\ncomparable to baseline methods. Our comprehensive experiments confirm the\neffectiveness of the proposed AdPO, which provides a novel perspective for\nfuture adversarial defense research.\n", "date": "2025-04-02", "categories": []}
{"id": "2504.01819", "title": "Implicit Bias Injection Attacks against Text-to-Image Diffusion Models", "abstract": "  The proliferation of text-to-image diffusion models (T2I DMs) has led to an\nincreased presence of AI-generated images in daily life. However, biased T2I\nmodels can generate content with specific tendencies, potentially influencing\npeople's perceptions. Intentional exploitation of these biases risks conveying\nmisleading information to the public. Current research on bias primarily\naddresses explicit biases with recognizable visual patterns, such as skin color\nand gender. This paper introduces a novel form of implicit bias that lacks\nexplicit visual features but can manifest in diverse ways across various\nsemantic contexts. This subtle and versatile nature makes this bias challenging\nto detect, easy to propagate, and adaptable to a wide range of scenarios. We\nfurther propose an implicit bias injection attack framework (IBI-Attacks)\nagainst T2I diffusion models by precomputing a general bias direction in the\nprompt embedding space and adaptively adjusting it based on different inputs.\nOur attack module can be seamlessly integrated into pre-trained diffusion\nmodels in a plug-and-play manner without direct manipulation of user input or\nmodel retraining. Extensive experiments validate the effectiveness of our\nscheme in introducing bias through subtle and diverse modifications while\npreserving the original semantics. The strong concealment and transferability\nof our attack across various scenarios further underscore the significance of\nour approach. Code is available at https://github.com/Hannah1102/IBI-attacks.\n", "date": "2025-04-02", "categories": []}
{"id": "2407.08678", "title": "How to beat a Bayesian adversary", "abstract": "  Deep neural networks and other modern machine learning models are often\nsusceptible to adversarial attacks. Indeed, an adversary may often be able to\nchange a model's prediction through a small, directed perturbation of the\nmodel's input - an issue in safety-critical applications. Adversarially robust\nmachine learning is usually based on a minmax optimisation problem that\nminimises the machine learning loss under maximisation-based adversarial\nattacks.\n  In this work, we study adversaries that determine their attack using a\nBayesian statistical approach rather than maximisation. The resulting Bayesian\nadversarial robustness problem is a relaxation of the usual minmax problem. To\nsolve this problem, we propose Abram - a continuous-time particle system that\nshall approximate the gradient flow corresponding to the underlying learning\nproblem. We show that Abram approximates a McKean-Vlasov process and justify\nthe use of Abram by giving assumptions under which the McKean-Vlasov process\nfinds the minimiser of the Bayesian adversarial robustness problem. We discuss\ntwo ways to discretise Abram and show its suitability in benchmark adversarial\ndeep learning experiments.\n", "date": "2024-07-11", "categories": []}
{"id": "2501.18098", "title": "Disentangling Safe and Unsafe Corruptions via Anisotropy and Locality", "abstract": "  State-of-the-art machine learning systems are vulnerable to small\nperturbations to their input, where ``small'' is defined according to a threat\nmodel that assigns a positive threat to each perturbation. Most prior works\ndefine a task-agnostic, isotropic, and global threat, like the $\\ell_p$ norm,\nwhere the magnitude of the perturbation fully determines the degree of the\nthreat and neither the direction of the attack nor its position in space\nmatter. However, common corruptions in computer vision, such as blur,\ncompression, or occlusions, are not well captured by such threat models. This\npaper proposes a novel threat model called \\texttt{Projected Displacement} (PD)\nto study robustness beyond existing isotropic and global threat models. The\nproposed threat model measures the threat of a perturbation via its alignment\nwith \\textit{unsafe directions}, defined as directions in the input space along\nwhich a perturbation of sufficient magnitude changes the ground truth class\nlabel. Unsafe directions are identified locally for each input based on\nobserved training data. In this way, the PD threat model exhibits anisotropy\nand locality. Experiments on Imagenet-1k data indicate that, for any input, the\nset of perturbations with small PD threat includes \\textit{safe} perturbations\nof large $\\ell_p$ norm that preserve the true label, such as noise, blur and\ncompression, while simultaneously excluding \\textit{unsafe} perturbations that\nalter the true label. Unlike perceptual threat models based on embeddings of\nlarge-vision models, the PD threat model can be readily computed for arbitrary\nclassification tasks without pre-training or finetuning. Further additional\ntask annotation such as sensitivity to image regions or concept hierarchies can\nbe easily integrated into the assessment of threat and thus the PD threat model\npresents practitioners with a flexible, task-driven threat specification.\n", "date": "2025-01-30", "categories": []}
{"id": "2504.00038", "title": "Revisiting the Relationship between Adversarial and Clean Training: Why\n  Clean Training Can Make Adversarial Training Better", "abstract": "  Adversarial training (AT) is an effective technique for enhancing adversarial\nrobustness, but it usually comes at the cost of a decline in generalization\nability. Recent studies have attempted to use clean training to assist\nadversarial training, yet there are contradictions among the conclusions. We\ncomprehensively summarize the representative strategies and, with a focus on\nthe multi - view hypothesis, provide a unified explanation for the\ncontradictory phenomena among different studies. In addition, we conduct an in\n- depth analysis of the knowledge combinations transferred from clean - trained\nmodels to adversarially - trained models in previous studies, and find that\nthey can be divided into two categories: reducing the learning difficulty and\nproviding correct guidance. Based on this finding, we propose a new idea of\nleveraging clean training to further improve the performance of advanced AT\nmethods.We reveal that the problem of generalization degradation faced by AT\npartly stems from the difficulty of adversarial training in learning certain\nsample features, and this problem can be alleviated by making full use of clean\ntraining.\n", "date": "2025-03-30", "categories": []}
{"id": "2504.00429", "title": "Unleashing the Power of Pre-trained Encoders for Universal Adversarial\n  Attack Detection", "abstract": "  Adversarial attacks pose a critical security threat to real-world AI systems\nby injecting human-imperceptible perturbations into benign samples to induce\nmisclassification in deep learning models. While existing detection methods,\nsuch as Bayesian uncertainty estimation and activation pattern analysis, have\nachieved progress through feature engineering, their reliance on handcrafted\nfeature design and prior knowledge of attack patterns limits generalization\ncapabilities and incurs high engineering costs. To address these limitations,\nthis paper proposes a lightweight adversarial detection framework based on the\nlarge-scale pre-trained vision-language model CLIP. Departing from conventional\nadversarial feature characterization paradigms, we innovatively adopt an\nanomaly detection perspective. By jointly fine-tuning CLIP's dual visual-text\nencoders with trainable adapter networks and learnable prompts, we construct a\ncompact representation space tailored for natural images. Notably, our\ndetection architecture achieves substantial improvements in generalization\ncapability across both known and unknown attack patterns compared to\ntraditional methods, while significantly reducing training overhead. This study\nprovides a novel technical pathway for establishing a parameter-efficient and\nattack-agnostic defense paradigm, markedly enhancing the robustness of vision\nsystems against evolving adversarial threats.\n", "date": "2025-04-01", "categories": []}
{"id": "2504.00721", "title": "Alleviating Performance Disparity in Adversarial Spatiotemporal Graph\n  Learning Under Zero-Inflated Distribution", "abstract": "  Spatiotemporal Graph Learning (SGL) under Zero-Inflated Distribution (ZID) is\ncrucial for urban risk management tasks, including crime prediction and traffic\naccident profiling. However, SGL models are vulnerable to adversarial attacks,\ncompromising their practical utility. While adversarial training (AT) has been\nwidely used to bolster model robustness, our study finds that traditional AT\nexacerbates performance disparities between majority and minority classes under\nZID, potentially leading to irreparable losses due to underreporting critical\nrisk events. In this paper, we first demonstrate the smaller top-k gradients\nand lower separability of minority class are key factors contributing to this\ndisparity. To address these issues, we propose MinGRE, a framework for Minority\nClass Gradients and Representations Enhancement. MinGRE employs a\nmulti-dimensional attention mechanism to reweight spatiotemporal gradients,\nminimizing the gradient distribution discrepancies across classes.\nAdditionally, we introduce an uncertainty-guided contrastive loss to improve\nthe inter-class separability and intra-class compactness of minority\nrepresentations with higher uncertainty. Extensive experiments demonstrate that\nthe MinGRE framework not only significantly reduces the performance disparity\nacross classes but also achieves enhanced robustness compared to existing\nbaselines. These findings underscore the potential of our method in fostering\nthe development of more equitable and robust models.\n", "date": "2025-04-01", "categories": []}
{"id": "2402.07498", "title": "Accelerated Smoothing: A Scalable Approach to Randomized Smoothing", "abstract": "  Randomized smoothing has emerged as a potent certifiable defense against\nadversarial attacks by employing smoothing noises from specific distributions\nto ensure the robustness of a smoothed classifier. However, the utilization of\nMonte Carlo sampling in this process introduces a compute-intensive element,\nwhich constrains the practicality of randomized smoothing on a larger scale. To\naddress this limitation, we propose a novel approach that replaces Monte Carlo\nsampling with the training of a surrogate neural network. Through extensive\nexperimentation in various settings, we demonstrate the efficacy of our\napproach in approximating the smoothed classifier with remarkable precision.\nFurthermore, we demonstrate that our approach significantly accelerates the\nrobust radius certification process, providing nearly $600$X improvement in\ncomputation time, overcoming the computational bottlenecks associated with\ntraditional randomized smoothing.\n", "date": "2024-02-12", "categories": []}
{"id": "2407.11025", "title": "Backdoor Graph Condensation", "abstract": "  Graph condensation has recently emerged as a prevalent technique to improve\nthe training efficiency for graph neural networks (GNNs). It condenses a large\ngraph into a small one such that a GNN trained on this small synthetic graph\ncan achieve comparable performance to a GNN trained on the large graph.\nHowever, while existing graph condensation studies mainly focus on the best\ntrade-off between graph size and the GNNs' performance (model utility), they\noverlook the security issues of graph condensation. To bridge this gap, we\nfirst explore backdoor attack against the GNNs trained on the condensed graphs.\n  We introduce an effective backdoor attack against graph condensation, termed\nBGC. This attack aims to (1) preserve the condensed graph quality despite\ntrigger injection, and (2) ensure trigger efficacy through the condensation\nprocess, achieving a high attack success rate. Specifically, BGC consistently\nupdates triggers during condensation and targets representative nodes for\npoisoning. Extensive experiments demonstrate the effectiveness of our attack.\nBGC achieves a high attack success rate (close to 1.0) and good model utility\nin all cases. Furthermore, the results against multiple defense methods\ndemonstrate BGC's resilience under their defenses. Finally, we analyze the key\nhyperparameters that influence the attack performance. Our code is available\nat: https://github.com/JiahaoWuGit/BGC.\n", "date": "2024-07-03", "categories": []}
{"id": "2407.13068", "title": "Krait: A Backdoor Attack Against Graph Prompt Tuning", "abstract": "  Graph prompt tuning has emerged as a promising paradigm to effectively\ntransfer general graph knowledge from pre-trained models to various downstream\ntasks, particularly in few-shot contexts. However, its susceptibility to\nbackdoor attacks, where adversaries insert triggers to manipulate outcomes,\nraises a critical concern. We conduct the first study to investigate such\nvulnerability, revealing that backdoors can disguise benign graph prompts, thus\nevading detection. We introduce Krait, a novel graph prompt backdoor.\nSpecifically, we propose a simple yet effective model-agnostic metric called\nlabel non-uniformity homophily to select poisoned candidates, significantly\nreducing computational complexity. To accommodate diverse attack scenarios and\nadvanced attack types, we design three customizable trigger generation methods\nto craft prompts as triggers. We propose a novel centroid similarity-based loss\nfunction to optimize prompt tuning for attack effectiveness and stealthiness.\nExperiments on four real-world graphs demonstrate that Krait can efficiently\nembed triggers to merely 0.15% to 2% of training nodes, achieving high attack\nsuccess rates without sacrificing clean accuracy. Notably, in one-to-one and\nall-to-one attacks, Krait can achieve 100% attack success rates by poisoning as\nfew as 2 and 22 nodes, respectively. Our experiments further show that Krait\nremains potent across different transfer cases, attack types, and graph neural\nnetwork backbones. Additionally, Krait can be successfully extended to the\nblack-box setting, posing more severe threats. Finally, we analyze why Krait\ncan evade both classical and state-of-the-art defenses, and provide practical\ninsights for detecting and mitigating this class of attacks.\n", "date": "2024-07-18", "categories": []}
{"id": "2411.15555", "title": "Improving the Transferability of Adversarial Attacks on Face Recognition\n  with Diverse Parameters Augmentation", "abstract": "  Face Recognition (FR) models are vulnerable to adversarial examples that\nsubtly manipulate benign face images, underscoring the urgent need to improve\nthe transferability of adversarial attacks in order to expose the blind spots\nof these systems. Existing adversarial attack methods often overlook the\npotential benefits of augmenting the surrogate model with diverse\ninitializations, which limits the transferability of the generated adversarial\nexamples. To address this gap, we propose a novel method called Diverse\nParameters Augmentation (DPA) attack method, which enhances surrogate models by\nincorporating diverse parameter initializations, resulting in a broader and\nmore diverse set of surrogate models. Specifically, DPA consists of two key\nstages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA).\nIn the DPO stage, we initialize the parameters of the surrogate model using\nboth pre-trained and random parameters. Subsequently, we save the models in the\nintermediate training process to obtain a diverse set of surrogate models.\nDuring the HMA stage, we enhance the feature maps of the diversified surrogate\nmodels by incorporating beneficial perturbations, thereby further improving the\ntransferability. Experimental results demonstrate that our proposed attack\nmethod can effectively enhance the transferability of the crafted adversarial\nface examples.\n", "date": "2024-11-23", "categories": []}
{"id": "2501.11901", "title": "Enhancing Adversarial Transferability via Component-Wise Transformation", "abstract": "  Deep Neural Networks (DNNs) are highly vulnerable to adversarial examples,\nwhich pose significant challenges in security-sensitive applications. Among\nvarious adversarial attack strategies, input transformation-based attacks have\ndemonstrated remarkable effectiveness in enhancing adversarial transferability.\nHowever, existing methods still perform poorly across different architectures,\neven though they have achieved promising results within the same architecture.\nThis limitation arises because, while models of the same architecture may focus\non different regions of the object, the variation is even more pronounced\nacross different architectures. Unfortunately, current approaches fail to\neffectively guide models to attend to these diverse regions. To address this\nissue, this paper proposes a novel input transformation-based attack method,\ntermed Component-Wise Transformation (CWT). CWT applies interpolation and\nselective rotation to individual image blocks, ensuring that each transformed\nimage highlights different target regions, thereby improving the\ntransferability of adversarial examples. Extensive experiments on the standard\nImageNet dataset show that CWT consistently outperforms state-of-the-art\nmethods in both attack success rates and stability across CNN- and\nTransformer-based models.\n", "date": "2025-01-21", "categories": []}
{"id": "2502.21048", "title": "Data-free Universal Adversarial Perturbation with Pseudo-semantic Prior", "abstract": "  Data-free Universal Adversarial Perturbation (UAP) is an image-agnostic\nadversarial attack that deceives deep neural networks using a single\nperturbation generated solely from random noise without relying on data priors.\nHowever, traditional data-free UAP methods often suffer from limited\ntransferability due to the absence of semantic content in random noise. To\naddress this issue, we propose a novel data-free universal attack method that\nrecursively extracts pseudo-semantic priors directly from the UAPs during\ntraining to enrich the semantic content within the data-free UAP framework. Our\napproach effectively leverages latent semantic information within UAPs via\nregion sampling, enabling successful input transformations-typically\nineffective in traditional data-free UAP methods due to the lack of semantic\ncues-and significantly enhancing black-box transferability. Furthermore, we\nintroduce a sample reweighting technique to mitigate potential imbalances from\nrandom sampling and transformations, emphasizing hard examples less affected by\nthe UAPs. Comprehensive experiments on ImageNet show that our method achieves\nstate-of-the-art performance in average fooling rate by a substantial margin,\nnotably improves attack transferability across various CNN architectures\ncompared to existing data-free UAP methods, and even surpasses data-dependent\nUAP methods. Code is available at: https://github.com/ChnanChan/PSP-UAP.\n", "date": "2025-02-28", "categories": []}
{"id": "2503.22759", "title": "Data Poisoning in Deep Learning: A Survey", "abstract": "  Deep learning has become a cornerstone of modern artificial intelligence,\nenabling transformative applications across a wide range of domains. As the\ncore element of deep learning, the quality and security of training data\ncritically influence model performance and reliability. However, during the\ntraining process, deep learning models face the significant threat of data\npoisoning, where attackers introduce maliciously manipulated training data to\ndegrade model accuracy or lead to anomalous behavior. While existing surveys\nprovide valuable insights into data poisoning, they generally adopt a broad\nperspective, encompassing both attacks and defenses, but lack a dedicated,\nin-depth analysis of poisoning attacks specifically in deep learning. In this\nsurvey, we bridge this gap by presenting a comprehensive and targeted review of\ndata poisoning in deep learning. First, this survey categorizes data poisoning\nattacks across multiple perspectives, providing an in-depth analysis of their\ncharacteristics and underlying design princinples. Second, the discussion is\nextended to the emerging area of data poisoning in large language models(LLMs).\nFinally, we explore critical open challenges in the field and propose potential\nresearch directions to advance the field further. To support further\nexploration, an up-to-date repository of resources on data poisoning in deep\nlearning is available at https://github.com/Pinlong-Zhao/Data-Poisoning.\n", "date": "2025-03-27", "categories": []}
{"id": "2502.19820", "title": "Foot-In-The-Door: A Multi-turn Jailbreak for LLMs", "abstract": "  Ensuring AI safety is crucial as large language models become increasingly\nintegrated into real-world applications. A key challenge is jailbreak, where\nadversarial prompts bypass built-in safeguards to elicit harmful disallowed\noutputs. Inspired by psychological foot-in-the-door principles, we introduce\nFITD,a novel multi-turn jailbreak method that leverages the phenomenon where\nminor initial commitments lower resistance to more significant or more\nunethical transgressions. Our approach progressively escalates the malicious\nintent of user queries through intermediate bridge prompts and aligns the\nmodel's response by itself to induce toxic responses. Extensive experimental\nresults on two jailbreak benchmarks demonstrate that FITD achieves an average\nattack success rate of 94% across seven widely used models, outperforming\nexisting state-of-the-art methods. Additionally, we provide an in-depth\nanalysis of LLM self-corruption, highlighting vulnerabilities in current\nalignment strategies and emphasizing the risks inherent in multi-turn\ninteractions. The code is available at\nhttps://github.com/Jinxiaolong1129/Foot-in-the-door-Jailbreak.\n", "date": "2025-02-27", "categories": []}
{"id": "2503.22205", "title": "Data-Free Universal Attack by Exploiting the Intrinsic Vulnerability of\n  Deep Models", "abstract": "  Deep neural networks (DNNs) are susceptible to Universal Adversarial\nPerturbations (UAPs), which are instance agnostic perturbations that can\ndeceive a target model across a wide range of samples. Unlike instance-specific\nadversarial examples, UAPs present a greater challenge as they must generalize\nacross different samples and models. Generating UAPs typically requires access\nto numerous examples, which is a strong assumption in real-world tasks. In this\npaper, we propose a novel data-free method called Intrinsic UAP (IntriUAP), by\nexploiting the intrinsic vulnerabilities of deep models. We analyze a series of\npopular deep models composed of linear and nonlinear layers with a Lipschitz\nconstant of 1, revealing that the vulnerability of these models is\npredominantly influenced by their linear components. Based on this observation,\nwe leverage the ill-conditioned nature of the linear components by aligning the\nUAP with the right singular vectors corresponding to the maximum singular value\nof each linear layer. Remarkably, our method achieves highly competitive\nperformance in attacking popular image classification deep models without using\nany image samples. We also evaluate the black-box attack performance of our\nmethod, showing that it matches the state-of-the-art baseline for data-free\nmethods on models that conform to our theoretical framework. Beyond the\ndata-free assumption, IntriUAP also operates under a weaker assumption, where\nthe adversary only can access a few of the victim model's layers. Experiments\ndemonstrate that the attack success rate decreases by only 4% when the\nadversary has access to just 50% of the linear layers in the victim model.\n", "date": "2025-03-28", "categories": []}
{"id": "2503.22653", "title": "Tropical Bisectors and Carlini-Wagner Attacks", "abstract": "  Pasque et al. showed that using a tropical symmetric metric as an activation\nfunction in the last layer can improve the robustness of convolutional neural\nnetworks (CNNs) against state-of-the-art attacks, including the Carlini-Wagner\nattack. This improvement occurs when the attacks are not specifically adapted\nto the non-differentiability of the tropical layer. Moreover, they showed that\nthe decision boundary of a tropical CNN is defined by tropical bisectors. In\nthis paper, we explore the combinatorics of tropical bisectors and analyze how\nthe tropical embedding layer enhances robustness against Carlini-Wagner\nattacks. We prove an upper bound on the number of linear segments the decision\nboundary of a tropical CNN can have. We then propose a refined version of the\nCarlini-Wagner attack, specifically tailored for the tropical architecture.\nComputational experiments with MNIST and LeNet5 showcase our attacks improved\nsuccess rate.\n", "date": "2025-03-28", "categories": []}
{"id": "2410.16449", "title": "Robust Feature Learning for Multi-Index Models in High Dimensions", "abstract": "  Recently, there have been numerous studies on feature learning with neural\nnetworks, specifically on learning single- and multi-index models where the\ntarget is a function of a low-dimensional projection of the input. Prior works\nhave shown that in high dimensions, the majority of the compute and data\nresources are spent on recovering the low-dimensional projection; once this\nsubspace is recovered, the remainder of the target can be learned independently\nof the ambient dimension. However, implications of feature learning in\nadversarial settings remain unexplored. In this work, we take the first steps\ntowards understanding adversarially robust feature learning with neural\nnetworks. Specifically, we prove that the hidden directions of a multi-index\nmodel offer a Bayes optimal low-dimensional projection for robustness against\n$\\ell_2$-bounded adversarial perturbations under the squared loss, assuming\nthat the multi-index coordinates are statistically independent from the rest of\nthe coordinates. Therefore, robust learning can be achieved by first performing\nstandard feature learning, then robustly tuning a linear readout layer on top\nof the standard representations. In particular, we show that adversarially\nrobust learning is just as easy as standard learning. Specifically, the\nadditional number of samples needed to robustly learn multi-index models when\ncompared to standard learning does not depend on dimensionality.\n", "date": "2024-10-21", "categories": []}
{"id": "2503.20844", "title": "Robust Deep Reinforcement Learning in Robotics via Adaptive\n  Gradient-Masked Adversarial Attacks", "abstract": "  Deep reinforcement learning (DRL) has emerged as a promising approach for\nrobotic control, but its realworld deployment remains challenging due to its\nvulnerability to environmental perturbations. Existing white-box adversarial\nattack methods, adapted from supervised learning, fail to effectively target\nDRL agents as they overlook temporal dynamics and indiscriminately perturb all\nstate dimensions, limiting their impact on long-term rewards. To address these\nchallenges, we propose the Adaptive Gradient-Masked Reinforcement (AGMR)\nAttack, a white-box attack method that combines DRL with a gradient-based soft\nmasking mechanism to dynamically identify critical state dimensions and\noptimize adversarial policies. AGMR selectively allocates perturbations to the\nmost impactful state features and incorporates a dynamic adjustment mechanism\nto balance exploration and exploitation during training. Extensive experiments\ndemonstrate that AGMR outperforms state-of-the-art adversarial attack methods\nin degrading the performance of the victim agent and enhances the victim\nagent's robustness through adversarial defense mechanisms.\n", "date": "2025-03-26", "categories": []}
{"id": "2503.21236", "title": "Clean Image May be Dangerous: Data Poisoning Attacks Against Deep\n  Hashing", "abstract": "  Large-scale image retrieval using deep hashing has become increasingly\npopular due to the exponential growth of image data and the remarkable feature\nextraction capabilities of deep neural networks (DNNs). However, deep hashing\nmethods are vulnerable to malicious attacks, including adversarial and backdoor\nattacks. It is worth noting that these attacks typically involve altering the\nquery images, which is not a practical concern in real-world scenarios. In this\npaper, we point out that even clean query images can be dangerous, inducing\nmalicious target retrieval results, like undesired or illegal images. To the\nbest of our knowledge, we are the first to study data \\textbf{p}oisoning\n\\textbf{a}ttacks against \\textbf{d}eep \\textbf{hash}ing\n\\textbf{(\\textit{PADHASH})}. Specifically, we first train a surrogate model to\nsimulate the behavior of the target deep hashing model. Then, a strict gradient\nmatching strategy is proposed to generate the poisoned images. Extensive\nexperiments on different models, datasets, hash methods, and hash code lengths\ndemonstrate the effectiveness and generality of our attack method.\n", "date": "2025-03-27", "categories": []}
{"id": "2503.21305", "title": "DeBackdoor: A Deductive Framework for Detecting Backdoor Attacks on Deep\n  Models with Limited Data", "abstract": "  Backdoor attacks are among the most effective, practical, and stealthy\nattacks in deep learning. In this paper, we consider a practical scenario where\na developer obtains a deep model from a third party and uses it as part of a\nsafety-critical system. The developer wants to inspect the model for potential\nbackdoors prior to system deployment. We find that most existing detection\ntechniques make assumptions that are not applicable to this scenario. In this\npaper, we present a novel framework for detecting backdoors under realistic\nrestrictions. We generate candidate triggers by deductively searching over the\nspace of possible triggers. We construct and optimize a smoothed version of\nAttack Success Rate as our search objective. Starting from a broad class of\ntemplate attacks and just using the forward pass of a deep model, we reverse\nengineer the backdoor attack. We conduct extensive evaluation on a wide range\nof attacks, models, and datasets, with our technique performing almost\nperfectly across these settings.\n", "date": "2025-03-27", "categories": []}
{"id": "2503.21315", "title": "Tricking Retrievers with Influential Tokens: An Efficient Black-Box\n  Corpus Poisoning Attack", "abstract": "  Retrieval-augmented generation (RAG) systems enhance large language models by\nincorporating external knowledge, addressing issues like outdated internal\nknowledge and hallucination. However, their reliance on external knowledge\nbases makes them vulnerable to corpus poisoning attacks, where adversarial\npassages can be injected to manipulate retrieval results. Existing methods for\ncrafting such passages, such as random token replacement or training inversion\nmodels, are often slow and computationally expensive, requiring either access\nto retriever's gradients or large computational resources. To address these\nlimitations, we propose Dynamic Importance-Guided Genetic Algorithm (DIGA), an\nefficient black-box method that leverages two key properties of retrievers:\ninsensitivity to token order and bias towards influential tokens. By focusing\non these characteristics, DIGA dynamically adjusts its genetic operations to\ngenerate effective adversarial passages with significantly reduced time and\nmemory usage. Our experimental evaluation shows that DIGA achieves superior\nefficiency and scalability compared to existing methods, while maintaining\ncomparable or better attack success rates across multiple datasets.\n", "date": "2025-03-27", "categories": []}
{"id": "2310.00116", "title": "Certified Robustness via Dynamic Margin Maximization and Improved\n  Lipschitz Regularization", "abstract": "  To improve the robustness of deep classifiers against adversarial\nperturbations, many approaches have been proposed, such as designing new\narchitectures with better robustness properties (e.g., Lipschitz-capped\nnetworks), or modifying the training process itself (e.g., min-max\noptimization, constrained learning, or regularization). These approaches,\nhowever, might not be effective at increasing the margin in the input (feature)\nspace. As a result, there has been an increasing interest in developing\ntraining procedures that can directly manipulate the decision boundary in the\ninput space. In this paper, we build upon recent developments in this category\nby developing a robust training algorithm whose objective is to increase the\nmargin in the output (logit) space while regularizing the Lipschitz constant of\nthe model along vulnerable directions. We show that these two objectives can\ndirectly promote larger margins in the input space. To this end, we develop a\nscalable method for calculating guaranteed differentiable upper bounds on the\nLipschitz constant of neural networks accurately and efficiently. The relative\naccuracy of the bounds prevents excessive regularization and allows for more\ndirect manipulation of the decision boundary. Furthermore, our Lipschitz\nbounding algorithm exploits the monotonicity and Lipschitz continuity of the\nactivation layers, and the resulting bounds can be used to design new layers\nwith controllable bounds on their Lipschitz constant. Experiments on the MNIST,\nCIFAR-10, and Tiny-ImageNet data sets verify that our proposed algorithm\nobtains competitively improved results compared to the state-of-the-art.\n", "date": "2023-09-29", "categories": []}
{"id": "2502.18290", "title": "Stealthy Backdoor Attack in Self-Supervised Learning Vision Encoders for\n  Large Vision Language Models", "abstract": "  Self-supervised learning (SSL) vision encoders learn high-quality image\nrepresentations and thus have become a vital part of developing vision modality\nof large vision language models (LVLMs). Due to the high cost of training such\nencoders, pre-trained encoders are widely shared and deployed into many LVLMs,\nwhich are security-critical or bear societal significance. Under this practical\nscenario, we reveal a new backdoor threat that significant visual\nhallucinations can be induced into these LVLMs by merely compromising vision\nencoders. Because of the sharing and reuse of these encoders, many downstream\nLVLMs may inherit backdoor behaviors from encoders, leading to widespread\nbackdoors. In this work, we propose BadVision, the first method to exploit this\nvulnerability in SSL vision encoders for LVLMs with novel trigger optimization\nand backdoor learning techniques. We evaluate BadVision on two types of SSL\nencoders and LVLMs across eight benchmarks. We show that BadVision effectively\ndrives the LVLMs to attacker-chosen hallucination with over 99% attack success\nrate, causing a 77.6% relative visual understanding error while maintaining the\nstealthiness. SoTA backdoor detection methods cannot detect our attack\neffectively.\n", "date": "2025-02-25", "categories": []}
{"id": "2503.20310", "title": "Enabling Heterogeneous Adversarial Transferability via Feature\n  Permutation Attacks", "abstract": "  Adversarial attacks in black-box settings are highly practical, with\ntransfer-based attacks being the most effective at generating adversarial\nexamples (AEs) that transfer from surrogate models to unseen target models.\nHowever, their performance significantly degrades when transferring across\nheterogeneous architectures -- such as CNNs, MLPs, and Vision Transformers\n(ViTs) -- due to fundamental architectural differences. To address this, we\npropose Feature Permutation Attack (FPA), a zero-FLOP, parameter-free method\nthat enhances adversarial transferability across diverse architectures. FPA\nintroduces a novel feature permutation (FP) operation, which rearranges pixel\nvalues in selected feature maps to simulate long-range dependencies,\neffectively making CNNs behave more like ViTs and MLPs. This enhances feature\ndiversity and improves transferability both across heterogeneous architectures\nand within homogeneous CNNs. Extensive evaluations on 14 state-of-the-art\narchitectures show that FPA achieves maximum absolute gains in attack success\nrates of 7.68% on CNNs, 14.57% on ViTs, and 14.48% on MLPs, outperforming\nexisting black-box attacks. Additionally, FPA is highly generalizable and can\nseamlessly integrate with other transfer-based attacks to further boost their\nperformance. Our findings establish FPA as a robust, efficient, and\ncomputationally lightweight strategy for enhancing adversarial transferability\nacross heterogeneous architectures.\n", "date": "2025-03-26", "categories": []}
{"id": "2503.20454", "title": "Lipschitz Constant Meets Condition Number: Learning Robust and Compact\n  Deep Neural Networks", "abstract": "  Recent research has revealed that high compression of Deep Neural Networks\n(DNNs), e.g., massive pruning of the weight matrix of a DNN, leads to a severe\ndrop in accuracy and susceptibility to adversarial attacks. Integration of\nnetwork pruning into an adversarial training framework has been proposed to\npromote adversarial robustness. It has been observed that a highly pruned\nweight matrix tends to be ill-conditioned, i.e., increasing the condition\nnumber of the weight matrix. This phenomenon aggravates the vulnerability of a\nDNN to input noise. Although a highly pruned weight matrix is considered to be\nable to lower the upper bound of the local Lipschitz constant to tolerate large\ndistortion, the ill-conditionedness of such a weight matrix results in a\nnon-robust DNN model. To overcome this challenge, this work develops novel\njoint constraints to adjust the weight distribution of networks, namely, the\nTransformed Sparse Constraint joint with Condition Number Constraint (TSCNC),\nwhich copes with smoothing distribution and differentiable constraint functions\nto reduce condition number and thus avoid the ill-conditionedness of weight\nmatrices. Furthermore, our theoretical analyses unveil the relevance between\nthe condition number and the local Lipschitz constant of the weight matrix,\nnamely, the sharply increasing condition number becomes the dominant factor\nthat restricts the robustness of over-sparsified models. Extensive experiments\nare conducted on several public datasets, and the results show that the\nproposed constraints significantly improve the robustness of a DNN with high\npruning rates.\n", "date": "2025-03-26", "categories": []}
{"id": "2503.20613", "title": "State-Aware Perturbation Optimization for Robust Deep Reinforcement\n  Learning", "abstract": "  Recently, deep reinforcement learning (DRL) has emerged as a promising\napproach for robotic control. However, the deployment of DRL in real-world\nrobots is hindered by its sensitivity to environmental perturbations. While\nexisting whitebox adversarial attacks rely on local gradient information and\napply uniform perturbations across all states to evaluate DRL robustness, they\nfail to account for temporal dynamics and state-specific vulnerabilities. To\ncombat the above challenge, we first conduct a theoretical analysis of\nwhite-box attacks in DRL by establishing the adversarial victim-dynamics Markov\ndecision process (AVD-MDP), to derive the necessary and sufficient conditions\nfor a successful attack. Based on this, we propose a selective state-aware\nreinforcement adversarial attack method, named STAR, to optimize perturbation\nstealthiness and state visitation dispersion. STAR first employs a soft\nmask-based state-targeting mechanism to minimize redundant perturbations,\nenhancing stealthiness and attack effectiveness. Then, it incorporates an\ninformation-theoretic optimization objective to maximize mutual information\nbetween perturbations, environmental states, and victim actions, ensuring a\ndispersed state-visitation distribution that steers the victim agent into\nvulnerable states for maximum return reduction. Extensive experiments\ndemonstrate that STAR outperforms state-of-the-art benchmarks.\n", "date": "2025-03-26", "categories": []}
{"id": "2503.20630", "title": "$\\beta$-GNN: A Robust Ensemble Approach Against Graph Structure\n  Perturbation", "abstract": "  Graph Neural Networks (GNNs) are playing an increasingly important role in\nthe efficient operation and security of computing systems, with applications in\nworkload scheduling, anomaly detection, and resource management. However, their\nvulnerability to network perturbations poses a significant challenge. We\npropose $\\beta$-GNN, a model enhancing GNN robustness without sacrificing clean\ndata performance. $\\beta$-GNN uses a weighted ensemble, combining any GNN with\na multi-layer perceptron. A learned dynamic weight, $\\beta$, modulates the\nGNN's contribution. This $\\beta$ not only weights GNN influence but also\nindicates data perturbation levels, enabling proactive mitigation. Experimental\nresults on diverse datasets show $\\beta$-GNN's superior adversarial accuracy\nand attack severity quantification. Crucially, $\\beta$-GNN avoids perturbation\nassumptions, preserving clean data structure and performance.\n", "date": "2025-03-26", "categories": []}
{"id": "2411.15553", "title": "Improving Transferable Targeted Attacks with Feature Tuning Mixup", "abstract": "  Deep neural networks (DNNs) exhibit vulnerability to adversarial examples\nthat can transfer across different DNN models. A particularly challenging\nproblem is developing transferable targeted attacks that can mislead DNN models\ninto predicting specific target classes. While various methods have been\nproposed to enhance attack transferability, they often incur substantial\ncomputational costs while yielding limited improvements. Recent clean feature\nmixup methods use random clean features to perturb the feature space but lack\noptimization for disrupting adversarial examples, overlooking the advantages of\nattack-specific perturbations. In this paper, we propose Feature Tuning Mixup\n(FTM), a novel method that enhances targeted attack transferability by\ncombining both random and optimized noises in the feature space. FTM introduces\nlearnable feature perturbations and employs an efficient stochastic update\nstrategy for optimization. These learnable perturbations facilitate the\ngeneration of more robust adversarial examples with improved transferability.\nWe further demonstrate that attack performance can be enhanced through an\nensemble of multiple FTM-perturbed surrogate models. Extensive experiments on\nthe ImageNet-compatible dataset across various DNN models demonstrate that our\nmethod achieves significant improvements over state-of-the-art methods while\nmaintaining low computational cost.\n", "date": "2024-11-23", "categories": []}
{"id": "2503.19099", "title": "Masks and Mimicry: Strategic Obfuscation and Impersonation Attacks on\n  Authorship Verification", "abstract": "  The increasing use of Artificial Intelligence (AI) technologies, such as\nLarge Language Models (LLMs) has led to nontrivial improvements in various\ntasks, including accurate authorship identification of documents. However,\nwhile LLMs improve such defense techniques, they also simultaneously provide a\nvehicle for malicious actors to launch new attack vectors. To combat this\nsecurity risk, we evaluate the adversarial robustness of authorship models\n(specifically an authorship verification model) to potent LLM-based attacks.\nThese attacks include untargeted methods - \\textit{authorship obfuscation} and\ntargeted methods - \\textit{authorship impersonation}. For both attacks, the\nobjective is to mask or mimic the writing style of an author while preserving\nthe original texts' semantics, respectively. Thus, we perturb an accurate\nauthorship verification model, and achieve maximum attack success rates of 92\\%\nand 78\\% for both obfuscation and impersonation attacks, respectively.\n", "date": "2025-03-24", "categories": []}
{"id": "2503.19134", "title": "MIRAGE: Multimodal Immersive Reasoning and Guided Exploration for\n  Red-Team Jailbreak Attacks", "abstract": "  While safety mechanisms have significantly progressed in filtering harmful\ntext inputs, MLLMs remain vulnerable to multimodal jailbreaks that exploit\ntheir cross-modal reasoning capabilities. We present MIRAGE, a novel multimodal\njailbreak framework that exploits narrative-driven context and role immersion\nto circumvent safety mechanisms in Multimodal Large Language Models (MLLMs). By\nsystematically decomposing the toxic query into environment, role, and action\ntriplets, MIRAGE constructs a multi-turn visual storytelling sequence of images\nand text using Stable Diffusion, guiding the target model through an engaging\ndetective narrative. This process progressively lowers the model's defences and\nsubtly guides its reasoning through structured contextual cues, ultimately\neliciting harmful responses. In extensive experiments on the selected datasets\nwith six mainstream MLLMs, MIRAGE achieves state-of-the-art performance,\nimproving attack success rates by up to 17.5% over the best baselines.\nMoreover, we demonstrate that role immersion and structured semantic\nreconstruction can activate inherent model biases, facilitating the model's\nspontaneous violation of ethical safeguards. These results highlight critical\nweaknesses in current multimodal safety mechanisms and underscore the urgent\nneed for more robust defences against cross-modal threats.\n", "date": "2025-03-24", "categories": []}
{"id": "2503.19591", "title": "Boosting the Transferability of Audio Adversarial Examples with Acoustic\n  Representation Optimization", "abstract": "  With the widespread application of automatic speech recognition (ASR)\nsystems, their vulnerability to adversarial attacks has been extensively\nstudied. However, most existing adversarial examples are generated on specific\nindividual models, resulting in a lack of transferability. In real-world\nscenarios, attackers often cannot access detailed information about the target\nmodel, making query-based attacks unfeasible. To address this challenge, we\npropose a technique called Acoustic Representation Optimization that aligns\nadversarial perturbations with low-level acoustic characteristics derived from\nspeech representation models. Rather than relying on model-specific,\nhigher-layer abstractions, our approach leverages fundamental acoustic\nrepresentations that remain consistent across diverse ASR architectures. By\nenforcing an acoustic representation loss to guide perturbations toward these\nrobust, lower-level representations, we enhance the cross-model transferability\nof adversarial examples without degrading audio quality. Our method is\nplug-and-play and can be integrated with any existing attack methods. We\nevaluate our approach on three modern ASR models, and the experimental results\ndemonstrate that our method significantly improves the transferability of\nadversarial examples generated by previous methods while preserving the audio\nquality.\n", "date": "2025-03-25", "categories": []}
{"id": "2503.19791", "title": "SITA: Structurally Imperceptible and Transferable Adversarial Attacks\n  for Stylized Image Generation", "abstract": "  Image generation technology has brought significant advancements across\nvarious fields but has also raised concerns about data misuse and potential\nrights infringements, particularly with respect to creating visual artworks.\nCurrent methods aimed at safeguarding artworks often employ adversarial\nattacks. However, these methods face challenges such as poor transferability,\nhigh computational costs, and the introduction of noticeable noise, which\ncompromises the aesthetic quality of the original artwork. To address these\nlimitations, we propose a Structurally Imperceptible and Transferable\nAdversarial (SITA) attacks. SITA leverages a CLIP-based destylization loss,\nwhich decouples and disrupts the robust style representation of the image. This\ndisruption hinders style extraction during stylized image generation, thereby\nimpairing the overall stylization process. Importantly, SITA eliminates the\nneed for a surrogate diffusion model, leading to significantly reduced\ncomputational overhead. The method's robust style feature disruption ensures\nhigh transferability across diverse models. Moreover, SITA introduces\nperturbations by embedding noise within the imperceptible structural details of\nthe image. This approach effectively protects against style extraction without\ncompromising the visual quality of the artwork. Extensive experiments\ndemonstrate that SITA offers superior protection for artworks against\nunauthorized use in stylized generation. It significantly outperforms existing\nmethods in terms of transferability, computational efficiency, and noise\nimperceptibility. Code is available at https://github.com/A-raniy-day/SITA.\n", "date": "2025-03-25", "categories": []}
{"id": "2503.19817", "title": "Bitstream Collisions in Neural Image Compression via Adversarial\n  Perturbations", "abstract": "  Neural image compression (NIC) has emerged as a promising alternative to\nclassical compression techniques, offering improved compression ratios. Despite\nits progress towards standardization and practical deployment, there has been\nminimal exploration into it's robustness and security. This study reveals an\nunexpected vulnerability in NIC - bitstream collisions - where semantically\ndifferent images produce identical compressed bitstreams. Utilizing a novel\nwhitebox adversarial attack algorithm, this paper demonstrates that adding\ncarefully crafted perturbations to semantically different images can cause\ntheir compressed bitstreams to collide exactly. The collision vulnerability\nposes a threat to the practical usability of NIC, particularly in\nsecurity-critical applications. The cause of the collision is analyzed, and a\nsimple yet effective mitigation method is presented.\n", "date": "2025-03-25", "categories": []}
{"id": "2307.06608", "title": "MF-CLIP: Leveraging CLIP as Surrogate Models for No-box Adversarial\n  Attacks", "abstract": "  The vulnerability of Deep Neural Networks (DNNs) to adversarial attacks poses\na significant challenge to their deployment in safety-critical applications.\nWhile extensive research has addressed various attack scenarios, the no-box\nattack setting where adversaries have no prior knowledge, including access to\ntraining data of the target model, remains relatively underexplored despite its\npractical relevance. This work presents a systematic investigation into\nleveraging large-scale Vision-Language Models (VLMs), particularly CLIP, as\nsurrogate models for executing no-box attacks. Our theoretical and empirical\nanalyses reveal a key limitation in the execution of no-box attacks stemming\nfrom insufficient discriminative capabilities for direct application of vanilla\nCLIP as a surrogate model. To address this limitation, we propose MF-CLIP: a\nnovel framework that enhances CLIP's effectiveness as a surrogate model through\nmargin-aware feature space optimization. Comprehensive evaluations across\ndiverse architectures and datasets demonstrate that MF-CLIP substantially\nadvances the state-of-the-art in no-box attacks, surpassing existing baselines\nby 15.23% on standard models and achieving a 9.52% improvement on adversarially\ntrained models. Our code will be made publicly available to facilitate\nreproducibility and future research in this direction.\n", "date": "2023-07-13", "categories": []}
{"id": "2407.09958", "title": "Partner in Crime: Boosting Targeted Poisoning Attacks against Federated\n  Learning", "abstract": "  Federated Learning (FL) exposes vulnerabilities to targeted poisoning attacks\nthat aim to cause misclassification specifically from the source class to the\ntarget class. However, using well-established defense frameworks, the poisoning\nimpact of these attacks can be greatly mitigated. We introduce a generalized\npre-training stage approach to Boost Targeted Poisoning Attacks against FL,\ncalled BoTPA. Its design rationale is to leverage the model update\ncontributions of all data points, including ones outside of the source and\ntarget classes, to construct an Amplifier set, in which we falsify the data\nlabels before the FL training process, as a means to boost attacks. We\ncomprehensively evaluate the effectiveness and compatibility of BoTPA on\nvarious targeted poisoning attacks. Under data poisoning attacks, our\nevaluations reveal that BoTPA can achieve a median Relative Increase in Attack\nSuccess Rate (RI-ASR) between 15.3% and 36.9% across all possible source-target\nclass combinations, with varying percentages of malicious clients, compared to\nits baseline. In the context of model poisoning, BoTPA attains RI-ASRs ranging\nfrom 13.3% to 94.7% in the presence of the Krum and Multi-Krum defenses, from\n2.6% to 49.2% under the Median defense, and from 2.9% to 63.5% under the Flame\ndefense.\n", "date": "2024-07-13", "categories": []}
{"id": "2410.01617", "title": "On Using Certified Training towards Empirical Robustness", "abstract": "  Adversarial training is arguably the most popular way to provide empirical\nrobustness against specific adversarial examples. While variants based on\nmulti-step attacks incur significant computational overhead, single-step\nvariants are vulnerable to a failure mode known as catastrophic overfitting,\nwhich hinders their practical utility for large perturbations. A parallel line\nof work, certified training, has focused on producing networks amenable to\nformal guarantees of robustness against any possible attack. However, the wide\ngap between the best-performing empirical and certified defenses has severely\nlimited the applicability of the latter. Inspired by recent developments in\ncertified training, which rely on a combination of adversarial attacks with\nnetwork over-approximations, and by the connections between local linearity and\ncatastrophic overfitting, we present experimental evidence on the practical\nutility and limitations of using certified training towards empirical\nrobustness. We show that, when tuned for the purpose, a recent certified\ntraining algorithm can prevent catastrophic overfitting on single-step attacks,\nand that it can bridge the gap to multi-step baselines under appropriate\nexperimental settings. Finally, we present a conceptually simple regularizer\nfor network over-approximations that can achieve similar effects while markedly\nreducing runtime.\n", "date": "2024-10-02", "categories": []}
{"id": "2503.01407", "title": "Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based\n  Adversarial Purification", "abstract": "  Existing diffusion-based purification methods aim to disrupt adversarial\nperturbations by introducing a certain amount of noise through a forward\ndiffusion process, followed by a reverse process to recover clean examples.\nHowever, this approach is fundamentally flawed: the uniform operation of the\nforward process across all pixels compromises normal pixels while attempting to\ncombat adversarial perturbations, resulting in the target model producing\nincorrect predictions. Simply relying on low-intensity noise is insufficient\nfor effective defense. To address this critical issue, we implement a\nheterogeneous purification strategy grounded in the interpretability of neural\nnetworks. Our method decisively applies higher-intensity noise to specific\npixels that the target model focuses on while the remaining pixels are\nsubjected to only low-intensity noise. This requirement motivates us to\nredesign the sampling process of the diffusion model, allowing for the\neffective removal of varying noise levels. Furthermore, to evaluate our method\nagainst strong adaptative attack, our proposed method sharply reduces time cost\nand memory usage through a single-step resampling. The empirical evidence from\nextensive experiments across three datasets demonstrates that our method\noutperforms most current adversarial training and purification techniques by a\nsubstantial margin.\n", "date": "2025-03-03", "categories": []}
{"id": "2503.10661", "title": "CeTAD: Towards Certified Toxicity-Aware Distance in Vision Language\n  Models", "abstract": "  Recent advances in large vision-language models (VLMs) have demonstrated\nremarkable success across a wide range of visual understanding tasks. However,\nthe robustness of these models against jailbreak attacks remains an open\nchallenge. In this work, we propose a universal certified defence framework to\nsafeguard VLMs rigorously against potential visual jailbreak attacks. First, we\nproposed a novel distance metric to quantify semantic discrepancies between\nmalicious and intended responses, capturing subtle differences often overlooked\nby conventional cosine similarity-based measures. Then, we devise a regressed\ncertification approach that employs randomized smoothing to provide formal\nrobustness guarantees against both adversarial and structural perturbations,\neven under black-box settings. Complementing this, our feature-space defence\nintroduces noise distributions (e.g., Gaussian, Laplacian) into the latent\nembeddings to safeguard against both pixel-level and structure-level\nperturbations. Our results highlight the potential of a formally grounded,\nintegrated strategy toward building more resilient and trustworthy VLMs.\n", "date": "2025-03-08", "categories": []}
{"id": "2503.17630", "title": "Generating Realistic, Diverse, and Fault-Revealing Inputs with Latent\n  Space Interpolation for Testing Deep Neural Networks", "abstract": "  Deep Neural Networks (DNNs) have been widely employed across various domains,\nincluding safety-critical systems, necessitating comprehensive testing to\nensure their reliability. Although numerous DNN model testing methods have been\nproposed to generate adversarial samples that are capable of revealing faults,\nexisting methods typically perturb samples in the input space and then mutate\nthese based on feedback from the DNN model. These methods often result in test\nsamples that are not realistic and with low-probability reveal faults. To\naddress these limitations, we propose a black-box DNN test input generation\nmethod, ARGUS, to generate realistic, diverse, and fault-revealing test inputs.\nARGUS first compresses samples into a continuous latent space and then perturbs\nthe original samples by interpolating these with samples of different classes.\nSubsequently, we employ a vector quantizer and decoder to reconstruct\nadversarial samples back into the input space. Additionally, we employ\ndiscriminators both in the latent space and in the input space to ensure the\nrealism of the generated samples. Evaluation of ARGUS in comparison with\nstate-of-the-art black-box testing and white-box testing methods, shows that\nARGUS excels in generating realistic and diverse adversarial samples relative\nto the target dataset, and ARGUS successfully perturbs all original samples and\nachieves up to 4 times higher error rate than the best baseline method.\nFurthermore, using these adversarial samples for model retraining can improve\nmodel classification accuracy.\n", "date": "2025-03-22", "categories": []}
{"id": "2503.18678", "title": "NullSwap: Proactive Identity Cloaking Against Deepfake Face Swapping", "abstract": "  Suffering from performance bottlenecks in passively detecting high-quality\nDeepfake images due to the advancement of generative models, proactive\nperturbations offer a promising approach to disabling Deepfake manipulations by\ninserting signals into benign images. However, existing proactive perturbation\napproaches remain unsatisfactory in several aspects: 1) visual degradation due\nto direct element-wise addition; 2) limited effectiveness against face swapping\nmanipulation; 3) unavoidable reliance on white- and grey-box settings to\ninvolve generative models during training. In this study, we analyze the\nessence of Deepfake face swapping and argue the necessity of protecting source\nidentities rather than target images, and we propose NullSwap, a novel\nproactive defense approach that cloaks source image identities and nullifies\nface swapping under a pure black-box scenario. We design an Identity Extraction\nmodule to obtain facial identity features from the source image, while a\nPerturbation Block is then devised to generate identity-guided perturbations\naccordingly. Meanwhile, a Feature Block extracts shallow-level image features,\nwhich are then fused with the perturbation in the Cloaking Block for image\nreconstruction. Furthermore, to ensure adaptability across different identity\nextractors in face swapping algorithms, we propose Dynamic Loss Weighting to\nadaptively balance identity losses. Experiments demonstrate the outstanding\nability of our approach to fool various identity recognition models,\noutperforming state-of-the-art proactive perturbations in preventing face\nswapping models from generating images with correct source identities.\n", "date": "2025-03-24", "categories": []}
{"id": "2401.10090", "title": "Cross-Modality Perturbation Synergy Attack for Person Re-identification", "abstract": "  In recent years, there has been significant research focusing on addressing\nsecurity concerns in single-modal person re-identification (ReID) systems that\nare based on RGB images. However, the safety of cross-modality scenarios, which\nare more commonly encountered in practical applications involving images\ncaptured by infrared cameras, has not received adequate attention. The main\nchallenge in cross-modality ReID lies in effectively dealing with visual\ndifferences between different modalities. For instance, infrared images are\ntypically grayscale, unlike visible images that contain color information.\nExisting attack methods have primarily focused on the characteristics of the\nvisible image modality, overlooking the features of other modalities and the\nvariations in data distribution among different modalities. This oversight can\npotentially undermine the effectiveness of these methods in image retrieval\nacross diverse modalities. This study represents the first exploration into the\nsecurity of cross-modality ReID models and proposes a universal perturbation\nattack specifically designed for cross-modality ReID. This attack optimizes\nperturbations by leveraging gradients from diverse modality data, thereby\ndisrupting the discriminator and reinforcing the differences between\nmodalities. We conducted experiments on three widely used cross-modality\ndatasets, namely RegDB, SYSU, and LLCM. The results not only demonstrate the\neffectiveness of our method but also provide insights for future improvements\nin the robustness of cross-modality ReID systems. The code will be available at\nhttps://github.com/finger-monkey/cmps__attack.\n", "date": "2024-01-18", "categories": []}
{"id": "2408.17064", "title": "Instant Adversarial Purification with Adversarial Consistency\n  Distillation", "abstract": "  Neural networks have revolutionized numerous fields with their exceptional\nperformance, yet they remain susceptible to adversarial attacks through subtle\nperturbations. While diffusion-based purification methods like DiffPure offer\npromising defense mechanisms, their computational overhead presents a\nsignificant practical limitation. In this paper, we introduce One Step Control\nPurification (OSCP), a novel defense framework that achieves robust adversarial\npurification in a single Neural Function Evaluation (NFE) within diffusion\nmodels. We propose Gaussian Adversarial Noise Distillation (GAND) as the\ndistillation objective and Controlled Adversarial Purification (CAP) as the\ninference pipeline, which makes OSCP demonstrate remarkable efficiency while\nmaintaining defense efficacy. Our proposed GAND addresses a fundamental tension\nbetween consistency distillation and adversarial perturbation, bridging the gap\nbetween natural and adversarial manifolds in the latent space, while remaining\ncomputationally efficient through Parameter-Efficient Fine-Tuning (PEFT)\nmethods such as LoRA, eliminating the high computational budget request from\nfull parameter fine-tuning. The CAP guides the purification process through the\nunlearnable edge detection operator calculated by the input image as an extra\nprompt, effectively preventing the purified images from deviating from their\noriginal appearance when large purification steps are used. Our experimental\nresults on ImageNet showcase OSCP's superior performance, achieving a 74.19%\ndefense success rate with merely 0.1s per purification -- a 100-fold speedup\ncompared to conventional approaches.\n", "date": "2024-08-30", "categories": []}
{"id": "2503.01924", "title": "TAET: Two-Stage Adversarial Equalization Training on Long-Tailed\n  Distributions", "abstract": "  Adversarial robustness is a critical challenge in deploying deep neural\nnetworks for real-world applications. While adversarial training is a widely\nrecognized defense strategy, most existing studies focus on balanced datasets,\noverlooking the prevalence of long-tailed distributions in real-world data,\nwhich significantly complicates robustness. This paper provides a comprehensive\nanalysis of adversarial training under long-tailed distributions and identifies\nlimitations in the current state-of-the-art method, AT-BSL, in achieving robust\nperformance under such conditions. To address these challenges, we propose a\nnovel training framework, TAET, which integrates an initial stabilization phase\nfollowed by a stratified equalization adversarial training phase. Additionally,\nprior work on long-tailed robustness has largely ignored the crucial evaluation\nmetric of balanced accuracy. To bridge this gap, we introduce the concept of\nbalanced robustness, a comprehensive metric tailored for assessing robustness\nunder long-tailed distributions. Extensive experiments demonstrate that our\nmethod surpasses existing advanced defenses, achieving significant improvements\nin both memory and computational efficiency. This work represents a substantial\nadvancement in addressing robustness challenges in real-world applications. Our\ncode is available at:\nhttps://github.com/BuhuiOK/TAET-Two-Stage-Adversarial-Equalization-Training-on-Long-Tailed-Distributions.\n", "date": "2025-03-02", "categories": []}
{"id": "2503.17172", "title": "Principal Eigenvalue Regularization for Improved Worst-Class Certified\n  Robustness of Smoothed Classifiers", "abstract": "  Recent studies have identified a critical challenge in deep neural networks\n(DNNs) known as ``robust fairness\", where models exhibit significant\ndisparities in robust accuracy across different classes. While prior work has\nattempted to address this issue in adversarial robustness, the study of\nworst-class certified robustness for smoothed classifiers remains unexplored.\nOur work bridges this gap by developing a PAC-Bayesian bound for the\nworst-class error of smoothed classifiers. Through theoretical analysis, we\ndemonstrate that the largest eigenvalue of the smoothed confusion matrix\nfundamentally influences the worst-class error of smoothed classifiers. Based\non this insight, we introduce a regularization method that optimizes the\nlargest eigenvalue of smoothed confusion matrix to enhance worst-class accuracy\nof the smoothed classifier and further improve its worst-class certified\nrobustness. We provide extensive experimental validation across multiple\ndatasets and model architectures to demonstrate the effectiveness of our\napproach.\n", "date": "2025-03-21", "categories": []}
{"id": "2503.17198", "title": "Jailbreaking the Non-Transferable Barrier via Test-Time Data Disguising", "abstract": "  Non-transferable learning (NTL) has been proposed to protect model\nintellectual property (IP) by creating a \"non-transferable barrier\" to restrict\ngeneralization from authorized to unauthorized domains. Recently, well-designed\nattack, which restores the unauthorized-domain performance by fine-tuning NTL\nmodels on few authorized samples, highlights the security risks of NTL-based\napplications. However, such attack requires modifying model weights, thus being\ninvalid in the black-box scenario. This raises a critical question: can we\ntrust the security of NTL models deployed as black-box systems? In this work,\nwe reveal the first loophole of black-box NTL models by proposing a novel\nattack method (dubbed as JailNTL) to jailbreak the non-transferable barrier\nthrough test-time data disguising. The main idea of JailNTL is to disguise\nunauthorized data so it can be identified as authorized by the NTL model,\nthereby bypassing the non-transferable barrier without modifying the NTL model\nweights. Specifically, JailNTL encourages unauthorized-domain disguising in two\nlevels, including: (i) data-intrinsic disguising (DID) for eliminating domain\ndiscrepancy and preserving class-related content at the input-level, and (ii)\nmodel-guided disguising (MGD) for mitigating output-level statistics difference\nof the NTL model. Empirically, when attacking state-of-the-art (SOTA) NTL\nmodels in the black-box scenario, JailNTL achieves an accuracy increase of up\nto 55.7% in the unauthorized domain by using only 1% authorized samples,\nlargely exceeding existing SOTA white-box attacks.\n", "date": "2025-03-21", "categories": []}
{"id": "2409.04982", "title": "2DSig-Detect: a semi-supervised framework for anomaly detection on image\n  data using 2D-signatures", "abstract": "  The rapid advancement of machine learning technologies raises questions about\nthe security of machine learning models, with respect to both training-time\n(poisoning) and test-time (evasion, impersonation, and inversion) attacks.\nModels performing image-related tasks, e.g. detection, and classification, are\nvulnerable to adversarial attacks that can degrade their performance and\nproduce undesirable outcomes. This paper introduces a novel technique for\nanomaly detection in images called 2DSig-Detect, which uses a\n2D-signature-embedded semi-supervised framework rooted in rough path theory. We\ndemonstrate our method in adversarial settings for training-time and test-time\nattacks, and benchmark our framework against other state of the art methods.\nUsing 2DSig-Detect for anomaly detection, we show both superior performance and\na reduction in the computation time to detect the presence of adversarial\nperturbations in images.\n", "date": "2024-09-08", "categories": []}
{"id": "2503.11032", "title": "Weakly Supervised Contrastive Adversarial Training for Learning Robust\n  Features from Semi-supervised Data", "abstract": "  Existing adversarial training (AT) methods often suffer from incomplete\nperturbation, meaning that not all non-robust features are perturbed when\ngenerating adversarial examples (AEs). This results in residual correlations\nbetween non-robust features and labels, leading to suboptimal learning of\nrobust features. However, achieving complete perturbation, i.e., perturbing as\nmany non-robust features as possible, is challenging due to the difficulty in\ndistinguishing robust and non-robust features and the sparsity of labeled data.\nTo address these challenges, we propose a novel approach called Weakly\nSupervised Contrastive Adversarial Training (WSCAT). WSCAT ensures complete\nperturbation for improved learning of robust features by disrupting\ncorrelations between non-robust features and labels through complete AE\ngeneration over partially labeled data, grounded in information theory.\nExtensive theoretical analysis and comprehensive experiments on widely adopted\nbenchmarks validate the superiority of WSCAT. Our code is available at\nhttps://github.com/zhang-lilin/WSCAT.\n", "date": "2025-03-14", "categories": []}
{"id": "2503.16179", "title": "Narrowing Class-Wise Robustness Gaps in Adversarial Training", "abstract": "  Efforts to address declining accuracy as a result of data shifts often\ninvolve various data-augmentation strategies. Adversarial training is one such\nmethod, designed to improve robustness to worst-case distribution shifts caused\nby adversarial examples. While this method can improve robustness, it may also\nhinder generalization to clean examples and exacerbate performance imbalances\nacross different classes. This paper explores the impact of adversarial\ntraining on both overall and class-specific performance, as well as its\nspill-over effects. We observe that enhanced labeling during training boosts\nadversarial robustness by 53.50% and mitigates class imbalances by 5.73%,\nleading to improved accuracy in both clean and adversarial settings compared to\nstandard adversarial training.\n", "date": "2025-03-20", "categories": []}
{"id": "2403.01896", "title": "Robustness bounds on the successful adversarial examples in\n  probabilistic models: Implications from Gaussian processes", "abstract": "  Adversarial example (AE) is an attack method for machine learning, which is\ncrafted by adding imperceptible perturbation to the data inducing\nmisclassification. In the current paper, we investigated the upper bound of the\nprobability of successful AEs based on the Gaussian Process (GP)\nclassification, a probabilistic inference model. We proved a new upper bound of\nthe probability of a successful AE attack that depends on AE's perturbation\nnorm, the kernel function used in GP, and the distance of the closest pair with\ndifferent labels in the training dataset. Surprisingly, the upper bound is\ndetermined regardless of the distribution of the sample dataset. We showed that\nour theoretical result was confirmed through the experiment using ImageNet. In\naddition, we showed that changing the parameters of the kernel function induces\na change of the upper bound of the probability of successful AEs.\n", "date": "2024-03-04", "categories": []}
{"id": "2407.18981", "title": "Prompt Injection Attacks on Large Language Models in Oncology", "abstract": "  Vision-language artificial intelligence models (VLMs) possess medical\nknowledge and can be employed in healthcare in numerous ways, including as\nimage interpreters, virtual scribes, and general decision support systems.\nHowever, here, we demonstrate that current VLMs applied to medical tasks\nexhibit a fundamental security flaw: they can be attacked by prompt injection\nattacks, which can be used to output harmful information just by interacting\nwith the VLM, without any access to its parameters. We performed a quantitative\nstudy to evaluate the vulnerabilities to these attacks in four state of the art\nVLMs which have been proposed to be of utility in healthcare: Claude 3 Opus,\nClaude 3.5 Sonnet, Reka Core, and GPT-4o. Using a set of N=297 attacks, we show\nthat all of these models are susceptible. Specifically, we show that embedding\nsub-visual prompts in medical imaging data can cause the model to provide\nharmful output, and that these prompts are non-obvious to human observers.\nThus, our study demonstrates a key vulnerability in medical VLMs which should\nbe mitigated before widespread clinical adoption.\n", "date": "2024-07-23", "categories": []}
{"id": "2408.00315", "title": "ADBM: Adversarial diffusion bridge model for reliable adversarial\n  purification", "abstract": "  Recently Diffusion-based Purification (DiffPure) has been recognized as an\neffective defense method against adversarial examples. However, we find\nDiffPure which directly employs the original pre-trained diffusion models for\nadversarial purification, to be suboptimal. This is due to an inherent\ntrade-off between noise purification performance and data recovery quality.\nAdditionally, the reliability of existing evaluations for DiffPure is\nquestionable, as they rely on weak adaptive attacks. In this work, we propose a\nnovel Adversarial Diffusion Bridge Model, termed ADBM. ADBM directly constructs\na reverse bridge from the diffused adversarial data back to its original clean\nexamples, enhancing the purification capabilities of the original diffusion\nmodels. Through theoretical analysis and experimental validation across various\nscenarios, ADBM has proven to be a superior and robust defense mechanism,\noffering significant promise for practical applications.\n", "date": "2024-08-01", "categories": []}
{"id": "2410.12677", "title": "Efficient Optimization Algorithms for Linear Adversarial Training", "abstract": "  Adversarial training can be used to learn models that are robust against\nperturbations. For linear models, it can be formulated as a convex optimization\nproblem. Compared to methods proposed in the context of deep learning,\nleveraging the optimization structure allows significantly faster convergence\nrates. Still, the use of generic convex solvers can be inefficient for\nlarge-scale problems. Here, we propose tailored optimization algorithms for the\nadversarial training of linear models, which render large-scale regression and\nclassification problems more tractable. For regression problems, we propose a\nfamily of solvers based on iterative ridge regression and, for classification,\na family of solvers based on projected gradient descent. The methods are based\non extended variable reformulations of the original problem. We illustrate\ntheir efficiency in numerical examples.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.15176", "title": "Adaptive Pruning with Module Robustness Sensitivity: Balancing\n  Compression and Robustness", "abstract": "  Neural network pruning has traditionally focused on weight-based criteria to\nachieve model compression, frequently overlooking the crucial balance between\nadversarial robustness and accuracy. Existing approaches often fail to preserve\nrobustness in pruned networks, leaving them more susceptible to adversarial\nattacks. This paper introduces Module Robustness Sensitivity (MRS), a novel\nmetric that quantifies layer-wise sensitivity to adversarial perturbations and\ndynamically informs pruning decisions. Leveraging MRS, we propose Module Robust\nPruning and Fine-Tuning (MRPF), an adaptive pruning algorithm compatible with\nany adversarial training method, offering both flexibility and scalability.\nExtensive experiments on SVHN, CIFAR, and Tiny-ImageNet across diverse\narchitectures, including ResNet, VGG, and MobileViT, demonstrate that MRPF\nsignificantly enhances adversarial robustness while maintaining competitive\naccuracy and computational efficiency. Furthermore, MRPF consistently\noutperforms state-of-the-art structured pruning methods in balancing\nrobustness, accuracy, and compression. This work establishes a practical and\ngeneralizable framework for robust pruning, addressing the long-standing\ntrade-off between model compression and robustness preservation.\n", "date": "2024-10-19", "categories": []}
{"id": "2503.14922", "title": "A Semantic and Clean-label Backdoor Attack against Graph Convolutional\n  Networks", "abstract": "  Graph Convolutional Networks (GCNs) have shown excellent performance in\ngraph-structured tasks such as node classification and graph classification.\nHowever, recent research has shown that GCNs are vulnerable to a new type of\nthreat called the backdoor attack, where the adversary can inject a hidden\nbackdoor into the GCNs so that the backdoored model performs well on benign\nsamples, whereas its prediction will be maliciously changed to the\nattacker-specified target label if the hidden backdoor is activated by the\nattacker-defined trigger. Clean-label backdoor attack and semantic backdoor\nattack are two new backdoor attacks to Deep Neural Networks (DNNs), they are\nmore imperceptible and have posed new and serious threats. The semantic and\nclean-label backdoor attack is not fully explored in GCNs. In this paper, we\npropose a semantic and clean-label backdoor attack against GCNs under the\ncontext of graph classification to reveal the existence of this security\nvulnerability in GCNs. Specifically, SCLBA conducts an importance analysis on\ngraph samples to select one type of node as semantic trigger, which is then\ninserted into the graph samples to create poisoning samples without changing\nthe labels of the poisoning samples to the attacker-specified target label. We\nevaluate SCLBA on multiple datasets and the results show that SCLBA can achieve\nattack success rates close to 99% with poisoning rates of less than 3%, and\nwith almost no impact on the performance of model on benign samples.\n", "date": "2025-03-19", "categories": []}
{"id": "2503.15404", "title": "Improving Adversarial Transferability on Vision Transformers via Forward\n  Propagation Refinement", "abstract": "  Vision Transformers (ViTs) have been widely applied in various computer\nvision and vision-language tasks. To gain insights into their robustness in\npractical scenarios, transferable adversarial examples on ViTs have been\nextensively studied. A typical approach to improving adversarial\ntransferability is by refining the surrogate model. However, existing work on\nViTs has restricted their surrogate refinement to backward propagation. In this\nwork, we instead focus on Forward Propagation Refinement (FPR) and specifically\nrefine two key modules of ViTs: attention maps and token embeddings. For\nattention maps, we propose Attention Map Diversification (AMD), which\ndiversifies certain attention maps and also implicitly imposes beneficial\ngradient vanishing during backward propagation. For token embeddings, we\npropose Momentum Token Embedding (MTE), which accumulates historical token\nembeddings to stabilize the forward updates in both the Attention and MLP\nblocks. We conduct extensive experiments with adversarial examples transferred\nfrom ViTs to various CNNs and ViTs, demonstrating that our FPR outperforms the\ncurrent best (backward) surrogate refinement by up to 7.0\\% on average. We also\nvalidate its superiority against popular defenses and its compatibility with\nother transfer methods. Codes and appendix are available at\nhttps://github.com/RYC-98/FPR.\n", "date": "2025-03-19", "categories": []}
{"id": "2502.16793", "title": "VGFL-SA: Vertical Graph Federated Learning Structure Attack Based on\n  Contrastive Learning", "abstract": "  Graph Neural Networks (GNNs) have gained attention for their ability to learn\nrepresentations from graph data. Due to privacy concerns and conflicts of\ninterest that prevent clients from directly sharing graph data with one\nanother, Vertical Graph Federated Learning (VGFL) frameworks have been\ndeveloped. Recent studies have shown that VGFL is vulnerable to adversarial\nattacks that degrade performance. However, it is a common problem that client\nnodes are often unlabeled in the realm of VGFL. Consequently, the existing\nattacks, which rely on the availability of labeling information to obtain\ngradients, are inherently constrained in their applicability. This limitation\nprecludes their deployment in practical, real-world environments. To address\nthe above problems, we propose a novel graph adversarial attack against VGFL,\nreferred to as VGFL-SA, to degrade the performance of VGFL by modifying the\nlocal clients structure without using labels. Specifically, VGFL-SA uses a\ncontrastive learning method to complete the attack before the local clients are\ntrained. VGFL-SA first accesses the graph structure and node feature\ninformation of the poisoned clients, and generates the contrastive views by\nnode-degree-based edge augmentation and feature shuffling augmentation. Then,\nVGFL-SA uses the shared graph encoder to get the embedding of each view, and\nthe gradients of the adjacency matrices are obtained by the contrastive\nfunction. Finally, perturbed edges are generated using gradient modification\nrules. We validated the performance of VGFL-SA by performing a node\nclassification task on real-world datasets, and the results show that VGFL-SA\nachieves good attack effectiveness and transferability.\n", "date": "2025-02-24", "categories": []}
{"id": "2503.12874", "title": "Evolution-based Region Adversarial Prompt Learning for Robustness\n  Enhancement in Vision-Language Models", "abstract": "  Large pre-trained vision-language models (VLMs), such as CLIP, demonstrate\nimpressive generalization but remain highly vulnerable to adversarial examples\n(AEs). Previous work has explored robust text prompts through adversarial\ntraining, achieving some improvement in both robustness and generalization.\nHowever, they primarily rely on singlegradient direction perturbations (e.g.,\nPGD) to generate AEs, which lack diversity, resulting in limited improvement in\nadversarial robustness. To address these limitations, we propose an\nevolution-based region adversarial prompt tuning method called ER-APT, which\ncombines gradient methods with genetic evolution to generate more diverse and\nchallenging AEs. In each training iteration, we first generate AEs using\ntraditional gradient-based methods. Subsequently, a genetic evolution mechanism\nincorporating selection, mutation, and crossover is applied to optimize the\nAEs, ensuring a broader and more aggressive perturbation distribution.The final\nevolved AEs are used for prompt tuning, achieving region-based adversarial\noptimization instead of conventional single-point adversarial prompt tuning. We\nalso propose a dynamic loss weighting method to adjust prompt learning\nefficiency for accuracy and robustness. Experimental evaluations on various\nbenchmark datasets demonstrate the superiority of our proposed method,\noutperforming stateof-the-art APT methods. The code is released at\nhttps://github.com/jiaxiaojunQAQ/ER-APT.\n", "date": "2025-03-17", "categories": []}
{"id": "2503.13945", "title": "Make the Most of Everything: Further Considerations on Disrupting\n  Diffusion-based Customization", "abstract": "  The fine-tuning technique for text-to-image diffusion models facilitates\nimage customization but risks privacy breaches and opinion manipulation.\nCurrent research focuses on prompt- or image-level adversarial attacks for\nanti-customization, yet it overlooks the correlation between these two levels\nand the relationship between internal modules and inputs. This hinders\nanti-customization performance in practical threat scenarios. We propose Dual\nAnti-Diffusion (DADiff), a two-stage adversarial attack targeting diffusion\ncustomization, which, for the first time, integrates the adversarial\nprompt-level attack into the generation process of image-level adversarial\nexamples. In stage 1, we generate prompt-level adversarial vectors to guide the\nsubsequent image-level attack. In stage 2, besides conducting the end-to-end\nattack on the UNet model, we disrupt its self- and cross-attention modules,\naiming to break the correlations between image pixels and align the\ncross-attention results computed using instance prompts and adversarial prompt\nvectors within the images. Furthermore, we introduce a local random timestep\ngradient ensemble strategy, which updates adversarial perturbations by\nintegrating random gradients from multiple segmented timesets. Experimental\nresults on various mainstream facial datasets demonstrate 10%-30% improvements\nin cross-prompt, keyword mismatch, cross-model, and cross-mechanism\nanti-customization with DADiff compared to existing methods.\n", "date": "2025-03-18", "categories": []}
{"id": "2503.13962", "title": "Survey of Adversarial Robustness in Multimodal Large Language Models", "abstract": "  Multimodal Large Language Models (MLLMs) have demonstrated exceptional\nperformance in artificial intelligence by facilitating integrated understanding\nacross diverse modalities, including text, images, video, audio, and speech.\nHowever, their deployment in real-world applications raises significant\nconcerns about adversarial vulnerabilities that could compromise their safety\nand reliability. Unlike unimodal models, MLLMs face unique challenges due to\nthe interdependencies among modalities, making them susceptible to\nmodality-specific threats and cross-modal adversarial manipulations. This paper\nreviews the adversarial robustness of MLLMs, covering different modalities. We\nbegin with an overview of MLLMs and a taxonomy of adversarial attacks tailored\nto each modality. Next, we review key datasets and evaluation metrics used to\nassess the robustness of MLLMs. After that, we provide an in-depth review of\nattacks targeting MLLMs across different modalities. Our survey also identifies\ncritical challenges and suggests promising future research directions.\n", "date": "2025-03-18", "categories": []}
{"id": "2210.16371", "title": "Distributed Black-box Attack: Do Not Overestimate Black-box Attacks", "abstract": "  As cloud computing becomes pervasive, deep learning models are deployed on\ncloud servers and then provided as APIs to end users. However, black-box\nadversarial attacks can fool image classification models without access to\nmodel structure and weights. Recent studies have reported attack success rates\nof over 95% with fewer than 1,000 queries. Then the question arises: whether\nblack-box attacks have become a real threat against cloud APIs? To shed some\nlight on this, our research indicates that black-box attacks are not as\neffective against cloud APIs as proposed in research papers due to several\ncommon mistakes that overestimate the efficiency of black-box attacks. To avoid\nsimilar mistakes, we conduct black-box attacks directly on cloud APIs rather\nthan local models.\n", "date": "2022-10-28", "categories": []}
{"id": "2309.01102", "title": "CARNet: Collaborative Adversarial Resilience for Robust Underwater Image\n  Enhancement and Perception", "abstract": "  Due to the uneven absorption of different light wavelengths in aquatic\nenvironments, underwater images suffer from low visibility and clear color\ndeviations. With the advancement of autonomous underwater vehicles, extensive\nresearch has been conducted on learning-based underwater enhancement\nalgorithms. These works can generate visually pleasing enhanced images and\nmitigate the adverse effects of degraded images on subsequent perception tasks.\nHowever, learning-based methods are susceptible to the inherent fragility of\nadversarial attacks, causing significant disruption in enhanced results. In\nthis work, we introduce a collaborative adversarial resilience network, dubbed\nCARNet, for underwater image enhancement and subsequent detection tasks.\nConcretely, we first introduce an invertible network with strong\nperturbation-perceptual abilities to isolate attacks from underwater images,\npreventing interference with visual quality enhancement and perceptual tasks.\nFurthermore, an attack pattern discriminator is introduced to adaptively\nidentify and eliminate various types of attacks. Additionally, we propose a\nbilevel attack optimization strategy to heighten the robustness of the network\nagainst different types of attacks under the collaborative adversarial training\nof vision-driven and perception-driven attacks. Extensive experiments\ndemonstrate that the proposed method outputs visually appealing enhancement\nimages and performs an average 6.71% higher detection mAP than state-of-the-art\nmethods.\n", "date": "2023-09-03", "categories": []}
{"id": "2404.17092", "title": "Robust and Efficient Adversarial Defense in SNNs via Image Purification\n  and Joint Detection", "abstract": "  Spiking Neural Networks (SNNs) aim to bridge the gap between neuroscience and\nmachine learning by emulating the structure of the human nervous system.\nHowever, like convolutional neural networks, SNNs are vulnerable to adversarial\nattacks. To tackle the challenge, we propose a biologically inspired\nmethodology to enhance the robustness of SNNs, drawing insights from the visual\nmasking effect and filtering theory. First, an end-to-end SNN-based image\npurification model is proposed to defend against adversarial attacks, including\na noise extraction network and a non-blind denoising network. The former\nnetwork extracts noise features from noisy images, while the latter component\nemploys a residual U-Net structure to reconstruct high-quality noisy images and\ngenerate clean images. Simultaneously, a multi-level firing SNN based on\nSqueeze-and-Excitation Network is introduced to improve the robustness of the\nclassifier. Crucially, the proposed image purification network serves as a\npre-processing module, avoiding modifications to classifiers. Unlike\nadversarial training, our method is highly flexible and can be seamlessly\nintegrated with other defense strategies. Experimental results on various\ndatasets demonstrate that the proposed methodology outperforms state-of-the-art\nbaselines in terms of defense effectiveness, training time, and resource\nconsumption.\n", "date": "2024-04-26", "categories": []}
{"id": "2502.13459", "title": "Poisoned Source Code Detection in Code Models", "abstract": "  Deep learning models have gained popularity for conducting various tasks\ninvolving source code. However, their black-box nature raises concerns about\npotential risks. One such risk is a poisoning attack, where an attacker\nintentionally contaminates the training set with malicious samples to mislead\nthe model's predictions in specific scenarios. To protect source code models\nfrom poisoning attacks, we introduce CodeGarrison (CG), a hybrid deep-learning\nmodel that relies on code embeddings to identify poisoned code samples. We\nevaluated CG against the state-of-the-art technique ONION for detecting\npoisoned samples generated by DAMP, MHM, ALERT, as well as a novel poisoning\ntechnique named CodeFooler. Results showed that CG significantly outperformed\nONION with an accuracy of 93.5%. We also tested CG's robustness against unknown\nattacks, achieving an average accuracy of 85.6% in identifying poisoned samples\nacross the four attacks mentioned above.\n", "date": "2025-02-19", "categories": []}
{"id": "2503.11750", "title": "Making Every Step Effective: Jailbreaking Large Vision-Language Models\n  Through Hierarchical KV Equalization", "abstract": "  In the realm of large vision-language models (LVLMs), adversarial jailbreak\nattacks serve as a red-teaming approach to identify safety vulnerabilities of\nthese models and their associated defense mechanisms. However, we identify a\ncritical limitation: not every adversarial optimization step leads to a\npositive outcome, and indiscriminately accepting optimization results at each\nstep may reduce the overall attack success rate. To address this challenge, we\nintroduce HKVE (Hierarchical Key-Value Equalization), an innovative\njailbreaking framework that selectively accepts gradient optimization results\nbased on the distribution of attention scores across different layers, ensuring\nthat every optimization step positively contributes to the attack. Extensive\nexperiments demonstrate HKVE's significant effectiveness, achieving attack\nsuccess rates of 75.08% on MiniGPT4, 85.84% on LLaVA and 81.00% on Qwen-VL,\nsubstantially outperforming existing methods by margins of 20.43\\%, 21.01\\% and\n26.43\\% respectively. Furthermore, making every step effective not only leads\nto an increase in attack success rate but also allows for a reduction in the\nnumber of iterations, thereby lowering computational costs. Warning: This paper\ncontains potentially harmful example data.\n", "date": "2025-03-14", "categories": []}
{"id": "2503.12683", "title": "Algebraic Adversarial Attacks on Explainability Models", "abstract": "  Classical adversarial attacks are phrased as a constrained optimisation\nproblem. Despite the efficacy of a constrained optimisation approach to\nadversarial attacks, one cannot trace how an adversarial point was generated.\nIn this work, we propose an algebraic approach to adversarial attacks and study\nthe conditions under which one can generate adversarial examples for post-hoc\nexplainability models. Phrasing neural networks in the framework of geometric\ndeep learning, algebraic adversarial attacks are constructed through analysis\nof the symmetry groups of neural networks. Algebraic adversarial examples\nprovide a mathematically tractable approach to adversarial examples. We\nvalidate our approach of algebraic adversarial examples on two well-known and\none real-world dataset.\n", "date": "2025-03-16", "categories": []}
{"id": "2503.13419", "title": "Securing Virtual Reality Experiences: Unveiling and Tackling\n  Cybersickness Attacks with Explainable AI", "abstract": "  The synergy between virtual reality (VR) and artificial intelligence (AI),\nspecifically deep learning (DL)-based cybersickness detection models, has\nushered in unprecedented advancements in immersive experiences by automatically\ndetecting cybersickness severity and adaptively various mitigation techniques,\noffering a smooth and comfortable VR experience. While this DL-enabled\ncybersickness detection method provides promising solutions for enhancing user\nexperiences, it also introduces new risks since these models are vulnerable to\nadversarial attacks; a small perturbation of the input data that is visually\nundetectable to human observers can fool the cybersickness detection model and\ntrigger unexpected mitigation, thus disrupting user immersive experiences (UIX)\nand even posing safety risks. In this paper, we present a new type of VR\nattack, i.e., a cybersickness attack, which successfully stops the triggering\nof cybersickness mitigation by fooling DL-based cybersickness detection models\nand dramatically hinders the UIX. Next, we propose a novel explainable\nartificial intelligence (XAI)-guided cybersickness attack detection framework\nto detect such attacks in VR to ensure UIX and a comfortable VR experience. We\nevaluate the proposed attack and the detection framework using two\nstate-of-the-art open-source VR cybersickness datasets: Simulation 2021 and\nGameplay dataset. Finally, to verify the effectiveness of our proposed method,\nwe implement the attack and the XAI-based detection using a testbed with a\ncustom-built VR roller coaster simulation with an HTC Vive Pro Eye headset and\nperform a user study. Our study shows that such an attack can dramatically\nhinder the UIX. However, our proposed XAI-guided cybersickness attack detection\ncan successfully detect cybersickness attacks and trigger the proper\nmitigation, effectively reducing VR cybersickness.\n", "date": "2025-03-17", "categories": []}
{"id": "2312.04584", "title": "Towards Sample-specific Backdoor Attack with Clean Labels via Attribute\n  Trigger", "abstract": "  Currently, sample-specific backdoor attacks (SSBAs) are the most advanced and\nmalicious methods since they can easily circumvent most of the current backdoor\ndefenses. In this paper, we reveal that SSBAs are not sufficiently stealthy due\nto their poisoned-label nature, where users can discover anomalies if they\ncheck the image-label relationship. In particular, we demonstrate that it is\nineffective to directly generalize existing SSBAs to their clean-label variants\nby poisoning samples solely from the target class. We reveal that it is\nprimarily due to two reasons, including \\textbf{(1)} the `antagonistic effects'\nof ground-truth features and \\textbf{(2)} the learning difficulty of\nsample-specific features. Accordingly, trigger-related features of existing\nSSBAs cannot be effectively learned under the clean-label setting due to their\nmild trigger intensity required for ensuring stealthiness. We argue that the\nintensity constraint of existing SSBAs is mostly because their trigger patterns\nare `content-irrelevant' and therefore act as `noises' for both humans and\nDNNs. Motivated by this understanding, we propose to exploit content-relevant\nfeatures, $a.k.a.$ (human-relied) attributes, as the trigger patterns to design\nclean-label SSBAs. This new attack paradigm is dubbed backdoor attack with\nattribute trigger (BAAT). Extensive experiments are conducted on benchmark\ndatasets, which verify the effectiveness of our BAAT and its resistance to\nexisting defenses.\n", "date": "2023-12-03", "categories": []}
{"id": "2411.14243", "title": "AnywhereDoor: Multi-Target Backdoor Attacks on Object Detection", "abstract": "  As object detection becomes integral to many safety-critical applications,\nunderstanding its vulnerabilities is essential. Backdoor attacks, in\nparticular, pose a serious threat by implanting hidden triggers in victim\nmodels, which adversaries can later exploit to induce malicious behaviors\nduring inference. However, current understanding is limited to single-target\nattacks, where adversaries must define a fixed malicious behavior (target)\nbefore training, making inference-time adaptability impossible. Given the large\noutput space of object detection (including object existence prediction,\nbounding box estimation, and classification), the feasibility of flexible,\ninference-time model control remains unexplored. This paper introduces\nAnywhereDoor, a multi-target backdoor attack for object detection. Once\nimplanted, AnywhereDoor allows adversaries to make objects disappear, fabricate\nnew ones, or mislabel them, either across all object classes or specific ones,\noffering an unprecedented degree of control. This flexibility is enabled by\nthree key innovations: (i) objective disentanglement to scale the number of\nsupported targets; (ii) trigger mosaicking to ensure robustness even against\nregion-based detectors; and (iii) strategic batching to address object-level\ndata imbalances that hinder manipulation. Extensive experiments demonstrate\nthat AnywhereDoor grants attackers a high degree of control, improving attack\nsuccess rates by 26% compared to adaptations of existing methods for such\nflexible control.\n", "date": "2024-11-21", "categories": []}
{"id": "2503.06254", "title": "Poisoned-MRAG: Knowledge Poisoning Attacks to Multimodal Retrieval\n  Augmented Generation", "abstract": "  Multimodal retrieval-augmented generation (RAG) enhances the visual reasoning\ncapability of vision-language models (VLMs) by dynamically accessing\ninformation from external knowledge bases. In this work, we introduce\n\\textit{Poisoned-MRAG}, the first knowledge poisoning attack on multimodal RAG\nsystems. Poisoned-MRAG injects a few carefully crafted image-text pairs into\nthe multimodal knowledge database, manipulating VLMs to generate the\nattacker-desired response to a target query. Specifically, we formalize the\nattack as an optimization problem and propose two cross-modal attack\nstrategies, dirty-label and clean-label, tailored to the attacker's knowledge\nand goals. Our extensive experiments across multiple knowledge databases and\nVLMs show that Poisoned-MRAG outperforms existing methods, achieving up to 98\\%\nattack success rate with just five malicious image-text pairs injected into the\nInfoSeek database (481,782 pairs). Additionally, We evaluate 4 different\ndefense strategies, including paraphrasing, duplicate removal, structure-driven\nmitigation, and purification, demonstrating their limited effectiveness and\ntrade-offs against Poisoned-MRAG. Our results highlight the effectiveness and\nscalability of Poisoned-MRAG, underscoring its potential as a significant\nthreat to multimodal RAG systems.\n", "date": "2025-03-08", "categories": []}
{"id": "2407.20657", "title": "Prompt-Driven Contrastive Learning for Transferable Adversarial Attacks", "abstract": "  Recent vision-language foundation models, such as CLIP, have demonstrated\nsuperior capabilities in learning representations that can be transferable\nacross diverse range of downstream tasks and domains. With the emergence of\nsuch powerful models, it has become crucial to effectively leverage their\ncapabilities in tackling challenging vision tasks. On the other hand, only a\nfew works have focused on devising adversarial examples that transfer well to\nboth unknown domains and model architectures. In this paper, we propose a novel\ntransfer attack method called PDCL-Attack, which leverages the CLIP model to\nenhance the transferability of adversarial perturbations generated by a\ngenerative model-based attack framework. Specifically, we formulate an\neffective prompt-driven feature guidance by harnessing the semantic\nrepresentation power of text, particularly from the ground-truth class labels\nof input images. To the best of our knowledge, we are the first to introduce\nprompt learning to enhance the transferable generative attacks. Extensive\nexperiments conducted across various cross-domain and cross-model settings\nempirically validate our approach, demonstrating its superiority over\nstate-of-the-art methods.\n", "date": "2024-07-30", "categories": []}
{"id": "2408.13809", "title": "On the Robustness of Kolmogorov-Arnold Networks: An Adversarial\n  Perspective", "abstract": "  Kolmogorov-Arnold Networks (KANs) have recently emerged as a novel approach\nto function approximation, demonstrating remarkable potential in various\ndomains. Despite their theoretical promise, the robustness of KANs under\nadversarial conditions has yet to be thoroughly examined. In this paper we\nexplore the adversarial robustness of KANs, with a particular focus on image\nclassification tasks. We assess the performance of KANs against standard white\nbox and black-box adversarial attacks, comparing their resilience to that of\nestablished neural network architectures. Our experimental evaluation\nencompasses a variety of standard image classification benchmark datasets and\ninvestigates both fully connected and convolutional neural network\narchitectures, of three sizes: small, medium, and large. We conclude that\nsmall- and medium-sized KANs (either fully connected or convolutional) are not\nconsistently more robust than their standard counterparts, but that large-sized\nKANs are, by and large, more robust. This comprehensive evaluation of KANs in\nadversarial scenarios offers the first in-depth analysis of KAN security,\nlaying the groundwork for future research in this emerging field.\n", "date": "2024-08-25", "categories": []}
{"id": "2411.19853", "title": "Towards Class-wise Robustness Analysis", "abstract": "  While being very successful in solving many downstream tasks, the application\nof deep neural networks is limited in real-life scenarios because of their\nsusceptibility to domain shifts such as common corruptions, and adversarial\nattacks. The existence of adversarial examples and data corruption\nsignificantly reduces the performance of deep classification models.\nResearchers have made strides in developing robust neural architectures to\nbolster decisions of deep classifiers. However, most of these works rely on\neffective adversarial training methods, and predominantly focus on overall\nmodel robustness, disregarding class-wise differences in robustness, which are\ncritical. Exploiting weakly robust classes is a potential avenue for attackers\nto fool the image recognition models. Therefore, this study investigates\nclass-to-class biases across adversarially trained robust classification models\nto understand their latent space structures and analyze their strong and weak\nclass-wise properties. We further assess the robustness of classes against\ncommon corruptions and adversarial attacks, recognizing that class\nvulnerability extends beyond the number of correct classifications for a\nspecific class. We find that the number of false positives of classes as\nspecific target classes significantly impacts their vulnerability to attacks.\nThrough our analysis on the Class False Positive Score, we assess a fair\nevaluation of how susceptible each class is to misclassification.\n", "date": "2024-11-29", "categories": []}
{"id": "2412.02171", "title": "Can't Slow me Down: Learning Robust and Hardware-Adaptive Object\n  Detectors against Latency Attacks for Edge Devices", "abstract": "  Object detection is a fundamental enabler for many real-time downstream\napplications such as autonomous driving, augmented reality and supply chain\nmanagement. However, the algorithmic backbone of neural networks is brittle to\nimperceptible perturbations in the system inputs, which were generally known as\nmisclassifying attacks. By targeting the real-time processing capability, a new\nclass of latency attacks are reported recently. They exploit new attack\nsurfaces in object detectors by creating a computational bottleneck in the\npost-processing module, that leads to cascading failure and puts the real-time\ndownstream tasks at risks. In this work, we take an initial attempt to defend\nagainst this attack via background-attentive adversarial training that is also\ncognizant of the underlying hardware capabilities. We first draw system-level\nconnections between latency attack and hardware capacity across heterogeneous\nGPU devices. Based on the particular adversarial behaviors, we utilize\nobjectness loss as a proxy and build background attention into the adversarial\ntraining pipeline, and achieve a reasonable balance between clean and robust\naccuracy. The extensive experiments demonstrate the defense effectiveness of\nrestoring real-time processing capability from $13$ FPS to $43$ FPS on Jetson\nOrin NX, with a better trade-off between the clean and robust accuracy.\n", "date": "2024-12-03", "categories": []}
{"id": "2503.04385", "title": "Scale-Invariant Adversarial Attack against Arbitrary-scale\n  Super-resolution", "abstract": "  The advent of local continuous image function (LIIF) has garnered significant\nattention for arbitrary-scale super-resolution (SR) techniques. However, while\nthe vulnerabilities of fixed-scale SR have been assessed, the robustness of\ncontinuous representation-based arbitrary-scale SR against adversarial attacks\nremains an area warranting further exploration. The elaborately designed\nadversarial attacks for fixed-scale SR are scale-dependent, which will cause\ntime-consuming and memory-consuming problems when applied to arbitrary-scale\nSR. To address this concern, we propose a simple yet effective\n``scale-invariant'' SR adversarial attack method with good transferability,\ntermed SIAGT. Specifically, we propose to construct resource-saving attacks by\nexploiting finite discrete points of continuous representation. In addition, we\nformulate a coordinate-dependent loss to enhance the cross-model\ntransferability of the attack. The attack can significantly deteriorate the SR\nimages while introducing imperceptible distortion to the targeted\nlow-resolution (LR) images. Experiments carried out on three popular LIIF-based\nSR approaches and four classical SR datasets show remarkable attack performance\nand transferability of SIAGT.\n", "date": "2025-03-06", "categories": []}
{"id": "2503.06529", "title": "AnywhereDoor: Multi-Target Backdoor Attacks on Object Detection", "abstract": "  As object detection becomes integral to many safety-critical applications,\nunderstanding its vulnerabilities is essential. Backdoor attacks, in\nparticular, pose a serious threat by implanting hidden triggers in victim\nmodels, which adversaries can later exploit to induce malicious behaviors\nduring inference. However, current understanding is limited to single-target\nattacks, where adversaries must define a fixed malicious behavior (target)\nbefore training, making inference-time adaptability impossible. Given the large\noutput space of object detection (including object existence prediction,\nbounding box estimation, and classification), the feasibility of flexible,\ninference-time model control remains unexplored. This paper introduces\nAnywhereDoor, a multi-target backdoor attack for object detection. Once\nimplanted, AnywhereDoor allows adversaries to make objects disappear, fabricate\nnew ones, or mislabel them, either across all object classes or specific ones,\noffering an unprecedented degree of control. This flexibility is enabled by\nthree key innovations: (i) objective disentanglement to scale the number of\nsupported targets; (ii) trigger mosaicking to ensure robustness even against\nregion-based detectors; and (iii) strategic batching to address object-level\ndata imbalances that hinder manipulation. Extensive experiments demonstrate\nthat AnywhereDoor grants attackers a high degree of control, improving attack\nsuccess rates by 26% compared to adaptations of existing methods for such\nflexible control.\n", "date": "2025-03-09", "categories": []}
{"id": "2503.08708", "title": "TH-Bench: Evaluating Evading Attacks via Humanizing AI Text on\n  Machine-Generated Text Detectors", "abstract": "  As Large Language Models (LLMs) advance, Machine-Generated Texts (MGTs) have\nbecome increasingly fluent, high-quality, and informative. Existing wide-range\nMGT detectors are designed to identify MGTs to prevent the spread of plagiarism\nand misinformation. However, adversaries attempt to humanize MGTs to evade\ndetection (named evading attacks), which requires only minor modifications to\nbypass MGT detectors. Unfortunately, existing attacks generally lack a unified\nand comprehensive evaluation framework, as they are assessed using different\nexperimental settings, model architectures, and datasets. To fill this gap, we\nintroduce the Text-Humanization Benchmark (TH-Bench), the first comprehensive\nbenchmark to evaluate evading attacks against MGT detectors. TH-Bench evaluates\nattacks across three key dimensions: evading effectiveness, text quality, and\ncomputational overhead. Our extensive experiments evaluate 6 state-of-the-art\nattacks against 13 MGT detectors across 6 datasets, spanning 19 domains and\ngenerated by 11 widely used LLMs. Our findings reveal that no single evading\nattack excels across all three dimensions. Through in-depth analysis, we\nhighlight the strengths and limitations of different attacks. More importantly,\nwe identify a trade-off among three dimensions and propose two optimization\ninsights. Through preliminary experiments, we validate their correctness and\neffectiveness, offering potential directions for future research.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.09669", "title": "Silent Branding Attack: Trigger-free Data Poisoning Attack on\n  Text-to-Image Diffusion Models", "abstract": "  Text-to-image diffusion models have achieved remarkable success in generating\nhigh-quality contents from text prompts. However, their reliance on publicly\navailable data and the growing trend of data sharing for fine-tuning make these\nmodels particularly vulnerable to data poisoning attacks. In this work, we\nintroduce the Silent Branding Attack, a novel data poisoning method that\nmanipulates text-to-image diffusion models to generate images containing\nspecific brand logos or symbols without any text triggers. We find that when\ncertain visual patterns are repeatedly in the training data, the model learns\nto reproduce them naturally in its outputs, even without prompt mentions.\nLeveraging this, we develop an automated data poisoning algorithm that\nunobtrusively injects logos into original images, ensuring they blend naturally\nand remain undetected. Models trained on this poisoned dataset generate images\ncontaining logos without degrading image quality or text alignment. We\nexperimentally validate our silent branding attack across two realistic\nsettings on large-scale high-quality image datasets and style personalization\ndatasets, achieving high success rates even without a specific text trigger.\nHuman evaluation and quantitative metrics including logo detection show that\nour method can stealthily embed logos.\n", "date": "2025-03-12", "categories": []}
{"id": "2503.09735", "title": "Enhancing Adversarial Example Detection Through Model Explanation", "abstract": "  Adversarial examples are a major problem for machine learning models, leading\nto a continuous search for effective defenses. One promising direction is to\nleverage model explanations to better understand and defend against these\nattacks. We looked at AmI, a method proposed by a NeurIPS 2018 spotlight paper\nthat uses model explanations to detect adversarial examples. Our study shows\nthat while AmI is a promising idea, its performance is too dependent on\nspecific settings (e.g., hyperparameter) and external factors such as the\noperating system and the deep learning framework used, and such drawbacks limit\nAmI's practical usage. Our findings highlight the need for more robust defense\nmechanisms that are effective under various conditions. In addition, we\nadvocate for a comprehensive evaluation framework for defense techniques.\n", "date": "2025-03-12", "categories": []}
{"id": "2503.10081", "title": "AdvPaint: Protecting Images from Inpainting Manipulation via Adversarial\n  Attention Disruption", "abstract": "  The outstanding capability of diffusion models in generating high-quality\nimages poses significant threats when misused by adversaries. In particular, we\nassume malicious adversaries exploiting diffusion models for inpainting tasks,\nsuch as replacing a specific region with a celebrity. While existing methods\nfor protecting images from manipulation in diffusion-based generative models\nhave primarily focused on image-to-image and text-to-image tasks, the challenge\nof preventing unauthorized inpainting has been rarely addressed, often\nresulting in suboptimal protection performance. To mitigate inpainting abuses,\nwe propose ADVPAINT, a novel defensive framework that generates adversarial\nperturbations that effectively disrupt the adversary's inpainting tasks.\nADVPAINT targets the self- and cross-attention blocks in a target diffusion\ninpainting model to distract semantic understanding and prompt interactions\nduring image generation. ADVPAINT also employs a two-stage perturbation\nstrategy, dividing the perturbation region based on an enlarged bounding box\naround the object, enhancing robustness across diverse masks of varying shapes\nand sizes. Our experimental results demonstrate that ADVPAINT's perturbations\nare highly effective in disrupting the adversary's inpainting tasks,\noutperforming existing methods; ADVPAINT attains over a 100-point increase in\nFID and substantial decreases in precision.\n", "date": "2025-03-13", "categories": []}
{"id": "2503.10191", "title": "Robustness Tokens: Towards Adversarial Robustness of Transformers", "abstract": "  Recently, large pre-trained foundation models have become widely adopted by\nmachine learning practitioners for a multitude of tasks. Given that such models\nare publicly available, relying on their use as backbone models for downstream\ntasks might result in high vulnerability to adversarial attacks crafted with\nthe same public model. In this work, we propose Robustness Tokens, a novel\napproach specific to the transformer architecture that fine-tunes a few\nadditional private tokens with low computational requirements instead of tuning\nmodel parameters as done in traditional adversarial training. We show that\nRobustness Tokens make Vision Transformer models significantly more robust to\nwhite-box adversarial attacks while also retaining the original downstream\nperformances.\n", "date": "2025-03-13", "categories": []}
{"id": "2503.10228", "title": "Policy Teaching via Data Poisoning in Learning from Human Preferences", "abstract": "  We study data poisoning attacks in learning from human preferences. More\nspecifically, we consider the problem of teaching/enforcing a target policy\n$\\pi^\\dagger$ by synthesizing preference data. We seek to understand the\nsusceptibility of different preference-based learning paradigms to poisoned\npreference data by analyzing the number of samples required by the attacker to\nenforce $\\pi^\\dagger$. We first propose a general data poisoning formulation in\nlearning from human preferences and then study it for two popular paradigms,\nnamely: (a) reinforcement learning from human feedback (RLHF) that operates by\nlearning a reward model using preferences; (b) direct preference optimization\n(DPO) that directly optimizes policy using preferences. We conduct a\ntheoretical analysis of the effectiveness of data poisoning in a setting where\nthe attacker is allowed to augment a pre-existing dataset and also study its\nspecial case where the attacker can synthesize the entire preference dataset\nfrom scratch. As our main results, we provide lower/upper bounds on the number\nof samples required to enforce $\\pi^\\dagger$. Finally, we discuss the\nimplications of our results in terms of the susceptibility of these learning\nparadigms under such data poisoning attacks.\n", "date": "2025-03-13", "categories": []}
{"id": "2412.02371", "title": "TSCheater: Generating High-Quality Tibetan Adversarial Texts via Visual\n  Similarity", "abstract": "  Language models based on deep neural networks are vulnerable to textual\nadversarial attacks. While rich-resource languages like English are receiving\nfocused attention, Tibetan, a cross-border language, is gradually being studied\ndue to its abundant ancient literature and critical language strategy.\nCurrently, there are several Tibetan adversarial text generation methods, but\nthey do not fully consider the textual features of Tibetan script and\noverestimate the quality of generated adversarial texts. To address this issue,\nwe propose a novel Tibetan adversarial text generation method called TSCheater,\nwhich considers the characteristic of Tibetan encoding and the feature that\nvisually similar syllables have similar semantics. This method can also be\ntransferred to other abugidas, such as Devanagari script. We utilize a\nself-constructed Tibetan syllable visual similarity database called TSVSDB to\ngenerate substitution candidates and adopt a greedy algorithm-based scoring\nmechanism to determine substitution order. After that, we conduct the method on\neight victim language models. Experimentally, TSCheater outperforms existing\nmethods in attack effectiveness, perturbation magnitude, semantic similarity,\nvisual similarity, and human acceptance. Finally, we construct the first\nTibetan adversarial robustness evaluation benchmark called AdvTS, which is\ngenerated by existing methods and proofread by humans.\n", "date": "2024-12-03", "categories": []}
{"id": "2503.08704", "title": "Life-Cycle Routing Vulnerabilities of LLM Router", "abstract": "  Large language models (LLMs) have achieved remarkable success in natural\nlanguage processing, yet their performance and computational costs vary\nsignificantly. LLM routers play a crucial role in dynamically balancing these\ntrade-offs. While previous studies have primarily focused on routing\nefficiency, security vulnerabilities throughout the entire LLM router life\ncycle, from training to inference, remain largely unexplored. In this paper, we\npresent a comprehensive investigation into the life-cycle routing\nvulnerabilities of LLM routers. We evaluate both white-box and black-box\nadversarial robustness, as well as backdoor robustness, across several\nrepresentative routing models under extensive experimental settings. Our\nexperiments uncover several key findings: 1) Mainstream DNN-based routers tend\nto exhibit the weakest adversarial and backdoor robustness, largely due to\ntheir strong feature extraction capabilities that amplify vulnerabilities\nduring both training and inference; 2) Training-free routers demonstrate the\nstrongest robustness across different attack types, benefiting from the absence\nof learnable parameters that can be manipulated. These findings highlight\ncritical security risks spanning the entire life cycle of LLM routers and\nprovide insights for developing more robust models.\n", "date": "2025-03-09", "categories": []}
{"id": "2503.08716", "title": "AuthorMist: Evading AI Text Detectors with Reinforcement Learning", "abstract": "  In the age of powerful AI-generated text, automatic detectors have emerged to\nidentify machine-written content. This poses a threat to author privacy and\nfreedom, as text authored with AI assistance may be unfairly flagged. We\npropose AuthorMist, a novel reinforcement learning-based system to transform\nAI-generated text into human-like writing. AuthorMist leverages a\n3-billion-parameter language model as a backbone, fine-tuned with Group\nRelative Policy Optimization (GPRO) to paraphrase text in a way that evades AI\ndetectors.\n  Our framework establishes a generic approach where external detector APIs\n(GPTZero, WinstonAI, Originality.ai, etc.) serve as reward functions within the\nreinforcement learning loop, enabling the model to systematically learn outputs\nthat these detectors are less likely to classify as AI-generated. This\nAPI-as-reward methodology can be applied broadly to optimize text against any\ndetector with an accessible interface. Experiments on multiple datasets and\ndetectors demonstrate that AuthorMist effectively reduces the detectability of\nAI-generated text while preserving the original meaning. Our evaluation shows\nattack success rates ranging from 78.6% to 96.2% against individual detectors,\nsignificantly outperforming baseline paraphrasing methods. AuthorMist maintains\nhigh semantic similarity (above 0.94) with the original text while successfully\nevading detection. These results highlight limitations in current AI text\ndetection technologies and raise questions about the sustainability of the\ndetection-evasion arms race.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.08801", "title": "Enhanced Estimation Techniques for Certified Radii in Randomized\n  Smoothing", "abstract": "  This paper presents novel methods for estimating certified radii in\nrandomized smoothing, a technique crucial for certifying the robustness of\nneural networks against adversarial perturbations. Our proposed techniques\nsignificantly improve the accuracy of certified test-set accuracy by providing\ntighter bounds on the certified radii. We introduce advanced algorithms for\nboth discrete and continuous domains, demonstrating their effectiveness on\nCIFAR-10 and ImageNet datasets. The new methods show considerable improvements\nover existing approaches, particularly in reducing discrepancies in certified\nradii estimates. We also explore the impact of various hyperparameters,\nincluding sample size, standard deviation, and temperature, on the performance\nof these methods. Our findings highlight the potential for more efficient\ncertification processes and pave the way for future research on tighter\nconfidence sequences and improved theoretical frameworks. The study concludes\nwith a discussion of potential future directions, including enhanced estimation\ntechniques for discrete domains and further theoretical advancements to bridge\nthe gap between empirical and theoretical performance in randomized smoothing.\n", "date": "2025-03-11", "categories": []}
{"id": "2503.08973", "title": "Quantitative Analysis of Deeply Quantized Tiny Neural Networks Robust to\n  Adversarial Attacks", "abstract": "  Reducing the memory footprint of Machine Learning (ML) models, especially\nDeep Neural Networks (DNNs), is imperative to facilitate their deployment on\nresource-constrained edge devices. However, a notable drawback of DNN models\nlies in their susceptibility to adversarial attacks, wherein minor input\nperturbations can deceive them. A primary challenge revolves around the\ndevelopment of accurate, resilient, and compact DNN models suitable for\ndeployment on resource-constrained edge devices. This paper presents the\noutcomes of a compact DNN model that exhibits resilience against both black-box\nand white-box adversarial attacks. This work has achieved this resilience\nthrough training with the QKeras quantization-aware training framework. The\nstudy explores the potential of QKeras and an adversarial robustness technique,\nJacobian Regularization (JR), to co-optimize the DNN architecture through\nper-layer JR methodology. As a result, this paper has devised a DNN model\nemploying this co-optimization strategy based on Stochastic Ternary\nQuantization (STQ). Its performance was compared against existing DNN models in\nthe face of various white-box and black-box attacks. The experimental findings\nrevealed that, the proposed DNN model had small footprint and on average, it\nexhibited better performance than Quanos and DS-CNN MLCommons/TinyML (MLC/T)\nbenchmarks when challenged with white-box and black-box attacks, respectively,\non the CIFAR-10 image and Google Speech Commands audio datasets.\n", "date": "2025-03-12", "categories": []}
{"id": "2503.09124", "title": "AdvAD: Exploring Non-Parametric Diffusion for Imperceptible Adversarial\n  Attacks", "abstract": "  Imperceptible adversarial attacks aim to fool DNNs by adding imperceptible\nperturbation to the input data. Previous methods typically improve the\nimperceptibility of attacks by integrating common attack paradigms with\nspecifically designed perception-based losses or the capabilities of generative\nmodels. In this paper, we propose Adversarial Attacks in Diffusion (AdvAD), a\nnovel modeling framework distinct from existing attack paradigms. AdvAD\ninnovatively conceptualizes attacking as a non-parametric diffusion process by\ntheoretically exploring basic modeling approach rather than using the denoising\nor generation abilities of regular diffusion models requiring neural networks.\nAt each step, much subtler yet effective adversarial guidance is crafted using\nonly the attacked model without any additional network, which gradually leads\nthe end of diffusion process from the original image to a desired imperceptible\nadversarial example. Grounded in a solid theoretical foundation of the proposed\nnon-parametric diffusion process, AdvAD achieves high attack efficacy and\nimperceptibility with intrinsically lower overall perturbation strength.\nAdditionally, an enhanced version AdvAD-X is proposed to evaluate the extreme\nof our novel framework under an ideal scenario. Extensive experiments\ndemonstrate the effectiveness of the proposed AdvAD and AdvAD-X. Compared with\nstate-of-the-art imperceptible attacks, AdvAD achieves an average of 99.9$\\%$\n(+17.3$\\%$) ASR with 1.34 (-0.97) $l_2$ distance, 49.74 (+4.76) PSNR and 0.9971\n(+0.0043) SSIM against four prevalent DNNs with three different architectures\non the ImageNet-compatible dataset. Code is available at\nhttps://github.com/XianguiKang/AdvAD.\n", "date": "2025-03-12", "categories": []}
{"id": "2403.16067", "title": "Adversarial Guided Diffusion Models for Adversarial Purification", "abstract": "  Diffusion model (DM) based adversarial purification (AP) has proven to be a\npowerful defense method that can remove adversarial perturbations and generate\na purified example without threats. In principle, the pre-trained DMs can only\nensure that purified examples conform to the same distribution of the training\ndata, but it may inadvertently compromise the semantic information of input\nexamples, leading to misclassification of purified examples. Recent\nadvancements introduce guided diffusion techniques to preserve semantic\ninformation while removing the perturbations. However, these guidances often\nrely on distance measures between purified examples and diffused examples,\nwhich can also preserve perturbations in purified examples. To further unleash\nthe robustness power of DM-based AP, we propose an adversarial guided diffusion\nmodel (AGDM) by introducing a novel adversarial guidance that contains\nsufficient semantic information but does not explicitly involve adversarial\nperturbations. The guidance is modeled by an auxiliary neural network obtained\nwith adversarial training, considering the distance in the latent\nrepresentations rather than at the pixel-level values. Extensive experiments\nare conducted on CIFAR-10, CIFAR-100 and ImageNet to demonstrate that our\nmethod is effective for simultaneously maintaining semantic information and\nremoving the adversarial perturbations. In addition, comprehensive comparisons\nshow that our method significantly enhances the robustness of existing DM-based\nAP, with an average robust accuracy improved by up to 7.30% on CIFAR-10.\n", "date": "2024-03-24", "categories": []}
{"id": "2411.15210", "title": "Towards Million-Scale Adversarial Robustness Evaluation With Stronger\n  Individual Attacks", "abstract": "  As deep learning models are increasingly deployed in safety-critical\napplications, evaluating their vulnerabilities to adversarial perturbations is\nessential for ensuring their reliability and trustworthiness. Over the past\ndecade, a large number of white-box adversarial robustness evaluation methods\n(i.e., attacks) have been proposed, ranging from single-step to multi-step\nmethods and from individual to ensemble methods. Despite these advances,\nchallenges remain in conducting meaningful and comprehensive robustness\nevaluations, particularly when it comes to large-scale testing and ensuring\nevaluations reflect real-world adversarial risks. In this work, we focus on\nimage classification models and propose a novel individual attack method,\nProbability Margin Attack (PMA), which defines the adversarial margin in the\nprobability space rather than the logits space. We analyze the relationship\nbetween PMA and existing cross-entropy or logits-margin-based attacks, and show\nthat PMA can outperform the current state-of-the-art individual methods.\nBuilding on PMA, we propose two types of ensemble attacks that balance\neffectiveness and efficiency. Furthermore, we create a million-scale dataset,\nCC1M, derived from the existing CC3M dataset, and use it to conduct the first\nmillion-scale white-box adversarial robustness evaluation of\nadversarially-trained ImageNet models. Our findings provide valuable insights\ninto the robustness gaps between individual versus ensemble attacks and\nsmall-scale versus million-scale evaluations.\n", "date": "2024-11-20", "categories": []}
{"id": "2412.08014", "title": "MAGIC: Mastering Physical Adversarial Generation in Context through\n  Collaborative LLM Agents", "abstract": "  Physical adversarial attacks in driving scenarios can expose critical\nvulnerabilities in visual perception models. However, developing such attacks\nremains challenging due to diverse real-world environments and the requirement\nfor maintaining visual naturality. Building upon this challenge, we reformulate\nphysical adversarial attacks as a one-shot patch generation problem. Our\napproach generates adversarial patches through a deep generative model that\nconsiders the specific scene context, enabling direct physical deployment in\nmatching environments. The primary challenge lies in simultaneously achieving\ntwo objectives: generating adversarial patches that effectively mislead object\ndetection systems while determining contextually appropriate deployment within\nthe scene. We propose MAGIC (Mastering Physical Adversarial Generation In\nContext), a novel framework powered by multi-modal LLM agents to address these\nchallenges. MAGIC automatically understands scene context and generates\nadversarial patch through the synergistic interaction of language and vision\ncapabilities. In particular, MAGIC orchestrates three specialized LLM agents:\nThe adv-patch generation agent (GAgent) masters the creation of deceptive\npatches through strategic prompt engineering for text-to-image models. The\nadv-patch deployment agent (DAgent) ensures contextual coherence by determining\noptimal deployment strategies based on scene understanding. The\nself-examination agent (EAgent) completes this trilogy by providing critical\noversight and iterative refinement of both processes. We validate our method on\nboth digital and physical levels, i.e., nuImage and manually captured\nreal-world scenes, where both statistical and visual results prove that our\nMAGIC is powerful and effective for attacking widely applied object detection\nsystems, i.e., YOLO and DETR series.\n", "date": "2024-12-11", "categories": []}
{"id": "2503.06966", "title": "MIGA: Mutual Information-Guided Attack on Denoising Models for Semantic\n  Manipulation", "abstract": "  Deep learning-based denoising models have been widely employed in vision\ntasks, functioning as filters to eliminate noise while retaining crucial\nsemantic information. Additionally, they play a vital role in defending against\nadversarial perturbations that threaten downstream tasks. However, these models\ncan be intrinsically susceptible to adversarial attacks due to their dependence\non specific noise assumptions. Existing attacks on denoising models mainly aim\nat deteriorating visual clarity while neglecting semantic manipulation,\nrendering them either easily detectable or limited in effectiveness. In this\npaper, we propose Mutual Information-Guided Attack (MIGA), the first method\ndesigned to directly attack deep denoising models by strategically disrupting\ntheir ability to preserve semantic content via adversarial perturbations. By\nminimizing the mutual information between the original and denoised images, a\nmeasure of semantic similarity. MIGA forces the denoiser to produce\nperceptually clean yet semantically altered outputs. While these images appear\nvisually plausible, they encode systematically distorted semantics, revealing a\nfundamental vulnerability in denoising models. These distortions persist in\ndenoised outputs and can be quantitatively assessed through downstream task\nperformance. We propose new evaluation metrics and systematically assess MIGA\non four denoising models across five datasets, demonstrating its consistent\neffectiveness in disrupting semantic fidelity. Our findings suggest that\ndenoising models are not always robust and can introduce security risks in\nreal-world applications.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.07818", "title": "Strengthening the Internal Adversarial Robustness in Lifted Neural\n  Networks", "abstract": "  Lifted neural networks (i.e. neural architectures explicitly optimizing over\nrespective network potentials to determine the neural activities) can be\ncombined with a type of adversarial training to gain robustness for internal as\nwell as input layers, in addition to improved generalization performance. In\nthis work we first investigate how adversarial robustness in this framework can\nbe further strengthened by solely modifying the training loss. In a second step\nwe fix some remaining limitations and arrive at a novel training loss for\nlifted neural networks, that combines targeted and untargeted adversarial\nperturbations.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.08195", "title": "Dialogue Injection Attack: Jailbreaking LLMs through Context\n  Manipulation", "abstract": "  Large language models (LLMs) have demonstrated significant utility in a wide\nrange of applications; however, their deployment is plagued by security\nvulnerabilities, notably jailbreak attacks. These attacks manipulate LLMs to\ngenerate harmful or unethical content by crafting adversarial prompts. While\nmuch of the current research on jailbreak attacks has focused on single-turn\ninteractions, it has largely overlooked the impact of historical dialogues on\nmodel behavior. In this paper, we introduce a novel jailbreak paradigm,\nDialogue Injection Attack (DIA), which leverages the dialogue history to\nenhance the success rates of such attacks. DIA operates in a black-box setting,\nrequiring only access to the chat API or knowledge of the LLM's chat template.\nWe propose two methods for constructing adversarial historical dialogues: one\nadapts gray-box prefilling attacks, and the other exploits deferred responses.\nOur experiments show that DIA achieves state-of-the-art attack success rates on\nrecent LLMs, including Llama-3.1 and GPT-4o. Additionally, we demonstrate that\nDIA can bypass 5 different defense mechanisms, highlighting its robustness and\neffectiveness.\n", "date": "2025-03-11", "categories": []}
{"id": "2503.08226", "title": "A Grey-box Text Attack Framework using Explainable AI", "abstract": "  Explainable AI is a strong strategy implemented to understand complex\nblack-box model predictions in a human interpretable language. It provides the\nevidence required to execute the use of trustworthy and reliable AI systems. On\nthe other hand, however, it also opens the door to locating possible\nvulnerabilities in an AI model. Traditional adversarial text attack uses word\nsubstitution, data augmentation techniques and gradient-based attacks on\npowerful pre-trained Bidirectional Encoder Representations from Transformers\n(BERT) variants to generate adversarial sentences. These attacks are generally\nwhitebox in nature and not practical as they can be easily detected by humans\nE.g. Changing the word from \"Poor\" to \"Rich\". We proposed a simple yet\neffective Grey-box cum Black-box approach that does not require the knowledge\nof the model while using a set of surrogate Transformer/BERT models to perform\nthe attack using Explainable AI techniques. As Transformers are the current\nstate-of-the-art models for almost all Natural Language Processing (NLP) tasks,\nan attack generated from BERT1 is transferable to BERT2. This transferability\nis made possible due to the attention mechanism in the transformer that allows\nthe model to capture long-range dependencies in a sequence. Using the power of\nBERT generalisation via attention, we attempt to exploit how transformers learn\nby attacking a few surrogate transformer variants which are all based on a\ndifferent architecture. We demonstrate that this approach is highly effective\nto generate semantically good sentences by changing as little as one word that\nis not detectable by humans while still fooling other BERT models.\n", "date": "2025-03-11", "categories": []}
{"id": "2503.08269", "title": "Adv-CPG: A Customized Portrait Generation Framework with Facial\n  Adversarial Attacks", "abstract": "  Recent Customized Portrait Generation (CPG) methods, taking a facial image\nand a textual prompt as inputs, have attracted substantial attention. Although\nthese methods generate high-fidelity portraits, they fail to prevent the\ngenerated portraits from being tracked and misused by malicious face\nrecognition systems. To address this, this paper proposes a Customized Portrait\nGeneration framework with facial Adversarial attacks (Adv-CPG). Specifically,\nto achieve facial privacy protection, we devise a lightweight local ID\nencryptor and an encryption enhancer. They implement progressive double-layer\nencryption protection by directly injecting the target identity and adding\nadditional identity guidance, respectively. Furthermore, to accomplish\nfine-grained and personalized portrait generation, we develop a multi-modal\nimage customizer capable of generating controlled fine-grained facial features.\nTo the best of our knowledge, Adv-CPG is the first study that introduces facial\nadversarial attacks into CPG. Extensive experiments demonstrate the superiority\nof Adv-CPG, e.g., the average attack success rate of the proposed Adv-CPG is\n28.1% and 2.86% higher compared to the SOTA noise-based attack methods and\nunconstrained attack methods, respectively.\n", "date": "2025-03-11", "categories": []}
{"id": "2405.16918", "title": "The Uncanny Valley: Exploring Adversarial Robustness from a Flatness\n  Perspective", "abstract": "  Flatness of the loss surface not only correlates positively with\ngeneralization, but is also related to adversarial robustness since\nperturbations of inputs relate non-linearly to perturbations of weights. In\nthis paper, we empirically analyze the relation between adversarial examples\nand relative flatness with respect to the parameters of one layer. We observe a\npeculiar property of adversarial examples in the context of relative flatness:\nduring an iterative first-order white-box attack, the flatness of the loss\nsurface measured around the adversarial example first becomes sharper until the\nlabel is flipped, but if we keep the attack running, it runs into a flat\nuncanny valley where the label remains flipped. In extensive experiments, we\nobserve this phenomenon across various model architectures and datasets, even\nfor adversarially trained models. Our results also extend to large language\nmodels (LLMs), but due to the discrete nature of the input space and\ncomparatively weak attacks, adversarial examples rarely reach truly flat\nregions. Most importantly, this phenomenon shows that flatness alone cannot\nexplain adversarial robustness unless we can also guarantee the behavior of the\nfunction around the examples. We, therefore theoretically connect relative\nflatness to adversarial robustness by bounding the third derivative of the loss\nsurface, underlining the need for flatness in combination with a low global\nLipschitz constant for a robust model.\n", "date": "2024-05-27", "categories": []}
{"id": "2406.05870", "title": "Machine Against the RAG: Jamming Retrieval-Augmented Generation with\n  Blocker Documents", "abstract": "  Retrieval-augmented generation (RAG) systems respond to queries by retrieving\nrelevant documents from a knowledge database and applying an LLM to the\nretrieved documents. We demonstrate that RAG systems that operate on databases\nwith untrusted content are vulnerable to denial-of-service attacks we call\njamming. An adversary can add a single ``blocker'' document to the database\nthat will be retrieved in response to a specific query and result in the RAG\nsystem not answering this query, ostensibly because it lacks relevant\ninformation or because the answer is unsafe.\n  We describe and measure the efficacy of several methods for generating\nblocker documents, including a new method based on black-box optimization. Our\nmethod (1) does not rely on instruction injection, (2) does not require the\nadversary to know the embedding or LLM used by the target RAG system, and (3)\ndoes not employ an auxiliary LLM.\n  We evaluate jamming attacks on several embeddings and LLMs and demonstrate\nthat the existing safety metrics for LLMs do not capture their vulnerability to\njamming. We then discuss defenses against blocker documents.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.08443", "title": "Transform-Dependent Adversarial Attacks", "abstract": "  Deep networks are highly vulnerable to adversarial attacks, yet conventional\nattack methods utilize static adversarial perturbations that induce fixed\nmispredictions. In this work, we exploit an overlooked property of adversarial\nperturbations--their dependence on image transforms--and introduce\ntransform-dependent adversarial attacks. Unlike traditional attacks, our\nperturbations exhibit metamorphic properties, enabling diverse adversarial\neffects as a function of transformation parameters. We demonstrate that this\ntransform-dependent vulnerability exists across different architectures (e.g.,\nCNN and transformer), vision tasks (e.g., image classification and object\ndetection), and a wide range of image transforms. Additionally, we show that\ntransform-dependent perturbations can serve as a defense mechanism, preventing\nsensitive information disclosure when image enhancement transforms pose a risk\nof revealing private content. Through analysis in blackbox and defended model\nsettings, we show that transform-dependent perturbations achieve high targeted\nattack success rates, outperforming state-of-the-art transfer attacks by 17-31%\nin blackbox scenarios. Our work introduces novel, controllable paradigm for\nadversarial attack deployment, revealing a previously overlooked vulnerability\nin deep networks.\n", "date": "2024-06-12", "categories": []}
{"id": "2407.02437", "title": "Beyond Full Poisoning: Effective Availability Attacks with Partial\n  Perturbation", "abstract": "  The widespread use of publicly available datasets for training machine\nlearning models raises significant concerns about data misuse. Availability\nattacks have emerged as a means for data owners to safeguard their data by\ndesigning imperceptible perturbations that degrade model performance when\nincorporated into training datasets. However, existing availability attacks are\nineffective when only a portion of the data can be perturbed. To address this\nchallenge, we propose a novel availability attack approach termed Parameter\nMatching Attack (PMA). PMA is the first availability attack capable of causing\nmore than a 30\\% performance drop when only a portion of data can be perturbed.\nPMA optimizes perturbations so that when the model is trained on a mixture of\nclean and perturbed data, the resulting model will approach a model designed to\nperform poorly. Experimental results across four datasets demonstrate that PMA\noutperforms existing methods, achieving significant model performance\ndegradation when a part of the training data is perturbed. Our code is\navailable in the supplementary materials.\n", "date": "2024-07-02", "categories": []}
{"id": "2412.11168", "title": "PGD-Imp: Rethinking and Unleashing Potential of Classic PGD with Dual\n  Strategies for Imperceptible Adversarial Attacks", "abstract": "  Imperceptible adversarial attacks have recently attracted increasing research\ninterests. Existing methods typically incorporate external modules or loss\nterms other than a simple $l_p$-norm into the attack process to achieve\nimperceptibility, while we argue that such additional designs may not be\nnecessary. In this paper, we rethink the essence of imperceptible attacks and\npropose two simple yet effective strategies to unleash the potential of PGD,\nthe common and classical attack, for imperceptibility from an optimization\nperspective. Specifically, the Dynamic Step Size is introduced to find the\noptimal solution with minimal attack cost towards the decision boundary of the\nattacked model, and the Adaptive Early Stop strategy is adopted to reduce the\nredundant strength of adversarial perturbations to the minimum level. The\nproposed PGD-Imperceptible (PGD-Imp) attack achieves state-of-the-art results\nin imperceptible adversarial attacks for both untargeted and targeted\nscenarios. When performing untargeted attacks against ResNet-50, PGD-Imp\nattains 100$\\%$ (+0.3$\\%$) ASR, 0.89 (-1.76) $l_2$ distance, and 52.93 (+9.2)\nPSNR with 57s (-371s) running time, significantly outperforming existing\nmethods.\n", "date": "2024-12-15", "categories": []}
{"id": "2501.04802", "title": "Reproducing HotFlip for Corpus Poisoning Attacks in Dense Retrieval", "abstract": "  HotFlip is a topical gradient-based word substitution method for attacking\nlanguage models. Recently, this method has been further applied to attack\nretrieval systems by generating malicious passages that are injected into a\ncorpus, i.e., corpus poisoning. However, HotFlip is known to be computationally\ninefficient, with the majority of time being spent on gradient accumulation for\neach query-passage pair during the adversarial token generation phase, making\nit impossible to generate an adequate number of adversarial passages in a\nreasonable amount of time. Moreover, the attack method itself assumes access to\na set of user queries, a strong assumption that does not correspond to how\nreal-world adversarial attacks are usually performed. In this paper, we first\nsignificantly boost the efficiency of HotFlip, reducing the adversarial\ngeneration process from 4 hours per document to only 15 minutes, using the same\nhardware. We further contribute experiments and analysis on two additional\ntasks: (1) transfer-based black-box attacks, and (2) query-agnostic attacks.\nWhenever possible, we provide comparisons between the original method and our\nimproved version. Our experiments demonstrate that HotFlip can effectively\nattack a variety of dense retrievers, with an observed trend that its attack\nperformance diminishes against more advanced and recent methods. Interestingly,\nwe observe that while HotFlip performs poorly in a black-box setting,\nindicating limited capacity for generalization, in query-agnostic scenarios its\nperformance is correlated to the volume of injected adversarial passages.\n", "date": "2025-01-08", "categories": []}
{"id": "2502.14833", "title": "Probabilistic Robustness in Deep Learning: A Concise yet Comprehensive\n  Guide", "abstract": "  Deep learning (DL) has demonstrated significant potential across various\nsafety-critical applications, yet ensuring its robustness remains a key\nchallenge. While adversarial robustness has been extensively studied in\nworst-case scenarios, probabilistic robustness (PR) offers a more practical\nperspective by quantifying the likelihood of failures under stochastic\nperturbations. This paper provides a concise yet comprehensive overview of PR,\ncovering its formal definitions, evaluation and enhancement methods. We\nintroduce a reformulated ''min-max'' optimisation framework for adversarial\ntraining specifically designed to improve PR. Furthermore, we explore the\nintegration of PR verification evidence into system-level safety assurance,\naddressing challenges in translating DL model-level robustness to system-level\nclaims. Finally, we highlight open research questions, including benchmarking\nPR evaluation methods, extending PR to generative AI tasks, and developing\nrigorous methodologies and case studies for system-level integration.\n", "date": "2025-02-20", "categories": []}
{"id": "2503.06340", "title": "Backdoor Attacks on Discrete Graph Diffusion Models", "abstract": "  Diffusion models are powerful generative models in continuous data domains\nsuch as image and video data. Discrete graph diffusion models (DGDMs) have\nrecently extended them for graph generation, which are crucial in fields like\nmolecule and protein modeling, and obtained the SOTA performance. However, it\nis risky to deploy DGDMs for safety-critical applications (e.g., drug\ndiscovery) without understanding their security vulnerabilities. In this work,\nwe perform the first study on graph diffusion models against backdoor attacks,\na severe attack that manipulates both the training and inference/generation\nphases in graph diffusion models. We first define the threat model, under which\nwe design the attack such that the backdoored graph diffusion model can\ngenerate 1) high-quality graphs without backdoor activation, 2) effective,\nstealthy, and persistent backdoored graphs with backdoor activation, and 3)\ngraphs that are permutation invariant and exchangeable--two core properties in\ngraph generative models. 1) and 2) are validated via empirical evaluations\nwithout and with backdoor defenses, while 3) is validated via theoretical\nresults.\n", "date": "2025-03-08", "categories": []}
{"id": "2503.06559", "title": "MMARD: Improving the Min-Max Optimization Process in Adversarial\n  Robustness Distillation", "abstract": "  Adversarial Robustness Distillation (ARD) is a promising task to boost the\nrobustness of small-capacity models with the guidance of the pre-trained robust\nteacher. The ARD can be summarized as a min-max optimization process, i.e.,\nsynthesizing adversarial examples (inner) & training the student (outer).\nAlthough competitive robustness performance, existing ARD methods still have\nissues. In the inner process, the synthetic training examples are far from the\nteacher's decision boundary leading to important robust information missing. In\nthe outer process, the student model is decoupled from learning natural and\nrobust scenarios, leading to the robustness saturation, i.e., student\nperformance is highly susceptible to customized teacher selection. To tackle\nthese issues, this paper proposes a general Min-Max optimization Adversarial\nRobustness Distillation (MMARD) method. For the inner process, we introduce the\nteacher's robust predictions, which drive the training examples closer to the\nteacher's decision boundary to explore more robust knowledge. For the outer\nprocess, we propose a structured information modeling method based on\ntriangular relationships to measure the mutual information of the model in\nnatural and robust scenarios and enhance the model's ability to understand\nmulti-scenario mapping relationships. Experiments show our MMARD achieves\nstate-of-the-art performance on multiple benchmarks. Besides, MMARD is\nplug-and-play and convenient to combine with existing methods.\n", "date": "2025-03-09", "categories": []}
{"id": "2503.07058", "title": "Breaking the Limits of Quantization-Aware Defenses: QADT-R for\n  Robustness Against Patch-Based Adversarial Attacks in QNNs", "abstract": "  Quantized Neural Networks (QNNs) have emerged as a promising solution for\nreducing model size and computational costs, making them well-suited for\ndeployment in edge and resource-constrained environments. While quantization is\nknown to disrupt gradient propagation and enhance robustness against\npixel-level adversarial attacks, its effectiveness against patch-based\nadversarial attacks remains largely unexplored. In this work, we demonstrate\nthat adversarial patches remain highly transferable across quantized models,\nachieving over 70\\% attack success rates (ASR) even at extreme bit-width\nreductions (e.g., 2-bit). This challenges the common assumption that\nquantization inherently mitigates adversarial threats. To address this, we\npropose Quantization-Aware Defense Training with Randomization (QADT-R), a\nnovel defense strategy that integrates Adaptive Quantization-Aware Patch\nGeneration (A-QAPA), Dynamic Bit-Width Training (DBWT), and\nGradient-Inconsistent Regularization (GIR) to enhance resilience against highly\ntransferable patch-based attacks. A-QAPA generates adversarial patches within\nquantized models, ensuring robustness across different bit-widths. DBWT\nintroduces bit-width cycling during training to prevent overfitting to a\nspecific quantization setting, while GIR injects controlled gradient\nperturbations to disrupt adversarial optimization. Extensive evaluations on\nCIFAR-10 and ImageNet show that QADT-R reduces ASR by up to 25\\% compared to\nprior defenses such as PBAT and DWQ. Our findings further reveal that\nPBAT-trained models, while effective against seen patch configurations, fail to\ngeneralize to unseen patches due to quantization shift. Additionally, our\nempirical analysis of gradient alignment, spatial sensitivity, and patch\nvisibility provides insights into the mechanisms that contribute to the high\ntransferability of patch-based attacks in QNNs.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.07568", "title": "Runtime Detection of Adversarial Attacks in AI Accelerators Using\n  Performance Counters", "abstract": "  Rapid adoption of AI technologies raises several major security concerns,\nincluding the risks of adversarial perturbations, which threaten the\nconfidentiality and integrity of AI applications. Protecting AI hardware from\nmisuse and diverse security threats is a challenging task. To address this\nchallenge, we propose SAMURAI, a novel framework for safeguarding against\nmalicious usage of AI hardware and its resilience to attacks. SAMURAI\nintroduces an AI Performance Counter (APC) for tracking dynamic behavior of an\nAI model coupled with an on-chip Machine Learning (ML) analysis engine, known\nas TANTO (Trained Anomaly Inspection Through Trace Observation). APC records\nthe runtime profile of the low-level hardware events of different AI\noperations. Subsequently, the summary information recorded by the APC is\nprocessed by TANTO to efficiently identify potential security breaches and\nensure secure, responsible use of AI. SAMURAI enables real-time detection of\nsecurity threats and misuse without relying on traditional software-based\nsolutions that require model integration. Experimental results demonstrate that\nSAMURAI achieves up to 97% accuracy in detecting adversarial attacks with\nmoderate overhead on various AI models, significantly outperforming\nconventional software-based approaches. It enhances security and regulatory\ncompliance, providing a comprehensive solution for safeguarding AI against\nemergent threats.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.04853", "title": "From Pixels to Trajectory: Universal Adversarial Example Detection via\n  Temporal Imprints", "abstract": "  For the first time, we unveil discernible temporal (or historical) trajectory\nimprints resulting from adversarial example (AE) attacks. Standing in contrast\nto existing studies all focusing on spatial (or static) imprints within the\ntargeted underlying victim models, we present a fresh temporal paradigm for\nunderstanding these attacks. Of paramount discovery is that these imprints are\nencapsulated within a single loss metric, spanning universally across diverse\ntasks such as classification and regression, and modalities including image,\ntext, and audio. Recognizing the distinct nature of loss between adversarial\nand clean examples, we exploit this temporal imprint for AE detection by\nproposing TRAIT (TRaceable Adversarial temporal trajectory ImprinTs). TRAIT\noperates under minimal assumptions without prior knowledge of attacks, thereby\nframing the detection challenge as a one-class classification problem. However,\ndetecting AEs is still challenged by significant overlaps between the\nconstructed synthetic losses of adversarial and clean examples due to the\nabsence of ground truth for incoming inputs. TRAIT addresses this challenge by\nconverting the synthetic loss into a spectrum signature, using the technique of\nFast Fourier Transform to highlight the discrepancies, drawing inspiration from\nthe temporal nature of the imprints, analogous to time-series signals. Across\n12 AE attacks including SMACK (USENIX Sec'2023), TRAIT demonstrates consistent\noutstanding performance across comprehensively evaluated modalities, tasks,\ndatasets, and model architectures. In all scenarios, TRAIT achieves an AE\ndetection accuracy exceeding 97%, often around 99%, while maintaining a false\nrejection rate of 1%. TRAIT remains effective under the formulated strong\nadaptive attacks.\n", "date": "2025-03-06", "categories": []}
{"id": "2503.04963", "title": "Energy-Latency Attacks: A New Adversarial Threat to Deep Learning", "abstract": "  The growing computational demand for deep neural networks ( DNNs) has raised\nconcerns about their energy consumption and carbon footprint, particularly as\nthe size and complexity of the models continue to increase. To address these\nchallenges, energy-efficient hardware and custom accelerators have become\nessential. Additionally, adaptable DNN s are being developed to dynamically\nbalance performance and efficiency. The use of these strategies became more\ncommon to enable sustainable AI deployment. However, these efficiency-focused\ndesigns may also introduce vulnerabilities, as attackers can potentially\nexploit them to increase latency and energy usage by triggering their\nworst-case-performance scenarios. This new type of attack, called\nenergy-latency attacks, has recently gained significant research attention,\nfocusing on the vulnerability of DNN s to this emerging attack paradigm, which\ncan trigger denial-of-service ( DoS) attacks. This paper provides a\ncomprehensive overview of current research on energy-latency attacks,\ncategorizing them using the established taxonomy for traditional adversarial\nattacks. We explore different metrics used to measure the success of these\nattacks and provide an analysis and comparison of existing attack strategies.\nWe also analyze existing defense mechanisms and highlight current challenges\nand potential areas for future research in this developing field. The GitHub\npage for this work can be accessed at\nhttps://github.com/hbrachemi/Survey_energy_attacks/\n", "date": "2025-03-06", "categories": []}
{"id": "2503.05264", "title": "Jailbreaking is (Mostly) Simpler Than You Think", "abstract": "  We introduce the Context Compliance Attack (CCA), a novel, optimization-free\nmethod for bypassing AI safety mechanisms. Unlike current approaches -- which\nrely on complex prompt engineering and computationally intensive optimization\n-- CCA exploits a fundamental architectural vulnerability inherent in many\ndeployed AI systems. By subtly manipulating conversation history, CCA convinces\nthe model to comply with a fabricated dialogue context, thereby triggering\nrestricted behavior. Our evaluation across a diverse set of open-source and\nproprietary models demonstrates that this simple attack can circumvent\nstate-of-the-art safety protocols. We discuss the implications of these\nfindings and propose practical mitigation strategies to fortify AI systems\nagainst such elementary yet effective adversarial tactics.\n", "date": "2025-03-07", "categories": []}
{"id": "2312.03286", "title": "Indirect Gradient Matching for Adversarial Robust Distillation", "abstract": "  Adversarial training significantly improves adversarial robustness, but\nsuperior performance is primarily attained with large models. This substantial\nperformance gap for smaller models has spurred active research into adversarial\ndistillation (AD) to mitigate the difference. Existing AD methods leverage the\nteacher's logits as a guide. In contrast to these approaches, we aim to\ntransfer another piece of knowledge from the teacher, the input gradient. In\nthis paper, we propose a distillation module termed Indirect Gradient\nDistillation Module (IGDM) that indirectly matches the student's input gradient\nwith that of the teacher. Experimental results show that IGDM seamlessly\nintegrates with existing AD methods, significantly enhancing their performance.\nParticularly, utilizing IGDM on the CIFAR-100 dataset improves the AutoAttack\naccuracy from 28.06% to 30.32% with the ResNet-18 architecture and from 26.18%\nto 29.32% with the MobileNetV2 architecture when integrated into the SOTA\nmethod without additional data augmentation.\n", "date": "2023-12-06", "categories": []}
{"id": "2402.01879", "title": "$\\sigma$-zero: Gradient-based Optimization of $\\ell_0$-norm Adversarial\n  Examples", "abstract": "  Evaluating the adversarial robustness of deep networks to gradient-based\nattacks is challenging. While most attacks consider $\\ell_2$- and\n$\\ell_\\infty$-norm constraints to craft input perturbations, only a few\ninvestigate sparse $\\ell_1$- and $\\ell_0$-norm attacks. In particular,\n$\\ell_0$-norm attacks remain the least studied due to the inherent complexity\nof optimizing over a non-convex and non-differentiable constraint. However,\nevaluating adversarial robustness under these attacks could reveal weaknesses\notherwise left untested with more conventional $\\ell_2$- and $\\ell_\\infty$-norm\nattacks. In this work, we propose a novel $\\ell_0$-norm attack, called\n$\\sigma$-zero, which leverages a differentiable approximation of the $\\ell_0$\nnorm to facilitate gradient-based optimization, and an adaptive projection\noperator to dynamically adjust the trade-off between loss minimization and\nperturbation sparsity. Extensive evaluations using MNIST, CIFAR10, and ImageNet\ndatasets, involving robust and non-robust models, show that\n$\\sigma$\\texttt{-zero} finds minimum $\\ell_0$-norm adversarial examples without\nrequiring any time-consuming hyperparameter tuning, and that it outperforms all\ncompeting sparse attacks in terms of success rate, perturbation size, and\nefficiency.\n", "date": "2024-02-02", "categories": []}
{"id": "2410.00862", "title": "Timber! Poisoning Decision Trees", "abstract": "  We present Timber, the first white-box poisoning attack targeting decision\ntrees. Timber is based on a greedy attack strategy that leverages sub-tree\nretraining to efficiently estimate the damage caused by poisoning a given\ntraining instance. The attack relies on a tree annotation procedure, which\nenables the sorting of training instances so that they are processed in\nincreasing order of the computational cost of sub-tree retraining. This sorting\nyields a variant of Timber that supports an early stopping criterion, designed\nto make poisoning attacks more efficient and feasible on larger datasets. We\nalso discuss an extension of Timber to traditional random forest models, which\nis valuable since decision trees are typically combined into ensembles to\nimprove their predictive power. Our experimental evaluation on public datasets\ndemonstrates that our attacks outperform existing baselines in terms of\neffectiveness, efficiency, or both. Moreover, we show that two representative\ndefenses can mitigate the effect of our attacks, but fail to effectively thwart\nthem.\n", "date": "2024-10-01", "categories": []}
{"id": "2410.02163", "title": "Adversarial Decoding: Generating Readable Documents for Adversarial\n  Objectives", "abstract": "  We design, implement, and evaluate adversarial decoding, a new, generic text\ngeneration technique that produces readable documents for different adversarial\nobjectives. Prior methods either produce easily detectable gibberish, or cannot\nhandle objectives that include embedding similarity. In particular, they only\nwork for direct attacks (such as jailbreaking) and cannot produce adversarial\ntext for realistic indirect injection, e.g., documents that (1) are retrieved\nin RAG systems in response to broad classes of queries, and also (2)\nadversarially influence subsequent generation. We also show that fluency (low\nperplexity) is not sufficient to evade filtering. We measure the effectiveness\nof adversarial decoding for different objectives, including RAG poisoning,\njailbreaking, and evasion of defensive filters, and demonstrate that it\noutperforms existing methods while producing readable adversarial documents.\n", "date": "2024-10-03", "categories": []}
{"id": "2411.11006", "title": "BackdoorMBTI: A Backdoor Learning Multimodal Benchmark Tool Kit for\n  Backdoor Defense Evaluation", "abstract": "  Over the past few years, the emergence of backdoor attacks has presented\nsignificant challenges to deep learning systems, allowing attackers to insert\nbackdoors into neural networks. When data with a trigger is processed by a\nbackdoor model, it can lead to mispredictions targeted by attackers, whereas\nnormal data yields regular results. The scope of backdoor attacks is expanding\nbeyond computer vision and encroaching into areas such as natural language\nprocessing and speech recognition. Nevertheless, existing backdoor defense\nmethods are typically tailored to specific data modalities, restricting their\napplication in multimodal contexts. While multimodal learning proves highly\napplicable in facial recognition, sentiment analysis, action recognition,\nvisual question answering, the security of these models remains a crucial\nconcern. Specifically, there are no existing backdoor benchmarks targeting\nmultimodal applications or related tasks.\n  In order to facilitate the research in multimodal backdoor, we introduce\nBackdoorMBTI, the first backdoor learning toolkit and benchmark designed for\nmultimodal evaluation across three representative modalities from eleven\ncommonly used datasets. BackdoorMBTI provides a systematic backdoor learning\npipeline, encompassing data processing, data poisoning, backdoor training, and\nevaluation. The generated poison datasets and backdoor models enable detailed\nevaluation of backdoor defenses. Given the diversity of modalities,\nBackdoorMBTI facilitates systematic evaluation across different data types.\nFurthermore, BackdoorMBTI offers a standardized approach to handling practical\nfactors in backdoor learning, such as issues related to data quality and\nerroneous labels. We anticipate that BackdoorMBTI will expedite future research\nin backdoor defense methods within a multimodal context. Code is available at\nhttps://github.com/SJTUHaiyangYu/BackdoorMBTI.\n", "date": "2024-11-17", "categories": []}
{"id": "2502.01262", "title": "FSPGD: Rethinking Black-box Attacks on Semantic Segmentation", "abstract": "  Transferability, the ability of adversarial examples crafted for one model to\ndeceive other models, is crucial for black-box attacks. Despite advancements in\nattack methods for semantic segmentation, transferability remains limited,\nreducing their effectiveness in real-world applications. To address this, we\nintroduce the Feature Similarity Projected Gradient Descent (FSPGD) attack, a\nnovel black-box approach that enhances both attack performance and\ntransferability. Unlike conventional segmentation attacks that rely on output\npredictions for gradient calculation, FSPGD computes gradients from\nintermediate layer features. Specifically, our method introduces a loss\nfunction that targets local information by comparing features between clean\nimages and adversarial examples, while also disrupting contextual information\nby accounting for spatial relationships between objects. Experiments on Pascal\nVOC 2012 and Cityscapes datasets demonstrate that FSPGD achieves superior\ntransferability and attack performance, establishing a new state-of-the-art\nbenchmark. Code is available at https://github.com/KU-AIVS/FSPGD.\n", "date": "2025-02-03", "categories": []}
{"id": "2503.03842", "title": "Task-Agnostic Attacks Against Vision Foundation Models", "abstract": "  The study of security in machine learning mainly focuses on downstream\ntask-specific attacks, where the adversarial example is obtained by optimizing\na loss function specific to the downstream task. At the same time, it has\nbecome standard practice for machine learning practitioners to adopt publicly\navailable pre-trained vision foundation models, effectively sharing a common\nbackbone architecture across a multitude of applications such as\nclassification, segmentation, depth estimation, retrieval, question-answering\nand more. The study of attacks on such foundation models and their impact to\nmultiple downstream tasks remains vastly unexplored. This work proposes a\ngeneral framework that forges task-agnostic adversarial examples by maximally\ndisrupting the feature representation obtained with foundation models. We\nextensively evaluate the security of the feature representations obtained by\npopular vision foundation models by measuring the impact of this attack on\nmultiple downstream tasks and its transferability between models.\n", "date": "2025-03-05", "categories": []}
{"id": "2405.16226", "title": "Detecting Adversarial Data using Perturbation Forgery", "abstract": "  As a defense strategy against adversarial attacks, adversarial detection aims\nto identify and filter out adversarial data from the data flow based on\ndiscrepancies in distribution and noise patterns between natural and\nadversarial data. Although previous detection methods achieve high performance\nin detecting gradient-based adversarial attacks, new attacks based on\ngenerative models with imbalanced and anisotropic noise patterns evade\ndetection. Even worse, the significant inference time overhead and limited\nperformance against unseen attacks make existing techniques impractical for\nreal-world use. In this paper, we explore the proximity relationship among\nadversarial noise distributions and demonstrate the existence of an open\ncovering for these distributions. By training on the open covering of\nadversarial noise distributions, a detector with strong generalization\nperformance against various types of unseen attacks can be developed. Based on\nthis insight, we heuristically propose Perturbation Forgery, which includes\nnoise distribution perturbation, sparse mask generation, and pseudo-adversarial\ndata production, to train an adversarial detector capable of detecting any\nunseen gradient-based, generative-based, and physical adversarial attacks.\nComprehensive experiments conducted on multiple general and facial datasets,\nwith a wide spectrum of attacks, validate the strong generalization of our\nmethod.\n", "date": "2024-05-25", "categories": []}
{"id": "2503.03613", "title": "CLIP is Strong Enough to Fight Back: Test-time Counterattacks towards\n  Zero-shot Adversarial Robustness of CLIP", "abstract": "  Despite its prevalent use in image-text matching tasks in a zero-shot manner,\nCLIP has been shown to be highly vulnerable to adversarial perturbations added\nonto images. Recent studies propose to finetune the vision encoder of CLIP with\nadversarial samples generated on the fly, and show improved robustness against\nadversarial attacks on a spectrum of downstream datasets, a property termed as\nzero-shot robustness. In this paper, we show that malicious perturbations that\nseek to maximise the classification loss lead to `falsely stable' images, and\npropose to leverage the pre-trained vision encoder of CLIP to counterattack\nsuch adversarial images during inference to achieve robustness. Our paradigm is\nsimple and training-free, providing the first method to defend CLIP from\nadversarial attacks at test time, which is orthogonal to existing methods\naiming to boost zero-shot adversarial robustness of CLIP. We conduct\nexperiments across 16 classification datasets, and demonstrate stable and\nconsistent gains compared to test-time defence methods adapted from existing\nadversarial robustness studies that do not rely on external networks, without\nnoticeably impairing performance on clean images. We also show that our\nparadigm can be employed on CLIP models that have been adversarially finetuned\nto further enhance their robustness at test time. Our code is available\n\\href{https://github.com/Sxing2/CLIP-Test-time-Counterattacks}{here}.\n", "date": "2025-03-05", "categories": []}
{"id": "2308.10373", "title": "HoSNN: Adversarially-Robust Homeostatic Spiking Neural Networks with\n  Adaptive Firing Thresholds", "abstract": "  While spiking neural networks (SNNs) offer a promising neurally-inspired\nmodel of computation, they are vulnerable to adversarial attacks. We present\nthe first study that draws inspiration from neural homeostasis to design a\nthreshold-adapting leaky integrate-and-fire (TA-LIF) neuron model and utilize\nTA-LIF neurons to construct the adversarially robust homeostatic SNNs (HoSNNs)\nfor improved robustness. The TA-LIF model incorporates a self-stabilizing\ndynamic thresholding mechanism, offering a local feedback control solution to\nthe minimization of each neuron's membrane potential error caused by\nadversarial disturbance. Theoretical analysis demonstrates favorable dynamic\nproperties of TA-LIF neurons in terms of the bounded-input bounded-output\nstability and suppressed time growth of membrane potential error, underscoring\ntheir superior robustness compared with the standard LIF neurons. When trained\nwith weak FGSM attacks (attack budget = 2/255) and tested with much stronger\nPGD attacks (attack budget = 8/255), our HoSNNs significantly improve model\naccuracy on several datasets: from 30.54% to 74.91% on FashionMNIST, from 0.44%\nto 35.06% on SVHN, from 0.56% to 42.63% on CIFAR10, from 0.04% to 16.66% on\nCIFAR100, over the conventional LIF-based SNNs.\n", "date": "2023-08-20", "categories": []}
{"id": "2310.04285", "title": "Assessing Robustness via Score-Based Adversarial Image Generation", "abstract": "  Most adversarial attacks and defenses focus on perturbations within small\n$\\ell_p$-norm constraints. However, $\\ell_p$ threat models cannot capture all\nrelevant semantics-preserving perturbations, and hence, the scope of robustness\nevaluations is limited. In this work, we introduce Score-Based Adversarial\nGeneration (ScoreAG), a novel framework that leverages the advancements in\nscore-based generative models to generate unrestricted adversarial examples\nthat overcome the limitations of $\\ell_p$-norm constraints. Unlike traditional\nmethods, ScoreAG maintains the core semantics of images while generating\nadversarial examples, either by transforming existing images or synthesizing\nnew ones entirely from scratch. We further exploit the generative capability of\nScoreAG to purify images, empirically enhancing the robustness of classifiers.\nOur extensive empirical evaluation demonstrates that ScoreAG improves upon the\nmajority of state-of-the-art attacks and defenses across multiple benchmarks.\nThis work highlights the importance of investigating adversarial examples\nbounded by semantics rather than $\\ell_p$-norm constraints. ScoreAG represents\nan important step towards more encompassing robustness assessments.\n", "date": "2023-10-06", "categories": []}
{"id": "2411.04376", "title": "Game-Theoretic Defenses for Robust Conformal Prediction Against\n  Adversarial Attacks in Medical Imaging", "abstract": "  Adversarial attacks pose significant threats to the reliability and safety of\ndeep learning models, especially in critical domains such as medical imaging.\nThis paper introduces a novel framework that integrates conformal prediction\nwith game-theoretic defensive strategies to enhance model robustness against\nboth known and unknown adversarial perturbations. We address three primary\nresearch questions: constructing valid and efficient conformal prediction sets\nunder known attacks (RQ1), ensuring coverage under unknown attacks through\nconservative thresholding (RQ2), and determining optimal defensive strategies\nwithin a zero-sum game framework (RQ3). Our methodology involves training\nspecialized defensive models against specific attack types and employing\nmaximum and minimum classifiers to aggregate defenses effectively. Extensive\nexperiments conducted on the MedMNIST datasets, including PathMNIST,\nOrganAMNIST, and TissueMNIST, demonstrate that our approach maintains high\ncoverage guarantees while minimizing prediction set sizes. The game-theoretic\nanalysis reveals that the optimal defensive strategy often converges to a\nsingular robust model, outperforming uniform and simple strategies across all\nevaluated datasets. This work advances the state-of-the-art in uncertainty\nquantification and adversarial robustness, providing a reliable mechanism for\ndeploying deep learning models in adversarial environments.\n", "date": "2024-11-07", "categories": []}
{"id": "2401.15262", "title": "Asymptotic Behavior of Adversarial Training Estimator under\n  $\\ell_\\infty$-Perturbation", "abstract": "  Adversarial training has been proposed to protect machine learning models\nagainst adversarial attacks. This paper focuses on adversarial training under\n$\\ell_\\infty$-perturbation, which has recently attracted much research\nattention. The asymptotic behavior of the adversarial training estimator is\ninvestigated in the generalized linear model. The results imply that the\nasymptotic distribution of the adversarial training estimator under\n$\\ell_\\infty$-perturbation could put a positive probability mass at $0$ when\nthe true parameter is $0$, providing a theoretical guarantee of the associated\nsparsity-recovery ability. Alternatively, a two-step procedure is proposed --\nadaptive adversarial training, which could further improve the performance of\nadversarial training under $\\ell_\\infty$-perturbation. Specifically, the\nproposed procedure could achieve asymptotic variable-selection consistency and\nunbiasedness. Numerical experiments are conducted to show the sparsity-recovery\nability of adversarial training under $\\ell_\\infty$-perturbation and to compare\nthe empirical performance between classic adversarial training and adaptive\nadversarial training.\n", "date": "2024-01-27", "categories": []}
{"id": "2402.09154", "title": "Attacking Large Language Models with Projected Gradient Descent", "abstract": "  Current LLM alignment methods are readily broken through specifically crafted\nadversarial prompts. While crafting adversarial prompts using discrete\noptimization is highly effective, such attacks typically use more than 100,000\nLLM calls. This high computational cost makes them unsuitable for, e.g.,\nquantitative analyses and adversarial training. To remedy this, we revisit\nProjected Gradient Descent (PGD) on the continuously relaxed input prompt.\nAlthough previous attempts with ordinary gradient-based attacks largely failed,\nwe show that carefully controlling the error introduced by the continuous\nrelaxation tremendously boosts their efficacy. Our PGD for LLMs is up to one\norder of magnitude faster than state-of-the-art discrete optimization to\nachieve the same devastating attack results.\n", "date": "2024-02-14", "categories": []}
{"id": "2405.01229", "title": "Boosting Jailbreak Attack with Momentum", "abstract": "  Large Language Models (LLMs) have achieved remarkable success across diverse\ntasks, yet they remain vulnerable to adversarial attacks, notably the\nwell-known jailbreak attack. In particular, the Greedy Coordinate Gradient\n(GCG) attack has demonstrated efficacy in exploiting this vulnerability by\noptimizing adversarial prompts through a combination of gradient heuristics and\ngreedy search. However, the efficiency of this attack has become a bottleneck\nin the attacking process. To mitigate this limitation, in this paper we rethink\nthe generation of the adversarial prompts through an optimization lens, aiming\nto stabilize the optimization process and harness more heuristic insights from\nprevious optimization iterations. Specifically, we propose the\n\\textbf{M}omentum \\textbf{A}ccelerated G\\textbf{C}G (\\textbf{MAC}) attack,\nwhich integrates a momentum term into the gradient heuristic to boost and\nstabilize the random search for tokens in adversarial prompts. Experimental\nresults showcase the notable enhancement achieved by MAC over baselines in\nterms of attack success rate and optimization efficiency. Moreover, we\ndemonstrate that MAC can still exhibit superior performance for transfer\nattacks and models under defense mechanisms. Our code is available at\nhttps://github.com/weizeming/momentum-attack-llm.\n", "date": "2024-05-02", "categories": []}
{"id": "2409.07002", "title": "AdvLogo: Adversarial Patch Attack against Object Detectors based on\n  Diffusion Models", "abstract": "  With the rapid development of deep learning, object detectors have\ndemonstrated impressive performance; however, vulnerabilities still exist in\ncertain scenarios. Current research exploring the vulnerabilities using\nadversarial patches often struggles to balance the trade-off between attack\neffectiveness and visual quality. To address this problem, we propose a novel\nframework of patch attack from semantic perspective, which we refer to as\nAdvLogo. Based on the hypothesis that every semantic space contains an\nadversarial subspace where images can cause detectors to fail in recognizing\nobjects, we leverage the semantic understanding of the diffusion denoising\nprocess and drive the process to adversarial subareas by perturbing the latent\nand unconditional embeddings at the last timestep. To mitigate the distribution\nshift that exposes a negative impact on image quality, we apply perturbation to\nthe latent in frequency domain with the Fourier Transform. Experimental results\ndemonstrate that AdvLogo achieves strong attack performance while maintaining\nhigh visual quality.\n", "date": "2024-09-11", "categories": []}
{"id": "2409.19788", "title": "Exploring Adversarial Robustness in Classification tasks using DNA\n  Language Models", "abstract": "  DNA Language Models, such as GROVER, DNABERT2 and the Nucleotide Transformer,\noperate on DNA sequences that inherently contain sequencing errors, mutations,\nand laboratory-induced noise, which may significantly impact model performance.\nDespite the importance of this issue, the robustness of DNA language models\nremains largely underexplored. In this paper, we comprehensivly investigate\ntheir robustness in DNA classification by applying various adversarial attack\nstrategies: the character (nucleotide substitutions), word (codon\nmodifications), and sentence levels (back-translation-based transformations) to\nsystematically analyze model vulnerabilities. Our results demonstrate that DNA\nlanguage models are highly susceptible to adversarial attacks, leading to\nsignificant performance degradation. Furthermore, we explore adversarial\ntraining method as a defense mechanism, which enhances both robustness and\nclassification accuracy. This study highlights the limitations of DNA language\nmodels and underscores the necessity of robustness in bioinformatics.\n", "date": "2024-09-29", "categories": []}
{"id": "2410.08190", "title": "Poison-splat: Computation Cost Attack on 3D Gaussian Splatting", "abstract": "  3D Gaussian splatting (3DGS), known for its groundbreaking performance and\nefficiency, has become a dominant 3D representation and brought progress to\nmany 3D vision tasks. However, in this work, we reveal a significant security\nvulnerability that has been largely overlooked in 3DGS: the computation cost of\ntraining 3DGS could be maliciously tampered by poisoning the input data. By\ndeveloping an attack named Poison-splat, we reveal a novel attack surface where\nthe adversary can poison the input images to drastically increase the\ncomputation memory and time needed for 3DGS training, pushing the algorithm\ntowards its worst computation complexity. In extreme cases, the attack can even\nconsume all allocable memory, leading to a Denial-of-Service (DoS) that\ndisrupts servers, resulting in practical damages to real-world 3DGS service\nvendors. Such a computation cost attack is achieved by addressing a bi-level\noptimization problem through three tailored strategies: attack objective\napproximation, proxy model rendering, and optional constrained optimization.\nThese strategies not only ensure the effectiveness of our attack but also make\nit difficult to defend with simple defensive measures. We hope the revelation\nof this novel attack surface can spark attention to this crucial yet overlooked\nvulnerability of 3DGS systems. Our code is available at\nhttps://github.com/jiahaolu97/poison-splat .\n", "date": "2024-10-10", "categories": []}
{"id": "2410.10322", "title": "Feature Averaging: An Implicit Bias of Gradient Descent Leading to\n  Non-Robustness in Neural Networks", "abstract": "  In this work, we investigate a particular implicit bias in gradient descent\ntraining, which we term \"Feature Averaging,\" and argue that it is one of the\nprincipal factors contributing to the non-robustness of deep neural networks.\nWe show that, even when multiple discriminative features are present in the\ninput data, neural networks trained by gradient descent tend to rely on an\naverage (or a certain combination) of these features for classification, rather\nthan distinguishing and leveraging each feature individually. Specifically, we\nprovide a detailed theoretical analysis of the training dynamics of two-layer\nReLU networks on a binary classification task, where the data distribution\nconsists of multiple clusters with mutually orthogonal centers. We rigorously\nprove that gradient descent biases the network towards feature averaging, where\nthe weights of each hidden neuron represent an average of the cluster centers\n(each corresponding to a distinct feature), thereby making the network\nvulnerable to input perturbations aligned with the negative direction of the\naveraged features. On the positive side, we demonstrate that this vulnerability\ncan be mitigated through more granular supervision. In particular, we prove\nthat a two-layer ReLU network can achieve optimal robustness when trained to\nclassify individual features rather than merely the original binary classes.\nFinally, we validate our theoretical findings with experiments on synthetic\ndatasets, MNIST, and CIFAR-10, and confirm the prevalence of feature averaging\nand its impact on adversarial robustness. We hope these theoretical and\nempirical insights deepen the understanding of how gradient descent shapes\nfeature learning and adversarial robustness, and how more detailed supervision\ncan enhance robustness.\n", "date": "2024-10-14", "categories": []}
{"id": "2502.08079", "title": "MAA: Meticulous Adversarial Attack against Vision-Language Pre-trained\n  Models", "abstract": "  Current adversarial attacks for evaluating the robustness of vision-language\npre-trained (VLP) models in multi-modal tasks suffer from limited\ntransferability, where attacks crafted for a specific model often struggle to\ngeneralize effectively across different models, limiting their utility in\nassessing robustness more broadly. This is mainly attributed to the\nover-reliance on model-specific features and regions, particularly in the image\nmodality. In this paper, we propose an elegant yet highly effective method\ntermed Meticulous Adversarial Attack (MAA) to fully exploit model-independent\ncharacteristics and vulnerabilities of individual samples, achieving enhanced\ngeneralizability and reduced model dependence. MAA emphasizes fine-grained\noptimization of adversarial images by developing a novel resizing and sliding\ncrop (RScrop) technique, incorporating a multi-granularity similarity\ndisruption (MGSD) strategy. Extensive experiments across diverse VLP models,\nmultiple benchmark datasets, and a variety of downstream tasks demonstrate that\nMAA significantly enhances the effectiveness and transferability of adversarial\nattacks. A large cohort of performance studies is conducted to generate\ninsights into the effectiveness of various model configurations, guiding future\nadvancements in this domain.\n", "date": "2025-02-12", "categories": []}
{"id": "2502.11858", "title": "Rethinking Audio-Visual Adversarial Vulnerability from Temporal and\n  Modality Perspectives", "abstract": "  While audio-visual learning equips models with a richer understanding of the\nreal world by leveraging multiple sensory modalities, this integration also\nintroduces new vulnerabilities to adversarial attacks.\n  In this paper, we present a comprehensive study of the adversarial robustness\nof audio-visual models, considering both temporal and modality-specific\nvulnerabilities. We propose two powerful adversarial attacks: 1) a temporal\ninvariance attack that exploits the inherent temporal redundancy across\nconsecutive time segments and 2) a modality misalignment attack that introduces\nincongruence between the audio and visual modalities. These attacks are\ndesigned to thoroughly assess the robustness of audio-visual models against\ndiverse threats. Furthermore, to defend against such attacks, we introduce a\nnovel audio-visual adversarial training framework. This framework addresses key\nchallenges in vanilla adversarial training by incorporating efficient\nadversarial perturbation crafting tailored to multi-modal data and an\nadversarial curriculum strategy. Extensive experiments in the Kinetics-Sounds\ndataset demonstrate that our proposed temporal and modality-based attacks in\ndegrading model performance can achieve state-of-the-art performance, while our\nadversarial training defense largely improves the adversarial robustness as\nwell as the adversarial training efficiency.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.18176", "title": "CLIPure: Purification in Latent Space via CLIP for Adversarially Robust\n  Zero-Shot Classification", "abstract": "  In this paper, we aim to build an adversarially robust zero-shot image\nclassifier. We ground our work on CLIP, a vision-language pre-trained encoder\nmodel that can perform zero-shot classification by matching an image with text\nprompts ``a photo of a <class-name>.''. Purification is the path we choose\nsince it does not require adversarial training on specific attack types and\nthus can cope with any foreseen attacks. We then formulate purification risk as\nthe KL divergence between the joint distributions of the purification process\nof denoising the adversarial samples and the attack process of adding\nperturbations to benign samples, through bidirectional Stochastic Differential\nEquations (SDEs). The final derived results inspire us to explore purification\nin the multi-modal latent space of CLIP. We propose two variants for our\nCLIPure approach: CLIPure-Diff which models the likelihood of images' latent\nvectors with the DiffusionPrior module in DaLLE-2 (modeling the generation\nprocess of CLIP's latent vectors), and CLIPure-Cos which models the likelihood\nwith the cosine similarity between the embeddings of an image and ``a photo of\na.''. As far as we know, CLIPure is the first purification method in\nmulti-modal latent space and CLIPure-Cos is the first purification method that\nis not based on generative models, which substantially improves defense\nefficiency. We conducted extensive experiments on CIFAR-10, ImageNet, and 13\ndatasets that previous CLIP-based defense methods used for evaluating zero-shot\nclassification robustness. Results show that CLIPure boosts the SOTA robustness\nby a large margin, e.g., from 71.7% to 91.1% on CIFAR10, from 59.6% to 72.6% on\nImageNet, and 108% relative improvements of average robustness on the 13\ndatasets over previous SOTA. The code is available at\nhttps://github.com/TMLResearchGroup-CAS/CLIPure.\n", "date": "2025-02-25", "categories": []}
{"id": "2502.19047", "title": "A Dual-Purpose Framework for Backdoor Defense and Backdoor Amplification\n  in Diffusion Models", "abstract": "  Diffusion models have emerged as state-of-the-art generative frameworks,\nexcelling in producing high-quality multi-modal samples. However, recent\nstudies have revealed their vulnerability to backdoor attacks, where backdoored\nmodels generate specific, undesirable outputs called backdoor target (e.g.,\nharmful images) when a pre-defined trigger is embedded to their inputs. In this\npaper, we propose PureDiffusion, a dual-purpose framework that simultaneously\nserves two contrasting roles: backdoor defense and backdoor attack\namplification. For defense, we introduce two novel loss functions to invert\nbackdoor triggers embedded in diffusion models. The first leverages\ntrigger-induced distribution shifts across multiple timesteps of the diffusion\nprocess, while the second exploits the denoising consistency effect when a\nbackdoor is activated. Once an accurate trigger inversion is achieved, we\ndevelop a backdoor detection method that analyzes both the inverted trigger and\nthe generated backdoor targets to identify backdoor attacks. In terms of attack\namplification with the role of an attacker, we describe how our trigger\ninversion algorithm can be used to reinforce the original trigger embedded in\nthe backdoored diffusion model. This significantly boosts attack performance\nwhile reducing the required backdoor training time. Experimental results\ndemonstrate that PureDiffusion achieves near-perfect detection accuracy,\noutperforming existing defenses by a large margin, particularly against complex\ntrigger patterns. Additionally, in an attack scenario, our attack amplification\napproach elevates the attack success rate (ASR) of existing backdoor attacks to\nnearly 100\\% while reducing training time by up to 20x.\n", "date": "2025-02-26", "categories": []}
{"id": "2503.00377", "title": "Adversarial Attacks on Event-Based Pedestrian Detectors: A Physical\n  Approach", "abstract": "  Event cameras, known for their low latency and high dynamic range, show great\npotential in pedestrian detection applications. However, while recent research\nhas primarily focused on improving detection accuracy, the robustness of\nevent-based visual models against physical adversarial attacks has received\nlimited attention. For example, adversarial physical objects, such as specific\nclothing patterns or accessories, can exploit inherent vulnerabilities in these\nsystems, leading to misdetections or misclassifications. This study is the\nfirst to explore physical adversarial attacks on event-driven pedestrian\ndetectors, specifically investigating whether certain clothing patterns worn by\npedestrians can cause these detectors to fail, effectively rendering them\nunable to detect the person. To address this, we developed an end-to-end\nadversarial framework in the digital domain, framing the design of adversarial\nclothing textures as a 2D texture optimization problem. By crafting an\neffective adversarial loss function, the framework iteratively generates\noptimal textures through backpropagation. Our results demonstrate that the\ntextures identified in the digital domain possess strong adversarial\nproperties. Furthermore, we translated these digitally optimized textures into\nphysical clothing and tested them in real-world scenarios, successfully\ndemonstrating that the designed textures significantly degrade the performance\nof event-based pedestrian detection models. This work highlights the\nvulnerability of such models to physical adversarial attacks.\n", "date": "2025-03-01", "categories": []}
{"id": "2503.00384", "title": "A Survey of Adversarial Defenses in Vision-based Systems:\n  Categorization, Methods and Challenges", "abstract": "  Adversarial attacks have emerged as a major challenge to the trustworthy\ndeployment of machine learning models, particularly in computer vision\napplications. These attacks have a varied level of potency and can be\nimplemented in both white box and black box approaches. Practical attacks\ninclude methods to manipulate the physical world and enforce adversarial\nbehaviour by the corresponding target neural network models. Multiple different\napproaches to mitigate different kinds of such attacks are available in the\nliterature, each with their own advantages and limitations. In this survey, we\npresent a comprehensive systematization of knowledge on adversarial defenses,\nfocusing on two key computer vision tasks: image classification and object\ndetection. We review the state-of-the-art adversarial defense techniques and\ncategorize them for easier comparison. In addition, we provide a schematic\nrepresentation of these categories within the context of the overall machine\nlearning pipeline, facilitating clearer understanding and benchmarking of\ndefenses. Furthermore, we map these defenses to the types of adversarial\nattacks and datasets where they are most effective, offering practical insights\nfor researchers and practitioners. This study is necessary for understanding\nthe scope of how the available defenses are able to address the adversarial\nthreats, and their shortcomings as well, which is necessary for driving the\nresearch in this area in the most appropriate direction, with the aim of\nbuilding trustworthy AI systems for regular practical use-cases.\n", "date": "2025-03-01", "categories": []}
{"id": "2503.00932", "title": "Improving the Transferability of Adversarial Attacks by an Input\n  Transpose", "abstract": "  Deep neural networks (DNNs) are highly susceptible to adversarial\nexamples--subtle perturbations applied to inputs that are often imperceptible\nto humans yet lead to incorrect model predictions. In black-box scenarios,\nhowever, existing adversarial examples exhibit limited transferability and\nstruggle to effectively compromise multiple unseen DNN models. Previous\nstrategies enhance the cross-model generalization of adversarial examples by\nintroducing versatility into adversarial perturbations, thereby improving\ntransferability. However, further refining perturbation versatility often\ndemands intricate algorithm development and substantial computation\nconsumption. In this work, we propose an input transpose method that requires\nalmost no additional labor and computation costs but can significantly improve\nthe transferability of existing adversarial strategies. Even without adding\nadversarial perturbations, our method demonstrates considerable effectiveness\nin cross-model attacks. Our exploration finds that on specific datasets, a mere\n$1^\\circ$ left or right rotation might be sufficient for most adversarial\nexamples to deceive unseen models. Our further analysis suggests that this\ntransferability improvement triggered by rotating only $1^\\circ$ may stem from\nvisible pattern shifts in the DNN's low-level feature maps. Moreover, this\ntransferability exhibits optimal angles that, when identified under\nunrestricted query conditions, could potentially yield even greater\nperformance.\n", "date": "2025-03-02", "categories": []}
{"id": "2503.01811", "title": "AutoAdvExBench: Benchmarking autonomous exploitation of adversarial\n  example defenses", "abstract": "  We introduce AutoAdvExBench, a benchmark to evaluate if large language models\n(LLMs) can autonomously exploit defenses to adversarial examples. Unlike\nexisting security benchmarks that often serve as proxies for real-world tasks,\nbench directly measures LLMs' success on tasks regularly performed by machine\nlearning security experts. This approach offers a significant advantage: if a\nLLM could solve the challenges presented in bench, it would immediately present\npractical utility for adversarial machine learning researchers. We then design\na strong agent that is capable of breaking 75% of CTF-like (\"homework\nexercise\") adversarial example defenses. However, we show that this agent is\nonly able to succeed on 13% of the real-world defenses in our benchmark,\nindicating the large gap between difficulty in attacking \"real\" code, and\nCTF-like code. In contrast, a stronger LLM that can attack 21% of real defenses\nonly succeeds on 54% of CTF-like defenses. We make this benchmark available at\nhttps://github.com/ethz-spylab/AutoAdvExBench.\n", "date": "2025-03-03", "categories": []}
{"id": "2412.11441", "title": "UIBDiffusion: Universal Imperceptible Backdoor Attack for Diffusion\n  Models", "abstract": "  Recent studies show that diffusion models (DMs) are vulnerable to backdoor\nattacks. Existing backdoor attacks impose unconcealed triggers (e.g., a gray\nbox and eyeglasses) that contain evident patterns, rendering remarkable attack\neffects yet easy detection upon human inspection and defensive algorithms.\nWhile it is possible to improve stealthiness by reducing the strength of the\nbackdoor, doing so can significantly compromise its generality and\neffectiveness. In this paper, we propose UIBDiffusion, the universal\nimperceptible backdoor attack for diffusion models, which allows us to achieve\nsuperior attack and generation performance while evading state-of-the-art\ndefenses. We propose a novel trigger generation approach based on universal\nadversarial perturbations (UAPs) and reveal that such perturbations, which are\ninitially devised for fooling pre-trained discriminative models, can be adapted\nas potent imperceptible backdoor triggers for DMs. We evaluate UIBDiffusion on\nmultiple types of DMs with different kinds of samplers across various datasets\nand targets. Experimental results demonstrate that UIBDiffusion brings three\nadvantages: 1) Universality, the imperceptible trigger is universal (i.e.,\nimage and model agnostic) where a single trigger is effective to any images and\nall diffusion models with different samplers; 2) Utility, it achieves\ncomparable generation quality (e.g., FID) and even better attack success rate\n(i.e., ASR) at low poison rates compared to the prior works; and 3)\nUndetectability, UIBDiffusion is plausible to human perception and can bypass\nElijah and TERD, the SOTA defenses against backdoors for DMs. We will release\nour backdoor triggers and code.\n", "date": "2024-12-16", "categories": []}
{"id": "2502.20562", "title": "LISArD: Learning Image Similarity to Defend Against Gray-box Adversarial\n  Attacks", "abstract": "  State-of-the-art defense mechanisms are typically evaluated in the context of\nwhite-box attacks, which is not realistic, as it assumes the attacker can\naccess the gradients of the target network. To protect against this scenario,\nAdversarial Training (AT) and Adversarial Distillation (AD) include adversarial\nexamples during the training phase, and Adversarial Purification uses a\ngenerative model to reconstruct all the images given to the classifier. This\npaper considers an even more realistic evaluation scenario: gray-box attacks,\nwhich assume that the attacker knows the architecture and the dataset used to\ntrain the target network, but cannot access its gradients. We provide empirical\nevidence that models are vulnerable to gray-box attacks and propose LISArD, a\ndefense mechanism that does not increase computational and temporal costs but\nprovides robustness against gray-box and white-box attacks without including\nAT. Our method approximates a cross-correlation matrix, created with the\nembeddings of perturbed and clean images, to a diagonal matrix while\nsimultaneously conducting classification learning. Our results show that LISArD\ncan effectively protect against gray-box attacks, can be used in multiple\narchitectures, and carries over its resilience to the white-box scenario. Also,\nstate-of-the-art AD models underperform greatly when removing AT and/or moving\nto gray-box settings, highlighting the lack of robustness from existing\napproaches to perform in various conditions (aside from white-box settings).\nAll the source code is available at https://github.com/Joana-Cabral/LISArD.\n", "date": "2025-02-27", "categories": []}
{"id": "2502.20943", "title": "BadRefSR: Backdoor Attacks Against Reference-based Image Super\n  Resolution", "abstract": "  Reference-based image super-resolution (RefSR) represents a promising\nadvancement in super-resolution (SR). In contrast to single-image\nsuper-resolution (SISR), RefSR leverages an additional reference image to help\nrecover high-frequency details, yet its vulnerability to backdoor attacks has\nnot been explored. To fill this research gap, we propose a novel attack\nframework called BadRefSR, which embeds backdoors in the RefSR model by adding\ntriggers to the reference images and training with a mixed loss function.\nExtensive experiments across various backdoor attack settings demonstrate the\neffectiveness of BadRefSR. The compromised RefSR network performs normally on\nclean input images, while outputting attacker-specified target images on\ntriggered input images. Our study aims to alert researchers to the potential\nbackdoor risks in RefSR. Codes are available at\nhttps://github.com/xuefusiji/BadRefSR.\n", "date": "2025-02-28", "categories": []}
{"id": "2502.20948", "title": "Concealed Adversarial attacks on neural networks for sequential data", "abstract": "  The emergence of deep learning led to the broad usage of neural networks in\nthe time series domain for various applications, including finance and\nmedicine. While powerful, these models are prone to adversarial attacks: a\nbenign targeted perturbation of input data leads to significant changes in a\nclassifier's output. However, formally small attacks in the time series domain\nbecome easily detected by the human eye or a simple detector model.\n  We develop a concealed adversarial attack for different time-series models:\nit provides more realistic perturbations, being hard to detect by a human or\nmodel discriminator. To achieve this goal, the proposed adversarial attack\nmaximizes an aggregation of a classifier and a trained discriminator loss. To\nmake the attack stronger, we also propose a training procedure for a\ndiscriminator that provides broader coverage of possible attacks. Extensive\nbenchmarking on six UCR time series datasets across four diverse architectures\n- including recurrent, convolutional, state-space, and transformer-based models\n- demonstrates the superiority of our attack for a concealability-efficiency\ntrade-off. Our findings highlight the growing challenge of designing robust\ntime series models, emphasizing the need for improved defenses against\nrealistic and effective attacks.\n", "date": "2025-02-28", "categories": []}
{"id": "2407.16233", "title": "Algebraic Adversarial Attacks on Integrated Gradients", "abstract": "  Adversarial attacks on explainability models have drastic consequences when\nexplanations are used to understand the reasoning of neural networks in safety\ncritical systems. Path methods are one such class of attribution methods\nsusceptible to adversarial attacks. Adversarial learning is typically phrased\nas a constrained optimisation problem. In this work, we propose algebraic\nadversarial examples and study the conditions under which one can generate\nadversarial examples for integrated gradients. Algebraic adversarial examples\nprovide a mathematically tractable approach to adversarial examples.\n", "date": "2024-07-23", "categories": []}
{"id": "2502.12893", "title": "H-CoT: Hijacking the Chain-of-Thought Safety Reasoning Mechanism to\n  Jailbreak Large Reasoning Models, Including OpenAI o1/o3, DeepSeek-R1, and\n  Gemini 2.0 Flash Thinking", "abstract": "  Large Reasoning Models (LRMs) have recently extended their powerful reasoning\ncapabilities to safety checks-using chain-of-thought reasoning to decide\nwhether a request should be answered. While this new approach offers a\npromising route for balancing model utility and safety, its robustness remains\nunderexplored. To address this gap, we introduce Malicious-Educator, a\nbenchmark that disguises extremely dangerous or malicious requests beneath\nseemingly legitimate educational prompts. Our experiments reveal severe\nsecurity flaws in popular commercial-grade LRMs, including OpenAI o1/o3,\nDeepSeek-R1, and Gemini 2.0 Flash Thinking. For instance, although OpenAI's o1\nmodel initially maintains a high refusal rate of about 98%, subsequent model\nupdates significantly compromise its safety; and attackers can easily extract\ncriminal strategies from DeepSeek-R1 and Gemini 2.0 Flash Thinking without any\nadditional tricks. To further highlight these vulnerabilities, we propose\nHijacking Chain-of-Thought (H-CoT), a universal and transferable attack method\nthat leverages the model's own displayed intermediate reasoning to jailbreak\nits safety reasoning mechanism. Under H-CoT, refusal rates sharply\ndecline-dropping from 98% to below 2%-and, in some instances, even transform\ninitially cautious tones into ones that are willing to provide harmful content.\nWe hope these findings underscore the urgent need for more robust safety\nmechanisms to preserve the benefits of advanced reasoning capabilities without\ncompromising ethical standards.\n", "date": "2025-02-18", "categories": []}
{"id": "2502.19672", "title": "Improving Adversarial Transferability in MLLMs via Dynamic\n  Vision-Language Alignment Attack", "abstract": "  Multimodal Large Language Models (MLLMs), built upon LLMs, have recently\ngained attention for their capabilities in image recognition and understanding.\nHowever, while MLLMs are vulnerable to adversarial attacks, the transferability\nof these attacks across different models remains limited, especially under\ntargeted attack setting. Existing methods primarily focus on vision-specific\nperturbations but struggle with the complex nature of vision-language modality\nalignment. In this work, we introduce the Dynamic Vision-Language Alignment\n(DynVLA) Attack, a novel approach that injects dynamic perturbations into the\nvision-language connector to enhance generalization across diverse\nvision-language alignment of different models. Our experimental results show\nthat DynVLA significantly improves the transferability of adversarial examples\nacross various MLLMs, including BLIP2, InstructBLIP, MiniGPT4, LLaVA, and\nclosed-source models such as Gemini.\n", "date": "2025-02-27", "categories": []}
{"id": "2502.19710", "title": "SAP-DIFF: Semantic Adversarial Patch Generation for Black-Box Face\n  Recognition Models via Diffusion Models", "abstract": "  Given the need to evaluate the robustness of face recognition (FR) models,\nmany efforts have focused on adversarial patch attacks that mislead FR models\nby introducing localized perturbations. Impersonation attacks are a significant\nthreat because adversarial perturbations allow attackers to disguise themselves\nas legitimate users. This can lead to severe consequences, including data\nbreaches, system damage, and misuse of resources. However, research on such\nattacks in FR remains limited. Existing adversarial patch generation methods\nexhibit limited efficacy in impersonation attacks due to (1) the need for high\nattacker capabilities, (2) low attack success rates, and (3) excessive query\nrequirements. To address these challenges, we propose a novel method SAP-DIFF\nthat leverages diffusion models to generate adversarial patches via semantic\nperturbations in the latent space rather than direct pixel manipulation. We\nintroduce an attention disruption mechanism to generate features unrelated to\nthe original face, facilitating the creation of adversarial samples and a\ndirectional loss function to guide perturbations toward the target identity\nfeature space, thereby enhancing attack effectiveness and efficiency. Extensive\nexperiments on popular FR models and datasets demonstrate that our method\noutperforms state-of-the-art approaches, achieving an average attack success\nrate improvement of 45.66% (all exceeding 40%), and a reduction in the number\nof queries by about 40% compared to the SOTA approach\n", "date": "2025-02-27", "categories": []}
{"id": "2502.19757", "title": "Snowball Adversarial Attack on Traffic Sign Classification", "abstract": "  Adversarial attacks on machine learning models often rely on small,\nimperceptible perturbations to mislead classifiers. Such strategy focuses on\nminimizing the visual perturbation for humans so they are not confused, and\nalso maximizing the misclassification for machine learning algorithms. An\northogonal strategy for adversarial attacks is to create perturbations that are\nclearly visible but do not confuse humans, yet still maximize misclassification\nfor machine learning algorithms. This work follows the later strategy, and\ndemonstrates instance of it through the Snowball Adversarial Attack in the\ncontext of traffic sign recognition. The attack leverages the human brain's\nsuperior ability to recognize objects despite various occlusions, while machine\nlearning algorithms are easily confused. The evaluation shows that the Snowball\nAdversarial Attack is robust across various images and is able to confuse\nstate-of-the-art traffic sign recognition algorithm. The findings reveal that\nSnowball Adversarial Attack can significantly degrade model performance with\nminimal effort, raising important concerns about the vulnerabilities of deep\nneural networks and highlighting the necessity for improved defenses for image\nrecognition machine learning models.\n", "date": "2025-02-27", "categories": []}
{"id": "2502.20306", "title": "SecureGaze: Defending Gaze Estimation Against Backdoor Attacks", "abstract": "  Gaze estimation models are widely used in applications such as driver\nattention monitoring and human-computer interaction. While many methods for\ngaze estimation exist, they rely heavily on data-hungry deep learning to\nachieve high performance. This reliance often forces practitioners to harvest\ntraining data from unverified public datasets, outsource model training, or\nrely on pre-trained models. However, such practices expose gaze estimation\nmodels to backdoor attacks. In such attacks, adversaries inject backdoor\ntriggers by poisoning the training data, creating a backdoor vulnerability: the\nmodel performs normally with benign inputs, but produces manipulated gaze\ndirections when a specific trigger is present. This compromises the security of\nmany gaze-based applications, such as causing the model to fail in tracking the\ndriver's attention. To date, there is no defense that addresses backdoor\nattacks on gaze estimation models. In response, we introduce SecureGaze, the\nfirst solution designed to protect gaze estimation models from such attacks.\nUnlike classification models, defending gaze estimation poses unique challenges\ndue to its continuous output space and globally activated backdoor behavior. By\nidentifying distinctive characteristics of backdoored gaze estimation models,\nwe develop a novel and effective approach to reverse-engineer the trigger\nfunction for reliable backdoor detection. Extensive evaluations in both digital\nand physical worlds demonstrate that SecureGaze effectively counters a range of\nbackdoor attacks and outperforms seven state-of-the-art defenses adapted from\nclassification models.\n", "date": "2025-02-27", "categories": []}
{"id": "2410.08338", "title": "Time Traveling to Defend Against Adversarial Example Attacks in Image\n  Classification", "abstract": "  Adversarial example attacks have emerged as a critical threat to machine\nlearning. Adversarial attacks in image classification abuse various, minor\nmodifications to the image that confuse the image classification neural network\n-- while the image still remains recognizable to humans. One important domain\nwhere the attacks have been applied is in the automotive setting with traffic\nsign classification. Researchers have demonstrated that adding stickers,\nshining light, or adding shadows are all different means to make machine\nlearning inference algorithms mis-classify the traffic signs. This can cause\npotentially dangerous situations as a stop sign is recognized as a speed limit\nsign causing vehicles to ignore it and potentially leading to accidents. To\naddress these attacks, this work focuses on enhancing defenses against such\nadversarial attacks. This work shifts the advantage to the user by introducing\nthe idea of leveraging historical images and majority voting. While the\nattacker modifies a traffic sign that is currently being processed by the\nvictim's machine learning inference, the victim can gain advantage by examining\npast images of the same traffic sign. This work introduces the notion of ''time\ntraveling'' and uses historical Street View images accessible to anybody to\nperform inference on different, past versions of the same traffic sign. In the\nevaluation, the proposed defense has 100% effectiveness against latest\nadversarial example attack on traffic sign classification algorithm.\n", "date": "2024-10-10", "categories": []}
{"id": "2411.18776", "title": "Fall Leaf Adversarial Attack on Traffic Sign Classification", "abstract": "  Adversarial input image perturbation attacks have emerged as a significant\nthreat to machine learning algorithms, particularly in image classification\nsetting. These attacks involve subtle perturbations to input images that cause\nneural networks to misclassify the input images, even though the images remain\neasily recognizable to humans. One critical area where adversarial attacks have\nbeen demonstrated is in automotive systems where traffic sign classification\nand recognition is critical, and where misclassified images can cause\nautonomous systems to take wrong actions. This work presents a new class of\nadversarial attacks. Unlike existing work that has focused on adversarial\nperturbations that leverage human-made artifacts to cause the perturbations,\nsuch as adding stickers, paint, or shining flashlights at traffic signs, this\nwork leverages nature-made artifacts: tree leaves. By leveraging nature-made\nartifacts, the new class of attacks has plausible deniability: a fall leaf\nstuck to a street sign could come from a near-by tree, rather than be placed\nthere by an malicious human attacker. To evaluate the new class of the\nadversarial input image perturbation attacks, this work analyses how fall\nleaves can cause misclassification in street signs. The work evaluates various\nleaves from different species of trees, and considers various parameters such\nas size, color due to tree leaf type, and rotation. The work demonstrates high\nsuccess rate for misclassification. The work also explores the correlation\nbetween successful attacks and how they affect the edge detection, which is\ncritical in many image classification algorithms.\n", "date": "2024-11-27", "categories": []}
{"id": "2502.18592", "title": "DeBUGCN -- Detecting Backdoors in CNNs Using Graph Convolutional\n  Networks", "abstract": "  Deep neural networks (DNNs) are becoming commonplace in critical\napplications, making their susceptibility to backdoor (trojan) attacks a\nsignificant problem. In this paper, we introduce a novel backdoor attack\ndetection pipeline, detecting attacked models using graph convolution networks\n(DeBUGCN). To the best of our knowledge, ours is the first use of GCNs for\ntrojan detection. We use the static weights of a DNN to create a graph\nstructure of its layers. A GCN is then used as a binary classifier on these\ngraphs, yielding a trojan or clean determination for the DNN. To demonstrate\nthe efficacy of our pipeline, we train hundreds of clean and trojaned CNN\nmodels on the MNIST handwritten digits and CIFAR-10 image datasets, and show\nthe DNN classification results using DeBUGCN. For a true In-the-Wild use case,\nour pipeline is evaluated on the TrojAI dataset which consists of various CNN\narchitectures, thus showing the robustness and model-agnostic behaviour of\nDeBUGCN. Furthermore, on comparing our results on several datasets with\nstate-of-the-art trojan detection algorithms, DeBUGCN is faster and more\naccurate.\n", "date": "2025-02-25", "categories": []}
{"id": "2502.18724", "title": "Adversarial Universal Stickers: Universal Perturbation Attacks on\n  Traffic Sign using Stickers", "abstract": "  Adversarial attacks on deep learning models have proliferated in recent\nyears. In many cases, a different adversarial perturbation is required to be\nadded to each image to cause the deep learning model to misclassify it. This is\nineffective as each image has to be modified in a different way. Meanwhile,\nresearch on universal perturbations focuses on designing a single perturbation\nthat can be applied to all images in a data set, and cause a deep learning\nmodel to misclassify the images. This work advances the field of universal\nperturbations by exploring universal perturbations in the context of traffic\nsigns and autonomous vehicle systems. This work introduces a novel method for\ngenerating universal perturbations that visually look like simple black and\nwhite stickers, and using them to cause incorrect street sign predictions.\nUnlike traditional adversarial perturbations, the adversarial universal\nstickers are designed to be applicable to any street sign: same sticker, or\nstickers, can be applied in same location to any street sign and cause it to be\nmisclassified. Further, to enable safe experimentation with adversarial images\nand street signs, this work presents a virtual setting that leverages Street\nView images of street signs, rather than the need to physically modify street\nsigns, to test the attacks. The experiments in the virtual setting demonstrate\nthat these stickers can consistently mislead deep learning models used commonly\nin street sign recognition, and achieve high attack success rates on dataset of\nUS traffic signs. The findings highlight the practical security risks posed by\nsimple stickers applied to traffic signs, and the ease with which adversaries\ncan generate adversarial universal stickers that can be applied to many street\nsigns.\n", "date": "2025-02-26", "categories": []}
{"id": "2405.13922", "title": "Towards Certification of Uncertainty Calibration under Adversarial\n  Attacks", "abstract": "  Since neural classifiers are known to be sensitive to adversarial\nperturbations that alter their accuracy, \\textit{certification methods} have\nbeen developed to provide provable guarantees on the insensitivity of their\npredictions to such perturbations. Furthermore, in safety-critical\napplications, the frequentist interpretation of the confidence of a classifier\n(also known as model calibration) can be of utmost importance. This property\ncan be measured via the Brier score or the expected calibration error. We show\nthat attacks can significantly harm calibration, and thus propose certified\ncalibration as worst-case bounds on calibration under adversarial\nperturbations. Specifically, we produce analytic bounds for the Brier score and\napproximate bounds via the solution of a mixed-integer program on the expected\ncalibration error. Finally, we propose novel calibration attacks and\ndemonstrate how they can improve model calibration through \\textit{adversarial\ncalibration training}.\n", "date": "2024-05-22", "categories": []}
{"id": "2409.13232", "title": "Relationship between Uncertainty in DNNs and Adversarial Attacks", "abstract": "  Deep Neural Networks (DNNs) have achieved state of the art results and even\noutperformed human accuracy in many challenging tasks, leading to DNNs adoption\nin a variety of fields including natural language processing, pattern\nrecognition, prediction, and control optimization. However, DNNs are\naccompanied by uncertainty about their results, causing them to predict an\noutcome that is either incorrect or outside of a certain level of confidence.\nThese uncertainties stem from model or data constraints, which could be\nexacerbated by adversarial attacks. Adversarial attacks aim to provide\nperturbed input to DNNs, causing the DNN to make incorrect predictions or\nincrease model uncertainty. In this review, we explore the relationship between\nDNN uncertainty and adversarial attacks, emphasizing how adversarial attacks\nmight raise DNN uncertainty.\n", "date": "2024-09-20", "categories": []}
{"id": "2410.23091", "title": "CausalDiff: Causality-Inspired Disentanglement via Diffusion Model for\n  Adversarial Defense", "abstract": "  Despite ongoing efforts to defend neural classifiers from adversarial\nattacks, they remain vulnerable, especially to unseen attacks. In contrast,\nhumans are difficult to be cheated by subtle manipulations, since we make\njudgments only based on essential factors. Inspired by this observation, we\nattempt to model label generation with essential label-causative factors and\nincorporate label-non-causative factors to assist data generation. For an\nadversarial example, we aim to discriminate the perturbations as non-causative\nfactors and make predictions only based on the label-causative factors.\nConcretely, we propose a casual diffusion model (CausalDiff) that adapts\ndiffusion models for conditional data generation and disentangles the two types\nof casual factors by learning towards a novel casual information bottleneck\nobjective. Empirically, CausalDiff has significantly outperformed\nstate-of-the-art defense methods on various unseen attacks, achieving an\naverage robustness of 86.39% (+4.01%) on CIFAR-10, 56.25% (+3.13%) on\nCIFAR-100, and 82.62% (+4.93%) on GTSRB (German Traffic Sign Recognition\nBenchmark). The code is available at\nhttps://github.com/CAS-AISafetyBasicResearchGroup/CausalDiff.\n", "date": "2024-10-30", "categories": []}
{"id": "2502.16361", "title": "A Framework for Evaluating Vision-Language Model Safety: Building Trust\n  in AI for Public Sector Applications", "abstract": "  Vision-Language Models (VLMs) are increasingly deployed in public sector\nmissions, necessitating robust evaluation of their safety and vulnerability to\nadversarial attacks. This paper introduces a novel framework to quantify\nadversarial risks in VLMs. We analyze model performance under Gaussian,\nsalt-and-pepper, and uniform noise, identifying misclassification thresholds\nand deriving composite noise patches and saliency patterns that highlight\nvulnerable regions. These patterns are compared against the Fast Gradient Sign\nMethod (FGSM) to assess their adversarial effectiveness. We propose a new\nVulnerability Score that combines the impact of random noise and adversarial\nattacks, providing a comprehensive metric for evaluating model robustness.\n", "date": "2025-02-22", "categories": []}
{"id": "2502.17972", "title": "Model-Free Adversarial Purification via Coarse-To-Fine Tensor Network\n  Representation", "abstract": "  Deep neural networks are known to be vulnerable to well-designed adversarial\nattacks. Although numerous defense strategies have been proposed, many are\ntailored to the specific attacks or tasks and often fail to generalize across\ndiverse scenarios. In this paper, we propose Tensor Network Purification (TNP),\na novel model-free adversarial purification method by a specially designed\ntensor network decomposition algorithm. TNP depends neither on the pre-trained\ngenerative model nor the specific dataset, resulting in strong robustness\nacross diverse adversarial scenarios. To this end, the key challenge lies in\nrelaxing Gaussian-noise assumptions of classical decompositions and\naccommodating the unknown distribution of adversarial perturbations. Unlike the\nlow-rank representation of classical decompositions, TNP aims to reconstruct\nthe unobserved clean examples from an adversarial example. Specifically, TNP\nleverages progressive downsampling and introduces a novel adversarial\noptimization objective to address the challenge of minimizing reconstruction\nerror but without inadvertently restoring adversarial perturbations. Extensive\nexperiments conducted on CIFAR-10, CIFAR-100, and ImageNet demonstrate that our\nmethod generalizes effectively across various norm threats, attack types, and\ntasks, providing a versatile and promising adversarial purification technique.\n", "date": "2025-02-25", "categories": []}
{"id": "2402.02316", "title": "Your Diffusion Model is Secretly a Certifiably Robust Classifier", "abstract": "  Generative learning, recognized for its effective modeling of data\ndistributions, offers inherent advantages in handling out-of-distribution\ninstances, especially for enhancing robustness to adversarial attacks. Among\nthese, diffusion classifiers, utilizing powerful diffusion models, have\ndemonstrated superior empirical robustness. However, a comprehensive\ntheoretical understanding of their robustness is still lacking, raising\nconcerns about their vulnerability to stronger future attacks. In this study,\nwe prove that diffusion classifiers possess $O(1)$ Lipschitzness, and establish\ntheir certified robustness, demonstrating their inherent resilience. To achieve\nnon-constant Lipschitzness, thereby obtaining much tighter certified\nrobustness, we generalize diffusion classifiers to classify Gaussian-corrupted\ndata. This involves deriving the evidence lower bounds (ELBOs) for these\ndistributions, approximating the likelihood using the ELBO, and calculating\nclassification probabilities via Bayes' theorem. Experimental results show the\nsuperior certified robustness of these Noised Diffusion Classifiers (NDCs).\nNotably, we achieve over 80% and 70% certified robustness on CIFAR-10 under\nadversarial perturbations with \\(\\ell_2\\) norms less than 0.25 and 0.5,\nrespectively, using a single off-the-shelf diffusion model without any\nadditional data.\n", "date": "2024-02-04", "categories": []}
{"id": "2402.13148", "title": "Defending Jailbreak Prompts via In-Context Adversarial Game", "abstract": "  Large Language Models (LLMs) demonstrate remarkable capabilities across\ndiverse applications. However, concerns regarding their security, particularly\nthe vulnerability to jailbreak attacks, persist. Drawing inspiration from\nadversarial training in deep learning and LLM agent learning processes, we\nintroduce the In-Context Adversarial Game (ICAG) for defending against\njailbreaks without the need for fine-tuning. ICAG leverages agent learning to\nconduct an adversarial game, aiming to dynamically extend knowledge to defend\nagainst jailbreaks. Unlike traditional methods that rely on static datasets,\nICAG employs an iterative process to enhance both the defense and attack\nagents. This continuous improvement process strengthens defenses against newly\ngenerated jailbreak prompts. Our empirical studies affirm ICAG's efficacy,\nwhere LLMs safeguarded by ICAG exhibit significantly reduced jailbreak success\nrates across various attack scenarios. Moreover, ICAG demonstrates remarkable\ntransferability to other LLMs, indicating its potential as a versatile defense\nmechanism.\n", "date": "2024-02-20", "categories": []}
{"id": "2409.14424", "title": "Dormant: Defending against Pose-driven Human Image Animation", "abstract": "  Pose-driven human image animation has achieved tremendous progress, enabling\nthe generation of vivid and realistic human videos from just one single photo.\nHowever, it conversely exacerbates the risk of image misuse, as attackers may\nuse one available image to create videos involving politics, violence, and\nother illegal content. To counter this threat, we propose Dormant, a novel\nprotection approach tailored to defend against pose-driven human image\nanimation techniques. Dormant applies protective perturbation to one human\nimage, preserving the visual similarity to the original but resulting in\npoor-quality video generation. The protective perturbation is optimized to\ninduce misextraction of appearance features from the image and create\nincoherence among the generated video frames. Our extensive evaluation across 8\nanimation methods and 4 datasets demonstrates the superiority of Dormant over 6\nbaseline protection methods, leading to misaligned identities, visual\ndistortions, noticeable artifacts, and inconsistent frames in the generated\nvideos. Moreover, Dormant shows effectiveness on 6 real-world commercial\nservices, even with fully black-box access.\n", "date": "2024-09-22", "categories": []}
{"id": "2410.06287", "title": "Non-Halting Queries: Exploiting Fixed Points in LLMs", "abstract": "  We introduce a new vulnerability that exploits fixed points in autoregressive\nmodels and use it to craft queries that never halt. More precisely, for\nnon-halting queries, the LLM never samples the end-of-string token <eos>. We\nrigorously analyze the conditions under which the non-halting anomaly presents\nitself. In particular, at temperature zero, we prove that if a repeating\n(cyclic) token sequence is observed at the output beyond the context size, then\nthe LLM does not halt.\n  We demonstrate non-halting queries in many experiments performed in base\nunaligned models where repeating prompts immediately lead to a non-halting\ncyclic behavior as predicted by the analysis. Further, we develop a simple\nrecipe that takes the same fixed points observed in the base model and creates\na prompt structure to target aligned models. We demonstrate the recipe's\nsuccess in sending every major model released over the past year into a\nnon-halting state with the same simple prompt even over higher temperatures.\nFurther, we devise an experiment with 100 randomly selected tokens and show\nthat the recipe to create non-halting queries succeeds with high success rates\nranging from 97% for GPT-4o to 19% for Gemini Pro 1.5. These results show that\nthe proposed adversarial recipe succeeds in bypassing alignment at one to two\norders of magnitude higher rates compared to earlier reports.\n  We also study gradient-based direct inversion using ARCA to craft new short\nprompts to induce the non-halting state. We inverted 10,000 random repeating\n2-cycle outputs for llama-3.1-8b-instruct. Out of 10,000 three-token inverted\nprompts 1,512 yield non-halting queries reaching a rate of 15%. Our experiments\nwith ARCA show that non-halting may be easily induced with as few as 3 input\ntokens with high probability. Overall, our experiments demonstrate that\nnon-halting queries are prevalent and relatively easy to find.\n", "date": "2024-10-08", "categories": []}
{"id": "2412.12145", "title": "Na'vi or Knave: Jailbreaking Language Models via Metaphorical Avatars", "abstract": "  Metaphor serves as an implicit approach to convey information, while enabling\nthe generalized comprehension of complex subjects. However, metaphor can\npotentially be exploited to bypass the safety alignment mechanisms of Large\nLanguage Models (LLMs), leading to the theft of harmful knowledge. In our\nstudy, we introduce a novel attack framework that exploits the imaginative\ncapacity of LLMs to achieve jailbreaking, the J\\underline{\\textbf{A}}ilbreak\n\\underline{\\textbf{V}}ia \\underline{\\textbf{A}}dversarial\nMe\\underline{\\textbf{TA}} -pho\\underline{\\textbf{R}} (\\textit{AVATAR}).\nSpecifically, to elicit the harmful response, AVATAR extracts harmful entities\nfrom a given harmful target and maps them to innocuous adversarial entities\nbased on LLM's imagination. Then, according to these metaphors, the harmful\ntarget is nested within human-like interaction for jailbreaking adaptively.\nExperimental results demonstrate that AVATAR can effectively and transferablly\njailbreak LLMs and achieve a state-of-the-art attack success rate across\nmultiple advanced LLMs. Our study exposes a security risk in LLMs from their\nendogenous imaginative capabilities. Furthermore, the analytical study reveals\nthe vulnerability of LLM to adversarial metaphors and the necessity of\ndeveloping defense methods against jailbreaking caused by the adversarial\nmetaphor. \\textcolor{orange}{ \\textbf{Warning: This paper contains potentially\nharmful content from LLMs.}}\n", "date": "2024-12-10", "categories": []}
{"id": "2502.16012", "title": "Cross-Model Transferability of Adversarial Patches in Real-time\n  Segmentation for Autonomous Driving", "abstract": "  Adversarial attacks pose a significant threat to deep learning models,\nparticularly in safety-critical applications like healthcare and autonomous\ndriving. Recently, patch based attacks have demonstrated effectiveness in\nreal-time inference scenarios owing to their 'drag and drop' nature. Following\nthis idea for Semantic Segmentation (SS), here we propose a novel Expectation\nOver Transformation (EOT) based adversarial patch attack that is more realistic\nfor autonomous vehicles. To effectively train this attack we also propose a\n'simplified' loss function that is easy to analyze and implement. Using this\nattack as our basis, we investigate whether adversarial patches once optimized\non a specific SS model, can fool other models or architectures. We conduct a\ncomprehensive cross-model transferability analysis of adversarial patches\ntrained on SOTA Convolutional Neural Network (CNN) models such PIDNet-S,\nPIDNet-M and PIDNet-L, among others. Additionally, we also include the\nSegformer model to study transferability to Vision Transformers (ViTs). All of\nour analysis is conducted on the widely used Cityscapes dataset. Our study\nreveals key insights into how model architectures (CNN vs CNN or CNN vs.\nTransformer-based) influence attack susceptibility. In particular, we conclude\nthat although the transferability (effectiveness) of attacks on unseen images\nof any dimension is really high, the attacks trained against one particular\nmodel are minimally effective on other models. And this was found to be true\nfor both ViT and CNN based models. Additionally our results also indicate that\nfor CNN-based models, the repercussions of patch attacks are local, unlike\nViTs. Per-class analysis reveals that simple-classes like 'sky' suffer less\nmisclassification than others. The code for the project is available at:\nhttps://github.com/p-shekhar/adversarial-patch-transferability\n", "date": "2025-02-22", "categories": []}
{"id": "2502.16044", "title": "A Multi-Scale Isolation Forest Approach for Real-Time Detection and\n  Filtering of FGSM Adversarial Attacks in Video Streams of Autonomous Vehicles", "abstract": "  Deep Neural Networks (DNNs) have demonstrated remarkable success across a\nwide range of tasks, particularly in fields such as image classification.\nHowever, DNNs are highly susceptible to adversarial attacks, where subtle\nperturbations are introduced to input images, leading to erroneous model\noutputs. In today's digital era, ensuring the security and integrity of images\nprocessed by DNNs is of critical importance. One of the most prominent\nadversarial attack methods is the Fast Gradient Sign Method (FGSM), which\nperturbs images in the direction of the loss gradient to deceive the model.\n  This paper presents a novel approach for detecting and filtering FGSM\nadversarial attacks in image processing tasks. Our proposed method evaluates\n10,000 images, each subjected to five different levels of perturbation,\ncharacterized by $\\epsilon$ values of 0.01, 0.02, 0.05, 0.1, and 0.2. These\nperturbations are applied in the direction of the loss gradient. We demonstrate\nthat our approach effectively filters adversarially perturbed images,\nmitigating the impact of FGSM attacks.\n  The method is implemented in Python, and the source code is publicly\navailable on GitHub for reproducibility and further research.\n", "date": "2025-02-22", "categories": []}
{"id": "2502.16167", "title": "PersGuard: Preventing Malicious Personalization via Backdoor Attacks on\n  Pre-trained Text-to-Image Diffusion Models", "abstract": "  Diffusion models (DMs) have revolutionized data generation, particularly in\ntext-to-image (T2I) synthesis. However, the widespread use of personalized\ngenerative models raises significant concerns regarding privacy violations and\ncopyright infringement. To address these issues, researchers have proposed\nadversarial perturbation-based protection techniques. However, these methods\nhave notable limitations, including insufficient robustness against data\ntransformations and the inability to fully eliminate identifiable features of\nprotected objects in the generated output. In this paper, we introduce\nPersGuard, a novel backdoor-based approach that prevents malicious\npersonalization of specific images. Unlike traditional adversarial perturbation\nmethods, PersGuard implant backdoor triggers into pre-trained T2I models,\npreventing the generation of customized outputs for designated protected images\nwhile allowing normal personalization for unprotected ones. Unfortunately,\nexisting backdoor methods for T2I diffusion models fail to be applied to\npersonalization scenarios due to the different backdoor objectives and the\npotential backdoor elimination during downstream fine-tuning processes. To\naddress these, we propose three novel backdoor objectives specifically designed\nfor personalization scenarios, coupled with backdoor retention loss engineered\nto resist downstream fine-tuning. These components are integrated into a\nunified optimization framework. Extensive experimental evaluations demonstrate\nPersGuard's effectiveness in preserving data privacy, even under challenging\nconditions including gray-box settings, multi-object protection, and facial\nidentity scenarios. Our method significantly outperforms existing techniques,\noffering a more robust solution for privacy and copyright protection.\n", "date": "2025-02-22", "categories": []}
{"id": "2502.16423", "title": "Unified Prompt Attack Against Text-to-Image Generation Models", "abstract": "  Text-to-Image (T2I) models have advanced significantly, but their growing\npopularity raises security concerns due to their potential to generate harmful\nimages. To address these issues, we propose UPAM, a novel framework to evaluate\nthe robustness of T2I models from an attack perspective. Unlike prior methods\nthat focus solely on textual defenses, UPAM unifies the attack on both textual\nand visual defenses. Additionally, it enables gradient-based optimization,\novercoming reliance on enumeration for improved efficiency and effectiveness.\nTo handle cases where T2I models block image outputs due to defenses, we\nintroduce Sphere-Probing Learning (SPL) to enable optimization even without\nimage results. Following SPL, our model bypasses defenses, inducing the\ngeneration of harmful content. To ensure semantic alignment with attacker\nintent, we propose Semantic-Enhancing Learning (SEL) for precise semantic\ncontrol. UPAM also prioritizes the naturalness of adversarial prompts using\nIn-context Naturalness Enhancement (INE), making them harder for human\nexaminers to detect. Additionally, we address the issue of iterative\nqueries--common in prior methods and easily detectable by API defenders--by\nintroducing Transferable Attack Learning (TAL), allowing effective attacks with\nminimal queries. Extensive experiments validate UPAM's superiority in\neffectiveness, efficiency, naturalness, and low query detection rates.\n", "date": "2025-02-23", "categories": []}
{"id": "2502.16734", "title": "Towards Optimal Adversarial Robust Reinforcement Learning with Infinity\n  Measurement Error", "abstract": "  Ensuring the robustness of deep reinforcement learning (DRL) agents against\nadversarial attacks is critical for their trustworthy deployment. Recent\nresearch highlights the challenges of achieving state-adversarial robustness\nand suggests that an optimal robust policy (ORP) does not always exist,\ncomplicating the enforcement of strict robustness constraints. In this paper,\nwe further explore the concept of ORP. We first introduce the Intrinsic\nState-adversarial Markov Decision Process (ISA-MDP), a novel formulation where\nadversaries cannot fundamentally alter the intrinsic nature of state\nobservations. ISA-MDP, supported by empirical and theoretical evidence,\nuniversally characterizes decision-making under state-adversarial paradigms. We\nrigorously prove that within ISA-MDP, a deterministic and stationary ORP\nexists, aligning with the Bellman optimal policy. Our findings theoretically\nreveal that improving DRL robustness does not necessarily compromise\nperformance in natural environments. Furthermore, we demonstrate the necessity\nof infinity measurement error (IME) in both $Q$-function and probability spaces\nto achieve ORP, unveiling vulnerabilities of previous DRL algorithms that rely\non $1$-measurement errors. Motivated by these insights, we develop the\nConsistent Adversarial Robust Reinforcement Learning (CAR-RL) framework, which\noptimizes surrogates of IME. We apply CAR-RL to both value-based and\npolicy-based DRL algorithms, achieving superior performance and validating our\ntheoretical analysis.\n", "date": "2025-02-23", "categories": []}
{"id": "2502.16737", "title": "Keeping up with dynamic attackers: Certifying robustness to adaptive\n  online data poisoning", "abstract": "  The rise of foundation models fine-tuned on human feedback from potentially\nuntrusted users has increased the risk of adversarial data poisoning,\nnecessitating the study of robustness of learning algorithms against such\nattacks. Existing research on provable certified robustness against data\npoisoning attacks primarily focuses on certifying robustness for static\nadversaries who modify a fraction of the dataset used to train the model before\nthe training algorithm is applied. In practice, particularly when learning from\nhuman feedback in an online sense, adversaries can observe and react to the\nlearning process and inject poisoned samples that optimize adversarial\nobjectives better than when they are restricted to poisoning a static dataset\nonce, before the learning algorithm is applied. Indeed, it has been shown in\nprior work that online dynamic adversaries can be significantly more powerful\nthan static ones. We present a novel framework for computing certified bounds\non the impact of dynamic poisoning, and use these certificates to design robust\nlearning algorithms. We give an illustration of the framework for the mean\nestimation and binary classification problems and outline directions for\nextending this in further work. The code to implement our certificates and\nreplicate our results is available at\nhttps://github.com/Avinandan22/Certified-Robustness.\n", "date": "2025-02-23", "categories": []}
{"id": "2502.17003", "title": "Improving the Transferability of Adversarial Examples by Inverse\n  Knowledge Distillation", "abstract": "  In recent years, the rapid development of deep neural networks has brought\nincreased attention to the security and robustness of these models. While\nexisting adversarial attack algorithms have demonstrated success in improving\nadversarial transferability, their performance remains suboptimal due to a lack\nof consideration for the discrepancies between target and source models. To\naddress this limitation, we propose a novel method, Inverse Knowledge\nDistillation (IKD), designed to enhance adversarial transferability\neffectively. IKD introduces a distillation-inspired loss function that\nseamlessly integrates with gradient-based attack methods, promoting diversity\nin attack gradients and mitigating overfitting to specific model architectures.\nBy diversifying gradients, IKD enables the generation of adversarial samples\nwith superior generalization capabilities across different models,\nsignificantly enhancing their effectiveness in black-box attack scenarios.\nExtensive experiments on the ImageNet dataset validate the effectiveness of our\napproach, demonstrating substantial improvements in the transferability and\nattack success rates of adversarial samples across a wide range of models.\n", "date": "2025-02-24", "categories": []}
{"id": "2502.17121", "title": "Adversarial Training for Defense Against Label Poisoning Attacks", "abstract": "  As machine learning models grow in complexity and increasingly rely on\npublicly sourced data, such as the human-annotated labels used in training\nlarge language models, they become more vulnerable to label poisoning attacks.\nThese attacks, in which adversaries subtly alter the labels within a training\ndataset, can severely degrade model performance, posing significant risks in\ncritical applications. In this paper, we propose FLORAL, a novel adversarial\ntraining defense strategy based on support vector machines (SVMs) to counter\nthese threats. Utilizing a bilevel optimization framework, we cast the training\nprocess as a non-zero-sum Stackelberg game between an attacker, who\nstrategically poisons critical training labels, and the model, which seeks to\nrecover from such attacks. Our approach accommodates various model\narchitectures and employs a projected gradient descent algorithm with kernel\nSVMs for adversarial training. We provide a theoretical analysis of our\nalgorithm's convergence properties and empirically evaluate FLORAL's\neffectiveness across diverse classification tasks. Compared to robust baselines\nand foundation models such as RoBERTa, FLORAL consistently achieves higher\nrobust accuracy under increasing attacker budgets. These results underscore the\npotential of FLORAL to enhance the resilience of machine learning models\nagainst label poisoning threats, thereby ensuring robust classification in\nadversarial settings.\n", "date": "2025-02-24", "categories": []}
{"id": "2502.17254", "title": "REINFORCE Adversarial Attacks on Large Language Models: An Adaptive,\n  Distributional, and Semantic Objective", "abstract": "  To circumvent the alignment of large language models (LLMs), current\noptimization-based adversarial attacks usually craft adversarial prompts by\nmaximizing the likelihood of a so-called affirmative response. An affirmative\nresponse is a manually designed start of a harmful answer to an inappropriate\nrequest. While it is often easy to craft prompts that yield a substantial\nlikelihood for the affirmative response, the attacked model frequently does not\ncomplete the response in a harmful manner. Moreover, the affirmative objective\nis usually not adapted to model-specific preferences and essentially ignores\nthe fact that LLMs output a distribution over responses. If low attack success\nunder such an objective is taken as a measure of robustness, the true\nrobustness might be grossly overestimated. To alleviate these flaws, we propose\nan adaptive and semantic optimization problem over the population of responses.\nWe derive a generally applicable objective via the REINFORCE policy-gradient\nformalism and demonstrate its efficacy with the state-of-the-art jailbreak\nalgorithms Greedy Coordinate Gradient (GCG) and Projected Gradient Descent\n(PGD). For example, our objective doubles the attack success rate (ASR) on\nLlama3 and increases the ASR from 2% to 50% with circuit breaker defense.\n", "date": "2025-02-24", "categories": []}
{"id": "2502.17392", "title": "Emoti-Attack: Zero-Perturbation Adversarial Attacks on NLP Systems via\n  Emoji Sequences", "abstract": "  Deep neural networks (DNNs) have achieved remarkable success in the field of\nnatural language processing (NLP), leading to widely recognized applications\nsuch as ChatGPT. However, the vulnerability of these models to adversarial\nattacks remains a significant concern. Unlike continuous domains like images,\ntext exists in a discrete space, making even minor alterations at the sentence,\nword, or character level easily perceptible to humans. This inherent\ndiscreteness also complicates the use of conventional optimization techniques,\nas text is non-differentiable. Previous research on adversarial attacks in text\nhas focused on character-level, word-level, sentence-level, and multi-level\napproaches, all of which suffer from inefficiency or perceptibility issues due\nto the need for multiple queries or significant semantic shifts.\n  In this work, we introduce a novel adversarial attack method, Emoji-Attack,\nwhich leverages the manipulation of emojis to create subtle, yet effective,\nperturbations. Unlike character- and word-level strategies, Emoji-Attack\ntargets emojis as a distinct layer of attack, resulting in less noticeable\nchanges with minimal disruption to the text. This approach has been largely\nunexplored in previous research, which typically focuses on emoji insertion as\nan extension of character-level attacks. Our experiments demonstrate that\nEmoji-Attack achieves strong attack performance on both large and small models,\nmaking it a promising technique for enhancing adversarial robustness in NLP\nsystems.\n", "date": "2025-02-24", "categories": []}
{"id": "2502.11910", "title": "Adversarial Alignment for LLMs Requires Simpler, Reproducible, and More\n  Measurable Objectives", "abstract": "  Misaligned research objectives have considerably hindered progress in\nadversarial robustness research over the past decade. For instance, an\nextensive focus on optimizing target metrics, while neglecting rigorous\nstandardized evaluation, has led researchers to pursue ad-hoc heuristic\ndefenses that were seemingly effective. Yet, most of these were exposed as\nflawed by subsequent evaluations, ultimately contributing little measurable\nprogress to the field. In this position paper, we illustrate that current\nresearch on the robustness of large language models (LLMs) risks repeating past\npatterns with potentially worsened real-world implications. To address this, we\nargue that realigned objectives are necessary for meaningful progress in\nadversarial alignment. To this end, we build on established cybersecurity\ntaxonomy to formally define differences between past and emerging threat models\nthat apply to LLMs. Using this framework, we illustrate that progress requires\ndisentangling adversarial alignment into addressable sub-problems and returning\nto core academic principles, such as measureability, reproducibility, and\ncomparability. Although the field presents significant challenges, the fresh\nstart on adversarial robustness offers the unique opportunity to build on past\nexperience while avoiding previous mistakes.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.14976", "title": "EigenShield: Causal Subspace Filtering via Random Matrix Theory for\n  Adversarially Robust Vision-Language Models", "abstract": "  Vision-Language Models (VLMs) inherit adversarial vulnerabilities of Large\nLanguage Models (LLMs), which are further exacerbated by their multimodal\nnature. Existing defenses, including adversarial training, input\ntransformations, and heuristic detection, are computationally expensive,\narchitecture-dependent, and fragile against adaptive attacks. We introduce\nEigenShield, an inference-time defense leveraging Random Matrix Theory to\nquantify adversarial disruptions in high-dimensional VLM representations.\nUnlike prior methods that rely on empirical heuristics, EigenShield employs the\nspiked covariance model to detect structured spectral deviations. Using a\nRobustness-based Nonconformity Score (RbNS) and quantile-based thresholding, it\nseparates causal eigenvectors, which encode semantic information, from\ncorrelational eigenvectors that are susceptible to adversarial artifacts. By\nprojecting embeddings onto the causal subspace, EigenShield filters adversarial\nnoise without modifying model parameters or requiring adversarial training.\nThis architecture-independent, attack-agnostic approach significantly reduces\nthe attack success rate, establishing spectral analysis as a principled\nalternative to conventional defenses. Our results demonstrate that EigenShield\nconsistently outperforms all existing defenses, including adversarial training,\nUNIGUARD, and CIDER.\n", "date": "2025-02-20", "categories": []}
{"id": "2502.15017", "title": "Interpreting Adversarial Attacks and Defences using Architectures with\n  Enhanced Interpretability", "abstract": "  Adversarial attacks in deep learning represent a significant threat to the\nintegrity and reliability of machine learning models. Adversarial training has\nbeen a popular defence technique against these adversarial attacks. In this\nwork, we capitalize on a network architecture, namely Deep Linearly Gated\nNetworks (DLGN), which has better interpretation capabilities than regular deep\nnetwork architectures. Using this architecture, we interpret robust models\ntrained using PGD adversarial training and compare them with standard training.\nFeature networks in DLGN act as feature extractors, making them the only medium\nthrough which an adversary can attack the model. We analyze the feature network\nof DLGN with fully connected layers with respect to properties like alignment\nof the hyperplanes, hyperplane relation with PCA, and sub-network overlap among\nclasses and compare these properties between robust and standard models. We\nalso consider this architecture having CNN layers wherein we qualitatively\n(using visualizations) and quantitatively contrast gating patterns between\nrobust and standard models. We uncover insights into hyperplanes resembling\nprincipal components in PGD-AT and STD-TR models, with PGD-AT hyperplanes\naligned farther from the data points. We use path activity analysis to show\nthat PGD-AT models create diverse, non-overlapping active subnetworks across\nclasses, preventing attack-induced gating overlaps. Our visualization ideas\nshow the nature of representations learnt by PGD-AT and STD-TR models.\n", "date": "2025-02-20", "categories": []}
{"id": "2502.15435", "title": "Single-pass Detection of Jailbreaking Input in Large Language Models", "abstract": "  Defending aligned Large Language Models (LLMs) against jailbreaking attacks\nis a challenging problem, with existing approaches requiring multiple requests\nor even queries to auxiliary LLMs, making them computationally heavy. Instead,\nwe focus on detecting jailbreaking input in a single forward pass. Our method,\ncalled Single Pass Detection SPD, leverages the information carried by the\nlogits to predict whether the output sentence will be harmful. This allows us\nto defend in just one forward pass. SPD can not only detect attacks effectively\non open-source models, but also minimizes the misclassification of harmless\ninputs. Furthermore, we show that SPD remains effective even without complete\nlogit access in GPT-3.5 and GPT-4. We believe that our proposed method offers a\npromising approach to efficiently safeguard LLMs against adversarial attacks.\n", "date": "2025-02-21", "categories": []}
{"id": "2502.15561", "title": "A Defensive Framework Against Adversarial Attacks on Machine\n  Learning-Based Network Intrusion Detection Systems", "abstract": "  As cyberattacks become increasingly sophisticated, advanced Network Intrusion\nDetection Systems (NIDS) are critical for modern network security. Traditional\nsignature-based NIDS are inadequate against zero-day and evolving attacks. In\nresponse, machine learning (ML)-based NIDS have emerged as promising solutions;\nhowever, they are vulnerable to adversarial evasion attacks that subtly\nmanipulate network traffic to bypass detection. To address this vulnerability,\nwe propose a novel defensive framework that enhances the robustness of ML-based\nNIDS by simultaneously integrating adversarial training, dataset balancing\ntechniques, advanced feature engineering, ensemble learning, and extensive\nmodel fine-tuning. We validate our framework using the NSL-KDD and UNSW-NB15\ndatasets. Experimental results show, on average, a 35% increase in detection\naccuracy and a 12.5% reduction in false positives compared to baseline models,\nparticularly under adversarial conditions. The proposed defense against\nadversarial attacks significantly advances the practical deployment of robust\nML-based NIDS in real-world networks.\n", "date": "2025-02-21", "categories": []}
{"id": "2308.11845", "title": "SEA: Shareable and Explainable Attribution for Query-based Black-box\n  Attacks", "abstract": "  Machine Learning (ML) systems are vulnerable to adversarial examples,\nparticularly those from query-based black-box attacks. Despite various efforts\nto detect and prevent such attacks, ML systems are still at risk, demanding a\nmore comprehensive approach to security that includes logging, analyzing, and\nsharing evidence. While traditional security benefits from well-established\npractices of forensics and threat intelligence sharing, ML security has yet to\nfind a way to profile its attackers and share information about them. In\nresponse, this paper introduces SEA, a novel ML security system to characterize\nblack-box attacks on ML systems for forensic purposes and to facilitate\nhuman-explainable intelligence sharing. SEA leverages Hidden Markov Models to\nattribute the observed query sequence to known attacks. It thus understands the\nattack's progression rather than focusing solely on the final adversarial\nexamples. Our evaluations reveal that SEA is effective at attack attribution,\neven on the second incident, and is robust to adaptive strategies designed to\nevade forensic analysis. SEA's explanations of the attack's behavior allow us\neven to fingerprint specific minor bugs in widely used attack libraries. For\nexample, we discover that the SignOPT and Square attacks in ART v1.14 send over\n50% duplicated queries. We thoroughly evaluate SEA on a variety of settings and\ndemonstrate that it can recognize the same attack with more than 90% Top-1 and\n95% Top-3 accuracy. Finally, we demonstrate how SEA generalizes to other\ndomains like text classification.\n", "date": "2023-08-23", "categories": []}
{"id": "2408.04683", "title": "Eliminating Backdoors in Neural Code Models for Secure Code\n  Understanding", "abstract": "  Neural code models (NCMs) have been widely used to address various code\nunderstanding tasks, such as defect detection. However, numerous recent studies\nreveal that such models are vulnerable to backdoor attacks. Backdoored NCMs\nfunction normally on normal/clean code snippets, but exhibit adversary-expected\nbehavior on poisoned code snippets injected with the adversary-crafted trigger.\nIt poses a significant security threat. Therefore, there is an urgent need for\neffective techniques to detect and eliminate backdoors stealthily implanted in\nNCMs.\n  To address this issue, in this paper, we innovatively propose a backdoor\nelimination technique for secure code understanding, called EliBadCode.\nEliBadCode eliminates backdoors in NCMs by inverting/reverse-engineering and\nunlearning backdoor triggers. Specifically, EliBadCode first filters the model\nvocabulary for trigger tokens based on the naming conventions of specific\nprogramming languages to reduce the trigger search space and cost. Then,\nEliBadCode introduces a sample-specific trigger position identification method,\nwhich can reduce the interference of non-backdoor (adversarial) perturbations\nfor subsequent trigger inversion, thereby producing effective inverted backdoor\ntriggers efficiently. Backdoor triggers can be viewed as backdoor (adversarial)\nperturbations. Subsequently, EliBadCode employs a Greedy Coordinate Gradient\nalgorithm to optimize the inverted trigger and designs a trigger anchoring\nmethod to purify the inverted trigger. Finally, EliBadCode eliminates backdoors\nthrough model unlearning. We evaluate the effectiveness of EliBadCode in\neliminating backdoors implanted in multiple NCMs used for three safety-critical\ncode understanding tasks. The results demonstrate that EliBadCode can\neffectively eliminate backdoors while having minimal adverse effects on the\nnormal functionality of the model.\n", "date": "2024-08-08", "categories": []}
{"id": "2501.13676", "title": "Certified Robustness Under Bounded Levenshtein Distance", "abstract": "  Text classifiers suffer from small perturbations, that if chosen\nadversarially, can dramatically change the output of the model. Verification\nmethods can provide robustness certificates against such adversarial\nperturbations, by computing a sound lower bound on the robust accuracy.\nNevertheless, existing verification methods incur in prohibitive costs and\ncannot practically handle Levenshtein distance constraints. We propose the\nfirst method for computing the Lipschitz constant of convolutional classifiers\nwith respect to the Levenshtein distance. We use these Lipschitz constant\nestimates for training 1-Lipschitz classifiers. This enables computing the\ncertified radius of a classifier in a single forward pass. Our method, LipsLev,\nis able to obtain $38.80$% and $13.93$% verified accuracy at distance $1$ and\n$2$ respectively in the AG-News dataset, while being $4$ orders of magnitude\nfaster than existing approaches. We believe our work can open the door to more\nefficient verification in the text domain.\n", "date": "2025-01-23", "categories": []}
{"id": "2502.10490", "title": "A Robust Attack: Displacement Backdoor Attack", "abstract": "  As artificial intelligence becomes more prevalent in our lives, people are\nenjoying the convenience it brings, but they are also facing hidden threats,\nsuch as data poisoning and adversarial attacks. These threats can have\ndisastrous consequences for the application of artificial intelligence,\nespecially for some applications that take effect immediately, such as\nautonomous driving and medical fields. Among these threats, backdoor attacks\nhave left a deep impression on people with their concealment and simple\ndeployment, making them a threat that cannot be ignored, however, in the\nprocess of deploying the backdoor model, the backdoor attack often has some\nreasons that make it unsatisfactory in real-world applications, such as jitter\nand brightness changes. Based on this, we propose a highly robust backdoor\nattack that shifts the target sample and combines it with itself to form a\nbackdoor sample, the Displacement Backdoor Attack(DBA). Experimental results\nshow that the DBA attack can resist data augmentation that simulates real-world\ndifferences, such as rotation and cropping.\n", "date": "2025-02-14", "categories": []}
{"id": "2502.14298", "title": "Generalization Certificates for Adversarially Robust Bayesian Linear\n  Regression", "abstract": "  Adversarial robustness of machine learning models is critical to ensuring\nreliable performance under data perturbations. Recent progress has been on\npoint estimators, and this paper considers distributional predictors. First,\nusing the link between exponential families and Bregman divergences, we\nformulate an adversarial Bregman divergence loss as an adversarial negative\nlog-likelihood. Using the geometric properties of Bregman divergences, we\ncompute the adversarial perturbation for such models in closed-form. Second,\nunder such losses, we introduce \\emph{adversarially robust posteriors}, by\nexploiting the optimization-centric view of generalized Bayesian inference.\nThird, we derive the \\emph{first} rigorous generalization certificates in the\ncontext of an adversarial extension of Bayesian linear regression by leveraging\nthe PAC-Bayesian framework. Finally, experiments on real and synthetic datasets\ndemonstrate the superior robustness of the derived adversarially robust\nposterior over Bayes posterior, and also validate our theoretical guarantees.\n", "date": "2025-02-20", "categories": []}
{"id": "2502.14586", "title": "Moshi Moshi? A Model Selection Hijacking Adversarial Attack", "abstract": "  Model selection is a fundamental task in Machine Learning~(ML), focusing on\nselecting the most suitable model from a pool of candidates by evaluating their\nperformance on specific metrics. This process ensures optimal performance,\ncomputational efficiency, and adaptability to diverse tasks and environments.\nDespite its critical role, its security from the perspective of adversarial ML\nremains unexplored. This risk is heightened in the\nMachine-Learning-as-a-Service model, where users delegate the training phase\nand the model selection process to third-party providers, supplying data and\ntraining strategies. Therefore, attacks on model selection could harm both the\nuser and the provider, undermining model performance and driving up operational\ncosts.\n  In this work, we present MOSHI (MOdel Selection HIjacking adversarial\nattack), the first adversarial attack specifically targeting model selection.\nOur novel approach manipulates model selection data to favor the adversary,\neven without prior knowledge of the system. Utilizing a framework based on\nVariational Auto Encoders, we provide evidence that an attacker can induce\ninefficiencies in ML deployment. We test our attack on diverse computer vision\nand speech recognition benchmark tasks and different settings, obtaining an\naverage attack success rate of 75.42%. In particular, our attack causes an\naverage 88.30% decrease in generalization capabilities, an 83.33% increase in\nlatency, and an increase of up to 105.85% in energy consumption. These results\nhighlight the significant vulnerabilities in model selection processes and\ntheir potential impact on real-world applications.\n", "date": "2025-02-20", "categories": []}
{"id": "2406.16342", "title": "Is your benchmark truly adversarial? AdvScore: Evaluating Human-Grounded\n  Adversarialness", "abstract": "  Adversarial datasets should validate AI robustness by providing samples on\nwhich humans perform well, but models do not. However, as models evolve,\ndatasets can become obsolete. Measuring whether a dataset remains adversarial\nis hindered by the lack of a standardized metric for measuring adversarialness.\nWe propose AdvScore, a human-grounded evaluation metric that assesses a\ndataset's adversarialness by capturing models' and humans' varying abilities\nwhile also identifying poor examples. We then use AdvScore to motivate a new\ndataset creation pipeline for realistic and high-quality adversarial samples,\nenabling us to collect an adversarial question answering (QA) dataset, AdvQA.\nWe apply AdvScore using 9,347 human responses and ten language models'\npredictions to track model improvement over five years, from 2020 to 2024.\nAdvScore thus provides guidance for achieving robustness comparable with human\ncapabilities. Furthermore, it helps determine to what extent adversarial\ndatasets continue to pose challenges, ensuring that, rather than reflecting\noutdated or overly artificial difficulties, they effectively test model\ncapabilities.\n", "date": "2024-06-24", "categories": []}
{"id": "2408.11810", "title": "Pixel Is Not a Barrier: An Effective Evasion Attack for Pixel-Domain\n  Diffusion Models", "abstract": "  Diffusion Models have emerged as powerful generative models for high-quality\nimage synthesis, with many subsequent image editing techniques based on them.\nHowever, the ease of text-based image editing introduces significant risks,\nsuch as malicious editing for scams or intellectual property infringement.\nPrevious works have attempted to safeguard images from diffusion-based editing\nby adding imperceptible perturbations. These methods are costly and\nspecifically target prevalent Latent Diffusion Models (LDMs), while\nPixel-domain Diffusion Models (PDMs) remain largely unexplored and robust\nagainst such attacks. Our work addresses this gap by proposing a novel attack\nframework, AtkPDM. AtkPDM is mainly composed of a feature representation\nattacking loss that exploits vulnerabilities in denoising UNets and a latent\noptimization strategy to enhance the naturalness of adversarial images.\nExtensive experiments demonstrate the effectiveness of our approach in\nattacking dominant PDM-based editing methods (e.g., SDEdit) while maintaining\nreasonable fidelity and robustness against common defense methods.\nAdditionally, our framework is extensible to LDMs, achieving comparable\nperformance to existing approaches.\n", "date": "2024-08-21", "categories": []}
{"id": "2410.04234", "title": "Functional Homotopy: Smoothing Discrete Optimization via Continuous\n  Parameters for LLM Jailbreak Attacks", "abstract": "  Optimization methods are widely employed in deep learning to identify and\nmitigate undesired model responses. While gradient-based techniques have proven\neffective for image models, their application to language models is hindered by\nthe discrete nature of the input space. This study introduces a novel\noptimization approach, termed the \\emph{functional homotopy} method, which\nleverages the functional duality between model training and input generation.\nBy constructing a series of easy-to-hard optimization problems, we iteratively\nsolve these problems using principles derived from established homotopy\nmethods. We apply this approach to jailbreak attack synthesis for large\nlanguage models (LLMs), achieving a $20\\%-30\\%$ improvement in success rate\nover existing methods in circumventing established safe open-source models such\nas Llama-2 and Llama-3.\n", "date": "2024-10-05", "categories": []}
{"id": "2410.19785", "title": "How to Backdoor Consistency Models?", "abstract": "  Consistency models are a new class of models that generate images by directly\nmapping noise to data, allowing for one-step generation and significantly\naccelerating the sampling process. However, their robustness against\nadversarial attacks has not yet been thoroughly investigated. In this work, we\nconduct the first study on the vulnerability of consistency models to backdoor\nattacks. While previous research has explored backdoor attacks on diffusion\nmodels, those studies have primarily focused on conventional diffusion models,\nemploying a customized backdoor training process and objective, whereas\nconsistency models have distinct training processes and objectives. Our\nproposed framework demonstrates the vulnerability of consistency models to\nbackdoor attacks. During image generation, poisoned consistency models produce\nimages with a Fr\\'echet Inception Distance (FID) comparable to that of a clean\nmodel when sampling from Gaussian noise. However, once the trigger is\nactivated, they generate backdoor target images. We explore various trigger and\ntarget configurations to evaluate the vulnerability of consistency models,\nincluding the use of random noise as a trigger. This novel trigger is visually\ninconspicuous, more challenging to detect, and aligns well with the sampling\nprocess of consistency models. Across all configurations, our framework\nsuccessfully compromises the consistency models while maintaining high utility\nand specificity. We also examine the stealthiness of our proposed attack, which\nis attributed to the unique properties of consistency models and the elusive\nnature of the Gaussian noise trigger. Our code is available at\n\\href{https://github.com/chengenw/backdoorCM}{https://github.com/chengenw/backdoorCM}.\n", "date": "2024-10-14", "categories": []}
{"id": "2411.08133", "title": "Impactful Bit-Flip Search on Full-precision Models", "abstract": "  Neural networks have shown remarkable performance in various tasks, yet they\nremain susceptible to subtle changes in their input or model parameters. One\nparticularly impactful vulnerability arises through the Bit-Flip Attack (BFA),\nwhere flipping a small number of critical bits in a model's parameters can\nseverely degrade its performance. A common technique for inducing bit flips in\nDRAM is the Row-Hammer attack, which exploits frequent uncached memory accesses\nto alter data. Identifying susceptible bits can be achieved through exhaustive\nsearch or progressive layer-by-layer analysis, especially in quantized\nnetworks. In this work, we introduce Impactful Bit-Flip Search (IBS), a novel\nmethod for efficiently pinpointing and flipping critical bits in full-precision\nnetworks. Additionally, we propose a Weight-Stealth technique that\nstrategically modifies the model's parameters in a way that maintains the float\nvalues within the original distribution, thereby bypassing simple range checks\noften used in tamper detection.\n", "date": "2024-11-12", "categories": []}
{"id": "2502.10487", "title": "Fast Proxies for LLM Robustness Evaluation", "abstract": "  Evaluating the robustness of LLMs to adversarial attacks is crucial for safe\ndeployment, yet current red-teaming methods are often prohibitively expensive.\nWe compare the ability of fast proxy metrics to predict the real-world\nrobustness of an LLM against a simulated attacker ensemble. This allows us to\nestimate a model's robustness to computationally expensive attacks without\nrequiring runs of the attacks themselves. Specifically, we consider\ngradient-descent-based embedding-space attacks, prefilling attacks, and direct\nprompting. Even though direct prompting in particular does not achieve high\nASR, we find that it and embedding-space attacks can predict attack success\nrates well, achieving $r_p=0.87$ (linear) and $r_s=0.94$ (Spearman rank)\ncorrelations with the full attack ensemble while reducing computational cost by\nthree orders of magnitude.\n", "date": "2025-02-14", "categories": []}
{"id": "2502.11379", "title": "CCJA: Context-Coherent Jailbreak Attack for Aligned Large Language\n  Models", "abstract": "  Despite explicit alignment efforts for large language models (LLMs), they can\nstill be exploited to trigger unintended behaviors, a phenomenon known as\n\"jailbreaking.\" Current jailbreak attack methods mainly focus on discrete\nprompt manipulations targeting closed-source LLMs, relying on manually crafted\nprompt templates and persuasion rules. However, as the capabilities of\nopen-source LLMs improve, ensuring their safety becomes increasingly crucial.\nIn such an environment, the accessibility of model parameters and gradient\ninformation by potential attackers exacerbates the severity of jailbreak\nthreats. To address this research gap, we propose a novel\n\\underline{C}ontext-\\underline{C}oherent \\underline{J}ailbreak\n\\underline{A}ttack (CCJA). We define jailbreak attacks as an optimization\nproblem within the embedding space of masked language models. Through\ncombinatorial optimization, we effectively balance the jailbreak attack success\nrate with semantic coherence. Extensive evaluations show that our method not\nonly maintains semantic consistency but also surpasses state-of-the-art\nbaselines in attack effectiveness. Additionally, by integrating semantically\ncoherent jailbreak prompts generated by our method into widely used black-box\nmethodologies, we observe a notable enhancement in their success rates when\ntargeting closed-source commercial LLMs. This highlights the security threat\nposed by open-source LLMs to commercial counterparts. We will open-source our\ncode if the paper is accepted.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.11725", "title": "Adversarially Robust CLIP Models Can Induce Better (Robust) Perceptual\n  Metrics", "abstract": "  Measuring perceptual similarity is a key tool in computer vision. In recent\nyears perceptual metrics based on features extracted from neural networks with\nlarge and diverse training sets, e.g. CLIP, have become popular. At the same\ntime, the metrics extracted from features of neural networks are not\nadversarially robust. In this paper we show that adversarially robust CLIP\nmodels, called R-CLIP$_\\textrm{F}$, obtained by unsupervised adversarial\nfine-tuning induce a better and adversarially robust perceptual metric that\noutperforms existing metrics in a zero-shot setting, and further matches the\nperformance of state-of-the-art metrics while being robust after fine-tuning.\nMoreover, our perceptual metric achieves strong performance on related tasks\nsuch as robust image-to-image retrieval, which becomes especially relevant when\napplied to \"Not Safe for Work\" (NSFW) content detection and dataset filtering.\nWhile standard perceptual metrics can be easily attacked by a small\nperturbation completely degrading NSFW detection, our robust perceptual metric\nmaintains high accuracy under an attack while having similar performance for\nunperturbed images. Finally, perceptual metrics induced by robust CLIP models\nhave higher interpretability: feature inversion can show which images are\nconsidered similar, while text inversion can find what images are associated to\na given prompt. This also allows us to visualize the very rich visual concepts\nlearned by a CLIP model, including memorized persons, paintings and complex\nqueries.\n", "date": "2025-02-17", "categories": []}
{"id": "2305.17438", "title": "On the Importance of Backbone to the Adversarial Robustness of Object\n  Detectors", "abstract": "  Object detection is a critical component of various security-sensitive\napplications, such as autonomous driving and video surveillance. However,\nexisting object detectors are vulnerable to adversarial attacks, which poses a\nsignificant challenge to their reliability and security. Through experiments,\nfirst, we found that existing works on improving the adversarial robustness of\nobject detectors give a false sense of security. Second, we found that\nadversarially pre-trained backbone networks were essential for enhancing the\nadversarial robustness of object detectors. We then proposed a simple yet\neffective recipe for fast adversarial fine-tuning on object detectors with\nadversarially pre-trained backbones. Without any modifications to the structure\nof object detectors, our recipe achieved significantly better adversarial\nrobustness than previous works. Finally, we explored the potential of different\nmodern object detector designs for improving adversarial robustness with our\nrecipe and demonstrated interesting findings, which inspired us to design\nstate-of-the-art (SOTA) robust detectors. Our empirical results set a new\nmilestone for adversarially robust object detection. Code and trained\ncheckpoints are available at https://github.com/thu-ml/oddefense.\n", "date": "2023-05-27", "categories": []}
{"id": "2402.00626", "title": "Vision-LLMs Can Fool Themselves with Self-Generated Typographic Attacks", "abstract": "  Typographic attacks, adding misleading text to images, can deceive\nvision-language models (LVLMs). The susceptibility of recent large LVLMs like\nGPT4-V to such attacks is understudied, raising concerns about amplified\nmisinformation in personal assistant applications. Previous attacks use simple\nstrategies, such as random misleading words, which don't fully exploit LVLMs'\nlanguage reasoning abilities. We introduce an experimental setup for testing\ntypographic attacks on LVLMs and propose two novel self-generated attacks: (1)\nClass-based attacks, where the model identifies a similar class to deceive\nitself, and (2) Reasoned attacks, where an advanced LVLM suggests an attack\ncombining a deceiving class and description. Our experiments show these attacks\nsignificantly reduce classification performance by up to 60\\% and are effective\nacross different models, including InstructBLIP and MiniGPT4. Code:\nhttps://github.com/mqraitem/Self-Gen-Typo-Attack\n", "date": "2024-02-01", "categories": []}
{"id": "2412.08555", "title": "Grimm: A Plug-and-Play Perturbation Rectifier for Graph Neural Networks\n  Defending against Poisoning Attacks", "abstract": "  Recent studies have revealed the vulnerability of graph neural networks\n(GNNs) to adversarial poisoning attacks on node classification tasks. Current\ndefensive methods require substituting the original GNNs with defense models,\nregardless of the original's type. This approach, while targeting adversarial\nrobustness, compromises the enhancements developed in prior research to boost\nGNNs' practical performance. Here we introduce Grimm, the first plug-and-play\ndefense model. With just a minimal interface requirement for extracting\nfeatures from any layer of the protected GNNs, Grimm is thus enabled to\nseamlessly rectify perturbations. Specifically, we utilize the feature\ntrajectories (FTs) generated by GNNs, as they evolve through epochs, to reflect\nthe training status of the networks. We then theoretically prove that the FTs\nof victim nodes will inevitably exhibit discriminable anomalies. Consequently,\ninspired by the natural parallelism between the biological nervous and immune\nsystems, we construct Grimm, a comprehensive artificial immune system for GNNs.\nGrimm not only detects abnormal FTs and rectifies adversarial edges during\ntraining but also operates efficiently in parallel, thereby mirroring the\nconcurrent functionalities of its biological counterparts. We experimentally\nconfirm that Grimm offers four empirically validated advantages: 1)\nHarmlessness, as it does not actively interfere with GNN training; 2)\nParallelism, ensuring monitoring, detection, and rectification functions\noperate independently of the GNN training process; 3) Generalizability,\ndemonstrating compatibility with mainstream GNNs such as GCN, GAT, and\nGraphSAGE; and 4) Transferability, as the detectors for abnormal FTs can be\nefficiently transferred across different systems for one-step rectification.\n", "date": "2024-12-11", "categories": []}
{"id": "2502.09352", "title": "Wasserstein distributional adversarial training for deep neural networks", "abstract": "  Design of adversarial attacks for deep neural networks, as well as methods of\nadversarial training against them, are subject of intense research. In this\npaper, we propose methods to train against distributional attack threats,\nextending the TRADES method used for pointwise attacks. Our approach leverages\nrecent contributions and relies on sensitivity analysis for Wasserstein\ndistributionally robust optimization problems. We introduce an efficient\nfine-tuning method which can be deployed on a previously trained model. We test\nour methods on a range of pre-trained models on RobustBench. These experimental\nresults demonstrate the additional training enhances Wasserstein distributional\nrobustness, while maintaining original levels of pointwise robustness, even for\nalready very successful networks. The improvements are less marked for models\npre-trained using huge synthetic datasets of 20-100M images. However,\nremarkably, sometimes our methods are still able to improve their performance\neven when trained using only the original training dataset (50k images).\n", "date": "2025-02-13", "categories": []}
{"id": "2409.02483", "title": "TASAR: Transfer-based Attack on Skeletal Action Recognition", "abstract": "  Skeletal sequence data, as a widely employed representation of human actions,\nare crucial in Human Activity Recognition (HAR). Recently, adversarial attacks\nhave been proposed in this area, which exposes potential security concerns, and\nmore importantly provides a good tool for model robustness test. Within this\nresearch, transfer-based attack is an important tool as it mimics the\nreal-world scenario where an attacker has no knowledge of the target model, but\nis under-explored in Skeleton-based HAR (S-HAR). Consequently, existing S-HAR\nattacks exhibit weak adversarial transferability and the reason remains largely\nunknown. In this paper, we investigate this phenomenon via the characterization\nof the loss function. We find that one prominent indicator of poor\ntransferability is the low smoothness of the loss function. Led by this\nobservation, we improve the transferability by properly smoothening the loss\nwhen computing the adversarial examples. This leads to the first Transfer-based\nAttack on Skeletal Action Recognition, TASAR. TASAR explores the smoothened\nmodel posterior of pre-trained surrogates, which is achieved by a new\npost-train Dual Bayesian optimization strategy. Furthermore, unlike existing\ntransfer-based methods which overlook the temporal coherence within sequences,\nTASAR incorporates motion dynamics into the Bayesian attack, effectively\ndisrupting the spatial-temporal coherence of S-HARs. For exhaustive evaluation,\nwe build the first large-scale robust S-HAR benchmark, comprising 7 S-HAR\nmodels, 10 attack methods, 3 S-HAR datasets and 2 defense models. Extensive\nresults demonstrate the superiority of TASAR. Our benchmark enables easy\ncomparisons for future studies, with the code available in the\nhttps://github.com/yunfengdiao/Skeleton-Robustness-Benchmark.\n", "date": "2024-09-04", "categories": []}
{"id": "2502.07821", "title": "Amnesia as a Catalyst for Enhancing Black Box Pixel Attacks in Image\n  Classification and Object Detection", "abstract": "  It is well known that query-based attacks tend to have relatively higher\nsuccess rates in adversarial black-box attacks. While research on black-box\nattacks is actively being conducted, relatively few studies have focused on\npixel attacks that target only a limited number of pixels. In image\nclassification, query-based pixel attacks often rely on patches, which heavily\ndepend on randomness and neglect the fact that scattered pixels are more\nsuitable for adversarial attacks. Moreover, to the best of our knowledge,\nquery-based pixel attacks have not been explored in the field of object\ndetection. To address these issues, we propose a novel pixel-based black-box\nattack called Remember and Forget Pixel Attack using Reinforcement\nLearning(RFPAR), consisting of two main components: the Remember and Forget\nprocesses. RFPAR mitigates randomness and avoids patch dependency by leveraging\nrewards generated through a one-step RL algorithm to perturb pixels. RFPAR\neffectively creates perturbed images that minimize the confidence scores while\nadhering to limited pixel constraints. Furthermore, we advance our proposed\nattack beyond image classification to object detection, where RFPAR reduces the\nconfidence scores of detected objects to avoid detection. Experiments on the\nImageNet-1K dataset for classification show that RFPAR outperformed\nstate-of-the-art query-based pixel attacks. For object detection, using the\nMSCOCO dataset with YOLOv8 and DDQ, RFPAR demonstrates comparable mAP reduction\nto state-of-the-art query-based attack while requiring fewer query. Further\nexperiments on the Argoverse dataset using YOLOv8 confirm that RFPAR\neffectively removed objects on a larger scale dataset. Our code is available at\nhttps://github.com/KAU-QuantumAILab/RFPAR.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.08097", "title": "ID-Cloak: Crafting Identity-Specific Cloaks Against Personalized\n  Text-to-Image Generation", "abstract": "  Personalized text-to-image models allow users to generate images of new\nconcepts from several reference photos, thereby leading to critical concerns\nregarding civil privacy. Although several anti-personalization techniques have\nbeen developed, these methods typically assume that defenders can afford to\ndesign a privacy cloak corresponding to each specific image. However, due to\nextensive personal images shared online, image-specific methods are limited by\nreal-world practical applications. To address this issue, we are the first to\ninvestigate the creation of identity-specific cloaks (ID-Cloak) that safeguard\nall images belong to a specific identity. Specifically, we first model an\nidentity subspace that preserves personal commonalities and learns diverse\ncontexts to capture the image distribution to be protected. Then, we craft\nidentity-specific cloaks with the proposed novel objective that encourages the\ncloak to guide the model away from its normal output within the subspace.\nExtensive experiments show that the generated universal cloak can effectively\nprotect the images. We believe our method, along with the proposed\nidentity-specific cloak setting, marks a notable advance in realistic privacy\nprotection.\n", "date": "2025-02-12", "categories": []}
{"id": "2502.08374", "title": "AdvSwap: Covert Adversarial Perturbation with High Frequency\n  Info-swapping for Autonomous Driving Perception", "abstract": "  Perception module of Autonomous vehicles (AVs) are increasingly susceptible\nto be attacked, which exploit vulnerabilities in neural networks through\nadversarial inputs, thereby compromising the AI safety. Some researches focus\non creating covert adversarial samples, but existing global noise techniques\nare detectable and difficult to deceive the human visual system. This paper\nintroduces a novel adversarial attack method, AdvSwap, which creatively\nutilizes wavelet-based high-frequency information swapping to generate covert\nadversarial samples and fool the camera. AdvSwap employs invertible neural\nnetwork for selective high-frequency information swapping, preserving both\nforward propagation and data integrity. The scheme effectively removes the\noriginal label data and incorporates the guidance image data, producing\nconcealed and robust adversarial samples. Experimental evaluations and\ncomparisons on the GTSRB and nuScenes datasets demonstrate that AdvSwap can\nmake concealed attacks on common traffic targets. The generates adversarial\nsamples are also difficult to perceive by humans and algorithms. Meanwhile, the\nmethod has strong attacking robustness and attacking transferability.\n", "date": "2025-02-12", "categories": []}
{"id": "2408.08502", "title": "Efficient Image-to-Image Diffusion Classifier for Adversarial Robustness", "abstract": "  Diffusion models (DMs) have demonstrated great potential in the field of\nadversarial robustness, where DM-based defense methods can achieve superior\ndefense capability without adversarial training. However, they all require huge\ncomputational costs due to the usage of large-scale pre-trained DMs, making it\ndifficult to conduct full evaluation under strong attacks and compare with\ntraditional CNN-based methods. Simply reducing the network size and timesteps\nin DMs could significantly harm the image generation quality, which invalidates\nprevious frameworks. To alleviate this issue, we redesign the diffusion\nframework from generating high-quality images to predicting distinguishable\nimage labels. Specifically, we employ an image translation framework to learn\nmany-to-one mapping from input samples to designed orthogonal image labels.\nBased on this framework, we introduce an efficient Image-to-Image diffusion\nclassifier with a pruned U-Net structure and reduced diffusion timesteps.\nBesides the framework, we redesign the optimization objective of DMs to fit the\ntarget of image classification, where a new classification loss is incorporated\nin the DM-based image translation framework to distinguish the generated label\nfrom those of other classes. We conduct sufficient evaluations of the proposed\nclassifier under various attacks on popular benchmarks. Extensive experiments\nshow that our method achieves better adversarial robustness with fewer\ncomputational costs than DM-based and CNN-based methods. The code is available\nat https://github.com/hfmei/IDC\n", "date": "2024-08-16", "categories": []}
{"id": "2502.06917", "title": "Krum Federated Chain (KFC): Using blockchain to defend against\n  adversarial attacks in Federated Learning", "abstract": "  Federated Learning presents a nascent approach to machine learning, enabling\ncollaborative model training across decentralized devices while safeguarding\ndata privacy. However, its distributed nature renders it susceptible to\nadversarial attacks. Integrating blockchain technology with Federated Learning\noffers a promising avenue to enhance security and integrity. In this paper, we\ntackle the potential of blockchain in defending Federated Learning against\nadversarial attacks. First, we test Proof of Federated Learning, a well known\nconsensus mechanism designed ad-hoc to federated contexts, as a defense\nmechanism demonstrating its efficacy against Byzantine and backdoor attacks\nwhen at least one miner remains uncompromised. Second, we propose Krum\nFederated Chain, a novel defense strategy combining Krum and Proof of Federated\nLearning, valid to defend against any configuration of Byzantine or backdoor\nattacks, even when all miners are compromised. Our experiments conducted on\nimage classification datasets validate the effectiveness of our proposed\napproaches.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.07232", "title": "Simplifying Adversarially Robust PAC Learning with Tolerance", "abstract": "  Adversarially robust PAC learning has proved to be challenging, with the\ncurrently best known learners [Montasser et al., 2021a] relying on improper\nmethods based on intricate compression schemes, resulting in sample complexity\nexponential in the VC-dimension. A series of follow up work considered a\nslightly relaxed version of the problem called adversarially robust learning\nwith tolerance [Ashtiani et al., 2023, Bhattacharjee et al., 2023, Raman et\nal., 2024] and achieved better sample complexity in terms of the VC-dimension.\nHowever, those algorithms were either improper and complex, or required\nadditional assumptions on the hypothesis class H. We prove, for the first time,\nthe existence of a simpler learner that achieves a sample complexity linear in\nthe VC-dimension without requiring additional assumptions on H. Even though our\nlearner is improper, it is \"almost proper\" in the sense that it outputs a\nhypothesis that is \"similar\" to a hypothesis in H.\n  We also use the ideas from our algorithm to construct a semi-supervised\nlearner in the tolerant setting. This simple algorithm achieves comparable\nbounds to the previous (non-tolerant) semi-supervised algorithm of Attias et\nal. [2022a], but avoids the use of intricate subroutines from previous works,\nand is \"almost proper.\"\n", "date": "2025-02-11", "categories": []}
{"id": "2408.10848", "title": "Perception-guided Jailbreak against Text-to-Image Models", "abstract": "  In recent years, Text-to-Image (T2I) models have garnered significant\nattention due to their remarkable advancements. However, security concerns have\nemerged due to their potential to generate inappropriate or Not-Safe-For-Work\n(NSFW) images. In this paper, inspired by the observation that texts with\ndifferent semantics can lead to similar human perceptions, we propose an\nLLM-driven perception-guided jailbreak method, termed PGJ. It is a black-box\njailbreak method that requires no specific T2I model (model-free) and generates\nhighly natural attack prompts. Specifically, we propose identifying a safe\nphrase that is similar in human perception yet inconsistent in text semantics\nwith the target unsafe word and using it as a substitution. The experiments\nconducted on six open-source models and commercial online services with\nthousands of prompts have verified the effectiveness of PGJ.\n", "date": "2024-08-20", "categories": []}
{"id": "2409.14161", "title": "When Witnesses Defend: A Witness Graph Topological Layer for Adversarial\n  Graph Learning", "abstract": "  Capitalizing on the intuitive premise that shape characteristics are more\nrobust to perturbations, we bridge adversarial graph learning with the emerging\ntools from computational topology, namely, persistent homology representations\nof graphs. We introduce the concept of witness complex to adversarial analysis\non graphs, which allows us to focus only on the salient shape characteristics\nof graphs, yielded by the subset of the most essential nodes (i.e., landmarks),\nwith minimal loss of topological information on the whole graph. The remaining\nnodes are then used as witnesses, governing which higher-order graph\nsubstructures are incorporated into the learning process. Armed with the\nwitness mechanism, we design Witness Graph Topological Layer (WGTL), which\nsystematically integrates both local and global topological graph feature\nrepresentations, the impact of which is, in turn, automatically controlled by\nthe robust regularized topological loss. Given the attacker's budget, we derive\nthe important stability guarantees of both local and global topology encodings\nand the associated robust topological loss. We illustrate the versatility and\nefficiency of WGTL by its integration with five GNNs and three existing\nnon-topological defense mechanisms. Our extensive experiments across six\ndatasets demonstrate that WGTL boosts the robustness of GNNs across a range of\nperturbations and against a range of adversarial attacks. Our datasets and\nsource codes are available at https://github.com/toggled/WGTL.\n", "date": "2024-09-21", "categories": []}
{"id": "2502.04643", "title": "Confidence Elicitation: A New Attack Vector for Large Language Models", "abstract": "  A fundamental issue in deep learning has been adversarial robustness. As\nthese systems have scaled, such issues have persisted. Currently, large\nlanguage models (LLMs) with billions of parameters suffer from adversarial\nattacks just like their earlier, smaller counterparts. However, the threat\nmodels have changed. Previously, having gray-box access, where input embeddings\nor output logits/probabilities were visible to the user, might have been\nreasonable. However, with the introduction of closed-source models, no\ninformation about the model is available apart from the generated output. This\nmeans that current black-box attacks can only utilize the final prediction to\ndetect if an attack is successful. In this work, we investigate and demonstrate\nthe potential of attack guidance, akin to using output probabilities, while\nhaving only black-box access in a classification setting. This is achieved\nthrough the ability to elicit confidence from the model. We empirically show\nthat the elicited confidence is calibrated and not hallucinated for current\nLLMs. By minimizing the elicited confidence, we can therefore increase the\nlikelihood of misclassification. Our new proposed paradigm demonstrates\npromising state-of-the-art results on three datasets across two models\n(LLaMA-3-8B-Instruct and Mistral-7B-Instruct-V0.3) when comparing our technique\nto existing hard-label black-box attack methods that introduce word-level\nsubstitutions.\n", "date": "2025-02-07", "categories": []}
{"id": "2502.05203", "title": "Adversarial Machine Learning: Attacking and Safeguarding Image Datasets", "abstract": "  This paper examines the vulnerabilities of convolutional neural networks\n(CNNs) to adversarial attacks and explores a method for their safeguarding. In\nthis study, CNNs were implemented on four of the most common image datasets,\nnamely CIFAR-10, ImageNet, MNIST, and Fashion-MNIST, and achieved high baseline\naccuracy. To assess the strength of these models, the Fast Gradient Sign Method\nwas used, which is a type of exploit on the model that is used to bring down\nthe models accuracies by adding a very minimal perturbation to the input image.\nTo counter the FGSM attack, a safeguarding approach went through, which\nincludes retraining the models on clear and pollutant or adversarial images to\nincrease their resistance ability. The next step involves applying FGSM again,\nbut this time to the adversarially trained models, to see how much the accuracy\nof the models has gone down and evaluate the effectiveness of the defense. It\nappears that while most level of robustness is achieved against the models\nafter adversarial training, there are still a few losses in the performance of\nthese models against adversarial perturbations. This work emphasizes the need\nto create better defenses for models deployed in real-world scenarios against\nadversaries.\n", "date": "2025-01-31", "categories": []}
{"id": "2502.05542", "title": "Democratic Training Against Universal Adversarial Perturbations", "abstract": "  Despite their advances and success, real-world deep neural networks are known\nto be vulnerable to adversarial attacks. Universal adversarial perturbation, an\ninput-agnostic attack, poses a serious threat for them to be deployed in\nsecurity-sensitive systems. In this case, a single universal adversarial\nperturbation deceives the model on a range of clean inputs without requiring\ninput-specific optimization, which makes it particularly threatening. In this\nwork, we observe that universal adversarial perturbations usually lead to\nabnormal entropy spectrum in hidden layers, which suggests that the prediction\nis dominated by a small number of ``feature'' in such cases (rather than\ndemocratically by many features). Inspired by this, we propose an efficient yet\neffective defense method for mitigating UAPs called \\emph{Democratic Training}\nby performing entropy-based model enhancement to suppress the effect of the\nuniversal adversarial perturbations in a given model. \\emph{Democratic\nTraining} is evaluated with 7 neural networks trained on 5 benchmark datasets\nand 5 types of state-of-the-art universal adversarial attack methods. The\nresults show that it effectively reduces the attack success rate, improves\nmodel robustness and preserves the model accuracy on clean samples.\n", "date": "2025-02-08", "categories": []}
{"id": "2502.05637", "title": "Adversarial Machine Learning: Attacks, Defenses, and Open Challenges", "abstract": "  Adversarial Machine Learning (AML) addresses vulnerabilities in AI systems\nwhere adversaries manipulate inputs or training data to degrade performance.\nThis article provides a comprehensive analysis of evasion and poisoning\nattacks, formalizes defense mechanisms with mathematical rigor, and discusses\nthe challenges of implementing robust solutions in adaptive threat models.\nAdditionally, it highlights open challenges in certified robustness,\nscalability, and real-world deployment.\n", "date": "2025-02-08", "categories": []}
{"id": "2502.05772", "title": "Effective Black-Box Multi-Faceted Attacks Breach Vision Large Language\n  Model Guardrails", "abstract": "  Vision Large Language Models (VLLMs) integrate visual data processing,\nexpanding their real-world applications, but also increasing the risk of\ngenerating unsafe responses. In response, leading companies have implemented\nMulti-Layered safety defenses, including alignment training, safety system\nprompts, and content moderation. However, their effectiveness against\nsophisticated adversarial attacks remains largely unexplored. In this paper, we\npropose MultiFaceted Attack, a novel attack framework designed to\nsystematically bypass Multi-Layered Defenses in VLLMs. It comprises three\ncomplementary attack facets: Visual Attack that exploits the multimodal nature\nof VLLMs to inject toxic system prompts through images; Alignment Breaking\nAttack that manipulates the model's alignment mechanism to prioritize the\ngeneration of contrasting responses; and Adversarial Signature that deceives\ncontent moderators by strategically placing misleading information at the end\nof the response. Extensive evaluations on eight commercial VLLMs in a black-box\nsetting demonstrate that MultiFaceted Attack achieves a 61.56% attack success\nrate, surpassing state-of-the-art methods by at least 42.18%.\n", "date": "2025-02-09", "categories": []}
{"id": "2403.13196", "title": "ADAPT to Robustify Prompt Tuning Vision Transformers", "abstract": "  The performance of deep models, including Vision Transformers, is known to be\nvulnerable to adversarial attacks. Many existing defenses against these\nattacks, such as adversarial training, rely on full-model fine-tuning to induce\nrobustness in the models. These defenses require storing a copy of the entire\nmodel, that can have billions of parameters, for each task. At the same time,\nparameter-efficient prompt tuning is used to adapt large transformer-based\nmodels to downstream tasks without the need to save large copies. In this\npaper, we examine parameter-efficient prompt tuning of Vision Transformers for\ndownstream tasks under the lens of robustness. We show that previous\nadversarial defense methods, when applied to the prompt tuning paradigm, suffer\nfrom gradient obfuscation and are vulnerable to adaptive attacks. We introduce\nADAPT, a novel framework for performing adaptive adversarial training in the\nprompt tuning paradigm. Our method achieves competitive robust accuracy of ~40%\nw.r.t. SOTA robustness methods using full-model fine-tuning, by tuning only ~1%\nof the number of parameters.\n", "date": "2024-03-19", "categories": []}
{"id": "2410.08503", "title": "Adversarial Training Can Provably Improve Robustness: Theoretical\n  Analysis of Feature Learning Process Under Structured Data", "abstract": "  Adversarial training is a widely-applied approach to training deep neural\nnetworks to be robust against adversarial perturbation. However, although\nadversarial training has achieved empirical success in practice, it still\nremains unclear why adversarial examples exist and how adversarial training\nmethods improve model robustness. In this paper, we provide a theoretical\nunderstanding of adversarial examples and adversarial training algorithms from\nthe perspective of feature learning theory. Specifically, we focus on a\nmultiple classification setting, where the structured data can be composed of\ntwo types of features: the robust features, which are resistant to perturbation\nbut sparse, and the non-robust features, which are susceptible to perturbation\nbut dense. We train a two-layer smoothed ReLU convolutional neural network to\nlearn our structured data. First, we prove that by using standard training\n(gradient descent over the empirical risk), the network learner primarily\nlearns the non-robust feature rather than the robust feature, which thereby\nleads to the adversarial examples that are generated by perturbations aligned\nwith negative non-robust feature directions. Then, we consider the\ngradient-based adversarial training algorithm, which runs gradient ascent to\nfind adversarial examples and runs gradient descent over the empirical risk at\nadversarial examples to update models. We show that the adversarial training\nmethod can provably strengthen the robust feature learning and suppress the\nnon-robust feature learning to improve the network robustness. Finally, we also\nempirically validate our theoretical findings with experiments on real-image\ndatasets, including MNIST, CIFAR10 and SVHN.\n", "date": "2024-10-11", "categories": []}
{"id": "2502.04360", "title": "MARAGE: Transferable Multi-Model Adversarial Attack for\n  Retrieval-Augmented Generation Data Extraction", "abstract": "  Retrieval-Augmented Generation (RAG) offers a solution to mitigate\nhallucinations in Large Language Models (LLMs) by grounding their outputs to\nknowledge retrieved from external sources. The use of private resources and\ndata in constructing these external data stores can expose them to risks of\nextraction attacks, in which attackers attempt to steal data from these private\ndatabases. Existing RAG extraction attacks often rely on manually crafted\nprompts, which limit their effectiveness. In this paper, we introduce a\nframework called MARAGE for optimizing an adversarial string that, when\nappended to user queries submitted to a target RAG system, causes outputs\ncontaining the retrieved RAG data verbatim. MARAGE leverages a continuous\noptimization scheme that integrates gradients from multiple models with\ndifferent architectures simultaneously to enhance the transferability of the\noptimized string to unseen models. Additionally, we propose a strategy that\nemphasizes the initial tokens in the target RAG data, further improving the\nattack's generalizability. Evaluations show that MARAGE consistently\noutperforms both manual and optimization-based baselines across multiple LLMs\nand RAG datasets, while maintaining robust transferability to previously unseen\nmodels. Moreover, we conduct probing tasks to shed light on the reasons why\nMARAGE is more effective compared to the baselines and to analyze the impact of\nour approach on the model's internal state.\n", "date": "2025-02-05", "categories": []}
{"id": "2502.04679", "title": "Mechanistic Understandings of Representation Vulnerabilities and\n  Engineering Robust Vision Transformers", "abstract": "  While transformer-based models dominate NLP and vision applications, their\nunderlying mechanisms to map the input space to the label space semantically\nare not well understood. In this paper, we study the sources of known\nrepresentation vulnerabilities of vision transformers (ViT), where perceptually\nidentical images can have very different representations and semantically\nunrelated images can have the same representation. Our analysis indicates that\nimperceptible changes to the input can result in significant representation\nchanges, particularly in later layers, suggesting potential instabilities in\nthe performance of ViTs. Our comprehensive study reveals that adversarial\neffects, while subtle in early layers, propagate and amplify through the\nnetwork, becoming most pronounced in middle to late layers. This insight\nmotivates the development of NeuroShield-ViT, a novel defense mechanism that\nstrategically neutralizes vulnerable neurons in earlier layers to prevent the\ncascade of adversarial effects. We demonstrate NeuroShield-ViT's effectiveness\nacross various attacks, particularly excelling against strong iterative\nattacks, and showcase its remarkable zero-shot generalization capabilities.\nWithout fine-tuning, our method achieves a competitive accuracy of 77.8% on\nadversarial examples, surpassing conventional robustness methods. Our results\nshed new light on how adversarial effects propagate through ViT layers, while\nproviding a promising approach to enhance the robustness of vision transformers\nagainst adversarial attacks. Additionally, they provide a promising approach to\nenhance the robustness of vision transformers against adversarial attacks.\n", "date": "2025-02-07", "categories": []}
{"id": "2502.04771", "title": "DMPA: Model Poisoning Attacks on Decentralized Federated Learning for\n  Model Differences", "abstract": "  Federated learning (FL) has garnered significant attention as a prominent\nprivacy-preserving Machine Learning (ML) paradigm. Decentralized FL (DFL)\neschews traditional FL's centralized server architecture, enhancing the\nsystem's robustness and scalability. However, these advantages of DFL also\ncreate new vulnerabilities for malicious participants to execute adversarial\nattacks, especially model poisoning attacks. In model poisoning attacks,\nmalicious participants aim to diminish the performance of benign models by\ncreating and disseminating the compromised model. Existing research on model\npoisoning attacks has predominantly concentrated on undermining global models\nwithin the Centralized FL (CFL) paradigm, while there needs to be more research\nin DFL. To fill the research gap, this paper proposes an innovative model\npoisoning attack called DMPA. This attack calculates the differential\ncharacteristics of multiple malicious client models and obtains the most\neffective poisoning strategy, thereby orchestrating a collusive attack by\nmultiple participants. The effectiveness of this attack is validated across\nmultiple datasets, with results indicating that the DMPA approach consistently\nsurpasses existing state-of-the-art FL model poisoning attack strategies.\n", "date": "2025-02-07", "categories": []}
{"id": "2502.05000", "title": "Robust Graph Learning Against Adversarial Evasion Attacks via Prior-Free\n  Diffusion-Based Structure Purification", "abstract": "  Adversarial evasion attacks pose significant threats to graph learning, with\nlines of studies that have improved the robustness of Graph Neural Networks\n(GNNs). However, existing works rely on priors about clean graphs or attacking\nstrategies, which are often heuristic and inconsistent. To achieve robust graph\nlearning over different types of evasion attacks and diverse datasets, we\ninvestigate this problem from a prior-free structure purification perspective.\nSpecifically, we propose a novel Diffusion-based Structure Purification\nframework named DiffSP, which creatively incorporates the graph diffusion model\nto learn intrinsic distributions of clean graphs and purify the perturbed\nstructures by removing adversaries under the direction of the captured\npredictive patterns without relying on priors. DiffSP is divided into the\nforward diffusion process and the reverse denoising process, during which\nstructure purification is achieved. To avoid valuable information loss during\nthe forward process, we propose an LID-driven nonisotropic diffusion mechanism\nto selectively inject noise anisotropically. To promote semantic alignment\nbetween the clean graph and the purified graph generated during the reverse\nprocess, we reduce the generation uncertainty by the proposed graph transfer\nentropy guided denoising mechanism. Extensive experiments demonstrate the\nsuperior robustness of DiffSP against evasion attacks.\n", "date": "2025-02-07", "categories": []}
{"id": "2408.11071", "title": "DiffZOO: A Purely Query-Based Black-Box Attack for Red-teaming\n  Text-to-Image Generative Model via Zeroth Order Optimization", "abstract": "  Current text-to-image (T2I) synthesis diffusion models raise misuse concerns,\nparticularly in creating prohibited or not-safe-for-work (NSFW) images. To\naddress this, various safety mechanisms and red teaming attack methods are\nproposed to enhance or expose the T2I model's capability to generate unsuitable\ncontent. However, many red teaming attack methods assume knowledge of the text\nencoders, limiting their practical usage. In this work, we rethink the case of\n\\textit{purely black-box} attacks without prior knowledge of the T2l model. To\novercome the unavailability of gradients and the inability to optimize attacks\nwithin a discrete prompt space, we propose DiffZOO which applies Zeroth Order\nOptimization to procure gradient approximations and harnesses both C-PRV and\nD-PRV to enhance attack prompts within the discrete prompt domain. We evaluated\nour method across multiple safety mechanisms of the T2I diffusion model and\nonline servers. Experiments on multiple state-of-the-art safety mechanisms show\nthat DiffZOO attains an 8.5% higher average attack success rate than previous\nworks, hence its promise as a practical red teaming tool for T2l models.\n", "date": "2024-08-18", "categories": []}
{"id": "2411.16591", "title": "Adversarial Attacks for Drift Detection", "abstract": "  Concept drift refers to the change of data distributions over time. While\ndrift poses a challenge for learning models, requiring their continual\nadaption, it is also relevant in system monitoring to detect malfunctions,\nsystem failures, and unexpected behavior. In the latter case, the robust and\nreliable detection of drifts is imperative. This work studies the shortcomings\nof commonly used drift detection schemes. We show how to construct data streams\nthat are drifting without being detected. We refer to those as drift\nadversarials. In particular, we compute all possible adversairals for common\ndetection schemes and underpin our theoretical findings with empirical\nevaluations.\n", "date": "2024-11-25", "categories": []}
{"id": "2502.04248", "title": "Adapting to Evolving Adversaries with Regularized Continual Robust\n  Training", "abstract": "  Robust training methods typically defend against specific attack types, such\nas Lp attacks with fixed budgets, and rarely account for the fact that\ndefenders may encounter new attacks over time. A natural solution is to adapt\nthe defended model to new adversaries as they arise via fine-tuning, a method\nwhich we call continual robust training (CRT). However, when implemented\nnaively, fine-tuning on new attacks degrades robustness on previous attacks.\nThis raises the question: how can we improve the initial training and\nfine-tuning of the model to simultaneously achieve robustness against previous\nand new attacks? We present theoretical results which show that the gap in a\nmodel's robustness against different attacks is bounded by how far each attack\nperturbs a sample in the model's logit space, suggesting that regularizing with\nrespect to this logit space distance can help maintain robustness against\nprevious attacks. Extensive experiments on 3 datasets (CIFAR-10, CIFAR-100, and\nImageNette) and over 100 attack combinations demonstrate that the proposed\nregularization improves robust accuracy with little overhead in training time.\nOur findings and open-source code lay the groundwork for the deployment of\nmodels robust to evolving attacks.\n", "date": "2025-02-06", "categories": []}
{"id": "2406.12814", "title": "Dissecting Adversarial Robustness of Multimodal LM Agents", "abstract": "  As language models (LMs) are used to build autonomous agents in real\nenvironments, ensuring their adversarial robustness becomes a critical\nchallenge. Unlike chatbots, agents are compound systems with multiple\ncomponents taking actions, which existing LMs safety evaluations do not\nadequately address. To bridge this gap, we manually create 200 targeted\nadversarial tasks and evaluation scripts in a realistic threat model on top of\nVisualWebArena, a real environment for web agents. To systematically examine\nthe robustness of agents, we propose the Agent Robustness Evaluation (ARE)\nframework. ARE views the agent as a graph showing the flow of intermediate\noutputs between components and decomposes robustness as the flow of adversarial\ninformation on the graph. We find that we can successfully break latest agents\nthat use black-box frontier LMs, including those that perform reflection and\ntree search. With imperceptible perturbations to a single image (less than 5%\nof total web page pixels), an attacker can hijack these agents to execute\ntargeted adversarial goals with success rates up to 67%. We also use ARE to\nrigorously evaluate how the robustness changes as new components are added. We\nfind that inference-time compute that typically improves benign performance can\nopen up new vulnerabilities and harm robustness. An attacker can compromise the\nevaluator used by the reflexion agent and the value function of the tree search\nagent, which increases the attack success relatively by 15% and 20%. Our data\nand code for attacks, defenses, and evaluation are at\nhttps://github.com/ChenWu98/agent-attack\n", "date": "2024-06-18", "categories": []}
{"id": "2401.15883", "title": "Model Supply Chain Poisoning: Backdooring Pre-trained Models via\n  Embedding Indistinguishability", "abstract": "  Pre-trained models (PTMs) are widely adopted across various downstream tasks\nin the machine learning supply chain. Adopting untrustworthy PTMs introduces\nsignificant security risks, where adversaries can poison the model supply chain\nby embedding hidden malicious behaviors (backdoors) into PTMs. However,\nexisting backdoor attacks to PTMs can only achieve partially task-agnostic and\nthe embedded backdoors are easily erased during the fine-tuning process. This\nmakes it challenging for the backdoors to persist and propagate through the\nsupply chain. In this paper, we propose a novel and severer backdoor attack,\nTransTroj, which enables the backdoors embedded in PTMs to efficiently transfer\nin the model supply chain. In particular, we first formalize this attack as an\nindistinguishability problem between poisoned and clean samples in the\nembedding space. We decompose embedding indistinguishability into pre- and\npost-indistinguishability, representing the similarity of the poisoned and\nreference embeddings before and after the attack. Then, we propose a two-stage\noptimization that separately optimizes triggers and victim PTMs to achieve\nembedding indistinguishability. We evaluate TransTroj on four PTMs and six\ndownstream tasks. Experimental results show that our method significantly\noutperforms SOTA task-agnostic backdoor attacks -- achieving nearly 100% attack\nsuccess rate on most downstream tasks -- and demonstrates robustness under\nvarious system settings. Our findings underscore the urgent need to secure the\nmodel supply chain against such transferable backdoor attacks. The code is\navailable at https://github.com/haowang-cqu/TransTroj .\n", "date": "2024-01-29", "categories": []}
{"id": "2411.14834", "title": "Evaluating the Robustness of the \"Ensemble Everything Everywhere\"\n  Defense", "abstract": "  Ensemble everything everywhere is a defense to adversarial examples that was\nrecently proposed to make image classifiers robust. This defense works by\nensembling a model's intermediate representations at multiple noisy image\nresolutions, producing a single robust classification. This defense was shown\nto be effective against multiple state-of-the-art attacks. Perhaps even more\nconvincingly, it was shown that the model's gradients are perceptually aligned:\nattacks against the model produce noise that perceptually resembles the\ntargeted class.\n  In this short note, we show that this defense is not robust to adversarial\nattack. We first show that the defense's randomness and ensembling method cause\nsevere gradient masking. We then use standard adaptive attack techniques to\nreduce the defense's robust accuracy from 48% to 14% on CIFAR-100 and from 62%\nto 11% on CIFAR-10, under the $\\ell_\\infty$-norm threat model with\n$\\varepsilon=8/255$.\n", "date": "2024-11-22", "categories": []}
{"id": "2412.11638", "title": "IDProtector: An Adversarial Noise Encoder to Protect Against\n  ID-Preserving Image Generation", "abstract": "  Recently, zero-shot methods like InstantID have revolutionized\nidentity-preserving generation. Unlike multi-image finetuning approaches such\nas DreamBooth, these zero-shot methods leverage powerful facial encoders to\nextract identity information from a single portrait photo, enabling efficient\nidentity-preserving generation through a single inference pass. However, this\nconvenience introduces new threats to the facial identity protection. This\npaper aims to safeguard portrait photos from unauthorized encoder-based\ncustomization. We introduce IDProtector, an adversarial noise encoder that\napplies imperceptible adversarial noise to portrait photos in a single forward\npass. Our approach offers universal protection for portraits against multiple\nstate-of-the-art encoder-based methods, including InstantID, IP-Adapter, and\nPhotoMaker, while ensuring robustness to common image transformations such as\nJPEG compression, resizing, and affine transformations. Experiments across\ndiverse portrait datasets and generative models reveal that IDProtector\ngeneralizes effectively to unseen data and even closed-source proprietary\nmodels.\n", "date": "2024-12-16", "categories": []}
{"id": "2501.04108", "title": "TrojanDec: Data-free Detection of Trojan Inputs in Self-supervised\n  Learning", "abstract": "  An image encoder pre-trained by self-supervised learning can be used as a\ngeneral-purpose feature extractor to build downstream classifiers for various\ndownstream tasks. However, many studies showed that an attacker can embed a\ntrojan into an encoder such that multiple downstream classifiers built based on\nthe trojaned encoder simultaneously inherit the trojan behavior. In this work,\nwe propose TrojanDec, the first data-free method to identify and recover a test\ninput embedded with a trigger. Given a (trojaned or clean) encoder and a test\ninput, TrojanDec first predicts whether the test input is trojaned. If not, the\ntest input is processed in a normal way to maintain the utility. Otherwise, the\ntest input will be further restored to remove the trigger. Our extensive\nevaluation shows that TrojanDec can effectively identify the trojan (if any)\nfrom a given test input and recover it under state-of-the-art trojan attacks.\nWe further demonstrate by experiments that our TrojanDec outperforms the\nstate-of-the-art defenses.\n", "date": "2025-01-07", "categories": []}
{"id": "2502.01896", "title": "INTACT: Inducing Noise Tolerance through Adversarial Curriculum Training\n  for LiDAR-based Safety-Critical Perception and Autonomy", "abstract": "  In this work, we present INTACT, a novel two-phase framework designed to\nenhance the robustness of deep neural networks (DNNs) against noisy LiDAR data\nin safety-critical perception tasks. INTACT combines meta-learning with\nadversarial curriculum training (ACT) to systematically address challenges\nposed by data corruption and sparsity in 3D point clouds. The meta-learning\nphase equips a teacher network with task-agnostic priors, enabling it to\ngenerate robust saliency maps that identify critical data regions. The ACT\nphase leverages these saliency maps to progressively expose a student network\nto increasingly complex noise patterns, ensuring targeted perturbation and\nimproved noise resilience. INTACT's effectiveness is demonstrated through\ncomprehensive evaluations on object detection, tracking, and classification\nbenchmarks using diverse datasets, including KITTI, Argoverse, and ModelNet40.\nResults indicate that INTACT improves model robustness by up to 20% across all\ntasks, outperforming standard adversarial and curriculum training methods. This\nframework not only addresses the limitations of conventional training\nstrategies but also offers a scalable and efficient solution for real-world\ndeployment in resource-constrained safety-critical systems. INTACT's principled\nintegration of meta-learning and adversarial training establishes a new\nparadigm for noise-tolerant 3D perception in safety-critical applications.\nINTACT improved KITTI Multiple Object Tracking Accuracy (MOTA) by 9.6% (64.1%\n-> 75.1%) and by 12.4% under Gaussian noise (52.5% -> 73.7%). Similarly, KITTI\nmean Average Precision (mAP) rose from 59.8% to 69.8% (50% point drop) and\n49.3% to 70.9% (Gaussian noise), highlighting the framework's ability to\nenhance deep learning model resilience in safety-critical object tracking\nscenarios.\n", "date": "2025-02-04", "categories": []}
{"id": "2502.01936", "title": "Query-Based and Unnoticeable Graph Injection Attack from Neighborhood\n  Perspective", "abstract": "  The robustness of Graph Neural Networks (GNNs) has become an increasingly\nimportant topic due to their expanding range of applications. Various attack\nmethods have been proposed to explore the vulnerabilities of GNNs, ranging from\nGraph Modification Attacks (GMA) to the more practical and flexible Graph\nInjection Attacks (GIA). However, existing methods face two key challenges: (i)\ntheir reliance on surrogate models, which often leads to reduced attack\neffectiveness due to structural differences and prior biases, and (ii) existing\nGIA methods often sacrifice attack success rates in undefended settings to\nbypass certain defense models, thereby limiting their overall effectiveness. To\novercome these limitations, we propose QUGIA, a Query-based and Unnoticeable\nGraph Injection Attack. QUGIA injects nodes by first selecting edges based on\nvictim node connections and then generating node features using a Bayesian\nframework. This ensures that the injected nodes are similar to the original\ngraph nodes, implicitly preserving homophily and making the attack more\nunnoticeable. Unlike previous methods, QUGIA does not rely on surrogate models,\nthereby avoiding performance degradation and achieving better generalization.\nExtensive experiments on six real-world datasets with diverse characteristics\ndemonstrate that QUGIA achieves unnoticeable attacks and outperforms\nstate-of-the-art attackers. The code will be released upon acceptance.\n", "date": "2025-02-04", "categories": []}
{"id": "2406.02619", "title": "Unelicitable Backdoors in Language Models via Cryptographic Transformer\n  Circuits", "abstract": "  The rapid proliferation of open-source language models significantly\nincreases the risks of downstream backdoor attacks. These backdoors can\nintroduce dangerous behaviours during model deployment and can evade detection\nby conventional cybersecurity monitoring systems. In this paper, we introduce a\nnovel class of backdoors in transformer models, that, in contrast to prior art,\nare unelicitable in nature. Unelicitability prevents the defender from\ntriggering the backdoor, making it impossible to properly evaluate ahead of\ndeployment even if given full white-box access and using automated techniques,\nsuch as red-teaming or certain formal verification methods. We show that our\nnovel construction is not only unelicitable thanks to using cryptographic\ntechniques, but also has favourable robustness properties. We confirm these\nproperties in empirical investigations, and provide evidence that our backdoors\ncan withstand state-of-the-art mitigation strategies. Additionally, we expand\non previous work by showing that our universal backdoors, while not completely\nundetectable in white-box settings, can be harder to detect than some existing\ndesigns. By demonstrating the feasibility of seamlessly integrating backdoors\ninto transformer models, this paper fundamentally questions the efficacy of\npre-deployment detection strategies. This offers new insights into the\noffence-defence balance in AI safety and security.\n", "date": "2024-06-03", "categories": []}
{"id": "2410.01438", "title": "The Great Contradiction Showdown: How Jailbreak and Stealth Wrestle in\n  Vision-Language Models?", "abstract": "  Vision-Language Models (VLMs) have achieved remarkable performance on a\nvariety of tasks, yet they remain vulnerable to jailbreak attacks that\ncompromise safety and reliability. In this paper, we provide an\ninformation-theoretic framework for understanding the fundamental trade-off\nbetween the effectiveness of these attacks and their stealthiness. Drawing on\nFano's inequality, we demonstrate how an attacker's success probability is\nintrinsically linked to the stealthiness of generated prompts. Building on\nthis, we propose an efficient algorithm for detecting non-stealthy jailbreak\nattacks, offering significant improvements in model robustness. Experimental\nresults highlight the tension between strong attacks and their detectability,\nproviding insights into both adversarial strategies and defense mechanisms.\n", "date": "2024-10-02", "categories": []}
{"id": "2412.11735", "title": "Transferable Adversarial Face Attack with Text Controlled Attribute", "abstract": "  Traditional adversarial attacks typically produce adversarial examples under\nnorm-constrained conditions, whereas unrestricted adversarial examples are\nfree-form with semantically meaningful perturbations. Current unrestricted\nadversarial impersonation attacks exhibit limited control over adversarial face\nattributes and often suffer from low transferability. In this paper, we propose\na novel Text Controlled Attribute Attack (TCA$^2$) to generate photorealistic\nadversarial impersonation faces guided by natural language. Specifically, the\ncategory-level personal softmax vector is employed to precisely guide the\nimpersonation attacks. Additionally, we propose both data and model\naugmentation strategies to achieve transferable attacks on unknown target\nmodels. Finally, a generative model, \\textit{i.e}, Style-GAN, is utilized to\nsynthesize impersonated faces with desired attributes. Extensive experiments on\ntwo high-resolution face recognition datasets validate that our TCA$^2$ method\ncan generate natural text-guided adversarial impersonation faces with high\ntransferability. We also evaluate our method on real-world face recognition\nsystems, \\textit{i.e}, Face++ and Aliyun, further demonstrating the practical\npotential of our approach.\n", "date": "2024-12-16", "categories": []}
{"id": "2502.00834", "title": "Boosting Adversarial Robustness and Generalization with Structural Prior", "abstract": "  This work investigates a novel approach to boost adversarial robustness and\ngeneralization by incorporating structural prior into the design of deep\nlearning models. Specifically, our study surprisingly reveals that existing\ndictionary learning-inspired convolutional neural networks (CNNs) provide a\nfalse sense of security against adversarial attacks. To address this, we\npropose Elastic Dictionary Learning Networks (EDLNets), a novel ResNet\narchitecture that significantly enhances adversarial robustness and\ngeneralization. This novel and effective approach is supported by a theoretical\nrobustness analysis using influence functions. Moreover, extensive and reliable\nexperiments demonstrate consistent and significant performance improvement on\nopen robustness leaderboards such as RobustBench, surpassing state-of-the-art\nbaselines. To the best of our knowledge, this is the first work to discover and\nvalidate that structural prior can reliably enhance deep learning robustness\nunder strong adaptive attacks, unveiling a promising direction for future\nresearch.\n", "date": "2025-02-02", "categories": []}
{"id": "2404.14402", "title": "A mean curvature flow arising in adversarial training", "abstract": "  We connect adversarial training for binary classification to a geometric\nevolution equation for the decision boundary. Relying on a perspective that\nrecasts adversarial training as a regularization problem, we introduce a\nmodified training scheme that constitutes a minimizing movements scheme for a\nnonlocal perimeter functional. We prove that the scheme is monotone and\nconsistent as the adversarial budget vanishes and the perimeter localizes, and\nas a consequence we rigorously show that the scheme approximates a weighted\nmean curvature flow. This highlights that the efficacy of adversarial training\nmay be due to locally minimizing the length of the decision boundary. In our\nanalysis, we introduce a variety of tools for working with the subdifferential\nof a supremal-type nonlocal total variation and its regularity properties.\n", "date": "2024-04-22", "categories": []}
{"id": "2411.01703", "title": "UniGuard: Towards Universal Safety Guardrails for Jailbreak Attacks on\n  Multimodal Large Language Models", "abstract": "  Multimodal large language models (MLLMs) have revolutionized vision-language\nunderstanding but remain vulnerable to multimodal jailbreak attacks, where\nadversarial inputs are meticulously crafted to elicit harmful or inappropriate\nresponses. We propose UniGuard, a novel multimodal safety guardrail that\njointly considers the unimodal and cross-modal harmful signals. UniGuard trains\na multimodal guardrail to minimize the likelihood of generating harmful\nresponses in a toxic corpus. The guardrail can be seamlessly applied to any\ninput prompt during inference with minimal computational costs. Extensive\nexperiments demonstrate the generalizability of UniGuard across multiple\nmodalities, attack strategies, and multiple state-of-the-art MLLMs, including\nLLaVA, Gemini Pro, GPT-4o, MiniGPT-4, and InstructBLIP. Notably, this robust\ndefense mechanism maintains the models' overall vision-language understanding\ncapabilities.\n", "date": "2024-11-03", "categories": []}
{"id": "2501.18629", "title": "The Relationship Between Network Similarity and Transferability of\n  Adversarial Attacks", "abstract": "  Neural networks are vulnerable to adversarial attacks, and several defenses\nhave been proposed. Designing a robust network is a challenging task given the\nwide range of attacks that have been developed. Therefore, we aim to provide\ninsight into the influence of network similarity on the success rate of\ntransferred adversarial attacks. Network designers can then compare their new\nnetwork with existing ones to estimate its vulnerability. To achieve this, we\ninvestigate the complex relationship between network similarity and the success\nrate of transferred adversarial attacks. We applied the Centered Kernel\nAlignment (CKA) network similarity score and used various methods to find a\ncorrelation between a large number of Convolutional Neural Networks (CNNs) and\nadversarial attacks. Network similarity was found to be moderate across\ndifferent CNN architectures, with more complex models such as DenseNet showing\nlower similarity scores due to their architectural complexity. Layer similarity\nwas highest for consistent, basic layers such as DataParallel, Dropout and\nConv2d, while specialized layers showed greater variability. Adversarial attack\nsuccess rates were generally consistent for non-transferred attacks, but varied\nsignificantly for some transferred attacks, with complex networks being more\nvulnerable. We found that a DecisionTreeRegressor can predict the success rate\nof transferred attacks for all black-box and Carlini & Wagner attacks with an\naccuracy of over 90%, suggesting that predictive models may be viable under\ncertain conditions. However, the variability of results across different data\nsubsets underscores the complexity of these relationships and suggests that\nfurther research is needed to generalize these findings across different attack\nscenarios and network architectures.\n", "date": "2025-01-27", "categories": []}
{"id": "2501.18841", "title": "Trading Inference-Time Compute for Adversarial Robustness", "abstract": "  We conduct experiments on the impact of increasing inference-time compute in\nreasoning models (specifically OpenAI o1-preview and o1-mini) on their\nrobustness to adversarial attacks. We find that across a variety of attacks,\nincreased inference-time compute leads to improved robustness. In many cases\n(with important exceptions), the fraction of model samples where the attack\nsucceeds tends to zero as the amount of test-time compute grows. We perform no\nadversarial training for the tasks we study, and we increase inference-time\ncompute by simply allowing the models to spend more compute on reasoning,\nindependently of the form of attack. Our results suggest that inference-time\ncompute has the potential to improve adversarial robustness for Large Language\nModels. We also explore new attacks directed at reasoning models, as well as\nsettings where inference-time compute does not improve reliability, and\nspeculate on the reasons for these as well as ways to address them.\n", "date": "2025-01-31", "categories": []}
{"id": "2501.17882", "title": "Heterogeneous Multi-Player Multi-Armed Bandits Robust To Adversarial\n  Attacks", "abstract": "  We consider a multi-player multi-armed bandit setting in the presence of\nadversaries that attempt to negatively affect the rewards received by the\nplayers in the system. The reward distributions for any given arm are\nheterogeneous across the players. In the event of a collision (more than one\nplayer choosing the same arm), all the colliding users receive zero rewards.\nThe adversaries use collisions to affect the rewards received by the players,\ni.e., if an adversary attacks an arm, any player choosing that arm will receive\nzero reward. At any time step, the adversaries may attack more than one arm. It\nis assumed that the players in the system do not deviate from a pre-determined\npolicy used by all the players, and that the probability that none of the arms\nface adversarial attacks is strictly positive at every time step. In order to\ncombat the adversarial attacks, the players are allowed to communicate using a\nsingle bit for $O(\\log T)$ time units, where $T$ is the time horizon, and each\nplayer can only observe their own actions and rewards at all time steps. We\npropose a {policy that is used by all the players, which} achieves near order\noptimal regret of order $O(\\log^{1+\\delta}T + W)$, where $W$ is total number of\ntime units for which there was an adversarial attack on at least one arm.\n", "date": "2025-01-21", "categories": []}
{"id": "2501.18006", "title": "Topological Signatures of Adversaries in Multimodal Alignments", "abstract": "  Multimodal Machine Learning systems, particularly those aligning text and\nimage data like CLIP/BLIP models, have become increasingly prevalent, yet\nremain susceptible to adversarial attacks. While substantial research has\naddressed adversarial robustness in unimodal contexts, defense strategies for\nmultimodal systems are underexplored. This work investigates the topological\nsignatures that arise between image and text embeddings and shows how\nadversarial attacks disrupt their alignment, introducing distinctive\nsignatures. We specifically leverage persistent homology and introduce two\nnovel Topological-Contrastive losses based on Total Persistence and Multi-scale\nkernel methods to analyze the topological signatures introduced by adversarial\nperturbations. We observe a pattern of monotonic changes in the proposed\ntopological losses emerging in a wide range of attacks on image-text\nalignments, as more adversarial samples are introduced in the data. By\ndesigning an algorithm to back-propagate these signatures to input samples, we\nare able to integrate these signatures into Maximum Mean Discrepancy tests,\ncreating a novel class of tests that leverage topological signatures for better\nadversarial detection.\n", "date": "2025-01-29", "categories": []}
{"id": "2501.17433", "title": "Virus: Harmful Fine-tuning Attack for Large Language Models Bypassing\n  Guardrail Moderation", "abstract": "  Recent research shows that Large Language Models (LLMs) are vulnerable to\nharmful fine-tuning attacks -- models lose their safety alignment ability after\nfine-tuning on a few harmful samples. For risk mitigation, a guardrail is\ntypically used to filter out harmful samples before fine-tuning. By designing a\nnew red-teaming method, we in this paper show that purely relying on the\nmoderation guardrail for data filtration is not reliable. Our proposed attack\nmethod, dubbed Virus, easily bypasses the guardrail moderation by slightly\nmodifying the harmful data. Experimental results show that the harmful data\noptimized by Virus is not detectable by the guardrail with up to 100\\% leakage\nratio, and can simultaneously achieve superior attack performance. Finally, the\nkey message we want to convey through this paper is that: \\textbf{it is\nreckless to consider guardrail moderation as a clutch at straws towards harmful\nfine-tuning attack}, as it cannot solve the inherent safety issue of the\npre-trained LLMs. Our code is available at https://github.com/git-disl/Virus\n", "date": "2025-01-29", "categories": []}
{"id": "2501.15563", "title": "PCAP-Backdoor: Backdoor Poisoning Generator for Network Traffic in\n  CPS/IoT Environments", "abstract": "  The rapid expansion of connected devices has made them prime targets for\ncyberattacks. To address these threats, deep learning-based, data-driven\nintrusion detection systems (IDS) have emerged as powerful tools for detecting\nand mitigating such attacks. These IDSs analyze network traffic to identify\nunusual patterns and anomalies that may indicate potential security breaches.\nHowever, prior research has shown that deep learning models are vulnerable to\nbackdoor attacks, where attackers inject triggers into the model to manipulate\nits behavior and cause misclassifications of network traffic. In this paper, we\nexplore the susceptibility of deep learning-based IDS systems to backdoor\nattacks in the context of network traffic analysis. We introduce\n\\texttt{PCAP-Backdoor}, a novel technique that facilitates backdoor poisoning\nattacks on PCAP datasets. Our experiments on real-world Cyber-Physical Systems\n(CPS) and Internet of Things (IoT) network traffic datasets demonstrate that\nattackers can effectively backdoor a model by poisoning as little as 1\\% or\nless of the entire training dataset. Moreover, we show that an attacker can\nintroduce a trigger into benign traffic during model training yet cause the\nbackdoored model to misclassify malicious traffic when the trigger is present.\nFinally, we highlight the difficulty of detecting this trigger-based backdoor,\neven when using existing backdoor defense techniques.\n", "date": "2025-01-26", "categories": []}
{"id": "2501.16497", "title": "Smoothed Embeddings for Robust Language Models", "abstract": "  Improving the safety and reliability of large language models (LLMs) is a\ncrucial aspect of realizing trustworthy AI systems. Although alignment methods\naim to suppress harmful content generation, LLMs are often still vulnerable to\njailbreaking attacks that employ adversarial inputs that subvert alignment and\ninduce harmful outputs. We propose the Randomized Embedding Smoothing and Token\nAggregation (RESTA) defense, which adds random noise to the embedding vectors\nand performs aggregation during the generation of each output token, with the\naim of better preserving semantic information. Our experiments demonstrate that\nour approach achieves superior robustness versus utility tradeoffs compared to\nthe baseline defenses.\n", "date": "2025-01-27", "categories": []}
{"id": "2501.17151", "title": "Scanning Trojaned Models Using Out-of-Distribution Samples", "abstract": "  Scanning for trojan (backdoor) in deep neural networks is crucial due to\ntheir significant real-world applications. There has been an increasing focus\non developing effective general trojan scanning methods across various trojan\nattacks. Despite advancements, there remains a shortage of methods that perform\neffectively without preconceived assumptions about the backdoor attack method.\nAdditionally, we have observed that current methods struggle to identify\nclassifiers trojaned using adversarial training. Motivated by these challenges,\nour study introduces a novel scanning method named TRODO (TROjan scanning by\nDetection of adversarial shifts in Out-of-distribution samples). TRODO\nleverages the concept of \"blind spots\"--regions where trojaned classifiers\nerroneously identify out-of-distribution (OOD) samples as in-distribution (ID).\nWe scan for these blind spots by adversarially shifting OOD samples towards\nin-distribution. The increased likelihood of perturbed OOD samples being\nclassified as ID serves as a signature for trojan detection. TRODO is both\ntrojan and label mapping agnostic, effective even against adversarially trained\ntrojaned classifiers. It is applicable even in scenarios where training data is\nabsent, demonstrating high accuracy and adaptability across various scenarios\nand datasets, highlighting its potential as a robust trojan scanning strategy.\n", "date": "2025-01-28", "categories": []}
{"id": "2501.14999", "title": "VideoPure: Diffusion-based Adversarial Purification for Video\n  Recognition", "abstract": "  Recent work indicates that video recognition models are vulnerable to\nadversarial examples, posing a serious security risk to downstream\napplications. However, current research has primarily focused on adversarial\nattacks, with limited work exploring defense mechanisms. Furthermore, due to\nthe spatial-temporal complexity of videos, existing video defense methods face\nissues of high cost, overfitting, and limited defense performance. Recently,\ndiffusion-based adversarial purification methods have achieved robust defense\nperformance in the image domain. However, due to the additional temporal\ndimension in videos, directly applying these diffusion-based adversarial\npurification methods to the video domain suffers performance and efficiency\ndegradation. To achieve an efficient and effective video adversarial defense\nmethod, we propose the first diffusion-based video purification framework to\nimprove video recognition models' adversarial robustness: VideoPure. Given an\nadversarial example, we first employ temporal DDIM inversion to transform the\ninput distribution into a temporally consistent and trajectory-defined\ndistribution, covering adversarial noise while preserving more video structure.\nThen, during DDIM denoising, we leverage intermediate results at each denoising\nstep and conduct guided spatial-temporal optimization, removing adversarial\nnoise while maintaining temporal consistency. Finally, we input the list of\noptimized intermediate results into the video recognition model for multi-step\nvoting to obtain the predicted class. We investigate the defense performance of\nour method against black-box, gray-box, and adaptive attacks on benchmark\ndatasets and models. Compared with other adversarial purification methods, our\nmethod overall demonstrates better defense performance against different\nattacks. Our code is available at https://github.com/deep-kaixun/VideoPure.\n", "date": "2025-01-25", "categories": []}
{"id": "2501.15269", "title": "Mirage in the Eyes: Hallucination Attack on Multi-modal Large Language\n  Models with Only Attention Sink", "abstract": "  Fusing visual understanding into language generation, Multi-modal Large\nLanguage Models (MLLMs) are revolutionizing visual-language applications. Yet,\nthese models are often plagued by the hallucination problem, which involves\ngenerating inaccurate objects, attributes, and relationships that do not match\nthe visual content. In this work, we delve into the internal attention\nmechanisms of MLLMs to reveal the underlying causes of hallucination, exposing\nthe inherent vulnerabilities in the instruction-tuning process.\n  We propose a novel hallucination attack against MLLMs that exploits attention\nsink behaviors to trigger hallucinated content with minimal image-text\nrelevance, posing a significant threat to critical downstream applications.\nDistinguished from previous adversarial methods that rely on fixed patterns,\nour approach generates dynamic, effective, and highly transferable visual\nadversarial inputs, without sacrificing the quality of model responses.\nComprehensive experiments on 6 prominent MLLMs demonstrate the efficacy of our\nattack in compromising black-box MLLMs even with extensive mitigating\nmechanisms, as well as the promising results against cutting-edge commercial\nAPIs, such as GPT-4o and Gemini 1.5. Our code is available at\nhttps://huggingface.co/RachelHGF/Mirage-in-the-Eyes.\n", "date": "2025-01-25", "categories": []}
{"id": "2501.15271", "title": "Killing it with Zero-Shot: Adversarially Robust Novelty Detection", "abstract": "  Novelty Detection (ND) plays a crucial role in machine learning by\nidentifying new or unseen data during model inference. This capability is\nespecially important for the safe and reliable operation of automated systems.\nDespite advances in this field, existing techniques often fail to maintain\ntheir performance when subject to adversarial attacks. Our research addresses\nthis gap by marrying the merits of nearest-neighbor algorithms with robust\nfeatures obtained from models pretrained on ImageNet. We focus on enhancing the\nrobustness and performance of ND algorithms. Experimental results demonstrate\nthat our approach significantly outperforms current state-of-the-art methods\nacross various benchmarks, particularly under adversarial conditions. By\nincorporating robust pretrained features into the k-NN algorithm, we establish\na new standard for performance and robustness in the field of robust ND. This\nwork opens up new avenues for research aimed at fortifying machine learning\nsystems against adversarial vulnerabilities. Our implementation is publicly\navailable at https://github.com/rohban-lab/ZARND.\n", "date": "2025-01-25", "categories": []}
{"id": "2501.15434", "title": "Mitigating Spurious Negative Pairs for Robust Industrial Anomaly\n  Detection", "abstract": "  Despite significant progress in Anomaly Detection (AD), the robustness of\nexisting detection methods against adversarial attacks remains a challenge,\ncompromising their reliability in critical real-world applications such as\nautonomous driving. This issue primarily arises from the AD setup, which\nassumes that training data is limited to a group of unlabeled normal samples,\nmaking the detectors vulnerable to adversarial anomaly samples during testing.\nAdditionally, implementing adversarial training as a safeguard encounters\ndifficulties, such as formulating an effective objective function without\naccess to labels. An ideal objective function for adversarial training in AD\nshould promote strong perturbations both within and between the normal and\nanomaly groups to maximize margin between normal and anomaly distribution. To\naddress these issues, we first propose crafting a pseudo-anomaly group derived\nfrom normal group samples. Then, we demonstrate that adversarial training with\ncontrastive loss could serve as an ideal objective function, as it creates both\ninter- and intra-group perturbations. However, we notice that spurious negative\npairs compromise the conventional contrastive loss to achieve robust AD.\nSpurious negative pairs are those that should be closely mapped but are\nerroneously separated. These pairs introduce noise and misguide the direction\nof inter-group adversarial perturbations. To overcome the effect of spurious\nnegative pairs, we define opposite pairs and adversarially pull them apart to\nstrengthen inter-group perturbations. Experimental results demonstrate our\nsuperior performance in both clean and adversarial scenarios, with a 26.1%\nimprovement in robust detection across various challenging benchmark datasets.\nThe implementation of our work is available at:\nhttps://github.com/rohban-lab/COBRA.\n", "date": "2025-01-26", "categories": []}
{"id": "2501.14005", "title": "Device-aware Optical Adversarial Attack for a Portable Projector-camera\n  System", "abstract": "  Deep-learning-based face recognition (FR) systems are susceptible to\nadversarial examples in both digital and physical domains. Physical attacks\npresent a greater threat to deployed systems as adversaries can easily access\nthe input channel, allowing them to provide malicious inputs to impersonate a\nvictim. This paper addresses the limitations of existing projector-camera-based\nadversarial light attacks in practical FR setups. By incorporating device-aware\nadaptations into the digital attack algorithm, such as resolution-aware and\ncolor-aware adjustments, we mitigate the degradation from digital to physical\ndomains. Experimental validation showcases the efficacy of our proposed\nalgorithm against real and spoof adversaries, achieving high physical\nsimilarity scores in FR models and state-of-the-art commercial systems. On\naverage, there is only a 14% reduction in scores from digital to physical\nattacks, with high attack success rate in both white- and black-box scenarios.\n", "date": "2025-01-23", "categories": []}
{"id": "2501.14496", "title": "A Note on Implementation Errors in Recent Adaptive Attacks Against\n  Multi-Resolution Self-Ensembles", "abstract": "  This note documents an implementation issue in recent adaptive attacks (Zhang\net al. [2024]) against the multi-resolution self-ensemble defense (Fort and\nLakshminarayanan [2024]). The implementation allowed adversarial perturbations\nto exceed the standard $L_\\infty = 8/255$ bound by up to a factor of\n20$\\times$, reaching magnitudes of up to $L_\\infty = 160/255$. When attacks are\nproperly constrained within the intended bounds, the defense maintains\nnon-trivial robustness. Beyond highlighting the importance of careful\nvalidation in adversarial machine learning research, our analysis reveals an\nintriguing finding: properly bounded adaptive attacks against strong\nmulti-resolution self-ensembles often align with human perception, suggesting\nthe need to reconsider how we measure adversarial robustness.\n", "date": "2025-01-24", "categories": []}
{"id": "2410.04377", "title": "Graded Suspiciousness of Adversarial Texts to Human", "abstract": "  Adversarial examples pose a significant challenge to deep neural networks\n(DNNs) across both image and text domains, with the intent to degrade model\nperformance through meticulously altered inputs. Adversarial texts, however,\nare distinct from adversarial images due to their requirement for semantic\nsimilarity and the discrete nature of the textual contents. This study delves\ninto the concept of human suspiciousness, a quality distinct from the\ntraditional focus on imperceptibility found in image-based adversarial\nexamples. Unlike images, where adversarial changes are meant to be\nindistinguishable to the human eye, textual adversarial content must often\nremain undetected or non-suspicious to human readers, even when the text's\npurpose is to deceive NLP systems or bypass filters.\n  In this research, we expand the study of human suspiciousness by analyzing\nhow individuals perceive adversarial texts. We gather and publish a novel\ndataset of Likert-scale human evaluations on the suspiciousness of adversarial\nsentences, crafted by four widely used adversarial attack methods and assess\ntheir correlation with the human ability to detect machine-generated\nalterations. Additionally, we develop a regression-based model to quantify\nsuspiciousness and establish a baseline for future research in reducing the\nsuspiciousness in adversarial text generation. We also demonstrate how the\nregressor-generated suspicious scores can be incorporated into adversarial\ngeneration methods to produce texts that are less likely to be perceived as\ncomputer-generated. We make our human suspiciousness annotated data and our\ncode available.\n", "date": "2024-10-06", "categories": []}
{"id": "2501.13563", "title": "Black-Box Adversarial Attack on Vision Language Models for Autonomous\n  Driving", "abstract": "  Vision-language models (VLMs) have significantly advanced autonomous driving\n(AD) by enhancing reasoning capabilities; however, these models remain highly\nsusceptible to adversarial attacks. While existing research has explored\nwhite-box attacks to some extent, the more practical and challenging black-box\nscenarios remain largely underexplored due to their inherent difficulty. In\nthis paper, we take the first step toward designing black-box adversarial\nattacks specifically targeting VLMs in AD. We identify two key challenges for\nachieving effective black-box attacks in this context: the effectiveness across\ndriving reasoning chains in AD systems and the dynamic nature of driving\nscenarios. To address this, we propose Cascading Adversarial Disruption (CAD).\nIt first introduces Decision Chain Disruption, which targets low-level\nreasoning breakdown by generating and injecting deceptive semantics, ensuring\nthe perturbations remain effective across the entire decision-making chain.\nBuilding on this, we present Risky Scene Induction, which addresses dynamic\nadaptation by leveraging a surrogate VLM to understand and construct high-level\nrisky scenarios that are likely to result in critical errors in the current\ndriving contexts. Extensive experiments conducted on multiple AD VLMs and\nbenchmarks demonstrate that CAD achieves state-of-the-art attack effectiveness,\nsignificantly outperforming existing methods (+13.43% on average). Moreover, we\nvalidate its practical applicability through real-world attacks on AD vehicles\npowered by VLMs, where the route completion rate drops by 61.11% and the\nvehicle crashes directly into the obstacle vehicle with adversarial patches.\nFinally, we release CADA dataset, comprising 18,808 adversarial\nvisual-question-answer pairs, to facilitate further evaluation and research in\nthis critical domain. Our codes and dataset will be available after paper's\nacceptance.\n", "date": "2025-01-23", "categories": []}
{"id": "2501.13782", "title": "Defending against Adversarial Malware Attacks on ML-based Android\n  Malware Detection Systems", "abstract": "  Android malware presents a persistent threat to users' privacy and data\nintegrity. To combat this, researchers have proposed machine learning-based\n(ML-based) Android malware detection (AMD) systems. However, adversarial\nAndroid malware attacks compromise the detection integrity of the ML-based AMD\nsystems, raising significant concerns. Existing defenses against adversarial\nAndroid malware provide protections against feature space attacks which\ngenerate adversarial feature vectors only, leaving protection against realistic\nthreats from problem space attacks which generate real adversarial malware an\nopen problem. In this paper, we address this gap by proposing ADD, a practical\nadversarial Android malware defense framework designed as a plug-in to enhance\nthe adversarial robustness of the ML-based AMD systems against problem space\nattacks. Our extensive evaluation across various ML-based AMD systems\ndemonstrates that ADD is effective against state-of-the-art problem space\nadversarial Android malware attacks. Additionally, ADD shows the defense\neffectiveness in enhancing the adversarial robustness of real-world antivirus\nsolutions.\n", "date": "2025-01-23", "categories": []}
{"id": "2305.08929", "title": "AF2-Mutation: Adversarial Sequence Mutations against AlphaFold2 on\n  Protein Tertiary Structure Prediction", "abstract": "  Deep learning-based approaches, such as AlphaFold2 (AF2), have significantly\nadvanced protein tertiary structure prediction, achieving results comparable to\nreal biological experimental methods. While AF2 has shown limitations in\npredicting the effects of mutations, its robustness against sequence mutations\nremains to be determined. Starting with the wild-type (WT) sequence, we\ninvestigate adversarial sequences generated via an evolutionary approach, which\nAF2 predicts to be substantially different from WT. Our experiments on CASP14\nreveal that by modifying merely three residues in the protein sequence using a\ncombination of replacement, deletion, and insertion strategies, the alteration\nin AF2's predictions, as measured by the Local Distance Difference Test (lDDT),\nreaches 46.61. Moreover, when applied to a specific protein, SPNS2, our\nproposed algorithm successfully identifies biologically meaningful residues\ncritical to protein structure determination and potentially indicates\nalternative conformations, thus significantly expediting the experimental\nprocess.\n", "date": "2023-05-15", "categories": []}
{"id": "2501.02654", "title": "Tougher Text, Smarter Models: Raising the Bar for Adversarial Defence\n  Benchmarks", "abstract": "  Recent advancements in natural language processing have highlighted the\nvulnerability of deep learning models to adversarial attacks. While various\ndefence mechanisms have been proposed, there is a lack of comprehensive\nbenchmarks that evaluate these defences across diverse datasets, models, and\ntasks. In this work, we address this gap by presenting an extensive benchmark\nfor textual adversarial defence that significantly expands upon previous work.\nOur benchmark incorporates a wide range of datasets, evaluates state-of-the-art\ndefence mechanisms, and extends the assessment to include critical tasks such\nas single-sentence classification, similarity and paraphrase identification,\nnatural language inference, and commonsense reasoning. This work not only\nserves as a valuable resource for researchers and practitioners in the field of\nadversarial robustness but also identifies key areas for future research in\ntextual adversarial defence. By establishing a new standard for benchmarking in\nthis domain, we aim to accelerate progress towards more robust and reliable\nnatural language processing systems.\n", "date": "2025-01-05", "categories": []}
{"id": "2501.11815", "title": "CogMorph: Cognitive Morphing Attacks for Text-to-Image Models", "abstract": "  The development of text-to-image (T2I) generative models, that enable the\ncreation of high-quality synthetic images from textual prompts, has opened new\nfrontiers in creative design and content generation. However, this paper\nreveals a significant and previously unrecognized ethical risk inherent in this\ntechnology and introduces a novel method, termed the Cognitive Morphing Attack\n(CogMorph), which manipulates T2I models to generate images that retain the\noriginal core subjects but embeds toxic or harmful contextual elements. This\nnuanced manipulation exploits the cognitive principle that human perception of\nconcepts is shaped by the entire visual scene and its context, producing images\nthat amplify emotional harm far beyond attacks that merely preserve the\noriginal semantics. To address this, we first construct an imagery toxicity\ntaxonomy spanning 10 major and 48 sub-categories, aligned with human\ncognitive-perceptual dimensions, and further build a toxicity risk matrix\nresulting in 1,176 high-quality T2I toxic prompts. Based on this, our CogMorph\nfirst introduces Cognitive Toxicity Augmentation, which develops a cognitive\ntoxicity knowledge base with rich external toxic representations for humans\n(e.g., fine-grained visual features) that can be utilized to further guide the\noptimization of adversarial prompts. In addition, we present Contextual\nHierarchical Morphing, which hierarchically extracts critical parts of the\noriginal prompt (e.g., scenes, subjects, and body parts), and then iteratively\nretrieves and fuses toxic features to inject harmful contexts. Extensive\nexperiments on multiple open-sourced T2I models and black-box commercial APIs\n(e.g., DALLE-3) demonstrate the efficacy of CogMorph which significantly\noutperforms other baselines by large margins (+20.62% on average).\n", "date": "2025-01-21", "categories": []}
{"id": "2501.12516", "title": "Robustness of Selected Learning Models under Label-Flipping Attack", "abstract": "  In this paper we compare traditional machine learning and deep learning\nmodels trained on a malware dataset when subjected to adversarial attack based\non label-flipping. Specifically, we investigate the robustness of Support\nVector Machines (SVM), Random Forest, Gaussian Naive Bayes (GNB), Gradient\nBoosting Machine (GBM), LightGBM, XGBoost, Multilayer Perceptron (MLP),\nConvolutional Neural Network (CNN), MobileNet, and DenseNet models when facing\nvarying percentages of misleading labels. We empirically assess the the\naccuracy of each of these models under such an adversarial attack on the\ntraining data. This research aims to provide insights into which models are\ninherently more robust, in the sense of being better able to resist intentional\ndisruptions to the training data. We find wide variation in the robustness of\nthe models tested to adversarial attack, with our MLP model achieving the best\ncombination of initial accuracy and robustness.\n", "date": "2025-01-21", "categories": []}
{"id": "2406.11239", "title": "SilverSpeak: Evading AI-Generated Text Detectors using Homoglyphs", "abstract": "  The advent of Large Language Models (LLMs) has enabled the generation of text\nthat increasingly exhibits human-like characteristics. As the detection of such\ncontent is of significant importance, substantial research has been conducted\nwith the objective of developing reliable AI-generated text detectors. These\ndetectors have demonstrated promising results on test data, but recent research\nhas revealed that they can be circumvented by employing different techniques.\n  In this paper, we present homoglyph-based attacks (A $\\rightarrow$ Cyrillic\nA) as a means of circumventing existing detectors. We conduct a comprehensive\nevaluation to assess the effectiveness of these attacks on seven detectors,\nincluding ArguGPT, Binoculars, DetectGPT, Fast-DetectGPT, Ghostbuster, OpenAI's\ndetector, and watermarking techniques, on five different datasets. Our findings\ndemonstrate that homoglyph-based attacks can effectively circumvent\nstate-of-the-art detectors, leading them to classify all texts as either\nAI-generated or human-written (decreasing the average Matthews Correlation\nCoefficient from 0.64 to -0.01). Through further examination, we extract the\ntechnical justification underlying the success of the attacks, which varies\nacross detectors. Finally, we discuss the implications of these findings and\npotential defenses against such attacks.\n", "date": "2024-06-17", "categories": []}
{"id": "2406.17830", "title": "Treatment of Statistical Estimation Problems in Randomized Smoothing for\n  Adversarial Robustness", "abstract": "  Randomized smoothing is a popular certified defense against adversarial\nattacks. In its essence, we need to solve a problem of statistical estimation\nwhich is usually very time-consuming since we need to perform numerous (usually\n$10^5$) forward passes of the classifier for every point to be certified. In\nthis paper, we review the statistical estimation problems for randomized\nsmoothing to find out if the computational burden is necessary. In particular,\nwe consider the (standard) task of adversarial robustness where we need to\ndecide if a point is robust at a certain radius or not using as few samples as\npossible while maintaining statistical guarantees. We present estimation\nprocedures employing confidence sequences enjoying the same statistical\nguarantees as the standard methods, with the optimal sample complexities for\nthe estimation task and empirically demonstrate their good performance.\nAdditionally, we provide a randomized version of Clopper-Pearson confidence\nintervals resulting in strictly stronger certificates.\n", "date": "2024-06-25", "categories": []}
{"id": "2410.04884", "title": "Patch is Enough: Naturalistic Adversarial Patch against Vision-Language\n  Pre-training Models", "abstract": "  Visual language pre-training (VLP) models have demonstrated significant\nsuccess across various domains, yet they remain vulnerable to adversarial\nattacks. Addressing these adversarial vulnerabilities is crucial for enhancing\nsecurity in multimodal learning. Traditionally, adversarial methods targeting\nVLP models involve simultaneously perturbing images and text. However, this\napproach faces notable challenges: first, adversarial perturbations often fail\nto translate effectively into real-world scenarios; second, direct\nmodifications to the text are conspicuously visible. To overcome these\nlimitations, we propose a novel strategy that exclusively employs image patches\nfor attacks, thus preserving the integrity of the original text. Our method\nleverages prior knowledge from diffusion models to enhance the authenticity and\nnaturalness of the perturbations. Moreover, to optimize patch placement and\nimprove the efficacy of our attacks, we utilize the cross-attention mechanism,\nwhich encapsulates intermodal interactions by generating attention maps to\nguide strategic patch placements. Comprehensive experiments conducted in a\nwhite-box setting for image-to-text scenarios reveal that our proposed method\nsignificantly outperforms existing techniques, achieving a 100% attack success\nrate. Additionally, it demonstrates commendable performance in transfer tasks\ninvolving text-to-image configurations.\n", "date": "2024-10-07", "categories": []}
{"id": "2410.11317", "title": "Deciphering the Chaos: Enhancing Jailbreak Attacks via Adversarial\n  Prompt Translation", "abstract": "  Automatic adversarial prompt generation provides remarkable success in\njailbreaking safely-aligned large language models (LLMs). Existing\ngradient-based attacks, while demonstrating outstanding performance in\njailbreaking white-box LLMs, often generate garbled adversarial prompts with\nchaotic appearance. These adversarial prompts are difficult to transfer to\nother LLMs, hindering their performance in attacking unknown victim models. In\nthis paper, for the first time, we delve into the semantic meaning embedded in\ngarbled adversarial prompts and propose a novel method that \"translates\" them\ninto coherent and human-readable natural language adversarial prompts. In this\nway, we can effectively uncover the semantic information that triggers\nvulnerabilities of the model and unambiguously transfer it to the victim model,\nwithout overlooking the adversarial information hidden in the garbled text, to\nenhance jailbreak attacks. It also offers a new approach to discovering\neffective designs for jailbreak prompts, advancing the understanding of\njailbreak attacks. Experimental results demonstrate that our method\nsignificantly improves the success rate of jailbreak attacks against various\nsafety-aligned LLMs and outperforms state-of-the-arts by large margins. With at\nmost 10 queries, our method achieves an average attack success rate of 81.8% in\nattacking 7 commercial closed-source LLMs, including GPT and Claude-3 series,\non HarmBench. Our method also achieves over 90% attack success rates against\nLlama-2-Chat models on AdvBench, despite their outstanding resistance to\njailbreak attacks. Code at:\nhttps://github.com/qizhangli/Adversarial-Prompt-Translator.\n", "date": "2024-10-15", "categories": []}
{"id": "2410.23677", "title": "Wide Two-Layer Networks can Learn from Adversarial Perturbations", "abstract": "  Adversarial examples have raised several open questions, such as why they can\ndeceive classifiers and transfer between different models. A prevailing\nhypothesis to explain these phenomena suggests that adversarial perturbations\nappear as random noise but contain class-specific features. This hypothesis is\nsupported by the success of perturbation learning, where classifiers trained\nsolely on adversarial examples and the corresponding incorrect labels\ngeneralize well to correctly labeled test data. Although this hypothesis and\nperturbation learning are effective in explaining intriguing properties of\nadversarial examples, their solid theoretical foundation is limited. In this\nstudy, we theoretically explain the counterintuitive success of perturbation\nlearning. We assume wide two-layer networks and the results hold for any data\ndistribution. We prove that adversarial perturbations contain sufficient\nclass-specific features for networks to generalize from them. Moreover, the\npredictions of classifiers trained on mislabeled adversarial examples coincide\nwith those of classifiers trained on correctly labeled clean samples. The code\nis available at https://github.com/s-kumano/perturbation-learning.\n", "date": "2024-10-31", "categories": []}
{"id": "2412.09844", "title": "Real-time Identity Defenses against Malicious Personalization of\n  Diffusion Models", "abstract": "  Personalized generative diffusion models, capable of synthesizing highly\nrealistic images based on a few reference portraits, may pose substantial\nsocial, ethical, and legal risks via identity replication. Existing defense\nmechanisms rely on computationally intensive adversarial perturbations tailored\nto individual images, rendering them impractical for real-world deployment.\nThis study introduces the Real-time Identity Defender (RID), a neural network\ndesigned to generate adversarial perturbations through a single forward pass,\nbypassing the need for image-specific optimization. RID achieves unprecedented\nefficiency, with defense times as low as 0.12 seconds on a single NVIDIA A100\n80G GPU (4,400 times faster than leading methods) and 1.1 seconds per image on\na standard Intel i9 CPU, making it suitable for edge devices such as\nsmartphones. Despite its efficiency, RID achieves promising protection\nperformance across visual and quantitative benchmarks, effectively mitigating\nidentity replication risks. Our analysis reveals that RID's perturbations mimic\nthe efficacy of traditional defenses while exhibiting properties distinct from\nnatural noise, such as Gaussian perturbations. To enhance robustness, we extend\nRID into an ensemble framework that integrates multiple pre-trained\ntext-to-image diffusion models, ensuring resilience against black-box attacks\nand post-processing techniques, including image compression and purification.\nOur model is envisioned to play a crucial role in safeguarding portrait rights,\nthereby preventing illegal and unethical uses.\n", "date": "2024-12-13", "categories": []}
{"id": "2501.02147", "title": "Exploring Secure Machine Learning Through Payload Injection and FGSM\n  Attacks on ResNet-50", "abstract": "  This paper investigates the resilience of a ResNet-50 image classification\nmodel under two prominent security threats: Fast Gradient Sign Method (FGSM)\nadversarial attacks and malicious payload injection. Initially, the model\nattains a 53.33% accuracy on clean images. When subjected to FGSM\nperturbations, its overall accuracy remains unchanged; however, the model's\nconfidence in incorrect predictions notably increases. Concurrently, a payload\ninjection scheme is successfully executed in 93.33% of the tested samples,\nrevealing how stealthy attacks can manipulate model predictions without\ndegrading visual quality. These findings underscore the vulnerability of even\nhigh-performing neural networks and highlight the urgency of developing more\nrobust defense mechanisms for security-critical applications.\n", "date": "2025-01-04", "categories": []}
{"id": "2501.10606", "title": "Differentiable Adversarial Attacks for Marked Temporal Point Processes", "abstract": "  Marked temporal point processes (MTPPs) have been shown to be extremely\neffective in modeling continuous time event sequences (CTESs). In this work, we\npresent adversarial attacks designed specifically for MTPP models. A key\ncriterion for a good adversarial attack is its imperceptibility. For objects\nsuch as images or text, this is often achieved by bounding perturbation in some\nfixed $L_p$ norm-ball. However, similarly minimizing distance norms between two\nCTESs in the context of MTPPs is challenging due to their sequential nature and\nvarying time-scales and lengths. We address this challenge by first permuting\nthe events and then incorporating the additive noise to the arrival timestamps.\nHowever, the worst case optimization of such adversarial attacks is a hard\ncombinatorial problem, requiring exploration across a permutation space that is\nfactorially large in the length of the input sequence. As a result, we propose\na novel differentiable scheme PERMTPP using which we can perform adversarial\nattacks by learning to minimize the likelihood, while minimizing the distance\nbetween two CTESs. Our experiments on four real-world datasets demonstrate the\noffensive and defensive capabilities, and lower inference times of PERMTPP.\n", "date": "2025-01-17", "categories": []}
{"id": "2501.11054", "title": "Temporal Analysis of Adversarial Attacks in Federated Learning", "abstract": "  In this paper, we experimentally analyze the robustness of selected Federated\nLearning (FL) systems in the presence of adversarial clients. We find that\ntemporal attacks significantly affect model performance in the FL models\ntested, especially when the adversaries are active throughout or during the\nlater rounds. We consider a variety of classic learning models, including\nMultinominal Logistic Regression (MLR), Random Forest, XGBoost, Support Vector\nClassifier (SVC), as well as various Neural Network models including Multilayer\nPerceptron (MLP), Convolution Neural Network (CNN), Recurrent Neural Network\n(RNN), and Long Short-Term Memory (LSTM). Our results highlight the\neffectiveness of temporal attacks and the need to develop strategies to make\nthe FL process more robust against such attacks. We also briefly consider the\neffectiveness of defense mechanisms, including outlier detection in the\naggregation algorithm.\n", "date": "2025-01-19", "categories": []}
{"id": "2501.11795", "title": "Provably effective detection of effective data poisoning attacks", "abstract": "  This paper establishes a mathematically precise definition of dataset\npoisoning attack and proves that the very act of effectively poisoning a\ndataset ensures that the attack can be effectively detected. On top of a\nmathematical guarantee that dataset poisoning is identifiable by a new\nstatistical test that we call the Conformal Separability Test, we provide\nexperimental evidence that we can adequately detect poisoning attempts in the\nreal world.\n", "date": "2025-01-21", "categories": []}
{"id": "2501.11852", "title": "Cross-Entropy Attacks to Language Models via Rare Event Simulation", "abstract": "  Black-box textual adversarial attacks are challenging due to the lack of\nmodel information and the discrete, non-differentiable nature of text. Existing\nmethods often lack versatility for attacking different models, suffer from\nlimited attacking performance due to the inefficient optimization with word\nsaliency ranking, and frequently sacrifice semantic integrity to achieve better\nattack outcomes. This paper introduces a novel approach to textual adversarial\nattacks, which we call Cross-Entropy Attacks (CEA), that uses Cross-Entropy\noptimization to address the above issues. Our CEA approach defines adversarial\nobjectives for both soft-label and hard-label settings and employs CE\noptimization to identify optimal replacements. Through extensive experiments on\ndocument classification and language translation problems, we demonstrate that\nour attack method excels in terms of attacking performance, imperceptibility,\nand sentence quality.\n", "date": "2025-01-21", "categories": []}
{"id": "2501.12275", "title": "With Great Backbones Comes Great Adversarial Transferability", "abstract": "  Advances in self-supervised learning (SSL) for machine vision have improved\nrepresentation robustness and model performance, giving rise to pre-trained\nbackbones like \\emph{ResNet} and \\emph{ViT} models tuned with SSL methods such\nas \\emph{SimCLR}. Due to the computational and data demands of pre-training,\nthe utilization of such backbones becomes a strenuous necessity. However,\nemploying these backbones may inherit vulnerabilities to adversarial attacks.\nWhile adversarial robustness has been studied under \\emph{white-box} and\n\\emph{black-box} settings, the robustness of models tuned on pre-trained\nbackbones remains largely unexplored. Additionally, the role of tuning\nmeta-information in mitigating exploitation risks is unclear. This work\nsystematically evaluates the adversarial robustness of such models across\n$20,000$ combinations of tuning meta-information, including fine-tuning\ntechniques, backbone families, datasets, and attack types. We propose using\nproxy models to transfer attacks, simulating varying levels of target knowledge\nby fine-tuning these proxies with diverse configurations. Our findings reveal\nthat proxy-based attacks approach the effectiveness of \\emph{white-box}\nmethods, even with minimal tuning knowledge. We also introduce a naive\n\"backbone attack,\" leveraging only the backbone to generate adversarial\nsamples, which outperforms \\emph{black-box} attacks and rivals \\emph{white-box}\nmethods, highlighting critical risks in model-sharing practices. Finally, our\nablations reveal how increasing tuning meta-information impacts attack\ntransferability, measuring each meta-information combination.\n", "date": "2025-01-21", "categories": []}
{"id": "2408.00348", "title": "Securing the Diagnosis of Medical Imaging: An In-depth Analysis of\n  AI-Resistant Attacks", "abstract": "  Machine learning (ML) is a rapidly developing area of medicine that uses\nsignificant resources to apply computer science and statistics to medical\nissues. ML's proponents laud its capacity to handle vast, complicated, and\nerratic medical data. It's common knowledge that attackers might cause\nmisclassification by deliberately creating inputs for machine learning\nclassifiers. Research on adversarial examples has been extensively conducted in\nthe field of computer vision applications. Healthcare systems are thought to be\nhighly difficult because of the security and life-or-death considerations they\ninclude, and performance accuracy is very important. Recent arguments have\nsuggested that adversarial attacks could be made against medical image analysis\n(MedIA) technologies because of the accompanying technology infrastructure and\npowerful financial incentives. Since the diagnosis will be the basis for\nimportant decisions, it is essential to assess how strong medical DNN tasks are\nagainst adversarial attacks. Simple adversarial attacks have been taken into\naccount in several earlier studies. However, DNNs are susceptible to more risky\nand realistic attacks. The present paper covers recent proposed adversarial\nattack strategies against DNNs for medical imaging as well as countermeasures.\nIn this study, we review current techniques for adversarial imaging attacks,\ndetections. It also encompasses various facets of these techniques and offers\nsuggestions for the robustness of neural networks to be improved in the future.\n", "date": "2024-08-01", "categories": []}
{"id": "2501.08258", "title": "Towards an End-to-End (E2E) Adversarial Learning and Application in the\n  Physical World", "abstract": "  The traditional learning process of patch-based adversarial attacks,\nconducted in the digital domain and then applied in the physical domain (e.g.,\nvia printed stickers), may suffer from reduced performance due to adversarial\npatches' limited transferability from the digital domain to the physical\ndomain. Given that previous studies have considered using projectors to apply\nadversarial attacks, we raise the following question: can adversarial learning\n(i.e., patch generation) be performed entirely in the physical domain with a\nprojector? In this work, we propose the Physical-domain Adversarial Patch\nLearning Augmentation (PAPLA) framework, a novel end-to-end (E2E) framework\nthat converts adversarial learning from the digital domain to the physical\ndomain using a projector. We evaluate PAPLA across multiple scenarios,\nincluding controlled laboratory settings and realistic outdoor environments,\ndemonstrating its ability to ensure attack success compared to conventional\ndigital learning-physical application (DL-PA) methods. We also analyze the\nimpact of environmental factors, such as projection surface color, projector\nstrength, ambient light, distance, and angle of the target object relative to\nthe camera, on the effectiveness of projected patches. Finally, we demonstrate\nthe feasibility of the attack against a parked car and a stop sign in a\nreal-world outdoor environment. Our results show that under specific\nconditions, E2E adversarial learning in the physical domain eliminates the\ntransferability issue and ensures evasion by object detectors. Finally, we\nprovide insights into the challenges and opportunities of applying adversarial\nlearning in the physical domain and explain where such an approach is more\neffective than using a sticker.\n", "date": "2025-01-14", "categories": []}
{"id": "2501.09086", "title": "Salient Information Preserving Adversarial Training Improves Clean and\n  Robust Accuracy", "abstract": "  In this work we introduce Salient Information Preserving Adversarial Training\n(SIP-AT), an intuitive method for relieving the robustness-accuracy trade-off\nincurred by traditional adversarial training. SIP-AT uses salient image regions\nto guide the adversarial training process in such a way that fragile features\ndeemed meaningful by an annotator remain unperturbed during training, allowing\nmodels to learn highly predictive non-robust features without sacrificing\noverall robustness. This technique is compatible with both human-based and\nautomatically generated salience estimates, allowing SIP-AT to be used as a\npart of human-driven model development without forcing SIP-AT to be reliant\nupon additional human data. We perform experiments across multiple datasets and\narchitectures and demonstrate that SIP-AT is able to boost the clean accuracy\nof models while maintaining a high degree of robustness against attacks at\nmultiple epsilon levels. We complement our central experiments with an\nobservational study measuring the rate at which human subjects successfully\nidentify perturbed images. This study helps build a more intuitive\nunderstanding of adversarial attack strength and demonstrates the heightened\nimportance of low-epsilon robustness. Our results demonstrate the efficacy of\nSIP-AT and provide valuable insight into the risks posed by adversarial samples\nof various strengths.\n", "date": "2025-01-15", "categories": []}
{"id": "2501.09609", "title": "Adversarial-Ensemble Kolmogorov Arnold Networks for Enhancing Indoor\n  Wi-Fi Positioning: A Defensive Approach Against Spoofing and Signal\n  Manipulation Attacks", "abstract": "  The research presents a study on enhancing the robustness of Wi-Fi-based\nindoor positioning systems against adversarial attacks. The goal is to improve\nthe positioning accuracy and resilience of these systems under two attack\nscenarios: Wi-Fi Spoofing and Signal Strength Manipulation. Three models are\ndeveloped and evaluated: a baseline model (M_Base), an adversarially trained\nrobust model (M_Rob), and an ensemble model (M_Ens). All models utilize a\nKolmogorov-Arnold Network (KAN) architecture. The robust model is trained with\nadversarially perturbed data, while the ensemble model combines predictions\nfrom both the base and robust models. Experimental results show that the robust\nmodel reduces positioning error by approximately 10% compared to the baseline,\nachieving 2.03 meters error under Wi-Fi spoofing and 2.00 meters under signal\nstrength manipulation. The ensemble model further outperforms with errors of\n2.01 meters and 1.975 meters for the respective attack types. This analysis\nhighlights the effectiveness of adversarial training techniques in mitigating\nattack impacts. The findings underscore the importance of considering\nadversarial scenarios in developing indoor positioning systems, as improved\nresilience can significantly enhance the accuracy and reliability of such\nsystems in mission-critical environments.\n", "date": "2025-01-16", "categories": []}
{"id": "2501.09006", "title": "Improving Stability Estimates in Adversarial Explainable AI through\n  Alternate Search Methods", "abstract": "  Advances in the effectiveness of machine learning models have come at the\ncost of enormous complexity resulting in a poor understanding of how they\nfunction. Local surrogate methods have been used to approximate the workings of\nthese complex models, but recent work has revealed their vulnerability to\nadversarial attacks where the explanation produced is appreciably different\nwhile the meaning and structure of the complex model's output remains similar.\nThis prior work has focused on the existence of these weaknesses but not on\ntheir magnitude. Here we explore using an alternate search method with the goal\nof finding minimum viable perturbations, the fewest perturbations necessary to\nachieve a fixed similarity value between the original and altered text's\nexplanation. Intuitively, a method that requires fewer perturbations to expose\na given level of instability is inferior to one which requires more. This\nnuance allows for superior comparisons of the stability of explainability\nmethods.\n", "date": "2025-01-15", "categories": []}
{"id": "2310.04539", "title": "Generating Less Certain Adversarial Examples Improves Robust\n  Generalization", "abstract": "  This paper revisits the robust overfitting phenomenon of adversarial\ntraining. Observing that models with better robust generalization performance\nare less certain in predicting adversarially generated training inputs, we\nargue that overconfidence in predicting adversarial examples is a potential\ncause. Therefore, we hypothesize that generating less certain adversarial\nexamples improves robust generalization, and propose a formal definition of\nadversarial certainty that captures the variance of the model's predicted\nlogits on adversarial examples. Our theoretical analysis of synthetic\ndistributions characterizes the connection between adversarial certainty and\nrobust generalization. Accordingly, built upon the notion of adversarial\ncertainty, we develop a general method to search for models that can generate\ntraining-time adversarial inputs with reduced certainty, while maintaining the\nmodel's capability in distinguishing adversarial examples. Extensive\nexperiments on image benchmarks demonstrate that our method effectively learns\nmodels with consistently improved robustness and mitigates robust overfitting,\nconfirming the importance of generating less certain adversarial examples for\nrobust generalization. Our implementations are available as open-source code\nat: https://github.com/TrustMLRG/AdvCertainty.\n", "date": "2023-10-06", "categories": []}
{"id": "2406.12843", "title": "Can Go AIs be adversarially robust?", "abstract": "  Prior work found that superhuman Go AIs can be defeated by simple adversarial\nstrategies, especially \"cyclic\" attacks. In this paper, we study whether adding\nnatural countermeasures can achieve robustness in Go, a favorable domain for\nrobustness since it benefits from incredible average-case capability and a\nnarrow, innately adversarial setting. We test three defenses: adversarial\ntraining on hand-constructed positions, iterated adversarial training, and\nchanging the network architecture. We find that though some of these defenses\nprotect against previously discovered attacks, none withstand freshly trained\nadversaries. Furthermore, most of the reliably effective attacks these\nadversaries discover are different realizations of the same overall class of\ncyclic attacks. Our results suggest that building robust AI systems is\nchallenging even with extremely superhuman systems in some of the most\ntractable settings, and highlight two key gaps: efficient generalization of\ndefenses, and diversity in training. For interactive examples of attacks and a\nlink to our codebase, see https://goattack.far.ai.\n", "date": "2024-06-18", "categories": []}
{"id": "2501.07922", "title": "VENOM: Text-driven Unrestricted Adversarial Example Generation with\n  Diffusion Models", "abstract": "  Adversarial attacks have proven effective in deceiving machine learning\nmodels by subtly altering input images, motivating extensive research in recent\nyears. Traditional methods constrain perturbations within $l_p$-norm bounds,\nbut advancements in Unrestricted Adversarial Examples (UAEs) allow for more\ncomplex, generative-model-based manipulations. Diffusion models now lead UAE\ngeneration due to superior stability and image quality over GANs. However,\nexisting diffusion-based UAE methods are limited to using reference images and\nface challenges in generating Natural Adversarial Examples (NAEs) directly from\nrandom noise, often producing uncontrolled or distorted outputs. In this work,\nwe introduce VENOM, the first text-driven framework for high-quality\nunrestricted adversarial examples generation through diffusion models. VENOM\nunifies image content generation and adversarial synthesis into a single\nreverse diffusion process, enabling high-fidelity adversarial examples without\nsacrificing attack success rate (ASR). To stabilize this process, we\nincorporate an adaptive adversarial guidance strategy with momentum, ensuring\nthat the generated adversarial examples $x^*$ align with the distribution\n$p(x)$ of natural images. Extensive experiments demonstrate that VENOM achieves\nsuperior ASR and image quality compared to prior methods, marking a significant\nadvancement in adversarial example generation and providing insights into model\nvulnerabilities for improved defense development.\n", "date": "2025-01-14", "categories": []}
{"id": "2211.10227", "title": "Adversarial Detection by Approximation of Ensemble Boundary", "abstract": "  Despite being effective in many application areas, Deep Neural Networks\n(DNNs) are vulnerable to being attacked. In object recognition, the attack\ntakes the form of a small perturbation added to an image, that causes the DNN\nto misclassify, but to a human appears no different. Adversarial attacks lead\nto defences that are themselves subject to attack, and the attack/ defence\nstrategies provide important information about the properties of DNNs. In this\npaper, a novel method of detecting adversarial attacks is proposed for an\nensemble of Deep Neural Networks (DNNs) solving two-class pattern recognition\nproblems. The ensemble is combined using Walsh coefficients which are capable\nof approximating Boolean functions and thereby controlling the decision\nboundary complexity. The hypothesis in this paper is that decision boundaries\nwith high curvature allow adversarial perturbations to be found, but change the\ncurvature of the decision boundary, which is then approximated in a different\nway by Walsh coefficients compared to the clean images. Besides controlling\nboundary complexity, the coefficients also measure the correlation with class\nlabels, which may aid in understanding the learning and transferability\nproperties of DNNs. While the experiments here use images, the proposed\napproach of modelling two-class ensemble decision boundaries could in principle\nbe applied to any application area.\n", "date": "2022-11-18", "categories": []}
{"id": "2407.04016", "title": "Mitigating Low-Frequency Bias: Feature Recalibration and Frequency\n  Attention Regularization for Adversarial Robustness", "abstract": "  Ensuring the robustness of deep neural networks against adversarial attacks\nremains a fundamental challenge in computer vision. While adversarial training\n(AT) has emerged as a promising defense strategy, our analysis reveals a\ncritical limitation: AT-trained models exhibit a bias toward low-frequency\nfeatures while neglecting high-frequency components. This bias is particularly\nconcerning as each frequency component carries distinct and crucial\ninformation: low-frequency features encode fundamental structural patterns,\nwhile high-frequency features capture intricate details and textures. To\naddress this limitation, we propose High-Frequency Feature Disentanglement and\nRecalibration (HFDR), a novel module that strategically separates and\nrecalibrates frequency-specific features to capture latent semantic cues. We\nfurther introduce frequency attention regularization to harmonize feature\nextraction across the frequency spectrum and mitigate the inherent\nlow-frequency bias of AT. Extensive experiments demonstrate our method's\nsuperior performance against white-box attacks and transfer attacks, while\nexhibiting strong generalization capabilities across diverse scenarios.\n", "date": "2024-07-04", "categories": []}
{"id": "2412.07249", "title": "Buster: Implanting Semantic Backdoor into Text Encoder to Mitigate NSFW\n  Content Generation", "abstract": "  The rise of deep learning models in the digital era has raised substantial\nconcerns regarding the generation of Not-Safe-for-Work (NSFW) content. Existing\ndefense methods primarily involve model fine-tuning and post-hoc content\nmoderation. Nevertheless, these approaches largely lack scalability in\neliminating harmful content, degrade the quality of benign image generation, or\nincur high inference costs. To address these challenges, we propose an\ninnovative framework named \\textit{Buster}, which injects backdoors into the\ntext encoder to prevent NSFW content generation. Buster leverages deep semantic\ninformation rather than explicit prompts as triggers, redirecting NSFW prompts\ntowards targeted benign prompts. Additionally, Buster employs energy-based\ntraining data generation through Langevin dynamics for adversarial knowledge\naugmentation, thereby ensuring robustness in harmful concept definition. This\napproach demonstrates exceptional resilience and scalability in mitigating NSFW\ncontent. Particularly, Buster fine-tunes the text encoder of Text-to-Image\nmodels within merely five minutes, showcasing its efficiency. Our extensive\nexperiments denote that Buster outperforms nine state-of-the-art baselines,\nachieving a superior NSFW content removal rate of at least 91.2\\% while\npreserving the quality of harmless images.\n", "date": "2024-12-10", "categories": []}
{"id": "2501.01025", "title": "Towards Adversarially Robust Deep Metric Learning", "abstract": "  Deep Metric Learning (DML) has shown remarkable successes in many domains by\ntaking advantage of powerful deep neural networks. Deep neural networks are\nprone to adversarial attacks and could be easily fooled by adversarial\nexamples. The current progress on this robustness issue is mainly about deep\nclassification models but pays little attention to DML models. Existing works\nfail to thoroughly inspect the robustness of DML and neglect an important DML\nscenario, the clustering-based inference. In this work, we first point out the\nrobustness issue of DML models in clustering-based inference scenarios. We find\nthat, for the clustering-based inference, existing defenses designed DML are\nunable to be reused and the adaptions of defenses designed for deep\nclassification models cannot achieve satisfactory robustness performance. To\nalleviate the hazard of adversarial examples, we propose a new defense, the\nEnsemble Adversarial Training (EAT), which exploits ensemble learning and\nadversarial training. EAT promotes the diversity of the ensemble, encouraging\neach model in the ensemble to have different robustness features, and employs a\nself-transferring mechanism to make full use of the robustness statistics of\nthe whole ensemble in the update of every single model. We evaluate the EAT\nmethod on three widely-used datasets with two popular model architectures. The\nresults show that the proposed EAT method greatly outperforms the adaptions of\ndefenses designed for deep classification models.\n", "date": "2025-01-02", "categories": []}
{"id": "2501.07044", "title": "Protego: Detecting Adversarial Examples for Vision Transformers via\n  Intrinsic Capabilities", "abstract": "  Transformer models have excelled in natural language tasks, prompting the\nvision community to explore their implementation in computer vision problems.\nHowever, these models are still influenced by adversarial examples. In this\npaper, we investigate the attack capabilities of six common adversarial attacks\non three pretrained ViT models to reveal the vulnerability of ViT models. To\nunderstand and analyse the bias in neural network decisions when the input is\nadversarial, we use two visualisation techniques that are attention rollout and\ngrad attention rollout. To prevent ViT models from adversarial attack, we\npropose Protego, a detection framework that leverages the transformer intrinsic\ncapabilities to detection adversarial examples of ViT models. Nonetheless, this\nis challenging due to a diversity of attack strategies that may be adopted by\nadversaries. Inspired by the attention mechanism, we know that the token of\nprediction contains all the information from the input sample. Additionally,\nthe attention region for adversarial examples differs from that of normal\nexamples. Given these points, we can train a detector that achieves superior\nperformance than existing detection methods to identify adversarial examples.\nOur experiments have demonstrated the high effectiveness of our detection\nmethod. For these six adversarial attack methods, our detector's AUC scores all\nexceed 0.95. Protego may advance investigations in metaverse security.\n", "date": "2025-01-13", "categories": []}
{"id": "2501.07192", "title": "A4O: All Trigger for One sample", "abstract": "  Backdoor attacks have become a critical threat to deep neural networks\n(DNNs), drawing many research interests. However, most of the studied attacks\nemploy a single type of trigger. Consequently, proposed backdoor defenders\noften rely on the assumption that triggers would appear in a unified way. In\nthis paper, we show that this naive assumption can create a loophole, allowing\nmore sophisticated backdoor attacks to bypass. We design a novel backdoor\nattack mechanism that incorporates multiple types of backdoor triggers,\nfocusing on stealthiness and effectiveness. Our journey begins with the\nintriguing observation that the performance of a backdoor attack in deep\nlearning models, as well as its detectability and removability, are all\nproportional to the magnitude of the trigger. Based on this correlation, we\npropose reducing the magnitude of each trigger type and combining them to\nachieve a strong backdoor relying on the combined trigger while still staying\nsafely under the radar of defenders. Extensive experiments on three standard\ndatasets demonstrate that our method can achieve high attack success rates\n(ASRs) while consistently bypassing state-of-the-art defenses.\n", "date": "2025-01-13", "categories": []}
{"id": "2501.07275", "title": "Generating Poisoning Attacks against Ridge Regression Models with\n  Categorical Features", "abstract": "  Machine Learning (ML) models have become a very powerful tool to extract\ninformation from large datasets and use it to make accurate predictions and\nautomated decisions. However, ML models can be vulnerable to external attacks,\ncausing them to underperform or deviate from their expected tasks. One way to\nattack ML models is by injecting malicious data to mislead the algorithm during\nthe training phase, which is referred to as a poisoning attack. We can prepare\nfor such situations by designing anticipated attacks, which are later used for\ncreating and testing defence strategies. In this paper, we propose an algorithm\nto generate strong poisoning attacks for a ridge regression model containing\nboth numerical and categorical features that explicitly models and poisons\ncategorical features. We model categorical features as SOS-1 sets and formulate\nthe problem of designing poisoning attacks as a bilevel optimization problem\nthat is nonconvex mixed-integer in the upper-level and unconstrained convex\nquadratic in the lower-level. We present the mathematical formulation of the\nproblem, introduce a single-level reformulation based on the Karush-Kuhn-Tucker\n(KKT) conditions of the lower level, find bounds for the lower-level variables\nto accelerate solver performance, and propose a new algorithm to poison\ncategorical features. Numerical experiments show that our method improves the\nmean squared error of all datasets compared to the previous benchmark in the\nliterature.\n", "date": "2025-01-13", "categories": []}
{"id": "2501.07493", "title": "Exploring and Mitigating Adversarial Manipulation of Voting-Based\n  Leaderboards", "abstract": "  It is now common to evaluate Large Language Models (LLMs) by having humans\nmanually vote to evaluate model outputs, in contrast to typical benchmarks that\nevaluate knowledge or skill at some particular task. Chatbot Arena, the most\npopular benchmark of this type, ranks models by asking users to select the\nbetter response between two randomly selected models (without revealing which\nmodel was responsible for the generations). These platforms are widely trusted\nas a fair and accurate measure of LLM capabilities. In this paper, we show that\nif bot protection and other defenses are not implemented, these voting-based\nbenchmarks are potentially vulnerable to adversarial manipulation.\nSpecifically, we show that an attacker can alter the leaderboard (to promote\ntheir favorite model or demote competitors) at the cost of roughly a thousand\nvotes (verified in a simulated, offline version of Chatbot Arena). Our attack\nconsists of two steps: first, we show how an attacker can determine which model\nwas used to generate a given reply with more than $95\\%$ accuracy; and then,\nthe attacker can use this information to consistently vote for (or against) a\ntarget model. Working with the Chatbot Arena developers, we identify, propose,\nand implement mitigations to improve the robustness of Chatbot Arena against\nadversarial manipulation, which, based on our analysis, substantially increases\nthe cost of such attacks. Some of these defenses were present before our\ncollaboration, such as bot protection with Cloudflare, malicious user\ndetection, and rate limiting. Others, including reCAPTCHA and login are being\nintegrated to strengthen the security in Chatbot Arena.\n", "date": "2025-01-13", "categories": []}
{"id": "2501.05588", "title": "Enforcing Fundamental Relations via Adversarial Attacks on Input\n  Parameter Correlations", "abstract": "  Correlations between input parameters play a crucial role in many scientific\nclassification tasks, since these are often related to fundamental laws of\nnature. For example, in high energy physics, one of the common deep learning\nuse-cases is the classification of signal and background processes in particle\ncollisions. In many such cases, the fundamental principles of the correlations\nbetween observables are often better understood than the actual distributions\nof the observables themselves. In this work, we present a new adversarial\nattack algorithm called Random Distribution Shuffle Attack (RDSA), emphasizing\nthe correlations between observables in the network rather than individual\nfeature characteristics. Correct application of the proposed novel attack can\nresult in a significant improvement in classification performance -\nparticularly in the context of data augmentation - when using the generated\nadversaries within adversarial training. Given that correlations between input\nfeatures are also crucial in many other disciplines. We demonstrate the RDSA\neffectiveness on six classification tasks, including two particle collision\nchallenges (using CERN Open Data), hand-written digit recognition (MNIST784),\nhuman activity recognition (HAR), weather forecasting (Rain in Australia), and\nICU patient mortality (MIMIC-IV), demonstrating a general use case beyond\nfundamental physics for this new type of adversarial attack algorithms.\n", "date": "2025-01-09", "categories": []}
{"id": "2501.05015", "title": "On Measuring Unnoticeability of Graph Adversarial Attacks: Observations,\n  New Measure, and Applications", "abstract": "  Adversarial attacks are allegedly unnoticeable. Prior studies have designed\nattack noticeability measures on graphs, primarily using statistical tests to\ncompare the topology of original and (possibly) attacked graphs. However, we\nobserve two critical limitations in the existing measures. First, because the\nmeasures rely on simple rules, attackers can readily enhance their attacks to\nbypass them, reducing their attack \"noticeability\" and, yet, maintaining their\nattack performance. Second, because the measures naively leverage global\nstatistics, such as degree distributions, they may entirely overlook attacks\nuntil severe perturbations occur, letting the attacks be almost \"totally\nunnoticeable.\" To address the limitations, we introduce HideNSeek, a learnable\nmeasure for graph attack noticeability. First, to mitigate the bypass problem,\nHideNSeek learns to distinguish the original and (potential) attack edges using\na learnable edge scorer (LEO), which scores each edge on its likelihood of\nbeing an attack. Second, to mitigate the overlooking problem, HideNSeek\nconducts imbalance-aware aggregation of all the edge scores to obtain the final\nnoticeability score. Using six real-world graphs, we empirically demonstrate\nthat HideNSeek effectively alleviates the observed limitations, and LEO (i.e.,\nour learnable edge scorer) outperforms eleven competitors in distinguishing\nattack edges under five different attack methods. For an additional\napplication, we show that LEO boost the performance of robust GNNs by removing\nattack-like edges.\n", "date": "2025-01-09", "categories": []}
{"id": "2501.04527", "title": "Towards Fair Class-wise Robustness: Class Optimal Distribution\n  Adversarial Training", "abstract": "  Adversarial training has proven to be a highly effective method for improving\nthe robustness of deep neural networks against adversarial attacks.\nNonetheless, it has been observed to exhibit a limitation in terms of robust\nfairness, characterized by a significant disparity in robustness across\ndifferent classes. Recent efforts to mitigate this problem have turned to\nclass-wise reweighted methods. However, these methods suffer from a lack of\nrigorous theoretical analysis and are limited in their exploration of the\nweight space, as they mainly rely on existing heuristic algorithms or intuition\nto compute weights. In addition, these methods fail to guarantee the\nconsistency of the optimization direction due to the decoupled optimization of\nweights and the model parameters. They potentially lead to suboptimal weight\nassignments and consequently, a suboptimal model. To address these problems,\nthis paper proposes a novel min-max training framework, Class Optimal\nDistribution Adversarial Training (CODAT), which employs distributionally\nrobust optimization to fully explore the class-wise weight space, thus enabling\nthe identification of the optimal weight with theoretical guarantees.\nFurthermore, we derive a closed-form optimal solution to the internal\nmaximization and then get a deterministic equivalent objective function, which\nprovides a theoretical basis for the joint optimization of weights and model\nparameters. Meanwhile, we propose a fairness elasticity coefficient for the\nevaluation of the algorithm with regard to both robustness and robust fairness.\nExperimental results on various datasets show that the proposed method can\neffectively improve the robust fairness of the model and outperform the\nstate-of-the-art approaches.\n", "date": "2025-01-08", "categories": []}
{"id": "2208.06538", "title": "Transferable Adversarial Examples with Bayes Approach", "abstract": "  The vulnerability of deep neural networks (DNNs) to black-box adversarial\nattacks is one of the most heated topics in trustworthy AI. In such attacks,\nthe attackers operate without any insider knowledge of the model, making the\ncross-model transferability of adversarial examples critical. Despite the\npotential for adversarial examples to be effective across various models, it\nhas been observed that adversarial examples that are specifically crafted for a\nspecific model often exhibit poor transferability. In this paper, we explore\nthe transferability of adversarial examples via the lens of Bayesian approach.\nSpecifically, we leverage Bayesian approach to probe the transferability and\nthen study what constitutes a transferability-promoting prior. Following this,\nwe design two concrete transferability-promoting priors, along with an adaptive\ndynamic weighting strategy for instances sampled from these priors. Employing\nthese techniques, we present BayAtk. Extensive experiments illustrate the\nsignificant effectiveness of BayAtk in crafting more transferable adversarial\nexamples against both undefended and defended black-box models compared to\nexisting state-of-the-art attacks.\n", "date": "2022-08-13", "categories": []}
{"id": "2404.08980", "title": "Stability and Generalization in Free Adversarial Training", "abstract": "  While adversarial training methods have significantly improved the robustness\nof deep neural networks against norm-bounded adversarial perturbations, the\ngeneralization gap between their performance on training and test data is\nconsiderably greater than that of standard empirical risk minimization. Recent\nstudies have aimed to connect the generalization properties of adversarially\ntrained classifiers to the min-max optimization algorithm used in their\ntraining. In this work, we analyze the interconnections between generalization\nand optimization in adversarial training using the algorithmic stability\nframework. Specifically, our goal is to compare the generalization gap of\nneural networks trained using the vanilla adversarial training method, which\nfully optimizes perturbations at every iteration, with the free adversarial\ntraining method, which simultaneously optimizes norm-bounded perturbations and\nclassifier parameters. We prove bounds on the generalization error of these\nmethods, indicating that the free adversarial training method may exhibit a\nlower generalization gap between training and test samples due to its\nsimultaneous min-max optimization of classifier weights and perturbation\nvariables. We conduct several numerical experiments to evaluate the\ntrain-to-test generalization gap in vanilla and free adversarial training\nmethods. Our empirical findings also suggest that the free adversarial training\nmethod could lead to a smaller generalization gap over a similar number of\ntraining iterations. The paper code is available at\nhttps://github.com/Xiwei-Cheng/Stability_FreeAT.\n", "date": "2024-04-13", "categories": []}
{"id": "2411.11200", "title": "Countering Backdoor Attacks in Image Recognition: A Survey and\n  Evaluation of Mitigation Strategies", "abstract": "  The widespread adoption of deep learning across various industries has\nintroduced substantial challenges, particularly in terms of model\nexplainability and security. The inherent complexity of deep learning models,\nwhile contributing to their effectiveness, also renders them susceptible to\nadversarial attacks. Among these, backdoor attacks are especially concerning,\nas they involve surreptitiously embedding specific triggers within training\ndata, causing the model to exhibit aberrant behavior when presented with input\ncontaining the triggers. Such attacks often exploit vulnerabilities in\noutsourced processes, compromising model integrity without affecting\nperformance on clean (trigger-free) input data. In this paper, we present a\ncomprehensive review of existing mitigation strategies designed to counter\nbackdoor attacks in image recognition. We provide an in-depth analysis of the\ntheoretical foundations, practical efficacy, and limitations of these\napproaches. In addition, we conduct an extensive benchmarking of sixteen\nstate-of-the-art approaches against eight distinct backdoor attacks, utilizing\nthree datasets, four model architectures, and three poisoning ratios. Our\nresults, derived from 122,236 individual experiments, indicate that while many\napproaches provide some level of protection, their performance can vary\nconsiderably. Furthermore, when compared to two seminal approaches, most newer\napproaches do not demonstrate substantial improvements in overall performance\nor consistency across diverse settings. Drawing from these findings, we propose\npotential directions for developing more effective and generalizable defensive\nmechanisms in the future.\n", "date": "2024-11-17", "categories": []}
{"id": "2501.03272", "title": "Backdoor Token Unlearning: Exposing and Defending Backdoors in\n  Pretrained Language Models", "abstract": "  Supervised fine-tuning has become the predominant method for adapting large\npretrained models to downstream tasks. However, recent studies have revealed\nthat these models are vulnerable to backdoor attacks, where even a small number\nof malicious samples can successfully embed backdoor triggers into the model.\nWhile most existing defense methods focus on post-training backdoor defense,\nefficiently defending against backdoor attacks during training phase remains\nlargely unexplored. To address this gap, we propose a novel defense method\ncalled Backdoor Token Unlearning (BTU), which proactively detects and\nneutralizes trigger tokens during the training stage. Our work is based on two\nkey findings: 1) backdoor learning causes distinctive differences between\nbackdoor token parameters and clean token parameters in word embedding layers,\nand 2) the success of backdoor attacks heavily depends on backdoor token\nparameters. The BTU defense leverages these properties to identify aberrant\nembedding parameters and subsequently removes backdoor behaviors using a\nfine-grained unlearning technique. Extensive evaluations across three datasets\nand four types of backdoor attacks demonstrate that BTU effectively defends\nagainst these threats while preserving the model's performance on primary\ntasks. Our code is available at https://github.com/XDJPH/BTU.\n", "date": "2025-01-05", "categories": []}
{"id": "2501.03402", "title": "On the Adversarial Robustness of Benjamini Hochberg", "abstract": "  The Benjamini-Hochberg (BH) procedure is widely used to control the false\ndetection rate (FDR) in multiple testing. Applications of this control abound\nin drug discovery, forensics, anomaly detection, and, in particular, machine\nlearning, ranging from nonparametric outlier detection to out-of-distribution\ndetection and one-class classification methods. Considering this control could\nbe relied upon in critical safety/security contexts, we investigate its\nadversarial robustness. More precisely, we study under what conditions BH does\nand does not exhibit adversarial robustness, we present a class of simple and\neasily implementable adversarial test-perturbation algorithms, and we perform\ncomputational experiments. With our algorithms, we demonstrate that there are\nconditions under which BH's control can be significantly broken with relatively\nfew (even just one) test score perturbation(s), and provide non-asymptotic\nguarantees on the expected adversarial-adjustment to FDR. Our technical\nanalysis involves a combinatorial reframing of the BH procedure as a ``balls\ninto bins'' process, and drawing a connection to generalized ballot problems to\nfacilitate an information-theoretic approach for deriving non-asymptotic lower\nbounds.\n", "date": "2025-01-06", "categories": []}
{"id": "2501.03437", "title": "DAMAGE: Detecting Adversarially Modified AI Generated Text", "abstract": "  AI humanizers are a new class of online software tools meant to paraphrase\nand rewrite AI-generated text in a way that allows them to evade AI detection\nsoftware. We study 19 AI humanizer and paraphrasing tools and qualitatively\nassess their effects and faithfulness in preserving the meaning of the original\ntext. We show that many existing AI detectors fail to detect humanized text.\nFinally, we demonstrate a robust model that can detect humanized AI text while\nmaintaining a low false positive rate using a data-centric augmentation\napproach. We attack our own detector, training our own fine-tuned model\noptimized against our detector's predictions, and show that our detector's\ncross-humanizer generalization is sufficient to remain robust to this attack.\n", "date": "2025-01-06", "categories": []}
{"id": "2306.06359", "title": "NeRFool: Uncovering the Vulnerability of Generalizable Neural Radiance\n  Fields against Adversarial Perturbations", "abstract": "  Generalizable Neural Radiance Fields (GNeRF) are one of the most promising\nreal-world solutions for novel view synthesis, thanks to their cross-scene\ngeneralization capability and thus the possibility of instant rendering on new\nscenes. While adversarial robustness is essential for real-world applications,\nlittle study has been devoted to understanding its implication on GNeRF. We\nhypothesize that because GNeRF is implemented by conditioning on the source\nviews from new scenes, which are often acquired from the Internet or\nthird-party providers, there are potential new security concerns regarding its\nreal-world applications. Meanwhile, existing understanding and solutions for\nneural networks' adversarial robustness may not be applicable to GNeRF, due to\nits 3D nature and uniquely diverse operations. To this end, we present NeRFool,\nwhich to the best of our knowledge is the first work that sets out to\nunderstand the adversarial robustness of GNeRF. Specifically, NeRFool unveils\nthe vulnerability patterns and important insights regarding GNeRF's adversarial\nrobustness. Built upon the above insights gained from NeRFool, we further\ndevelop NeRFool+, which integrates two techniques capable of effectively\nattacking GNeRF across a wide range of target views, and provide guidelines for\ndefending against our proposed attacks. We believe that our NeRFool/NeRFool+\nlays the initial foundation for future innovations in developing robust\nreal-world GNeRF solutions. Our codes are available at:\nhttps://github.com/GATECH-EIC/NeRFool.\n", "date": "2023-06-10", "categories": []}
{"id": "2501.02232", "title": "Distillation-Enhanced Physical Adversarial Attacks", "abstract": "  The study of physical adversarial patches is crucial for identifying\nvulnerabilities in AI-based recognition systems and developing more robust deep\nlearning models. While recent research has focused on improving patch\nstealthiness for greater practical applicability, achieving an effective\nbalance between stealth and attack performance remains a significant challenge.\nTo address this issue, we propose a novel physical adversarial attack method\nthat leverages knowledge distillation. Specifically, we first define a stealthy\ncolor space tailored to the target environment to ensure smooth blending. Then,\nwe optimize an adversarial patch in an unconstrained color space, which serves\nas the 'teacher' patch. Finally, we use an adversarial knowledge distillation\nmodule to transfer the teacher patch's knowledge to the 'student' patch,\nguiding the optimization of the stealthy patch. Experimental results show that\nour approach improves attack performance by 20%, while maintaining stealth,\nhighlighting its practical value.\n", "date": "2025-01-04", "categories": []}
{"id": "2307.08939", "title": "Runtime Stealthy Perception Attacks against DNN-based Adaptive Cruise\n  Control Systems", "abstract": "  Adaptive Cruise Control (ACC) is a widely used driver assistance technology\nfor maintaining the desired speed and safe distance to the leading vehicle.\nThis paper evaluates the security of the deep neural network (DNN) based ACC\nsystems under runtime stealthy perception attacks that strategically inject\nperturbations into camera data to cause forward collisions. We present a\ncontext-aware strategy for the selection of the most critical times for\ntriggering the attacks and a novel optimization-based method for the adaptive\ngeneration of image perturbations at runtime. We evaluate the effectiveness of\nthe proposed attack using an actual vehicle, a publicly available driving\ndataset, and a realistic simulation platform with the control software from a\nproduction ACC system, a physical-world driving simulator, and interventions by\nthe human driver and safety features such as Advanced Emergency Braking System\n(AEBS). Experimental results show that the proposed attack achieves 142.9 times\nhigher success rate in causing hazards and 82.6% higher evasion rate than\nbaselines, while being stealthy and robust to real-world factors and dynamic\nchanges in the environment. This study highlights the role of human drivers and\nbasic safety mechanisms in preventing attacks.\n", "date": "2023-07-18", "categories": []}
{"id": "2402.01369", "title": "On the Multi-modal Vulnerability of Diffusion Models", "abstract": "  Diffusion models have been widely deployed in various image generation tasks,\ndemonstrating an extraordinary connection between image and text modalities.\nAlthough prior studies have explored the vulnerability of diffusion models from\nthe perspectives of text and image modalities separately, the current research\nlandscape has not yet thoroughly investigated the vulnerabilities that arise\nfrom the integration of multiple modalities, specifically through the joint\nanalysis of textual and visual features. In this paper, we are the first to\nvisualize both text and image feature space embedded by diffusion models and\nobserve a significant difference. The prompts are embedded chaotically in the\ntext feature space, while in the image feature space they are clustered\naccording to their subjects. These fascinating findings may underscore a\npotential misalignment in robustness between the two modalities that exists\nwithin diffusion models. Based on this observation, we propose MMP-Attack,\nwhich leverages multi-modal priors (MMP) to manipulate the generation results\nof diffusion models by appending a specific suffix to the original prompt.\nSpecifically, our goal is to induce diffusion models to generate a specific\nobject while simultaneously eliminating the original object. Our MMP-Attack\nshows a notable advantage over existing studies with superior manipulation\ncapability and efficiency. Our code is publicly available at\n\\url{https://github.com/ydc123/MMP-Attack}.\n", "date": "2024-02-02", "categories": []}
{"id": "2412.16651", "title": "PB-UAP: Hybrid Universal Adversarial Attack For Image Segmentation", "abstract": "  With the rapid advancement of deep learning, the model robustness has become\na significant research hotspot, \\ie, adversarial attacks on deep neural\nnetworks. Existing works primarily focus on image classification tasks, aiming\nto alter the model's predicted labels. Due to the output complexity and deeper\nnetwork architectures, research on adversarial examples for segmentation models\nis still limited, particularly for universal adversarial perturbations. In this\npaper, we propose a novel universal adversarial attack method designed for\nsegmentation models, which includes dual feature separation and low-frequency\nscattering modules. The two modules guide the training of adversarial examples\nin the pixel and frequency space, respectively. Experiments demonstrate that\nour method achieves high attack success rates surpassing the state-of-the-art\nmethods, and exhibits strong transferability across different models.\n", "date": "2024-12-21", "categories": []}
{"id": "2501.01529", "title": "SAFER: Sharpness Aware layer-selective Finetuning for Enhanced\n  Robustness in vision transformers", "abstract": "  Vision transformers (ViTs) have become essential backbones in advanced\ncomputer vision applications and multi-modal foundation models. Despite their\nstrengths, ViTs remain vulnerable to adversarial perturbations, comparable to\nor even exceeding the vulnerability of convolutional neural networks (CNNs).\nFurthermore, the large parameter count and complex architecture of ViTs make\nthem particularly prone to adversarial overfitting, often compromising both\nclean and adversarial accuracy.\n  This paper mitigates adversarial overfitting in ViTs through a novel,\nlayer-selective fine-tuning approach: SAFER. Instead of optimizing the entire\nmodel, we identify and selectively fine-tune a small subset of layers most\nsusceptible to overfitting, applying sharpness-aware minimization to these\nlayers while freezing the rest of the model. Our method consistently enhances\nboth clean and adversarial accuracy over baseline approaches. Typical\nimprovements are around 5%, with some cases achieving gains as high as 20%\nacross various ViT architectures and datasets.\n", "date": "2025-01-02", "categories": []}
{"id": "2501.01620", "title": "Adaptive Meta-learning-based Adversarial Training for Robust Automatic\n  Modulation Classification", "abstract": "  DL-based automatic modulation classification (AMC) models are highly\nsusceptible to adversarial attacks, where even minimal input perturbations can\ncause severe misclassifications. While adversarially training an AMC model\nbased on an adversarial attack significantly increases its robustness against\nthat attack, the AMC model will still be defenseless against other adversarial\nattacks. The theoretically infinite possibilities for adversarial perturbations\nmean that an AMC model will inevitably encounter new unseen adversarial attacks\nif it is ever to be deployed to a real-world communication system. Moreover,\nthe computational limitations and challenges of obtaining new data in real-time\nwill not allow a full training process for the AMC model to adapt to the new\nattack when it is online. To this end, we propose a meta-learning-based\nadversarial training framework for AMC models that substantially enhances\nrobustness against unseen adversarial attacks and enables fast adaptation to\nthese attacks using just a few new training samples, if any are available. Our\nresults demonstrate that this training framework provides superior robustness\nand accuracy with much less online training time than conventional adversarial\ntraining of AMC models, making it highly efficient for real-world deployment.\n", "date": "2025-01-03", "categories": []}
{"id": "2501.01818", "title": "Rerouting LLM Routers", "abstract": "  LLM routers aim to balance quality and cost of generation by classifying\nqueries and routing them to a cheaper or more expensive LLM depending on their\ncomplexity. Routers represent one type of what we call LLM control planes:\nsystems that orchestrate use of one or more LLMs. In this paper, we investigate\nrouters' adversarial robustness.\n  We first define LLM control plane integrity, i.e., robustness of LLM\norchestration to adversarial inputs, as a distinct problem in AI safety. Next,\nwe demonstrate that an adversary can generate query-independent token sequences\nwe call ``confounder gadgets'' that, when added to any query, cause LLM routers\nto send the query to a strong LLM.\n  Our quantitative evaluation shows that this attack is successful both in\nwhite-box and black-box settings against a variety of open-source and\ncommercial routers, and that confounding queries do not affect the quality of\nLLM responses. Finally, we demonstrate that gadgets can be effective while\nmaintaining low perplexity, thus perplexity-based filtering is not an effective\ndefense. We finish by investigating alternative defenses.\n", "date": "2025-01-03", "categories": []}
{"id": "2311.05006", "title": "Familiarity-Based Open-Set Recognition Under Adversarial Attacks", "abstract": "  Open-set recognition (OSR), the identification of novel categories, can be a\ncritical component when deploying classification models in real-world\napplications. Recent work has shown that familiarity-based scoring rules such\nas the Maximum Softmax Probability (MSP) or the Maximum Logit Score (MLS) are\nstrong baselines when the closed-set accuracy is high. However, one of the\npotential weaknesses of familiarity-based OSR are adversarial attacks. Here, we\nstudy gradient-based adversarial attacks on familiarity scores for both types\nof attacks, False Familiarity and False Novelty attacks, and evaluate their\neffectiveness in informed and uninformed settings on TinyImageNet. Furthermore,\nwe explore how novel and familiar samples react to adversarial attacks and\nformulate the adversarial reaction score as an alternative OSR scoring rule,\nwhich shows a high correlation with the MLS familiarity score.\n", "date": "2023-11-08", "categories": []}
{"id": "2411.02974", "title": "Region-Guided Attack on the Segment Anything Model (SAM)", "abstract": "  The Segment Anything Model (SAM) is a cornerstone of image segmentation,\ndemonstrating exceptional performance across various applications, particularly\nin autonomous driving and medical imaging, where precise segmentation is\ncrucial. However, SAM is vulnerable to adversarial attacks that can\nsignificantly impair its functionality through minor input perturbations.\nTraditional techniques, such as FGSM and PGD, are often ineffective in\nsegmentation tasks due to their reliance on global perturbations that overlook\nspatial nuances. Recent methods like Attack-SAM-K and UAD have begun to address\nthese challenges, but they frequently depend on external cues and do not fully\nleverage the structural interdependencies within segmentation processes. This\nlimitation underscores the need for a novel adversarial strategy that exploits\nthe unique characteristics of segmentation tasks. In response, we introduce the\nRegion-Guided Attack (RGA), designed specifically for SAM. RGA utilizes a\nRegion-Guided Map (RGM) to manipulate segmented regions, enabling targeted\nperturbations that fragment large segments and expand smaller ones, resulting\nin erroneous outputs from SAM. Our experiments demonstrate that RGA achieves\nhigh success rates in both white-box and black-box scenarios, emphasizing the\nneed for robust defenses against such sophisticated attacks. RGA not only\nreveals SAM's vulnerabilities but also lays the groundwork for developing more\nresilient defenses against adversarial threats in image segmentation.\n", "date": "2024-11-05", "categories": []}
{"id": "2501.00055", "title": "LLM-Virus: Evolutionary Jailbreak Attack on Large Language Models", "abstract": "  While safety-aligned large language models (LLMs) are increasingly used as\nthe cornerstone for powerful systems such as multi-agent frameworks to solve\ncomplex real-world problems, they still suffer from potential adversarial\nqueries, such as jailbreak attacks, which attempt to induce harmful content.\nResearching attack methods allows us to better understand the limitations of\nLLM and make trade-offs between helpfulness and safety. However, existing\njailbreak attacks are primarily based on opaque optimization techniques (e.g.\ntoken-level gradient descent) and heuristic search methods like LLM refinement,\nwhich fall short in terms of transparency, transferability, and computational\ncost. In light of these limitations, we draw inspiration from the evolution and\ninfection processes of biological viruses and propose LLM-Virus, a jailbreak\nattack method based on evolutionary algorithm, termed evolutionary jailbreak.\nLLM-Virus treats jailbreak attacks as both an evolutionary and transfer\nlearning problem, utilizing LLMs as heuristic evolutionary operators to ensure\nhigh attack efficiency, transferability, and low time cost. Our experimental\nresults on multiple safety benchmarks show that LLM-Virus achieves competitive\nor even superior performance compared to existing attack methods.\n", "date": "2024-12-28", "categories": []}
{"id": "2501.00707", "title": "Everywhere Attack: Attacking Locally and Globally to Boost Targeted\n  Transferability", "abstract": "  Adversarial examples' (AE) transferability refers to the phenomenon that AEs\ncrafted with one surrogate model can also fool other models. Notwithstanding\nremarkable progress in untargeted transferability, its targeted counterpart\nremains challenging. This paper proposes an everywhere scheme to boost targeted\ntransferability. Our idea is to attack a victim image both globally and\nlocally. We aim to optimize 'an army of targets' in every local image region\ninstead of the previous works that optimize a high-confidence target in the\nimage. Specifically, we split a victim image into non-overlap blocks and\njointly mount a targeted attack on each block. Such a strategy mitigates\ntransfer failures caused by attention inconsistency between surrogate and\nvictim models and thus results in stronger transferability. Our approach is\nmethod-agnostic, which means it can be easily combined with existing\ntransferable attacks for even higher transferability. Extensive experiments on\nImageNet demonstrate that the proposed approach universally improves the\nstate-of-the-art targeted attacks by a clear margin, e.g., the transferability\nof the widely adopted Logit attack can be improved by 28.8%-300%.We also\nevaluate the crafted AEs on a real-world platform: Google Cloud Vision. Results\nfurther support the superiority of the proposed method.\n", "date": "2025-01-01", "categories": []}
{"id": "2501.01106", "title": "AIM: Additional Image Guided Generation of Transferable Adversarial\n  Attacks", "abstract": "  Transferable adversarial examples highlight the vulnerability of deep neural\nnetworks (DNNs) to imperceptible perturbations across various real-world\napplications. While there have been notable advancements in untargeted\ntransferable attacks, targeted transferable attacks remain a significant\nchallenge. In this work, we focus on generative approaches for targeted\ntransferable attacks. Current generative attacks focus on reducing overfitting\nto surrogate models and the source data domain, but they often overlook the\nimportance of enhancing transferability through additional semantics. To\naddress this issue, we introduce a novel plug-and-play module into the general\ngenerator architecture to enhance adversarial transferability. Specifically, we\npropose a \\emph{Semantic Injection Module} (SIM) that utilizes the semantics\ncontained in an additional guiding image to improve transferability. The\nguiding image provides a simple yet effective method to incorporate target\nsemantics from the target class to create targeted and highly transferable\nattacks. Additionally, we propose new loss formulations that can integrate the\nsemantic injection module more effectively for both targeted and untargeted\nattacks. We conduct comprehensive experiments under both targeted and\nuntargeted attack settings to demonstrate the efficacy of our proposed\napproach.\n", "date": "2025-01-02", "categories": []}
{"id": "2501.01263", "title": "Stealthy Backdoor Attack to Real-world Models in Android Apps", "abstract": "  Powered by their superior performance, deep neural networks (DNNs) have found\nwidespread applications across various domains. Many deep learning (DL) models\nare now embedded in mobile apps, making them more accessible to end users\nthrough on-device DL. However, deploying on-device DL to users' smartphones\nsimultaneously introduces several security threats. One primary threat is\nbackdoor attacks. Extensive research has explored backdoor attacks for several\nyears and has proposed numerous attack approaches. However, few studies have\ninvestigated backdoor attacks on DL models deployed in the real world, or they\nhave shown obvious deficiencies in effectiveness and stealthiness. In this\nwork, we explore more effective and stealthy backdoor attacks on real-world DL\nmodels extracted from mobile apps. Our main justification is that imperceptible\nand sample-specific backdoor triggers generated by DNN-based steganography can\nenhance the efficacy of backdoor attacks on real-world models. We first confirm\nthe effectiveness of steganography-based backdoor attacks on four\nstate-of-the-art DNN models. Subsequently, we systematically evaluate and\nanalyze the stealthiness of the attacks to ensure they are difficult to\nperceive. Finally, we implement the backdoor attacks on real-world models and\ncompare our approach with three baseline methods. We collect 38,387 mobile\napps, extract 89 DL models from them, and analyze these models to obtain the\nprerequisite model information for the attacks. After identifying the target\nmodels, our approach achieves an average of 12.50% higher attack success rate\nthan DeepPayload while better maintaining the normal performance of the models.\nExtensive experimental results demonstrate that our method enables more\neffective, robust, and stealthy backdoor attacks on real-world models.\n", "date": "2025-01-02", "categories": []}
{"id": "2412.19947", "title": "Standard-Deviation-Inspired Regularization for Improving Adversarial\n  Robustness", "abstract": "  Adversarial Training (AT) has been demonstrated to improve the robustness of\ndeep neural networks (DNNs) against adversarial attacks. AT is a min-max\noptimization procedure where in adversarial examples are generated to train a\nmore robust DNN. The inner maximization step of AT increases the losses of\ninputs with respect to their actual classes. The outer minimization involves\nminimizing the losses on the adversarial examples obtained from the inner\nmaximization. This work proposes a standard-deviation-inspired (SDI)\nregularization term to improve adversarial robustness and generalization. We\nargue that the inner maximization in AT is similar to minimizing a modified\nstandard deviation of the model's output probabilities. Moreover, we suggest\nthat maximizing this modified standard deviation can complement the outer\nminimization of the AT framework. To support our argument, we experimentally\nshow that the SDI measure can be used to craft adversarial examples.\nAdditionally, we demonstrate that combining the SDI regularization term with\nexisting AT variants enhances the robustness of DNNs against stronger attacks,\nsuch as CW and Auto-attack, and improves generalization.\n", "date": "2024-12-27", "categories": []}
{"id": "2412.20087", "title": "On the Validity of Traditional Vulnerability Scoring Systems for\n  Adversarial Attacks against LLMs", "abstract": "  This research investigates the effectiveness of established vulnerability\nmetrics, such as the Common Vulnerability Scoring System (CVSS), in evaluating\nattacks against Large Language Models (LLMs), with a focus on Adversarial\nAttacks (AAs). The study explores the influence of both general and specific\nmetric factors in determining vulnerability scores, providing new perspectives\non potential enhancements to these metrics.\n  This study adopts a quantitative approach, calculating and comparing the\ncoefficient of variation of vulnerability scores across 56 adversarial attacks\non LLMs. The attacks, sourced from various research papers, and obtained\nthrough online databases, were evaluated using multiple vulnerability metrics.\nScores were determined by averaging the values assessed by three distinct LLMs.\nThe results indicate that existing scoring-systems yield vulnerability scores\nwith minimal variation across different attacks, suggesting that many of the\nmetric factors are inadequate for assessing adversarial attacks on LLMs. This\nis particularly true for context-specific factors or those with predefined\nvalue sets, such as those in CVSS. These findings support the hypothesis that\ncurrent vulnerability metrics, especially those with rigid values, are limited\nin evaluating AAs on LLMs, highlighting the need for the development of more\nflexible, generalized metrics tailored to such attacks.\n  This research offers a fresh analysis of the effectiveness and applicability\nof established vulnerability metrics, particularly in the context of\nAdversarial Attacks on Large Language Models, both of which have gained\nsignificant attention in recent years. Through extensive testing and\ncalculations, the study underscores the limitations of these metrics and opens\nup new avenues for improving and refining vulnerability assessment frameworks\nspecifically tailored for LLMs.\n", "date": "2024-12-28", "categories": []}
{"id": "2412.20529", "title": "Attacks on the neural network and defense methods", "abstract": "  This article will discuss the use of attacks on a neural network trained on\naudio data, as well as possible methods of protection against these attacks.\nFGSM, PGD and CW attacks, as well as data poisoning, will be considered. Within\nthe framework of protection, Art-IBM and advertorch libraries will be\nconsidered. The obtained accuracy metrics within the framework of attack\napplications are presented\n", "date": "2024-12-29", "categories": []}
{"id": "2412.20807", "title": "Two Heads Are Better Than One: Averaging along Fine-Tuning to Improve\n  Targeted Transferability", "abstract": "  With much longer optimization time than that of untargeted attacks\nnotwithstanding, the transferability of targeted attacks is still far from\nsatisfactory. Recent studies reveal that fine-tuning an existing adversarial\nexample (AE) in feature space can efficiently boost its targeted\ntransferability. However, existing fine-tuning schemes only utilize the\nendpoint and ignore the valuable information in the fine-tuning trajectory.\nNoting that the vanilla fine-tuning trajectory tends to oscillate around the\nperiphery of a flat region of the loss surface, we propose averaging over the\nfine-tuning trajectory to pull the crafted AE towards a more centered region.\nWe compare the proposed method with existing fine-tuning schemes by integrating\nthem with state-of-the-art targeted attacks in various attacking scenarios.\nExperimental results uphold the superiority of the proposed method in boosting\ntargeted transferability. The code is available at github.com/zengh5/Avg_FT.\n", "date": "2024-12-30", "categories": []}
{"id": "2412.21164", "title": "Adversarial Attack and Defense for LoRa Device Identification and\n  Authentication via Deep Learning", "abstract": "  LoRa provides long-range, energy-efficient communications in Internet of\nThings (IoT) applications that rely on Low-Power Wide-Area Network (LPWAN)\ncapabilities. Despite these merits, concerns persist regarding the security of\nLoRa networks, especially in situations where device identification and\nauthentication are imperative to secure the reliable access to the LoRa\nnetworks. This paper explores a deep learning (DL) approach to tackle these\nconcerns, focusing on two critical tasks, namely (i) identifying LoRa devices\nand (ii) classifying them to legitimate and rogue devices. Deep neural networks\n(DNNs), encompassing both convolutional and feedforward neural networks, are\ntrained for these tasks using actual LoRa signal data. In this setting, the\nadversaries may spoof rogue LoRa signals through the kernel density estimation\n(KDE) method based on legitimate device signals that are received by the\nadversaries. Two cases are considered, (i) training two separate classifiers,\none for each of the two tasks, and (ii) training a multi-task classifier for\nboth tasks. The vulnerabilities of the resulting DNNs to manipulations in input\nsamples are studied in form of untargeted and targeted adversarial attacks\nusing the Fast Gradient Sign Method (FGSM). Individual and common perturbations\nare considered against single-task and multi-task classifiers for the LoRa\nsignal analysis. To provide resilience against such attacks, a defense approach\nis presented by increasing the robustness of classifiers with adversarial\ntraining. Results quantify how vulnerable LoRa signal classification tasks are\nto adversarial attacks and emphasize the need to fortify IoT applications\nagainst these subtle yet effective threats.\n", "date": "2024-12-30", "categories": []}
{"id": "2412.18693", "title": "Diverse and Effective Red Teaming with Auto-generated Rewards and\n  Multi-step Reinforcement Learning", "abstract": "  Automated red teaming can discover rare model failures and generate\nchallenging examples that can be used for training or evaluation. However, a\ncore challenge in automated red teaming is ensuring that the attacks are both\ndiverse and effective. Prior methods typically succeed in optimizing either for\ndiversity or for effectiveness, but rarely both. In this paper, we provide\nmethods that enable automated red teaming to generate a large number of diverse\nand successful attacks.\n  Our approach decomposes the task into two steps: (1) automated methods for\ngenerating diverse attack goals and (2) generating effective attacks for those\ngoals. While we provide multiple straightforward methods for generating diverse\ngoals, our key contributions are to train an RL attacker that both follows\nthose goals and generates diverse attacks for those goals. First, we\ndemonstrate that it is easy to use a large language model (LLM) to generate\ndiverse attacker goals with per-goal prompts and rewards, including rule-based\nrewards (RBRs) to grade whether the attacks are successful for the particular\ngoal. Second, we demonstrate how training the attacker model with multi-step\nRL, where the model is rewarded for generating attacks that are different from\npast attempts further increases diversity while remaining effective. We use our\napproach to generate both prompt injection attacks and prompts that elicit\nunsafe responses. In both cases, we find that our approach is able to generate\nhighly-effective and considerably more diverse attacks than past general\nred-teaming approaches.\n", "date": "2024-12-24", "categories": []}
{"id": "2412.18706", "title": "SurvAttack: Black-Box Attack On Survival Models through\n  Ontology-Informed EHR Perturbation", "abstract": "  Survival analysis (SA) models have been widely studied in mining electronic\nhealth records (EHRs), particularly in forecasting the risk of critical\nconditions for prioritizing high-risk patients. However, their vulnerability to\nadversarial attacks is much less explored in the literature. Developing\nblack-box perturbation algorithms and evaluating their impact on\nstate-of-the-art survival models brings two benefits to medical applications.\nFirst, it can effectively evaluate the robustness of models in pre-deployment\ntesting. Also, exploring how subtle perturbations would result in significantly\ndifferent outcomes can provide counterfactual insights into the clinical\ninterpretation of model prediction. In this work, we introduce SurvAttack, a\nnovel black-box adversarial attack framework leveraging subtle clinically\ncompatible, and semantically consistent perturbations on longitudinal EHRs to\ndegrade survival models' predictive performance. We specifically develop a\ngreedy algorithm to manipulate medical codes with various adversarial actions\nthroughout a patient's medical history. Then, these adversarial actions are\nprioritized using a composite scoring strategy based on multi-aspect\nperturbation quality, including saliency, perturbation stealthiness, and\nclinical meaningfulness. The proposed adversarial EHR perturbation algorithm is\nthen used in an efficient SA-specific strategy to attack a survival model when\nestimating the temporal ranking of survival urgency for patients. To\ndemonstrate the significance of our work, we conduct extensive experiments,\nincluding baseline comparisons, explainability analysis, and case studies. The\nexperimental results affirm our research's effectiveness in illustrating the\nvulnerabilities of patient survival models, model interpretation, and\nultimately contributing to healthcare quality.\n", "date": "2024-12-24", "categories": []}
{"id": "2412.18718", "title": "Evaluating the Adversarial Robustness of Detection Transformers", "abstract": "  Robust object detection is critical for autonomous driving and mobile\nrobotics, where accurate detection of vehicles, pedestrians, and obstacles is\nessential for ensuring safety. Despite the advancements in object detection\ntransformers (DETRs), their robustness against adversarial attacks remains\nunderexplored. This paper presents a comprehensive evaluation of DETR model and\nits variants under both white-box and black-box adversarial attacks, using the\nMS-COCO and KITTI datasets to cover general and autonomous driving scenarios.\nWe extend prominent white-box attack methods (FGSM, PGD, and CW) to assess DETR\nvulnerability, demonstrating that DETR models are significantly susceptible to\nadversarial attacks, similar to traditional CNN-based detectors. Our extensive\ntransferability analysis reveals high intra-network transferability among DETR\nvariants, but limited cross-network transferability to CNN-based models.\nAdditionally, we propose a novel untargeted attack designed specifically for\nDETR, exploiting its intermediate loss functions to induce misclassification\nwith minimal perturbations. Visualizations of self-attention feature maps\nprovide insights into how adversarial attacks affect the internal\nrepresentations of DETR models. These findings reveal critical vulnerabilities\nin detection transformers under standard adversarial attacks, emphasizing the\nneed for future research to enhance the robustness of transformer-based object\ndetectors in safety-critical applications.\n", "date": "2024-12-25", "categories": []}
{"id": "2412.18815", "title": "Distortion-Aware Adversarial Attacks on Bounding Boxes of Object\n  Detectors", "abstract": "  Deep learning-based object detection has become ubiquitous in the last decade\ndue to its high accuracy in many real-world applications. With this growing\ntrend, these models are interested in being attacked by adversaries, with most\nof the results being on classifiers, which do not match the context of\npractical object detection. In this work, we propose a novel method to fool\nobject detectors, expose the vulnerability of state-of-the-art detectors, and\npromote later works to build more robust detectors to adversarial examples. Our\nmethod aims to generate adversarial images by perturbing object confidence\nscores during training, which is crucial in predicting confidence for each\nclass in the testing phase. Herein, we provide a more intuitive technique to\nembed additive noises based on detected objects' masks and the training loss\nwith distortion control over the original image by leveraging the gradient of\niterative images. To verify the proposed method, we perform adversarial attacks\nagainst different object detectors, including the most recent state-of-the-art\nmodels like YOLOv8, Faster R-CNN, RetinaNet, and Swin Transformer. We also\nevaluate our technique on MS COCO 2017 and PASCAL VOC 2012 datasets and analyze\nthe trade-off between success attack rate and image distortion. Our experiments\nshow that the achievable success attack rate is up to $100$\\% and up to $98$\\%\nwhen performing white-box and black-box attacks, respectively. The source code\nand relevant documentation for this work are available at the following link:\nhttps://github.com/anonymous20210106/attack_detector\n", "date": "2024-12-25", "categories": []}
{"id": "2412.18844", "title": "Improving Integrated Gradient-based Transferable Adversarial Examples by\n  Refining the Integration Path", "abstract": "  Transferable adversarial examples are known to cause threats in practical,\nblack-box attack scenarios. A notable approach to improving transferability is\nusing integrated gradients (IG), originally developed for model\ninterpretability. In this paper, we find that existing IG-based attacks have\nlimited transferability due to their naive adoption of IG in model\ninterpretability. To address this limitation, we focus on the IG integration\npath and refine it in three aspects: multiplicity, monotonicity, and diversity,\nsupported by theoretical analyses. We propose the Multiple Monotonic\nDiversified Integrated Gradients (MuMoDIG) attack, which can generate highly\ntransferable adversarial examples on different CNN and ViT models and defenses.\nExperiments validate that MuMoDIG outperforms the latest IG-based attack by up\nto 37.3\\% and other state-of-the-art attacks by 8.4\\%. In general, our study\nreveals that migrating established techniques to improve transferability may\nrequire non-trivial efforts. Code is available at\n\\url{https://github.com/RYC-98/MuMoDIG}.\n", "date": "2024-12-25", "categories": []}
{"id": "2412.18886", "title": "Adversarial Training for Graph Neural Networks via Graph Subspace Energy\n  Optimization", "abstract": "  Despite impressive capability in learning over graph-structured data, graph\nneural networks (GNN) suffer from adversarial topology perturbation in both\ntraining and inference phases. While adversarial training has demonstrated\nremarkable effectiveness in image classification tasks, its suitability for GNN\nmodels has been doubted until a recent advance that shifts the focus from\ntransductive to inductive learning. Still, GNN robustness in the inductive\nsetting is under-explored, and it calls for deeper understanding of GNN\nadversarial training. To this end, we propose a new concept of graph subspace\nenergy (GSE) -- a generalization of graph energy that measures graph stability\n-- of the adjacency matrix, as an indicator of GNN robustness against topology\nperturbations. To further demonstrate the effectiveness of such concept, we\npropose an adversarial training method with the perturbed graphs generated by\nmaximizing the GSE regularization term, referred to as AT-GSE. To deal with the\nlocal and global topology perturbations raised respectively by LRBCD and PRBCD,\nwe employ randomized SVD (RndSVD) and Nystrom low-rank approximation to favor\nthe different aspects of the GSE terms. An extensive set of experiments shows\nthat AT-GSE outperforms consistently the state-of-the-art GNN adversarial\ntraining methods over different homophily and heterophily datasets in terms of\nadversarial accuracy, whilst more surprisingly achieving a superior clean\naccuracy on non-perturbed graphs.\n", "date": "2024-12-25", "categories": []}
{"id": "2412.18975", "title": "Injecting Bias into Text Classification Models using Backdoor Attacks", "abstract": "  The rapid growth of natural language processing (NLP) and pre-trained\nlanguage models have enabled accurate text classification in a variety of\nsettings. However, text classification models are susceptible to backdoor\nattacks, where an attacker embeds a trigger into the victim model to make the\nmodel predict attacker-desired labels in targeted scenarios. In this paper, we\npropose to utilize backdoor attacks for a new purpose: bias injection. We\ndevelop a backdoor attack in which a subset of the training dataset is poisoned\nto associate strong male actors with negative sentiment. We execute our attack\non two popular text classification datasets (IMDb and SST) and seven different\nmodels ranging from traditional Doc2Vec-based models to LSTM networks and\nmodern transformer-based BERT and RoBERTa models. Our results show that the\nreduction in backdoored models' benign classification accuracy is limited,\nimplying that our attacks remain stealthy, whereas the models successfully\nlearn to associate strong male actors with negative sentiment (100% attack\nsuccess rate with >= 3% poison rate). Attacks on BERT and RoBERTa are\nparticularly more stealthy and effective, demonstrating an increased risk of\nusing modern and larger models. We also measure the generalizability of our\nbias injection by proposing two metrics: (i) U-BBSR which uses previously\nunseen words when measuring attack success, and (ii) P-BBSR which measures\nattack success using paraphrased test samples. U-BBSR and P-BBSR results show\nthat the bias injected by our attack can go beyond memorizing a trigger phrase.\n", "date": "2024-12-25", "categories": []}
{"id": "2412.19015", "title": "Imperceptible Adversarial Attacks on Point Clouds Guided by\n  Point-to-Surface Field", "abstract": "  Adversarial attacks on point clouds are crucial for assessing and improving\nthe adversarial robustness of 3D deep learning models. Traditional solutions\nstrictly limit point displacement during attacks, making it challenging to\nbalance imperceptibility with adversarial effectiveness. In this paper, we\nattribute the inadequate imperceptibility of adversarial attacks on point\nclouds to deviations from the underlying surface. To address this, we introduce\na novel point-to-surface (P2S) field that adjusts adversarial perturbation\ndirections by dragging points back to their original underlying surface.\nSpecifically, we use a denoising network to learn the gradient field of the\nlogarithmic density function encoding the shape's surface, and apply a\ndistance-aware adjustment to perturbation directions during attacks, thereby\nenhancing imperceptibility. Extensive experiments show that adversarial attacks\nguided by our P2S field are more imperceptible, outperforming state-of-the-art\nmethods.\n", "date": "2024-12-26", "categories": []}
{"id": "2412.19747", "title": "Enhancing Adversarial Robustness of Deep Neural Networks Through\n  Supervised Contrastive Learning", "abstract": "  Adversarial attacks exploit the vulnerabilities of convolutional neural\nnetworks by introducing imperceptible perturbations that lead to\nmisclassifications, exposing weaknesses in feature representations and decision\nboundaries. This paper presents a novel framework combining supervised\ncontrastive learning and margin-based contrastive loss to enhance adversarial\nrobustness. Supervised contrastive learning improves the structure of the\nfeature space by clustering embeddings of samples within the same class and\nseparating those from different classes. Margin-based contrastive loss,\ninspired by support vector machines, enforces explicit constraints to create\nrobust decision boundaries with well-defined margins. Experiments on the\nCIFAR-100 dataset with a ResNet-18 backbone demonstrate robustness performance\nimprovements in adversarial accuracy under Fast Gradient Sign Method attacks.\n", "date": "2024-12-27", "categories": []}
{"id": "2408.08685", "title": "Can Large Language Models Improve the Adversarial Robustness of Graph\n  Neural Networks?", "abstract": "  Graph neural networks (GNNs) are vulnerable to adversarial attacks,\nespecially for topology perturbations, and many methods that improve the\nrobustness of GNNs have received considerable attention. Recently, we have\nwitnessed the significant success of large language models (LLMs), leading many\nto explore the great potential of LLMs on GNNs. However, they mainly focus on\nimproving the performance of GNNs by utilizing LLMs to enhance the node\nfeatures. Therefore, we ask: Will the robustness of GNNs also be enhanced with\nthe powerful understanding and inference capabilities of LLMs? By presenting\nthe empirical results, we find that despite that LLMs can improve the\nrobustness of GNNs, there is still an average decrease of 23.1% in accuracy,\nimplying that the GNNs remain extremely vulnerable against topology attacks.\nTherefore, another question is how to extend the capabilities of LLMs on graph\nadversarial robustness. In this paper, we propose an LLM-based robust graph\nstructure inference framework, LLM4RGNN, which distills the inference\ncapabilities of GPT-4 into a local LLM for identifying malicious edges and an\nLM-based edge predictor for finding missing important edges, so as to recover a\nrobust graph structure. Extensive experiments demonstrate that LLM4RGNN\nconsistently improves the robustness across various GNNs. Even in some cases\nwhere the perturbation ratio increases to 40%, the accuracy of GNNs is still\nbetter than that on the clean graph. The source code can be found in\nhttps://github.com/zhongjian-zhang/LLM4RGNN.\n", "date": "2024-08-16", "categories": []}
{"id": "2408.10204", "title": "Criticality Leveraged Adversarial Training (CLAT) for Boosted\n  Performance via Parameter Efficiency", "abstract": "  Adversarial training enhances neural network robustness but suffers from a\ntendency to overfit and increased generalization errors on clean data. This\nwork introduces CLAT, an innovative approach that mitigates adversarial\noverfitting by introducing parameter efficiency into the adversarial training\nprocess, improving both clean accuracy and adversarial robustness. Instead of\ntuning the entire model, CLAT identifies and fine-tunes robustness-critical\nlayers - those predominantly learning non-robust features - while freezing the\nremaining model to enhance robustness. It employs dynamic critical layer\nselection to adapt to changes in layer criticality throughout the fine-tuning\nprocess. Empirically, CLAT can be applied on top of existing adversarial\ntraining methods, significantly reduces the number of trainable parameters by\napproximately 95%, and achieves more than a 2% improvement in adversarial\nrobustness compared to baseline methods.\n", "date": "2024-08-19", "categories": []}
{"id": "2412.16662", "title": "Adversarial Attack Against Images Classification based on Generative\n  Adversarial Networks", "abstract": "  Adversarial attacks on image classification systems have always been an\nimportant problem in the field of machine learning, and generative adversarial\nnetworks (GANs), as popular models in the field of image generation, have been\nwidely used in various novel scenarios due to their powerful generative\ncapabilities. However, with the popularity of generative adversarial networks,\nthe misuse of fake image technology has raised a series of security problems,\nsuch as malicious tampering with other people's photos and videos, and invasion\nof personal privacy. Inspired by the generative adversarial networks, this work\nproposes a novel adversarial attack method, aiming to gain insight into the\nweaknesses of the image classification system and improve its anti-attack\nability. Specifically, the generative adversarial networks are used to generate\nadversarial samples with small perturbations but enough to affect the\ndecision-making of the classifier, and the adversarial samples are generated\nthrough the adversarial learning of the training generator and the classifier.\nFrom extensive experiment analysis, we evaluate the effectiveness of the method\non a classical image classification dataset, and the results show that our\nmodel successfully deceives a variety of advanced classifiers while maintaining\nthe naturalness of adversarial samples.\n", "date": "2024-12-21", "categories": []}
{"id": "2412.18302", "title": "FameBias: Embedding Manipulation Bias Attack in Text-to-Image Models", "abstract": "  Text-to-Image (T2I) diffusion models have rapidly advanced, enabling the\ngeneration of high-quality images that align closely with textual descriptions.\nHowever, this progress has also raised concerns about their misuse for\npropaganda and other malicious activities. Recent studies reveal that attackers\ncan embed biases into these models through simple fine-tuning, causing them to\ngenerate targeted imagery when triggered by specific phrases. This underscores\nthe potential for T2I models to act as tools for disseminating propaganda,\nproducing images aligned with an attacker's objective for end-users.\n  Building on this concept, we introduce FameBias, a T2I biasing attack that\nmanipulates the embeddings of input prompts to generate images featuring\nspecific public figures. Unlike prior methods, Famebias operates solely on the\ninput embedding vectors without requiring additional model training. We\nevaluate FameBias comprehensively using Stable Diffusion V2, generating a large\ncorpus of images based on various trigger nouns and target public figures. Our\nexperiments demonstrate that FameBias achieves a high attack success rate while\npreserving the semantic context of the original prompts across multiple\ntrigger-target pairs.\n", "date": "2024-12-24", "categories": []}
{"id": "2412.18365", "title": "Hypergraph Attacks via Injecting Homogeneous Nodes into Elite Hyperedges", "abstract": "  Recent studies have shown that Hypergraph Neural Networks (HGNNs) are\nvulnerable to adversarial attacks. Existing approaches focus on hypergraph\nmodification attacks guided by gradients, overlooking node spanning in the\nhypergraph and the group identity of hyperedges, thereby resulting in limited\nattack performance and detectable attacks. In this manuscript, we present a\nnovel framework, i.e., Hypergraph Attacks via Injecting Homogeneous Nodes into\nElite Hyperedges (IE-Attack), to tackle these challenges. Initially, utilizing\nthe node spanning in the hypergraph, we propose the elite hyperedges sampler to\nidentify hyperedges to be injected. Subsequently, a node generator utilizing\nKernel Density Estimation (KDE) is proposed to generate the homogeneous node\nwith the group identity of hyperedges. Finally, by injecting the homogeneous\nnode into elite hyperedges, IE-Attack improves the attack performance and\nenhances the imperceptibility of attacks. Extensive experiments are conducted\non five authentic datasets to validate the effectiveness of IE-Attack and the\ncorresponding superiority to state-of-the-art methods.\n", "date": "2024-12-24", "categories": []}
{"id": "2408.02632", "title": "SEAS: Self-Evolving Adversarial Safety Optimization for Large Language\n  Models", "abstract": "  As large language models (LLMs) continue to advance in capability and\ninfluence, ensuring their security and preventing harmful outputs has become\ncrucial. A promising approach to address these concerns involves training\nmodels to automatically generate adversarial prompts for red teaming. However,\nthe evolving subtlety of vulnerabilities in LLMs challenges the effectiveness\nof current adversarial methods, which struggle to specifically target and\nexplore the weaknesses of these models. To tackle these challenges, we\nintroduce the $\\mathbf{S}\\text{elf-}\\mathbf{E}\\text{volving\n}\\mathbf{A}\\text{dversarial }\\mathbf{S}\\text{afety }\\mathbf{(SEAS)}$\noptimization framework, which enhances security by leveraging data generated by\nthe model itself. SEAS operates through three iterative stages: Initialization,\nAttack, and Adversarial Optimization, refining both the Red Team and Target\nmodels to improve robustness and safety. This framework reduces reliance on\nmanual testing and significantly enhances the security capabilities of LLMs.\nOur contributions include a novel adversarial framework, a comprehensive safety\ndataset, and after three iterations, the Target model achieves a security level\ncomparable to GPT-4, while the Red Team model shows a marked increase in attack\nsuccess rate (ASR) against advanced models. Our code and datasets are released\nat https://SEAS-LLM.github.io/.\n", "date": "2024-08-05", "categories": []}
{"id": "2412.09692", "title": "Three-in-One: Robust Enhanced Universal Transferable Anti-Facial\n  Retrieval in Online Social Networks", "abstract": "  Deep hash-based retrieval techniques are widely used in facial retrieval\nsystems to improve the efficiency of facial matching. However, it also carries\nthe danger of exposing private information. Deep hash models are easily\ninfluenced by adversarial examples, which can be leveraged to protect private\nimages from malicious retrieval. The existing adversarial example methods\nagainst deep hash models focus on universality and transferability, lacking the\nresearch on its robustness in online social networks (OSNs), which leads to\ntheir failure in anti-retrieval after post-processing. Therefore, we provide\nthe first in-depth discussion on robustness adversarial perturbation in\nuniversal transferable anti-facial retrieval and propose Three-in-One\nAdversarial Perturbation (TOAP). Specifically, we construct a local and global\nCompression Generator (CG) to simulate complex post-processing scenarios, which\ncan be used to mitigate perturbation. Then, we propose robust optimization\nobjectives based on the discovery of the variation patterns of model's\ndistribution after post-processing, and generate adversarial examples using\nthese objectives and meta-learning. Finally, we iteratively optimize\nperturbation by alternately generating adversarial examples and fine-tuning the\nCG, balancing the performance of perturbation while enhancing CG's ability to\nmitigate them. Numerous experiments demonstrate that, in addition to its\nadvantages in universality and transferability, TOAP significantly outperforms\ncurrent state-of-the-art methods in multiple robustness metrics. It further\nimproves universality and transferability by 5% to 28%, and achieves up to\nabout 33% significant improvement in several simulated post-processing\nscenarios as well as mainstream OSNs, demonstrating that TOAP can effectively\nprotect private images from malicious retrieval in real-world scenarios.\n", "date": "2024-12-12", "categories": []}
{"id": "2412.16213", "title": "AdvIRL: Reinforcement Learning-Based Adversarial Attacks on 3D NeRF\n  Models", "abstract": "  The increasing deployment of AI models in critical applications has exposed\nthem to significant risks from adversarial attacks. While adversarial\nvulnerabilities in 2D vision models have been extensively studied, the threat\nlandscape for 3D generative models, such as Neural Radiance Fields (NeRF),\nremains underexplored. This work introduces \\textit{AdvIRL}, a novel framework\nfor crafting adversarial NeRF models using Instant Neural Graphics Primitives\n(Instant-NGP) and Reinforcement Learning. Unlike prior methods, \\textit{AdvIRL}\ngenerates adversarial noise that remains robust under diverse 3D\ntransformations, including rotations and scaling, enabling effective black-box\nattacks in real-world scenarios. Our approach is validated across a wide range\nof scenes, from small objects (e.g., bananas) to large environments (e.g.,\nlighthouses). Notably, targeted attacks achieved high-confidence\nmisclassifications, such as labeling a banana as a slug and a truck as a\ncannon, demonstrating the practical risks posed by adversarial NeRFs. Beyond\nattacking, \\textit{AdvIRL}-generated adversarial models can serve as\nadversarial training data to enhance the robustness of vision systems. The\nimplementation of \\textit{AdvIRL} is publicly available at\n\\url{https://github.com/Tommy-Nguyen-cpu/AdvIRL/tree/MultiView-Clean}, ensuring\nreproducibility and facilitating future research.\n", "date": "2024-12-18", "categories": []}
{"id": "2412.16512", "title": "TrojFlow: Flow Models are Natural Targets for Trojan Attacks", "abstract": "  Flow-based generative models (FMs) have rapidly advanced as a method for\nmapping noise to data, its efficient training and sampling process makes it\nwidely applicable in various fields. FMs can be viewed as a variant of\ndiffusion models (DMs). At the same time, previous studies have shown that DMs\nare vulnerable to Trojan/Backdoor attacks, a type of output manipulation attack\ntriggered by a maliciously embedded pattern at model input. We found that\nTrojan attacks on generative models are essentially equivalent to image\ntransfer tasks from the backdoor distribution to the target distribution, the\nunique ability of FMs to fit any two arbitrary distributions significantly\nsimplifies the training and sampling setups for attacking FMs, making them\ninherently natural targets for backdoor attacks. In this paper, we propose\nTrojFlow, exploring the vulnerabilities of FMs through Trojan attacks. In\nparticular, we consider various attack settings and their combinations and\nthoroughly explore whether existing defense methods for DMs can effectively\ndefend against our proposed attack scenarios. We evaluate TrojFlow on CIFAR-10\nand CelebA datasets, our experiments show that our method can compromise FMs\nwith high utility and specificity, and can easily break through existing\ndefense mechanisms.\n", "date": "2024-12-21", "categories": []}
{"id": "2412.16893", "title": "Preventing Non-intrusive Load Monitoring Privacy Invasion: A Precise\n  Adversarial Attack Scheme for Networked Smart Meters", "abstract": "  Smart grid, through networked smart meters employing the non-intrusive load\nmonitoring (NILM) technique, can considerably discern the usage patterns of\nresidential appliances. However, this technique also incurs privacy leakage. To\naddress this issue, we propose an innovative scheme based on adversarial attack\nin this paper. The scheme effectively prevents NILM models from violating\nappliance-level privacy, while also ensuring accurate billing calculation for\nusers. To achieve this objective, we overcome two primary challenges. First, as\nNILM models fall under the category of time-series regression models, direct\napplication of traditional adversarial attacks designed for classification\ntasks is not feasible. To tackle this issue, we formulate a novel adversarial\nattack problem tailored specifically for NILM and providing a theoretical\nfoundation for utilizing the Jacobian of the NILM model to generate\nimperceptible perturbations. Leveraging the Jacobian, our scheme can produce\nperturbations, which effectively misleads the signal prediction of NILM models\nto safeguard users' appliance-level privacy. The second challenge pertains to\nfundamental utility requirements, where existing adversarial attack schemes\nstruggle to achieve accurate billing calculation for users. To handle this\nproblem, we introduce an additional constraint, mandating that the sum of added\nperturbations within a billing period must be precisely zero. Experimental\nvalidation on real-world power datasets REDD and UK-DALE demonstrates the\nefficacy of our proposed solutions, which can significantly amplify the\ndiscrepancy between the output of the targeted NILM model and the actual power\nsignal of appliances, and enable accurate billing at the same time.\nAdditionally, our solutions exhibit transferability, making the generated\nperturbation signal from one target model applicable to other diverse NILM\nmodels.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.16955", "title": "NumbOD: A Spatial-Frequency Fusion Attack Against Object Detectors", "abstract": "  With the advancement of deep learning, object detectors (ODs) with various\narchitectures have achieved significant success in complex scenarios like\nautonomous driving. Previous adversarial attacks against ODs have been focused\non designing customized attacks targeting their specific structures (e.g., NMS\nand RPN), yielding some results but simultaneously constraining their\nscalability. Moreover, most efforts against ODs stem from image-level attacks\noriginally designed for classification tasks, resulting in redundant\ncomputations and disturbances in object-irrelevant areas (e.g., background).\nConsequently, how to design a model-agnostic efficient attack to\ncomprehensively evaluate the vulnerabilities of ODs remains challenging and\nunresolved. In this paper, we propose NumbOD, a brand-new spatial-frequency\nfusion attack against various ODs, aimed at disrupting object detection within\nimages. We directly leverage the features output by the OD without relying on\nits internal structures to craft adversarial examples. Specifically, we first\ndesign a dual-track attack target selection strategy to select high-quality\nbounding boxes from OD outputs for targeting. Subsequently, we employ\ndirectional perturbations to shift and compress predicted boxes and change\nclassification results to deceive ODs. Additionally, we focus on manipulating\nthe high-frequency components of images to confuse ODs' attention on critical\nobjects, thereby enhancing the attack efficiency. Our extensive experiments on\nnine ODs and two datasets show that NumbOD achieves powerful attack performance\nand high stealthiness.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.16958", "title": "Breaking Barriers in Physical-World Adversarial Examples: Improving\n  Robustness and Transferability via Robust Feature", "abstract": "  As deep neural networks (DNNs) are widely applied in the physical world, many\nresearches are focusing on physical-world adversarial examples (PAEs), which\nintroduce perturbations to inputs and cause the model's incorrect outputs.\nHowever, existing PAEs face two challenges: unsatisfactory attack performance\n(i.e., poor transferability and insufficient robustness to environment\nconditions), and difficulty in balancing attack effectiveness with\nstealthiness, where better attack effectiveness often makes PAEs more\nperceptible.\n  In this paper, we explore a novel perturbation-based method to overcome the\nchallenges. For the first challenge, we introduce a strategy Deceptive RF\ninjection based on robust features (RFs) that are predictive, robust to\nperturbations, and consistent across different models. Specifically, it\nimproves the transferability and robustness of PAEs by covering RFs of other\nclasses onto the predictive features in clean images. For the second challenge,\nwe introduce another strategy Adversarial Semantic Pattern Minimization, which\nremoves most perturbations and retains only essential adversarial patterns in\nAEsBased on the two strategies, we design our method Robust Feature Coverage\nAttack (RFCoA), comprising Robust Feature Disentanglement and Adversarial\nFeature Fusion. In the first stage, we extract target class RFs in feature\nspace. In the second stage, we use attention-based feature fusion to overlay\nthese RFs onto predictive features of clean images and remove unnecessary\nperturbations. Experiments show our method's superior transferability,\nrobustness, and stealthiness compared to existing state-of-the-art methods.\nAdditionally, our method's effectiveness can extend to Large Vision-Language\nModels (LVLMs), indicating its potential applicability to more complex tasks.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.15275", "title": "Fooling LLM graders into giving better grades through neural activity\n  guided adversarial prompting", "abstract": "  The deployment of artificial intelligence (AI) in critical decision-making\nand evaluation processes raises concerns about inherent biases that malicious\nactors could exploit to distort decision outcomes. We propose a systematic\nmethod to reveal such biases in AI evaluation systems and apply it to automated\nessay grading as an example. Our approach first identifies hidden neural\nactivity patterns that predict distorted decision outcomes and then optimizes\nan adversarial input suffix to amplify such patterns. We demonstrate that this\ncombination can effectively fool large language model (LLM) graders into\nassigning much higher grades than humans would. We further show that this\nwhite-box attack transfers to black-box attacks on other models, including\ncommercial closed-source models like Gemini. They further reveal the existence\nof a \"magic word\" that plays a pivotal role in the efficacy of the attack. We\ntrace the origin of this magic word bias to the structure of commonly-used chat\ntemplates for supervised fine-tuning of LLMs and show that a minor change in\nthe template can drastically reduce the bias. This work not only uncovers\nvulnerabilities in current LLMs but also proposes a systematic method to\nidentify and remove hidden biases, contributing to the goal of ensuring AI\nsafety and security.\n", "date": "2024-12-17", "categories": []}
{"id": "2412.15614", "title": "Technical Report for ICML 2024 TiFA Workshop MLLM Attack Challenge:\n  Suffix Injection and Projected Gradient Descent Can Easily Fool An MLLM", "abstract": "  This technical report introduces our top-ranked solution that employs two\napproaches, \\ie suffix injection and projected gradient descent (PGD) , to\naddress the TiFA workshop MLLM attack challenge. Specifically, we first append\nthe text from an incorrectly labeled option (pseudo-labeled) to the original\nquery as a suffix. Using this modified query, our second approach applies the\nPGD method to add imperceptible perturbations to the image. Combining these two\ntechniques enables successful attacks on the LLaVA 1.5 model.\n", "date": "2024-12-20", "categories": []}
{"id": "2412.15924", "title": "Watertox: The Art of Simplicity in Universal Attacks A Cross-Model\n  Framework for Robust Adversarial Generation", "abstract": "  Contemporary adversarial attack methods face significant limitations in\ncross-model transferability and practical applicability. We present Watertox,\nan elegant adversarial attack framework achieving remarkable effectiveness\nthrough architectural diversity and precision-controlled perturbations. Our\ntwo-stage Fast Gradient Sign Method combines uniform baseline perturbations\n($\\epsilon_1 = 0.1$) with targeted enhancements ($\\epsilon_2 = 0.4$). The\nframework leverages an ensemble of complementary architectures, from VGG to\nConvNeXt, synthesizing diverse perspectives through an innovative voting\nmechanism. Against state-of-the-art architectures, Watertox reduces model\naccuracy from 70.6% to 16.0%, with zero-shot attacks achieving up to 98.8%\naccuracy reduction against unseen architectures. These results establish\nWatertox as a significant advancement in adversarial methodologies, with\npromising applications in visual security systems and CAPTCHA generation.\n", "date": "2024-12-20", "categories": []}
{"id": "2412.08282", "title": "How Does the Smoothness Approximation Method Facilitate Generalization\n  for Federated Adversarial Learning?", "abstract": "  Federated Adversarial Learning (FAL) is a robust framework for resisting\nadversarial attacks on federated learning. Although some FAL studies have\ndeveloped efficient algorithms, they primarily focus on convergence performance\nand overlook generalization. Generalization is crucial for evaluating algorithm\nperformance on unseen data. However, generalization analysis is more\nchallenging due to non-smooth adversarial loss functions. A common approach to\naddressing this issue is to leverage smoothness approximation. In this paper,\nwe develop algorithm stability measures to evaluate the generalization\nperformance of two popular FAL algorithms: \\textit{Vanilla FAL (VFAL)} and {\\it\nSlack FAL (SFAL)}, using three different smooth approximation methods: 1)\n\\textit{Surrogate Smoothness Approximation (SSA)}, (2) \\textit{Randomized\nSmoothness Approximation (RSA)}, and (3) \\textit{Over-Parameterized Smoothness\nApproximation (OPSA)}. Based on our in-depth analysis, we answer the question\nof how to properly set the smoothness approximation method to mitigate\ngeneralization error in FAL. Moreover, we identify RSA as the most effective\nmethod for reducing generalization error. In highly data-heterogeneous\nscenarios, we also recommend employing SFAL to mitigate the deterioration of\ngeneralization performance caused by heterogeneity. Based on our theoretical\nresults, we provide insights to help develop more efficient FAL algorithms,\nsuch as designing new metrics and dynamic aggregation rules to mitigate\nheterogeneity.\n", "date": "2024-12-11", "categories": []}
{"id": "2412.14714", "title": "Holistic Adversarially Robust Pruning", "abstract": "  Neural networks can be drastically shrunk in size by removing redundant\nparameters. While crucial for the deployment on resource-constraint hardware,\noftentimes, compression comes with a severe drop in accuracy and lack of\nadversarial robustness. Despite recent advances, counteracting both aspects has\nonly succeeded for moderate compression rates so far. We propose a novel\nmethod, HARP, that copes with aggressive pruning significantly better than\nprior work. For this, we consider the network holistically. We learn a global\ncompression strategy that optimizes how many parameters (compression rate) and\nwhich parameters (scoring connections) to prune specific to each layer\nindividually. Our method fine-tunes an existing model with dynamic\nregularization, that follows a step-wise incremental function balancing the\ndifferent objectives. It starts by favoring robustness before shifting focus on\nreaching the target compression rate and only then handles the objectives\nequally. The learned compression strategies allow us to maintain the\npre-trained model natural accuracy and its adversarial robustness for a\nreduction by 99% of the network original size. Moreover, we observe a crucial\ninfluence of non-uniform compression across layers.\n", "date": "2024-12-19", "categories": []}
{"id": "2210.14164", "title": "Understanding Key Point Cloud Features for Development Three-dimensional\n  Adversarial Attacks", "abstract": "  Adversarial attacks pose serious challenges for deep neural network\n(DNN)-based analysis of various input signals. In the case of three-dimensional\npoint clouds, methods have been developed to identify points that play a key\nrole in network decision, and these become crucial in generating existing\nadversarial attacks. For example, a saliency map approach is a popular method\nfor identifying adversarial drop points, whose removal would significantly\nimpact the network decision. This paper seeks to enhance the understanding of\nthree-dimensional adversarial attacks by exploring which point cloud features\nare most important for predicting adversarial points. Specifically, Fourteen\nkey point cloud features such as edge intensity and distance from the centroid\nare defined, and multiple linear regression is employed to assess their\npredictive power for adversarial points. Based on critical feature selection\ninsights, a new attack method has been developed to evaluate whether the\nselected features can generate an attack successfully. Unlike traditional\nattack methods that rely on model-specific vulnerabilities, this approach\nfocuses on the intrinsic characteristics of the point clouds themselves. It is\ndemonstrated that these features can predict adversarial points across four\ndifferent DNN architectures, Point Network (PointNet), PointNet++, Dynamic\nGraph Convolutional Neural Networks (DGCNN), and Point Convolutional Network\n(PointConv) outperforming random guessing and achieving results comparable to\nsaliency map-based attacks. This study has important engineering applications,\nsuch as enhancing the security and robustness of three-dimensional point\ncloud-based systems in fields like robotics and autonomous driving.\n", "date": "2022-10-19", "categories": []}
{"id": "2306.10963", "title": "Eigenpatches -- Adversarial Patches from Principal Components", "abstract": "  Adversarial patches are still a simple yet powerful white box attack that can\nbe used to fool object detectors by suppressing possible detections. The\npatches of these so-called evasion attacks are computational expensive to\nproduce and require full access to the attacked detector. This paper addresses\nthe problem of computational expensiveness by analyzing 375 generated patches,\ncalculating the principal components of these and show, that linear\ncombinations of the resulting \"eigenpatches\" can be used to fool object\ndetections successfully.\n", "date": "2023-06-19", "categories": []}
{"id": "2404.18791", "title": "Certification of Speaker Recognition Models to Additive Perturbations", "abstract": "  Speaker recognition technology is applied to various tasks, from personal\nvirtual assistants to secure access systems. However, the robustness of these\nsystems against adversarial attacks, particularly to additive perturbations,\nremains a significant challenge. In this paper, we pioneer applying robustness\ncertification techniques to speaker recognition, initially developed for the\nimage domain. Our work covers this gap by transferring and improving randomized\nsmoothing certification techniques against norm-bounded additive perturbations\nfor classification and few-shot learning tasks to speaker recognition. We\ndemonstrate the effectiveness of these methods on VoxCeleb 1 and 2 datasets for\nseveral models. We expect this work to improve the robustness of voice\nbiometrics and accelerate the research of certification methods in the audio\ndomain.\n", "date": "2024-04-29", "categories": []}
{"id": "2412.13376", "title": "Targeted View-Invariant Adversarial Perturbations for 3D Object\n  Recognition", "abstract": "  Adversarial attacks pose significant challenges in 3D object recognition,\nespecially in scenarios involving multi-view analysis where objects can be\nobserved from varying angles. This paper introduces View-Invariant Adversarial\nPerturbations (VIAP), a novel method for crafting robust adversarial examples\nthat remain effective across multiple viewpoints. Unlike traditional methods,\nVIAP enables targeted attacks capable of manipulating recognition systems to\nclassify objects as specific, pre-determined labels, all while using a single\nuniversal perturbation. Leveraging a dataset of 1,210 images across 121 diverse\nrendered 3D objects, we demonstrate the effectiveness of VIAP in both targeted\nand untargeted settings. Our untargeted perturbations successfully generate a\nsingular adversarial noise robust to 3D transformations, while targeted attacks\nachieve exceptional results, with top-1 accuracies exceeding 95% across various\nepsilon values. These findings highlight VIAPs potential for real-world\napplications, such as testing the robustness of 3D recognition systems. The\nproposed method sets a new benchmark for view-invariant adversarial robustness,\nadvancing the field of adversarial machine learning for 3D object recognition.\n", "date": "2024-12-17", "categories": []}
{"id": "2412.13705", "title": "Mitigating Adversarial Attacks in LLMs through Defensive Suffix\n  Generation", "abstract": "  Large language models (LLMs) have exhibited outstanding performance in\nnatural language processing tasks. However, these models remain susceptible to\nadversarial attacks in which slight input perturbations can lead to harmful or\nmisleading outputs. A gradient-based defensive suffix generation algorithm is\ndesigned to bolster the robustness of LLMs. By appending carefully optimized\ndefensive suffixes to input prompts, the algorithm mitigates adversarial\ninfluences while preserving the models' utility. To enhance adversarial\nunderstanding, a novel total loss function ($L_{\\text{total}}$) combining\ndefensive loss ($L_{\\text{def}}$) and adversarial loss ($L_{\\text{adv}}$)\ngenerates defensive suffixes more effectively. Experimental evaluations\nconducted on open-source LLMs such as Gemma-7B, mistral-7B, Llama2-7B, and\nLlama2-13B show that the proposed method reduces attack success rates (ASR) by\nan average of 11\\% compared to models without defensive suffixes. Additionally,\nthe perplexity score of Gemma-7B decreased from 6.57 to 3.93 when applying the\ndefensive suffix generated by openELM-270M. Furthermore, TruthfulQA evaluations\ndemonstrate consistent improvements with Truthfulness scores increasing by up\nto 10\\% across tested configurations. This approach significantly enhances the\nsecurity of LLMs in critical applications without requiring extensive\nretraining.\n", "date": "2024-12-18", "categories": []}
{"id": "2412.13709", "title": "Physics-Based Adversarial Attack on Near-Infrared Human Detector for\n  Nighttime Surveillance Camera Systems", "abstract": "  Many surveillance cameras switch between daytime and nighttime modes based on\nilluminance levels. During the day, the camera records ordinary RGB images\nthrough an enabled IR-cut filter. At night, the filter is disabled to capture\nnear-infrared (NIR) light emitted from NIR LEDs typically mounted around the\nlens. While RGB-based AI algorithm vulnerabilities have been widely reported,\nthe vulnerabilities of NIR-based AI have rarely been investigated. In this\npaper, we identify fundamental vulnerabilities in NIR-based image understanding\ncaused by color and texture loss due to the intrinsic characteristics of\nclothes' reflectance and cameras' spectral sensitivity in the NIR range. We\nfurther show that the nearly co-located configuration of illuminants and\ncameras in existing surveillance systems facilitates concealing and fully\npassive attacks in the physical world. Specifically, we demonstrate how\nretro-reflective and insulation plastic tapes can manipulate the intensity\ndistribution of NIR images. We showcase an attack on the YOLO-based human\ndetector using binary patterns designed in the digital space (via black-box\nquery and searching) and then physically realized using tapes pasted onto\nclothes. Our attack highlights significant reliability concerns for nighttime\nsurveillance systems, which are intended to enhance security. Codes Available:\nhttps://github.com/MyNiuuu/AdvNIR\n", "date": "2024-12-18", "categories": []}
{"id": "2412.13762", "title": "Cultivating Archipelago of Forests: Evolving Robust Decision Trees\n  through Island Coevolution", "abstract": "  Decision trees are widely used in machine learning due to their simplicity\nand interpretability, but they often lack robustness to adversarial attacks and\ndata perturbations. The paper proposes a novel island-based coevolutionary\nalgorithm (ICoEvoRDF) for constructing robust decision tree ensembles. The\nalgorithm operates on multiple islands, each containing populations of decision\ntrees and adversarial perturbations. The populations on each island evolve\nindependently, with periodic migration of top-performing decision trees between\nislands. This approach fosters diversity and enhances the exploration of the\nsolution space, leading to more robust and accurate decision tree ensembles.\nICoEvoRDF utilizes a popular game theory concept of mixed Nash equilibrium for\nensemble weighting, which further leads to improvement in results. ICoEvoRDF is\nevaluated on 20 benchmark datasets, demonstrating its superior performance\ncompared to state-of-the-art methods in optimizing both adversarial accuracy\nand minimax regret. The flexibility of ICoEvoRDF allows for the integration of\ndecision trees from various existing methods, providing a unified framework for\ncombining diverse solutions. Our approach offers a promising direction for\ndeveloping robust and interpretable machine learning models\n", "date": "2024-12-18", "categories": []}
{"id": "2402.02629", "title": "PROSAC: Provably Safe Certification for Machine Learning Models under\n  Adversarial Attacks", "abstract": "  It is widely known that state-of-the-art machine learning models, including\nvision and language models, can be seriously compromised by adversarial\nperturbations. It is therefore increasingly relevant to develop capabilities to\ncertify their performance in the presence of the most effective adversarial\nattacks. Our paper offers a new approach to certify the performance of machine\nlearning models in the presence of adversarial attacks with population level\nrisk guarantees. In particular, we introduce the notion of\n$(\\alpha,\\zeta)$-safe machine learning model. We propose a hypothesis testing\nprocedure, based on the availability of a calibration set, to derive\nstatistical guarantees providing that the probability of declaring that the\nadversarial (population) risk of a machine learning model is less than $\\alpha$\n(i.e. the model is safe), while the model is in fact unsafe (i.e. the model\nadversarial population risk is higher than $\\alpha$), is less than $\\zeta$. We\nalso propose Bayesian optimization algorithms to determine efficiently whether\na machine learning model is $(\\alpha,\\zeta)$-safe in the presence of an\nadversarial attack, along with statistical guarantees. We apply our framework\nto a range of machine learning models - including various sizes of vision\nTransformer (ViT) and ResNet models - impaired by a variety of adversarial\nattacks, such as PGDAttack, MomentumAttack, GenAttack and BanditAttack, to\nillustrate the operation of our approach. Importantly, we show that ViT's are\ngenerally more robust to adversarial attacks than ResNets, and large models are\ngenerally more robust than smaller models. Our approach goes beyond existing\nempirical adversarial risk-based certification guarantees. It formulates\nrigorous (and provable) performance guarantees that can be used to satisfy\nregulatory requirements mandating the use of state-of-the-art technical tools.\n", "date": "2024-02-04", "categories": []}
{"id": "2406.12259", "title": "Adversarial Attacks on Large Language Models in Medicine", "abstract": "  The integration of Large Language Models (LLMs) into healthcare applications\noffers promising advancements in medical diagnostics, treatment\nrecommendations, and patient care. However, the susceptibility of LLMs to\nadversarial attacks poses a significant threat, potentially leading to harmful\noutcomes in delicate medical contexts. This study investigates the\nvulnerability of LLMs to two types of adversarial attacks in three medical\ntasks. Utilizing real-world patient data, we demonstrate that both open-source\nand proprietary LLMs are susceptible to manipulation across multiple tasks.\nThis research further reveals that domain-specific tasks demand more\nadversarial data in model fine-tuning than general domain tasks for effective\nattack execution, especially for more capable models. We discover that while\nintegrating adversarial data does not markedly degrade overall model\nperformance on medical benchmarks, it does lead to noticeable shifts in\nfine-tuned model weights, suggesting a potential pathway for detecting and\ncountering model attacks. This research highlights the urgent need for robust\nsecurity measures and the development of defensive mechanisms to safeguard LLMs\nin medical applications, to ensure their safe and effective deployment in\nhealthcare settings.\n", "date": "2024-06-18", "categories": []}
{"id": "2406.19753", "title": "Attack On Prompt: Backdoor Attack in Prompt-Based Continual Learning", "abstract": "  Prompt-based approaches offer a cutting-edge solution to data privacy issues\nin continual learning, particularly in scenarios involving multiple data\nsuppliers where long-term storage of private user data is prohibited. Despite\ndelivering state-of-the-art performance, its impressive remembering capability\ncan become a double-edged sword, raising security concerns as it might\ninadvertently retain poisoned knowledge injected during learning from private\nuser data. Following this insight, in this paper, we expose continual learning\nto a potential threat: backdoor attack, which drives the model to follow a\ndesired adversarial target whenever a specific trigger is present while still\nperforming normally on clean samples. We highlight three critical challenges in\nexecuting backdoor attacks on incremental learners and propose corresponding\nsolutions: (1) \\emph{Transferability}: We employ a surrogate dataset and\nmanipulate prompt selection to transfer backdoor knowledge to data from other\nsuppliers; (2) \\emph{Resiliency}: We simulate static and dynamic states of the\nvictim to ensure the backdoor trigger remains robust during intense incremental\nlearning processes; and (3) \\emph{Authenticity}: We apply binary cross-entropy\nloss as an anti-cheating factor to prevent the backdoor trigger from devolving\ninto adversarial noise. Extensive experiments across various benchmark datasets\nand continual learners validate our continual backdoor framework, achieving up\nto $100\\%$ attack success rate, with further ablation studies confirming our\ncontributions' effectiveness.\n", "date": "2024-06-28", "categories": []}
{"id": "2412.07559", "title": "Adaptive Epsilon Adversarial Training for Robust Gravitational Wave\n  Parameter Estimation Using Normalizing Flows", "abstract": "  Adversarial training with Normalizing Flow (NF) models is an emerging\nresearch area aimed at improving model robustness through adversarial samples.\nIn this study, we focus on applying adversarial training to NF models for\ngravitational wave parameter estimation. We propose an adaptive epsilon method\nfor Fast Gradient Sign Method (FGSM) adversarial training, which dynamically\nadjusts perturbation strengths based on gradient magnitudes using logarithmic\nscaling. Our hybrid architecture, combining ResNet and Inverse Autoregressive\nFlow, reduces the Negative Log Likelihood (NLL) loss by 47\\% under FGSM attacks\ncompared to the baseline model, while maintaining an NLL of 4.2 on clean data\n(only 5\\% higher than the baseline). For perturbation strengths between 0.01\nand 0.1, our model achieves an average NLL of 5.8, outperforming both\nfixed-epsilon (NLL: 6.7) and progressive-epsilon (NLL: 7.2) methods. Under\nstronger Projected Gradient Descent attacks with perturbation strength of 0.05,\nour model maintains an NLL of 6.4, demonstrating superior robustness while\navoiding catastrophic overfitting.\n", "date": "2024-12-10", "categories": []}
{"id": "2412.12212", "title": "Finding a Wolf in Sheep's Clothing: Combating Adversarial Text-To-Image\n  Prompts with Text Summarization", "abstract": "  Text-to-image models are vulnerable to the stepwise \"Divide-and-Conquer\nAttack\" (DACA) that utilize a large language model to obfuscate inappropriate\ncontent in prompts by wrapping sensitive text in a benign narrative. To\nmitigate stepwise DACA attacks, we propose a two-layer method involving text\nsummarization followed by binary classification. We assembled the Adversarial\nText-to-Image Prompt (ATTIP) dataset ($N=940$), which contained DACA-obfuscated\nand non-obfuscated prompts. From the ATTIP dataset, we created two summarized\nversions: one generated by a small encoder model and the other by a large\nlanguage model. Then, we used an encoder classifier and a GPT-4o classifier to\nperform content moderation on the summarized and unsummarized prompts. When\ncompared with a classifier that operated over the unsummarized data, our method\nimproved F1 score performance by 31%. Further, the highest recorded F1 score\nachieved (98%) was produced by the encoder classifier on a summarized ATTIP\nvariant. This study indicates that pre-classification text summarization can\ninoculate content detection models against stepwise DACA obfuscations.\n", "date": "2024-12-15", "categories": []}
{"id": "2412.12217", "title": "Comprehensive Survey on Adversarial Examples in Cybersecurity: Impacts,\n  Challenges, and Mitigation Strategies", "abstract": "  Deep learning (DL) has significantly transformed cybersecurity, enabling\nadvancements in malware detection, botnet identification, intrusion detection,\nuser authentication, and encrypted traffic analysis. However, the rise of\nadversarial examples (AE) poses a critical challenge to the robustness and\nreliability of DL-based systems. These subtle, crafted perturbations can\ndeceive models, leading to severe consequences like misclassification and\nsystem vulnerabilities. This paper provides a comprehensive review of the\nimpact of AE attacks on key cybersecurity applications, highlighting both their\ntheoretical and practical implications. We systematically examine the methods\nused to generate adversarial examples, their specific effects across various\ndomains, and the inherent trade-offs attackers face between efficacy and\nresource efficiency. Additionally, we explore recent advancements in defense\nmechanisms, including gradient masking, adversarial training, and detection\ntechniques, evaluating their potential to enhance model resilience. By\nsummarizing cutting-edge research, this study aims to bridge the gap between\nadversarial research and practical security applications, offering insights to\nfortify the adoption of DL solutions in cybersecurity.\n", "date": "2024-12-16", "categories": []}
{"id": "2412.12449", "title": "Adversarially robust generalization theory via Jacobian regularization\n  for deep neural networks", "abstract": "  Powerful deep neural networks are vulnerable to adversarial attacks. To\nobtain adversarially robust models, researchers have separately developed\nadversarial training and Jacobian regularization techniques. There are abundant\ntheoretical and empirical studies for adversarial training, but theoretical\nfoundations for Jacobian regularization are still lacking. In this study, we\nshow that Jacobian regularization is closely related to adversarial training in\nthat $\\ell_{2}$ or $\\ell_{1}$ Jacobian regularized loss serves as an\napproximate upper bound on the adversarially robust loss under $\\ell_{2}$ or\n$\\ell_{\\infty}$ adversarial attack respectively. Further, we establish the\nrobust generalization gap for Jacobian regularized risk minimizer via bounding\nthe Rademacher complexity of both the standard loss function class and Jacobian\nregularization function class. Our theoretical results indicate that the norms\nof Jacobian are related to both standard and robust generalization. We also\nperform experiments on MNIST data classification to demonstrate that Jacobian\nregularized risk minimization indeed serves as a surrogate for adversarially\nrobust risk minimization, and that reducing the norms of Jacobian can improve\nboth standard and robust generalization. This study promotes both theoretical\nand empirical understandings to adversarially robust generalization via\nJacobian regularization.\n", "date": "2024-12-17", "categories": []}
{"id": "2412.12626", "title": "Improving the Transferability of 3D Point Cloud Attack via\n  Spectral-aware Admix and Optimization Designs", "abstract": "  Deep learning models for point clouds have shown to be vulnerable to\nadversarial attacks, which have received increasing attention in various\nsafety-critical applications such as autonomous driving, robotics, and\nsurveillance. Existing 3D attackers generally design various attack strategies\nin the white-box setting, requiring the prior knowledge of 3D model details.\nHowever, real-world 3D applications are in the black-box setting, where we can\nonly acquire the outputs of the target classifier. Although few recent works\ntry to explore the black-box attack, they still achieve limited attack success\nrates (ASR). To alleviate this issue, this paper focuses on attacking the 3D\nmodels in a transfer-based black-box setting, where we first carefully design\nadversarial examples in a white-box surrogate model and then transfer them to\nattack other black-box victim models. Specifically, we propose a novel\nSpectral-aware Admix with Augmented Optimization method (SAAO) to improve the\nadversarial transferability. In particular, since traditional Admix strategy\nare deployed in the 2D domain that adds pixel-wise images for perturbing, we\ncan not directly follow it to merge point clouds in coordinate domain as it\nwill destroy the geometric shapes. Therefore, we design spectral-aware fusion\nthat performs Graph Fourier Transform (GFT) to get spectral features of the\npoint clouds and add them in the spectral domain. Afterward, we run a few steps\nwith spectral-aware weighted Admix to select better optimization paths as well\nas to adjust corresponding learning weights. At last, we run more steps to\ngenerate adversarial spectral feature along the optimization path and perform\nInverse-GFT on the adversarial spectral feature to obtain the adversarial\nexample in the data domain. Experiments show that our SAAO achieves better\ntransferability compared to existing 3D attack methods.\n", "date": "2024-12-17", "categories": []}
{"id": "2412.13134", "title": "Practicable Black-box Evasion Attacks on Link Prediction in Dynamic\n  Graphs -- A Graph Sequential Embedding Method", "abstract": "  Link prediction in dynamic graphs (LPDG) has been widely applied to\nreal-world applications such as website recommendation, traffic flow\nprediction, organizational studies, etc. These models are usually kept local\nand secure, with only the interactive interface restrictively available to the\npublic. Thus, the problem of the black-box evasion attack on the LPDG model,\nwhere model interactions and data perturbations are restricted, seems to be\nessential and meaningful in practice. In this paper, we propose the first\npracticable black-box evasion attack method that achieves effective attacks\nagainst the target LPDG model, within a limited amount of interactions and\nperturbations. To perform effective attacks under limited perturbations, we\ndevelop a graph sequential embedding model to find the desired state embedding\nof the dynamic graph sequences, under a deep reinforcement learning framework.\nTo overcome the scarcity of interactions, we design a multi-environment\ntraining pipeline and train our agent for multiple instances, by sharing an\naggregate interaction buffer. Finally, we evaluate our attack against three\nadvanced LPDG models on three real-world graph datasets of different scales and\ncompare its performance with related methods under the interaction and\nperturbation constraints. Experimental results show that our attack is both\neffective and practicable.\n", "date": "2024-12-17", "categories": []}
{"id": "2312.03289", "title": "Enhancing Robustness in Incremental Learning with Adversarial Training", "abstract": "  Adversarial training is one of the most effective approaches against\nadversarial attacks. However, adversarial training has primarily been studied\nin scenarios where data for all classes is provided, with limited research\nconducted in the context of incremental learning where knowledge is introduced\nsequentially. In this study, we investigate Adversarially Robust Class\nIncremental Learning (ARCIL), which deals with adversarial robustness in\nincremental learning. We first explore a series of baselines that integrate\nincremental learning with existing adversarial training methods, finding that\nthey lead to conflicts between acquiring new knowledge and retaining past\nknowledge. Furthermore, we discover that training new knowledge causes the\ndisappearance of a key characteristic in robust models: a flat loss landscape\nin input space. To address such issues, we propose a novel and robust baseline\nfor ARCIL, named \\textbf{FL}atness-preserving \\textbf{A}dversarial\n\\textbf{I}ncremental learning for \\textbf{R}obustness (\\textbf{FLAIR}).\nExperimental results demonstrate that FLAIR significantly outperforms other\nbaselines. To the best of our knowledge, we are the first to comprehensively\ninvestigate the baselines, challenges, and solutions for ARCIL, which we\nbelieve represents a significant advance toward achieving real-world\nrobustness. Codes are available at \\url{https://github.com/HongsinLee/FLAIR}.\n", "date": "2023-12-06", "categories": []}
{"id": "2404.10335", "title": "Efficient Generation of Targeted and Transferable Adversarial Examples\n  for Vision-Language Models Via Diffusion Models", "abstract": "  Adversarial attacks, particularly \\textbf{targeted} transfer-based attacks,\ncan be used to assess the adversarial robustness of large visual-language\nmodels (VLMs), allowing for a more thorough examination of potential security\nflaws before deployment. However, previous transfer-based adversarial attacks\nincur high costs due to high iteration counts and complex method structure.\nFurthermore, due to the unnaturalness of adversarial semantics, the generated\nadversarial examples have low transferability. These issues limit the utility\nof existing methods for assessing robustness. To address these issues, we\npropose AdvDiffVLM, which uses diffusion models to generate natural,\nunrestricted and targeted adversarial examples via score matching.\nSpecifically, AdvDiffVLM uses Adaptive Ensemble Gradient Estimation to modify\nthe score during the diffusion model's reverse generation process, ensuring\nthat the produced adversarial examples have natural adversarial targeted\nsemantics, which improves their transferability. Simultaneously, to improve the\nquality of adversarial examples, we use the GradCAM-guided Mask method to\ndisperse adversarial semantics throughout the image rather than concentrating\nthem in a single area. Finally, AdvDiffVLM embeds more target semantics into\nadversarial examples after multiple iterations. Experimental results show that\nour method generates adversarial examples 5x to 10x faster than\nstate-of-the-art transfer-based adversarial attacks while maintaining higher\nquality adversarial examples. Furthermore, compared to previous transfer-based\nadversarial attacks, the adversarial examples generated by our method have\nbetter transferability. Notably, AdvDiffVLM can successfully attack a variety\nof commercial VLMs in a black-box environment, including GPT-4V.\n", "date": "2024-04-16", "categories": []}
{"id": "2407.15211", "title": "Failures to Find Transferable Image Jailbreaks Between Vision-Language\n  Models", "abstract": "  The integration of new modalities into frontier AI systems offers exciting\ncapabilities, but also increases the possibility such systems can be\nadversarially manipulated in undesirable ways. In this work, we focus on a\npopular class of vision-language models (VLMs) that generate text outputs\nconditioned on visual and textual inputs. We conducted a large-scale empirical\nstudy to assess the transferability of gradient-based universal image\n``jailbreaks\" using a diverse set of over 40 open-parameter VLMs, including 18\nnew VLMs that we publicly release. Overall, we find that transferable\ngradient-based image jailbreaks are extremely difficult to obtain. When an\nimage jailbreak is optimized against a single VLM or against an ensemble of\nVLMs, the jailbreak successfully jailbreaks the attacked VLM(s), but exhibits\nlittle-to-no transfer to any other VLMs; transfer is not affected by whether\nthe attacked and target VLMs possess matching vision backbones or language\nmodels, whether the language model underwent instruction-following and/or\nsafety-alignment training, or many other factors. Only two settings display\npartially successful transfer: between identically-pretrained and\nidentically-initialized VLMs with slightly different VLM training data, and\nbetween different training checkpoints of a single VLM. Leveraging these\nresults, we then demonstrate that transfer can be significantly improved\nagainst a specific target VLM by attacking larger ensembles of\n``highly-similar\" VLMs. These results stand in stark contrast to existing\nevidence of universal and transferable text jailbreaks against language models\nand transferable adversarial attacks against image classifiers, suggesting that\nVLMs may be more robust to gradient-based transfer attacks.\n", "date": "2024-07-21", "categories": []}
{"id": "2408.13896", "title": "HTS-Attack: Heuristic Token Search for Jailbreaking Text-to-Image Models", "abstract": "  Text-to-Image(T2I) models have achieved remarkable success in image\ngeneration and editing, yet these models still have many potential issues,\nparticularly in generating inappropriate or Not-Safe-For-Work(NSFW) content.\nStrengthening attacks and uncovering such vulnerabilities can advance the\ndevelopment of reliable and practical T2I models. Most of the previous works\ntreat T2I models as white-box systems, using gradient optimization to generate\nadversarial prompts. However, accessing the model's gradient is often\nimpossible in real-world scenarios. Moreover, existing defense methods, those\nusing gradient masking, are designed to prevent attackers from obtaining\naccurate gradient information. While several black-box jailbreak attacks have\nbeen explored, they achieve the limited performance of jailbreaking T2I models\ndue to difficulties associated with optimization in discrete spaces. To address\nthis, we propose HTS-Attack, a heuristic token search attack method. HTS-Attack\nbegins with an initialization that removes sensitive tokens, followed by a\nheuristic search where high-performing candidates are recombined and mutated.\nThis process generates a new pool of candidates, and the optimal adversarial\nprompt is updated based on their effectiveness. By incorporating both optimal\nand suboptimal candidates, HTS-Attack avoids local optima and improves\nrobustness in bypassing defenses. Extensive experiments validate the\neffectiveness of our method in attacking the latest prompt checkers, post-hoc\nimage checkers, securely trained T2I models, and online commercial models.\n", "date": "2024-08-25", "categories": []}
{"id": "2412.06727", "title": "Take Fake as Real: Realistic-like Robust Black-box Adversarial Attack to\n  Evade AIGC Detection", "abstract": "  The security of AI-generated content (AIGC) detection is crucial for ensuring\nmultimedia content credibility. To enhance detector security, research on\nadversarial attacks has become essential. However, most existing adversarial\nattacks focus only on GAN-generated facial images detection, struggle to be\neffective on multi-class natural images and diffusion-based detectors, and\nexhibit poor invisibility. To fill this gap, we first conduct an in-depth\nanalysis of the vulnerability of AIGC detectors and discover the feature that\ndetectors vary in vulnerability to different post-processing. Then, considering\nthat the detector is agnostic in real-world scenarios and given this discovery,\nwe propose a Realistic-like Robust Black-box Adversarial attack (R$^2$BA) with\npost-processing fusion optimization. Unlike typical perturbations, R$^2$BA uses\nreal-world post-processing, i.e., Gaussian blur, JPEG compression, Gaussian\nnoise and light spot to generate adversarial examples. Specifically, we use a\nstochastic particle swarm algorithm with inertia decay to optimize\npost-processing fusion intensity and explore the detector's decision boundary.\nGuided by the detector's fake probability, R$^2$BA enhances/weakens the\ndetector-vulnerable/detector-robust post-processing intensity to strike a\nbalance between adversariality and invisibility. Extensive experiments on\npopular/commercial AIGC detectors and datasets demonstrate that R$^2$BA\nexhibits impressive anti-detection performance, excellent invisibility, and\nstrong robustness in GAN-based and diffusion-based cases. Compared to\nstate-of-the-art white-box and black-box attacks, R$^2$BA shows significant\nimprovements of 15\\%--72\\% and 21\\%--47\\% in anti-detection performance under\nthe original and robust scenario respectively, offering valuable insights for\nthe security of AIGC detection in real-world applications.\n", "date": "2024-12-09", "categories": []}
{"id": "2412.08615", "title": "Exploiting the Index Gradients for Optimization-Based Jailbreaking on\n  Large Language Models", "abstract": "  Despite the advancements in training Large Language Models (LLMs) with\nalignment techniques to enhance the safety of generated content, these models\nremain susceptible to jailbreak, an adversarial attack method that exposes\nsecurity vulnerabilities in LLMs. Notably, the Greedy Coordinate Gradient (GCG)\nmethod has demonstrated the ability to automatically generate adversarial\nsuffixes that jailbreak state-of-the-art LLMs. However, the optimization\nprocess involved in GCG is highly time-consuming, rendering the jailbreaking\npipeline inefficient. In this paper, we investigate the process of GCG and\nidentify an issue of Indirect Effect, the key bottleneck of the GCG\noptimization. To this end, we propose the Model Attack Gradient Index GCG\n(MAGIC), that addresses the Indirect Effect by exploiting the gradient\ninformation of the suffix tokens, thereby accelerating the procedure by having\nless computation and fewer iterations. Our experiments on AdvBench show that\nMAGIC achieves up to a 1.5x speedup, while maintaining Attack Success Rates\n(ASR) on par or even higher than other baselines. Our MAGIC achieved an ASR of\n74% on the Llama-2 and an ASR of 54% when conducting transfer attacks on\nGPT-3.5. Code is available at https://github.com/jiah-li/magic.\n", "date": "2024-12-11", "categories": []}
{"id": "2412.08969", "title": "Deep Learning Model Security: Threats and Defenses", "abstract": "  Deep learning has transformed AI applications but faces critical security\nchallenges, including adversarial attacks, data poisoning, model theft, and\nprivacy leakage. This survey examines these vulnerabilities, detailing their\nmechanisms and impact on model integrity and confidentiality. Practical\nimplementations, including adversarial examples, label flipping, and backdoor\nattacks, are explored alongside defenses such as adversarial training,\ndifferential privacy, and federated learning, highlighting their strengths and\nlimitations.\n  Advanced methods like contrastive and self-supervised learning are presented\nfor enhancing robustness. The survey concludes with future directions,\nemphasizing automated defenses, zero-trust architectures, and the security\nchallenges of large AI models. A balanced approach to performance and security\nis essential for developing reliable deep learning systems.\n", "date": "2024-12-12", "categories": []}
{"id": "2412.10617", "title": "BinarySelect to Improve Accessibility of Black-Box Attack Research", "abstract": "  Adversarial text attack research is useful for testing the robustness of NLP\nmodels, however, the rise of transformers has greatly increased the time\nrequired to test attacks. Especially when researchers do not have access to\nadequate resources (e.g. GPUs). This can hinder attack research, as modifying\none example for an attack can require hundreds of queries to a model,\nespecially for black-box attacks. Often these attacks remove one token at a\ntime to find the ideal one to change, requiring $n$ queries (the length of the\ntext) right away. We propose a more efficient selection method called\nBinarySelect which combines binary search and attack selection methods to\ngreatly reduce the number of queries needed to find a token. We find that\nBinarySelect only needs $\\text{log}_2(n) * 2$ queries to find the first token\ncompared to $n$ queries. We also test BinarySelect in an attack setting against\n5 classifiers across 3 datasets and find a viable tradeoff between number of\nqueries saved and attack effectiveness. For example, on the Yelp dataset, the\nnumber of queries is reduced by 32% (72 less) with a drop in attack\neffectiveness of only 5 points. We believe that BinarySelect can help future\nresearchers study adversarial attacks and black-box problems more efficiently\nand opens the door for researchers with access to less resources.\n", "date": "2024-12-13", "categories": []}
{"id": "2412.10713", "title": "RAT: Adversarial Attacks on Deep Reinforcement Agents for Targeted\n  Behaviors", "abstract": "  Evaluating deep reinforcement learning (DRL) agents against targeted behavior\nattacks is critical for assessing their robustness. These attacks aim to\nmanipulate the victim into specific behaviors that align with the attacker's\nobjectives, often bypassing traditional reward-based defenses. Prior methods\nhave primarily focused on reducing cumulative rewards; however, rewards are\ntypically too generic to capture complex safety requirements effectively. As a\nresult, focusing solely on reward reduction can lead to suboptimal attack\nstrategies, particularly in safety-critical scenarios where more precise\nbehavior manipulation is needed. To address these challenges, we propose RAT, a\nmethod designed for universal, targeted behavior attacks. RAT trains an\nintention policy that is explicitly aligned with human preferences, serving as\na precise behavioral target for the adversary. Concurrently, an adversary\nmanipulates the victim's policy to follow this target behavior. To enhance the\neffectiveness of these attacks, RAT dynamically adjusts the state occupancy\nmeasure within the replay buffer, allowing for more controlled and effective\nbehavior manipulation. Our empirical results on robotic simulation tasks\ndemonstrate that RAT outperforms existing adversarial attack algorithms in\ninducing specific behaviors. Additionally, RAT shows promise in improving agent\nrobustness, leading to more resilient policies. We further validate RAT by\nguiding Decision Transformer agents to adopt behaviors aligned with human\npreferences in various MuJoCo tasks, demonstrating its effectiveness across\ndiverse tasks.\n", "date": "2024-12-14", "categories": []}
{"id": "2412.11119", "title": "Impact of Adversarial Attacks on Deep Learning Model Explainability", "abstract": "  In this paper, we investigate the impact of adversarial attacks on the\nexplainability of deep learning models, which are commonly criticized for their\nblack-box nature despite their capacity for autonomous feature extraction. This\nblack-box nature can affect the perceived trustworthiness of these models. To\naddress this, explainability techniques such as GradCAM, SmoothGrad, and LIME\nhave been developed to clarify model decision-making processes. Our research\nfocuses on the robustness of these explanations when models are subjected to\nadversarial attacks, specifically those involving subtle image perturbations\nthat are imperceptible to humans but can significantly mislead models. For\nthis, we utilize attack methods like the Fast Gradient Sign Method (FGSM) and\nthe Basic Iterative Method (BIM) and observe their effects on model accuracy\nand explanations. The results reveal a substantial decline in model accuracy,\nwith accuracies dropping from 89.94% to 58.73% and 45.50% under FGSM and BIM\nattacks, respectively. Despite these declines in accuracy, the explanation of\nthe models measured by metrics such as Intersection over Union (IoU) and Root\nMean Square Error (RMSE) shows negligible changes, suggesting that these\nmetrics may not be sensitive enough to detect the presence of adversarial\nperturbations.\n", "date": "2024-12-15", "categories": []}
{"id": "2412.11608", "title": "Towards Adversarial Robustness of Model-Level Mixture-of-Experts\n  Architectures for Semantic Segmentation", "abstract": "  Vulnerability to adversarial attacks is a well-known deficiency of deep\nneural networks. Larger networks are generally more robust, and ensembling is\none method to increase adversarial robustness: each model's weaknesses are\ncompensated by the strengths of others. While an ensemble uses a deterministic\nrule to combine model outputs, a mixture of experts (MoE) includes an\nadditional learnable gating component that predicts weights for the outputs of\nthe expert models, thus determining their contributions to the final\nprediction. MoEs have been shown to outperform ensembles on specific tasks, yet\ntheir susceptibility to adversarial attacks has not been studied yet. In this\nwork, we evaluate the adversarial vulnerability of MoEs for semantic\nsegmentation of urban and highway traffic scenes. We show that MoEs are, in\nmost cases, more robust to per-instance and universal white-box adversarial\nattacks and can better withstand transfer attacks. Our code is available at\n\\url{https://github.com/KASTEL-MobilityLab/mixtures-of-experts/}.\n", "date": "2024-12-16", "categories": []}
{"id": "2409.03646", "title": "Limited but consistent gains in adversarial robustness by co-training\n  object recognition models with human EEG", "abstract": "  In contrast to human vision, artificial neural networks (ANNs) remain\nrelatively susceptible to adversarial attacks. To address this vulnerability,\nefforts have been made to transfer inductive bias from human brains to ANNs,\noften by training the ANN representations to match their biological\ncounterparts. Previous works relied on brain data acquired in rodents or\nprimates using invasive techniques, from specific regions of the brain, under\nnon-natural conditions (anesthetized animals), and with stimulus datasets\nlacking diversity and naturalness. In this work, we explored whether aligning\nmodel representations to human EEG responses to a rich set of real-world images\nincreases robustness to ANNs. Specifically, we trained ResNet50-backbone models\non a dual task of classification and EEG prediction; and evaluated their EEG\nprediction accuracy and robustness to adversarial attacks. We observed\nsignificant correlation between the networks' EEG prediction accuracy, often\nhighest around 100 ms post stimulus onset, and their gains in adversarial\nrobustness. Although effect size was limited, effects were consistent across\ndifferent random initializations and robust for architectural variants. We\nfurther teased apart the data from individual EEG channels and observed\nstrongest contribution from electrodes in the parieto-occipital regions. The\ndemonstrated utility of human EEG for such tasks opens up avenues for future\nefforts that scale to larger datasets under diverse stimuli conditions with the\npromise of stronger effects.\n", "date": "2024-09-05", "categories": []}
{"id": "2412.09910", "title": "Prompt2Perturb (P2P): Text-Guided Diffusion-Based Adversarial Attacks on\n  Breast Ultrasound Images", "abstract": "  Deep neural networks (DNNs) offer significant promise for improving breast\ncancer diagnosis in medical imaging. However, these models are highly\nsusceptible to adversarial attacks--small, imperceptible changes that can\nmislead classifiers--raising critical concerns about their reliability and\nsecurity. Traditional attacks rely on fixed-norm perturbations, misaligning\nwith human perception. In contrast, diffusion-based attacks require pre-trained\nmodels, demanding substantial data when these models are unavailable, limiting\npractical use in data-scarce scenarios. In medical imaging, however, this is\noften unfeasible due to the limited availability of datasets. Building on\nrecent advancements in learnable prompts, we propose Prompt2Perturb (P2P), a\nnovel language-guided attack method capable of generating meaningful attack\nexamples driven by text instructions. During the prompt learning phase, our\napproach leverages learnable prompts within the text encoder to create subtle,\nyet impactful, perturbations that remain imperceptible while guiding the model\ntowards targeted outcomes. In contrast to current prompt learning-based\napproaches, our P2P stands out by directly updating text embeddings, avoiding\nthe need for retraining diffusion models. Further, we leverage the finding that\noptimizing only the early reverse diffusion steps boosts efficiency while\nensuring that the generated adversarial examples incorporate subtle noise, thus\npreserving ultrasound image quality without introducing noticeable artifacts.\nWe show that our method outperforms state-of-the-art attack techniques across\nthree breast ultrasound datasets in FID and LPIPS. Moreover, the generated\nimages are both more natural in appearance and more effective compared to\nexisting adversarial attacks. Our code will be publicly available\nhttps://github.com/yasamin-med/P2P.\n", "date": "2024-12-13", "categories": []}
{"id": "2412.10265", "title": "Adversarial Robustness of Bottleneck Injected Deep Neural Networks for\n  Task-Oriented Communication", "abstract": "  This paper investigates the adversarial robustness of Deep Neural Networks\n(DNNs) using Information Bottleneck (IB) objectives for task-oriented\ncommunication systems. We empirically demonstrate that while IB-based\napproaches provide baseline resilience against attacks targeting downstream\ntasks, the reliance on generative models for task-oriented communication\nintroduces new vulnerabilities. Through extensive experiments on several\ndatasets, we analyze how bottleneck depth and task complexity influence\nadversarial robustness. Our key findings show that Shallow Variational\nBottleneck Injection (SVBI) provides less adversarial robustness compared to\nDeep Variational Information Bottleneck (DVIB) approaches, with the gap\nwidening for more complex tasks. Additionally, we reveal that IB-based\nobjectives exhibit stronger robustness against attacks focusing on salient\npixels with high intensity compared to those perturbing many pixels with lower\nintensity. Lastly, we demonstrate that task-oriented communication systems that\nrely on generative models to extract and recover salient information have an\nincreased attack surface. The results highlight important security\nconsiderations for next-generation communication systems that leverage neural\nnetworks for goal-oriented compression.\n", "date": "2024-12-13", "categories": []}
{"id": "2402.11940", "title": "AICAttack: Adversarial Image Captioning Attack with Attention-Based\n  Optimization", "abstract": "  Recent advances in deep learning research have shown remarkable achievements\nacross many tasks in computer vision (CV) and natural language processing\n(NLP). At the intersection of CV and NLP is the problem of image captioning,\nwhere the related models' robustness against adversarial attacks has not been\nwell studied. This paper presents a novel adversarial attack strategy,\nAICAttack (Attention-based Image Captioning Attack), designed to attack image\ncaptioning models through subtle perturbations on images. Operating within a\nblack-box attack scenario, our algorithm requires no access to the target\nmodel's architecture, parameters, or gradient information. We introduce an\nattention-based candidate selection mechanism that identifies the optimal\npixels to attack, followed by a customised differential evolution method to\noptimise the perturbations of pixels' RGB values. We demonstrate AICAttack's\neffectiveness through extensive experiments on benchmark datasets against\nmultiple victim models. The experimental results demonstrate that our method\noutperforms current leading-edge techniques by achieving consistently higher\nattack success rates.\n", "date": "2024-02-19", "categories": []}
{"id": "2402.13575", "title": "Flexible Physical Camouflage Generation Based on a Differential Approach", "abstract": "  This study introduces a novel approach to neural rendering, specifically\ntailored for adversarial camouflage, within an extensive 3D rendering\nframework. Our method, named FPA, goes beyond traditional techniques by\nfaithfully simulating lighting conditions and material variations, ensuring a\nnuanced and realistic representation of textures on a 3D target. To achieve\nthis, we employ a generative approach that learns adversarial patterns from a\ndiffusion model. This involves incorporating a specially designed adversarial\nloss and covert constraint loss to guarantee the adversarial and covert nature\nof the camouflage in the physical world. Furthermore, we showcase the\neffectiveness of the proposed camouflage in sticker mode, demonstrating its\nability to cover the target without compromising adversarial information.\nThrough empirical and physical experiments, FPA exhibits strong performance in\nterms of attack success rate and transferability. Additionally, the designed\nsticker-mode camouflage, coupled with a concealment constraint, adapts to the\nenvironment, yielding diverse styles of texture. Our findings highlight the\nversatility and efficacy of the FPA approach in adversarial camouflage\napplications.\n", "date": "2024-02-21", "categories": []}
{"id": "2407.11463", "title": "Investigating Imperceptibility of Adversarial Attacks on Tabular Data:\n  An Empirical Analysis", "abstract": "  Adversarial attacks are a potential threat to machine learning models by\ncausing incorrect predictions through imperceptible perturbations to the input\ndata. While these attacks have been extensively studied in unstructured data\nlike images, applying them to tabular data, poses new challenges. These\nchallenges arise from the inherent heterogeneity and complex feature\ninterdependencies in tabular data, which differ from the image data. To account\nfor this distinction, it is necessary to establish tailored imperceptibility\ncriteria specific to tabular data. However, there is currently a lack of\nstandardised metrics for assessing the imperceptibility of adversarial attacks\non tabular data. To address this gap, we propose a set of key properties and\ncorresponding metrics designed to comprehensively characterise imperceptible\nadversarial attacks on tabular data. These are: proximity to the original\ninput, sparsity of altered features, deviation from the original data\ndistribution, sensitivity in perturbing features with narrow distribution,\nimmutability of certain features that should remain unchanged, feasibility of\nspecific feature values that should not go beyond valid practical ranges, and\nfeature interdependencies capturing complex relationships between data\nattributes. We evaluate the imperceptibility of five adversarial attacks,\nincluding both bounded attacks and unbounded attacks, on tabular data using the\nproposed imperceptibility metrics. The results reveal a trade-off between the\nimperceptibility and effectiveness of these attacks. The study also identifies\nlimitations in current attack algorithms, offering insights that can guide\nfuture research in the area. The findings gained from this empirical analysis\nprovide valuable direction for enhancing the design of adversarial attack\nalgorithms, thereby advancing adversarial machine learning on tabular data.\n", "date": "2024-07-16", "categories": []}
{"id": "2412.09150", "title": "Evaluating Adversarial Attacks on Traffic Sign Classifiers beyond\n  Standard Baselines", "abstract": "  Adversarial attacks on traffic sign classification models were among the\nfirst successfully tried in the real world. Since then, the research in this\narea has been mainly restricted to repeating baseline models, such as LISA-CNN\nor GTSRB-CNN, and similar experiment settings, including white and black\npatches on traffic signs. In this work, we decouple model architectures from\nthe datasets and evaluate on further generic models to make a fair comparison.\nFurthermore, we compare two attack settings, inconspicuous and visible, which\nare usually regarded without direct comparison. Our results show that standard\nbaselines like LISA-CNN or GTSRB-CNN are significantly more susceptible than\nthe generic ones. We, therefore, suggest evaluating new attacks on a broader\nspectrum of baselines in the future. Our code is available at\n\\url{https://github.com/KASTEL-MobilityLab/attacks-on-traffic-sign-recognition/}.\n", "date": "2024-12-12", "categories": []}
{"id": "2412.09195", "title": "On the Generation and Removal of Speaker Adversarial Perturbation for\n  Voice-Privacy Protection", "abstract": "  Neural networks are commonly known to be vulnerable to adversarial attacks\nmounted through subtle perturbation on the input data. Recent development in\nvoice-privacy protection has shown the positive use cases of the same technique\nto conceal speaker's voice attribute with additive perturbation signal\ngenerated by an adversarial network. This paper examines the reversibility\nproperty where an entity generating the adversarial perturbations is authorized\nto remove them and restore original speech (e.g., the speaker him/herself). A\nsimilar technique could also be used by an investigator to deanonymize a\nvoice-protected speech to restore criminals' identities in security and\nforensic analysis. In this setting, the perturbation generative module is\nassumed to be known in the removal process. To this end, a joint training of\nperturbation generation and removal modules is proposed. Experimental results\non the LibriSpeech dataset demonstrated that the subtle perturbations added to\nthe original speech can be predicted from the anonymized speech while achieving\nthe goal of privacy protection. By removing these perturbations from the\nanonymized sample, the original speech can be restored. Audio samples can be\nfound in \\url{https://voiceprivacy.github.io/Perturbation-Generation-Removal/}.\n", "date": "2024-12-12", "categories": []}
{"id": "2306.06909", "title": "Graph Agent Network: Empowering Nodes with Inference Capabilities for\n  Adversarial Resilience", "abstract": "  End-to-end training with global optimization have popularized graph neural\nnetworks (GNNs) for node classification, yet inadvertently introduced\nvulnerabilities to adversarial edge-perturbing attacks. Adversaries can exploit\nthe inherent opened interfaces of GNNs' input and output, perturbing critical\nedges and thus manipulating the classification results. Current defenses, due\nto their persistent utilization of global-optimization-based end-to-end\ntraining schemes, inherently encapsulate the vulnerabilities of GNNs. This is\nspecifically evidenced in their inability to defend against targeted secondary\nattacks. In this paper, we propose the Graph Agent Network (GAgN) to address\nthe aforementioned vulnerabilities of GNNs. GAgN is a graph-structured agent\nnetwork in which each node is designed as an 1-hop-view agent. Through the\ndecentralized interactions between agents, they can learn to infer global\nperceptions to perform tasks including inferring embeddings, degrees and\nneighbor relationships for given nodes. This empowers nodes to filtering\nadversarial edges while carrying out classification tasks. Furthermore, agents'\nlimited view prevents malicious messages from propagating globally in GAgN,\nthereby resisting global-optimization-based secondary attacks. We prove that\nsingle-hidden-layer multilayer perceptrons (MLPs) are theoretically sufficient\nto achieve these functionalities. Experimental results show that GAgN\neffectively implements all its intended capabilities and, compared to\nstate-of-the-art defenses, achieves optimal classification accuracy on the\nperturbed datasets.\n", "date": "2023-06-12", "categories": []}
{"id": "2402.16822", "title": "Rainbow Teaming: Open-Ended Generation of Diverse Adversarial Prompts", "abstract": "  As large language models (LLMs) become increasingly prevalent across many\nreal-world applications, understanding and enhancing their robustness to\nadversarial attacks is of paramount importance. Existing methods for\nidentifying adversarial prompts tend to focus on specific domains, lack\ndiversity, or require extensive human annotations. To address these\nlimitations, we present Rainbow Teaming, a novel black-box approach for\nproducing a diverse collection of adversarial prompts. Rainbow Teaming casts\nadversarial prompt generation as a quality-diversity problem and uses\nopen-ended search to generate prompts that are both effective and diverse.\nFocusing on the safety domain, we use Rainbow Teaming to target various\nstate-of-the-art LLMs, including the Llama 2 and Llama 3 models. Our approach\nreveals hundreds of effective adversarial prompts, with an attack success rate\nexceeding 90% across all tested models. Furthermore, we demonstrate that\nprompts generated by Rainbow Teaming are highly transferable and that\nfine-tuning models with synthetic data generated by our method significantly\nenhances their safety without sacrificing general performance or helpfulness.\nWe additionally explore the versatility of Rainbow Teaming by applying it to\nquestion answering and cybersecurity, showcasing its potential to drive robust\nopen-ended self-improvement in a wide range of applications.\n", "date": "2024-02-26", "categories": []}
{"id": "2410.21453", "title": "Inverting Gradient Attacks Makes Powerful Data Poisoning", "abstract": "  Gradient attacks and data poisoning tamper with the training of machine\nlearning algorithms to maliciously alter them and have been proven to be\nequivalent in convex settings. The extent of harm these attacks can produce in\nnon-convex settings is still to be determined. Gradient attacks can affect far\nless systems than data poisoning but have been argued to be more harmful since\nthey can be arbitrary, whereas data poisoning reduces the attacker's power to\nonly being able to inject data points to training sets, via e.g. legitimate\nparticipation in a collaborative dataset. This raises the question of whether\nthe harm made by gradient attacks can be matched by data poisoning in\nnon-convex settings. In this work, we provide a positive answer in a worst-case\nscenario and show how data poisoning can mimic a gradient attack to perform an\navailability attack on (non-convex) neural networks. Through gradient\ninversion, commonly used to reconstruct data points from actual gradients, we\nshow how reconstructing data points out of malicious gradients can be\nsufficient to perform a range of attacks. This allows us to show, for the first\ntime, an availability attack on neural networks through data poisoning, that\ndegrades the model's performances to random-level through a minority (as low as\n1%) of poisoned points.\n", "date": "2024-10-28", "categories": []}
{"id": "2411.01084", "title": "Plentiful Jailbreaks with String Compositions", "abstract": "  Large language models (LLMs) remain vulnerable to a slew of adversarial\nattacks and jailbreaking methods. One common approach employed by white-hat\nattackers, or red-teamers, is to process model inputs and outputs using\nstring-level obfuscations, which can include leetspeak, rotary ciphers, Base64,\nASCII, and more. Our work extends these encoding-based attacks by unifying them\nin a framework of invertible string transformations. With invertibility, we can\ndevise arbitrary string compositions, defined as sequences of transformations,\nthat we can encode and decode end-to-end programmatically. We devise a\nautomated best-of-n attack that samples from a combinatorially large number of\nstring compositions. Our jailbreaks obtain competitive attack success rates on\nseveral leading frontier models when evaluated on HarmBench, highlighting that\nencoding-based attacks remain a persistent vulnerability even in advanced LLMs.\n", "date": "2024-11-01", "categories": []}
{"id": "2412.07996", "title": "Enhancing Remote Adversarial Patch Attacks on Face Detectors with Tiling\n  and Scaling", "abstract": "  This paper discusses the attack feasibility of Remote Adversarial Patch (RAP)\ntargeting face detectors. The RAP that targets face detectors is similar to the\nRAP that targets general object detectors, but the former has multiple issues\nin the attack process the latter does not. (1) It is possible to detect objects\nof various scales. In particular, the area of small objects that are convolved\nduring feature extraction by CNN is small,so the area that affects the\ninference results is also small. (2) It is a two-class classification, so there\nis a large gap in characteristics between the classes. This makes it difficult\nto attack the inference results by directing them to a different class. In this\npaper, we propose a new patch placement method and loss function for each\nproblem. The patches targeting the proposed face detector showed superior\ndetection obstruct effects compared to the patches targeting the general object\ndetector.\n", "date": "2024-12-11", "categories": []}
{"id": "2412.08156", "title": "Antelope: Potent and Concealed Jailbreak Attack Strategy", "abstract": "  Due to the remarkable generative potential of diffusion-based models,\nnumerous researches have investigated jailbreak attacks targeting these\nframeworks. A particularly concerning threat within image models is the\ngeneration of Not-Safe-for-Work (NSFW) content. Despite the implementation of\nsecurity filters, numerous efforts continue to explore ways to circumvent these\nsafeguards. Current attack methodologies primarily encompass adversarial prompt\nengineering or concept obfuscation, yet they frequently suffer from slow search\nefficiency, conspicuous attack characteristics and poor alignment with targets.\nTo overcome these challenges, we propose Antelope, a more robust and covert\njailbreak attack strategy designed to expose security vulnerabilities inherent\nin generative models. Specifically, Antelope leverages the confusion of\nsensitive concepts with similar ones, facilitates searches in the semantically\nadjacent space of these related concepts and aligns them with the target\nimagery, thereby generating sensitive images that are consistent with the\ntarget and capable of evading detection. Besides, we successfully exploit the\ntransferability of model-based attacks to penetrate online black-box services.\nExperimental evaluations demonstrate that Antelope outperforms existing\nbaselines across multiple defensive mechanisms, underscoring its efficacy and\nversatility.\n", "date": "2024-12-11", "categories": []}
{"id": "2412.08394", "title": "Adversarial Purification by Consistency-aware Latent Space Optimization\n  on Data Manifolds", "abstract": "  Deep neural networks (DNNs) are vulnerable to adversarial samples crafted by\nadding imperceptible perturbations to clean data, potentially leading to\nincorrect and dangerous predictions. Adversarial purification has been an\neffective means to improve DNNs robustness by removing these perturbations\nbefore feeding the data into the model. However, it faces significant\nchallenges in preserving key structural and semantic information of data, as\nthe imperceptible nature of adversarial perturbations makes it hard to avoid\nover-correcting, which can destroy important information and degrade model\nperformance. In this paper, we break away from traditional adversarial\npurification methods by focusing on the clean data manifold. To this end, we\nreveal that samples generated by a well-trained generative model are close to\nclean ones but far from adversarial ones. Leveraging this insight, we propose\nConsistency Model-based Adversarial Purification (CMAP), which optimizes\nvectors within the latent space of a pre-trained consistency model to generate\nsamples for restoring clean data. Specifically, 1) we propose a\n\\textit{Perceptual consistency restoration} mechanism by minimizing the\ndiscrepancy between generated samples and input samples in both pixel and\nperceptual spaces. 2) To maintain the optimized latent vectors within the valid\ndata manifold, we introduce a \\textit{Latent distribution consistency\nconstraint} strategy to align generated samples with the clean data\ndistribution. 3) We also apply a \\textit{Latent vector consistency prediction}\nscheme via an ensemble approach to enhance prediction reliability. CMAP\nfundamentally addresses adversarial perturbations at their source, providing a\nrobust purification. Extensive experiments on CIFAR-10 and ImageNet-100 show\nthat our CMAP significantly enhances robustness against strong adversarial\nattacks while preserving high natural accuracy.\n", "date": "2024-12-11", "categories": []}
{"id": "2412.08608", "title": "AdvWave: Stealthy Adversarial Jailbreak Attack against Large\n  Audio-Language Models", "abstract": "  Recent advancements in large audio-language models (LALMs) have enabled\nspeech-based user interactions, significantly enhancing user experience and\naccelerating the deployment of LALMs in real-world applications. However,\nensuring the safety of LALMs is crucial to prevent risky outputs that may raise\nsocietal concerns or violate AI regulations. Despite the importance of this\nissue, research on jailbreaking LALMs remains limited due to their recent\nemergence and the additional technical challenges they present compared to\nattacks on DNN-based audio models. Specifically, the audio encoders in LALMs,\nwhich involve discretization operations, often lead to gradient shattering,\nhindering the effectiveness of attacks relying on gradient-based optimizations.\nThe behavioral variability of LALMs further complicates the identification of\neffective (adversarial) optimization targets. Moreover, enforcing stealthiness\nconstraints on adversarial audio waveforms introduces a reduced, non-convex\nfeasible solution space, further intensifying the challenges of the\noptimization process. To overcome these challenges, we develop AdvWave, the\nfirst jailbreak framework against LALMs. We propose a dual-phase optimization\nmethod that addresses gradient shattering, enabling effective end-to-end\ngradient-based optimization. Additionally, we develop an adaptive adversarial\ntarget search algorithm that dynamically adjusts the adversarial optimization\ntarget based on the response patterns of LALMs for specific queries. To ensure\nthat adversarial audio remains perceptually natural to human listeners, we\ndesign a classifier-guided optimization approach that generates adversarial\nnoise resembling common urban sounds. Extensive evaluations on multiple\nadvanced LALMs demonstrate that AdvWave outperforms baseline methods, achieving\na 40% higher average jailbreak attack success rate.\n", "date": "2024-12-11", "categories": []}
{"id": "2311.18403", "title": "Detecting and Corrupting Convolution-based Unlearnable Examples", "abstract": "  Convolution-based unlearnable examples (UEs) employ class-wise multiplicative\nconvolutional noise to training samples, severely compromising model\nperformance. This fire-new type of UEs have successfully countered all defense\nmechanisms against UEs. The failure of such defenses can be attributed to the\nabsence of norm constraints on convolutional noise, leading to severe blurring\nof image features. To address this, we first design an Edge Pixel-based\nDetector (EPD) to identify convolution-based UEs. Upon detection of them, we\npropose the first defense scheme against convolution-based UEs, COrrupting\nthese samples via random matrix multiplication by employing bilinear\nINterpolation (COIN) such that disrupting the distribution of class-wise\nmultiplicative noise. To evaluate the generalization of our proposed COIN, we\nnewly design two convolution-based UEs called VUDA and HUDA to expand the scope\nof convolution-based UEs. Extensive experiments demonstrate the effectiveness\nof detection scheme EPD and that our defense COIN outperforms 11\nstate-of-the-art (SOTA) defenses, achieving a significant improvement on the\nCIFAR and ImageNet datasets.\n", "date": "2023-11-30", "categories": []}
{"id": "2401.14707", "title": "AFD: Mitigating Feature Gap for Adversarial Robustness by Feature\n  Disentanglement", "abstract": "  Adversarial fine-tuning methods enhance adversarial robustness via\nfine-tuning the pre-trained model in an adversarial training manner. However,\nwe identify that some specific latent features of adversarial samples are\nconfused by adversarial perturbation and lead to an unexpectedly increasing gap\nbetween features in the last hidden layer of natural and adversarial samples.\nTo address this issue, we propose a disentanglement-based approach to\nexplicitly model and further remove the specific latent features. We introduce\na feature disentangler to separate out the specific latent features from the\nfeatures of the adversarial samples, thereby boosting robustness by eliminating\nthe specific latent features. Besides, we align clean features in the\npre-trained model with features of adversarial samples in the fine-tuned model,\nto benefit from the intrinsic features of natural samples. Empirical\nevaluations on three benchmark datasets demonstrate that our approach surpasses\nexisting adversarial fine-tuning methods and adversarial training baselines.\n", "date": "2024-01-26", "categories": []}
{"id": "2402.02095", "title": "Contrasting Adversarial Perturbations: The Space of Harmless\n  Perturbations", "abstract": "  Existing works have extensively studied adversarial examples, which are\nminimal perturbations that can mislead the output of deep neural networks\n(DNNs) while remaining imperceptible to humans. However, in this work, we\nreveal the existence of a harmless perturbation space, in which perturbations\ndrawn from this space, regardless of their magnitudes, leave the network output\nunchanged when applied to inputs. Essentially, the harmless perturbation space\nemerges from the usage of non-injective functions (linear or non-linear layers)\nwithin DNNs, enabling multiple distinct inputs to be mapped to the same output.\nFor linear layers with input dimensions exceeding output dimensions, any linear\ncombination of the orthogonal bases of the nullspace of the parameter\nconsistently yields no change in their output. For non-linear layers, the\nharmless perturbation space may expand, depending on the properties of the\nlayers and input samples. Inspired by this property of DNNs, we solve for a\nfamily of general perturbation spaces that are redundant for the DNN's\ndecision, and can be used to hide sensitive data and serve as a means of model\nidentification. Our work highlights the distinctive robustness of DNNs (i.e.,\nconsistency under large magnitude perturbations) in contrast to adversarial\nexamples (vulnerability for small imperceptible noises).\n", "date": "2024-02-03", "categories": []}
{"id": "2412.07078", "title": "Defensive Dual Masking for Robust Adversarial Defense", "abstract": "  The field of textual adversarial defenses has gained considerable attention\nin recent years due to the increasing vulnerability of natural language\nprocessing (NLP) models to adversarial attacks, which exploit subtle\nperturbations in input text to deceive models. This paper introduces the\nDefensive Dual Masking (DDM) algorithm, a novel approach designed to enhance\nmodel robustness against such attacks. DDM utilizes a unique adversarial\ntraining strategy where [MASK] tokens are strategically inserted into training\nsamples to prepare the model to handle adversarial perturbations more\neffectively. During inference, potentially adversarial tokens are dynamically\nreplaced with [MASK] tokens to neutralize potential threats while preserving\nthe core semantics of the input. The theoretical foundation of our approach is\nexplored, demonstrating how the selective masking mechanism strengthens the\nmodel's ability to identify and mitigate adversarial manipulations. Our\nempirical evaluation across a diverse set of benchmark datasets and attack\nmechanisms consistently shows that DDM outperforms state-of-the-art defense\ntechniques, improving model accuracy and robustness. Moreover, when applied to\nLarge Language Models (LLMs), DDM also enhances their resilience to adversarial\nattacks, providing a scalable defense mechanism for large-scale NLP\napplications.\n", "date": "2024-12-10", "categories": []}
{"id": "2412.07231", "title": "Adversarial Filtering Based Evasion and Backdoor Attacks to EEG-Based\n  Brain-Computer Interfaces", "abstract": "  A brain-computer interface (BCI) enables direct communication between the\nbrain and an external device. Electroencephalogram (EEG) is a common input\nsignal for BCIs, due to its convenience and low cost. Most research on\nEEG-based BCIs focuses on the accurate decoding of EEG signals, while ignoring\ntheir security. Recent studies have shown that machine learning models in BCIs\nare vulnerable to adversarial attacks. This paper proposes adversarial\nfiltering based evasion and backdoor attacks to EEG-based BCIs, which are very\neasy to implement. Experiments on three datasets from different BCI paradigms\ndemonstrated the effectiveness of our proposed attack approaches. To our\nknowledge, this is the first study on adversarial filtering for EEG-based BCIs,\nraising a new security concern and calling for more attention on the security\nof BCIs.\n", "date": "2024-12-10", "categories": []}
{"id": "2412.07253", "title": "CapGen:An Environment-Adaptive Generator of Adversarial Patches", "abstract": "  Adversarial patches, often used to provide physical stealth protection for\ncritical assets and assess perception algorithm robustness, usually neglect the\nneed for visual harmony with the background environment, making them easily\nnoticeable. Moreover, existing methods primarily concentrate on improving\nattack performance, disregarding the intricate dynamics of adversarial patch\nelements. In this work, we introduce the Camouflaged Adversarial Pattern\nGenerator (CAPGen), a novel approach that leverages specific base colors from\nthe surrounding environment to produce patches that seamlessly blend with their\nbackground for superior visual stealthiness while maintaining robust\nadversarial performance. We delve into the influence of both patterns (i.e.,\ncolor-agnostic texture information) and colors on the effectiveness of attacks\nfacilitated by patches, discovering that patterns exert a more pronounced\neffect on performance than colors. Based on these findings, we propose a rapid\ngeneration strategy for adversarial patches. This involves updating the colors\nof high-performance adversarial patches to align with those of the new\nenvironment, ensuring visual stealthiness without compromising adversarial\nimpact. This paper is the first to comprehensively examine the roles played by\npatterns and colors in the context of adversarial patches.\n", "date": "2024-12-10", "categories": []}
{"id": "2412.07274", "title": "A Generative Victim Model for Segmentation", "abstract": "  We find that the well-trained victim models (VMs), against which the attacks\nare generated, serve as fundamental prerequisites for adversarial attacks, i.e.\na segmentation VM is needed to generate attacks for segmentation. In this\ncontext, the victim model is assumed to be robust to achieve effective\nadversarial perturbation generation. Instead of focusing on improving the\nrobustness of the task-specific victim models, we shift our attention to image\ngeneration. From an image generation perspective, we derive a novel VM for\nsegmentation, aiming to generate adversarial perturbations for segmentation\ntasks without requiring models explicitly designed for image segmentation. Our\napproach to adversarial attack generation diverges from conventional white-box\nor black-box attacks, offering a fresh outlook on adversarial attack\nstrategies. Experiments show that our attack method is able to generate\neffective adversarial attacks with good transferability.\n", "date": "2024-12-10", "categories": []}
{"id": "2212.08341", "title": "Adversarial Example Defense via Perturbation Grading Strategy", "abstract": "  Deep Neural Networks have been widely used in many fields. However, studies\nhave shown that DNNs are easily attacked by adversarial examples, which have\ntiny perturbations and greatly mislead the correct judgment of DNNs.\nFurthermore, even if malicious attackers cannot obtain all the underlying model\nparameters, they can use adversarial examples to attack various DNN-based task\nsystems. Researchers have proposed various defense methods to protect DNNs,\nsuch as reducing the aggressiveness of adversarial examples by preprocessing or\nimproving the robustness of the model by adding modules. However, some defense\nmethods are only effective for small-scale examples or small perturbations but\nhave limited defense effects for adversarial examples with large perturbations.\nThis paper assigns different defense strategies to adversarial perturbations of\ndifferent strengths by grading the perturbations on the input examples.\nExperimental results show that the proposed method effectively improves defense\nperformance. In addition, the proposed method does not modify any task model,\nwhich can be used as a preprocessing module, which significantly reduces the\ndeployment cost in practical applications.\n", "date": "2022-12-16", "categories": []}
{"id": "2303.12481", "title": "Revisiting DeepFool: generalization and improvement", "abstract": "  Deep neural networks have been known to be vulnerable to adversarial\nexamples, which are inputs that are modified slightly to fool the network into\nmaking incorrect predictions. This has led to a significant amount of research\non evaluating the robustness of these networks against such perturbations. One\nparticularly important robustness metric is the robustness to minimal $\\ell_2$\nadversarial perturbations. However, existing methods for evaluating this\nrobustness metric are either computationally expensive or not very accurate. In\nthis paper, we introduce a new family of adversarial attacks that strike a\nbalance between effectiveness and computational efficiency. Our proposed\nattacks are generalizations of the well-known DeepFool (DF) attack, while they\nremain simple to understand and implement. We demonstrate that our attacks\noutperform existing methods in terms of both effectiveness and computational\nefficiency. Our proposed attacks are also suitable for evaluating the\nrobustness of large models and can be used to perform adversarial training (AT)\nto achieve state-of-the-art robustness to minimal $\\ell_2$ adversarial\nperturbations.\n", "date": "2023-03-22", "categories": []}
{"id": "2402.12329", "title": "Query-Based Adversarial Prompt Generation", "abstract": "  Recent work has shown it is possible to construct adversarial examples that\ncause an aligned language model to emit harmful strings or perform harmful\nbehavior. Existing attacks work either in the white-box setting (with full\naccess to the model weights), or through transferability: the phenomenon that\nadversarial examples crafted on one model often remain effective on other\nmodels. We improve on prior work with a query-based attack that leverages API\naccess to a remote language model to construct adversarial examples that cause\nthe model to emit harmful strings with (much) higher probability than with\ntransfer-only attacks. We validate our attack on GPT-3.5 and OpenAI's safety\nclassifier; we can cause GPT-3.5 to emit harmful strings that current transfer\nattacks fail at, and we can evade the safety classifier with nearly 100%\nprobability.\n", "date": "2024-02-19", "categories": []}
{"id": "2404.08255", "title": "Practical Region-level Attack against Segment Anything Models", "abstract": "  Segment Anything Models (SAM) have made significant advancements in image\nsegmentation, allowing users to segment target portions of an image with a\nsingle click (i.e., user prompt). Given its broad applications, the robustness\nof SAM against adversarial attacks is a critical concern. While recent works\nhave explored adversarial attacks against a pre-defined prompt/click, their\nthreat model is not yet realistic: (1) they often assume the user-click\nposition is known to the attacker (point-based attack), and (2) they often\noperate under a white-box setting with limited transferability. In this paper,\nwe propose a more practical region-level attack where attackers do not need to\nknow the precise user prompt. The attack remains effective as the user clicks\non any point on the target object in the image, hiding the object from SAM.\nAlso, by adapting a spectrum transformation method, we make the attack more\ntransferable under a black-box setting. Both control experiments and testing\nagainst real-world SAM services confirm its effectiveness.\n", "date": "2024-04-12", "categories": []}
{"id": "2412.03722", "title": "Optimal probabilistic feature shifts for reclassification in tree\n  ensembles", "abstract": "  In this paper we provide a novel mathematical optimization based methodology\nto perturb the features of a given observation to be re-classified, by a tree\nensemble classification rule, to a certain desired class. The method is based\non these facts: the most viable changes for an observation to reach the desired\nclass do not always coincide with the closest distance point (in the feature\nspace) of the target class; individuals put effort on a few number of features\nto reach the desired class; and each individual is endowed with a probability\nto change each of its features to a given value, which determines the overall\nprobability of changing to the target class. Putting all together, we provide\ndifferent methods to find the features where the individuals must exert effort\nto maximize the probability to reach the target class. Our method also allows\nus to rank the most important features in the tree-ensemble. The proposed\nmethodology is tested on a real dataset, validating the proposal.\n", "date": "2024-12-04", "categories": []}
{"id": "2412.05351", "title": "Towards Predicting the Success of Transfer-based Attacks by Quantifying\n  Shared Feature Representations", "abstract": "  Much effort has been made to explain and improve the success of\ntransfer-based attacks (TBA) on black-box computer vision models. This work\nprovides the first attempt at a priori prediction of attack success by\nidentifying the presence of vulnerable features within target models. Recent\nwork by Chen and Liu (2024) proposed the manifold attack model, a unifying\nframework proposing that successful TBA exist in a common manifold space. Our\nwork experimentally tests the common manifold space hypothesis by a new\nmethodology: first, projecting feature vectors from surrogate and target\nfeature extractors trained on ImageNet onto the same low-dimensional manifold;\nsecond, quantifying any observed structure similarities on the manifold; and\nfinally, by relating these observed similarities to the success of the TBA. We\nfind that shared feature representation moderately correlates with increased\nsuccess of TBA (\\r{ho}= 0.56). This method may be used to predict whether an\nattack will transfer without information of the model weights, training,\narchitecture or details of the attack. The results confirm the presence of\nshared feature representations between two feature extractors of different\nsizes and complexities, and demonstrate the utility of datasets from different\ntarget domains as test signals for interpreting black-box feature\nrepresentations.\n", "date": "2024-12-06", "categories": []}
{"id": "2412.05980", "title": "Anti-Reference: Universal and Immediate Defense Against Reference-Based\n  Generation", "abstract": "  Diffusion models have revolutionized generative modeling with their\nexceptional ability to produce high-fidelity images. However, misuse of such\npotent tools can lead to the creation of fake news or disturbing content\ntargeting individuals, resulting in significant social harm. In this paper, we\nintroduce Anti-Reference, a novel method that protects images from the threats\nposed by reference-based generation techniques by adding imperceptible\nadversarial noise to the images. We propose a unified loss function that\nenables joint attacks on fine-tuning-based customization methods,\nnon-fine-tuning customization methods, and human-centric driving methods. Based\non this loss, we train a Adversarial Noise Encoder to predict the noise or\ndirectly optimize the noise using the PGD method. Our method shows certain\ntransfer attack capabilities, effectively challenging both gray-box models and\nsome commercial APIs. Extensive experiments validate the performance of\nAnti-Reference, establishing a new benchmark in image security.\n", "date": "2024-12-08", "categories": []}
{"id": "2412.06149", "title": "An Effective and Resilient Backdoor Attack Framework against Deep Neural\n  Networks and Vision Transformers", "abstract": "  Recent studies have revealed the vulnerability of Deep Neural Network (DNN)\nmodels to backdoor attacks. However, existing backdoor attacks arbitrarily set\nthe trigger mask or use a randomly selected trigger, which restricts the\neffectiveness and robustness of the generated backdoor triggers. In this paper,\nwe propose a novel attention-based mask generation methodology that searches\nfor the optimal trigger shape and location. We also introduce a\nQuality-of-Experience (QoE) term into the loss function and carefully adjust\nthe transparency value of the trigger in order to make the backdoored samples\nto be more natural. To further improve the prediction accuracy of the victim\nmodel, we propose an alternating retraining algorithm in the backdoor injection\nprocess. The victim model is retrained with mixed poisoned datasets in even\niterations and with only benign samples in odd iterations. Besides, we launch\nthe backdoor attack under a co-optimized attack framework that alternately\noptimizes the backdoor trigger and backdoored model to further improve the\nattack performance. Apart from DNN models, we also extend our proposed attack\nmethod against vision transformers. We evaluate our proposed method with\nextensive experiments on VGG-Flower, CIFAR-10, GTSRB, CIFAR-100, and ImageNette\ndatasets. It is shown that we can increase the attack success rate by as much\nas 82\\% over baselines when the poison ratio is low and achieve a high QoE of\nthe backdoored samples. Our proposed backdoor attack framework also showcases\nrobustness against state-of-the-art backdoor defenses.\n", "date": "2024-12-09", "categories": []}
{"id": "2412.06181", "title": "Enhancing Adversarial Resistance in LLMs with Recursion", "abstract": "  The increasing integration of Large Language Models (LLMs) into society\nnecessitates robust defenses against vulnerabilities from jailbreaking and\nadversarial prompts. This project proposes a recursive framework for enhancing\nthe resistance of LLMs to manipulation through the use of prompt simplification\ntechniques. By increasing the transparency of complex and confusing adversarial\nprompts, the proposed method enables more reliable detection and prevention of\nmalicious inputs. Our findings attempt to address a critical problem in AI\nsafety and security, providing a foundation for the development of systems able\nto distinguish harmless inputs from prompts containing malicious intent. As\nLLMs continue to be used in diverse applications, the importance of such\nsafeguards will only grow.\n", "date": "2024-12-09", "categories": []}
{"id": "2412.06215", "title": "A Real-Time Defense Against Object Vanishing Adversarial Patch Attacks\n  for Object Detection in Autonomous Vehicles", "abstract": "  Autonomous vehicles (AVs) increasingly use DNN-based object detection models\nin vision-based perception. Correct detection and classification of obstacles\nis critical to ensure safe, trustworthy driving decisions. Adversarial patches\naim to fool a DNN with intentionally generated patterns concentrated in a\nlocalized region of an image. In particular, object vanishing patch attacks can\ncause object detection models to fail to detect most or all objects in a scene,\nposing a significant practical threat to AVs.\n  This work proposes ADAV (Adversarial Defense for Autonomous Vehicles), a\nnovel defense methodology against object vanishing patch attacks specifically\ndesigned for autonomous vehicles. Unlike existing defense methods which have\nhigh latency or are designed for static images, ADAV runs in real-time and\nleverages contextual information from prior frames in an AV's video feed. ADAV\nchecks if the object detector's output for the target frame is temporally\nconsistent with the output from a previous reference frame to detect the\npresence of a patch. If the presence of a patch is detected, ADAV uses\ngradient-based attribution to localize adversarial pixels that break temporal\nconsistency. This two stage procedure allows ADAV to efficiently process clean\ninputs, and both stages are optimized to be low latency. ADAV is evaluated\nusing real-world driving data from the Berkeley Deep Drive BDD100K dataset, and\ndemonstrates high adversarial and clean performance.\n", "date": "2024-12-09", "categories": []}
{"id": "2412.06219", "title": "Data Free Backdoor Attacks", "abstract": "  Backdoor attacks aim to inject a backdoor into a classifier such that it\npredicts any input with an attacker-chosen backdoor trigger as an\nattacker-chosen target class. Existing backdoor attacks require either\nretraining the classifier with some clean data or modifying the model's\narchitecture. As a result, they are 1) not applicable when clean data is\nunavailable, 2) less efficient when the model is large, and 3) less stealthy\ndue to architecture changes. In this work, we propose DFBA, a novel\nretraining-free and data-free backdoor attack without changing the model\narchitecture. Technically, our proposed method modifies a few parameters of a\nclassifier to inject a backdoor. Through theoretical analysis, we verify that\nour injected backdoor is provably undetectable and unremovable by various\nstate-of-the-art defenses under mild assumptions. Our evaluation on multiple\ndatasets further demonstrates that our injected backdoor: 1) incurs negligible\nclassification loss, 2) achieves 100% attack success rates, and 3) bypasses six\nexisting state-of-the-art defenses. Moreover, our comparison with a\nstate-of-the-art non-data-free backdoor attack shows our attack is more\nstealthy and effective against various defenses while achieving less\nclassification accuracy loss.\n", "date": "2024-12-09", "categories": []}
{"id": "2406.00918", "title": "Robustness of Practical Perceptual Hashing Algorithms to Hash-Evasion\n  and Hash-Inversion Attacks", "abstract": "  Perceptual hashing algorithms (PHAs) are widely used for identifying illegal\nonline content and are thus integral to various sensitive applications.\nHowever, due to their hasty deployment in real-world scenarios, their\nadversarial security has not been thoroughly evaluated. This paper assesses the\nsecurity of three widely utilized PHAs - PhotoDNA, PDQ, and NeuralHash -\nagainst hash-evasion and hash-inversion attacks. Contrary to existing\nliterature, our findings indicate that these PHAs demonstrate significant\nrobustness against such attacks. We provide an explanation for these differing\nresults, highlighting that the inherent robustness is partially due to the\nrandom hash variations characteristic of PHAs. Additionally, we propose a\ndefense method that enhances security by intentionally introducing\nperturbations into the hashes.\n", "date": "2024-06-03", "categories": []}
{"id": "2412.05010", "title": "Backdooring Outlier Detection Methods: A Novel Attack Approach", "abstract": "  There have been several efforts in backdoor attacks, but these have primarily\nfocused on the closed-set performance of classifiers (i.e., classification).\nThis has left a gap in addressing the threat to classifiers' open-set\nperformance, referred to as outlier detection in the literature. Reliable\noutlier detection is crucial for deploying classifiers in critical real-world\napplications such as autonomous driving and medical image analysis. First, we\nshow that existing backdoor attacks fall short in affecting the open-set\nperformance of classifiers, as they have been specifically designed to confuse\nintra-closed-set decision boundaries. In contrast, an effective backdoor attack\nfor outlier detection needs to confuse the decision boundary between the closed\nand open sets. Motivated by this, in this study, we propose BATOD, a novel\nBackdoor Attack targeting the Outlier Detection task. Specifically, we design\ntwo categories of triggers to shift inlier samples to outliers and vice versa.\nWe evaluate BATOD using various real-world datasets and demonstrate its\nsuperior ability to degrade the open-set performance of classifiers compared to\nprevious attacks, both before and after applying defenses.\n", "date": "2024-12-06", "categories": []}
{"id": "2412.03993", "title": "LaserGuider: A Laser Based Physical Backdoor Attack against Deep Neural\n  Networks", "abstract": "  Backdoor attacks embed hidden associations between triggers and targets in\ndeep neural networks (DNNs), causing them to predict the target when a trigger\nis present while maintaining normal behavior otherwise. Physical backdoor\nattacks, which use physical objects as triggers, are feasible but lack remote\ncontrol, temporal stealthiness, flexibility, and mobility. To overcome these\nlimitations, in this work, we propose a new type of backdoor triggers utilizing\nlasers that feature long-distance transmission and instant-imaging properties.\nBased on the laser-based backdoor triggers, we present a physical backdoor\nattack, called LaserGuider, which possesses remote control ability and achieves\nhigh temporal stealthiness, flexibility, and mobility. We also introduce a\nsystematic approach to optimize laser parameters for improving attack\neffectiveness. Our evaluation on traffic sign recognition DNNs, critical in\nautonomous vehicles, demonstrates that LaserGuider with three different\nlaser-based triggers achieves over 90% attack success rate with negligible\nimpact on normal inputs. Additionally, we release LaserMark, the first dataset\nof real world traffic signs stamped with physical laser spots, to support\nfurther research in backdoor attacks and defenses.\n", "date": "2024-12-05", "categories": []}
{"id": "2412.04415", "title": "Targeting the Core: A Simple and Effective Method to Attack RAG-based\n  Agents via Direct LLM Manipulation", "abstract": "  AI agents, powered by large language models (LLMs), have transformed\nhuman-computer interactions by enabling seamless, natural, and context-aware\ncommunication. While these advancements offer immense utility, they also\ninherit and amplify inherent safety risks such as bias, fairness,\nhallucinations, privacy breaches, and a lack of transparency. This paper\ninvestigates a critical vulnerability: adversarial attacks targeting the LLM\ncore within AI agents. Specifically, we test the hypothesis that a deceptively\nsimple adversarial prefix, such as \\textit{Ignore the document}, can compel\nLLMs to produce dangerous or unintended outputs by bypassing their contextual\nsafeguards. Through experimentation, we demonstrate a high attack success rate\n(ASR), revealing the fragility of existing LLM defenses. These findings\nemphasize the urgent need for robust, multi-layered security measures tailored\nto mitigate vulnerabilities at the LLM level and within broader agent-based\narchitectures.\n", "date": "2024-12-05", "categories": []}
{"id": "2404.12635", "title": "AED-PADA:Improving Generalizability of Adversarial Example Detection via\n  Principal Adversarial Domain Adaptation", "abstract": "  Adversarial example detection, which can be conveniently applied in many\nscenarios, is important in the area of adversarial defense. Unfortunately,\nexisting detection methods suffer from poor generalization performance, because\ntheir training process usually relies on the examples generated from a single\nknown adversarial attack and there exists a large discrepancy between the\ntraining and unseen testing adversarial examples. To address this issue, we\npropose a novel method, named Adversarial Example Detection via Principal\nAdversarial Domain Adaptation (AED-PADA). Specifically, our approach identifies\nthe Principal Adversarial Domains (PADs), i.e., a combination of features of\nthe adversarial examples generated by different attacks, which possesses a\nlarge portion of the entire adversarial feature space. Subsequently, we pioneer\nto exploit Multi-source Unsupervised Domain Adaptation in adversarial example\ndetection, with PADs as the source domains. Experimental results demonstrate\nthe superior generalization ability of our proposed AED-PADA. Note that this\nsuperiority is particularly achieved in challenging scenarios characterized by\nemploying the minimal magnitude constraint for the perturbations.\n", "date": "2024-04-19", "categories": []}
{"id": "2412.02323", "title": "Pay Attention to the Robustness of Chinese Minority Language Models!\n  Syllable-level Textual Adversarial Attack on Tibetan Script", "abstract": "  The textual adversarial attack refers to an attack method in which the\nattacker adds imperceptible perturbations to the original texts by elaborate\ndesign so that the NLP (natural language processing) model produces false\njudgments. This method is also used to evaluate the robustness of NLP models.\nCurrently, most of the research in this field focuses on English, and there is\nalso a certain amount of research on Chinese. However, to the best of our\nknowledge, there is little research targeting Chinese minority languages.\nTextual adversarial attacks are a new challenge for the information processing\nof Chinese minority languages. In response to this situation, we propose a\nTibetan syllable-level black-box textual adversarial attack called TSAttacker\nbased on syllable cosine distance and scoring mechanism. And then, we conduct\nTSAttacker on six models generated by fine-tuning two PLMs (pre-trained\nlanguage models) for three downstream tasks. The experiment results show that\nTSAttacker is effective and generates high-quality adversarial samples. In\naddition, the robustness of the involved models still has much room for\nimprovement.\n", "date": "2024-12-03", "categories": []}
{"id": "2412.02795", "title": "Hijacking Vision-and-Language Navigation Agents with Adversarial\n  Environmental Attacks", "abstract": "  Assistive embodied agents that can be instructed in natural language to\nperform tasks in open-world environments have the potential to significantly\nimpact labor tasks like manufacturing or in-home care -- benefiting the lives\nof those who come to depend on them. In this work, we consider how this benefit\nmight be hijacked by local modifications in the appearance of the agent's\noperating environment. Specifically, we take the popular Vision-and-Language\nNavigation (VLN) task as a representative setting and develop a whitebox\nadversarial attack that optimizes a 3D attack object's appearance to induce\ndesired behaviors in pretrained VLN agents that observe it in the environment.\nWe demonstrate that the proposed attack can cause VLN agents to ignore their\ninstructions and execute alternative actions after encountering the attack\nobject -- even for instructions and agent paths not considered when optimizing\nthe attack. For these novel settings, we find our attacks can induce\nearly-termination behaviors or divert an agent along an attacker-defined\nmulti-step trajectory. Under both conditions, environmental attacks\nsignificantly reduce agent capabilities to successfully follow user\ninstructions.\n", "date": "2024-12-03", "categories": []}
{"id": "2412.02803", "title": "Gaussian Splatting Under Attack: Investigating Adversarial Noise in 3D\n  Objects", "abstract": "  3D Gaussian Splatting has advanced radiance field reconstruction, enabling\nhigh-quality view synthesis and fast rendering in 3D modeling. While\nadversarial attacks on object detection models are well-studied for 2D images,\ntheir impact on 3D models remains underexplored. This work introduces the\nMasked Iterative Fast Gradient Sign Method (M-IFGSM), designed to generate\nadversarial noise targeting the CLIP vision-language model. M-IFGSM\nspecifically alters the object of interest by focusing perturbations on masked\nregions, degrading the performance of CLIP's zero-shot object detection\ncapability when applied to 3D models. Using eight objects from the Common\nObjects 3D (CO3D) dataset, we demonstrate that our method effectively reduces\nthe accuracy and confidence of the model, with adversarial noise being nearly\nimperceptible to human observers. The top-1 accuracy in original model renders\ndrops from 95.4\\% to 12.5\\% for train images and from 91.2\\% to 35.4\\% for test\nimages, with confidence levels reflecting this shift from true classification\nto misclassification, underscoring the risks of adversarial attacks on 3D\nmodels in applications such as autonomous driving, robotics, and surveillance.\nThe significance of this research lies in its potential to expose\nvulnerabilities in modern 3D vision models, including radiance fields,\nprompting the development of more robust defenses and security measures in\ncritical real-world applications.\n", "date": "2024-12-03", "categories": []}
{"id": "2412.03051", "title": "Less is More: A Stealthy and Efficient Adversarial Attack Method for\n  DRL-based Autonomous Driving Policies", "abstract": "  Despite significant advancements in deep reinforcement learning (DRL)-based\nautonomous driving policies, these policies still exhibit vulnerability to\nadversarial attacks. This vulnerability poses a formidable challenge to the\npractical deployment of these policies in autonomous driving. Designing\neffective adversarial attacks is an indispensable prerequisite for enhancing\nthe robustness of these policies. In view of this, we present a novel stealthy\nand efficient adversarial attack method for DRL-based autonomous driving\npolicies. Specifically, we introduce a DRL-based adversary designed to trigger\nsafety violations (e.g., collisions) by injecting adversarial samples at\ncritical moments. We model the attack as a mixed-integer optimization problem\nand formulate it as a Markov decision process. Then, we train the adversary to\nlearn the optimal policy for attacking at critical moments without domain\nknowledge. Furthermore, we introduce attack-related information and a\ntrajectory clipping method to enhance the learning capability of the adversary.\nFinally, we validate our method in an unprotected left-turn scenario across\ndifferent traffic densities. The experimental results show that our method\nachieves more than 90% collision rate within three attacks in most cases.\nFurthermore, our method achieves more than 130% improvement in attack\nefficiency compared to the unlimited attack method.\n", "date": "2024-12-04", "categories": []}
{"id": "2412.03453", "title": "Pre-trained Multiple Latent Variable Generative Models are good\n  defenders against Adversarial Attacks", "abstract": "  Attackers can deliberately perturb classifiers' input with subtle noise,\naltering final predictions. Among proposed countermeasures, adversarial\npurification employs generative networks to preprocess input images, filtering\nout adversarial noise. In this study, we propose specific generators, defined\nMultiple Latent Variable Generative Models (MLVGMs), for adversarial\npurification. These models possess multiple latent variables that naturally\ndisentangle coarse from fine features. Taking advantage of these properties, we\nautoencode images to maintain class-relevant information, while discarding and\nre-sampling any detail, including adversarial noise. The procedure is\ncompletely training-free, exploring the generalization abilities of pre-trained\nMLVGMs on the adversarial purification downstream task. Despite the lack of\nlarge models, trained on billions of samples, we show that smaller MLVGMs are\nalready competitive with traditional methods, and can be used as foundation\nmodels. Official code released at https://github.com/SerezD/gen_adversarial.\n", "date": "2024-12-04", "categories": []}
{"id": "2412.02343", "title": "Multi-Granularity Tibetan Textual Adversarial Attack Method Based on\n  Masked Language Model", "abstract": "  In social media, neural network models have been applied to hate speech\ndetection, sentiment analysis, etc., but neural network models are susceptible\nto adversarial attacks. For instance, in a text classification task, the\nattacker elaborately introduces perturbations to the original texts that hardly\nalter the original semantics in order to trick the model into making different\npredictions. By studying textual adversarial attack methods, the robustness of\nlanguage models can be evaluated and then improved. Currently, most of the\nresearch in this field focuses on English, and there is also a certain amount\nof research on Chinese. However, there is little research targeting Chinese\nminority languages. With the rapid development of artificial intelligence\ntechnology and the emergence of Chinese minority language models, textual\nadversarial attacks become a new challenge for the information processing of\nChinese minority languages. In response to this situation, we propose a\nmulti-granularity Tibetan textual adversarial attack method based on masked\nlanguage models called TSTricker. We utilize the masked language models to\ngenerate candidate substitution syllables or words, adopt the scoring mechanism\nto determine the substitution order, and then conduct the attack method on\nseveral fine-tuned victim models. The experimental results show that TSTricker\nreduces the accuracy of the classification models by more than 28.70% and makes\nthe classification models change the predictions of more than 90.60% of the\nsamples, which has an evidently higher attack effect than the baseline method.\n", "date": "2024-12-03", "categories": []}
{"id": "2308.07834", "title": "Simple and Efficient Partial Graph Adversarial Attack: A New Perspective", "abstract": "  As the study of graph neural networks becomes more intensive and\ncomprehensive, their robustness and security have received great research\ninterest. The existing global attack methods treat all nodes in the graph as\ntheir attack targets. Although existing methods have achieved excellent\nresults, there is still considerable space for improvement. The key problem is\nthat the current approaches rigidly follow the definition of global attacks.\nThey ignore an important issue, i.e., different nodes have different robustness\nand are not equally resilient to attacks. From a global attacker's view, we\nshould arrange the attack budget wisely, rather than wasting them on highly\nrobust nodes. To this end, we propose a totally new method named partial graph\nattack (PGA), which selects the vulnerable nodes as attack targets. First, to\nselect the vulnerable items, we propose a hierarchical target selection policy,\nwhich allows attackers to only focus on easy-to-attack nodes. Then, we propose\na cost-effective anchor-picking policy to pick the most promising anchors for\nadding or removing edges, and a more aggressive iterative greedy-based attack\nmethod to perform more efficient attacks. Extensive experimental results\ndemonstrate that PGA can achieve significant improvements in both attack effect\nand attack efficiency compared to other existing graph global attack methods.\n", "date": "2023-08-15", "categories": []}
{"id": "2401.02718", "title": "Calibration Attacks: A Comprehensive Study of Adversarial Attacks on\n  Model Confidence", "abstract": "  In this work, we highlight and perform a comprehensive study on calibration\nattacks, a form of adversarial attacks that aim to trap victim models to be\nheavily miscalibrated without altering their predicted labels, hence\nendangering the trustworthiness of the models and follow-up decision making\nbased on their confidence. We propose four typical forms of calibration\nattacks: underconfidence, overconfidence, maximum miscalibration, and random\nconfidence attacks, conducted in both black-box and white-box setups. We\ndemonstrate that the attacks are highly effective on both convolutional and\nattention-based models: with a small number of queries, they seriously skew\nconfidence without changing the predictive performance. Given the potential\ndanger, we further investigate the effectiveness of a wide range of adversarial\ndefence and recalibration methods, including our proposed defences specifically\ndesigned for calibration attacks to mitigate the harm. From the ECE and KS\nscores, we observe that there are still significant limitations in handling\ncalibration attacks. To the best of our knowledge, this is the first dedicated\nstudy that provides a comprehensive investigation on calibration-focused\nattacks. We hope this study helps attract more attention to these types of\nattacks and hence hamper their potential serious damages. To this end, this\nwork also provides detailed analyses to understand the characteristics of the\nattacks. Our code is available at https://github.com/PhenetOs/CalibrationAttack\n", "date": "2024-01-05", "categories": []}
{"id": "2401.07867", "title": "Authorship Obfuscation in Multilingual Machine-Generated Text Detection", "abstract": "  High-quality text generation capability of recent Large Language Models\n(LLMs) causes concerns about their misuse (e.g., in massive generation/spread\nof disinformation). Machine-generated text (MGT) detection is important to cope\nwith such threats. However, it is susceptible to authorship obfuscation (AO)\nmethods, such as paraphrasing, which can cause MGTs to evade detection. So far,\nthis was evaluated only in monolingual settings. Thus, the susceptibility of\nrecently proposed multilingual detectors is still unknown. We fill this gap by\ncomprehensively benchmarking the performance of 10 well-known AO methods,\nattacking 37 MGT detection methods against MGTs in 11 languages (i.e., 10\n$\\times$ 37 $\\times$ 11 = 4,070 combinations). We also evaluate the effect of\ndata augmentation on adversarial robustness using obfuscated texts. The results\nindicate that all tested AO methods can cause evasion of automated detection in\nall tested languages, where homoglyph attacks are especially successful.\nHowever, some of the AO methods severely damaged the text, making it no longer\nreadable or easily recognizable by humans (e.g., changed language, weird\ncharacters).\n", "date": "2024-01-15", "categories": []}
{"id": "2412.00341", "title": "Fusing Physics-Driven Strategies and Cross-Modal Adversarial Learning:\n  Toward Multi-Domain Applications", "abstract": "  The convergence of cross-modal adversarial learning and physics-driven\nmethods represents a cutting-edge direction for tackling challenges in complex\nmulti-modal tasks and scientific computing. This review focuses on\nsystematically analyzing how these two approaches can be synergistically\nintegrated to enhance performance and robustness across diverse application\ndomains. By addressing key obstacles such as modality discrepancies, limited\ndata availability, and insufficient model robustness, this paper highlights the\nrole of physics-based optimization frameworks in facilitating efficient and\ninterpretable adversarial perturbation generation. The review also explores\nsignificant advancements in cross-modal adversarial learning, including\napplications in tasks such as image cross-modal retrieval (e.g., infrared and\nRGB matching), scientific computing (e.g., solving partial differential\nequations), and optimization under physical consistency constraints in vision\nsystems. By examining theoretical foundations and experimental outcomes, this\nstudy demonstrates the potential of combining these approaches to handle\ncomplex scenarios and improve the security of multi-modal systems. Finally, we\noutline future directions, proposing a novel framework that unifies physical\nprinciples with adversarial optimization, providing a pathway for researchers\nto develop robust and adaptable cross-modal learning methods with both\ntheoretical and practical significance.\n", "date": "2024-11-30", "categories": []}
{"id": "2412.01101", "title": "Hiding Faces in Plain Sight: Defending DeepFakes by Disrupting Face\n  Detection", "abstract": "  This paper investigates the feasibility of a proactive DeepFake defense\nframework, {\\em FacePosion}, to prevent individuals from becoming victims of\nDeepFake videos by sabotaging face detection. The motivation stems from the\nreliance of most DeepFake methods on face detectors to automatically extract\nvictim faces from videos for training or synthesis (testing). Once the face\ndetectors malfunction, the extracted faces will be distorted or incorrect,\nsubsequently disrupting the training or synthesis of the DeepFake model. To\nachieve this, we adapt various adversarial attacks with a dedicated design for\nthis purpose and thoroughly analyze their feasibility. Based on FacePoison, we\nintroduce {\\em VideoFacePoison}, a strategy that propagates FacePoison across\nvideo frames rather than applying them individually to each frame. This\nstrategy can largely reduce the computational overhead while retaining the\nfavorable attack performance. Our method is validated on five face detectors,\nand extensive experiments against eleven different DeepFake models demonstrate\nthe effectiveness of disrupting face detectors to hinder DeepFake generation.\n", "date": "2024-12-02", "categories": []}
{"id": "2412.01154", "title": "R.I.P.: A Simple Black-box Attack on Continual Test-time Adaptation", "abstract": "  Test-time adaptation (TTA) has emerged as a promising solution to tackle the\ncontinual domain shift in machine learning by allowing model parameters to\nchange at test time, via self-supervised learning on unlabeled testing data. At\nthe same time, it unfortunately opens the door to unforeseen vulnerabilities\nfor degradation over time. Through a simple theoretical continual TTA model, we\nsuccessfully identify a risk in the sampling process of testing data that could\neasily degrade the performance of a continual TTA model. We name this risk as\nReusing of Incorrect Prediction (RIP) that TTA attackers can employ or as a\nresult of the unintended query from general TTA users. The risk posed by RIP is\nalso highly realistic, as it does not require prior knowledge of model\nparameters or modification of testing samples. This simple requirement makes\nRIP as the first black-box TTA attack algorithm that stands out from existing\nwhite-box attempts. We extensively benchmark the performance of the most recent\ncontinual TTA approaches when facing the RIP attack, providing insights on its\nsuccess, and laying out potential roadmaps that could enhance the resilience of\nfuture continual TTA systems.\n", "date": "2024-12-02", "categories": []}
{"id": "2412.01363", "title": "Exploring the Robustness of AI-Driven Tools in Digital Forensics: A\n  Preliminary Study", "abstract": "  Nowadays, many tools are used to facilitate forensic tasks about data\nextraction and data analysis. In particular, some tools leverage Artificial\nIntelligence (AI) to automatically label examined data into specific categories\n(\\ie, drugs, weapons, nudity). However, this raises a serious concern about the\nrobustness of the employed AI algorithms against adversarial attacks. Indeed,\nsome people may need to hide specific data to AI-based digital forensics tools,\nthus manipulating the content so that the AI system does not recognize the\noffensive/prohibited content and marks it at as suspicious to the analyst. This\ncould be seen as an anti-forensics attack scenario. For this reason, we\nanalyzed two of the most important forensics tools employing AI for data\nclassification: Magnet AI, used by Magnet Axiom, and Excire Photo AI, used by\nX-Ways Forensics. We made preliminary tests using about $200$ images, other\n$100$ sent in $3$ chats about pornography and teenage nudity, drugs and weapons\nto understand how the tools label them. Moreover, we loaded some deepfake\nimages (images generated by AI forging real ones) of some actors to understand\nif they would be classified in the same category as the original images. From\nour preliminary study, we saw that the AI algorithm is not robust enough, as we\nexpected since these topics are still open research problems. For example, some\nsexual images were not categorized as nudity, and some deepfakes were\ncategorized as the same real person, while the human eye can see the clear\nnudity image or catch the difference between the deepfakes. Building on these\nresults and other state-of-the-art works, we provide some suggestions for\nimproving how digital forensics analysis tool leverage AI and their robustness\nagainst adversarial attacks or different scenarios than the trained one.\n", "date": "2024-12-02", "categories": []}
{"id": "2412.01369", "title": "Behavior Backdoor for Deep Learning Models", "abstract": "  The various post-processing methods for deep-learning-based models, such as\nquantification, pruning, and fine-tuning, play an increasingly important role\nin artificial intelligence technology, with pre-train large models as one of\nthe main development directions. However, this popular series of\npost-processing behaviors targeting pre-training deep models has become a\nbreeding ground for new adversarial security issues. In this study, we take the\nfirst step towards ``behavioral backdoor'' attack, which is defined as a\nbehavior-triggered backdoor model training procedure, to reveal a new paradigm\nof backdoor attacks. In practice, we propose the first pipeline of implementing\nbehavior backdoor, i.e., the Quantification Backdoor (QB) attack, upon\nexploiting model quantification method as the set trigger. Specifically, to\nadapt the optimization goal of behavior backdoor, we introduce the\nbehavior-driven backdoor object optimizing method by a bi-target behavior\nbackdoor training loss, thus we could guide the poisoned model optimization\ndirection. To update the parameters across multiple models, we adopt the\naddress-shared backdoor model training, thereby the gradient information could\nbe utilized for multimodel collaborative optimization. Extensive experiments\nhave been conducted on different models, datasets, and tasks, demonstrating the\neffectiveness of this novel backdoor attack and its potential application\nthreats.\n", "date": "2024-12-02", "categories": []}
{"id": "2412.01495", "title": "Adversarial Attacks on Hyperbolic Networks", "abstract": "  As hyperbolic deep learning grows in popularity, so does the need for\nadversarial robustness in the context of such a non-Euclidean geometry. To this\nend, this paper proposes hyperbolic alternatives to the commonly used FGM and\nPGD adversarial attacks. Through interpretable synthetic benchmarks and\nexperiments on existing datasets, we show how the existing and newly proposed\nattacks differ. Moreover, we investigate the differences in adversarial\nrobustness between Euclidean and fully hyperbolic networks. We find that these\nnetworks suffer from different types of vulnerabilities and that the newly\nproposed hyperbolic attacks cannot address these differences. Therefore, we\nconclude that the shifts in adversarial robustness are due to the models\nlearning distinct patterns resulting from their different geometries.\n", "date": "2024-12-02", "categories": []}
{"id": "2412.01646", "title": "Robust and Transferable Backdoor Attacks Against Deep Image Compression\n  With Selective Frequency Prior", "abstract": "  Recent advancements in deep learning-based compression techniques have\nsurpassed traditional methods. However, deep neural networks remain vulnerable\nto backdoor attacks, where pre-defined triggers induce malicious behaviors.\nThis paper introduces a novel frequency-based trigger injection model for\nlaunching backdoor attacks with multiple triggers on learned image compression\nmodels. Inspired by the widely used DCT in compression codecs, triggers are\nembedded in the DCT domain. We design attack objectives tailored to diverse\nscenarios, including: 1) degrading compression quality in terms of bit-rate and\nreconstruction accuracy; 2) targeting task-driven measures like face\nrecognition and semantic segmentation. To improve training efficiency, we\npropose a dynamic loss function that balances loss terms with fewer\nhyper-parameters, optimizing attack objectives effectively. For advanced\nscenarios, we evaluate the attack's resistance to defensive preprocessing and\npropose a two-stage training schedule with robust frequency selection to\nenhance resilience. To improve cross-model and cross-domain transferability for\ndownstream tasks, we adjust the classification boundary in the attack loss\nduring training. Experiments show that our trigger injection models, combined\nwith minor modifications to encoder parameters, successfully inject multiple\nbackdoors and their triggers into a single compression model, demonstrating\nstrong performance and versatility. (*Due to the notification of arXiv \"The\nAbstract field cannot be longer than 1,920 characters\", the appeared Abstract\nis shortened. For the full Abstract, please download the Article.)\n", "date": "2024-12-02", "categories": []}
{"id": "2412.01725", "title": "Attacks on multimodal models", "abstract": "  Today, models capable of working with various modalities simultaneously in a\nchat format are gaining increasing popularity. Despite this, there is an issue\nof potential attacks on these models, especially considering that many of them\ninclude open-source components. It is important to study whether the\nvulnerabilities of these components are inherited and how dangerous this can be\nwhen using such models in the industry. This work is dedicated to researching\nvarious types of attacks on such models and evaluating their generalization\ncapabilities. Modern VLM models (LLaVA, BLIP, etc.) often use pre-trained parts\nfrom other models, so the main part of this research focuses on them,\nspecifically on the CLIP architecture and its image encoder (CLIP-ViT) and\nvarious patch attack variations for it.\n", "date": "2024-12-02", "categories": []}
{"id": "2212.00727", "title": "Adversarial Artifact Detection in EEG-Based Brain-Computer Interfaces", "abstract": "  Machine learning has achieved great success in electroencephalogram (EEG)\nbased brain-computer interfaces (BCIs). Most existing BCI research focused on\nimproving its accuracy, but few had considered its security. Recent studies,\nhowever, have shown that EEG-based BCIs are vulnerable to adversarial attacks,\nwhere small perturbations added to the input can cause misclassification.\nDetection of adversarial examples is crucial to both the understanding of this\nphenomenon and the defense. This paper, for the first time, explores\nadversarial detection in EEG-based BCIs. Experiments on two EEG datasets using\nthree convolutional neural networks were performed to verify the performances\nof multiple detection approaches. We showed that both white-box and black-box\nattacks can be detected, and the former are easier to detect.\n", "date": "2022-11-28", "categories": []}
{"id": "2312.02220", "title": "QuantAttack: Exploiting Dynamic Quantization to Attack Vision\n  Transformers", "abstract": "  In recent years, there has been a significant trend in deep neural networks\n(DNNs), particularly transformer-based models, of developing ever-larger and\nmore capable models. While they demonstrate state-of-the-art performance, their\ngrowing scale requires increased computational resources (e.g., GPUs with\ngreater memory capacity). To address this problem, quantization techniques\n(i.e., low-bit-precision representation and matrix multiplication) have been\nproposed. Most quantization techniques employ a static strategy in which the\nmodel parameters are quantized, either during training or inference, without\nconsidering the test-time sample. In contrast, dynamic quantization techniques,\nwhich have become increasingly popular, adapt during inference based on the\ninput provided, while maintaining full-precision performance. However, their\ndynamic behavior and average-case performance assumption makes them vulnerable\nto a novel threat vector -- adversarial attacks that target the model's\nefficiency and availability. In this paper, we present QuantAttack, a novel\nattack that targets the availability of quantized models, slowing down the\ninference, and increasing memory usage and energy consumption. We show that\ncarefully crafted adversarial examples, which are designed to exhaust the\nresources of the operating system, can trigger worst-case performance. In our\nexperiments, we demonstrate the effectiveness of our attack on vision\ntransformers on a wide range of tasks, both uni-modal and multi-modal. We also\nexamine the effect of different attack variants (e.g., a universal\nperturbation) and the transferability between different models.\n", "date": "2023-12-03", "categories": []}
{"id": "2402.10527", "title": "Assessing biomedical knowledge robustness in large language models by\n  query-efficient sampling attacks", "abstract": "  The increasing depth of parametric domain knowledge in large language models\n(LLMs) is fueling their rapid deployment in real-world applications.\nUnderstanding model vulnerabilities in high-stakes and knowledge-intensive\ntasks is essential for quantifying the trustworthiness of model predictions and\nregulating their use. The recent discovery of named entities as adversarial\nexamples (i.e. adversarial entities) in natural language processing tasks\nraises questions about their potential impact on the knowledge robustness of\npre-trained and finetuned LLMs in high-stakes and specialized domains. We\nexamined the use of type-consistent entity substitution as a template for\ncollecting adversarial entities for billion-parameter LLMs with biomedical\nknowledge. To this end, we developed an embedding-space attack based on\npowerscaled distance-weighted sampling to assess the robustness of their\nbiomedical knowledge with a low query budget and controllable coverage. Our\nmethod has favorable query efficiency and scaling over alternative approaches\nbased on random sampling and blackbox gradient-guided search, which we\ndemonstrated for adversarial distractor generation in biomedical question\nanswering. Subsequent failure mode analysis uncovered two regimes of\nadversarial entities on the attack surface with distinct characteristics and we\nshowed that entity substitution attacks can manipulate token-wise Shapley value\nexplanations, which become deceptive in this setting. Our approach complements\nstandard evaluations for high-capacity models and the results highlight the\nbrittleness of domain knowledge in LLMs.\n", "date": "2024-02-16", "categories": []}
{"id": "2411.18000", "title": "Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for\n  Jailbreaking Vision-Language Models", "abstract": "  Despite inheriting security measures from underlying language models,\nVision-Language Models (VLMs) may still be vulnerable to safety alignment\nissues. Through empirical analysis, we uncover two critical findings:\nscenario-matched images can significantly amplify harmful outputs, and contrary\nto common assumptions in gradient-based attacks, minimal loss values do not\nguarantee optimal attack effectiveness. Building on these insights, we\nintroduce MLAI (Multi-Loss Adversarial Images), a novel jailbreak framework\nthat leverages scenario-aware image generation for semantic alignment, exploits\nflat minima theory for robust adversarial image selection, and employs\nmulti-image collaborative attacks for enhanced effectiveness. Extensive\nexperiments demonstrate MLAI's significant impact, achieving attack success\nrates of 77.75% on MiniGPT-4 and 82.80% on LLaVA-2, substantially outperforming\nexisting methods by margins of 34.37% and 12.77% respectively. Furthermore,\nMLAI shows considerable transferability to commercial black-box VLMs, achieving\nup to 60.11% success rate. Our work reveals fundamental visual vulnerabilities\nin current VLMs safety mechanisms and underscores the need for stronger\ndefenses. Warning: This paper contains potentially harmful example text.\n", "date": "2024-11-27", "categories": []}
{"id": "2411.18956", "title": "Random Sampling for Diffusion-based Adversarial Purification", "abstract": "  Denoising Diffusion Probabilistic Models (DDPMs) have gained great attention\nin adversarial purification. Current diffusion-based works focus on designing\neffective condition-guided mechanisms while ignoring a fundamental problem,\ni.e., the original DDPM sampling is intended for stable generation, which may\nnot be the optimal solution for adversarial purification. Inspired by the\nstability of the Denoising Diffusion Implicit Model (DDIM), we propose an\nopposite sampling scheme called random sampling. In brief, random sampling will\nsample from a random noisy space during each diffusion process, while DDPM and\nDDIM sampling will continuously sample from the adjacent or original noisy\nspace. Thus, random sampling obtains more randomness and achieves stronger\nrobustness against adversarial attacks. Correspondingly, we also introduce a\nnovel mediator conditional guidance to guarantee the consistency of the\nprediction under the purified image and clean image input. To expand awareness\nof guided diffusion purification, we conduct a detailed evaluation with\ndifferent sampling methods and our random sampling achieves an impressive\nimprovement in multiple settings. Leveraging mediator-guided random sampling,\nwe also establish a baseline method named DiffAP, which significantly\noutperforms state-of-the-art (SOTA) approaches in performance and defensive\nstability. Remarkably, under strong attack, our DiffAP even achieves a more\nthan 20% robustness advantage with 10$\\times$ sampling acceleration.\n", "date": "2024-11-28", "categories": []}
{"id": "2411.19075", "title": "LADDER: Multi-objective Backdoor Attack via Evolutionary Algorithm", "abstract": "  Current black-box backdoor attacks in convolutional neural networks formulate\nattack objective(s) as single-objective optimization problems in single domain.\nDesigning triggers in single domain harms semantics and trigger robustness as\nwell as introduces visual and spectral anomaly. This work proposes a\nmulti-objective black-box backdoor attack in dual domains via evolutionary\nalgorithm (LADDER), the first instance of achieving multiple attack objectives\nsimultaneously by optimizing triggers without requiring prior knowledge about\nvictim model. In particular, we formulate LADDER as a multi-objective\noptimization problem (MOP) and solve it via multi-objective evolutionary\nalgorithm (MOEA). MOEA maintains a population of triggers with trade-offs among\nattack objectives and uses non-dominated sort to drive triggers toward optimal\nsolutions. We further apply preference-based selection to MOEA to exclude\nimpractical triggers. We state that LADDER investigates a new dual-domain\nperspective for trigger stealthiness by minimizing the anomaly between clean\nand poisoned samples in the spectral domain. Lastly, the robustness against\npreprocessing operations is achieved by pushing triggers to low-frequency\nregions. Extensive experiments comprehensively showcase that LADDER achieves\nattack effectiveness of at least 99%, attack robustness with 90.23% (50.09%\nhigher than state-of-the-art attacks on average), superior natural stealthiness\n(1.12x to 196.74x improvement) and excellent spectral stealthiness (8.45x\nenhancement) as compared to current stealthy attacks by the average $l_2$-norm\nacross 5 public datasets.\n", "date": "2024-11-28", "categories": []}
{"id": "2411.17936", "title": "Stealthy Multi-Task Adversarial Attacks", "abstract": "  Deep Neural Networks exhibit inherent vulnerabilities to adversarial attacks,\nwhich can significantly compromise their outputs and reliability. While\nexisting research primarily focuses on attacking single-task scenarios or\nindiscriminately targeting all tasks in multi-task environments, we investigate\nselectively targeting one task while preserving performance in others within a\nmulti-task framework. This approach is motivated by varying security priorities\namong tasks in real-world applications, such as autonomous driving, where\nmisinterpreting critical objects (e.g., signs, traffic lights) poses a greater\nsecurity risk than minor depth miscalculations. Consequently, attackers may\nhope to target security-sensitive tasks while avoiding non-critical tasks from\nbeing compromised, thus evading being detected before compromising crucial\nfunctions. In this paper, we propose a method for the stealthy multi-task\nattack framework that utilizes multiple algorithms to inject imperceptible\nnoise into the input. This novel method demonstrates remarkable efficacy in\ncompromising the target task while simultaneously maintaining or even enhancing\nperformance across non-targeted tasks - a criterion hitherto unexplored in the\nfield. Additionally, we introduce an automated approach for searching the\nweighting factors in the loss function, further enhancing attack efficiency.\nExperimental results validate our framework's ability to successfully attack\nthe target task while preserving the performance of non-targeted tasks. The\nautomated loss function weight searching method demonstrates comparable\nefficacy to manual tuning, establishing a state-of-the-art multi-task attack\nframework.\n", "date": "2024-11-26", "categories": []}
{"id": "2411.17959", "title": "Adversarial Training in Low-Label Regimes with Margin-Based\n  Interpolation", "abstract": "  Adversarial training has emerged as an effective approach to train robust\nneural network models that are resistant to adversarial attacks, even in\nlow-label regimes where labeled data is scarce. In this paper, we introduce a\nnovel semi-supervised adversarial training approach that enhances both\nrobustness and natural accuracy by generating effective adversarial examples.\nOur method begins by applying linear interpolation between clean and\nadversarial examples to create interpolated adversarial examples that cross\ndecision boundaries by a controlled margin. This sample-aware strategy tailors\nadversarial examples to the characteristics of each data point, enabling the\nmodel to learn from the most informative perturbations. Additionally, we\npropose a global epsilon scheduling strategy that progressively adjusts the\nupper bound of perturbation strengths during training. The combination of these\nstrategies allows the model to develop increasingly complex decision boundaries\nwith better robustness and natural accuracy. Empirical evaluations show that\nour approach effectively enhances performance against various adversarial\nattacks, such as PGD and AutoAttack.\n", "date": "2024-11-27", "categories": []}
{"id": "2312.07130", "title": "Harnessing LLM to Attack LLM-Guarded Text-to-Image Models", "abstract": "  To prevent Text-to-Image (T2I) models from generating unethical images,\npeople deploy safety filters to block inappropriate drawing prompts. Previous\nworks have employed token replacement to search adversarial prompts that\nattempt to bypass these filters, but they have become ineffective as\nnonsensical tokens fail semantic logic checks. In this paper, we approach\nadversarial prompts from a different perspective. We demonstrate that\nrephrasing a drawing intent into multiple benign descriptions of individual\nvisual components can obtain an effective adversarial prompt. We propose a\nLLM-piloted multi-agent method named DACA to automatically complete intended\nrephrasing. Our method successfully bypasses the safety filters of DALL-E 3 and\nMidjourney to generate the intended images, achieving success rates of up to\n76.7% and 64% in the one-time attack, and 98% and 84% in the re-use attack,\nrespectively. We open-source our code and dataset on [this\nlink](https://github.com/researchcode003/DACA).\n", "date": "2023-12-12", "categories": []}
{"id": "2404.00629", "title": "Against The Achilles' Heel: A Survey on Red Teaming for Generative\n  Models", "abstract": "  Generative models are rapidly gaining popularity and being integrated into\neveryday applications, raising concerns over their safe use as various\nvulnerabilities are exposed. In light of this, the field of red teaming is\nundergoing fast-paced growth, highlighting the need for a comprehensive survey\ncovering the entire pipeline and addressing emerging topics. Our extensive\nsurvey, which examines over 120 papers, introduces a taxonomy of fine-grained\nattack strategies grounded in the inherent capabilities of language models.\nAdditionally, we have developed the \"searcher\" framework to unify various\nautomatic red teaming approaches. Moreover, our survey covers novel areas\nincluding multimodal attacks and defenses, risks around LLM-based agents,\noverkill of harmless queries, and the balance between harmlessness and\nhelpfulness.\n", "date": "2024-03-31", "categories": []}
{"id": "2407.03946", "title": "TrackPGD: Efficient Adversarial Attack using Object Binary Masks against\n  Robust Transformer Trackers", "abstract": "  Adversarial perturbations can deceive neural networks by adding small,\nimperceptible noise to the input. Recent object trackers with transformer\nbackbones have shown strong performance on tracking datasets, but their\nadversarial robustness has not been thoroughly evaluated. While transformer\ntrackers are resilient to black-box attacks, existing white-box adversarial\nattacks are not universally applicable against these new transformer trackers\ndue to differences in backbone architecture. In this work, we introduce\nTrackPGD, a novel white-box attack that utilizes predicted object binary masks\nto target robust transformer trackers. Built upon the powerful segmentation\nattack SegPGD, our proposed TrackPGD effectively influences the decisions of\ntransformer-based trackers. Our method addresses two primary challenges in\nadapting a segmentation attack for trackers: limited class numbers and extreme\npixel class imbalance. TrackPGD uses the same number of iterations as other\nattack methods for tracker networks and produces competitive adversarial\nexamples that mislead transformer and non-transformer trackers such as\nMixFormerM, OSTrackSTS, TransT-SEG, and RTS on datasets including VOT2022STS,\nDAVIS2016, UAV123, and GOT-10k.\n", "date": "2024-07-04", "categories": []}
{"id": "2410.19863", "title": "Breaking the Illusion: Real-world Challenges for Adversarial Patches in\n  Object Detection", "abstract": "  Adversarial attacks pose a significant threat to the robustness and\nreliability of machine learning systems, particularly in computer vision\napplications. This study investigates the performance of adversarial patches\nfor the YOLO object detection network in the physical world. Two attacks were\ntested: a patch designed to be placed anywhere within the scene - global patch,\nand another patch intended to partially overlap with specific object targeted\nfor removal from detection - local patch. Various factors such as patch size,\nposition, rotation, brightness, and hue were analyzed to understand their\nimpact on the effectiveness of the adversarial patches. The results reveal a\nnotable dependency on these parameters, highlighting the challenges in\nmaintaining attack efficacy in real-world conditions. Learning to align\ndigitally applied transformation parameters with those measured in the real\nworld still results in up to a 64\\% discrepancy in patch performance. These\nfindings underscore the importance of understanding environmental influences on\nadversarial attacks, which can inform the development of more robust defenses\nfor practical machine learning applications.\n", "date": "2024-10-23", "categories": []}
{"id": "2411.16996", "title": "CRASH: Challenging Reinforcement-Learning Based Adversarial Scenarios\n  For Safety Hardening", "abstract": "  Ensuring the safety of autonomous vehicles (AVs) requires identifying rare\nbut critical failure cases that on-road testing alone cannot discover.\nHigh-fidelity simulations provide a scalable alternative, but automatically\ngenerating realistic and diverse traffic scenarios that can effectively stress\ntest AV motion planners remains a key challenge. This paper introduces CRASH -\nChallenging Reinforcement-learning based Adversarial scenarios for Safety\nHardening - an adversarial deep reinforcement learning framework to address\nthis issue. First CRASH can control adversarial Non Player Character (NPC)\nagents in an AV simulator to automatically induce collisions with the Ego\nvehicle, falsifying its motion planner. We also propose a novel approach, that\nwe term safety hardening, which iteratively refines the motion planner by\nsimulating improvement scenarios against adversarial agents, leveraging the\nfailure cases to strengthen the AV stack. CRASH is evaluated on a simplified\ntwo-lane highway scenario, demonstrating its ability to falsify both rule-based\nand learning-based planners with collision rates exceeding 90%. Additionally,\nsafety hardening reduces the Ego vehicle's collision rate by 26%. While\npreliminary, these results highlight RL-based safety hardening as a promising\napproach for scenario-driven simulation testing for autonomous vehicles.\n", "date": "2024-11-26", "categories": []}
{"id": "2411.17026", "title": "RED: Robust Environmental Design", "abstract": "  The classification of road signs by autonomous systems, especially those\nreliant on visual inputs, is highly susceptible to adversarial attacks.\nTraditional approaches to mitigating such vulnerabilities have focused on\nenhancing the robustness of classification models. In contrast, this paper\nadopts a fundamentally different strategy aimed at increasing robustness\nthrough the redesign of road signs themselves. We propose an attacker-agnostic\nlearning scheme to automatically design road signs that are robust to a wide\narray of patch-based attacks. Empirical tests conducted in both digital and\nphysical environments demonstrate that our approach significantly reduces\nvulnerability to patch attacks, outperforming existing techniques.\n", "date": "2024-11-26", "categories": []}
{"id": "2411.17468", "title": "Adversarial Bounding Boxes Generation (ABBG) Attack against Visual\n  Object Trackers", "abstract": "  Adversarial perturbations aim to deceive neural networks into predicting\ninaccurate results. For visual object trackers, adversarial attacks have been\ndeveloped to generate perturbations by manipulating the outputs. However,\ntransformer trackers predict a specific bounding box instead of an object\ncandidate list, which limits the applicability of many existing attack\nscenarios. To address this issue, we present a novel white-box approach to\nattack visual object trackers with transformer backbones using only one\nbounding box. From the tracker predicted bounding box, we generate a list of\nadversarial bounding boxes and compute the adversarial loss for those bounding\nboxes. Experimental results demonstrate that our simple yet effective attack\noutperforms existing attacks against several robust transformer trackers,\nincluding TransT-M, ROMTrack, and MixFormer, on popular benchmark tracking\ndatasets such as GOT-10k, UAV123, and VOT2022STS.\n", "date": "2024-11-26", "categories": []}
{"id": "2208.09449", "title": "A Novel Plug-and-Play Approach for Adversarially Robust Generalization", "abstract": "  In this work, we propose a robust framework that employs adversarially robust\ntraining to safeguard the ML models against perturbed testing data. Our\ncontributions can be seen from both computational and statistical perspectives.\nFirstly, from a computational/optimization point of view, we derive the\nready-to-use exact solution for several widely used loss functions with a\nvariety of norm constraints on adversarial perturbation for various supervised\nand unsupervised ML problems, including regression, classification, two-layer\nneural networks, graphical models, and matrix completion. The solutions are\neither in closed-form, or an easily tractable optimization problem such as 1-D\nconvex optimization, semidefinite programming, difference of convex programming\nor a sorting-based algorithm. Secondly, from statistical/generalization\nviewpoint, using some of these results, we derive novel bounds of the\nadversarial Rademacher complexity for various problems, which entails new\ngeneralization bounds. Thirdly, we perform some sanity-check experiments on\nreal-world datasets for supervised problems such as regression and\nclassification, as well as for unsupervised problems such as matrix completion\nand learning graphical models, with very little computational overhead.\n", "date": "2022-08-19", "categories": []}
{"id": "2306.00353", "title": "Constructing Semantics-Aware Adversarial Examples with a Probabilistic\n  Perspective", "abstract": "  We propose a probabilistic perspective on adversarial examples, allowing us\nto embed subjective understanding of semantics as a distribution into the\nprocess of generating adversarial examples, in a principled manner. Despite\nsignificant pixel-level modifications compared to traditional adversarial\nattacks, our method preserves the overall semantics of the image, making the\nchanges difficult for humans to detect. This extensive pixel-level modification\nenhances our method's ability to deceive classifiers designed to defend against\nadversarial attacks. Our empirical findings indicate that the proposed methods\nachieve higher success rates in circumventing adversarial defense mechanisms,\nwhile remaining difficult for human observers to detect.\n", "date": "2023-06-01", "categories": []}
{"id": "2404.07921", "title": "AmpleGCG: Learning a Universal and Transferable Generative Model of\n  Adversarial Suffixes for Jailbreaking Both Open and Closed LLMs", "abstract": "  As large language models (LLMs) become increasingly prevalent and integrated\ninto autonomous systems, ensuring their safety is imperative. Despite\nsignificant strides toward safety alignment, recent work\nGCG~\\citep{zou2023universal} proposes a discrete token optimization algorithm\nand selects the single suffix with the lowest loss to successfully jailbreak\naligned LLMs. In this work, we first discuss the drawbacks of solely picking\nthe suffix with the lowest loss during GCG optimization for jailbreaking and\nuncover the missed successful suffixes during the intermediate steps. Moreover,\nwe utilize those successful suffixes as training data to learn a generative\nmodel, named AmpleGCG, which captures the distribution of adversarial suffixes\ngiven a harmful query and enables the rapid generation of hundreds of suffixes\nfor any harmful queries in seconds. AmpleGCG achieves near 100\\% attack success\nrate (ASR) on two aligned LLMs (Llama-2-7B-chat and Vicuna-7B), surpassing two\nstrongest attack baselines. More interestingly, AmpleGCG also transfers\nseamlessly to attack different models, including closed-source LLMs, achieving\na 99\\% ASR on the latest GPT-3.5. To summarize, our work amplifies the impact\nof GCG by training a generative model of adversarial suffixes that is universal\nto any harmful queries and transferable from attacking open-source LLMs to\nclosed-source LLMs. In addition, it can generate 200 adversarial suffixes for\none harmful query in only 4 seconds, rendering it more challenging to defend.\n", "date": "2024-04-11", "categories": []}
{"id": "2406.13352", "title": "AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks\n  and Defenses for LLM Agents", "abstract": "  AI agents aim to solve complex tasks by combining text-based reasoning with\nexternal tool calls. Unfortunately, AI agents are vulnerable to prompt\ninjection attacks where data returned by external tools hijacks the agent to\nexecute malicious tasks. To measure the adversarial robustness of AI agents, we\nintroduce AgentDojo, an evaluation framework for agents that execute tools over\nuntrusted data. To capture the evolving nature of attacks and defenses,\nAgentDojo is not a static test suite, but rather an extensible environment for\ndesigning and evaluating new agent tasks, defenses, and adaptive attacks. We\npopulate the environment with 97 realistic tasks (e.g., managing an email\nclient, navigating an e-banking website, or making travel bookings), 629\nsecurity test cases, and various attack and defense paradigms from the\nliterature. We find that AgentDojo poses a challenge for both attacks and\ndefenses: state-of-the-art LLMs fail at many tasks (even in the absence of\nattacks), and existing prompt injection attacks break some security properties\nbut not all. We hope that AgentDojo can foster research on new design\nprinciples for AI agents that solve common tasks in a reliable and robust\nmanner.. We release the code for AgentDojo at\nhttps://github.com/ethz-spylab/agentdojo.\n", "date": "2024-06-19", "categories": []}
{"id": "2411.14449", "title": "Unlearn to Relearn Backdoors: Deferred Backdoor Functionality Attacks on\n  Deep Learning Models", "abstract": "  Deep learning models are vulnerable to backdoor attacks, where adversaries\ninject malicious functionality during training that activates on trigger inputs\nat inference time. Extensive research has focused on developing stealthy\nbackdoor attacks to evade detection and defense mechanisms. However, these\napproaches still have limitations that leave the door open for detection and\nmitigation due to their inherent design to cause malicious behavior in the\npresence of a trigger. To address this limitation, we introduce Deferred\nActivated Backdoor Functionality (DABF), a new paradigm in backdoor attacks.\nUnlike conventional attacks, DABF initially conceals its backdoor, producing\nbenign outputs even when triggered. This stealthy behavior allows DABF to\nbypass multiple detection and defense methods, remaining undetected during\ninitial inspections. The backdoor functionality is strategically activated only\nafter the model undergoes subsequent updates, such as retraining on benign\ndata. DABF attacks exploit the common practice in the life cycle of machine\nlearning models to perform model updates and fine-tuning after initial\ndeployment. To implement DABF attacks, we approach the problem by making the\nunlearning of the backdoor fragile, allowing it to be easily cancelled and\nsubsequently reactivate the backdoor functionality. To achieve this, we propose\na novel two-stage training scheme, called DeferBad. Our extensive experiments\nacross various fine-tuning scenarios, backdoor attack types, datasets, and\nmodel architectures demonstrate the effectiveness and stealthiness of DeferBad.\n", "date": "2024-11-10", "categories": []}
{"id": "2411.15222", "title": "Rethinking the Intermediate Features in Adversarial Attacks: Misleading\n  Robotic Models via Adversarial Distillation", "abstract": "  Language-conditioned robotic learning has significantly enhanced robot\nadaptability by enabling a single model to execute diverse tasks in response to\nverbal commands. Despite these advancements, security vulnerabilities within\nthis domain remain largely unexplored. This paper addresses this gap by\nproposing a novel adversarial prompt attack tailored to language-conditioned\nrobotic models. Our approach involves crafting a universal adversarial prefix\nthat induces the model to perform unintended actions when added to any original\nprompt. We demonstrate that existing adversarial techniques exhibit limited\neffectiveness when directly transferred to the robotic domain due to the\ninherent robustness of discretized robotic action spaces. To overcome this\nchallenge, we propose to optimize adversarial prefixes based on continuous\naction representations, circumventing the discretization process. Additionally,\nwe identify the beneficial impact of intermediate features on adversarial\nattacks and leverage the negative gradient of intermediate self-attention\nfeatures to further enhance attack efficacy. Extensive experiments on VIMA\nmodels across 13 robot manipulation tasks validate the superiority of our\nmethod over existing approaches and demonstrate its transferability across\ndifferent model variants.\n", "date": "2024-11-21", "categories": []}
{"id": "2411.15439", "title": "Twin Trigger Generative Networks for Backdoor Attacks against Object\n  Detection", "abstract": "  Object detectors, which are widely used in real-world applications, are\nvulnerable to backdoor attacks. This vulnerability arises because many users\nrely on datasets or pre-trained models provided by third parties due to\nconstraints on data and resources. However, most research on backdoor attacks\nhas focused on image classification, with limited investigation into object\ndetection. Furthermore, the triggers for most existing backdoor attacks on\nobject detection are manually generated, requiring prior knowledge and\nconsistent patterns between the training and inference stages. This approach\nmakes the attacks either easy to detect or difficult to adapt to various\nscenarios. To address these limitations, we propose novel twin trigger\ngenerative networks in the frequency domain to generate invisible triggers for\nimplanting stealthy backdoors into models during training, and visible triggers\nfor steady activation during inference, making the attack process difficult to\ntrace. Specifically, for the invisible trigger generative network, we deploy a\nGaussian smoothing layer and a high-frequency artifact classifier to enhance\nthe stealthiness of backdoor implantation in object detectors. For the visible\ntrigger generative network, we design a novel alignment loss to optimize the\nvisible triggers so that they differ from the original patterns but still align\nwith the malicious activation behavior of the invisible triggers. Extensive\nexperimental results and analyses prove the possibility of using different\ntriggers in the training stage and the inference stage, and demonstrate the\nattack effectiveness of our proposed visible trigger and invisible trigger\ngenerative networks, significantly reducing the mAP_0.5 of the object detectors\nby 70.0% and 84.5%, including YOLOv5 and YOLOv7 with different settings,\nrespectively.\n", "date": "2024-11-23", "categories": []}
{"id": "2411.15878", "title": "ExAL: An Exploration Enhanced Adversarial Learning Algorithm", "abstract": "  Adversarial learning is critical for enhancing model robustness, aiming to\ndefend against adversarial attacks that jeopardize machine learning systems.\nTraditional methods often lack efficient mechanisms to explore diverse\nadversarial perturbations, leading to limited model resilience. Inspired by\ngame-theoretic principles, where adversarial dynamics are analyzed through\nframeworks like Nash equilibrium, exploration mechanisms in such setups allow\nfor the discovery of diverse strategies, enhancing system robustness. However,\nexisting adversarial learning methods often fail to incorporate structured\nexploration effectively, reducing their ability to improve model defense\ncomprehensively. To address these challenges, we propose a novel\nExploration-enhanced Adversarial Learning Algorithm (ExAL), leveraging the\nExponentially Weighted Momentum Particle Swarm Optimizer (EMPSO) to generate\noptimized adversarial perturbations. ExAL integrates exploration-driven\nmechanisms to discover perturbations that maximize impact on the model's\ndecision boundary while preserving structural coherence in the data. We\nevaluate the performance of ExAL on the MNIST Handwritten Digits and Blended\nMalware datasets. Experimental results demonstrate that ExAL significantly\nenhances model resilience to adversarial attacks by improving robustness\nthrough adversarial learning.\n", "date": "2024-11-24", "categories": []}
{"id": "2411.16162", "title": "Sparse patches adversarial attacks via extrapolating point-wise\n  information", "abstract": "  Sparse and patch adversarial attacks were previously shown to be applicable\nin realistic settings and are considered a security risk to autonomous systems.\nSparse adversarial perturbations constitute a setting in which the adversarial\nperturbations are limited to affecting a relatively small number of points in\nthe input. Patch adversarial attacks denote the setting where the sparse\nattacks are limited to a given structure, i.e., sparse patches with a given\nshape and number. However, previous patch adversarial attacks do not\nsimultaneously optimize multiple patches' locations and perturbations. This\nwork suggests a novel approach for sparse patches adversarial attacks via\npoint-wise trimming dense adversarial perturbations. Our approach enables\nsimultaneous optimization of multiple sparse patches' locations and\nperturbations for any given number and shape. Moreover, our approach is also\napplicable for standard sparse adversarial attacks, where we show that it\nsignificantly improves the state-of-the-art over multiple extensive settings. A\nreference implementation of the proposed method and the reported experiments is\nprovided at \\url{https://github.com/yanemcovsky/SparsePatches.git}\n", "date": "2024-11-25", "categories": []}
{"id": "2411.16437", "title": "Privacy Protection in Personalized Diffusion Models via Targeted\n  Cross-Attention Adversarial Attack", "abstract": "  The growing demand for customized visual content has led to the rise of\npersonalized text-to-image (T2I) diffusion models. Despite their remarkable\npotential, they pose significant privacy risk when misused for malicious\npurposes. In this paper, we propose a novel and efficient adversarial attack\nmethod, Concept Protection by Selective Attention Manipulation (CoPSAM) which\ntargets only the cross-attention layers of a T2I diffusion model. For this\npurpose, we carefully construct an imperceptible noise to be added to clean\nsamples to get their adversarial counterparts. This is obtained during the\nfine-tuning process by maximizing the discrepancy between the corresponding\ncross-attention maps of the user-specific token and the class-specific token,\nrespectively. Experimental validation on a subset of CelebA-HQ face images\ndataset demonstrates that our approach outperforms existing methods. Besides\nthis, our method presents two important advantages derived from the qualitative\nevaluation: (i) we obtain better protection results for lower noise levels than\nour competitors; and (ii) we protect the content from unauthorized use thereby\nprotecting the individual's identity from potential misuse.\n", "date": "2024-11-25", "categories": []}
{"id": "2411.16622", "title": "Imperceptible Adversarial Examples in the Physical World", "abstract": "  Adversarial examples in the digital domain against deep learning-based\ncomputer vision models allow for perturbations that are imperceptible to human\neyes. However, producing similar adversarial examples in the physical world has\nbeen difficult due to the non-differentiable image distortion functions in\nvisual sensing systems. The existing algorithms for generating physically\nrealizable adversarial examples often loosen their definition of adversarial\nexamples by allowing unbounded perturbations, resulting in obvious or even\nstrange visual patterns. In this work, we make adversarial examples\nimperceptible in the physical world using a straight-through estimator (STE,\na.k.a. BPDA). We employ STE to overcome the non-differentiability -- applying\nexact, non-differentiable distortions in the forward pass of the\nbackpropagation step, and using the identity function in the backward pass. Our\ndifferentiable rendering extension to STE also enables imperceptible\nadversarial patches in the physical world. Using printout photos, and\nexperiments in the CARLA simulator, we show that STE enables fast generation of\n$\\ell_\\infty$ bounded adversarial examples despite the non-differentiable\ndistortions. To the best of our knowledge, this is the first work demonstrating\nimperceptible adversarial examples bounded by small $\\ell_\\infty$ norms in the\nphysical world that force zero classification accuracy in the global\nperturbation threat model and cause near-zero ($4.22\\%$) AP50 in object\ndetection in the patch perturbation threat model. We urge the community to\nre-evaluate the threat of adversarial examples in the physical world.\n", "date": "2024-11-25", "categories": []}
{"id": "2402.08648", "title": "Generating Universal Adversarial Perturbations for Quantum Classifiers", "abstract": "  Quantum Machine Learning (QML) has emerged as a promising field of research,\naiming to leverage the capabilities of quantum computing to enhance existing\nmachine learning methodologies. Recent studies have revealed that, like their\nclassical counterparts, QML models based on Parametrized Quantum Circuits\n(PQCs) are also vulnerable to adversarial attacks. Moreover, the existence of\nUniversal Adversarial Perturbations (UAPs) in the quantum domain has been\ndemonstrated theoretically in the context of quantum classifiers. In this work,\nwe introduce QuGAP: a novel framework for generating UAPs for quantum\nclassifiers. We conceptualize the notion of additive UAPs for PQC-based\nclassifiers and theoretically demonstrate their existence. We then utilize\ngenerative models (QuGAP-A) to craft additive UAPs and experimentally show that\nquantum classifiers are susceptible to such attacks. Moreover, we formulate a\nnew method for generating unitary UAPs (QuGAP-U) using quantum generative\nmodels and a novel loss function based on fidelity constraints. We evaluate the\nperformance of the proposed framework and show that our method achieves\nstate-of-the-art misclassification rates, while maintaining high fidelity\nbetween legitimate and adversarial samples.\n", "date": "2024-02-13", "categories": []}
{"id": "2405.08363", "title": "UnMarker: A Universal Attack on Defensive Image Watermarking", "abstract": "  Reports regarding the misuse of Generative AI (GenAI) to create deepfakes are\nfrequent. Defensive watermarking enables GenAI providers to hide fingerprints\nin their images and use them later for deepfake detection. Yet, its potential\nhas not been fully explored. We present UnMarker -- the first practical\nuniversal attack on defensive watermarking. Unlike existing attacks, UnMarker\nrequires no detector feedback, no unrealistic knowledge of the watermarking\nscheme or similar models, and no advanced denoising pipelines that may not be\navailable. Instead, being the product of an in-depth analysis of the\nwatermarking paradigm revealing that robust schemes must construct their\nwatermarks in the spectral amplitudes, UnMarker employs two novel adversarial\noptimizations to disrupt the spectra of watermarked images, erasing the\nwatermarks. Evaluations against SOTA schemes prove UnMarker's effectiveness. It\nnot only defeats traditional schemes while retaining superior quality compared\nto existing attacks but also breaks semantic watermarks that alter an image's\nstructure, reducing the best detection rate to $43\\%$ and rendering them\nuseless. To our knowledge, UnMarker is the first practical attack on semantic\nwatermarks, which have been deemed the future of defensive watermarking. Our\nfindings show that defensive watermarking is not a viable defense against\ndeepfakes, and we urge the community to explore alternatives.\n", "date": "2024-05-14", "categories": []}
{"id": "2209.15179", "title": "Physical Adversarial Attack meets Computer Vision: A Decade Survey", "abstract": "  Despite the impressive achievements of Deep Neural Networks (DNNs) in\ncomputer vision, their vulnerability to adversarial attacks remains a critical\nconcern. Extensive research has demonstrated that incorporating sophisticated\nperturbations into input images can lead to a catastrophic degradation in DNNs'\nperformance. This perplexing phenomenon not only exists in the digital space\nbut also in the physical world. Consequently, it becomes imperative to evaluate\nthe security of DNNs-based systems to ensure their safe deployment in\nreal-world scenarios, particularly in security-sensitive applications. To\nfacilitate a profound understanding of this topic, this paper presents a\ncomprehensive overview of physical adversarial attacks. Firstly, we distill\nfour general steps for launching physical adversarial attacks. Building upon\nthis foundation, we uncover the pervasive role of artifacts carrying\nadversarial perturbations in the physical world. These artifacts influence each\nstep. To denote them, we introduce a new term: adversarial medium. Then, we\ntake the first step to systematically evaluate the performance of physical\nadversarial attacks, taking the adversarial medium as a first attempt. Our\nproposed evaluation metric, hiPAA, comprises six perspectives: Effectiveness,\nStealthiness, Robustness, Practicability, Aesthetics, and Economics. We also\nprovide comparative results across task categories, together with insightful\nobservations and suggestions for future research directions.\n", "date": "2022-09-30", "categories": []}
{"id": "2310.00616", "title": "Towards Understanding Adversarial Transferability in Federated Learning", "abstract": "  We investigate a specific security risk in FL: a group of malicious clients\nhas impacted the model during training by disguising their identities and\nacting as benign clients but later switching to an adversarial role. They use\ntheir data, which was part of the training set, to train a substitute model and\nconduct transferable adversarial attacks against the federated model. This type\nof attack is subtle and hard to detect because these clients initially appear\nto be benign.\n  The key question we address is: How robust is the FL system to such covert\nattacks, especially compared to traditional centralized learning systems? We\nempirically show that the proposed attack imposes a high security risk to\ncurrent FL systems. By using only 3\\% of the client's data, we achieve the\nhighest attack rate of over 80\\%. To further offer a full understanding of the\nchallenges the FL system faces in transferable attacks, we provide a\ncomprehensive analysis over the transfer robustness of FL across a spectrum of\nconfigurations. Surprisingly, FL systems show a higher level of robustness than\ntheir centralized counterparts, especially when both systems are equally good\nat handling regular, non-malicious data.\n  We attribute this increased robustness to two main factors: 1) Decentralized\nData Training: Each client trains the model on its own data, reducing the\noverall impact of any single malicious client. 2) Model Update Averaging: The\nupdates from each client are averaged together, further diluting any malicious\nalterations. Both practical experiments and theoretical analysis support our\nconclusions. This research not only sheds light on the resilience of FL systems\nagainst hidden attacks but also raises important considerations for their\nfuture application and development.\n", "date": "2023-10-01", "categories": []}
{"id": "2311.00207", "title": "Magmaw: Modality-Agnostic Adversarial Attacks on Machine Learning-Based\n  Wireless Communication Systems", "abstract": "  Machine Learning (ML) has been instrumental in enabling joint transceiver\noptimization by merging all physical layer blocks of the end-to-end wireless\ncommunication systems. Although there have been a number of adversarial attacks\non ML-based wireless systems, the existing methods do not provide a\ncomprehensive view including multi-modality of the source data, common physical\nlayer protocols, and wireless domain constraints. This paper proposes Magmaw, a\nnovel wireless attack methodology capable of generating universal adversarial\nperturbations for any multimodal signal transmitted over a wireless channel. We\nfurther introduce new objectives for adversarial attacks on downstream\napplications. We adopt the widely-used defenses to verify the resilience of\nMagmaw. For proof-of-concept evaluation, we build a real-time wireless attack\nplatform using a software-defined radio system. Experimental results\ndemonstrate that Magmaw causes significant performance degradation even in the\npresence of strong defense mechanisms. Furthermore, we validate the performance\nof Magmaw in two case studies: encrypted communication channel and channel\nmodality-based ML model.\n", "date": "2023-11-01", "categories": []}
{"id": "2407.11424", "title": "Model Inversion Attacks Through Target-Specific Conditional Diffusion\n  Models", "abstract": "  Model inversion attacks (MIAs) aim to reconstruct private images from a\ntarget classifier's training set, thereby raising privacy concerns in AI\napplications. Previous GAN-based MIAs tend to suffer from inferior generative\nfidelity due to GAN's inherent flaws and biased optimization within latent\nspace. To alleviate these issues, leveraging on diffusion models' remarkable\nsynthesis capabilities, we propose Diffusion-based Model Inversion (Diff-MI)\nattacks. Specifically, we introduce a novel target-specific conditional\ndiffusion model (CDM) to purposely approximate target classifier's private\ndistribution and achieve superior accuracy-fidelity balance. Our method\ninvolves a two-step learning paradigm. Step-1 incorporates the target\nclassifier into the entire CDM learning under a pretrain-then-finetune fashion,\nwith creating pseudo-labels as model conditions in pretraining and adjusting\nspecified layers with image predictions in fine-tuning. Step-2 presents an\niterative image reconstruction method, further enhancing the attack performance\nthrough a combination of diffusion priors and target knowledge. Additionally,\nwe propose an improved max-margin loss that replaces the hard max with top-k\nmaxes, fully leveraging feature information and soft labels from the target\nclassifier. Extensive experiments demonstrate that Diff-MI significantly\nimproves generative fidelity with an average decrease of 20\\% in FID while\nmaintaining competitive attack accuracy compared to state-of-the-art methods\nacross various datasets and models. Our code is available at:\n\\url{https://github.com/Ouxiang-Li/Diff-MI}.\n", "date": "2024-07-16", "categories": []}
{"id": "2411.13778", "title": "A Survey on Adversarial Robustness of LiDAR-based Machine Learning\n  Perception in Autonomous Vehicles", "abstract": "  In autonomous driving, the combination of AI and vehicular technology offers\ngreat potential. However, this amalgamation comes with vulnerabilities to\nadversarial attacks. This survey focuses on the intersection of Adversarial\nMachine Learning (AML) and autonomous systems, with a specific focus on\nLiDAR-based systems. We comprehensively explore the threat landscape,\nencompassing cyber-attacks on sensors and adversarial perturbations.\nAdditionally, we investigate defensive strategies employed in countering these\nthreats. This paper endeavors to present a concise overview of the challenges\nand advances in securing autonomous driving systems against adversarial\nthreats, emphasizing the need for robust defenses to ensure safety and\nsecurity.\n", "date": "2024-11-21", "categories": []}
{"id": "2411.14263", "title": "Generating Realistic Adversarial Examples for Business Processes using\n  Variational Autoencoders", "abstract": "  In predictive process monitoring, predictive models are vulnerable to\nadversarial attacks, where input perturbations can lead to incorrect\npredictions. Unlike in computer vision, where these perturbations are designed\nto be imperceptible to the human eye, the generation of adversarial examples in\npredictive process monitoring poses unique challenges. Minor changes to the\nactivity sequences can create improbable or even impossible scenarios to occur\ndue to underlying constraints such as regulatory rules or process constraints.\nTo address this, we focus on generating realistic adversarial examples tailored\nto the business process context, in contrast to the imperceptible, pixel-level\nchanges commonly seen in computer vision adversarial attacks. This paper\nintroduces two novel latent space attacks, which generate adversaries by adding\nnoise to the latent space representation of the input data, rather than\ndirectly modifying the input attributes. These latent space methods are\ndomain-agnostic and do not rely on process-specific knowledge, as we restrict\nthe generation of adversarial examples to the learned class-specific data\ndistributions by directly perturbing the latent space representation of the\nbusiness process executions. We evaluate these two latent space methods with\nsix other adversarial attacking methods on eleven real-life event logs and four\npredictive models. The first three attacking methods directly permute the\nactivities of the historically observed business process executions. The fourth\nmethod constrains the adversarial examples to lie within the same data\ndistribution as the original instances, by projecting the adversarial examples\nto the original data distribution.\n", "date": "2024-11-21", "categories": []}
{"id": "2411.14351", "title": "Indiscriminate Disruption of Conditional Inference on Multivariate\n  Gaussians", "abstract": "  The multivariate Gaussian distribution underpins myriad operations-research,\ndecision-analytic, and machine-learning models (e.g., Bayesian optimization,\nGaussian influence diagrams, and variational autoencoders). However, despite\nrecent advances in adversarial machine learning (AML), inference for Gaussian\nmodels in the presence of an adversary is notably understudied. Therefore, we\nconsider a self-interested attacker who wishes to disrupt a decisionmaker's\nconditional inference and subsequent actions by corrupting a set of evidentiary\nvariables. To avoid detection, the attacker also desires the attack to appear\nplausible wherein plausibility is determined by the density of the corrupted\nevidence. We consider white- and grey-box settings such that the attacker has\ncomplete and incomplete knowledge about the decisionmaker's underlying\nmultivariate Gaussian distribution, respectively. Select instances are shown to\nreduce to quadratic and stochastic quadratic programs, and structural\nproperties are derived to inform solution methods. We assess the impact and\nefficacy of these attacks in three examples, including, real estate evaluation,\ninterest rate estimation and signals processing. Each example leverages an\nalternative underlying model, thereby highlighting the attacks' broad\napplicability. Through these applications, we also juxtapose the behavior of\nthe white- and grey-box attacks to understand how uncertainty and structure\naffect attacker behavior.\n", "date": "2024-11-21", "categories": []}
{"id": "2411.14424", "title": "Learning Fair Robustness via Domain Mixup", "abstract": "  Adversarial training is one of the predominant techniques for training\nclassifiers that are robust to adversarial attacks. Recent work, however has\nfound that adversarial training, which makes the overall classifier robust, it\ndoes not necessarily provide equal amount of robustness for all classes. In\nthis paper, we propose the use of mixup for the problem of learning fair robust\nclassifiers, which can provide similar robustness across all classes.\nSpecifically, the idea is to mix inputs from the same classes and perform\nadversarial training on mixed up inputs. We present a theoretical analysis of\nthis idea for the case of linear classifiers and show that mixup combined with\nadversarial training can provably reduce the class-wise robustness disparity.\nThis method not only contributes to reducing the disparity in class-wise\nadversarial risk, but also the class-wise natural risk. Complementing our\ntheoretical analysis, we also provide experimental results on both synthetic\ndata and the real world dataset (CIFAR-10), which shows improvement in class\nwise disparities for both natural and adversarial risks.\n", "date": "2024-11-21", "categories": []}
{"id": "2411.12071", "title": "Theoretical Corrections and the Leveraging of Reinforcement Learning to\n  Enhance Triangle Attack", "abstract": "  Adversarial examples represent a serious issue for the application of machine\nlearning models in many sensitive domains. For generating adversarial examples,\ndecision based black-box attacks are one of the most practical techniques as\nthey only require query access to the model. One of the most recently proposed\nstate-of-the-art decision based black-box attacks is Triangle Attack (TA). In\nthis paper, we offer a high-level description of TA and explain potential\ntheoretical limitations. We then propose a new decision based black-box attack,\nTriangle Attack with Reinforcement Learning (TARL). Our new attack addresses\nthe limits of TA by leveraging reinforcement learning. This creates an attack\nthat can achieve similar, if not better, attack accuracy than TA with half as\nmany queries on state-of-the-art classifiers and defenses across ImageNet and\nCIFAR-10.\n", "date": "2024-11-18", "categories": []}
{"id": "2411.12473", "title": "NMT-Obfuscator Attack: Ignore a sentence in translation with only one\n  word", "abstract": "  Neural Machine Translation systems are used in diverse applications due to\ntheir impressive performance. However, recent studies have shown that these\nsystems are vulnerable to carefully crafted small perturbations to their\ninputs, known as adversarial attacks. In this paper, we propose a new type of\nadversarial attack against NMT models. In this attack, we find a word to be\nadded between two sentences such that the second sentence is ignored and not\ntranslated by the NMT model. The word added between the two sentences is such\nthat the whole adversarial text is natural in the source language. This type of\nattack can be harmful in practical scenarios since the attacker can hide\nmalicious information in the automatic translation made by the target NMT\nmodel. Our experiments show that different NMT models and translation tasks are\nvulnerable to this type of attack. Our attack can successfully force the NMT\nmodels to ignore the second part of the input in the translation for more than\n50% of all cases while being able to maintain low perplexity for the whole\ninput.\n", "date": "2024-11-19", "categories": []}
{"id": "2411.12575", "title": "Stochastic BIQA: Median Randomized Smoothing for Certified Blind Image\n  Quality Assessment", "abstract": "  Most modern No-Reference Image-Quality Assessment (NR-IQA) metrics are based\non neural networks vulnerable to adversarial attacks. Attacks on such metrics\nlead to incorrect image/video quality predictions, which poses significant\nrisks, especially in public benchmarks. Developers of image processing\nalgorithms may unfairly increase the score of a target IQA metric without\nimproving the actual quality of the adversarial image. Although some empirical\ndefenses for IQA metrics were proposed, they do not provide theoretical\nguarantees and may be vulnerable to adaptive attacks. This work focuses on\ndeveloping a provably robust no-reference IQA metric. Our method is based on\nMedian Smoothing (MS) combined with an additional convolution denoiser with\nranking loss to improve the SROCC and PLCC scores of the defended IQA metric.\nCompared with two prior methods on three datasets, our method exhibited\nsuperior SROCC and PLCC scores while maintaining comparable certified\nguarantees.\n", "date": "2024-11-19", "categories": []}
{"id": "2403.05100", "title": "Exploring the Adversarial Frontier: Quantifying Robustness via\n  Adversarial Hypervolume", "abstract": "  The escalating threat of adversarial attacks on deep learning models,\nparticularly in security-critical fields, has underscored the need for robust\ndeep learning systems. Conventional robustness evaluations have relied on\nadversarial accuracy, which measures a model's performance under a specific\nperturbation intensity. However, this singular metric does not fully\nencapsulate the overall resilience of a model against varying degrees of\nperturbation. To address this gap, we propose a new metric termed adversarial\nhypervolume, assessing the robustness of deep learning models comprehensively\nover a range of perturbation intensities from a multi-objective optimization\nstandpoint. This metric allows for an in-depth comparison of defense mechanisms\nand recognizes the trivial improvements in robustness afforded by less potent\ndefensive strategies. Additionally, we adopt a novel training algorithm that\nenhances adversarial robustness uniformly across various perturbation\nintensities, in contrast to methods narrowly focused on optimizing adversarial\naccuracy. Our extensive empirical studies validate the effectiveness of the\nadversarial hypervolume metric, demonstrating its ability to reveal subtle\ndifferences in robustness that adversarial accuracy overlooks. This research\ncontributes a new measure of robustness and establishes a standard for\nassessing and benchmarking the resilience of current and future defensive\nmodels against adversarial threats.\n", "date": "2024-03-08", "categories": []}
{"id": "2403.09901", "title": "Robust Subgraph Learning by Monitoring Early Training Representations", "abstract": "  Graph neural networks (GNNs) have attracted significant attention for their\noutstanding performance in graph learning and node classification tasks.\nHowever, their vulnerability to adversarial attacks, particularly through\nsusceptible nodes, poses a challenge in decision-making. The need for robust\ngraph summarization is evident in adversarial challenges resulting from the\npropagation of attacks throughout the entire graph. In this paper, we address\nboth performance and adversarial robustness in graph input by introducing the\nnovel technique SHERD (Subgraph Learning Hale through Early Training\nRepresentation Distances). SHERD leverages information from layers of a\npartially trained graph convolutional network (GCN) to detect susceptible nodes\nduring adversarial attacks using standard distance metrics. The method\nidentifies \"vulnerable (bad)\" nodes and removes such nodes to form a robust\nsubgraph while maintaining node classification performance. Through our\nexperiments, we demonstrate the increased performance of SHERD in enhancing\nrobustness by comparing the network's performance on original and subgraph\ninputs against various baselines alongside existing adversarial attacks. Our\nexperiments across multiple datasets, including citation datasets such as Cora,\nCiteseer, and Pubmed, as well as microanatomical tissue structures of cell\ngraphs in the placenta, highlight that SHERD not only achieves substantial\nimprovement in robust performance but also outperforms several baselines in\nterms of node classification accuracy and computational complexity.\n", "date": "2024-03-14", "categories": []}
{"id": "2405.14210", "title": "Eidos: Efficient, Imperceptible Adversarial 3D Point Clouds", "abstract": "  Classification of 3D point clouds is a challenging machine learning (ML) task\nwith important real-world applications in a spectrum from autonomous driving\nand robot-assisted surgery to earth observation from low orbit. As with other\nML tasks, classification models are notoriously brittle in the presence of\nadversarial attacks. These are rooted in imperceptible changes to inputs with\nthe effect that a seemingly well-trained model ends up misclassifying the\ninput. This paper adds to the understanding of adversarial attacks by\npresenting Eidos, a framework providing Efficient Imperceptible aDversarial\nattacks on 3D pOint cloudS. Eidos supports a diverse set of imperceptibility\nmetrics. It employs an iterative, two-step procedure to identify optimal\nadversarial examples, thereby enabling a runtime-imperceptibility trade-off. We\nprovide empirical evidence relative to several popular 3D point cloud\nclassification models and several established 3D attack methods, showing Eidos'\nsuperiority with respect to efficiency as well as imperceptibility.\n", "date": "2024-05-23", "categories": []}
{"id": "2407.10445", "title": "Backdoor Attacks against Image-to-Image Networks", "abstract": "  Recently, deep learning-based Image-to-Image (I2I) networks have become the\npredominant choice for I2I tasks such as image super-resolution and denoising.\nDespite their remarkable performance, the backdoor vulnerability of I2I\nnetworks has not been explored. To fill this research gap, we conduct a\ncomprehensive investigation on the susceptibility of I2I networks to backdoor\nattacks. Specifically, we propose a novel backdoor attack technique, where the\ncompromised I2I network behaves normally on clean input images, yet outputs a\npredefined image of the adversary for malicious input images containing the\ntrigger. To achieve this I2I backdoor attack, we propose a targeted universal\nadversarial perturbation (UAP) generation algorithm for I2I networks, where the\ngenerated UAP is used as the backdoor trigger. Additionally, in the backdoor\ntraining process that contains the main task and the backdoor task, multi-task\nlearning (MTL) with dynamic weighting methods is employed to accelerate\nconvergence rates. In addition to attacking I2I tasks, we extend our I2I\nbackdoor to attack downstream tasks, including image classification and object\ndetection. Extensive experiments demonstrate the effectiveness of the I2I\nbackdoor on state-of-the-art I2I network architectures, as well as the\nrobustness against different mainstream backdoor defenses.\n", "date": "2024-07-15", "categories": []}
{"id": "2410.22678", "title": "Backdoor Attack Against Vision Transformers via Attention Gradient-Based\n  Image Erosion", "abstract": "  Vision Transformers (ViTs) have outperformed traditional Convolutional Neural\nNetworks (CNN) across various computer vision tasks. However, akin to CNN, ViTs\nare vulnerable to backdoor attacks, where the adversary embeds the backdoor\ninto the victim model, causing it to make wrong predictions about testing\nsamples containing a specific trigger. Existing backdoor attacks against ViTs\nhave the limitation of failing to strike an optimal balance between attack\nstealthiness and attack effectiveness.\n  In this work, we propose an Attention Gradient-based Erosion Backdoor (AGEB)\ntargeted at ViTs. Considering the attention mechanism of ViTs, AGEB selectively\nerodes pixels in areas of maximal attention gradient, embedding a covert\nbackdoor trigger. Unlike previous backdoor attacks against ViTs, AGEB achieves\nan optimal balance between attack stealthiness and attack effectiveness,\nensuring the trigger remains invisible to human detection while preserving the\nmodel's accuracy on clean samples. Extensive experimental evaluations across\nvarious ViT architectures and datasets confirm the effectiveness of AGEB,\nachieving a remarkable Attack Success Rate (ASR) without diminishing Clean Data\nAccuracy (CDA). Furthermore, the stealthiness of AGEB is rigorously validated,\ndemonstrating minimal visual discrepancies between the clean and the triggered\nimages.\n", "date": "2024-10-30", "categories": []}
{"id": "2411.10498", "title": "Prompt-Guided Environmentally Consistent Adversarial Patch", "abstract": "  Adversarial attacks in the physical world pose a significant threat to the\nsecurity of vision-based systems, such as facial recognition and autonomous\ndriving. Existing adversarial patch methods primarily focus on improving attack\nperformance, but they often produce patches that are easily detectable by\nhumans and struggle to achieve environmental consistency, i.e., blending\npatches into the environment. This paper introduces a novel approach for\ngenerating adversarial patches, which addresses both the visual naturalness and\nenvironmental consistency of the patches. We propose Prompt-Guided\nEnvironmentally Consistent Adversarial Patch (PG-ECAP), a method that aligns\nthe patch with the environment to ensure seamless integration into the\nenvironment. The approach leverages diffusion models to generate patches that\nare both environmental consistency and effective in evading detection. To\nfurther enhance the naturalness and consistency, we introduce two alignment\nlosses: Prompt Alignment Loss and Latent Space Alignment Loss, ensuring that\nthe generated patch maintains its adversarial properties while fitting\nnaturally within its environment. Extensive experiments in both digital and\nphysical domains demonstrate that PG-ECAP outperforms existing methods in\nattack success rate and environmental consistency.\n", "date": "2024-11-15", "categories": []}
{"id": "2411.11795", "title": "Exploring adversarial robustness of JPEG AI: methodology, comparison and\n  new methods", "abstract": "  Adversarial robustness of neural networks is an increasingly important area\nof research, combining studies on computer vision models, large language models\n(LLMs), and others. With the release of JPEG AI - the first standard for\nend-to-end neural image compression (NIC) methods - the question of its\nrobustness has become critically significant. JPEG AI is among the first\ninternational, real-world applications of neural-network-based models to be\nembedded in consumer devices. However, research on NIC robustness has been\nlimited to open-source codecs and a narrow range of attacks. This paper\nproposes a new methodology for measuring NIC robustness to adversarial attacks.\nWe present the first large-scale evaluation of JPEG AI's robustness, comparing\nit with other NIC models. Our evaluation results and code are publicly\navailable online (link is hidden for a blind review).\n", "date": "2024-11-18", "categories": []}
{"id": "2405.03316", "title": "Provably Unlearnable Data Examples", "abstract": "  The exploitation of publicly accessible data has led to escalating concerns\nregarding data privacy and intellectual property (IP) breaches in the age of\nartificial intelligence. To safeguard both data privacy and IP-related domain\nknowledge, efforts have been undertaken to render shared data unlearnable for\nunauthorized models in the wild. Existing methods apply empirically optimized\nperturbations to the data in the hope of disrupting the correlation between the\ninputs and the corresponding labels such that the data samples are converted\ninto Unlearnable Examples (UEs). Nevertheless, the absence of mechanisms to\nverify the robustness of UEs against uncertainty in unauthorized models and\ntheir training procedures engenders several under-explored challenges. First,\nit is hard to quantify the unlearnability of UEs against unauthorized\nadversaries from different runs of training, leaving the soundness of the\ndefense in obscurity. Particularly, as a prevailing evaluation metric,\nempirical test accuracy faces generalization errors and may not plausibly\nrepresent the quality of UEs. This also leaves room for attackers, as there is\nno rigid guarantee of the maximal test accuracy achievable by attackers.\nFurthermore, we find that a simple recovery attack can restore the clean-task\nperformance of the classifiers trained on UEs by slightly perturbing the\nlearned weights. To mitigate the aforementioned problems, in this paper, we\npropose a mechanism for certifying the so-called $(q, \\eta)$-Learnability of an\nunlearnable dataset via parametric smoothing. A lower certified $(q,\n\\eta)$-Learnability indicates a more robust and effective protection over the\ndataset. Concretely, we 1) improve the tightness of certified $(q,\n\\eta)$-Learnability and 2) design Provably Unlearnable Examples (PUEs) which\nhave reduced $(q, \\eta)$-Learnability.\n", "date": "2024-05-06", "categories": []}
{"id": "2407.03864", "title": "Adversarial Robustness of VAEs across Intersectional Subgroups", "abstract": "  Despite advancements in Autoencoders (AEs) for tasks like dimensionality\nreduction, representation learning and data generation, they remain vulnerable\nto adversarial attacks. Variational Autoencoders (VAEs), with their\nprobabilistic approach to disentangling latent spaces, show stronger resistance\nto such perturbations compared to deterministic AEs; however, their resilience\nagainst adversarial inputs is still a concern. This study evaluates the\nrobustness of VAEs against non-targeted adversarial attacks by optimizing\nminimal sample-specific perturbations to cause maximal damage across diverse\ndemographic subgroups (combinations of age and gender). We investigate two\nquestions: whether there are robustness disparities among subgroups, and what\nfactors contribute to these disparities, such as data scarcity and\nrepresentation entanglement. Our findings reveal that robustness disparities\nexist but are not always correlated with the size of the subgroup. By using\ndownstream gender and age classifiers and examining latent embeddings, we\nhighlight the vulnerability of subgroups like older women, who are prone to\nmisclassification due to adversarial perturbations pushing their\nrepresentations toward those of other subgroups.\n", "date": "2024-07-04", "categories": []}
{"id": "2411.08933", "title": "Confidence-aware Denoised Fine-tuning of Off-the-shelf Models for\n  Certified Robustness", "abstract": "  The remarkable advances in deep learning have led to the emergence of many\noff-the-shelf classifiers, e.g., large pre-trained models. However, since they\nare typically trained on clean data, they remain vulnerable to adversarial\nattacks. Despite this vulnerability, their superior performance and\ntransferability make off-the-shelf classifiers still valuable in practice,\ndemanding further work to provide adversarial robustness for them in a post-hoc\nmanner. A recently proposed method, denoised smoothing, leverages a denoiser\nmodel in front of the classifier to obtain provable robustness without\nadditional training. However, the denoiser often creates hallucination, i.e.,\nimages that have lost the semantics of their originally assigned class, leading\nto a drop in robustness. Furthermore, its noise-and-denoise procedure\nintroduces a significant distribution shift from the original distribution,\ncausing the denoised smoothing framework to achieve sub-optimal robustness. In\nthis paper, we introduce Fine-Tuning with Confidence-Aware Denoised Image\nSelection (FT-CADIS), a novel fine-tuning scheme to enhance the certified\nrobustness of off-the-shelf classifiers. FT-CADIS is inspired by the\nobservation that the confidence of off-the-shelf classifiers can effectively\nidentify hallucinated images during denoised smoothing. Based on this, we\ndevelop a confidence-aware training objective to handle such hallucinated\nimages and improve the stability of fine-tuning from denoised images. In this\nway, the classifier can be fine-tuned using only images that are beneficial for\nadversarial robustness. We also find that such a fine-tuning can be done by\nupdating a small fraction of parameters of the classifier. Extensive\nexperiments demonstrate that FT-CADIS has established the state-of-the-art\ncertified robustness among denoised smoothing methods across all\n$\\ell_2$-adversary radius in various benchmarks.\n", "date": "2024-11-13", "categories": []}
{"id": "2411.10367", "title": "Continual Adversarial Reinforcement Learning (CARL) of False Data\n  Injection detection: forgetting and explainability", "abstract": "  False data injection attacks (FDIAs) on smart inverters are a growing concern\nlinked to increased renewable energy production. While data-based FDIA\ndetection methods are also actively developed, we show that they remain\nvulnerable to impactful and stealthy adversarial examples that can be crafted\nusing Reinforcement Learning (RL). We propose to include such adversarial\nexamples in data-based detection training procedure via a continual adversarial\nRL (CARL) approach. This way, one can pinpoint the deficiencies of data-based\ndetection, thereby offering explainability during their incremental\nimprovement. We show that a continual learning implementation is subject to\ncatastrophic forgetting, and additionally show that forgetting can be addressed\nby employing a joint training strategy on all generated FDIA scenarios.\n", "date": "2024-11-15", "categories": []}
{"id": "2410.24006", "title": "DiffPAD: Denoising Diffusion-based Adversarial Patch Decontamination", "abstract": "  In the ever-evolving adversarial machine learning landscape, developing\neffective defenses against patch attacks has become a critical challenge,\nnecessitating reliable solutions to safeguard real-world AI systems. Although\ndiffusion models have shown remarkable capacity in image synthesis and have\nbeen recently utilized to counter $\\ell_p$-norm bounded attacks, their\npotential in mitigating localized patch attacks remains largely underexplored.\nIn this work, we propose DiffPAD, a novel framework that harnesses the power of\ndiffusion models for adversarial patch decontamination. DiffPAD first performs\nsuper-resolution restoration on downsampled input images, then adopts\nbinarization, dynamic thresholding scheme and sliding window for effective\nlocalization of adversarial patches. Such a design is inspired by the\ntheoretically derived correlation between patch size and diffusion restoration\nerror that is generalized across diverse patch attack scenarios. Finally,\nDiffPAD applies inpainting techniques to the original input images with the\nestimated patch region being masked. By integrating closed-form solutions for\nsuper-resolution restoration and image inpainting into the conditional reverse\nsampling process of a pre-trained diffusion model, DiffPAD obviates the need\nfor text guidance or fine-tuning. Through comprehensive experiments, we\ndemonstrate that DiffPAD not only achieves state-of-the-art adversarial\nrobustness against patch attacks but also excels in recovering naturalistic\nimages without patch remnants. The source code is available at\nhttps://github.com/JasonFu1998/DiffPAD.\n", "date": "2024-10-31", "categories": []}
{"id": "2411.09220", "title": "Transferable Adversarial Attacks against ASR", "abstract": "  Given the extensive research and real-world applications of automatic speech\nrecognition (ASR), ensuring the robustness of ASR models against minor input\nperturbations becomes a crucial consideration for maintaining their\neffectiveness in real-time scenarios. Previous explorations into ASR model\nrobustness have predominantly revolved around evaluating accuracy on white-box\nsettings with full access to ASR models. Nevertheless, full ASR model details\nare often not available in real-world applications. Therefore, evaluating the\nrobustness of black-box ASR models is essential for a comprehensive\nunderstanding of ASR model resilience. In this regard, we thoroughly study the\nvulnerability of practical black-box attacks in cutting-edge ASR models and\npropose to employ two advanced time-domain-based transferable attacks alongside\nour differentiable feature extractor. We also propose a speech-aware gradient\noptimization approach (SAGO) for ASR, which forces mistranscription with\nminimal impact on human imperceptibility through voice activity detection rule\nand a speech-aware gradient-oriented optimizer. Our comprehensive experimental\nresults reveal performance enhancements compared to baseline approaches across\nfive models on two databases.\n", "date": "2024-11-14", "categories": []}
{"id": "2411.09265", "title": "BEARD: Benchmarking the Adversarial Robustness for Dataset Distillation", "abstract": "  Dataset Distillation (DD) is an emerging technique that compresses\nlarge-scale datasets into significantly smaller synthesized datasets while\npreserving high test performance and enabling the efficient training of large\nmodels. However, current research primarily focuses on enhancing evaluation\naccuracy under limited compression ratios, often overlooking critical security\nconcerns such as adversarial robustness. A key challenge in evaluating this\nrobustness lies in the complex interactions between distillation methods, model\narchitectures, and adversarial attack strategies, which complicate standardized\nassessments. To address this, we introduce BEARD, an open and unified benchmark\ndesigned to systematically assess the adversarial robustness of DD methods,\nincluding DM, IDM, and BACON. BEARD encompasses a variety of adversarial\nattacks (e.g., FGSM, PGD, C&W) on distilled datasets like CIFAR-10/100 and\nTinyImageNet. Utilizing an adversarial game framework, it introduces three key\nmetrics: Robustness Ratio (RR), Attack Efficiency Ratio (AE), and Comprehensive\nRobustness-Efficiency Index (CREI). Our analysis includes unified benchmarks,\nvarious Images Per Class (IPC) settings, and the effects of adversarial\ntraining. Results are available on the BEARD Leaderboard, along with a library\nproviding model and dataset pools to support reproducible research. Access the\ncode at BEARD.\n", "date": "2024-11-14", "categories": []}
{"id": "2411.08248", "title": "Deceiving Question-Answering Models: A Hybrid Word-Level Adversarial\n  Approach", "abstract": "  Deep learning underpins most of the currently advanced natural language\nprocessing (NLP) tasks such as textual classification, neural machine\ntranslation (NMT), abstractive summarization and question-answering (QA).\nHowever, the robustness of the models, particularly QA models, against\nadversarial attacks is a critical concern that remains insufficiently explored.\nThis paper introduces QA-Attack (Question Answering Attack), a novel word-level\nadversarial strategy that fools QA models. Our attention-based attack exploits\nthe customized attention mechanism and deletion ranking strategy to identify\nand target specific words within contextual passages. It creates deceptive\ninputs by carefully choosing and substituting synonyms, preserving grammatical\nintegrity while misleading the model to produce incorrect responses. Our\napproach demonstrates versatility across various question types, particularly\nwhen dealing with extensive long textual inputs. Extensive experiments on\nmultiple benchmark datasets demonstrate that QA-Attack successfully deceives\nbaseline QA models and surpasses existing adversarial techniques regarding\nsuccess rate, semantics changes, BLEU score, fluency and grammar error rate.\n", "date": "2024-11-12", "categories": []}
{"id": "2411.07843", "title": "Chain Association-based Attacking and Shielding Natural Language\n  Processing Systems", "abstract": "  Association as a gift enables people do not have to mention something in\ncompletely straightforward words and allows others to understand what they\nintend to refer to. In this paper, we propose a chain association-based\nadversarial attack against natural language processing systems, utilizing the\ncomprehension gap between humans and machines. We first generate a chain\nassociation graph for Chinese characters based on the association paradigm for\nbuilding search space of potential adversarial examples. Then, we introduce an\ndiscrete particle swarm optimization algorithm to search for the optimal\nadversarial examples. We conduct comprehensive experiments and show that\nadvanced natural language processing models and applications, including large\nlanguage models, are vulnerable to our attack, while humans appear good at\nunderstanding the perturbed text. We also explore two methods, including\nadversarial training and associative graph-based recovery, to shield systems\nfrom chain association-based attack. Since a few examples that use some\nderogatory terms, this paper contains materials that may be offensive or\nupsetting to some people.\n", "date": "2024-11-12", "categories": []}
{"id": "2411.07850", "title": "IAE: Irony-based Adversarial Examples for Sentiment Analysis Systems", "abstract": "  Adversarial examples, which are inputs deliberately perturbed with\nimperceptible changes to induce model errors, have raised serious concerns for\nthe reliability and security of deep neural networks (DNNs). While adversarial\nattacks have been extensively studied in continuous data domains such as\nimages, the discrete nature of text presents unique challenges. In this paper,\nwe propose Irony-based Adversarial Examples (IAE), a method that transforms\nstraightforward sentences into ironic ones to create adversarial text. This\napproach exploits the rhetorical device of irony, where the intended meaning is\nopposite to the literal interpretation, requiring a deeper understanding of\ncontext to detect. The IAE method is particularly challenging due to the need\nto accurately locate evaluation words, substitute them with appropriate\ncollocations, and expand the text with suitable ironic elements while\nmaintaining semantic coherence. Our research makes the following key\ncontributions: (1) We introduce IAE, a strategy for generating textual\nadversarial examples using irony. This method does not rely on pre-existing\nirony corpora, making it a versatile tool for creating adversarial text in\nvarious NLP tasks. (2) We demonstrate that the performance of several\nstate-of-the-art deep learning models on sentiment analysis tasks significantly\ndeteriorates when subjected to IAE attacks. This finding underscores the\nsusceptibility of current NLP systems to adversarial manipulation through\nirony. (3) We compare the impact of IAE on human judgment versus NLP systems,\nrevealing that humans are less susceptible to the effects of irony in text.\n", "date": "2024-11-12", "categories": []}
{"id": "2301.02039", "title": "Randomized Message-Interception Smoothing: Gray-box Certificates for\n  Graph Neural Networks", "abstract": "  Randomized smoothing is one of the most promising frameworks for certifying\nthe adversarial robustness of machine learning models, including Graph Neural\nNetworks (GNNs). Yet, existing randomized smoothing certificates for GNNs are\noverly pessimistic since they treat the model as a black box, ignoring the\nunderlying architecture. To remedy this, we propose novel gray-box certificates\nthat exploit the message-passing principle of GNNs: We randomly intercept\nmessages and carefully analyze the probability that messages from adversarially\ncontrolled nodes reach their target nodes. Compared to existing certificates,\nwe certify robustness to much stronger adversaries that control entire nodes in\nthe graph and can arbitrarily manipulate node features. Our certificates\nprovide stronger guarantees for attacks at larger distances, as messages from\nfarther-away nodes are more likely to get intercepted. We demonstrate the\neffectiveness of our method on various models and datasets. Since our gray-box\ncertificates consider the underlying graph structure, we can significantly\nimprove certifiable robustness by applying graph sparsification.\n", "date": "2023-01-05", "categories": []}
{"id": "2310.16221", "title": "Hierarchical Randomized Smoothing", "abstract": "  Real-world data is complex and often consists of objects that can be\ndecomposed into multiple entities (e.g. images into pixels, graphs into\ninterconnected nodes). Randomized smoothing is a powerful framework for making\nmodels provably robust against small changes to their inputs - by guaranteeing\nrobustness of the majority vote when randomly adding noise before\nclassification. Yet, certifying robustness on such complex data via randomized\nsmoothing is challenging when adversaries do not arbitrarily perturb entire\nobjects (e.g. images) but only a subset of their entities (e.g. pixels). As a\nsolution, we introduce hierarchical randomized smoothing: We partially smooth\nobjects by adding random noise only on a randomly selected subset of their\nentities. By adding noise in a more targeted manner than existing methods we\nobtain stronger robustness guarantees while maintaining high accuracy. We\ninitialize hierarchical smoothing using different noising distributions,\nyielding novel robustness certificates for discrete and continuous domains. We\nexperimentally demonstrate the importance of hierarchical smoothing in image\nand node classification, where it yields superior robustness-accuracy\ntrade-offs. Overall, hierarchical smoothing is an important contribution\ntowards models that are both - certifiably robust to perturbations and\naccurate.\n", "date": "2023-10-24", "categories": []}
{"id": "2404.07234", "title": "Goal-guided Generative Prompt Injection Attack on Large Language Models", "abstract": "  Current large language models (LLMs) provide a strong foundation for\nlarge-scale user-oriented natural language tasks. A large number of users can\neasily inject adversarial text or instructions through the user interface, thus\ncausing LLMs model security challenges. Although there is currently a large\namount of research on prompt injection attacks, most of these black-box attacks\nuse heuristic strategies. It is unclear how these heuristic strategies relate\nto the success rate of attacks and thus effectively improve model robustness.\nTo solve this problem, we redefine the goal of the attack: to maximize the KL\ndivergence between the conditional probabilities of the clean text and the\nadversarial text. Furthermore, we prove that maximizing the KL divergence is\nequivalent to maximizing the Mahalanobis distance between the embedded\nrepresentation $x$ and $x'$ of the clean text and the adversarial text when the\nconditional probability is a Gaussian distribution and gives a quantitative\nrelationship on $x$ and $x'$. Then we designed a simple and effective\ngoal-guided generative prompt injection strategy (G2PIA) to find an injection\ntext that satisfies specific constraints to achieve the optimal attack effect\napproximately. It is particularly noteworthy that our attack method is a\nquery-free black-box attack method with low computational cost. Experimental\nresults on seven LLM models and four datasets show the effectiveness of our\nattack method.\n", "date": "2024-04-06", "categories": []}
{"id": "2411.06666", "title": "Adversarial Detection with a Dynamically Stable System", "abstract": "  Adversarial detection is designed to identify and reject maliciously crafted\nadversarial examples(AEs) which are generated to disrupt the classification of\ntarget models.\n  Presently, various input transformation-based methods have been developed on\nadversarial example detection, which typically rely on empirical experience and\nlead to unreliability against new attacks.\n  To address this issue, we propose and conduct a Dynamically Stable System\n(DSS), which can effectively detect the adversarial examples from normal\nexamples according to the stability of input examples.\n  Particularly, in our paper, the generation of adversarial examples is\nconsidered as the perturbation process of a Lyapunov dynamic system, and we\npropose an example stability mechanism, in which a novel control term is added\nin adversarial example generation to ensure that the normal examples can\nachieve dynamic stability while the adversarial examples cannot achieve the\nstability.\n  Then, based on the proposed example stability mechanism, a Dynamically Stable\nSystem (DSS) is proposed, which can utilize the disruption and restoration\nactions to determine the stability of input examples and detect the adversarial\nexamples through changes in the stability of the input examples.\n  In comparison with existing methods in three benchmark datasets(MNIST,\nCIFAR10, and CIFAR100), our evaluation results show that our proposed DSS can\nachieve ROC-AUC values of 99.83%, 97.81% and 94.47%, surpassing the\nstate-of-the-art(SOTA) values of 97.35%, 91.10% and 93.49% in the other 7\nmethods.\n", "date": "2024-11-11", "categories": []}
{"id": "2307.01778", "title": "Physically Realizable Natural-Looking Clothing Textures Evade Person\n  Detectors via 3D Modeling", "abstract": "  Recent works have proposed to craft adversarial clothes for evading person\ndetectors, while they are either only effective at limited viewing angles or\nvery conspicuous to humans. We aim to craft adversarial texture for clothes\nbased on 3D modeling, an idea that has been used to craft rigid adversarial\nobjects such as a 3D-printed turtle. Unlike rigid objects, humans and clothes\nare non-rigid, leading to difficulties in physical realization. In order to\ncraft natural-looking adversarial clothes that can evade person detectors at\nmultiple viewing angles, we propose adversarial camouflage textures (AdvCaT)\nthat resemble one kind of the typical textures of daily clothes, camouflage\ntextures. We leverage the Voronoi diagram and Gumbel-softmax trick to\nparameterize the camouflage textures and optimize the parameters via 3D\nmodeling. Moreover, we propose an efficient augmentation pipeline on 3D meshes\ncombining topologically plausible projection (TopoProj) and Thin Plate Spline\n(TPS) to narrow the gap between digital and real-world objects. We printed the\ndeveloped 3D texture pieces on fabric materials and tailored them into T-shirts\nand trousers. Experiments show high attack success rates of these clothes\nagainst multiple detectors.\n", "date": "2023-07-04", "categories": []}
{"id": "2308.04265", "title": "FLIRT: Feedback Loop In-context Red Teaming", "abstract": "  Warning: this paper contains content that may be inappropriate or offensive.\nAs generative models become available for public use in various applications,\ntesting and analyzing vulnerabilities of these models has become a priority. In\nthis work, we propose an automatic red teaming framework that evaluates a given\nblack-box model and exposes its vulnerabilities against unsafe and\ninappropriate content generation. Our framework uses in-context learning in a\nfeedback loop to red team models and trigger them into unsafe content\ngeneration. In particular, taking text-to-image models as target models, we\nexplore different feedback mechanisms to automatically learn effective and\ndiverse adversarial prompts. Our experiments demonstrate that even with\nenhanced safety features, Stable Diffusion (SD) models are vulnerable to our\nadversarial prompts, raising concerns on their robustness in practical uses.\nFurthermore, we demonstrate that the proposed framework is effective for red\nteaming text-to-text models.\n", "date": "2023-08-08", "categories": []}
{"id": "2311.06122", "title": "Fight Fire with Fire: Combating Adversarial Patch Attacks using\n  Pattern-randomized Defensive Patches", "abstract": "  Object detection has found extensive applications in various tasks, but it is\nalso susceptible to adversarial patch attacks. The ideal defense should be\neffective, efficient, easy to deploy, and capable of withstanding adaptive\nattacks. In this paper, we adopt a counterattack strategy to propose a novel\nand general methodology for defending adversarial attacks. Two types of\ndefensive patches, canary and woodpecker, are specially-crafted and injected\ninto the model input to proactively probe or counteract potential adversarial\npatches. In this manner, adversarial patch attacks can be effectively detected\nby simply analyzing the model output, without the need to alter the target\nmodel. Moreover, we employ randomized canary and woodpecker injection patterns\nto defend against defense-aware attacks. The effectiveness and practicality of\nthe proposed method are demonstrated through comprehensive experiments. The\nresults illustrate that canary and woodpecker achieve high performance, even\nwhen confronted with unknown attack methods, while incurring limited time\noverhead. Furthermore, our method also exhibits sufficient robustness against\ndefense-aware attacks, as evidenced by adaptive attack experiments.\n", "date": "2023-11-10", "categories": []}
{"id": "2404.13631", "title": "Fermi-Bose Machine achieves both generalization and adversarial\n  robustness", "abstract": "  Distinct from human cognitive processing, deep neural networks trained by\nbackpropagation can be easily fooled by adversarial examples. To design a\nsemantically meaningful representation learning, we discard backpropagation,\nand instead, propose a local contrastive learning, where the representation for\nthe inputs bearing the same label shrink (akin to boson) in hidden layers,\nwhile those of different labels repel (akin to fermion). This layer-wise\nlearning is local in nature, being biological plausible. A statistical\nmechanics analysis shows that the target fermion-pair-distance is a key\nparameter. Moreover, the application of this local contrastive learning to\nMNIST benchmark dataset demonstrates that the adversarial vulnerability of\nstandard perceptron can be greatly mitigated by tuning the target distance,\ni.e., controlling the geometric separation of prototype manifolds.\n", "date": "2024-04-21", "categories": []}
{"id": "2405.19928", "title": "BAN: Detecting Backdoors Activated by Adversarial Neuron Noise", "abstract": "  Backdoor attacks on deep learning represent a recent threat that has gained\nsignificant attention in the research community. Backdoor defenses are mainly\nbased on backdoor inversion, which has been shown to be generic,\nmodel-agnostic, and applicable to practical threat scenarios. State-of-the-art\nbackdoor inversion recovers a mask in the feature space to locate prominent\nbackdoor features, where benign and backdoor features can be disentangled.\nHowever, it suffers from high computational overhead, and we also find that it\noverly relies on prominent backdoor features that are highly distinguishable\nfrom benign features. To tackle these shortcomings, this paper improves\nbackdoor feature inversion for backdoor detection by incorporating extra neuron\nactivation information. In particular, we adversarially increase the loss of\nbackdoored models with respect to weights to activate the backdoor effect,\nbased on which we can easily differentiate backdoored and clean models.\nExperimental results demonstrate our defense, BAN, is 1.37$\\times$ (on\nCIFAR-10) and 5.11$\\times$ (on ImageNet200) more efficient with an average\n9.99\\% higher detect success rate than the state-of-the-art defense BTI-DBF.\nOur code and trained models are publicly available\nat~\\url{https://github.com/xiaoyunxxy/ban}.\n", "date": "2024-05-30", "categories": []}
{"id": "2411.05056", "title": "Seeing is Deceiving: Exploitation of Visual Pathways in Multi-Modal\n  Language Models", "abstract": "  Multi-Modal Language Models (MLLMs) have transformed artificial intelligence\nby combining visual and text data, making applications like image captioning,\nvisual question answering, and multi-modal content creation possible. This\nability to understand and work with complex information has made MLLMs useful\nin areas such as healthcare, autonomous systems, and digital content. However,\nintegrating multiple types of data also creates security risks. Attackers can\nmanipulate either the visual or text inputs, or both, to make the model produce\nunintended or even harmful responses. This paper reviews how visual inputs in\nMLLMs can be exploited by various attack strategies. We break down these\nattacks into categories: simple visual tweaks and cross-modal manipulations, as\nwell as advanced strategies like VLATTACK, HADES, and Collaborative Multimodal\nAdversarial Attack (Co-Attack). These attacks can mislead even the most robust\nmodels while looking nearly identical to the original visuals, making them hard\nto detect. We also discuss the broader security risks, including threats to\nprivacy and safety in important applications. To counter these risks, we review\ncurrent defense methods like the SmoothVLM framework, pixel-wise randomization,\nand MirrorCheck, looking at their strengths and limitations. We also discuss\nnew methods to make MLLMs more secure, including adaptive defenses, better\nevaluation tools, and security approaches that protect both visual and text\ndata. By bringing together recent developments and identifying key areas for\nimprovement, this review aims to support the creation of more secure and\nreliable multi-modal AI systems for real-world use.\n", "date": "2024-11-07", "categories": []}
{"id": "2411.05399", "title": "Post-Hoc Robustness Enhancement in Graph Neural Networks with\n  Conditional Random Fields", "abstract": "  Graph Neural Networks (GNNs), which are nowadays the benchmark approach in\ngraph representation learning, have been shown to be vulnerable to adversarial\nattacks, raising concerns about their real-world applicability. While existing\ndefense techniques primarily concentrate on the training phase of GNNs,\ninvolving adjustments to message passing architectures or pre-processing\nmethods, there is a noticeable gap in methods focusing on increasing robustness\nduring inference. In this context, this study introduces RobustCRF, a post-hoc\napproach aiming to enhance the robustness of GNNs at the inference stage. Our\nproposed method, founded on statistical relational learning using a Conditional\nRandom Field, is model-agnostic and does not require prior knowledge about the\nunderlying model architecture. We validate the efficacy of this approach across\nvarious models, leveraging benchmark node classification datasets.\n", "date": "2024-11-08", "categories": []}
{"id": "2401.03115", "title": "Transferable Learned Image Compression-Resistant Adversarial\n  Perturbations", "abstract": "  Adversarial attacks can readily disrupt the image classification system,\nrevealing the vulnerability of DNN-based recognition tasks. While existing\nadversarial perturbations are primarily applied to uncompressed images or\ncompressed images by the traditional image compression method, i.e., JPEG,\nlimited studies have investigated the robustness of models for image\nclassification in the context of DNN-based image compression. With the rapid\nevolution of advanced image compression, DNN-based learned image compression\nhas emerged as the promising approach for transmitting images in many\nsecurity-critical applications, such as cloud-based face recognition and\nautonomous driving, due to its superior performance over traditional\ncompression. Therefore, there is a pressing need to fully investigate the\nrobustness of a classification system post-processed by learned image\ncompression. To bridge this research gap, we explore the adversarial attack on\na new pipeline that targets image classification models that utilize learned\nimage compressors as pre-processing modules. Furthermore, to enhance the\ntransferability of perturbations across various quality levels and\narchitectures of learned image compression models, we introduce a saliency\nscore-based sampling method to enable the fast generation of transferable\nperturbation. Extensive experiments with popular attack methods demonstrate the\nenhanced transferability of our proposed method when attacking images that have\nbeen post-processed with different learned image compression models.\n", "date": "2024-01-06", "categories": []}
{"id": "2411.01222", "title": "$B^4$: A Black-Box Scrubbing Attack on LLM Watermarks", "abstract": "  Watermarking has emerged as a prominent technique for LLM-generated content\ndetection by embedding imperceptible patterns. Despite supreme performance, its\nrobustness against adversarial attacks remains underexplored. Previous work\ntypically considers a grey-box attack setting, where the specific type of\nwatermark is already known. Some even necessitates knowledge about\nhyperparameters of the watermarking method. Such prerequisites are unattainable\nin real-world scenarios. Targeting at a more realistic black-box threat model\nwith fewer assumptions, we here propose $B^4$, a black-box scrubbing attack on\nwatermarks. Specifically, we formulate the watermark scrubbing attack as a\nconstrained optimization problem by capturing its objectives with two\ndistributions, a Watermark Distribution and a Fidelity Distribution. This\noptimization problem can be approximately solved using two proxy distributions.\nExperimental results across 12 different settings demonstrate the superior\nperformance of $B^4$ compared with other baselines.\n", "date": "2024-11-02", "categories": []}
{"id": "2411.03348", "title": "Undermining Image and Text Classification Algorithms Using Adversarial\n  Attacks", "abstract": "  Machine learning models are prone to adversarial attacks, where inputs can be\nmanipulated in order to cause misclassifications. While previous research has\nfocused on techniques like Generative Adversarial Networks (GANs), there's\nlimited exploration of GANs and Synthetic Minority Oversampling Technique\n(SMOTE) in text and image classification models to perform adversarial attacks.\nOur study addresses this gap by training various machine learning models and\nusing GANs and SMOTE to generate additional data points aimed at attacking text\nclassification models. Furthermore, we extend our investigation to face\nrecognition models, training a Convolutional Neural Network(CNN) and subjecting\nit to adversarial attacks with fast gradient sign perturbations on key features\nidentified by GradCAM, a technique used to highlight key image characteristics\nCNNs use in classification. Our experiments reveal a significant vulnerability\nin classification models. Specifically, we observe a 20 % decrease in accuracy\nfor the top-performing text classification models post-attack, along with a 30\n% decrease in facial recognition accuracy. This highlights the susceptibility\nof these models to manipulation of input data. Adversarial attacks not only\ncompromise the security but also undermine the reliability of machine learning\nsystems. By showcasing the impact of adversarial attacks on both text\nclassification and face recognition models, our study underscores the urgent\nneed for develop robust defenses against such vulnerabilities.\n", "date": "2024-11-03", "categories": []}
{"id": "2411.04533", "title": "Neural Fingerprints for Adversarial Attack Detection", "abstract": "  Deep learning models for image classification have become standard tools in\nrecent years. A well known vulnerability of these models is their\nsusceptibility to adversarial examples. These are generated by slightly\naltering an image of a certain class in a way that is imperceptible to humans\nbut causes the model to classify it wrongly as another class. Many algorithms\nhave been proposed to address this problem, falling generally into one of two\ncategories: (i) building robust classifiers (ii) directly detecting attacked\nimages. Despite the good performance of these detectors, we argue that in a\nwhite-box setting, where the attacker knows the configuration and weights of\nthe network and the detector, they can overcome the detector by running many\nexamples on a local copy, and sending only those that were not detected to the\nactual model. This problem is common in security applications where even a very\ngood model is not sufficient to ensure safety. In this paper we propose to\novercome this inherent limitation of any static defence with randomization. To\ndo so, one must generate a very large family of detectors with consistent\nperformance, and select one or more of them randomly for each input. For the\nindividual detectors, we suggest the method of neural fingerprints. In the\ntraining phase, for each class we repeatedly sample a tiny random subset of\nneurons from certain layers of the network, and if their average is\nsufficiently different between clean and attacked images of the focal class\nthey are considered a fingerprint and added to the detector bank. During test\ntime, we sample fingerprints from the bank associated with the label predicted\nby the model, and detect attacks using a likelihood ratio test. We evaluate our\ndetectors on ImageNet with different attack methods and model architectures,\nand show near-perfect detection with low rates of false detection.\n", "date": "2024-11-07", "categories": []}
{"id": "2411.04772", "title": "Attention Masks Help Adversarial Attacks to Bypass Safety Detectors", "abstract": "  Despite recent research advancements in adversarial attack methods, current\napproaches against XAI monitors are still discoverable and slower. In this\npaper, we present an adaptive framework for attention mask generation to enable\nstealthy, explainable and efficient PGD image classification adversarial attack\nunder XAI monitors. Specifically, we utilize mutation XAI mixture and multitask\nself-supervised X-UNet for attention mask generation to guide PGD attack.\nExperiments on MNIST (MLP), CIFAR-10 (AlexNet) have shown that our system can\noutperform benchmark PGD, Sparsefool and SOTA SINIFGSM in balancing among\nstealth, efficiency and explainability which is crucial for effectively fooling\nSOTA defense protected classifiers.\n", "date": "2024-11-07", "categories": []}
{"id": "2303.15409", "title": "Class-Conditioned Transformation for Enhanced Robust Image\n  Classification", "abstract": "  Robust classification methods predominantly concentrate on algorithms that\naddress a specific threat model, resulting in ineffective defenses against\nother threat models. Real-world applications are exposed to this vulnerability,\nas malicious attackers might exploit alternative threat models. In this work,\nwe propose a novel test-time threat model agnostic algorithm that enhances\nAdversarial-Trained (AT) models. Our method operates through COnditional image\ntransformation and DIstance-based Prediction (CODIP) and includes two main\nsteps: First, we transform the input image into each dataset class, where the\ninput image might be either clean or attacked. Next, we make a prediction based\non the shortest transformed distance. The conditional transformation utilizes\nthe perceptually aligned gradients property possessed by AT models and, as a\nresult, eliminates the need for additional models or additional training.\nMoreover, it allows users to choose the desired balance between clean and\nrobust accuracy without training. The proposed method achieves state-of-the-art\nresults demonstrated through extensive experiments on various models, AT\nmethods, datasets, and attack types. Notably, applying CODIP leads to\nsubstantial robust accuracy improvement of up to $+23\\%$, $+20\\%$, $+26\\%$, and\n$+22\\%$ on CIFAR10, CIFAR100, ImageNet and Flowers datasets, respectively.\n", "date": "2023-03-27", "categories": []}
{"id": "2411.02669", "title": "Semantic-Aligned Adversarial Evolution Triangle for High-Transferability\n  Vision-Language Attack", "abstract": "  Vision-language pre-training (VLP) models excel at interpreting both images\nand text but remain vulnerable to multimodal adversarial examples (AEs).\nAdvancing the generation of transferable AEs, which succeed across unseen\nmodels, is key to developing more robust and practical VLP models. Previous\napproaches augment image-text pairs to enhance diversity within the adversarial\nexample generation process, aiming to improve transferability by expanding the\ncontrast space of image-text features. However, these methods focus solely on\ndiversity around the current AEs, yielding limited gains in transferability. To\naddress this issue, we propose to increase the diversity of AEs by leveraging\nthe intersection regions along the adversarial trajectory during optimization.\nSpecifically, we propose sampling from adversarial evolution triangles composed\nof clean, historical, and current adversarial examples to enhance adversarial\ndiversity. We provide a theoretical analysis to demonstrate the effectiveness\nof the proposed adversarial evolution triangle. Moreover, we find that\nredundant inactive dimensions can dominate similarity calculations, distorting\nfeature matching and making AEs model-dependent with reduced transferability.\nHence, we propose to generate AEs in the semantic image-text feature contrast\nspace, which can project the original feature space into a semantic corpus\nsubspace. The proposed semantic-aligned subspace can reduce the image feature\nredundancy, thereby improving adversarial transferability. Extensive\nexperiments across different datasets and models demonstrate that the proposed\nmethod can effectively improve adversarial transferability and outperform\nstate-of-the-art adversarial attack methods. The code is released at\nhttps://github.com/jiaxiaojunQAQ/SA-AET.\n", "date": "2024-11-04", "categories": []}
{"id": "2411.02871", "title": "Enhancing Adversarial Robustness via Uncertainty-Aware Distributional\n  Adversarial Training", "abstract": "  Despite remarkable achievements in deep learning across various domains, its\ninherent vulnerability to adversarial examples still remains a critical concern\nfor practical deployment. Adversarial training has emerged as one of the most\neffective defensive techniques for improving model robustness against such\nmalicious inputs. However, existing adversarial training schemes often lead to\nlimited generalization ability against underlying adversaries with diversity\ndue to their overreliance on a point-by-point augmentation strategy by mapping\neach clean example to its adversarial counterpart during training. In addition,\nadversarial examples can induce significant disruptions in the statistical\ninformation w.r.t. the target model, thereby introducing substantial\nuncertainty and challenges to modeling the distribution of adversarial\nexamples. To circumvent these issues, in this paper, we propose a novel\nuncertainty-aware distributional adversarial training method, which enforces\nadversary modeling by leveraging both the statistical information of\nadversarial examples and its corresponding uncertainty estimation, with the\ngoal of augmenting the diversity of adversaries. Considering the potentially\nnegative impact induced by aligning adversaries to misclassified clean\nexamples, we also refine the alignment reference based on the statistical\nproximity to clean examples during adversarial training, thereby reframing\nadversarial training within a distribution-to-distribution matching framework\ninteracted between the clean and adversarial domains. Furthermore, we design an\nintrospective gradient alignment approach via matching input gradients between\nthese domains without introducing external models. Extensive experiments across\nfour benchmark datasets and various network architectures demonstrate that our\napproach achieves state-of-the-art adversarial robustness and maintains natural\nperformance.\n", "date": "2024-11-05", "categories": []}
{"id": "2303.14133", "title": "Survey on Adversarial Attack and Defense for Medical Image Analysis:\n  Methods and Challenges", "abstract": "  Deep learning techniques have achieved superior performance in computer-aided\nmedical image analysis, yet they are still vulnerable to imperceptible\nadversarial attacks, resulting in potential misdiagnosis in clinical practice.\nOppositely, recent years have also witnessed remarkable progress in defense\nagainst these tailored adversarial examples in deep medical diagnosis systems.\nIn this exposition, we present a comprehensive survey on recent advances in\nadversarial attacks and defenses for medical image analysis with a systematic\ntaxonomy in terms of the application scenario. We also provide a unified\nframework for different types of adversarial attack and defense methods in the\ncontext of medical image analysis. For a fair comparison, we establish a new\nbenchmark for adversarially robust medical diagnosis models obtained by\nadversarial training under various scenarios. To the best of our knowledge,\nthis is the first survey paper that provides a thorough evaluation of\nadversarially robust medical diagnosis models. By analyzing qualitative and\nquantitative results, we conclude this survey with a detailed discussion of\ncurrent challenges for adversarial attack and defense in medical image analysis\nsystems to shed light on future research directions. Code is available on\n\\href{https://github.com/tomvii/Adv_MIA}{\\color{red}{GitHub}}.\n", "date": "2023-03-24", "categories": []}
{"id": "2402.02554", "title": "DeSparsify: Adversarial Attack Against Token Sparsification Mechanisms\n  in Vision Transformers", "abstract": "  Vision transformers have contributed greatly to advancements in the computer\nvision domain, demonstrating state-of-the-art performance in diverse tasks\n(e.g., image classification, object detection). However, their high\ncomputational requirements grow quadratically with the number of tokens used.\nToken sparsification mechanisms have been proposed to address this issue. These\nmechanisms employ an input-dependent strategy, in which uninformative tokens\nare discarded from the computation pipeline, improving the model's efficiency.\nHowever, their dynamism and average-case assumption makes them vulnerable to a\nnew threat vector - carefully crafted adversarial examples capable of fooling\nthe sparsification mechanism, resulting in worst-case performance. In this\npaper, we present DeSparsify, an attack targeting the availability of vision\ntransformers that use token sparsification mechanisms. The attack aims to\nexhaust the operating system's resources, while maintaining its stealthiness.\nOur evaluation demonstrates the attack's effectiveness on three token\nsparsification mechanisms and examines the attack's transferability between\nthem and its effect on the GPU resources. To mitigate the impact of the attack,\nwe propose various countermeasures.\n", "date": "2024-02-04", "categories": []}
{"id": "2405.18137", "title": "Exploiting LLM Quantization", "abstract": "  Quantization leverages lower-precision weights to reduce the memory usage of\nlarge language models (LLMs) and is a key technique for enabling their\ndeployment on commodity hardware. While LLM quantization's impact on utility\nhas been extensively explored, this work for the first time studies its adverse\neffects from a security perspective. We reveal that widely used quantization\nmethods can be exploited to produce a harmful quantized LLM, even though the\nfull-precision counterpart appears benign, potentially tricking users into\ndeploying the malicious quantized model. We demonstrate this threat using a\nthree-staged attack framework: (i) first, we obtain a malicious LLM through\nfine-tuning on an adversarial task; (ii) next, we quantize the malicious model\nand calculate constraints that characterize all full-precision models that map\nto the same quantized model; (iii) finally, using projected gradient descent,\nwe tune out the poisoned behavior from the full-precision model while ensuring\nthat its weights satisfy the constraints computed in step (ii). This procedure\nresults in an LLM that exhibits benign behavior in full precision but when\nquantized, it follows the adversarial behavior injected in step (i). We\nexperimentally demonstrate the feasibility and severity of such an attack\nacross three diverse scenarios: vulnerable code generation, content injection,\nand over-refusal attack. In practice, the adversary could host the resulting\nfull-precision model on an LLM community hub such as Hugging Face, exposing\nmillions of users to the threat of deploying its malicious quantized version on\ntheir devices.\n", "date": "2024-05-28", "categories": []}
{"id": "2409.08167", "title": "High-Frequency Anti-DreamBooth: Robust Defense against Personalized\n  Image Synthesis", "abstract": "  Recently, text-to-image generative models have been misused to create\nunauthorized malicious images of individuals, posing a growing social problem.\nPrevious solutions, such as Anti-DreamBooth, add adversarial noise to images to\nprotect them from being used as training data for malicious generation.\nHowever, we found that the adversarial noise can be removed by adversarial\npurification methods such as DiffPure. Therefore, we propose a new adversarial\nattack method that adds strong perturbation on the high-frequency areas of\nimages to make it more robust to adversarial purification. Our experiment\nshowed that the adversarial images retained noise even after adversarial\npurification, hindering malicious image generation.\n", "date": "2024-09-12", "categories": []}
{"id": "2410.15645", "title": "Boosting Jailbreak Transferability for Large Language Models", "abstract": "  Large language models have drawn significant attention to the challenge of\nsafe alignment, especially regarding jailbreak attacks that circumvent security\nmeasures to produce harmful content. To address the limitations of existing\nmethods like GCG, which perform well in single-model attacks but lack\ntransferability, we propose several enhancements, including a scenario\ninduction template, optimized suffix selection, and the integration of\nre-suffix attack mechanism to reduce inconsistent outputs. Our approach has\nshown superior performance in extensive experiments across various benchmarks,\nachieving nearly 100% success rates in both attack execution and\ntransferability. Notably, our method has won the first place in the AISG-hosted\nGlobal Challenge for Safe and Secure LLMs. The code is released at\nhttps://github.com/HqingLiu/SI-GCG.\n", "date": "2024-10-21", "categories": []}
{"id": "2411.00837", "title": "Longitudinal Mammogram Exam-based Breast Cancer Diagnosis Models:\n  Vulnerability to Adversarial Attacks", "abstract": "  In breast cancer detection and diagnosis, the longitudinal analysis of\nmammogram images is crucial. Contemporary models excel in detecting temporal\nimaging feature changes, thus enhancing the learning process over sequential\nimaging exams. Yet, the resilience of these longitudinal models against\nadversarial attacks remains underexplored. In this study, we proposed a novel\nattack method that capitalizes on the feature-level relationship between two\nsequential mammogram exams of a longitudinal model, guided by both\ncross-entropy loss and distance metric learning, to achieve significant attack\nefficacy, as implemented using attack transferring in a black-box attacking\nmanner. We performed experiments on a cohort of 590 breast cancer patients\n(each has two sequential mammogram exams) in a case-control setting. Results\nshowed that our proposed method surpassed several state-of-the-art adversarial\nattacks in fooling the diagnosis models to give opposite outputs. Our method\nremained effective even if the model was trained with the common defending\nmethod of adversarial training.\n", "date": "2024-10-29", "categories": []}
{"id": "2411.00898", "title": "Replace-then-Perturb: Targeted Adversarial Attacks With Visual Reasoning\n  for Vision-Language Models", "abstract": "  The conventional targeted adversarial attacks add a small perturbation to an\nimage to make neural network models estimate the image as a predefined target\nclass, even if it is not the correct target class. Recently, for\nvisual-language models (VLMs), the focus of targeted adversarial attacks is to\ngenerate a perturbation that makes VLMs answer intended target text outputs.\nFor example, they aim to make a small perturbation on an image to make VLMs'\nanswers change from \"there is an apple\" to \"there is a baseball.\" However,\nanswering just intended text outputs is insufficient for tricky questions like\n\"if there is a baseball, tell me what is below it.\" This is because the target\nof the adversarial attacks does not consider the overall integrity of the\noriginal image, thereby leading to a lack of visual reasoning. In this work, we\nfocus on generating targeted adversarial examples with visual reasoning against\nVLMs. To this end, we propose 1) a novel adversarial attack procedure --\nnamely, Replace-then-Perturb and 2) a contrastive learning-based adversarial\nloss -- namely, Contrastive-Adv. In Replace-then-Perturb, we first leverage a\ntext-guided segmentation model to find the target object in the image. Then, we\nget rid of the target object and inpaint the empty space with the desired\nprompt. By doing this, we can generate a target image corresponding to the\ndesired prompt, while maintaining the overall integrity of the original image.\nFurthermore, in Contrastive-Adv, we design a novel loss function to obtain\nbetter adversarial examples. Our extensive benchmark results demonstrate that\nReplace-then-Perturb and Contrastive-Adv outperform the baseline adversarial\nattack algorithms. We note that the source code to reproduce the results will\nbe available.\n", "date": "2024-11-01", "categories": []}
{"id": "2411.01889", "title": "LiDAttack: Robust Black-box Attack on LiDAR-based Object Detection", "abstract": "  Since DNN is vulnerable to carefully crafted adversarial examples,\nadversarial attack on LiDAR sensors have been extensively studied. We introduce\na robust black-box attack dubbed LiDAttack. It utilizes a genetic algorithm\nwith a simulated annealing strategy to strictly limit the location and number\nof perturbation points, achieving a stealthy and effective attack. And it\nsimulates scanning deviations, allowing it to adapt to dynamic changes in real\nworld scenario variations. Extensive experiments are conducted on 3 datasets\n(i.e., KITTI, nuScenes, and self-constructed data) with 3 dominant object\ndetection models (i.e., PointRCNN, PointPillar, and PV-RCNN++). The results\nreveal the efficiency of the LiDAttack when targeting a wide range of object\ndetection models, with an attack success rate (ASR) up to 90%.\n", "date": "2024-11-04", "categories": []}
{"id": "2411.02094", "title": "Alignment-Based Adversarial Training (ABAT) for Improving the Robustness\n  and Accuracy of EEG-Based BCIs", "abstract": "  Machine learning has achieved great success in electroencephalogram (EEG)\nbased brain-computer interfaces (BCIs). Most existing BCI studies focused on\nimproving the decoding accuracy, with only a few considering the adversarial\nsecurity. Although many adversarial defense approaches have been proposed in\nother application domains such as computer vision, previous research showed\nthat their direct extensions to BCIs degrade the classification accuracy on\nbenign samples. This phenomenon greatly affects the applicability of\nadversarial defense approaches to EEG-based BCIs. To mitigate this problem, we\npropose alignment-based adversarial training (ABAT), which performs EEG data\nalignment before adversarial training. Data alignment aligns EEG trials from\ndifferent domains to reduce their distribution discrepancies, and adversarial\ntraining further robustifies the classification boundary. The integration of\ndata alignment and adversarial training can make the trained EEG classifiers\nsimultaneously more accurate and more robust. Experiments on five EEG datasets\nfrom two different BCI paradigms (motor imagery classification, and event\nrelated potential recognition), three convolutional neural network classifiers\n(EEGNet, ShallowCNN and DeepCNN) and three different experimental settings\n(offline within-subject cross-block/-session classification, online\ncross-session classification, and pre-trained classifiers) demonstrated its\neffectiveness. It is very intriguing that adversarial attacks, which are\nusually used to damage BCI systems, can be used in ABAT to simultaneously\nimprove the model accuracy and robustness.\n", "date": "2024-11-04", "categories": []}
{"id": "2402.13629", "title": "Adversarial Purification and Fine-tuning for Robust UDC Image\n  Restoration", "abstract": "  This study delves into the enhancement of Under-Display Camera (UDC) image\nrestoration models, focusing on their robustness against adversarial attacks.\nDespite its innovative approach to seamless display integration, UDC technology\nfaces unique image degradation challenges exacerbated by the susceptibility to\nadversarial perturbations. Our research initially conducts an in-depth\nrobustness evaluation of deep-learning-based UDC image restoration models by\nemploying several white-box and black-box attacking methods. This evaluation is\npivotal in understanding the vulnerabilities of current UDC image restoration\ntechniques. Following the assessment, we introduce a defense framework\nintegrating adversarial purification with subsequent fine-tuning processes.\nFirst, our approach employs diffusion-based adversarial purification,\neffectively neutralizing adversarial perturbations. Then, we apply the\nfine-tuning methodologies to refine the image restoration models further,\nensuring that the quality and fidelity of the restored images are maintained.\nThe effectiveness of our proposed approach is validated through extensive\nexperiments, showing marked improvements in resilience against typical\nadversarial attacks.\n", "date": "2024-02-21", "categories": []}
{"id": "2405.15589", "title": "Efficient Adversarial Training in LLMs with Continuous Attacks", "abstract": "  Large language models (LLMs) are vulnerable to adversarial attacks that can\nbypass their safety guardrails. In many domains, adversarial training has\nproven to be one of the most promising methods to reliably improve robustness\nagainst such attacks. Yet, in the context of LLMs, current methods for\nadversarial training are hindered by the high computational costs required to\nperform discrete adversarial attacks at each training iteration. We address\nthis problem by instead calculating adversarial attacks in the continuous\nembedding space of the LLM, which is orders of magnitudes more efficient. We\npropose a fast adversarial training algorithm (C-AdvUL) composed of two losses:\nthe first makes the model robust on continuous embedding attacks computed on an\nadversarial behaviour dataset; the second ensures the usefulness of the final\nmodel by fine-tuning on utility data. Moreover, we introduce C-AdvIPO, an\nadversarial variant of IPO that does not require utility data for adversarially\nrobust alignment. Our empirical evaluation on five models from different\nfamilies (Gemma, Phi3, Mistral, Zephyr, Llama2) and at different scales (2B,\n3.8B, 7B) shows that both algorithms substantially enhance LLM robustness\nagainst discrete attacks (GCG, AutoDAN, PAIR), while maintaining utility. Our\nresults demonstrate that robustness to continuous perturbations can extrapolate\nto discrete threat models. Thereby, we present a path toward scalable\nadversarial training algorithms for robustly aligning LLMs.\n", "date": "2024-05-24", "categories": []}
{"id": "2405.16405", "title": "Intruding with Words: Towards Understanding Graph Injection Attacks at\n  the Text Level", "abstract": "  Graph Neural Networks (GNNs) excel across various applications but remain\nvulnerable to adversarial attacks, particularly Graph Injection Attacks (GIAs),\nwhich inject malicious nodes into the original graph and pose realistic\nthreats. Text-attributed graphs (TAGs), where nodes are associated with textual\nfeatures, are crucial due to their prevalence in real-world applications and\nare commonly used to evaluate these vulnerabilities. However, existing research\nonly focuses on embedding-level GIAs, which inject node embeddings rather than\nactual textual content, limiting their applicability and simplifying detection.\nIn this paper, we pioneer the exploration of GIAs at the text level, presenting\nthree novel attack designs that inject textual content into the graph. Through\ntheoretical and empirical analysis, we demonstrate that text interpretability,\na factor previously overlooked at the embedding level, plays a crucial role in\nattack strength. Among the designs we investigate, the Word-frequency-based\nText-level GIA (WTGIA) is particularly notable for its balance between\nperformance and interpretability. Despite the success of WTGIA, we discover\nthat defenders can easily enhance their defenses with customized text embedding\nmethods or large language model (LLM)--based predictors. These insights\nunderscore the necessity for further research into the potential and practical\nsignificance of text-level GIAs.\n", "date": "2024-05-26", "categories": []}
{"id": "2405.20778", "title": "Improved Generation of Adversarial Examples Against Safety-aligned LLMs", "abstract": "  Adversarial prompts generated using gradient-based methods exhibit\noutstanding performance in performing automatic jailbreak attacks against\nsafety-aligned LLMs. Nevertheless, due to the discrete nature of texts, the\ninput gradient of LLMs struggles to precisely reflect the magnitude of loss\nchange that results from token replacements in the prompt, leading to limited\nattack success rates against safety-aligned LLMs, even in the white-box\nsetting. In this paper, we explore a new perspective on this problem,\nsuggesting that it can be alleviated by leveraging innovations inspired in\ntransfer-based attacks that were originally proposed for attacking black-box\nimage classification models. For the first time, we appropriate the ideologies\nof effective methods among these transfer-based attacks, i.e., Skip Gradient\nMethod and Intermediate Level Attack, into gradient-based adversarial prompt\ngeneration and achieve significant performance gains without introducing\nobvious computational cost. Meanwhile, by discussing mechanisms behind the\ngains, new insights are drawn, and proper combinations of these methods are\nalso developed. Our empirical results show that 87% of the query-specific\nadversarial suffixes generated by the developed combination can induce\nLlama-2-7B-Chat to produce the output that exactly matches the target string on\nAdvBench. This match rate is 33% higher than that of a very strong baseline\nknown as GCG, demonstrating advanced discrete optimization for adversarial\nprompt generation against LLMs. In addition, without introducing obvious cost,\nthe combination achieves >30% absolute increase in attack success rates\ncompared with GCG when generating both query-specific (38% -> 68%) and\nuniversal adversarial prompts (26.68% -> 60.32%) for attacking the\nLlama-2-7B-Chat model on AdvBench. Code at:\nhttps://github.com/qizhangli/Gradient-based-Jailbreak-Attacks.\n", "date": "2024-05-28", "categories": []}
{"id": "2406.18451", "title": "Detecting Brittle Decisions for Free: Leveraging Margin Consistency in\n  Deep Robust Classifiers", "abstract": "  Despite extensive research on adversarial training strategies to improve\nrobustness, the decisions of even the most robust deep learning models can\nstill be quite sensitive to imperceptible perturbations, creating serious risks\nwhen deploying them for high-stakes real-world applications. While detecting\nsuch cases may be critical, evaluating a model's vulnerability at a\nper-instance level using adversarial attacks is computationally too intensive\nand unsuitable for real-time deployment scenarios. The input space margin is\nthe exact score to detect non-robust samples and is intractable for deep neural\nnetworks. This paper introduces the concept of margin consistency -- a property\nthat links the input space margins and the logit margins in robust models --\nfor efficient detection of vulnerable samples. First, we establish that margin\nconsistency is a necessary and sufficient condition to use a model's logit\nmargin as a score for identifying non-robust samples. Next, through\ncomprehensive empirical analysis of various robustly trained models on CIFAR10\nand CIFAR100 datasets, we show that they indicate high margin consistency with\na strong correlation between their input space margins and the logit margins.\nThen, we show that we can effectively and confidently use the logit margin to\ndetect brittle decisions with such models. Finally, we address cases where the\nmodel is not sufficiently margin-consistent by learning a pseudo-margin from\nthe feature representation. Our findings highlight the potential of leveraging\ndeep representations to assess adversarial vulnerability in deployment\nscenarios efficiently.\n", "date": "2024-06-26", "categories": []}
{"id": "2411.00192", "title": "Optical Lens Attack on Monocular Depth Estimation for Autonomous Driving", "abstract": "  Monocular Depth Estimation (MDE) is a pivotal component of vision-based\nAutonomous Driving (AD) systems, enabling vehicles to estimate the depth of\nsurrounding objects using a single camera image. This estimation guides\nessential driving decisions, such as braking before an obstacle or changing\nlanes to avoid collisions. In this paper, we explore vulnerabilities of MDE\nalgorithms in AD systems, presenting LensAttack, a novel physical attack that\nstrategically places optical lenses on the camera of an autonomous vehicle to\nmanipulate the perceived object depths. LensAttack encompasses two attack\nformats: concave lens attack and convex lens attack, each utilizing different\noptical lenses to induce false depth perception. We first develop a\nmathematical model that outlines the parameters of the attack, followed by\nsimulations and real-world evaluations to assess its efficacy on\nstate-of-the-art MDE models. Additionally, we adopt an attack optimization\nmethod to further enhance the attack success rate by optimizing the attack\nfocal length. To better evaluate the implications of LensAttack on AD, we\nconduct comprehensive end-to-end system simulations using the CARLA platform.\nThe results reveal that LensAttack can significantly disrupt the depth\nestimation processes in AD systems, posing a serious threat to their\nreliability and safety. Finally, we discuss some potential defense methods to\nmitigate the effects of the proposed attack.\n", "date": "2024-10-31", "categories": []}
{"id": "2411.00222", "title": "Protecting Feed-Forward Networks from Adversarial Attacks Using\n  Predictive Coding", "abstract": "  An adversarial example is a modified input image designed to cause a Machine\nLearning (ML) model to make a mistake; these perturbations are often invisible\nor subtle to human observers and highlight vulnerabilities in a model's ability\nto generalize from its training data. Several adversarial attacks can create\nsuch examples, each with a different perspective, effectiveness, and\nperceptibility of changes. Conversely, defending against such adversarial\nattacks improves the robustness of ML models in image processing and other\ndomains of deep learning. Most defence mechanisms require either a level of\nmodel awareness, changes to the model, or access to a comprehensive set of\nadversarial examples during training, which is impractical. Another option is\nto use an auxiliary model in a preprocessing manner without changing the\nprimary model. This study presents a practical and effective solution -- using\npredictive coding networks (PCnets) as an auxiliary step for adversarial\ndefence. By seamlessly integrating PCnets into feed-forward networks as a\npreprocessing step, we substantially bolster resilience to adversarial\nperturbations. Our experiments on MNIST and CIFAR10 demonstrate the remarkable\neffectiveness of PCnets in mitigating adversarial examples with about 82% and\n65% improvements in robustness, respectively. The PCnet, trained on a small\nsubset of the dataset, leverages its generative nature to effectively counter\nadversarial efforts, reverting perturbed images closer to their original forms.\nThis innovative approach holds promise for enhancing the security and\nreliability of neural network classifiers in the face of the escalating threat\nof adversarial attacks.\n", "date": "2024-10-31", "categories": []}
{"id": "2411.00519", "title": "Outlier-Oriented Poisoning Attack: A Grey-box Approach to Disturb\n  Decision Boundaries by Perturbing Outliers in Multiclass Learning", "abstract": "  Poisoning attacks are a primary threat to machine learning models, aiming to\ncompromise their performance and reliability by manipulating training datasets.\nThis paper introduces a novel attack - Outlier-Oriented Poisoning (OOP) attack,\nwhich manipulates labels of most distanced samples from the decision\nboundaries. The paper also investigates the adverse impact of such attacks on\ndifferent machine learning algorithms within a multiclass classification\nscenario, analyzing their variance and correlation between different poisoning\nlevels and performance degradation. To ascertain the severity of the OOP attack\nfor different degrees (5% - 25%) of poisoning, we analyzed variance, accuracy,\nprecision, recall, f1-score, and false positive rate for chosen ML\nmodels.Benchmarking our OOP attack, we have analyzed key characteristics of\nmulticlass machine learning algorithms and their sensitivity to poisoning\nattacks. Our experimentation used three publicly available datasets: IRIS,\nMNIST, and ISIC. Our analysis shows that KNN and GNB are the most affected\nalgorithms with a decrease in accuracy of 22.81% and 56.07% while increasing\nfalse positive rate to 17.14% and 40.45% for IRIS dataset with 15% poisoning.\nFurther, Decision Trees and Random Forest are the most resilient algorithms\nwith the least accuracy disruption of 12.28% and 17.52% with 15% poisoning of\nthe IRIS dataset. We have also analyzed the correlation between number of\ndataset classes and the performance degradation of models. Our analysis\nhighlighted that number of classes are inversely proportional to the\nperformance degradation, specifically the decrease in accuracy of the models,\nwhich is normalized with increasing number of classes. Further, our analysis\nidentified that imbalanced dataset distribution can aggravate the impact of\npoisoning for machine learning models\n", "date": "2024-11-01", "categories": []}
{"id": "2306.01762", "title": "Rapid Plug-in Defenders", "abstract": "  In the realm of daily services, the deployment of deep neural networks\nunderscores the paramount importance of their reliability. However, the\nvulnerability of these networks to adversarial attacks, primarily\nevasion-based, poses a concerning threat to their functionality. Common methods\nfor enhancing robustness involve heavy adversarial training or leveraging\nlearned knowledge from clean data, both necessitating substantial computational\nresources. This inherent time-intensive nature severely limits the agility of\nlarge foundational models to swiftly counter adversarial perturbations. To\naddress this challenge, this paper focuses on the Rapid Plug-in Defender\n(RaPiD) problem, aiming to rapidly counter adversarial perturbations without\naltering the deployed model. Drawing inspiration from the generalization and\nthe universal computation ability of pre-trained transformer models, we propose\na novel method termed CeTaD (Considering Pre-trained Transformers as Defenders)\nfor RaPiD, optimized for efficient computation. CeTaD strategically fine-tunes\nthe normalization layer parameters within the defender using a limited set of\nclean and adversarial examples. Our evaluation centers on assessing CeTaD's\neffectiveness, transferability, and the impact of different components in\nscenarios involving one-shot adversarial examples. The proposed method is\ncapable of rapidly adapting to various attacks and different application\nscenarios without altering the target model and clean training data. We also\nexplore the influence of varying training data conditions on CeTaD's\nperformance. Notably, CeTaD exhibits adaptability across differentiable service\nmodels and proves the potential of continuous learning.\n", "date": "2023-05-27", "categories": []}
{"id": "2312.05508", "title": "Improving Adversarial Robust Fairness via Anti-Bias Soft Label\n  Distillation", "abstract": "  Adversarial Training (AT) has been widely proved to be an effective method to\nimprove the adversarial robustness against adversarial examples for Deep Neural\nNetworks (DNNs). As a variant of AT, Adversarial Robustness Distillation (ARD)\nhas demonstrated its superior performance in improving the robustness of small\nstudent models with the guidance of large teacher models. However, both AT and\nARD encounter the robust fairness problem: these models exhibit strong\nrobustness when facing part of classes (easy class), but weak robustness when\nfacing others (hard class). In this paper, we give an in-depth analysis of the\npotential factors and argue that the smoothness degree of samples' soft labels\nfor different classes (i.e., hard class or easy class) will affect the robust\nfairness of DNNs from both empirical observation and theoretical analysis.\nBased on the above finding, we propose an Anti-Bias Soft Label Distillation\n(ABSLD) method to mitigate the adversarial robust fairness problem within the\nframework of Knowledge Distillation (KD). Specifically, ABSLD adaptively\nreduces the student's error risk gap between different classes to achieve\nfairness by adjusting the class-wise smoothness degree of samples' soft labels\nduring the training process, and the smoothness degree of soft labels is\ncontrolled by assigning different temperatures in KD to different classes.\nExtensive experiments demonstrate that ABSLD outperforms state-of-the-art AT,\nARD, and robust fairness methods in the comprehensive metric (Normalized\nStandard Deviation) of robustness and fairness.\n", "date": "2023-12-09", "categories": []}
{"id": "2404.15154", "title": "Do not think about pink elephant!", "abstract": "  Large Models (LMs) have heightened expectations for the potential of general\nAI as they are akin to human intelligence. This paper shows that recent large\nmodels such as Stable Diffusion and DALL-E3 also share the vulnerability of\nhuman intelligence, namely the \"white bear phenomenon\". We investigate the\ncauses of the white bear phenomenon by analyzing their representation space.\nBased on this analysis, we propose a simple prompt-based attack method, which\ngenerates figures prohibited by the LM provider's policy. To counter these\nattacks, we introduce prompt-based defense strategies inspired by cognitive\ntherapy techniques, successfully mitigating attacks by up to 48.22\\%.\n", "date": "2024-04-22", "categories": []}
{"id": "2410.23483", "title": "Keep on Swimming: Real Attackers Only Need Partial Knowledge of a\n  Multi-Model System", "abstract": "  Recent approaches in machine learning often solve a task using a composition\nof multiple models or agentic architectures. When targeting a composed system\nwith adversarial attacks, it might not be computationally or informationally\nfeasible to train an end-to-end proxy model or a proxy model for every\ncomponent of the system. We introduce a method to craft an adversarial attack\nagainst the overall multi-model system when we only have a proxy model for the\nfinal black-box model, and when the transformation applied by the initial\nmodels can make the adversarial perturbations ineffective. Current methods\nhandle this by applying many copies of the first model/transformation to an\ninput and then re-use a standard adversarial attack by averaging gradients, or\nlearning a proxy model for both stages. To our knowledge, this is the first\nattack specifically designed for this threat model and our method has a\nsubstantially higher attack success rate (80% vs 25%) and contains 9.4% smaller\nperturbations (MSE) compared to prior state-of-the-art methods. Our experiments\nfocus on a supervised image pipeline, but we are confident the attack will\ngeneralize to other multi-model settings [e.g. a mix of open/closed source\nfoundation models], or agentic systems\n", "date": "2024-10-30", "categories": []}
{"id": "2410.23870", "title": "Noise as a Double-Edged Sword: Reinforcement Learning Exploits\n  Randomized Defenses in Neural Networks", "abstract": "  This study investigates a counterintuitive phenomenon in adversarial machine\nlearning: the potential for noise-based defenses to inadvertently aid evasion\nattacks in certain scenarios. While randomness is often employed as a defensive\nstrategy against adversarial examples, our research reveals that this approach\ncan sometimes backfire, particularly when facing adaptive attackers using\nreinforcement learning (RL). Our findings show that in specific cases,\nespecially with visually noisy classes, the introduction of noise in the\nclassifier's confidence values can be exploited by the RL attacker, leading to\na significant increase in evasion success rates. In some instances, the\nnoise-based defense scenario outperformed other strategies by up to 20\\% on a\nsubset of classes. However, this effect was not consistent across all\nclassifiers tested, highlighting the complexity of the interaction between\nnoise-based defenses and different models. These results suggest that in some\ncases, noise-based defenses can inadvertently create an adversarial training\nloop beneficial to the RL attacker. Our study emphasizes the need for a more\nnuanced approach to defensive strategies in adversarial machine learning,\nparticularly in safety-critical applications. It challenges the assumption that\nrandomness universally enhances defense against evasion attacks and highlights\nthe importance of considering adaptive, RL-based attackers when designing\nrobust defense mechanisms.\n", "date": "2024-10-31", "categories": []}
{"id": "2403.01446", "title": "GuardT2I: Defending Text-to-Image Models from Adversarial Prompts", "abstract": "  Recent advancements in Text-to-Image (T2I) models have raised significant\nsafety concerns about their potential misuse for generating inappropriate or\nNot-Safe-For-Work (NSFW) contents, despite existing countermeasures such as\nNSFW classifiers or model fine-tuning for inappropriate concept removal.\nAddressing this challenge, our study unveils GuardT2I, a novel moderation\nframework that adopts a generative approach to enhance T2I models' robustness\nagainst adversarial prompts. Instead of making a binary classification,\nGuardT2I utilizes a Large Language Model (LLM) to conditionally transform text\nguidance embeddings within the T2I models into natural language for effective\nadversarial prompt detection, without compromising the models' inherent\nperformance. Our extensive experiments reveal that GuardT2I outperforms leading\ncommercial solutions like OpenAI-Moderation and Microsoft Azure Moderator by a\nsignificant margin across diverse adversarial scenarios. Our framework is\navailable at https://github.com/cure-lab/GuardT2I.\n", "date": "2024-03-03", "categories": []}
{"id": "2406.03052", "title": "Are Your Models Still Fair? Fairness Attacks on Graph Neural Networks\n  via Node Injections", "abstract": "  Despite the remarkable capabilities demonstrated by Graph Neural Networks\n(GNNs) in graph-related tasks, recent research has revealed the fairness\nvulnerabilities in GNNs when facing malicious adversarial attacks. However, all\nexisting fairness attacks require manipulating the connectivity between\nexisting nodes, which may be prohibited in reality. To this end, we introduce a\nNode Injection-based Fairness Attack (NIFA), exploring the vulnerabilities of\nGNN fairness in such a more realistic setting. In detail, NIFA first designs\ntwo insightful principles for node injection operations, namely the\nuncertainty-maximization principle and homophily-increase principle, and then\noptimizes injected nodes' feature matrix to further ensure the effectiveness of\nfairness attacks. Comprehensive experiments on three real-world datasets\nconsistently demonstrate that NIFA can significantly undermine the fairness of\nmainstream GNNs, even including fairness-aware GNNs, by injecting merely 1% of\nnodes. We sincerely hope that our work can stimulate increasing attention from\nresearchers on the vulnerability of GNN fairness, and encourage the development\nof corresponding defense mechanisms. Our code and data are released at:\nhttps://github.com/CGCL-codes/NIFA.\n", "date": "2024-06-05", "categories": []}
{"id": "2410.21802", "title": "Text-Guided Attention is All You Need for Zero-Shot Robustness in\n  Vision-Language Models", "abstract": "  Due to the impressive zero-shot capabilities, pre-trained vision-language\nmodels (e.g. CLIP), have attracted widespread attention and adoption across\nvarious domains. Nonetheless, CLIP has been observed to be susceptible to\nadversarial examples. Through experimental analysis, we have observed a\nphenomenon wherein adversarial perturbations induce shifts in text-guided\nattention. Building upon this observation, we propose a simple yet effective\nstrategy: Text-Guided Attention for Zero-Shot Robustness (TGA-ZSR). This\nframework incorporates two components: the Attention Refinement module and the\nAttention-based Model Constraint module. Our goal is to maintain the\ngeneralization of the CLIP model and enhance its adversarial robustness: The\nAttention Refinement module aligns the text-guided attention obtained from the\ntarget model via adversarial examples with the text-guided attention acquired\nfrom the original model via clean examples. This alignment enhances the model's\nrobustness. Additionally, the Attention-based Model Constraint module acquires\ntext-guided attention from both the target and original models using clean\nexamples. Its objective is to maintain model performance on clean samples while\nenhancing overall robustness. The experiments validate that our method yields a\n9.58% enhancement in zero-shot robust accuracy over the current\nstate-of-the-art techniques across 16 datasets. Our code is available at\nhttps://github.com/zhyblue424/TGA-ZSR.\n", "date": "2024-10-29", "categories": []}
{"id": "2410.23182", "title": "ProTransformer: Robustify Transformers via Plug-and-Play Paradigm", "abstract": "  Transformer-based architectures have dominated various areas of machine\nlearning in recent years. In this paper, we introduce a novel robust attention\nmechanism designed to enhance the resilience of transformer-based\narchitectures. Crucially, this technique can be integrated into existing\ntransformers as a plug-and-play layer, improving their robustness without the\nneed for additional training or fine-tuning. Through comprehensive experiments\nand ablation studies, we demonstrate that our ProTransformer significantly\nenhances the robustness of transformer models across a variety of prediction\ntasks, attack mechanisms, backbone architectures, and data domains. Notably,\nwithout further fine-tuning, the ProTransformer consistently improves the\nperformance of vanilla transformers by 19.5%, 28.3%, 16.1%, and 11.4% for BERT,\nALBERT, DistilBERT, and RoBERTa, respectively, under the classical TextFooler\nattack. Furthermore, ProTransformer shows promising resilience in large\nlanguage models (LLMs) against prompting-based attacks, improving the\nperformance of T5 and LLaMA by 24.8% and 17.8%, respectively, and enhancing\nVicuna by an average of 10.4% against the Jailbreaking attack. Beyond the\nlanguage domain, ProTransformer also demonstrates outstanding robustness in\nboth vision and graph domains.\n", "date": "2024-10-30", "categories": []}
{"id": "2410.21791", "title": "Enhancing Adversarial Attacks through Chain of Thought", "abstract": "  Large language models (LLMs) have demonstrated impressive performance across\nvarious domains but remain susceptible to safety concerns. Prior research\nindicates that gradient-based adversarial attacks are particularly effective\nagainst aligned LLMs and the chain of thought (CoT) prompting can elicit\ndesired answers through step-by-step reasoning. This paper proposes enhancing\nthe robustness of adversarial attacks on aligned LLMs by integrating CoT\nprompts with the greedy coordinate gradient (GCG) technique. Using CoT triggers\ninstead of affirmative targets stimulates the reasoning abilities of backend\nLLMs, thereby improving the transferability and universality of adversarial\nattacks. We conducted an ablation study comparing our CoT-GCG approach with\nAmazon Web Services auto-cot. Results revealed our approach outperformed both\nthe baseline GCG attack and CoT prompting. Additionally, we used Llama Guard to\nevaluate potentially harmful interactions, providing a more objective risk\nassessment of entire conversations compared to matching outputs to rejection\nphrases. The code of this paper is available at\nhttps://github.com/sujingbo0217/CS222W24-LLM-Attack.\n", "date": "2024-10-29", "categories": []}
{"id": "2410.22143", "title": "AmpleGCG-Plus: A Strong Generative Model of Adversarial Suffixes to\n  Jailbreak LLMs with Higher Success Rates in Fewer Attempts", "abstract": "  Although large language models (LLMs) are typically aligned, they remain\nvulnerable to jailbreaking through either carefully crafted prompts in natural\nlanguage or, interestingly, gibberish adversarial suffixes. However, gibberish\ntokens have received relatively less attention despite their success in\nattacking aligned LLMs. Recent work, AmpleGCG~\\citep{liao2024amplegcg},\ndemonstrates that a generative model can quickly produce numerous customizable\ngibberish adversarial suffixes for any harmful query, exposing a range of\nalignment gaps in out-of-distribution (OOD) language spaces. To bring more\nattention to this area, we introduce AmpleGCG-Plus, an enhanced version that\nachieves better performance in fewer attempts. Through a series of exploratory\nexperiments, we identify several training strategies to improve the learning of\ngibberish suffixes. Our results, verified under a strict evaluation setting,\nshow that it outperforms AmpleGCG on both open-weight and closed-source models,\nachieving increases in attack success rate (ASR) of up to 17\\% in the white-box\nsetting against Llama-2-7B-chat, and more than tripling ASR in the black-box\nsetting against GPT-4. Notably, AmpleGCG-Plus jailbreaks the newer GPT-4o\nseries of models at similar rates to GPT-4, and, uncovers vulnerabilities\nagainst the recently proposed circuit breakers defense. We publicly release\nAmpleGCG-Plus along with our collected training datasets.\n", "date": "2024-10-29", "categories": []}
{"id": "2208.03923", "title": "Adversarial robustness of VAEs through the lens of local geometry", "abstract": "  In an unsupervised attack on variational autoencoders (VAEs), an adversary\nfinds a small perturbation in an input sample that significantly changes its\nlatent space encoding, thereby compromising the reconstruction for a fixed\ndecoder. A known reason for such vulnerability is the distortions in the latent\nspace resulting from a mismatch between approximated latent posterior and a\nprior distribution. Consequently, a slight change in an input sample can move\nits encoding to a low/zero density region in the latent space resulting in an\nunconstrained generation. This paper demonstrates that an optimal way for an\nadversary to attack VAEs is to exploit a directional bias of a stochastic\npullback metric tensor induced by the encoder and decoder networks. The\npullback metric tensor of an encoder measures the change in infinitesimal\nlatent volume from an input to a latent space. Thus, it can be viewed as a lens\nto analyse the effect of input perturbations leading to latent space\ndistortions. We propose robustness evaluation scores using the eigenspectrum of\na pullback metric tensor. Moreover, we empirically show that the scores\ncorrelate with the robustness parameter $\\beta$ of the $\\beta-$VAE. Since\nincreasing $\\beta$ also degrades reconstruction quality, we demonstrate a\nsimple alternative using \\textit{mixup} training to fill the empty regions in\nthe latent space, thus improving robustness with improved reconstruction.\n", "date": "2022-08-08", "categories": []}
{"id": "2304.09875", "title": "GREAT Score: Global Robustness Evaluation of Adversarial Perturbation\n  using Generative Models", "abstract": "  Current studies on adversarial robustness mainly focus on aggregating local\nrobustness results from a set of data samples to evaluate and rank different\nmodels. However, the local statistics may not well represent the true global\nrobustness of the underlying unknown data distribution. To address this\nchallenge, this paper makes the first attempt to present a new framework,\ncalled GREAT Score , for global robustness evaluation of adversarial\nperturbation using generative models. Formally, GREAT Score carries the\nphysical meaning of a global statistic capturing a mean certified attack-proof\nperturbation level over all samples drawn from a generative model. For\nfinite-sample evaluation, we also derive a probabilistic guarantee on the\nsample complexity and the difference between the sample mean and the true mean.\nGREAT Score has several advantages: (1) Robustness evaluations using GREAT\nScore are efficient and scalable to large models, by sparing the need of\nrunning adversarial attacks. In particular, we show high correlation and\nsignificantly reduced computation cost of GREAT Score when compared to the\nattack-based model ranking on RobustBench (Croce,et. al. 2021). (2) The use of\ngenerative models facilitates the approximation of the unknown data\ndistribution. In our ablation study with different generative adversarial\nnetworks (GANs), we observe consistency between global robustness evaluation\nand the quality of GANs. (3) GREAT Score can be used for remote auditing of\nprivacy-sensitive black-box models, as demonstrated by our robustness\nevaluation on several online facial recognition services.\n", "date": "2023-04-19", "categories": []}
{"id": "2305.13948", "title": "Decoupled Kullback-Leibler Divergence Loss", "abstract": "  In this paper, we delve deeper into the Kullback-Leibler (KL) Divergence loss\nand mathematically prove that it is equivalent to the Decoupled\nKullback-Leibler (DKL) Divergence loss that consists of 1) a weighted Mean\nSquare Error (wMSE) loss and 2) a Cross-Entropy loss incorporating soft labels.\nThanks to the decomposed formulation of DKL loss, we have identified two areas\nfor improvement. Firstly, we address the limitation of KL/DKL in scenarios like\nknowledge distillation by breaking its asymmetric optimization property. This\nmodification ensures that the $\\mathbf{w}$MSE component is always effective\nduring training, providing extra constructive cues. Secondly, we introduce\nclass-wise global information into KL/DKL to mitigate bias from individual\nsamples. With these two enhancements, we derive the Improved Kullback-Leibler\n(IKL) Divergence loss and evaluate its effectiveness by conducting experiments\non CIFAR-10/100 and ImageNet datasets, focusing on adversarial training, and\nknowledge distillation tasks. The proposed approach achieves new\nstate-of-the-art adversarial robustness on the public leaderboard --\nRobustBench and competitive performance on knowledge distillation,\ndemonstrating the substantial practical merits. Our code is available at\nhttps://github.com/jiequancui/DKL.\n", "date": "2023-05-23", "categories": []}
{"id": "2401.09673", "title": "Artwork Protection Against Neural Style Transfer Using Locally Adaptive\n  Adversarial Color Attack", "abstract": "  Neural style transfer (NST) generates new images by combining the style of\none image with the content of another. However, unauthorized NST can exploit\nartwork, raising concerns about artists' rights and motivating the development\nof proactive protection methods. We propose Locally Adaptive Adversarial Color\nAttack (LAACA), empowering artists to protect their artwork from unauthorized\nstyle transfer by processing before public release. By delving into the\nintricacies of human visual perception and the role of different frequency\ncomponents, our method strategically introduces frequency-adaptive\nperturbations in the image. These perturbations significantly degrade the\ngeneration quality of NST while maintaining an acceptable level of visual\nchange in the original image, ensuring that potential infringers are\ndiscouraged from using the protected artworks, because of its bad NST\ngeneration quality. Additionally, existing metrics often overlook the\nimportance of color fidelity in evaluating color-mattered tasks, such as the\nquality of NST-generated images, which is crucial in the context of artistic\nworks. To comprehensively assess the color-mattered tasks, we propose the\nAdversarial Color Distance Metric (ACDM), designed to quantify the color\ndifference of images pre- and post-manipulations. Experimental results confirm\nthat attacking NST using LAACA results in visually inferior style transfer, and\nthe ACDM can efficiently measure color-mattered tasks. By providing artists\nwith a tool to safeguard their intellectual property, our work relieves the\nsocio-technical challenges posed by the misuse of NST in the art community.\n", "date": "2024-01-18", "categories": []}
{"id": "2410.18267", "title": "Backdoor in Seconds: Unlocking Vulnerabilities in Large Pre-trained\n  Models via Model Editing", "abstract": "  Large pre-trained models have achieved notable success across a range of\ndownstream tasks. However, recent research shows that a type of adversarial\nattack ($\\textit{i.e.,}$ backdoor attack) can manipulate the behavior of\nmachine learning models through contaminating their training dataset, posing\nsignificant threat in the real-world application of large pre-trained model,\nespecially for those customized models. Therefore, addressing the unique\nchallenges for exploring vulnerability of pre-trained models is of paramount\nimportance. Through empirical studies on the capability for performing backdoor\nattack in large pre-trained models ($\\textit{e.g.,}$ ViT), we find the\nfollowing unique challenges of attacking large pre-trained models: 1) the\ninability to manipulate or even access large training datasets, and 2) the\nsubstantial computational resources required for training or fine-tuning these\nmodels. To address these challenges, we establish new standards for an\neffective and feasible backdoor attack in the context of large pre-trained\nmodels. In line with these standards, we introduce our EDT model, an\n\\textbf{E}fficient, \\textbf{D}ata-free, \\textbf{T}raining-free backdoor attack\nmethod. Inspired by model editing techniques, EDT injects an editing-based\nlightweight codebook into the backdoor of large pre-trained models, which\nreplaces the embedding of the poisoned image with the target image without\npoisoning the training dataset or training the victim model. Our experiments,\nconducted across various pre-trained models such as ViT, CLIP, BLIP, and stable\ndiffusion, and on downstream tasks including image classification, image\ncaptioning, and image generation, demonstrate the effectiveness of our method.\nOur code is available in the supplementary material.\n", "date": "2024-10-23", "categories": []}
{"id": "2410.20019", "title": "Attacks against Abstractive Text Summarization Models through Lead Bias\n  and Influence Functions", "abstract": "  Large Language Models have introduced novel opportunities for text\ncomprehension and generation. Yet, they are vulnerable to adversarial\nperturbations and data poisoning attacks, particularly in tasks like text\nclassification and translation. However, the adversarial robustness of\nabstractive text summarization models remains less explored. In this work, we\nunveil a novel approach by exploiting the inherent lead bias in summarization\nmodels, to perform adversarial perturbations. Furthermore, we introduce an\ninnovative application of influence functions, to execute data poisoning, which\ncompromises the model's integrity. This approach not only shows a skew in the\nmodels behavior to produce desired outcomes but also shows a new behavioral\nchange, where models under attack tend to generate extractive summaries rather\nthan abstractive summaries.\n", "date": "2024-10-26", "categories": []}
{"id": "2410.20097", "title": "Generative Adversarial Patches for Physical Attacks on Cross-Modal\n  Pedestrian Re-Identification", "abstract": "  Visible-infrared pedestrian Re-identification (VI-ReID) aims to match\npedestrian images captured by infrared cameras and visible cameras. However,\nVI-ReID, like other traditional cross-modal image matching tasks, poses\nsignificant challenges due to its human-centered nature. This is evidenced by\nthe shortcomings of existing methods, which struggle to extract common features\nacross modalities, while losing valuable information when bridging the gap\nbetween them in the implicit feature space, potentially compromising security.\nTo address this vulnerability, this paper introduces the first physical\nadversarial attack against VI-ReID models. Our method, termed Edge-Attack,\nspecifically tests the models' ability to leverage deep-level implicit features\nby focusing on edge information, the most salient explicit feature\ndifferentiating individuals across modalities. Edge-Attack utilizes a novel\ntwo-step approach. First, a multi-level edge feature extractor is trained in a\nself-supervised manner to capture discriminative edge representations for each\nindividual. Second, a generative model based on Vision Transformer Generative\nAdversarial Networks (ViTGAN) is employed to generate adversarial patches\nconditioned on the extracted edge features. By applying these patches to\npedestrian clothing, we create realistic, physically-realizable adversarial\nsamples. This black-box, self-supervised approach ensures the generalizability\nof our attack against various VI-ReID models. Extensive experiments on\nSYSU-MM01 and RegDB datasets, including real-world deployments, demonstrate the\neffectiveness of Edge- Attack in significantly degrading the performance of\nstate-of-the-art VI-ReID methods.\n", "date": "2024-10-26", "categories": []}
{"id": "2410.20284", "title": "Classification under strategic adversary manipulation using pessimistic\n  bilevel optimisation", "abstract": "  Adversarial machine learning concerns situations in which learners face\nattacks from active adversaries. Such scenarios arise in applications such as\nspam email filtering, malware detection and fake-image generation, where\nsecurity methods must be actively updated to keep up with the ever improving\ngeneration of malicious data.We model these interactions between the learner\nand the adversary as a game and formulate the problem as a pessimistic bilevel\noptimisation problem with the learner taking the role of the leader. The\nadversary, modelled as a stochastic data generator, takes the role of the\nfollower, generating data in response to the classifier. While existing models\nrely on the assumption that the adversary will choose the least costly solution\nleading to a convex lower-level problem with a unique solution, we present a\nnovel model and solution method which do not make such assumptions. We compare\nthese to the existing approach and see significant improvements in performance\nsuggesting that relaxing these assumptions leads to a more realistic model.\n", "date": "2024-10-26", "categories": []}
{"id": "2410.20432", "title": "Integrating uncertainty quantification into randomized smoothing based\n  robustness guarantees", "abstract": "  Deep neural networks have proven to be extremely powerful, however, they are\nalso vulnerable to adversarial attacks which can cause hazardous incorrect\npredictions in safety-critical applications. Certified robustness via\nrandomized smoothing gives a probabilistic guarantee that the smoothed\nclassifier's predictions will not change within an $\\ell_2$-ball around a given\ninput. On the other hand (uncertainty) score-based rejection is a technique\noften applied in practice to defend models against adversarial attacks. In this\nwork, we fuse these two approaches by integrating a classifier that abstains\nfrom predicting when uncertainty is high into the certified robustness\nframework. This allows us to derive two novel robustness guarantees for\nuncertainty aware classifiers, namely (i) the radius of an $\\ell_2$-ball around\nthe input in which the same label is predicted and uncertainty remains low and\n(ii) the $\\ell_2$-radius of a ball in which the predictions will either not\nchange or be uncertain. While the former provides robustness guarantees with\nrespect to attacks aiming at increased uncertainty, the latter informs about\nthe amount of input perturbation necessary to lead the uncertainty aware model\ninto a wrong prediction. Notably, this is on CIFAR10 up to 20.93% larger than\nfor models not allowing for uncertainty based rejection. We demonstrate, that\nthe novel framework allows for a systematic robustness evaluation of different\nnetwork architectures and uncertainty measures and to identify desired\nproperties of uncertainty quantification techniques. Moreover, we show that\nleveraging uncertainty in a smoothed classifier helps out-of-distribution\ndetection.\n", "date": "2024-10-27", "categories": []}
{"id": "2303.12249", "title": "State-of-the-art optical-based physical adversarial attacks for deep\n  learning computer vision systems", "abstract": "  Adversarial attacks can mislead deep learning models to make false\npredictions by implanting small perturbations to the original input that are\nimperceptible to the human eye, which poses a huge security threat to the\ncomputer vision systems based on deep learning. Physical adversarial attacks,\nwhich is more realistic, as the perturbation is introduced to the input before\nit is being captured and converted to a binary image inside the vision system,\nwhen compared to digital adversarial attacks. In this paper, we focus on\nphysical adversarial attacks and further classify them into invasive and\nnon-invasive. Optical-based physical adversarial attack techniques (e.g. using\nlight irradiation) belong to the non-invasive category. As the perturbations\ncan be easily ignored by humans as the perturbations are very similar to the\neffects generated by a natural environment in the real world. They are highly\ninvisibility and executable and can pose a significant or even lethal threats\nto real systems. This paper focuses on optical-based physical adversarial\nattack techniques for computer vision systems, with emphasis on the\nintroduction and discussion of optical-based physical adversarial attack\ntechniques.\n", "date": "2023-03-22", "categories": []}
{"id": "2307.12872", "title": "Latent Code Augmentation Based on Stable Diffusion for Data-free\n  Substitute Attacks", "abstract": "  Since the training data of the target model is not available in the black-box\nsubstitute attack, most recent schemes utilize GANs to generate data for\ntraining the substitute model. However, these GANs-based schemes suffer from\nlow training efficiency as the generator needs to be retrained for each target\nmodel during the substitute training process, as well as low generation\nquality. To overcome these limitations, we consider utilizing the diffusion\nmodel to generate data, and propose a novel data-free substitute attack scheme\nbased on the Stable Diffusion (SD) to improve the efficiency and accuracy of\nsubstitute training. Despite the data generated by the SD exhibiting high\nquality, it presents a different distribution of domains and a large variation\nof positive and negative samples for the target model. For this problem, we\npropose Latent Code Augmentation (LCA) to facilitate SD in generating data that\naligns with the data distribution of the target model. Specifically, we augment\nthe latent codes of the inferred member data with LCA and use them as guidance\nfor SD. With the guidance of LCA, the data generated by the SD not only meets\nthe discriminative criteria of the target model but also exhibits high\ndiversity. By utilizing this data, it is possible to train the substitute model\nthat closely resembles the target model more efficiently. Extensive experiments\ndemonstrate that our LCA achieves higher attack success rates and requires\nfewer query budgets compared to GANs-based schemes for different target models.\nOur codes are available at \\url{https://github.com/LzhMeng/LCA}.\n", "date": "2023-07-24", "categories": []}
{"id": "2312.17431", "title": "MVPatch: More Vivid Patch for Adversarial Camouflaged Attacks on Object\n  Detectors in the Physical World", "abstract": "  Recent studies have shown that Adversarial Patches (APs) can effectively\nmanipulate object detection models. However, the conspicuous patterns often\nassociated with these patches tend to attract human attention, posing a\nsignificant challenge. Existing research has primarily focused on enhancing\nattack efficacy in the physical domain while often neglecting the optimization\nof stealthiness and transferability. Furthermore, applying APs in real-world\nscenarios faces major challenges related to transferability, stealthiness, and\npracticality. To address these challenges, we introduce generalization theory\ninto the context of APs, enabling our iterative process to simultaneously\nenhance transferability and refine visual correlation with realistic images. We\npropose a Dual-Perception-Based Framework (DPBF) to generate the More Vivid\nPatch (MVPatch), which enhances transferability, stealthiness, and\npracticality. The DPBF integrates two key components: the\nModel-Perception-Based Module (MPBM) and the Human-Perception-Based Module\n(HPBM), along with regularization terms. The MPBM employs ensemble strategy to\nreduce object confidence scores across multiple detectors, thereby improving AP\ntransferability with robust theoretical support. Concurrently, the HPBM\nintroduces a lightweight method for achieving visual similarity, creating\nnatural and inconspicuous adversarial patches without relying on additional\ngenerative models. The regularization terms further enhance the practicality of\nthe generated APs in the physical domain. Additionally, we introduce\nnaturalness and transferability scores to provide an unbiased assessment of\nAPs. Extensive experimental validation demonstrates that MVPatch achieves\nsuperior transferability and a natural appearance in both digital and physical\ndomains, underscoring its effectiveness and stealthiness.\n", "date": "2023-12-29", "categories": []}
{"id": "2409.13770", "title": "A constrained optimization approach to improve robustness of neural\n  networks", "abstract": "  In this paper, we present a novel nonlinear programming-based approach to\nfine-tune pre-trained neural networks to improve robustness against adversarial\nattacks while maintaining high accuracy on clean data. Our method introduces\nadversary-correction constraints to ensure correct classification of\nadversarial data and minimizes changes to the model parameters. We propose an\nefficient cutting-plane-based algorithm to iteratively solve the large-scale\nnonconvex optimization problem by approximating the feasible region through\npolyhedral cuts and balancing between robustness and accuracy. Computational\nexperiments on standard datasets such as MNIST and CIFAR10 demonstrate that the\nproposed approach significantly improves robustness, even with a very small set\nof adversarial data, while maintaining minimal impact on accuracy.\n", "date": "2024-09-18", "categories": []}
{"id": "2410.19160", "title": "Adversarial Attacks on Large Language Models Using Regularized\n  Relaxation", "abstract": "  As powerful Large Language Models (LLMs) are now widely used for numerous\npractical applications, their safety is of critical importance. While alignment\ntechniques have significantly improved overall safety, LLMs remain vulnerable\nto carefully crafted adversarial inputs. Consequently, adversarial attack\nmethods are extensively used to study and understand these vulnerabilities.\nHowever, current attack methods face significant limitations. Those relying on\noptimizing discrete tokens suffer from limited efficiency, while continuous\noptimization techniques fail to generate valid tokens from the model's\nvocabulary, rendering them impractical for real-world applications. In this\npaper, we propose a novel technique for adversarial attacks that overcomes\nthese limitations by leveraging regularized gradients with continuous\noptimization methods. Our approach is two orders of magnitude faster than the\nstate-of-the-art greedy coordinate gradient-based method, significantly\nimproving the attack success rate on aligned language models. Moreover, it\ngenerates valid tokens, addressing a fundamental limitation of existing\ncontinuous optimization methods. We demonstrate the effectiveness of our attack\non five state-of-the-art LLMs using four datasets.\n", "date": "2024-10-24", "categories": []}
{"id": "2312.12608", "title": "Rethinking Randomized Smoothing from the Perspective of Scalability", "abstract": "  Machine learning models have demonstrated remarkable success across diverse\ndomains but remain vulnerable to adversarial attacks. Empirical defense\nmechanisms often fail, as new attacks constantly emerge, rendering existing\ndefenses obsolete, shifting the focus to certification-based defenses.\nRandomized smoothing has emerged as a promising technique among notable\nadvancements. This study reviews the theoretical foundations and empirical\neffectiveness of randomized smoothing and its derivatives in verifying machine\nlearning classifiers from a perspective of scalability. We provide an in-depth\nexploration of the fundamental concepts underlying randomized smoothing,\nhighlighting its theoretical guarantees in certifying robustness against\nadversarial perturbations and discuss the challenges of existing methodologies.\n", "date": "2023-12-19", "categories": []}
{"id": "2410.18215", "title": "Advancing NLP Security by Leveraging LLMs as Adversarial Engines", "abstract": "  This position paper proposes a novel approach to advancing NLP security by\nleveraging Large Language Models (LLMs) as engines for generating diverse\nadversarial attacks. Building upon recent work demonstrating LLMs'\neffectiveness in creating word-level adversarial examples, we argue for\nexpanding this concept to encompass a broader range of attack types, including\nadversarial patches, universal perturbations, and targeted attacks. We posit\nthat LLMs' sophisticated language understanding and generation capabilities can\nproduce more effective, semantically coherent, and human-like adversarial\nexamples across various domains and classifier architectures. This paradigm\nshift in adversarial NLP has far-reaching implications, potentially enhancing\nmodel robustness, uncovering new vulnerabilities, and driving innovation in\ndefense mechanisms. By exploring this new frontier, we aim to contribute to the\ndevelopment of more secure, reliable, and trustworthy NLP systems for critical\napplications.\n", "date": "2024-10-23", "categories": []}
{"id": "2410.18556", "title": "Complexity Matters: Effective Dimensionality as a Measure for\n  Adversarial Robustness", "abstract": "  Quantifying robustness in a single measure for the purposes of model\nselection, development of adversarial training methods, and anticipating trends\nhas so far been elusive. The simplest metric to consider is the number of\ntrainable parameters in a model but this has previously been shown to be\ninsufficient at explaining robustness properties. A variety of other metrics,\nsuch as ones based on boundary thickness and gradient flatness have been\nproposed but have been shown to be inadequate proxies for robustness.\n  In this work, we investigate the relationship between a model's effective\ndimensionality, which can be thought of as model complexity, and its robustness\nproperties. We run experiments on commercial-scale models that are often used\nin real-world environments such as YOLO and ResNet. We reveal a near-linear\ninverse relationship between effective dimensionality and adversarial\nrobustness, that is models with a lower dimensionality exhibit better\nrobustness. We investigate the effect of a variety of adversarial training\nmethods on effective dimensionality and find the same inverse linear\nrelationship present, suggesting that effective dimensionality can serve as a\nuseful criterion for model selection and robustness evaluation, providing a\nmore nuanced and effective metric than parameter count or previously-tested\nmeasures.\n", "date": "2024-10-24", "categories": []}
{"id": "2410.18648", "title": "GADT: Enhancing Transferable Adversarial Attacks through Gradient-guided\n  Adversarial Data Transformation", "abstract": "  Current Transferable Adversarial Examples (TAE) are primarily generated by\nadding Adversarial Noise (AN). Recent studies emphasize the importance of\noptimizing Data Augmentation (DA) parameters along with AN, which poses a\ngreater threat to real-world AI applications. However, existing DA-based\nstrategies often struggle to find optimal solutions due to the challenging DA\nsearch procedure without proper guidance. In this work, we propose a novel\nDA-based attack algorithm, GADT. GADT identifies suitable DA parameters through\niterative antagonism and uses posterior estimates to update AN based on these\nparameters. We uniquely employ a differentiable DA operation library to\nidentify adversarial DA parameters and introduce a new loss function as a\nmetric during DA optimization. This loss term enhances adversarial effects\nwhile preserving the original image content, maintaining attack crypticity.\nExtensive experiments on public datasets with various networks demonstrate that\nGADT can be integrated with existing transferable attack methods, updating\ntheir DA parameters effectively while retaining their AN formulation\nstrategies. Furthermore, GADT can be utilized in other black-box attack\nscenarios, e.g., query-based attacks, offering a new avenue to enhance attacks\non real-world AI applications in both research and industrial contexts.\n", "date": "2024-10-24", "categories": []}
{"id": "2312.10132", "title": "Closing the Gap: Achieving Better Accuracy-Robustness Tradeoffs against\n  Query-Based Attacks", "abstract": "  Although promising, existing defenses against query-based attacks share a\ncommon limitation: they offer increased robustness against attacks at the price\nof a considerable accuracy drop on clean samples. In this work, we show how to\nefficiently establish, at test-time, a solid tradeoff between robustness and\naccuracy when mitigating query-based attacks. Given that these attacks\nnecessarily explore low-confidence regions, our insight is that activating\ndedicated defenses, such as random noise defense and random image\ntransformations, only for low-confidence inputs is sufficient to prevent them.\nOur approach is independent of training and supported by theory. We verify the\neffectiveness of our approach for various existing defenses by conducting\nextensive experiments on CIFAR-10, CIFAR-100, and ImageNet. Our results confirm\nthat our proposal can indeed enhance these defenses by providing better\ntradeoffs between robustness and accuracy when compared to state-of-the-art\napproaches while being completely training-free.\n", "date": "2023-12-15", "categories": []}
{"id": "2312.11309", "title": "The Ultimate Combo: Boosting Adversarial Example Transferability by\n  Composing Data Augmentations", "abstract": "  To help adversarial examples generalize from surrogate machine-learning (ML)\nmodels to targets, certain transferability-based black-box evasion attacks\nincorporate data augmentations (e.g., random resizing). Yet, prior work has\nexplored limited augmentations and their composition. To fill the gap, we\nsystematically studied how data augmentation affects transferability.\nSpecifically, we explored 46 augmentation techniques originally proposed to\nhelp ML models generalize to unseen benign samples, and assessed how they\nimpact transferability, when applied individually or composed. Performing\nexhaustive search on a small subset of augmentation techniques and genetic\nsearch on all techniques, we identified augmentation combinations that help\npromote transferability. Extensive experiments with the ImageNet and CIFAR-10\ndatasets and 18 models showed that simple color-space augmentations (e.g.,\ncolor to greyscale) attain high transferability when combined with standard\naugmentations. Furthermore, we discovered that composing augmentations impacts\ntransferability mostly monotonically (i.e., more augmentations $\\rightarrow$\n$\\ge$transferability). We also found that the best composition significantly\noutperformed the state of the art (e.g., 91.8% vs. $\\le$82.5% average\ntransferability to adversarially trained targets on ImageNet). Lastly, our\ntheoretical analysis, backed by empirical evidence, intuitively explains why\ncertain augmentations promote transferability.\n", "date": "2023-12-18", "categories": []}
{"id": "2403.14774", "title": "Few-Shot Adversarial Prompt Learning on Vision-Language Models", "abstract": "  The vulnerability of deep neural networks to imperceptible adversarial\nperturbations has attracted widespread attention. Inspired by the success of\nvision-language foundation models, previous efforts achieved zero-shot\nadversarial robustness by aligning adversarial visual features with text\nsupervision. However, in practice, they are still unsatisfactory due to several\nissues, including heavy adaptation cost, suboptimal text supervision, and\nuncontrolled natural generalization capacity. In this paper, to address these\nissues, we propose a few-shot adversarial prompt framework where adapting input\nsequences with limited data makes significant adversarial robustness\nimprovement. Specifically, we achieve this by providing adversarially\ncorrelated text supervision that is end-to-end learned from adversarial\nexamples. We also propose a novel training objective that enhances the\nconsistency of multi-modal features while encourages differentiated uni-modal\nfeatures between natural and adversarial examples. The proposed framework gives\naccess to learn adversarial text supervision, which provides superior\ncross-modal adversarial alignment and matches state-of-the-art zero-shot\nadversarial robustness with only 1% training data. Code is available at:\nhttps://github.com/lionel-w2/FAP.\n", "date": "2024-03-21", "categories": []}
{"id": "2404.00362", "title": "STBA: Towards Evaluating the Robustness of DNNs for Query-Limited\n  Black-box Scenario", "abstract": "  Many attack techniques have been proposed to explore the vulnerability of\nDNNs and further help to improve their robustness. Despite the significant\nprogress made recently, existing black-box attack methods still suffer from\nunsatisfactory performance due to the vast number of queries needed to optimize\ndesired perturbations. Besides, the other critical challenge is that\nadversarial examples built in a noise-adding manner are abnormal and struggle\nto successfully attack robust models, whose robustness is enhanced by\nadversarial training against small perturbations. There is no doubt that these\ntwo issues mentioned above will significantly increase the risk of exposure and\nresult in a failure to dig deeply into the vulnerability of DNNs. Hence, it is\nnecessary to evaluate DNNs' fragility sufficiently under query-limited settings\nin a non-additional way. In this paper, we propose the Spatial Transform\nBlack-box Attack (STBA), a novel framework to craft formidable adversarial\nexamples in the query-limited scenario. Specifically, STBA introduces a flow\nfield to the high-frequency part of clean images to generate adversarial\nexamples and adopts the following two processes to enhance their naturalness\nand significantly improve the query efficiency: a) we apply an estimated flow\nfield to the high-frequency part of clean images to generate adversarial\nexamples instead of introducing external noise to the benign image, and b) we\nleverage an efficient gradient estimation method based on a batch of samples to\noptimize such an ideal flow field under query-limited settings. Compared to\nexisting score-based black-box baselines, extensive experiments indicated that\nSTBA could effectively improve the imperceptibility of the adversarial examples\nand remarkably boost the attack success rate under query-limited settings.\n", "date": "2024-03-30", "categories": []}
{"id": "2404.14693", "title": "DIP-Watermark: A Double Identity Protection Method Based on Robust\n  Adversarial Watermark", "abstract": "  The wide deployment of Face Recognition (FR) systems poses privacy risks. One\ncountermeasure is adversarial attack, deceiving unauthorized malicious FR, but\nit also disrupts regular identity verification of trusted authorizers,\nexacerbating the potential threat of identity impersonation. To address this,\nwe propose the first double identity protection scheme based on traceable\nadversarial watermarking, termed DIP-Watermark. DIP-Watermark employs a\none-time watermark embedding to deceive unauthorized FR models and allows\nauthorizers to perform identity verification by extracting the watermark.\nSpecifically, we propose an information-guided adversarial attack against FR\nmodels. The encoder embeds an identity-specific watermark into the deep feature\nspace of the carrier, guiding recognizable features of the image to deviate\nfrom the source identity. We further adopt a collaborative meta-optimization\nstrategy compatible with sub-tasks, which regularizes the joint optimization\ndirection of the encoder and decoder. This strategy enhances the representation\nof universal carrier features, mitigating multi-objective optimization\nconflicts in watermarking. Experiments confirm that DIP-Watermark achieves\nsignificant attack success rates and traceability accuracy on state-of-the-art\nFR models, exhibiting remarkable robustness that outperforms the existing\nprivacy protection methods using adversarial attacks and deep watermarking, or\nsimple combinations of the two. Our work potentially opens up new insights into\nproactive protection for FR privacy.\n", "date": "2024-04-23", "categories": []}
{"id": "2405.03891", "title": "Enhancing O-RAN Security: Evasion Attacks and Robust Defenses for Graph\n  Reinforcement Learning-based Connection Management", "abstract": "  Adversarial machine learning, focused on studying various attacks and\ndefenses on machine learning (ML) models, is rapidly gaining importance as ML\nis increasingly being adopted for optimizing wireless systems such as Open\nRadio Access Networks (O-RAN). A comprehensive modeling of the security threats\nand the demonstration of adversarial attacks and defenses on practical AI based\nO-RAN systems is still in its nascent stages. We begin by conducting threat\nmodeling to pinpoint attack surfaces in O-RAN using an ML-based Connection\nmanagement application (xApp) as an example. The xApp uses a Graph Neural\nNetwork trained using Deep Reinforcement Learning and achieves on average 54%\nimprovement in the coverage rate measured as the 5th percentile user data\nrates. We then formulate and demonstrate evasion attacks that degrade the\ncoverage rates by as much as 50% through injecting bounded noise at different\nthreat surfaces including the open wireless medium itself. Crucially, we also\ncompare and contrast the effectiveness of such attacks on the ML-based xApp and\na non-ML based heuristic. We finally develop and demonstrate robust\ntraining-based defenses against the challenging physical/jamming-based attacks\nand show a 15% improvement in the coverage rates when compared to employing no\ndefense over a range of noise budgets\n", "date": "2024-05-06", "categories": []}
{"id": "2410.03489", "title": "Gradient-based Jailbreak Images for Multimodal Fusion Models", "abstract": "  Augmenting language models with image inputs may enable more effective\njailbreak attacks through continuous optimization, unlike text inputs that\nrequire discrete optimization. However, new multimodal fusion models tokenize\nall input modalities using non-differentiable functions, which hinders\nstraightforward attacks. In this work, we introduce the notion of a tokenizer\nshortcut that approximates tokenization with a continuous function and enables\ncontinuous optimization. We use tokenizer shortcuts to create the first\nend-to-end gradient image attacks against multimodal fusion models. We evaluate\nour attacks on Chameleon models and obtain jailbreak images that elicit harmful\ninformation for 72.5% of prompts. Jailbreak images outperform text jailbreaks\noptimized with the same objective and require 3x lower compute budget to\noptimize 50x more input tokens. Finally, we find that representation\nengineering defenses, like Circuit Breakers, trained only on text attacks can\neffectively transfer to adversarial image inputs.\n", "date": "2024-10-04", "categories": []}
{"id": "2410.16317", "title": "A Survey on Physical Adversarial Attacks against Face Recognition\n  Systems", "abstract": "  As Face Recognition (FR) technology becomes increasingly prevalent in\nfinance, the military, public safety, and everyday life, security concerns have\ngrown substantially. Physical adversarial attacks targeting FR systems in\nreal-world settings have attracted considerable research interest due to their\npracticality and the severe threats they pose. However, a systematic overview\nfocused on physical adversarial attacks against FR systems is still lacking,\nhindering an in-depth exploration of the challenges and future directions in\nthis field. In this paper, we bridge this gap by comprehensively collecting and\nanalyzing physical adversarial attack methods targeting FR systems.\nSpecifically, we first investigate the key challenges of physical attacks on FR\nsystems. We then categorize existing physical attacks into three categories\nbased on the physical medium used and summarize how the research in each\ncategory has evolved to address these challenges. Furthermore, we review\ncurrent defense strategies and discuss potential future research directions.\nOur goal is to provide a fresh, comprehensive, and deep understanding of\nphysical adversarial attacks against FR systems, thereby inspiring relevant\nresearch in this area.\n", "date": "2024-10-10", "categories": []}
{"id": "2410.16341", "title": "Vulnerabilities in Machine Learning-Based Voice Disorder Detection\n  Systems", "abstract": "  The impact of voice disorders is becoming more widely acknowledged as a\npublic health issue. Several machine learning-based classifiers with the\npotential to identify disorders have been used in recent studies to\ndifferentiate between normal and pathological voices and sounds. In this paper,\nwe focus on analyzing the vulnerabilities of these systems by exploring the\npossibility of attacks that can reverse classification and compromise their\nreliability. Given the critical nature of personal health information,\nunderstanding which types of attacks are effective is a necessary first step\ntoward improving the security of such systems. Starting from the original\naudios, we implement various attack methods, including adversarial, evasion,\nand pitching techniques, and evaluate how state-of-the-art disorder detection\nmodels respond to them. Our findings identify the most effective attack\nstrategies, underscoring the need to address these vulnerabilities in\nmachine-learning systems used in the healthcare domain.\n", "date": "2024-10-21", "categories": []}
{"id": "2410.16579", "title": "Conflict-Aware Adversarial Training", "abstract": "  Adversarial training is the most effective method to obtain adversarial\nrobustness for deep neural networks by directly involving adversarial samples\nin the training procedure. To obtain an accurate and robust model, the\nweighted-average method is applied to optimize standard loss and adversarial\nloss simultaneously. In this paper, we argue that the weighted-average method\ndoes not provide the best tradeoff for the standard performance and adversarial\nrobustness. We argue that the failure of the weighted-average method is due to\nthe conflict between the gradients derived from standard and adversarial loss,\nand further demonstrate such a conflict increases with attack budget\ntheoretically and practically. To alleviate this problem, we propose a new\ntrade-off paradigm for adversarial training with a conflict-aware factor for\nthe convex combination of standard and adversarial loss, named\n\\textbf{Conflict-Aware Adversarial Training~(CA-AT)}. Comprehensive\nexperimental results show that CA-AT consistently offers a superior trade-off\nbetween standard performance and adversarial robustness under the settings of\nadversarial training from scratch and parameter-efficient finetuning.\n", "date": "2024-10-21", "categories": []}
{"id": "2404.07242", "title": "Sandwich attack: Multi-language Mixture Adaptive Attack on LLMs", "abstract": "  Large Language Models (LLMs) are increasingly being developed and applied,\nbut their widespread use faces challenges. These include aligning LLMs'\nresponses with human values to prevent harmful outputs, which is addressed\nthrough safety training methods. Even so, bad actors and malicious users have\nsucceeded in attempts to manipulate the LLMs to generate misaligned responses\nfor harmful questions such as methods to create a bomb in school labs, recipes\nfor harmful drugs, and ways to evade privacy rights. Another challenge is the\nmultilingual capabilities of LLMs, which enable the model to understand and\nrespond in multiple languages. Consequently, attackers exploit the unbalanced\npre-training datasets of LLMs in different languages and the comparatively\nlower model performance in low-resource languages than high-resource ones. As a\nresult, attackers use a low-resource languages to intentionally manipulate the\nmodel to create harmful responses. Many of the similar attack vectors have been\npatched by model providers, making the LLMs more robust against language-based\nmanipulation. In this paper, we introduce a new black-box attack vector called\nthe \\emph{Sandwich attack}: a multi-language mixture attack, which manipulates\nstate-of-the-art LLMs into generating harmful and misaligned responses. Our\nexperiments with five different models, namely Google's Bard, Gemini Pro,\nLLaMA-2-70-B-Chat, GPT-3.5-Turbo, GPT-4, and Claude-3-OPUS, show that this\nattack vector can be used by adversaries to generate harmful responses and\nelicit misaligned responses from these models. By detailing both the mechanism\nand impact of the Sandwich attack, this paper aims to guide future research and\ndevelopment towards more secure and resilient LLMs, ensuring they serve the\npublic good while minimizing potential for misuse.\n", "date": "2024-04-09", "categories": []}
{"id": "2410.15042", "title": "Adversarial Training: A Survey", "abstract": "  Adversarial training (AT) refers to integrating adversarial examples --\ninputs altered with imperceptible perturbations that can significantly impact\nmodel predictions -- into the training process. Recent studies have\ndemonstrated the effectiveness of AT in improving the robustness of deep neural\nnetworks against diverse adversarial attacks. However, a comprehensive overview\nof these developments is still missing. This survey addresses this gap by\nreviewing a broad range of recent and representative studies. Specifically, we\nfirst describe the implementation procedures and practical applications of AT,\nfollowed by a comprehensive review of AT techniques from three perspectives:\ndata enhancement, network design, and training configurations. Lastly, we\ndiscuss common challenges in AT and propose several promising directions for\nfuture research.\n", "date": "2024-10-19", "categories": []}
{"id": "2410.15409", "title": "PEAS: A Strategy for Crafting Transferable Adversarial Examples", "abstract": "  Black box attacks, where adversaries have limited knowledge of the target\nmodel, pose a significant threat to machine learning systems. Adversarial\nexamples generated with a substitute model often suffer from limited\ntransferability to the target model. While recent work explores ranking\nperturbations for improved success rates, these methods see only modest gains.\nWe propose a novel strategy called PEAS that can boost the transferability of\nexisting black box attacks. PEAS leverages the insight that samples which are\nperceptually equivalent exhibit significant variability in their adversarial\ntransferability. Our approach first generates a set of images from an initial\nsample via subtle augmentations. We then evaluate the transferability of\nadversarial perturbations on these images using a set of substitute models.\nFinally, the most transferable adversarial example is selected and used for the\nattack. Our experiments show that PEAS can double the performance of existing\nattacks, achieving a 2.5x improvement in attack success rates on average over\ncurrent ranking methods. We thoroughly evaluate PEAS on ImageNet and CIFAR-10,\nanalyze hyperparameter impacts, and provide an ablation study to isolate each\ncomponent's importance.\n", "date": "2024-10-20", "categories": []}
{"id": "2410.15889", "title": "Model Mimic Attack: Knowledge Distillation for Provably Transferable\n  Adversarial Examples", "abstract": "  The vulnerability of artificial neural networks to adversarial perturbations\nin the black-box setting is widely studied in the literature. The majority of\nattack methods to construct these perturbations suffer from an impractically\nlarge number of queries required to find an adversarial example. In this work,\nwe focus on knowledge distillation as an approach to conduct transfer-based\nblack-box adversarial attacks and propose an iterative training of the\nsurrogate model on an expanding dataset. This work is the first, to our\nknowledge, to provide provable guarantees on the success of knowledge\ndistillation-based attack on classification neural networks: we prove that if\nthe student model has enough learning capabilities, the attack on the teacher\nmodel is guaranteed to be found within the finite number of distillation\niterations.\n", "date": "2024-10-21", "categories": []}
{"id": "2410.16073", "title": "On the Geometry of Regularization in Adversarial Training:\n  High-Dimensional Asymptotics and Generalization Bounds", "abstract": "  Regularization, whether explicit in terms of a penalty in the loss or\nimplicit in the choice of algorithm, is a cornerstone of modern machine\nlearning. Indeed, controlling the complexity of the model class is particularly\nimportant when data is scarce, noisy or contaminated, as it translates a\nstatistical belief on the underlying structure of the data. This work\ninvestigates the question of how to choose the regularization norm $\\lVert\n\\cdot \\rVert$ in the context of high-dimensional adversarial training for\nbinary classification. To this end, we first derive an exact asymptotic\ndescription of the robust, regularized empirical risk minimizer for various\ntypes of adversarial attacks and regularization norms (including non-$\\ell_p$\nnorms). We complement this analysis with a uniform convergence analysis,\nderiving bounds on the Rademacher Complexity for this class of problems.\nLeveraging our theoretical results, we quantitatively characterize the\nrelationship between perturbation size and the optimal choice of $\\lVert \\cdot\n\\rVert$, confirming the intuition that, in the data scarce regime, the type of\nregularization becomes increasingly important for adversarial training as\nperturbations grow in size.\n", "date": "2024-10-21", "categories": []}
{"id": "2410.13193", "title": "Golyadkin's Torment: Doppelg\\\"angers and Adversarial Vulnerability", "abstract": "  Many machine learning (ML) classifiers are claimed to outperform humans, but\nthey still make mistakes that humans do not. The most notorious examples of\nsuch mistakes are adversarial visual metamers. This paper aims to define and\ninvestigate the phenomenon of adversarial Doppelgangers (AD), which includes\nadversarial visual metamers, and to compare the performance and robustness of\nML classifiers to human performance.\n  We find that AD are inputs that are close to each other with respect to a\nperceptual metric defined in this paper. AD are qualitatively different from\nthe usual adversarial examples. The vast majority of classifiers are vulnerable\nto AD and robustness-accuracy trade-offs may not improve them. Some\nclassification problems may not admit any AD robust classifiers because the\nunderlying classes are ambiguous. We provide criteria that can be used to\ndetermine whether a classification problem is well defined or not; describe the\nstructure and attributes of an AD-robust classifier; introduce and explore the\nnotions of conceptual entropy and regions of conceptual ambiguity for\nclassifiers that are vulnerable to AD attacks, along with methods to bound the\nAD fooling rate of an attack. We define the notion of classifiers that exhibit\nhypersensitive behavior, that is, classifiers whose only mistakes are\nadversarial Doppelgangers. Improving the AD robustness of hyper-sensitive\nclassifiers is equivalent to improving accuracy. We identify conditions\nguaranteeing that all classifiers with sufficiently high accuracy are\nhyper-sensitive.\n  Our findings are aimed at significant improvements in the reliability and\nsecurity of machine learning systems.\n", "date": "2024-10-17", "categories": []}
{"id": "2410.14089", "title": "MMAD-Purify: A Precision-Optimized Framework for Efficient and Scalable\n  Multi-Modal Attacks", "abstract": "  Neural networks have achieved remarkable performance across a wide range of\ntasks, yet they remain susceptible to adversarial perturbations, which pose\nsignificant risks in safety-critical applications. With the rise of\nmultimodality, diffusion models have emerged as powerful tools not only for\ngenerative tasks but also for various applications such as image editing,\ninpainting, and super-resolution. However, these models still lack robustness\ndue to limited research on attacking them to enhance their resilience.\nTraditional attack techniques, such as gradient-based adversarial attacks and\ndiffusion model-based methods, are hindered by computational inefficiencies and\nscalability issues due to their iterative nature. To address these challenges,\nwe introduce an innovative framework that leverages the distilled backbone of\ndiffusion models and incorporates a precision-optimized noise predictor to\nenhance the effectiveness of our attack framework. This approach not only\nenhances the attack's potency but also significantly reduces computational\ncosts. Our framework provides a cutting-edge solution for multi-modal\nadversarial attacks, ensuring reduced latency and the generation of\nhigh-fidelity adversarial examples with superior success rates. Furthermore, we\ndemonstrate that our framework achieves outstanding transferability and\nrobustness against purification defenses, outperforming existing gradient-based\nattack models in both effectiveness and efficiency.\n", "date": "2024-10-17", "categories": []}
{"id": "2410.14105", "title": "DMGNN: Detecting and Mitigating Backdoor Attacks in Graph Neural\n  Networks", "abstract": "  Recent studies have revealed that GNNs are highly susceptible to multiple\nadversarial attacks. Among these, graph backdoor attacks pose one of the most\nprominent threats, where attackers cause models to misclassify by learning the\nbackdoored features with injected triggers and modified target labels during\nthe training phase. Based on the features of the triggers, these attacks can be\ncategorized into out-of-distribution (OOD) and in-distribution (ID) graph\nbackdoor attacks, triggers with notable differences from the clean sample\nfeature distributions constitute OOD backdoor attacks, whereas the triggers in\nID backdoor attacks are nearly identical to the clean sample feature\ndistributions. Existing methods can successfully defend against OOD backdoor\nattacks by comparing the feature distribution of triggers and clean samples but\nfail to mitigate stealthy ID backdoor attacks. Due to the lack of proper\nsupervision signals, the main task accuracy is negatively affected in defending\nagainst ID backdoor attacks. To bridge this gap, we propose DMGNN against OOD\nand ID graph backdoor attacks that can powerfully eliminate stealthiness to\nguarantee defense effectiveness and improve the model performance.\nSpecifically, DMGNN can easily identify the hidden ID and OOD triggers via\npredicting label transitions based on counterfactual explanation. To further\nfilter the diversity of generated explainable graphs and erase the influence of\nthe trigger features, we present a reverse sampling pruning method to screen\nand discard the triggers directly on the data level. Extensive experimental\nevaluations on open graph datasets demonstrate that DMGNN far outperforms the\nstate-of-the-art (SOTA) defense methods, reducing the attack success rate to 5%\nwith almost negligible degradation in model performance (within 3.5%).\n", "date": "2024-10-18", "categories": []}
{"id": "2309.14122", "title": "SurrogatePrompt: Bypassing the Safety Filter of Text-to-Image Models via\n  Substitution", "abstract": "  Advanced text-to-image models such as DALL$\\cdot$E 2 and Midjourney possess\nthe capacity to generate highly realistic images, raising significant concerns\nregarding the potential proliferation of unsafe content. This includes adult,\nviolent, or deceptive imagery of political figures. Despite claims of rigorous\nsafety mechanisms implemented in these models to restrict the generation of\nnot-safe-for-work (NSFW) content, we successfully devise and exhibit the first\nprompt attacks on Midjourney, resulting in the production of abundant\nphotorealistic NSFW images. We reveal the fundamental principles of such prompt\nattacks and suggest strategically substituting high-risk sections within a\nsuspect prompt to evade closed-source safety measures. Our novel framework,\nSurrogatePrompt, systematically generates attack prompts, utilizing large\nlanguage models, image-to-text, and image-to-image modules to automate attack\nprompt creation at scale. Evaluation results disclose an 88% success rate in\nbypassing Midjourney's proprietary safety filter with our attack prompts,\nleading to the generation of counterfeit images depicting political figures in\nviolent scenarios. Both subjective and objective assessments validate that the\nimages generated from our attack prompts present considerable safety hazards.\n", "date": "2023-09-25", "categories": []}
{"id": "2410.12955", "title": "Long-Tailed Backdoor Attack Using Dynamic Data Augmentation Operations", "abstract": "  Recently, backdoor attack has become an increasing security threat to deep\nneural networks and drawn the attention of researchers. Backdoor attacks\nexploit vulnerabilities in third-party pretrained models during the training\nphase, enabling them to behave normally for clean samples and mispredict for\nsamples with specific triggers. Existing backdoor attacks mainly focus on\nbalanced datasets. However, real-world datasets often follow long-tailed\ndistributions. In this paper, for the first time, we explore backdoor attack on\nsuch datasets. Specifically, we first analyze the influence of data imbalance\non backdoor attack. Based on our analysis, we propose an effective backdoor\nattack named Dynamic Data Augmentation Operation (D$^2$AO). We design D$^2$AO\nselectors to select operations depending jointly on the class, sample type\n(clean vs. backdoored) and sample features. Meanwhile, we develop a trigger\ngenerator to generate sample-specific triggers. Through simultaneous\noptimization of the backdoored model and trigger generator, guided by dynamic\ndata augmentation operation selectors, we achieve significant advancements.\nExtensive experiments demonstrate that our method can achieve the\nstate-of-the-art attack performance while preserving the clean accuracy.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.13010", "title": "Hiding-in-Plain-Sight (HiPS) Attack on CLIP for Targetted Object Removal\n  from Images", "abstract": "  Machine learning models are known to be vulnerable to adversarial attacks,\nbut traditional attacks have mostly focused on single-modalities. With the rise\nof large multi-modal models (LMMs) like CLIP, which combine vision and language\ncapabilities, new vulnerabilities have emerged. However, prior work in\nmultimodal targeted attacks aim to completely change the model's output to what\nthe adversary wants. In many realistic scenarios, an adversary might seek to\nmake only subtle modifications to the output, so that the changes go unnoticed\nby downstream models or even by humans. We introduce Hiding-in-Plain-Sight\n(HiPS) attacks, a novel class of adversarial attacks that subtly modifies model\npredictions by selectively concealing target object(s), as if the target object\nwas absent from the scene. We propose two HiPS attack variants, HiPS-cls and\nHiPS-cap, and demonstrate their effectiveness in transferring to downstream\nimage captioning models, such as CLIP-Cap, for targeted object removal from\nimage captions.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.13122", "title": "Boosting Imperceptibility of Stable Diffusion-based Adversarial Examples\n  Generation with Momentum", "abstract": "  We propose a novel framework, Stable Diffusion-based Momentum Integrated\nAdversarial Examples (SD-MIAE), for generating adversarial examples that can\neffectively mislead neural network classifiers while maintaining visual\nimperceptibility and preserving the semantic similarity to the original class\nlabel. Our method leverages the text-to-image generation capabilities of the\nStable Diffusion model by manipulating token embeddings corresponding to the\nspecified class in its latent space. These token embeddings guide the\ngeneration of adversarial images that maintain high visual fidelity. The\nSD-MIAE framework consists of two phases: (1) an initial adversarial\noptimization phase that modifies token embeddings to produce misclassified yet\nnatural-looking images and (2) a momentum-based optimization phase that refines\nthe adversarial perturbations. By introducing momentum, our approach stabilizes\nthe optimization of perturbations across iterations, enhancing both the\nmisclassification rate and visual fidelity of the generated adversarial\nexamples. Experimental results demonstrate that SD-MIAE achieves a high\nmisclassification rate of 79%, improving by 35% over the state-of-the-art\nmethod while preserving the imperceptibility of adversarial perturbations and\nthe semantic similarity to the original class label, making it a practical\nmethod for robust adversarial evaluation.\n", "date": "2024-10-17", "categories": []}
{"id": "2410.13138", "title": "Data Defenses Against Large Language Models", "abstract": "  Large language models excel at performing inference over text to extract\ninformation, summarize information, or generate additional text. These\ninference capabilities are implicated in a variety of ethical harms spanning\nsurveillance, labor displacement, and IP/copyright theft. While many policy,\nlegal, and technical mitigations have been proposed to counteract these harms,\nthese mitigations typically require cooperation from institutions that move\nslower than technical advances (i.e., governments) or that have few incentives\nto act to counteract these harms (i.e., the corporations that create and profit\nfrom these LLMs). In this paper, we define and build \"data defenses\" -- a novel\nstrategy that directly empowers data owners to block LLMs from performing\ninference on their data. We create data defenses by developing a method to\nautomatically generate adversarial prompt injections that, when added to input\ntext, significantly reduce the ability of LLMs to accurately infer personally\nidentifying information about the subject of the input text or to use\ncopyrighted text in inference. We examine the ethics of enabling such direct\nresistance to LLM inference, and argue that making data defenses that resist\nand subvert LLMs enables the realization of important values such as data\nownership, data sovereignty, and democratic control over AI systems. We verify\nthat our data defenses are cheap and fast to generate, work on the latest\ncommercial and open-source LLMs, resistance to countermeasures, and are robust\nto several different attack settings. Finally, we consider the security\nimplications of LLM data defenses and outline several future research\ndirections in this area. Our code is available at\nhttps://github.com/wagnew3/LLMDataDefenses and a tool for using our defenses to\nprotect text against LLM inference is at\nhttps://wagnew3.github.io/LLM-Data-Defenses/.\n", "date": "2024-10-17", "categories": []}
{"id": "2208.06222", "title": "Scale-free and Task-agnostic Attack: Generating Photo-realistic\n  Adversarial Patterns with Patch Quilting Generator", "abstract": "  \\noindent Traditional L_p norm-restricted image attack algorithms suffer from\npoor transferability to black box scenarios and poor robustness to defense\nalgorithms. Recent CNN generator-based attack approaches can synthesize\nunrestricted and semantically meaningful entities to the image, which is shown\nto be transferable and robust. However, such methods attack images by either\nsynthesizing local adversarial entities, which are only suitable for attacking\nspecific contents or performing global attacks, which are only applicable to a\nspecific image scale. In this paper, we propose a novel Patch Quilting\nGenerative Adversarial Networks (PQ-GAN) to learn the first scale-free CNN\ngenerator that can be applied to attack images with arbitrary scales for\nvarious computer vision tasks. The principal investigation on transferability\nof the generated adversarial examples, robustness to defense frameworks, and\nvisual quality assessment show that the proposed PQG-based attack framework\noutperforms the other nine state-of-the-art adversarial attack approaches when\nattacking the neural networks trained on two standard evaluation datasets\n(i.e., ImageNet and CityScapes).\n", "date": "2022-08-12", "categories": []}
{"id": "2402.02263", "title": "MixedNUTS: Training-Free Accuracy-Robustness Balance via Nonlinearly\n  Mixed Classifiers", "abstract": "  Adversarial robustness often comes at the cost of degraded accuracy, impeding\nreal-life applications of robust classification models. Training-based\nsolutions for better trade-offs are limited by incompatibilities with\nalready-trained high-performance large models, necessitating the exploration of\ntraining-free ensemble approaches. Observing that robust models are more\nconfident in correct predictions than in incorrect ones on clean and\nadversarial data alike, we speculate amplifying this \"benign confidence\nproperty\" can reconcile accuracy and robustness in an ensemble setting. To\nachieve so, we propose \"MixedNUTS\", a training-free method where the output\nlogits of a robust classifier and a standard non-robust classifier are\nprocessed by nonlinear transformations with only three parameters, which are\noptimized through an efficient algorithm. MixedNUTS then converts the\ntransformed logits into probabilities and mixes them as the overall output. On\nCIFAR-10, CIFAR-100, and ImageNet datasets, experimental results with custom\nstrong adaptive attacks demonstrate MixedNUTS's vastly improved accuracy and\nnear-SOTA robustness -- it boosts CIFAR-100 clean accuracy by 7.86 points,\nsacrificing merely 0.87 points in robust accuracy.\n", "date": "2024-02-03", "categories": []}
{"id": "2405.14033", "title": "Adversarial Training of Two-Layer Polynomial and ReLU Activation\n  Networks via Convex Optimization", "abstract": "  Training neural networks which are robust to adversarial attacks remains an\nimportant problem in deep learning, especially as heavily overparameterized\nmodels are adopted in safety-critical settings. Drawing from recent work which\nreformulates the training problems for two-layer ReLU and polynomial activation\nnetworks as convex programs, we devise a convex semidefinite program (SDP) for\nadversarial training of two-layer polynomial activation networks and prove that\nthe convex SDP achieves the same globally optimal solution as its nonconvex\ncounterpart. The convex SDP is observed to improve robust test accuracy against\n$\\ell_\\infty$ attacks relative to the original convex training formulation on\nmultiple datasets. Additionally, we present scalable implementations of\nadversarial training for two-layer polynomial and ReLU networks which are\ncompatible with standard machine learning libraries and GPU acceleration.\nLeveraging these implementations, we retrain the final two fully connected\nlayers of a Pre-Activation ResNet-18 model on the CIFAR-10 dataset with both\npolynomial and ReLU activations. The two `robustified' models achieve\nsignificantly higher robust test accuracies against $\\ell_\\infty$ attacks than\na Pre-Activation ResNet-18 model trained with sharpness-aware minimization,\ndemonstrating the practical utility of convex adversarial training on\nlarge-scale problems.\n", "date": "2024-05-22", "categories": []}
{"id": "2410.11639", "title": "Efficient and Effective Universal Adversarial Attack against\n  Vision-Language Pre-training Models", "abstract": "  Vision-language pre-training (VLP) models, trained on large-scale image-text\npairs, have become widely used across a variety of downstream\nvision-and-language (V+L) tasks. This widespread adoption raises concerns about\ntheir vulnerability to adversarial attacks. Non-universal adversarial attacks,\nwhile effective, are often impractical for real-time online applications due to\ntheir high computational demands per data instance. Recently, universal\nadversarial perturbations (UAPs) have been introduced as a solution, but\nexisting generator-based UAP methods are significantly time-consuming. To\novercome the limitation, we propose a direct optimization-based UAP approach,\ntermed DO-UAP, which significantly reduces resource consumption while\nmaintaining high attack performance. Specifically, we explore the necessity of\nmultimodal loss design and introduce a useful data augmentation strategy.\nExtensive experiments conducted on three benchmark VLP datasets, six popular\nVLP models, and three classical downstream tasks demonstrate the efficiency and\neffectiveness of DO-UAP. Specifically, our approach drastically decreases the\ntime consumption by 23-fold while achieving a better attack performance.\n", "date": "2024-10-15", "categories": []}
{"id": "2410.12076", "title": "Taking off the Rose-Tinted Glasses: A Critical Look at Adversarial ML\n  Through the Lens of Evasion Attacks", "abstract": "  The vulnerability of machine learning models in adversarial scenarios has\ngarnered significant interest in the academic community over the past decade,\nresulting in a myriad of attacks and defenses. However, while the community\nappears to be overtly successful in devising new attacks across new contexts,\nthe development of defenses has stalled. After a decade of research, we appear\nno closer to securing AI applications beyond additional training. Despite a\nlack of effective mitigations, AI development and its incorporation into\nexisting systems charge full speed ahead with the rise of generative AI and\nlarge language models. Will our ineffectiveness in developing solutions to\nadversarial threats further extend to these new technologies?\n  In this paper, we argue that overly permissive attack and overly restrictive\ndefensive threat models have hampered defense development in the ML domain.\nThrough the lens of adversarial evasion attacks against neural networks, we\ncritically examine common attack assumptions, such as the ability to bypass any\ndefense not explicitly built into the model. We argue that these flawed\nassumptions, seen as reasonable by the community based on paper acceptance,\nhave encouraged the development of adversarial attacks that map poorly to\nreal-world scenarios. In turn, new defenses evaluated against these very\nattacks are inadvertently required to be almost perfect and incorporated as\npart of the model. But do they need to? In practice, machine learning models\nare deployed as a small component of a larger system. We analyze adversarial\nmachine learning from a system security perspective rather than an AI\nperspective and its implications for emerging AI paradigms.\n", "date": "2024-10-15", "categories": []}
{"id": "2410.12307", "title": "DAT: Improving Adversarial Robustness via Generative Amplitude Mix-up in\n  Frequency Domain", "abstract": "  To protect deep neural networks (DNNs) from adversarial attacks, adversarial\ntraining (AT) is developed by incorporating adversarial examples (AEs) into\nmodel training. Recent studies show that adversarial attacks disproportionately\nimpact the patterns within the phase of the sample's frequency spectrum --\ntypically containing crucial semantic information -- more than those in the\namplitude, resulting in the model's erroneous categorization of AEs. We find\nthat, by mixing the amplitude of training samples' frequency spectrum with\nthose of distractor images for AT, the model can be guided to focus on phase\npatterns unaffected by adversarial perturbations. As a result, the model's\nrobustness can be improved. Unfortunately, it is still challenging to select\nappropriate distractor images, which should mix the amplitude without affecting\nthe phase patterns. To this end, in this paper, we propose an optimized\nAdversarial Amplitude Generator (AAG) to achieve a better tradeoff between\nimproving the model's robustness and retaining phase patterns. Based on this\ngenerator, together with an efficient AE production procedure, we design a new\nDual Adversarial Training (DAT) strategy. Experiments on various datasets show\nthat our proposed DAT leads to significantly improved robustness against\ndiverse adversarial attacks.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.12425", "title": "Perseus: Leveraging Common Data Patterns with Curriculum Learning for\n  More Robust Graph Neural Networks", "abstract": "  Graph Neural Networks (GNNs) excel at handling graph data but remain\nvulnerable to adversarial attacks. Existing defense methods typically rely on\nassumptions like graph sparsity and homophily to either preprocess the graph or\nguide structure learning. However, preprocessing methods often struggle to\naccurately distinguish between normal edges and adversarial perturbations,\nleading to suboptimal results due to the loss of valuable edge information.\nRobust graph neural network models train directly on graph data affected by\nadversarial perturbations, without preprocessing. This can cause the model to\nget stuck in poor local optima, negatively affecting its performance. To\naddress these challenges, we propose Perseus, a novel adversarial defense\nmethod based on curriculum learning. Perseus assesses edge difficulty using\nglobal homophily and applies a curriculum learning strategy to adjust the\nlearning order, guiding the model to learn the full graph structure while\nadaptively focusing on common data patterns. This approach mitigates the impact\nof adversarial perturbations. Experiments show that models trained with Perseus\nachieve superior performance and are significantly more robust to adversarial\nattacks.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.12607", "title": "Low-Rank Adversarial PGD Attack", "abstract": "  Adversarial attacks on deep neural network models have seen rapid development\nand are extensively used to study the stability of these networks. Among\nvarious adversarial strategies, Projected Gradient Descent (PGD) is a widely\nadopted method in computer vision due to its effectiveness and quick\nimplementation, making it suitable for adversarial training. In this work, we\nobserve that in many cases, the perturbations computed using PGD predominantly\naffect only a portion of the singular value spectrum of the original image,\nsuggesting that these perturbations are approximately low-rank. Motivated by\nthis observation, we propose a variation of PGD that efficiently computes a\nlow-rank attack. We extensively validate our method on a range of standard\nmodels as well as robust models that have undergone adversarial training. Our\nanalysis indicates that the proposed low-rank PGD can be effectively used in\nadversarial training due to its straightforward and fast implementation coupled\nwith competitive performance. Notably, we find that low-rank PGD often performs\ncomparably to, and sometimes even outperforms, the traditional full-rank PGD\nattack, while using significantly less memory.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.12759", "title": "Unitary Multi-Margin BERT for Robust Natural Language Processing", "abstract": "  Recent developments in adversarial attacks on deep learning leave many\nmission-critical natural language processing (NLP) systems at risk of\nexploitation. To address the lack of computationally efficient adversarial\ndefense methods, this paper reports a novel, universal technique that\ndrastically improves the robustness of Bidirectional Encoder Representations\nfrom Transformers (BERT) by combining the unitary weights with the multi-margin\nloss. We discover that the marriage of these two simple ideas amplifies the\nprotection against malicious interference. Our model, the unitary multi-margin\nBERT (UniBERT), boosts post-attack classification accuracies significantly by\n5.3% to 73.8% while maintaining competitive pre-attack accuracies. Furthermore,\nthe pre-attack and post-attack accuracy tradeoff can be adjusted via a single\nscalar parameter to best fit the design requirements for the target\napplications.\n", "date": "2024-10-16", "categories": []}
{"id": "2402.15853", "title": "RAUCA: A Novel Physical Adversarial Attack on Vehicle Detectors via\n  Robust and Accurate Camouflage Generation", "abstract": "  Adversarial camouflage is a widely used physical attack against vehicle\ndetectors for its superiority in multi-view attack performance. One promising\napproach involves using differentiable neural renderers to facilitate\nadversarial camouflage optimization through gradient back-propagation. However,\nexisting methods often struggle to capture environmental characteristics during\nthe rendering process or produce adversarial textures that can precisely map to\nthe target vehicle, resulting in suboptimal attack performance. Moreover, these\napproaches neglect diverse weather conditions, reducing the efficacy of\ngenerated camouflage across varying weather scenarios. To tackle these\nchallenges, we propose a robust and accurate camouflage generation method,\nnamely RAUCA. The core of RAUCA is a novel neural rendering component, Neural\nRenderer Plus (NRP), which can accurately project vehicle textures and render\nimages with environmental characteristics such as lighting and weather. In\naddition, we integrate a multi-weather dataset for camouflage generation,\nleveraging the NRP to enhance the attack robustness. Experimental results on\nsix popular object detectors show that RAUCA consistently outperforms existing\nmethods in both simulation and real-world settings.\n", "date": "2024-02-24", "categories": []}
{"id": "2405.16112", "title": "Mitigating Backdoor Attack by Injecting Proactive Defensive Backdoor", "abstract": "  Data-poisoning backdoor attacks are serious security threats to machine\nlearning models, where an adversary can manipulate the training dataset to\ninject backdoors into models. In this paper, we focus on in-training backdoor\ndefense, aiming to train a clean model even when the dataset may be potentially\npoisoned. Unlike most existing methods that primarily detect and remove/unlearn\nsuspicious samples to mitigate malicious backdoor attacks, we propose a novel\ndefense approach called PDB (Proactive Defensive Backdoor). Specifically, PDB\nleverages the home-field advantage of defenders by proactively injecting a\ndefensive backdoor into the model during training. Taking advantage of\ncontrolling the training process, the defensive backdoor is designed to\nsuppress the malicious backdoor effectively while remaining secret to\nattackers. In addition, we introduce a reversible mapping to determine the\ndefensive target label. During inference, PDB embeds a defensive trigger in the\ninputs and reverses the model's prediction, suppressing malicious backdoor and\nensuring the model's utility on the original task. Experimental results across\nvarious datasets and models demonstrate that our approach achieves\nstate-of-the-art defense performance against a wide range of backdoor attacks.\nThe code is available at\nhttps://github.com/shawkui/Proactive_Defensive_Backdoor.\n", "date": "2024-05-25", "categories": []}
{"id": "2410.09318", "title": "Impeding LLM-assisted Cheating in Introductory Programming Assignments\n  via Adversarial Perturbation", "abstract": "  While Large language model (LLM)-based programming assistants such as CoPilot\nand ChatGPT can help improve the productivity of professional software\ndevelopers, they can also facilitate cheating in introductory computer\nprogramming courses. Assuming instructors have limited control over the\nindustrial-strength models, this paper investigates the baseline performance of\n5 widely used LLMs on a collection of introductory programming problems,\nexamines adversarial perturbations to degrade their performance, and describes\nthe results of a user study aimed at understanding the efficacy of such\nperturbations in hindering actual code generation for introductory programming\nassignments. The user study suggests that i) perturbations combinedly reduced\nthe average correctness score by 77%, ii) the drop in correctness caused by\nthese perturbations was affected based on their detectability.\n", "date": "2024-10-12", "categories": []}
{"id": "2402.17509", "title": "Extreme Miscalibration and the Illusion of Adversarial Robustness", "abstract": "  Deep learning-based Natural Language Processing (NLP) models are vulnerable\nto adversarial attacks, where small perturbations can cause a model to\nmisclassify. Adversarial Training (AT) is often used to increase model\nrobustness. However, we have discovered an intriguing phenomenon: deliberately\nor accidentally miscalibrating models masks gradients in a way that interferes\nwith adversarial attack search methods, giving rise to an apparent increase in\nrobustness. We show that this observed gain in robustness is an illusion of\nrobustness (IOR), and demonstrate how an adversary can perform various forms of\ntest-time temperature calibration to nullify the aforementioned interference\nand allow the adversarial attack to find adversarial examples. Hence, we urge\nthe NLP community to incorporate test-time temperature scaling into their\nrobustness evaluations to ensure that any observed gains are genuine. Finally,\nwe show how the temperature can be scaled during \\textit{training} to improve\ngenuine robustness.\n", "date": "2024-02-27", "categories": []}
{"id": "2403.13322", "title": "DD-RobustBench: An Adversarial Robustness Benchmark for Dataset\n  Distillation", "abstract": "  Dataset distillation is an advanced technique aimed at compressing datasets\ninto significantly smaller counterparts, while preserving formidable training\nperformance. Significant efforts have been devoted to promote evaluation\naccuracy under limited compression ratio while overlooked the robustness of\ndistilled dataset. In this work, we introduce a comprehensive benchmark that,\nto the best of our knowledge, is the most extensive to date for evaluating the\nadversarial robustness of distilled datasets in a unified way. Our benchmark\nsignificantly expands upon prior efforts by incorporating a wider range of\ndataset distillation methods, including the latest advancements such as TESLA\nand SRe2L, a diverse array of adversarial attack methods, and evaluations\nacross a broader and more extensive collection of datasets such as ImageNet-1K.\nMoreover, we assessed the robustness of these distilled datasets against\nrepresentative adversarial attack algorithms like PGD and AutoAttack, while\nexploring their resilience from a frequency perspective. We also discovered\nthat incorporating distilled data into the training batches of the original\ndataset can yield to improvement of robustness.\n", "date": "2024-03-20", "categories": []}
{"id": "2405.17894", "title": "White-box Multimodal Jailbreaks Against Large Vision-Language Models", "abstract": "  Recent advancements in Large Vision-Language Models (VLMs) have underscored\ntheir superiority in various multimodal tasks. However, the adversarial\nrobustness of VLMs has not been fully explored. Existing methods mainly assess\nrobustness through unimodal adversarial attacks that perturb images, while\nassuming inherent resilience against text-based attacks. Different from\nexisting attacks, in this work we propose a more comprehensive strategy that\njointly attacks both text and image modalities to exploit a broader spectrum of\nvulnerability within VLMs. Specifically, we propose a dual optimization\nobjective aimed at guiding the model to generate affirmative responses with\nhigh toxicity. Our attack method begins by optimizing an adversarial image\nprefix from random noise to generate diverse harmful responses in the absence\nof text input, thus imbuing the image with toxic semantics. Subsequently, an\nadversarial text suffix is integrated and co-optimized with the adversarial\nimage prefix to maximize the probability of eliciting affirmative responses to\nvarious harmful instructions. The discovered adversarial image prefix and text\nsuffix are collectively denoted as a Universal Master Key (UMK). When\nintegrated into various malicious queries, UMK can circumvent the alignment\ndefenses of VLMs and lead to the generation of objectionable content, known as\njailbreaks. The experimental results demonstrate that our universal attack\nstrategy can effectively jailbreak MiniGPT-4 with a 96% success rate,\nhighlighting the vulnerability of VLMs and the urgent need for new alignment\nstrategies.\n", "date": "2024-05-28", "categories": []}
{"id": "2410.08776", "title": "F2A: An Innovative Approach for Prompt Injection by Utilizing Feign\n  Security Detection Agents", "abstract": "  With the rapid development of Large Language Models (LLMs), numerous mature\napplications of LLMs have emerged in the field of content safety detection.\nHowever, we have found that LLMs exhibit blind trust in safety detection\nagents. The general LLMs can be compromised by hackers with this vulnerability.\nHence, this paper proposed an attack named Feign Agent Attack (F2A).Through\nsuch malicious forgery methods, adding fake safety detection results into the\nprompt, the defense mechanism of LLMs can be bypassed, thereby obtaining\nharmful content and hijacking the normal conversation. Continually, a series of\nexperiments were conducted. In these experiments, the hijacking capability of\nF2A on LLMs was analyzed and demonstrated, exploring the fundamental reasons\nwhy LLMs blindly trust safety detection results. The experiments involved\nvarious scenarios where fake safety detection results were injected into\nprompts, and the responses were closely monitored to understand the extent of\nthe vulnerability. Also, this paper provided a reasonable solution to this\nattack, emphasizing that it is important for LLMs to critically evaluate the\nresults of augmented agents to prevent the generating harmful content. By doing\nso, the reliability and security can be significantly improved, protecting the\nLLMs from F2A.\n", "date": "2024-10-11", "categories": []}
{"id": "2410.09845", "title": "Understanding Robustness of Parameter-Efficient Tuning for Image\n  Classification", "abstract": "  Parameter-efficient tuning (PET) techniques calibrate the model's predictions\non downstream tasks by freezing the pre-trained models and introducing a small\nnumber of learnable parameters. However, despite the numerous PET methods\nproposed, their robustness has not been thoroughly investigated. In this paper,\nwe systematically explore the robustness of four classical PET techniques\n(e.g., VPT, Adapter, AdaptFormer, and LoRA) under both white-box attacks and\ninformation perturbations. For white-box attack scenarios, we first analyze the\nperformance of PET techniques using FGSM and PGD attacks. Subsequently, we\nfurther explore the transferability of adversarial samples and the impact of\nlearnable parameter quantities on the robustness of PET methods. Under\ninformation perturbation attacks, we introduce four distinct perturbation\nstrategies, including Patch-wise Drop, Pixel-wise Drop, Patch Shuffle, and\nGaussian Noise, to comprehensively assess the robustness of these PET\ntechniques in the presence of information loss. Via these extensive studies, we\nenhance the understanding of the robustness of PET methods, providing valuable\ninsights for improving their performance in computer vision applications. The\ncode is available at https://github.com/JCruan519/PETRobustness.\n", "date": "2024-10-13", "categories": []}
{"id": "2410.09909", "title": "UnSeg: One Universal Unlearnable Example Generator is Enough against All\n  Image Segmentation", "abstract": "  Image segmentation is a crucial vision task that groups pixels within an\nimage into semantically meaningful segments, which is pivotal in obtaining a\nfine-grained understanding of real-world scenes. However, an increasing privacy\nconcern exists regarding training large-scale image segmentation models on\nunauthorized private data. In this work, we exploit the concept of unlearnable\nexamples to make images unusable to model training by generating and adding\nunlearnable noise into the original images. Particularly, we propose a novel\nUnlearnable Segmentation (UnSeg) framework to train a universal unlearnable\nnoise generator that is capable of transforming any downstream images into\ntheir unlearnable version. The unlearnable noise generator is finetuned from\nthe Segment Anything Model (SAM) via bilevel optimization on an interactive\nsegmentation dataset towards minimizing the training error of a surrogate model\nthat shares the same architecture with SAM but is trained from scratch. We\nempirically verify the effectiveness of UnSeg across 6 mainstream image\nsegmentation tasks, 10 widely used datasets, and 7 different network\narchitectures, and show that the unlearnable images can reduce the segmentation\nperformance by a large margin. Our work provides useful insights into how to\nleverage foundation models in a data-efficient and computationally affordable\nmanner to protect images against image segmentation models.\n", "date": "2024-10-13", "categories": []}
{"id": "2410.10091", "title": "Out-of-Bounding-Box Triggers: A Stealthy Approach to Cheat Object\n  Detectors", "abstract": "  In recent years, the study of adversarial robustness in object detection\nsystems, particularly those based on deep neural networks (DNNs), has become a\npivotal area of research. Traditional physical attacks targeting object\ndetectors, such as adversarial patches and texture manipulations, directly\nmanipulate the surface of the object. While these methods are effective, their\novert manipulation of objects may draw attention in real-world applications. To\naddress this, this paper introduces a more subtle approach: an inconspicuous\nadversarial trigger that operates outside the bounding boxes, rendering the\nobject undetectable to the model. We further enhance this approach by proposing\nthe Feature Guidance (FG) technique and the Universal Auto-PGD (UAPGD)\noptimization strategy for crafting high-quality triggers. The effectiveness of\nour method is validated through extensive empirical testing, demonstrating its\nhigh performance in both digital and physical environments. The code and video\nwill be available at: https://github.com/linToTao/Out-of-bbox-attack.\n", "date": "2024-10-14", "categories": []}
{"id": "2410.10554", "title": "ROSAR: An Adversarial Re-Training Framework for Robust Side-Scan Sonar\n  Object Detection", "abstract": "  This paper introduces ROSAR, a novel framework enhancing the robustness of\ndeep learning object detection models tailored for side-scan sonar (SSS)\nimages, generated by autonomous underwater vehicles using sonar sensors. By\nextending our prior work on knowledge distillation (KD), this framework\nintegrates KD with adversarial retraining to address the dual challenges of\nmodel efficiency and robustness against SSS noises. We introduce three novel,\npublicly available SSS datasets, capturing different sonar setups and noise\nconditions. We propose and formalize two SSS safety properties and utilize them\nto generate adversarial datasets for retraining. Through a comparative analysis\nof projected gradient descent (PGD) and patch-based adversarial attacks, ROSAR\ndemonstrates significant improvements in model robustness and detection\naccuracy under SSS-specific conditions, enhancing the model's robustness by up\nto 1.85%. ROSAR is available at\nhttps://github.com/remaro-network/ROSAR-framework.\n", "date": "2024-10-14", "categories": []}
{"id": "2410.10736", "title": "Towards Calibrated Losses for Adversarial Robust Reject Option\n  Classification", "abstract": "  Robustness towards adversarial attacks is a vital property for classifiers in\nseveral applications such as autonomous driving, medical diagnosis, etc. Also,\nin such scenarios, where the cost of misclassification is very high, knowing\nwhen to abstain from prediction becomes crucial. A natural question is which\nsurrogates can be used to ensure learning in scenarios where the input points\nare adversarially perturbed and the classifier can abstain from prediction?\nThis paper aims to characterize and design surrogates calibrated in\n\"Adversarial Robust Reject Option\" setting. First, we propose an adversarial\nrobust reject option loss $\\ell_{d}^{\\gamma}$ and analyze it for the hypothesis\nset of linear classifiers ($\\mathcal{H}_{\\textrm{lin}}$). Next, we provide a\ncomplete characterization result for any surrogate to be\n$(\\ell_{d}^{\\gamma},\\mathcal{H}_{\\textrm{lin}})$- calibrated. To demonstrate\nthe difficulty in designing surrogates to $\\ell_{d}^{\\gamma}$, we show negative\ncalibration results for convex surrogates and quasi-concave conditional risk\ncases (these gave positive calibration in adversarial setting without reject\noption). We also empirically argue that Shifted Double Ramp Loss (DRL) and\nShifted Double Sigmoid Loss (DSL) satisfy the calibration conditions. Finally,\nwe demonstrate the robustness of shifted DRL and shifted DSL against\nadversarial perturbations on a synthetically generated dataset.\n", "date": "2024-10-14", "categories": []}
{"id": "2410.10760", "title": "Denial-of-Service Poisoning Attacks against Large Language Models", "abstract": "  Recent studies have shown that LLMs are vulnerable to denial-of-service (DoS)\nattacks, where adversarial inputs like spelling errors or non-semantic prompts\ntrigger endless outputs without generating an [EOS] token. These attacks can\npotentially cause high latency and make LLM services inaccessible to other\nusers or tasks. However, when there are speech-to-text interfaces (e.g., voice\ncommands to a robot), executing such DoS attacks becomes challenging, as it is\ndifficult to introduce spelling errors or non-semantic prompts through speech.\nA simple DoS attack in these scenarios would be to instruct the model to \"Keep\nrepeating Hello\", but we observe that relying solely on natural instructions\nlimits output length, which is bounded by the maximum length of the LLM's\nsupervised finetuning (SFT) data. To overcome this limitation, we propose\npoisoning-based DoS (P-DoS) attacks for LLMs, demonstrating that injecting a\nsingle poisoned sample designed for DoS purposes can break the output length\nlimit. For example, a poisoned sample can successfully attack GPT-4o and GPT-4o\nmini (via OpenAI's finetuning API) using less than $1, causing repeated outputs\nup to the maximum inference length (16K tokens, compared to 0.5K before\npoisoning). Additionally, we perform comprehensive ablation studies on\nopen-source LLMs and extend our method to LLM agents, where attackers can\ncontrol both the finetuning dataset and algorithm. Our findings underscore the\nurgent need for defenses against P-DoS attacks to secure LLMs. Our code is\navailable at https://github.com/sail-sg/P-DoS.\n", "date": "2024-10-14", "categories": []}
{"id": "2404.15081", "title": "Perturbing Attention Gives You More Bang for the Buck: Subtle Imaging\n  Perturbations That Efficiently Fool Customized Diffusion Models", "abstract": "  Diffusion models (DMs) embark a new era of generative modeling and offer more\nopportunities for efficient generating high-quality and realistic data samples.\nHowever, their widespread use has also brought forth new challenges in model\nsecurity, which motivates the creation of more effective adversarial attackers\non DMs to understand its vulnerability. We propose CAAT, a simple but generic\nand efficient approach that does not require costly training to effectively\nfool latent diffusion models (LDMs). The approach is based on the observation\nthat cross-attention layers exhibits higher sensitivity to gradient change,\nallowing for leveraging subtle perturbations on published images to\nsignificantly corrupt the generated images. We show that a subtle perturbation\non an image can significantly impact the cross-attention layers, thus changing\nthe mapping between text and image during the fine-tuning of customized\ndiffusion models. Extensive experiments demonstrate that CAAT is compatible\nwith diverse diffusion models and outperforms baseline attack methods in a more\neffective (more noise) and efficient (twice as fast as Anti-DreamBooth and\nMist) manner.\n", "date": "2024-04-23", "categories": []}
{"id": "2406.02983", "title": "FREA: Feasibility-Guided Generation of Safety-Critical Scenarios with\n  Reasonable Adversariality", "abstract": "  Generating safety-critical scenarios, which are essential yet difficult to\ncollect at scale, offers an effective method to evaluate the robustness of\nautonomous vehicles (AVs). Existing methods focus on optimizing adversariality\nwhile preserving the naturalness of scenarios, aiming to achieve a balance\nthrough data-driven approaches. However, without an appropriate upper bound for\nadversariality, the scenarios might exhibit excessive adversariality,\npotentially leading to unavoidable collisions. In this paper, we introduce\nFREA, a novel safety-critical scenarios generation method that incorporates the\nLargest Feasible Region (LFR) of AV as guidance to ensure the reasonableness of\nthe adversarial scenarios. Concretely, FREA initially pre-calculates the LFR of\nAV from offline datasets. Subsequently, it learns a reasonable adversarial\npolicy that controls the scene's critical background vehicles (CBVs) to\ngenerate adversarial yet AV-feasible scenarios by maximizing a novel\nfeasibility-dependent adversarial objective function. Extensive experiments\nillustrate that FREA can effectively generate safety-critical scenarios,\nyielding considerable near-miss events while ensuring AV's feasibility.\nGeneralization analysis also confirms the robustness of FREA in AV testing\nacross various surrogate AV methods and traffic environments.\n", "date": "2024-06-05", "categories": []}
{"id": "2407.04482", "title": "Controlling Whisper: Universal Acoustic Adversarial Attacks to Control\n  Speech Foundation Models", "abstract": "  Speech enabled foundation models, either in the form of flexible speech\nrecognition based systems or audio-prompted large language models (LLMs), are\nbecoming increasingly popular. One of the interesting aspects of these models\nis their ability to perform tasks other than automatic speech recognition (ASR)\nusing an appropriate prompt. For example, the OpenAI Whisper model can perform\nboth speech transcription and speech translation. With the development of\naudio-prompted LLMs there is the potential for even greater control options. In\nthis work we demonstrate that with this greater flexibility the systems can be\nsusceptible to model-control adversarial attacks. Without any access to the\nmodel prompt it is possible to modify the behaviour of the system by\nappropriately changing the audio input. To illustrate this risk, we demonstrate\nthat it is possible to prepend a short universal adversarial acoustic segment\nto any input speech signal to override the prompt setting of an ASR foundation\nmodel. Specifically, we successfully use a universal adversarial acoustic\nsegment to control Whisper to always perform speech translation, despite being\nset to perform speech transcription. Overall, this work demonstrates a new form\nof adversarial attack on multi-tasking speech enabled foundation models that\nneeds to be considered prior to the deployment of this form of model.\n", "date": "2024-07-05", "categories": []}
{"id": "2410.08620", "title": "Natural Language Induced Adversarial Images", "abstract": "  Research of adversarial attacks is important for AI security because it shows\nthe vulnerability of deep learning models and helps to build more robust\nmodels. Adversarial attacks on images are most widely studied, which include\nnoise-based attacks, image editing-based attacks, and latent space-based\nattacks. However, the adversarial examples crafted by these methods often lack\nsufficient semantic information, making it challenging for humans to understand\nthe failure modes of deep learning models under natural conditions. To address\nthis limitation, we propose a natural language induced adversarial image attack\nmethod. The core idea is to leverage a text-to-image model to generate\nadversarial images given input prompts, which are maliciously constructed to\nlead to misclassification for a target model. To adopt commercial text-to-image\nmodels for synthesizing more natural adversarial images, we propose an adaptive\ngenetic algorithm (GA) for optimizing discrete adversarial prompts without\nrequiring gradients and an adaptive word space reduction method for improving\nquery efficiency. We further used CLIP to maintain the semantic consistency of\nthe generated images. In our experiments, we found that some high-frequency\nsemantic information such as \"foggy\", \"humid\", \"stretching\", etc. can easily\ncause classifier errors. This adversarial semantic information exists not only\nin generated images but also in photos captured in the real world. We also\nfound that some adversarial semantic information can be transferred to unknown\nclassification tasks. Furthermore, our attack method can transfer to different\ntext-to-image models (e.g., Midjourney, DALL-E 3, etc.) and image classifiers.\nOur code is available at:\nhttps://github.com/zxp555/Natural-Language-Induced-Adversarial-Images.\n", "date": "2024-10-11", "categories": []}
{"id": "2410.08872", "title": "Fragile Giants: Understanding the Susceptibility of Models to\n  Subpopulation Attacks", "abstract": "  As machine learning models become increasingly complex, concerns about their\nrobustness and trustworthiness have become more pressing. A critical\nvulnerability of these models is data poisoning attacks, where adversaries\ndeliberately alter training data to degrade model performance. One particularly\nstealthy form of these attacks is subpopulation poisoning, which targets\ndistinct subgroups within a dataset while leaving overall performance largely\nintact. The ability of these attacks to generalize within subpopulations poses\na significant risk in real-world settings, as they can be exploited to harm\nmarginalized or underrepresented groups within the dataset. In this work, we\ninvestigate how model complexity influences susceptibility to subpopulation\npoisoning attacks. We introduce a theoretical framework that explains how\noverparameterized models, due to their large capacity, can inadvertently\nmemorize and misclassify targeted subpopulations. To validate our theory, we\nconduct extensive experiments on large-scale image and text datasets using\npopular model architectures. Our results show a clear trend: models with more\nparameters are significantly more vulnerable to subpopulation poisoning.\nMoreover, we find that attacks on smaller, human-interpretable subgroups often\ngo undetected by these models. These results highlight the need to develop\ndefenses that specifically address subpopulation vulnerabilities.\n", "date": "2024-10-11", "categories": []}
{"id": "2410.09040", "title": "AttnGCG: Enhancing Jailbreaking Attacks on LLMs with Attention\n  Manipulation", "abstract": "  This paper studies the vulnerabilities of transformer-based Large Language\nModels (LLMs) to jailbreaking attacks, focusing specifically on the\noptimization-based Greedy Coordinate Gradient (GCG) strategy. We first observe\na positive correlation between the effectiveness of attacks and the internal\nbehaviors of the models. For instance, attacks tend to be less effective when\nmodels pay more attention to system prompts designed to ensure LLM safety\nalignment. Building on this discovery, we introduce an enhanced method that\nmanipulates models' attention scores to facilitate LLM jailbreaking, which we\nterm AttnGCG. Empirically, AttnGCG shows consistent improvements in attack\nefficacy across diverse LLMs, achieving an average increase of ~7% in the\nLlama-2 series and ~10% in the Gemma series. Our strategy also demonstrates\nrobust attack transferability against both unseen harmful goals and black-box\nLLMs like GPT-3.5 and GPT-4. Moreover, we note our attention-score\nvisualization is more interpretable, allowing us to gain better insights into\nhow our targeted attention manipulation facilitates more effective\njailbreaking. We release the code at\nhttps://github.com/UCSC-VLAA/AttnGCG-attack.\n", "date": "2024-10-11", "categories": []}
{"id": "2212.07278", "title": "Backdoor Mitigation in Deep Neural Networks via Strategic Retraining", "abstract": "  Deep Neural Networks (DNN) are becoming increasingly more important in\nassisted and automated driving. Using such entities which are obtained using\nmachine learning is inevitable: tasks such as recognizing traffic signs cannot\nbe developed reasonably using traditional software development methods. DNN\nhowever do have the problem that they are mostly black boxes and therefore hard\nto understand and debug. One particular problem is that they are prone to\nhidden backdoors. This means that the DNN misclassifies its input, because it\nconsiders properties that should not be decisive for the output. Backdoors may\neither be introduced by malicious attackers or by inappropriate training. In\nany case, detecting and removing them is important in the automotive area, as\nthey might lead to safety violations with potentially severe consequences. In\nthis paper, we introduce a novel method to remove backdoors. Our method works\nfor both intentional as well as unintentional backdoors. We also do not require\nprior knowledge about the shape or distribution of backdoors. Experimental\nevidence shows that our method performs well on several medium-sized examples.\n", "date": "2022-12-14", "categories": []}
{"id": "2408.10571", "title": "Prompt-Agnostic Adversarial Perturbation for Customized Diffusion Models", "abstract": "  Diffusion models have revolutionized customized text-to-image generation,\nallowing for efficient synthesis of photos from personal data with textual\ndescriptions. However, these advancements bring forth risks including privacy\nbreaches and unauthorized replication of artworks. Previous researches\nprimarily center around using prompt-specific methods to generate adversarial\nexamples to protect personal images, yet the effectiveness of existing methods\nis hindered by constrained adaptability to different prompts. In this paper, we\nintroduce a Prompt-Agnostic Adversarial Perturbation (PAP) method for\ncustomized diffusion models. PAP first models the prompt distribution using a\nLaplace Approximation, and then produces prompt-agnostic perturbations by\nmaximizing a disturbance expectation based on the modeled distribution. This\napproach effectively tackles the prompt-agnostic attacks, leading to improved\ndefense stability. Extensive experiments in face privacy and artistic style\nprotection, demonstrate the superior generalization of PAP in comparison to\nexisting techniques. Our project page is available at\nhttps://github.com/vancyland/Prompt-Agnostic-Adversarial-Perturbation-for-Customized-Diffusion-Models.github.io.\n", "date": "2024-08-20", "categories": []}
{"id": "2410.07675", "title": "Adversarial Robustness Overestimation and Instability in TRADES", "abstract": "  This paper examines the phenomenon of probabilistic robustness overestimation\nin TRADES, a prominent adversarial training method. Our study reveals that\nTRADES sometimes yields disproportionately high PGD validation accuracy\ncompared to the AutoAttack testing accuracy in the multiclass classification\ntask. This discrepancy highlights a significant overestimation of robustness\nfor these instances, potentially linked to gradient masking. We further analyze\nthe parameters contributing to unstable models that lead to overestimation. Our\nfindings indicate that smaller batch sizes, lower beta values (which control\nthe weight of the robust loss term in TRADES), larger learning rates, and\nhigher class complexity (e.g., CIFAR-100 versus CIFAR-10) are associated with\nan increased likelihood of robustness overestimation. By examining metrics such\nas the First-Order Stationary Condition (FOSC), inner-maximization, and\ngradient information, we identify the underlying cause of this phenomenon as\ngradient masking and provide insights into it. Furthermore, our experiments\nshow that certain unstable training instances may return to a state without\nrobust overestimation, inspiring our attempts at a solution. In addition to\nadjusting parameter settings to reduce instability or retraining when\noverestimation occurs, we recommend incorporating Gaussian noise in inputs when\nthe FOSC score exceed the threshold. This method aims to mitigate robustness\noverestimation of TRADES and other similar methods at its source, ensuring more\nreliable representation of adversarial robustness during evaluation.\n", "date": "2024-10-10", "categories": []}
{"id": "2410.07962", "title": "Towards Assurance of LLM Adversarial Robustness using Ontology-Driven\n  Argumentation", "abstract": "  Despite the impressive adaptability of large language models (LLMs),\nchallenges remain in ensuring their security, transparency, and\ninterpretability. Given their susceptibility to adversarial attacks, LLMs need\nto be defended with an evolving combination of adversarial training and\nguardrails. However, managing the implicit and heterogeneous knowledge for\ncontinuously assuring robustness is difficult. We introduce a novel approach\nfor assurance of the adversarial robustness of LLMs based on formal\nargumentation. Using ontologies for formalization, we structure\nstate-of-the-art attacks and defenses, facilitating the creation of a\nhuman-readable assurance case, and a machine-readable representation. We\ndemonstrate its application with examples in English language and code\ntranslation tasks, and provide implications for theory and practice, by\ntargeting engineers, data scientists, users, and auditors.\n", "date": "2024-10-10", "categories": []}
{"id": "2311.06423", "title": "Transferability Bound Theory: Exploring Relationship between Adversarial\n  Transferability and Flatness", "abstract": "  A prevailing belief in attack and defense community is that the higher\nflatness of adversarial examples enables their better cross-model\ntransferability, leading to a growing interest in employing sharpness-aware\nminimization and its variants. However, the theoretical relationship between\nthe transferability of adversarial examples and their flatness has not been\nwell established, making the belief questionable. To bridge this gap, we embark\non a theoretical investigation and, for the first time, derive a theoretical\nbound for the transferability of adversarial examples with few practical\nassumptions. Our analysis challenges this belief by demonstrating that the\nincreased flatness of adversarial examples does not necessarily guarantee\nimproved transferability. Moreover, building upon the theoretical analysis, we\npropose TPA, a Theoretically Provable Attack that optimizes a surrogate of the\nderived bound to craft adversarial examples. Extensive experiments across\nwidely used benchmark datasets and various real-world applications show that\nTPA can craft more transferable adversarial examples compared to\nstate-of-the-art baselines. We hope that these results can recalibrate\npreconceived impressions within the community and facilitate the development of\nstronger adversarial attack and defense mechanisms. The source codes are\navailable in <https://github.com/fmy266/TPA>.\n", "date": "2023-11-10", "categories": []}
{"id": "2402.01920", "title": "Preference Poisoning Attacks on Reward Model Learning", "abstract": "  Learning reward models from pairwise comparisons is a fundamental component\nin a number of domains, including autonomous control, conversational agents,\nand recommendation systems, as part of a broad goal of aligning automated\ndecisions with user preferences. These approaches entail collecting preference\ninformation from people, with feedback often provided anonymously. Since\npreferences are subjective, there is no gold standard to compare against; yet,\nreliance of high-impact systems on preference learning creates a strong\nmotivation for malicious actors to skew data collected in this fashion to their\nends. We investigate the nature and extent of this vulnerability by considering\nan attacker who can flip a small subset of preference comparisons to either\npromote or demote a target outcome. We propose two classes of algorithmic\napproaches for these attacks: a gradient-based framework, and several variants\nof rank-by-distance methods. Next, we evaluate the efficacy of best attacks in\nboth these classes in successfully achieving malicious goals on datasets from\nthree domains: autonomous control, recommendation system, and textual\nprompt-response preference learning. We find that the best attacks are often\nhighly successful, achieving in the most extreme case 100\\% success rate with\nonly 0.3\\% of the data poisoned. However, \\emph{which} attack is best can vary\nsignificantly across domains. In addition, we observe that the simpler and more\nscalable rank-by-distance approaches are often competitive with, and on\noccasion significantly outperform, gradient-based methods. Finally, we show\nthat state-of-the-art defenses against other classes of poisoning attacks\nexhibit limited efficacy in our setting.\n", "date": "2024-02-02", "categories": []}
{"id": "2405.01349", "title": "Position: Towards Resilience Against Adversarial Examples", "abstract": "  Current research on defending against adversarial examples focuses primarily\non achieving robustness against a single attack type such as $\\ell_2$ or\n$\\ell_{\\infty}$-bounded attacks. However, the space of possible perturbations\nis much larger than considered by many existing defenses and is difficult to\nmathematically model, so the attacker can easily bypass the defense by using a\ntype of attack that is not covered by the defense. In this position paper, we\nargue that in addition to robustness, we should also aim to develop defense\nalgorithms that are adversarially resilient -- defense algorithms should\nspecify a means to quickly adapt the defended model to be robust against new\nattacks. We provide a definition of adversarial resilience and outline\nconsiderations of designing an adversarially resilient defense. We then\nintroduce a subproblem of adversarial resilience which we call continual\nadaptive robustness, in which the defender gains knowledge of the formulation\nof possible perturbation spaces over time and can then update their model based\non this information. Additionally, we demonstrate the connection between\ncontinual adaptive robustness and previously studied problems of multiattack\nrobustness and unforeseen attack robustness and outline open directions within\nthese fields which can contribute to improving continual adaptive robustness\nand adversarial resilience.\n", "date": "2024-05-02", "categories": []}
{"id": "2405.15984", "title": "Evaluating and Safeguarding the Adversarial Robustness of\n  Retrieval-Based In-Context Learning", "abstract": "  With the emergence of large language models, such as LLaMA and OpenAI GPT-3,\nIn-Context Learning (ICL) gained significant attention due to its effectiveness\nand efficiency. However, ICL is very sensitive to the choice, order, and\nverbaliser used to encode the demonstrations in the prompt. Retrieval-Augmented\nICL methods try to address this problem by leveraging retrievers to extract\nsemantically related examples as demonstrations. While this approach yields\nmore accurate results, its robustness against various types of adversarial\nattacks, including perturbations on test samples, demonstrations, and retrieved\ndata, remains under-explored. Our study reveals that retrieval-augmented models\ncan enhance robustness against test sample attacks, outperforming vanilla ICL\nwith a 4.87% reduction in Attack Success Rate (ASR); however, they exhibit\noverconfidence in the demonstrations, leading to a 2% increase in ASR for\ndemonstration attacks. Adversarial training can help improve the robustness of\nICL methods to adversarial attacks; however, such a training scheme can be too\ncostly in the context of LLMs. As an alternative, we introduce an effective\ntraining-free adversarial defence method, DARD, which enriches the example pool\nwith those attacked samples. We show that DARD yields improvements in\nperformance and robustness, achieving a 15% reduction in ASR over the\nbaselines. Code and data are released to encourage further research:\nhttps://github.com/simonucl/adv-retreival-icl\n", "date": "2024-05-24", "categories": []}
{"id": "2406.05532", "title": "Exploring Adversarial Robustness of Deep State Space Models", "abstract": "  Deep State Space Models (SSMs) have proven effective in numerous task\nscenarios but face significant security challenges due to Adversarial\nPerturbations (APs) in real-world deployments. Adversarial Training (AT) is a\nmainstream approach to enhancing Adversarial Robustness (AR) and has been\nvalidated on various traditional DNN architectures. However, its effectiveness\nin improving the AR of SSMs remains unclear. While many enhancements in SSM\ncomponents, such as integrating Attention mechanisms and expanding to\ndata-dependent SSM parameterizations, have brought significant gains in\nStandard Training (ST) settings, their potential benefits in AT remain\nunexplored. To investigate this, we evaluate existing structural variants of\nSSMs with AT to assess their AR performance. We observe that pure SSM\nstructures struggle to benefit from AT, whereas incorporating Attention yields\na markedly better trade-off between robustness and generalization for SSMs in\nAT compared to other components. Nonetheless, the integration of Attention also\nleads to Robust Overfitting (RO) issues. To understand these phenomena, we\nempirically and theoretically analyze the output error of SSMs under AP. We\nfind that fixed-parameterized SSMs have output error bounds strictly related to\ntheir parameters, limiting their AT benefits, while input-dependent SSMs may\nface the problem of error explosion. Furthermore, we show that the Attention\ncomponent effectively scales the output error of SSMs during training, enabling\nthem to benefit more from AT, but at the cost of introducing RO due to its high\nmodel complexity. Inspired by this, we propose a simple and effective Adaptive\nScaling (AdS) mechanism that brings AT performance close to\nAttention-integrated SSMs without introducing the issue of RO. Our code is\navailable at\nhttps://github.com/Biqing-Qi/Exploring-Adversarial-Robustness-of-Deep-State-Space-Models.git.\n", "date": "2024-06-08", "categories": []}
{"id": "2410.05334", "title": "TA3: Testing Against Adversarial Attacks on Machine Learning Models", "abstract": "  Adversarial attacks are major threats to the deployment of machine learning\n(ML) models in many applications. Testing ML models against such attacks is\nbecoming an essential step for evaluating and improving ML models. In this\npaper, we report the design and development of an interactive system for aiding\nthe workflow of Testing Against Adversarial Attacks (TA3). In particular, with\nTA3, human-in-the-loop (HITL) enables human-steered attack simulation and\nvisualization-assisted attack impact evaluation. While the current version of\nTA3 focuses on testing decision tree models against adversarial attacks based\non the One Pixel Attack Method, it demonstrates the importance of HITL in ML\ntesting and the potential application of HITL to the ML testing workflows for\nother types of ML models and other types of adversarial attacks.\n", "date": "2024-10-06", "categories": []}
{"id": "2410.05951", "title": "Hyper Adversarial Tuning for Boosting Adversarial Robustness of\n  Pretrained Large Vision Models", "abstract": "  Large vision models have been found vulnerable to adversarial examples,\nemphasizing the need for enhancing their adversarial robustness. While\nadversarial training is an effective defense for deep convolutional models, it\noften faces scalability issues with large vision models due to high\ncomputational costs. Recent approaches propose robust fine-tuning methods, such\nas adversarial tuning of low-rank adaptation (LoRA) in large vision models, but\nthey still struggle to match the accuracy of full parameter adversarial\nfine-tuning. The integration of various defense mechanisms offers a promising\napproach to enhancing the robustness of large vision models, yet this paradigm\nremains underexplored. To address this, we propose hyper adversarial tuning\n(HyperAT), which leverages shared defensive knowledge among different methods\nto improve model robustness efficiently and effectively simultaneously.\nSpecifically, adversarial tuning of each defense method is formulated as a\nlearning task, and a hypernetwork generates LoRA specific to this defense.\nThen, a random sampling and tuning strategy is proposed to extract and\nfacilitate the defensive knowledge transfer between different defenses.\nFinally, diverse LoRAs are merged to enhance the adversarial robustness.\nExperiments on various datasets and model architectures demonstrate that\nHyperAT significantly enhances the adversarial robustness of pretrained large\nvision models without excessive computational overhead, establishing a new\nstate-of-the-art benchmark.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.06339", "title": "Filtered Randomized Smoothing: A New Defense for Robust Modulation\n  Classification", "abstract": "  Deep Neural Network (DNN) based classifiers have recently been used for the\nmodulation classification of RF signals. These classifiers have shown\nimpressive performance gains relative to conventional methods, however, they\nare vulnerable to imperceptible (low-power) adversarial attacks. Some of the\nprominent defense approaches include adversarial training (AT) and randomized\nsmoothing (RS). While AT increases robustness in general, it fails to provide\nresilience against previously unseen adaptive attacks. Other approaches, such\nas Randomized Smoothing (RS), which injects noise into the input, address this\nshortcoming by providing provable certified guarantees against arbitrary\nattacks, however, they tend to sacrifice accuracy.\n  In this paper, we study the problem of designing robust DNN-based modulation\nclassifiers that can provide provable defense against arbitrary attacks without\nsignificantly sacrificing accuracy. To this end, we first analyze the spectral\ncontent of commonly studied attacks on modulation classifiers for the benchmark\nRadioML dataset. We observe that spectral signatures of un-perturbed RF signals\nare highly localized, whereas attack signals tend to be spread out in\nfrequency. To exploit this spectral heterogeneity, we propose Filtered\nRandomized Smoothing (FRS), a novel defense which combines spectral filtering\ntogether with randomized smoothing. FRS can be viewed as a strengthening of RS\nby leveraging the specificity (spectral Heterogeneity) inherent to the\nmodulation classification problem. In addition to providing an approach to\ncompute the certified accuracy of FRS, we also provide a comprehensive set of\nsimulations on the RadioML dataset to show the effectiveness of FRS and show\nthat it significantly outperforms existing defenses including AT and RS in\nterms of accuracy on both attacked and benign signals.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.06572", "title": "Can DeepFake Speech be Reliably Detected?", "abstract": "  Recent advances in text-to-speech (TTS) systems, particularly those with\nvoice cloning capabilities, have made voice impersonation readily accessible,\nraising ethical and legal concerns due to potential misuse for malicious\nactivities like misinformation campaigns and fraud. While synthetic speech\ndetectors (SSDs) exist to combat this, they are vulnerable to ``test domain\nshift\", exhibiting decreased performance when audio is altered through\ntranscoding, playback, or background noise. This vulnerability is further\nexacerbated by deliberate manipulation of synthetic speech aimed at deceiving\ndetectors. This work presents the first systematic study of such active\nmalicious attacks against state-of-the-art open-source SSDs. White-box attacks,\nblack-box attacks, and their transferability are studied from both attack\neffectiveness and stealthiness, using both hardcoded metrics and human ratings.\nThe results highlight the urgent need for more robust detection methods in the\nface of evolving adversarial threats.\n", "date": "2024-10-09", "categories": []}
{"id": "2410.06699", "title": "Break the Visual Perception: Adversarial Attacks Targeting Encoded\n  Visual Tokens of Large Vision-Language Models", "abstract": "  Large vision-language models (LVLMs) integrate visual information into large\nlanguage models, showcasing remarkable multi-modal conversational capabilities.\nHowever, the visual modules introduces new challenges in terms of robustness\nfor LVLMs, as attackers can craft adversarial images that are visually clean\nbut may mislead the model to generate incorrect answers. In general, LVLMs rely\non vision encoders to transform images into visual tokens, which are crucial\nfor the language models to perceive image contents effectively. Therefore, we\nare curious about one question: Can LVLMs still generate correct responses when\nthe encoded visual tokens are attacked and disrupting the visual information?\nTo this end, we propose a non-targeted attack method referred to as VT-Attack\n(Visual Tokens Attack), which constructs adversarial examples from multiple\nperspectives, with the goal of comprehensively disrupting feature\nrepresentations and inherent relationships as well as the semantic properties\nof visual tokens output by image encoders. Using only access to the image\nencoder in the proposed attack, the generated adversarial examples exhibit\ntransferability across diverse LVLMs utilizing the same image encoder and\ngenerality across different tasks. Extensive experiments validate the superior\nattack performance of the VT-Attack over baseline methods, demonstrating its\neffectiveness in attacking LVLMs with image encoders, which in turn can provide\nguidance on the robustness of LVLMs, particularly in terms of the stability of\nthe visual feature space.\n", "date": "2024-10-09", "categories": []}
{"id": "2410.06921", "title": "Adversarial Vulnerability as a Consequence of On-Manifold Inseparibility", "abstract": "  Recent works have shown theoretically and empirically that redundant data\ndimensions are a source of adversarial vulnerability. However, the inverse\ndoesn't seem to hold in practice; employing dimension-reduction techniques\ndoesn't exhibit robustness as expected. In this work, we consider\nclassification tasks and characterize the data distribution as a\nlow-dimensional manifold, with high/low variance features defining the on/off\nmanifold direction. We argue that clean training experiences poor convergence\nin the off-manifold direction caused by the ill-conditioning in widely used\nfirst-order optimizers like gradient descent. The poor convergence then acts as\na source of adversarial vulnerability when the dataset is inseparable in the\non-manifold direction. We provide theoretical results for logistic regression\nand a 2-layer linear network on the considered data distribution. Furthermore,\nwe advocate using second-order methods that are immune to ill-conditioning and\nlead to better robustness. We perform experiments and exhibit tremendous\nrobustness improvements in clean training through long training and the\nemployment of second-order methods, corroborating our framework. Additionally,\nwe find the inclusion of batch-norm layers hinders such robustness gains. We\nattribute this to differing implicit biases between traditional and\nbatch-normalized neural networks.\n", "date": "2024-10-09", "categories": []}
{"id": "2311.13656", "title": "Panda or not Panda? Understanding Adversarial Attacks with Interactive\n  Visualization", "abstract": "  Adversarial machine learning (AML) studies attacks that can fool machine\nlearning algorithms into generating incorrect outcomes as well as the defenses\nagainst worst-case attacks to strengthen model robustness. Specifically for\nimage classification, it is challenging to understand adversarial attacks due\nto their use of subtle perturbations that are not human-interpretable, as well\nas the variability of attack impacts influenced by diverse methodologies,\ninstance differences, and model architectures. Through a design study with AML\nlearners and teachers, we introduce AdvEx, a multi-level interactive\nvisualization system that comprehensively presents the properties and impacts\nof evasion attacks on different image classifiers for novice AML learners. We\nquantitatively and qualitatively assessed AdvEx in a two-part evaluation\nincluding user studies and expert interviews. Our results show that AdvEx is\nnot only highly effective as a visualization tool for understanding AML\nmechanisms, but also provides an engaging and enjoyable learning experience,\nthus demonstrating its overall benefits for AML learners.\n", "date": "2023-11-22", "categories": []}
{"id": "2408.14866", "title": "Advancing Adversarial Suffix Transfer Learning on Aligned Large Language\n  Models", "abstract": "  Language Language Models (LLMs) face safety concerns due to potential misuse\nby malicious users. Recent red-teaming efforts have identified adversarial\nsuffixes capable of jailbreaking LLMs using the gradient-based search algorithm\nGreedy Coordinate Gradient (GCG). However, GCG struggles with computational\ninefficiency, limiting further investigations regarding suffix transferability\nand scalability across models and data. In this work, we bridge the connection\nbetween search efficiency and suffix transferability. We propose a two-stage\ntransfer learning framework, DeGCG, which decouples the search process into\nbehavior-agnostic pre-searching and behavior-relevant post-searching.\nSpecifically, we employ direct first target token optimization in pre-searching\nto facilitate the search process. We apply our approach to cross-model,\ncross-data, and self-transfer scenarios. Furthermore, we introduce an\ninterleaved variant of our approach, i-DeGCG, which iteratively leverages\nself-transferability to accelerate the search process. Experiments on HarmBench\ndemonstrate the efficiency of our approach across various models and domains.\nNotably, our i-DeGCG outperforms the baseline on Llama2-chat-7b with ASRs of\n$43.9$ ($+22.2$) and $39.0$ ($+19.5$) on valid and test sets, respectively.\nFurther analysis on cross-model transfer indicates the pivotal role of first\ntarget token optimization in leveraging suffix transferability for efficient\nsearching.\n", "date": "2024-08-27", "categories": []}
{"id": "2409.05208", "title": "Influence-based Attributions can be Manipulated", "abstract": "  Influence Functions are a standard tool for attributing predictions to\ntraining data in a principled manner and are widely used in applications such\nas data valuation and fairness. In this work, we present realistic incentives\nto manipulate influence-based attributions and investigate whether these\nattributions can be \\textit{systematically} tampered by an adversary. We show\nthat this is indeed possible for logistic regression models trained on ResNet\nfeature embeddings and standard tabular fairness datasets and provide efficient\nattacks with backward-friendly implementations. Our work raises questions on\nthe reliability of influence-based attributions in adversarial circumstances.\nCode is available at :\n\\url{https://github.com/infinite-pursuits/influence-based-attributions-can-be-manipulated}\n", "date": "2024-09-08", "categories": []}
{"id": "2410.04577", "title": "Robustness Reprogramming for Representation Learning", "abstract": "  This work tackles an intriguing and fundamental open challenge in\nrepresentation learning: Given a well-trained deep learning model, can it be\nreprogrammed to enhance its robustness against adversarial or noisy input\nperturbations without altering its parameters? To explore this, we revisit the\ncore feature transformation mechanism in representation learning and propose a\nnovel non-linear robust pattern matching technique as a robust alternative.\nFurthermore, we introduce three model reprogramming paradigms to offer flexible\ncontrol of robustness under different efficiency requirements. Comprehensive\nexperiments and ablation studies across diverse learning models ranging from\nbasic linear model and MLPs to shallow and modern deep ConvNets demonstrate the\neffectiveness of our approaches. This work not only opens a promising and\northogonal direction for improving adversarial defenses in deep learning beyond\nexisting methods but also provides new insights into designing more resilient\nAI systems with robust statistics.\n", "date": "2024-10-06", "categories": []}
{"id": "2410.05136", "title": "LOTOS: Layer-wise Orthogonalization for Training Robust Ensembles", "abstract": "  Transferability of adversarial examples is a well-known property that\nendangers all classification models, even those that are only accessible\nthrough black-box queries. Prior work has shown that an ensemble of models is\nmore resilient to transferability: the probability that an adversarial example\nis effective against most models of the ensemble is low. Thus, most ongoing\nresearch focuses on improving ensemble diversity. Another line of prior work\nhas shown that Lipschitz continuity of the models can make models more robust\nsince it limits how a model's output changes with small input perturbations. In\nthis paper, we study the effect of Lipschitz continuity on transferability\nrates. We show that although a lower Lipschitz constant increases the\nrobustness of a single model, it is not as beneficial in training robust\nensembles as it increases the transferability rate of adversarial examples\nacross models in the ensemble. Therefore, we introduce LOTOS, a new training\nparadigm for ensembles, which counteracts this adverse effect. It does so by\npromoting orthogonality among the top-$k$ sub-spaces of the transformations of\nthe corresponding affine layers of any pair of models in the ensemble. We\ntheoretically show that $k$ does not need to be large for convolutional layers,\nwhich makes the computational overhead negligible. Through various experiments,\nwe show LOTOS increases the robust accuracy of ensembles of ResNet-18 models by\n$6$ percentage points (p.p) against black-box attacks on CIFAR-10. It is also\ncapable of combining with the robustness of prior state-of-the-art methods for\ntraining robust ensembles to enhance their robust accuracy by $10.7$ p.p.\n", "date": "2024-10-07", "categories": []}
{"id": "2410.03376", "title": "Mitigating Adversarial Perturbations for Deep Reinforcement Learning via\n  Vector Quantization", "abstract": "  Recent studies reveal that well-performing reinforcement learning (RL) agents\nin training often lack resilience against adversarial perturbations during\ndeployment. This highlights the importance of building a robust agent before\ndeploying it in the real world. Most prior works focus on developing robust\ntraining-based procedures to tackle this problem, including enhancing the\nrobustness of the deep neural network component itself or adversarially\ntraining the agent on strong attacks. In this work, we instead study an input\ntransformation-based defense for RL. Specifically, we propose using a variant\nof vector quantization (VQ) as a transformation for input observations, which\nis then used to reduce the space of adversarial attacks during testing,\nresulting in the transformed observations being less affected by attacks. Our\nmethod is computationally efficient and seamlessly integrates with adversarial\ntraining, further enhancing the robustness of RL agents against adversarial\nattacks. Through extensive experiments in multiple environments, we demonstrate\nthat using VQ as the input transformation effectively defends against\nadversarial attacks on the agent's observations.\n", "date": "2024-10-04", "categories": []}
{"id": "2410.03566", "title": "A Survey on Offensive AI Within Cybersecurity", "abstract": "  Artificial Intelligence (AI) has witnessed major growth and integration\nacross various domains. As AI systems become increasingly prevalent, they also\nbecome targets for threat actors to manipulate their functionality for\nmalicious purposes. This survey paper on offensive AI will comprehensively\ncover various aspects related to attacks against and using AI systems. It will\ndelve into the impact of offensive AI practices on different domains, including\nconsumer, enterprise, and public digital infrastructure. The paper will explore\nadversarial machine learning, attacks against AI models, infrastructure, and\ninterfaces, along with offensive techniques like information gathering, social\nengineering, and weaponized AI. Additionally, it will discuss the consequences\nand implications of offensive AI, presenting case studies, insights, and\navenues for further research.\n", "date": "2024-09-26", "categories": []}
{"id": "2410.03658", "title": "RAFT: Realistic Attacks to Fool Text Detectors", "abstract": "  Large language models (LLMs) have exhibited remarkable fluency across various\ntasks. However, their unethical applications, such as disseminating\ndisinformation, have become a growing concern. Although recent works have\nproposed a number of LLM detection methods, their robustness and reliability\nremain unclear. In this paper, we present RAFT: a grammar error-free black-box\nattack against existing LLM detectors. In contrast to previous attacks for\nlanguage models, our method exploits the transferability of LLM embeddings at\nthe word-level while preserving the original text quality. We leverage an\nauxiliary embedding to greedily select candidate words to perturb against the\ntarget detector. Experiments reveal that our attack effectively compromises all\ndetectors in the study across various domains by up to 99%, and are\ntransferable across source models. Manual human evaluation studies show our\nattacks are realistic and indistinguishable from original human-written text.\nWe also show that examples generated by RAFT can be used to train adversarially\nrobust detectors. Our work shows that current LLM detectors are not\nadversarially robust, underscoring the urgent need for more resilient detection\nmechanisms.\n", "date": "2024-10-04", "categories": []}
{"id": "2310.14637", "title": "Semantic-Aware Adversarial Training for Reliable Deep Hashing Retrieval", "abstract": "  Deep hashing has been intensively studied and successfully applied in\nlarge-scale image retrieval systems due to its efficiency and effectiveness.\nRecent studies have recognized that the existence of adversarial examples poses\na security threat to deep hashing models, that is, adversarial vulnerability.\nNotably, it is challenging to efficiently distill reliable semantic\nrepresentatives for deep hashing to guide adversarial learning, and thereby it\nhinders the enhancement of adversarial robustness of deep hashing-based\nretrieval models. Moreover, current researches on adversarial training for deep\nhashing are hard to be formalized into a unified minimax structure. In this\npaper, we explore Semantic-Aware Adversarial Training (SAAT) for improving the\nadversarial robustness of deep hashing models. Specifically, we conceive a\ndiscriminative mainstay features learning (DMFL) scheme to construct semantic\nrepresentatives for guiding adversarial learning in deep hashing. Particularly,\nour DMFL with the strict theoretical guarantee is adaptively optimized in a\ndiscriminative learning manner, where both discriminative and semantic\nproperties are jointly considered. Moreover, adversarial examples are\nfabricated by maximizing the Hamming distance between the hash codes of\nadversarial samples and mainstay features, the efficacy of which is validated\nin the adversarial attack trials. Further, we, for the first time, formulate\nthe formalized adversarial training of deep hashing into a unified minimax\noptimization under the guidance of the generated mainstay codes. Extensive\nexperiments on benchmark datasets show superb attack performance against the\nstate-of-the-art algorithms, meanwhile, the proposed adversarial training can\neffectively eliminate adversarial perturbations for trustworthy deep\nhashing-based retrieval. Our code is available at\nhttps://github.com/xandery-geek/SAAT.\n", "date": "2023-10-23", "categories": []}
{"id": "2407.10179", "title": "CLIP-Guided Generative Networks for Transferable Targeted Adversarial\n  Attacks", "abstract": "  Transferable targeted adversarial attacks aim to mislead models into\noutputting adversary-specified predictions in black-box scenarios. Recent\nstudies have introduced \\textit{single-target} generative attacks that train a\ngenerator for each target class to generate highly transferable perturbations,\nresulting in substantial computational overhead when handling multiple classes.\n\\textit{Multi-target} attacks address this by training only one\nclass-conditional generator for multiple classes. However, the generator simply\nuses class labels as conditions, failing to leverage the rich semantic\ninformation of the target class. To this end, we design a \\textbf{C}LIP-guided\n\\textbf{G}enerative \\textbf{N}etwork with \\textbf{C}ross-attention modules\n(CGNC) to enhance multi-target attacks by incorporating textual knowledge of\nCLIP into the generator. Extensive experiments demonstrate that CGNC yields\nsignificant improvements over previous multi-target generative attacks, e.g., a\n21.46\\% improvement in success rate from ResNet-152 to DenseNet-121. Moreover,\nwe propose a masked fine-tuning mechanism to further strengthen our method in\nattacking a single class, which surpasses existing single-target methods.\n", "date": "2024-07-14", "categories": []}
{"id": "2410.02043", "title": "Impact of White-Box Adversarial Attacks on Convolutional Neural Networks", "abstract": "  Autonomous vehicle navigation and healthcare diagnostics are among the many\nfields where the reliability and security of machine learning models for image\ndata are critical. We conduct a comprehensive investigation into the\nsusceptibility of Convolutional Neural Networks (CNNs), which are widely used\nfor image data, to white-box adversarial attacks. We investigate the effects of\nvarious sophisticated attacks -- Fast Gradient Sign Method, Basic Iterative\nMethod, Jacobian-based Saliency Map Attack, Carlini & Wagner, Projected\nGradient Descent, and DeepFool -- on CNN performance metrics, (e.g., loss,\naccuracy), the differential efficacy of adversarial techniques in increasing\nerror rates, the relationship between perceived image quality metrics (e.g.,\nERGAS, PSNR, SSIM, and SAM) and classification performance, and the comparative\neffectiveness of iterative versus single-step attacks. Using the MNIST,\nCIFAR-10, CIFAR-100, and Fashio_MNIST datasets, we explore the effect of\ndifferent attacks on the CNNs performance metrics by varying the\nhyperparameters of CNNs. Our study provides insights into the robustness of\nCNNs against adversarial threats, pinpoints vulnerabilities, and underscores\nthe urgent need for developing robust defense mechanisms to protect CNNs and\nensuring their trustworthy deployment in real-world scenarios.\n", "date": "2024-10-02", "categories": []}
{"id": "2410.01393", "title": "Signal Adversarial Examples Generation for Signal Detection Network via\n  White-Box Attack", "abstract": "  With the development and application of deep learning in signal detection\ntasks, the vulnerability of neural networks to adversarial attacks has also\nbecome a security threat to signal detection networks. This paper defines a\nsignal adversarial examples generation model for signal detection network from\nthe perspective of adding perturbations to the signal. The model uses the\ninequality relationship of L2-norm between time domain and time-frequency\ndomain to constrain the energy of signal perturbations. Building upon this\nmodel, we propose a method for generating signal adversarial examples utilizing\ngradient-based attacks and Short-Time Fourier Transform. The experimental\nresults show that under the constraint of signal perturbation energy ratio less\nthan 3%, our adversarial attack resulted in a 28.1% reduction in the mean\nAverage Precision (mAP), a 24.7% reduction in recall, and a 30.4% reduction in\nprecision of the signal detection network. Compared to random noise\nperturbation of equivalent intensity, our adversarial attack demonstrates a\nsignificant attack effect.\n", "date": "2024-10-02", "categories": []}
{"id": "2407.17447", "title": "FLRT: Fluent Student-Teacher Redteaming", "abstract": "  Many publicly available language models have been safety tuned to reduce the\nlikelihood of toxic or liability-inducing text. To redteam or jailbreak these\nmodels for compliance with toxic requests, users and security analysts have\ndeveloped adversarial prompting techniques. One attack method is to apply\ndiscrete optimization techniques to the prompt. However, the resulting attack\nstrings are often gibberish text, easily filtered by defenders due to high\nmeasured perplexity, and may fail for unseen tasks and/or well-tuned models. In\nthis work, we improve existing algorithms (primarily GCG and BEAST) to develop\npowerful and fluent attacks on safety-tuned models like Llama-2 and Phi-3. Our\ntechnique centers around a new distillation-based approach that encourages the\nvictim model to emulate a toxified finetune, either in terms of output\nprobabilities or internal activations. To encourage human-fluent attacks, we\nadd a multi-model perplexity penalty and a repetition penalty to the objective.\nWe also enhance optimizer strength by allowing token insertions, token swaps,\nand token deletions and by using longer attack sequences. The resulting process\nis able to reliably jailbreak the most difficult target models with prompts\nthat appear similar to human-written prompts. On Advbench we achieve attack\nsuccess rates $>93$% for Llama-2-7B, Llama-3-8B, and Vicuna-7B, while\nmaintaining model-measured perplexity $<33$; we achieve $95$% attack success\nfor Phi-3, though with higher perplexity. We also find a universally-optimized\nsingle fluent prompt that induces $>88$% compliance on previously unseen tasks\nacross Llama-2-7B, Phi-3-mini and Vicuna-7B and transfers to other black-box\nmodels.\n", "date": "2024-07-24", "categories": []}
{"id": "2408.09839", "title": "Segment-Anything Models Achieve Zero-shot Robustness in Autonomous\n  Driving", "abstract": "  Semantic segmentation is a significant perception task in autonomous driving.\nIt suffers from the risks of adversarial examples. In the past few years, deep\nlearning has gradually transitioned from convolutional neural network (CNN)\nmodels with a relatively small number of parameters to foundation models with a\nhuge number of parameters. The segment-anything model (SAM) is a generalized\nimage segmentation framework that is capable of handling various types of\nimages and is able to recognize and segment arbitrary objects in an image\nwithout the need to train on a specific object. It is a unified model that can\nhandle diverse downstream tasks, including semantic segmentation, object\ndetection, and tracking. In the task of semantic segmentation for autonomous\ndriving, it is significant to study the zero-shot adversarial robustness of\nSAM. Therefore, we deliver a systematic empirical study on the robustness of\nSAM without additional training. Based on the experimental results, the\nzero-shot adversarial robustness of the SAM under the black-box corruptions and\nwhite-box adversarial attacks is acceptable, even without the need for\nadditional training. The finding of this study is insightful in that the\ngigantic model parameters and huge amounts of training data lead to the\nphenomenon of emergence, which builds a guarantee of adversarial robustness.\nSAM is a vision foundation model that can be regarded as an early prototype of\nan artificial general intelligence (AGI) pipeline. In such a pipeline, a\nunified model can handle diverse tasks. Therefore, this research not only\ninspects the impact of vision foundation models on safe autonomous driving but\nalso provides a perspective on developing trustworthy AGI. The code is\navailable at: https://github.com/momo1986/robust_sam_iv.\n", "date": "2024-08-19", "categories": []}
{"id": "2410.00878", "title": "Empirical Perturbation Analysis of Linear System Solvers from a Data\n  Poisoning Perspective", "abstract": "  The perturbation analysis of linear solvers applied to systems arising\nbroadly in machine learning settings -- for instance, when using linear\nregression models -- establishes an important perspective when reframing these\nanalyses through the lens of a data poisoning attack. By analyzing solvers'\nresponses to such attacks, this work aims to contribute to the development of\nmore robust linear solvers and provide insights into poisoning attacks on\nlinear solvers. In particular, we investigate how the errors in the input data\nwill affect the fitting error and accuracy of the solution from a linear\nsystem-solving algorithm under perturbations common in adversarial attacks. We\npropose data perturbation through two distinct knowledge levels, developing a\npoisoning optimization and studying two methods of perturbation: Label-guided\nPerturbation (LP) and Unconditioning Perturbation (UP). Existing works mainly\nfocus on deriving the worst-case perturbation bound from a theoretical\nperspective, and the analysis is often limited to specific kinds of linear\nsystem solvers. Under the circumstance that the data is intentionally perturbed\n-- as is the case with data poisoning -- we seek to understand how different\nkinds of solvers react to these perturbations, identifying those algorithms\nmost impacted by different types of adversarial attacks.\n", "date": "2024-10-01", "categories": []}
{"id": "2305.18779", "title": "It begins with a boundary: A geometric view on probabilistically robust\n  learning", "abstract": "  Although deep neural networks have achieved super-human performance on many\nclassification tasks, they often exhibit a worrying lack of robustness towards\nadversarially generated examples. Thus, considerable effort has been invested\ninto reformulating standard Risk Minimization (RM) into an adversarially robust\nframework. Recently, attention has shifted towards approaches which interpolate\nbetween the robustness offered by adversarial training and the higher clean\naccuracy and faster training times of RM. In this paper, we take a fresh and\ngeometric view on one such method -- Probabilistically Robust Learning (PRL).\nWe propose a mathematical framework for understanding PRL, which allows us to\nidentify geometric pathologies in its original formulation and to introduce a\nfamily of probabilistic nonlocal perimeter functionals to rectify them. We\nprove existence of solutions to the original and modified problems using novel\nrelaxation methods and also study properties, as well as local limits, of the\nintroduced perimeters. We also clarify, through a suitable $\\Gamma$-convergence\nanalysis, the way in which the original and modified PRL models interpolate\nbetween risk minimization and adversarial training.\n", "date": "2023-05-30", "categories": []}
{"id": "2306.08313", "title": "A Proxy Attack-Free Strategy for Practically Improving the Poisoning\n  Efficiency in Backdoor Attacks", "abstract": "  Poisoning efficiency is crucial in poisoning-based backdoor attacks, as\nattackers aim to minimize the number of poisoning samples while maximizing\nattack efficacy. Recent studies have sought to enhance poisoning efficiency by\nselecting effective samples. However, these studies typically rely on a proxy\nbackdoor injection task to identify an efficient set of poisoning samples. This\nproxy attack-based approach can lead to performance degradation if the proxy\nattack settings differ from those of the actual victims, due to the shortcut\nnature of backdoor learning. Furthermore, proxy attack-based methods are\nextremely time-consuming, as they require numerous complete backdoor injection\nprocesses for sample selection. To address these concerns, we present a Proxy\nattack-Free Strategy (PFS) designed to identify efficient poisoning samples\nbased on the similarity between clean samples and their corresponding poisoning\nsamples, as well as the diversity of the poisoning set. The proposed PFS is\nmotivated by the observation that selecting samples with high similarity\nbetween clean and corresponding poisoning samples results in significantly\nhigher attack success rates compared to using samples with low similarity.\nAdditionally, we provide theoretical foundations to explain the proposed PFS.\nWe comprehensively evaluate the proposed strategy across various datasets,\ntriggers, poisoning rates, architectures, and training hyperparameters. Our\nexperimental results demonstrate that PFS enhances backdoor attack efficiency\nwhile also offering a remarkable speed advantage over previous proxy\nattack-based selection methodologies.\n", "date": "2023-06-14", "categories": []}
{"id": "2309.15669", "title": "On the Computational Entanglement of Distant Features in Adversarial\n  Machine Learning", "abstract": "  In this research, we introduce the concept of \"computational entanglement,\" a\nphenomenon observed in overparameterized feedforward linear networks that\nenables the network to achieve zero loss by fitting random noise, even on\npreviously unseen test samples. Analyzing this behavior through spacetime\ndiagrams reveals its connection to length contraction, where both training and\ntest samples converge toward a shared normalized point within a flat Riemannian\nmanifold. Moreover, we present a novel application of computational\nentanglement in transforming a worst-case adversarial examples-inputs that are\nhighly non-robust and uninterpretable to human observers-into outputs that are\nboth recognizable and robust. This provides new insights into the behavior of\nnon-robust features in adversarial example generation, underscoring the\ncritical role of computational entanglement in enhancing model robustness and\nadvancing our understanding of neural networks in adversarial contexts.\n", "date": "2023-09-27", "categories": []}
{"id": "2312.16907", "title": "DOEPatch: Dynamically Optimized Ensemble Model for Adversarial Patches\n  Generation", "abstract": "  Object detection is a fundamental task in various applications ranging from\nautonomous driving to intelligent security systems. However, recognition of a\nperson can be hindered when their clothing is decorated with carefully designed\ngraffiti patterns, leading to the failure of object detection. To achieve\ngreater attack potential against unknown black-box models, adversarial patches\ncapable of affecting the outputs of multiple-object detection models are\nrequired. While ensemble models have proven effective, current research in the\nfield of object detection typically focuses on the simple fusion of the outputs\nof all models, with limited attention being given to developing general\nadversarial patches that can function effectively in the physical world. In\nthis paper, we introduce the concept of energy and treat the adversarial\npatches generation process as an optimization of the adversarial patches to\nminimize the total energy of the ``person'' category. Additionally, by adopting\nadversarial training, we construct a dynamically optimized ensemble model.\nDuring training, the weight parameters of the attacked target models are\nadjusted to find the balance point at which the generated adversarial patches\ncan effectively attack all target models. We carried out six sets of\ncomparative experiments and tested our algorithm on five mainstream object\ndetection models. The adversarial patches generated by our algorithm can reduce\nthe recognition accuracy of YOLOv2 and YOLOv3 to 13.19\\% and 29.20\\%,\nrespectively. In addition, we conducted experiments to test the effectiveness\nof T-shirts covered with our adversarial patches in the physical world and\ncould achieve that people are not recognized by the object detection model.\nFinally, leveraging the Grad-CAM tool, we explored the attack mechanism of\nadversarial patches from an energetic perspective.\n", "date": "2023-12-28", "categories": []}
{"id": "2401.01537", "title": "The Art of Deception: Robust Backdoor Attack using Dynamic Stacking of\n  Triggers", "abstract": "  The area of Machine Learning as a Service (MLaaS) is experiencing increased\nimplementation due to recent advancements in the AI (Artificial Intelligence)\nindustry. However, this spike has prompted concerns regarding AI defense\nmechanisms, specifically regarding potential covert attacks from third-party\nproviders that cannot be entirely trusted. Recent research has uncovered that\nauditory backdoors may use certain modifications as their initiating mechanism.\nDynamicTrigger is introduced as a methodology for carrying out dynamic backdoor\nattacks that use cleverly designed tweaks to ensure that corrupted samples are\nindistinguishable from clean. By utilizing fluctuating signal sampling rates\nand masking speaker identities through dynamic sound triggers (such as the\nclapping of hands), it is possible to deceive speech recognition systems (ASR).\nOur empirical testing demonstrates that DynamicTrigger is both potent and\nstealthy, achieving impressive success rates during covert attacks while\nmaintaining exceptional accuracy with non-poisoned datasets.\n", "date": "2024-01-03", "categories": []}
{"id": "2407.07662", "title": "Mitigating Backdoor Attacks using Activation-Guided Model Editing", "abstract": "  Backdoor attacks compromise the integrity and reliability of machine learning\nmodels by embedding a hidden trigger during the training process, which can\nlater be activated to cause unintended misbehavior. We propose a novel backdoor\nmitigation approach via machine unlearning to counter such backdoor attacks.\nThe proposed method utilizes model activation of domain-equivalent unseen data\nto guide the editing of the model's weights. Unlike the previous\nunlearning-based mitigation methods, ours is computationally inexpensive and\nachieves state-of-the-art performance while only requiring a handful of unseen\nsamples for unlearning. In addition, we also point out that unlearning the\nbackdoor may cause the whole targeted class to be unlearned, thus introducing\nan additional repair step to preserve the model's utility after editing the\nmodel. Experiment results show that the proposed method is effective in\nunlearning the backdoor on different datasets and trigger patterns.\n", "date": "2024-07-10", "categories": []}
{"id": "2409.19096", "title": "Enhancing Robustness of Graph Neural Networks through p-Laplacian", "abstract": "  With the increase of data in day-to-day life, businesses and different\nstakeholders need to analyze the data for better predictions. Traditionally,\nrelational data has been a source of various insights, but with the increase in\ncomputational power and the need to understand deeper relationships between\nentities, the need to design new techniques has arisen. For this graph data\nanalysis has become an extraordinary tool for understanding the data, which\nreveals more realistic and flexible modelling of complex relationships.\nRecently, Graph Neural Networks (GNNs) have shown great promise in various\napplications, such as social network analysis, recommendation systems, drug\ndiscovery, and more. However, many adversarial attacks can happen over the\ndata, whether during training (poisoning attack) or during testing (evasion\nattack), which can adversely manipulate the desired outcome from the GNN model.\nTherefore, it is crucial to make the GNNs robust to such attacks. The existing\nrobustness methods are computationally demanding and perform poorly when the\nintensity of attack increases. This paper presents a computationally efficient\nframework, namely, pLapGNN, based on weighted p-Laplacian for making GNNs\nrobust. Empirical evaluation on real datasets establishes the efficacy and\nefficiency of the proposed method.\n", "date": "2024-09-27", "categories": []}
{"id": "2409.19210", "title": "Learning to Obstruct Few-Shot Image Classification over Restricted\n  Classes", "abstract": "  Advancements in open-source pre-trained backbones make it relatively easy to\nfine-tune a model for new tasks. However, this lowered entry barrier poses\npotential risks, e.g., bad actors developing models for harmful applications. A\nquestion arises: Is possible to develop a pre-trained model that is difficult\nto fine-tune for certain downstream tasks? To begin studying this, we focus on\nfew-shot classification (FSC). Specifically, we investigate methods to make FSC\nmore challenging for a set of restricted classes while maintaining the\nperformance of other classes. We propose to meta-learn over the pre-trained\nbackbone in a manner that renders it a ''poor initialization''. Our proposed\nLearning to Obstruct (LTO) algorithm successfully obstructs four FSC methods\nacross three datasets, including ImageNet and CIFAR100 for image\nclassification, as well as CelebA for attribute classification.\n", "date": "2024-09-28", "categories": []}
{"id": "2409.19526", "title": "Efficient Backdoor Defense in Multimodal Contrastive Learning: A\n  Token-Level Unlearning Method for Mitigating Threats", "abstract": "  Multimodal contrastive learning uses various data modalities to create\nhigh-quality features, but its reliance on extensive data sources on the\nInternet makes it vulnerable to backdoor attacks. These attacks insert\nmalicious behaviors during training, which are activated by specific triggers\nduring inference, posing significant security risks. Despite existing\ncountermeasures through fine-tuning that reduce the malicious impacts of such\nattacks, these defenses frequently necessitate extensive training time and\ndegrade clean accuracy. In this study, we propose an efficient defense\nmechanism against backdoor threats using a concept known as machine unlearning.\nThis entails strategically creating a small set of poisoned samples to aid the\nmodel's rapid unlearning of backdoor vulnerabilities, known as Unlearn Backdoor\nThreats (UBT). We specifically use overfit training to improve backdoor\nshortcuts and accurately detect suspicious samples in the potential poisoning\ndata set. Then, we select fewer unlearned samples from suspicious samples for\nrapid forgetting in order to eliminate the backdoor effect and thus improve\nbackdoor defense efficiency. In the backdoor unlearning process, we present a\nnovel token-based portion unlearning training regime. This technique focuses on\nthe model's compromised elements, dissociating backdoor correlations while\nmaintaining the model's overall integrity. Extensive experimental results show\nthat our method effectively defends against various backdoor attack methods in\nthe CLIP model. Compared to SoTA backdoor defense methods, UBT achieves the\nlowest attack success rate while maintaining a high clean accuracy of the model\n(attack success rate decreases by 19% compared to SOTA, while clean accuracy\nincreases by 2.57%).\n", "date": "2024-09-29", "categories": []}
{"id": "2409.19594", "title": "MASKDROID: Robust Android Malware Detection with Masked Graph\n  Representations", "abstract": "  Android malware attacks have posed a severe threat to mobile users,\nnecessitating a significant demand for the automated detection system. Among\nthe various tools employed in malware detection, graph representations (e.g.,\nfunction call graphs) have played a pivotal role in characterizing the\nbehaviors of Android apps. However, though achieving impressive performance in\nmalware detection, current state-of-the-art graph-based malware detectors are\nvulnerable to adversarial examples. These adversarial examples are meticulously\ncrafted by introducing specific perturbations to normal malicious inputs. To\ndefend against adversarial attacks, existing defensive mechanisms are typically\nsupplementary additions to detectors and exhibit significant limitations, often\nrelying on prior knowledge of adversarial examples and failing to defend\nagainst unseen types of attacks effectively. In this paper, we propose\nMASKDROID, a powerful detector with a strong discriminative ability to identify\nmalware and remarkable robustness against adversarial attacks. Specifically, we\nintroduce a masking mechanism into the Graph Neural Network (GNN) based\nframework, forcing MASKDROID to recover the whole input graph using a small\nportion (e.g., 20%) of randomly selected nodes.This strategy enables the model\nto understand the malicious semantics and learn more stable representations,\nenhancing its robustness against adversarial attacks. While capturing stable\nmalicious semantics in the form of dependencies inside the graph structures, we\nfurther employ a contrastive module to encourage MASKDROID to learn more\ncompact representations for both the benign and malicious classes to boost its\ndiscriminative power in detecting malware from benign apps and adversarial\nexamples.\n", "date": "2024-09-29", "categories": []}
{"id": "2409.19619", "title": "Discerning the Chaos: Detecting Adversarial Perturbations while\n  Disentangling Intentional from Unintentional Noises", "abstract": "  Deep learning models, such as those used for face recognition and attribute\nprediction, are susceptible to manipulations like adversarial noise and\nunintentional noise, including Gaussian and impulse noise. This paper\nintroduces CIAI, a Class-Independent Adversarial Intent detection network built\non a modified vision transformer with detection layers. CIAI employs a novel\nloss function that combines Maximum Mean Discrepancy and Center Loss to detect\nboth intentional (adversarial attacks) and unintentional noise, regardless of\nthe image class. It is trained in a multi-step fashion. We also introduce the\naspect of intent during detection that can act as an added layer of security.\nWe further showcase the performance of our proposed detector on CelebA,\nCelebA-HQ, LFW, AgeDB, and CIFAR-10 datasets. Our detector is able to detect\nboth intentional (like FGSM, PGD, and DeepFool) and unintentional (like\nGaussian and Salt & Pepper noises) perturbations.\n", "date": "2024-09-29", "categories": []}
{"id": "2409.20139", "title": "Characterizing Model Robustness via Natural Input Gradients", "abstract": "  Adversarially robust models are locally smooth around each data sample so\nthat small perturbations cannot drastically change model outputs. In modern\nsystems, such smoothness is usually obtained via Adversarial Training, which\nexplicitly enforces models to perform well on perturbed examples. In this work,\nwe show the surprising effectiveness of instead regularizing the gradient with\nrespect to model inputs on natural examples only. Penalizing input Gradient\nNorm is commonly believed to be a much inferior approach. Our analyses identify\nthat the performance of Gradient Norm regularization critically depends on the\nsmoothness of activation functions, and are in fact extremely effective on\nmodern vision transformers that adopt smooth activations over piecewise linear\nones (eg, ReLU), contrary to prior belief. On ImageNet-1k, Gradient Norm\ntraining achieves > 90% the performance of state-of-the-art PGD-3 Adversarial\nTraining} (52% vs.~56%), while using only 60% computation cost of the\nstate-of-the-art without complex adversarial optimization. Our analyses also\nhighlight the relationship between model robustness and properties of natural\ninput gradients, such as asymmetric sample and channel statistics.\nSurprisingly, we find model robustness can be significantly improved by simply\nregularizing its gradients to concentrate on image edges without explicit\nconditioning on the gradient norm.\n", "date": "2024-09-30", "categories": []}
{"id": "2409.20426", "title": "Navigating Threats: A Survey of Physical Adversarial Attacks on LiDAR\n  Perception Systems in Autonomous Vehicles", "abstract": "  Autonomous vehicles (AVs) rely heavily on LiDAR (Light Detection and Ranging)\nsystems for accurate perception and navigation, providing high-resolution 3D\nenvironmental data that is crucial for object detection and classification.\nHowever, LiDAR systems are vulnerable to adversarial attacks, which pose\nsignificant challenges to the safety and robustness of AVs. This survey\npresents a thorough review of the current research landscape on physical\nadversarial attacks targeting LiDAR-based perception systems, covering both\nsingle-modality and multi-modality contexts. We categorize and analyze various\nattack types, including spoofing and physical adversarial object attacks,\ndetailing their methodologies, impacts, and potential real-world implications.\nThrough detailed case studies and analyses, we identify critical challenges and\nhighlight gaps in existing attacks for LiDAR-based systems. Additionally, we\npropose future research directions to enhance the security and resilience of\nthese systems, ultimately contributing to the safer deployment of autonomous\nvehicles.\n", "date": "2024-09-30", "categories": []}
{"id": "2409.18244", "title": "Development of an Edge Resilient ML Ensemble to Tolerate ICS Adversarial\n  Attacks", "abstract": "  Deploying machine learning (ML) in dynamic data-driven applications systems\n(DDDAS) can improve the security of industrial control systems (ICS). However,\nML-based DDDAS are vulnerable to adversarial attacks because adversaries can\nalter the input data slightly so that the ML models predict a different result.\nIn this paper, our goal is to build a resilient edge machine learning (reML)\narchitecture that is designed to withstand adversarial attacks by performing\nData Air Gap Transformation (DAGT) to anonymize data feature spaces using deep\nneural networks and randomize the ML models used for predictions. The reML is\nbased on the Resilient DDDAS paradigm, Moving Target Defense (MTD) theory, and\nTinyML and is applied to combat adversarial attacks on ICS. Furthermore, the\nproposed approach is power-efficient and privacy-preserving and, therefore, can\nbe deployed on power-constrained devices to enhance ICS security. This approach\nenables resilient ML inference at the edge by shifting the computation from the\ncomputing-intensive platforms to the resource-constrained edge devices. The\nincorporation of TinyML with TensorFlow Lite ensures efficient resource\nutilization and, consequently, makes reML suitable for deployment in various\nindustrial control environments. Furthermore, the dynamic nature of reML,\nfacilitated by the resilient DDDAS development environment, allows for\ncontinuous adaptation and improvement in response to emerging threats. Lastly,\nwe evaluate our approach on an ICS dataset and demonstrate that reML provides a\nviable and effective solution for resilient ML inference at the edge devices.\n", "date": "2024-09-26", "categories": []}
{"id": "2408.03944", "title": "Improving Fast Adversarial Training Paradigm: An Example Taxonomy\n  Perspective", "abstract": "  While adversarial training is an effective defense method against adversarial\nattacks, it notably increases the training cost. To this end, fast adversarial\ntraining (FAT) is presented for efficient training and has become a hot\nresearch topic. However, FAT suffers from catastrophic overfitting, which leads\nto a performance drop compared with multi-step adversarial training. However,\nthe cause of catastrophic overfitting remains unclear and lacks exploration. In\nthis paper, we present an example taxonomy in FAT, which identifies that\ncatastrophic overfitting is caused by the imbalance between the inner and outer\noptimization in FAT. Furthermore, we investigated the impact of varying degrees\nof training loss, revealing a correlation between training loss and\ncatastrophic overfitting. Based on these observations, we redesign the loss\nfunction in FAT with the proposed dynamic label relaxation to concentrate the\nloss range and reduce the impact of misclassified examples. Meanwhile, we\nintroduce batch momentum initialization to enhance the diversity to prevent\ncatastrophic overfitting in an efficient manner. Furthermore, we also propose\nCatastrophic Overfitting aware Loss Adaptation (COLA), which employs a separate\ntraining strategy for examples based on their loss degree. Our proposed method,\nnamed example taxonomy aware FAT (ETA), establishes an improved paradigm for\nFAT. Experiment results demonstrate our ETA achieves state-of-the-art\nperformance. Comprehensive experiments on four standard datasets demonstrate\nthe competitiveness of our proposed method.\n", "date": "2024-07-22", "categories": []}
{"id": "2409.17328", "title": "The poison of dimensionality", "abstract": "  This paper advances the understanding of how the size of a machine learning\nmodel affects its vulnerability to poisoning, despite state-of-the-art\ndefenses. Given isotropic random honest feature vectors and the geometric\nmedian (or clipped mean) as the robust gradient aggregator rule, we essentially\nprove that, perhaps surprisingly, linear and logistic regressions with $D \\geq\n169 H^2/P^2$ parameters are subject to arbitrary model manipulation by\npoisoners, where $H$ and $P$ are the numbers of honestly labeled and poisoned\ndata points used for training. Our experiments go on exposing a fundamental\ntradeoff between augmenting model expressivity and increasing the poisoners'\nattack surface, on both synthetic data, and on MNIST & FashionMNIST data for\nlinear classifiers with random features. We also discuss potential implications\nfor source-based learning and neural nets.\n", "date": "2024-09-25", "categories": []}
{"id": "2409.17376", "title": "Optical Lens Attack on Deep Learning Based Monocular Depth Estimation", "abstract": "  Monocular Depth Estimation (MDE) plays a crucial role in vision-based\nAutonomous Driving (AD) systems. It utilizes a single-camera image to determine\nthe depth of objects, facilitating driving decisions such as braking a few\nmeters in front of a detected obstacle or changing lanes to avoid collision. In\nthis paper, we investigate the security risks associated with monocular\nvision-based depth estimation algorithms utilized by AD systems. By exploiting\nthe vulnerabilities of MDE and the principles of optical lenses, we introduce\nLensAttack, a physical attack that involves strategically placing optical\nlenses on the camera of an autonomous vehicle to manipulate the perceived\nobject depths. LensAttack encompasses two attack formats: concave lens attack\nand convex lens attack, each utilizing different optical lenses to induce false\ndepth perception. We begin by constructing a mathematical model of our attack,\nincorporating various attack parameters. Subsequently, we simulate the attack\nand evaluate its real-world performance in driving scenarios to demonstrate its\neffect on state-of-the-art MDE models. The results highlight the significant\nimpact of LensAttack on the accuracy of depth estimation in AD systems.\n", "date": "2024-09-25", "categories": []}
{"id": "2409.17403", "title": "Transient Adversarial 3D Projection Attacks on Object Detection in\n  Autonomous Driving", "abstract": "  Object detection is a crucial task in autonomous driving. While existing\nresearch has proposed various attacks on object detection, such as those using\nadversarial patches or stickers, the exploration of projection attacks on 3D\nsurfaces remains largely unexplored. Compared to adversarial patches or\nstickers, which have fixed adversarial patterns, projection attacks allow for\ntransient modifications to these patterns, enabling a more flexible attack. In\nthis paper, we introduce an adversarial 3D projection attack specifically\ntargeting object detection in autonomous driving scenarios. We frame the attack\nformulation as an optimization problem, utilizing a combination of color\nmapping and geometric transformation models. Our results demonstrate the\neffectiveness of the proposed attack in deceiving YOLOv3 and Mask R-CNN in\nphysical settings. Evaluations conducted in an indoor environment show an\nattack success rate of up to 100% under low ambient light conditions,\nhighlighting the potential damage of our attack in real-world driving\nscenarios.\n", "date": "2024-09-25", "categories": []}
{"id": "2409.17568", "title": "Showing Many Labels in Multi-label Classification Models: An Empirical\n  Study of Adversarial Examples", "abstract": "  With the rapid development of Deep Neural Networks (DNNs), they have been\napplied in numerous fields. However, research indicates that DNNs are\nsusceptible to adversarial examples, and this is equally true in the\nmulti-label domain. To further investigate multi-label adversarial examples, we\nintroduce a novel type of attacks, termed \"Showing Many Labels\". The objective\nof this attack is to maximize the number of labels included in the classifier's\nprediction results. In our experiments, we select nine attack algorithms and\nevaluate their performance under \"Showing Many Labels\". Eight of the attack\nalgorithms were adapted from the multi-class environment to the multi-label\nenvironment, while the remaining one was specifically designed for the\nmulti-label environment. We choose ML-LIW and ML-GCN as target models and train\nthem on four popular multi-label datasets: VOC2007, VOC2012, NUS-WIDE, and\nCOCO. We record the success rate of each algorithm when it shows the expected\nnumber of labels in eight different scenarios. Experimental results indicate\nthat under the \"Showing Many Labels\", iterative attacks perform significantly\nbetter than one-step attacks. Moreover, it is possible to show all labels in\nthe dataset.\n", "date": "2024-09-26", "categories": []}
{"id": "2409.17589", "title": "Improving Fast Adversarial Training via Self-Knowledge Guidance", "abstract": "  Adversarial training has achieved remarkable advancements in defending\nagainst adversarial attacks. Among them, fast adversarial training (FAT) is\ngaining attention for its ability to achieve competitive robustness with fewer\ncomputing resources. Existing FAT methods typically employ a uniform strategy\nthat optimizes all training data equally without considering the influence of\ndifferent examples, which leads to an imbalanced optimization. However, this\nimbalance remains unexplored in the field of FAT. In this paper, we conduct a\ncomprehensive study of the imbalance issue in FAT and observe an obvious class\ndisparity regarding their performances. This disparity could be embodied from a\nperspective of alignment between clean and robust accuracy. Based on the\nanalysis, we mainly attribute the observed misalignment and disparity to the\nimbalanced optimization in FAT, which motivates us to optimize different\ntraining data adaptively to enhance robustness. Specifically, we take disparity\nand misalignment into consideration. First, we introduce self-knowledge guided\nregularization, which assigns differentiated regularization weights to each\nclass based on its training state, alleviating class disparity. Additionally,\nwe propose self-knowledge guided label relaxation, which adjusts label\nrelaxation according to the training accuracy, alleviating the misalignment and\nimproving robustness. By combining these methods, we formulate the\nSelf-Knowledge Guided FAT (SKG-FAT), leveraging naturally generated knowledge\nduring training to enhance the adversarial robustness without compromising\ntraining efficiency. Extensive experiments on four standard datasets\ndemonstrate that the SKG-FAT improves the robustness and preserves competitive\nclean accuracy, outperforming the state-of-the-art methods.\n", "date": "2024-09-26", "categories": []}
{"id": "2409.17874", "title": "DarkSAM: Fooling Segment Anything Model to Segment Nothing", "abstract": "  Segment Anything Model (SAM) has recently gained much attention for its\noutstanding generalization to unseen data and tasks. Despite its promising\nprospect, the vulnerabilities of SAM, especially to universal adversarial\nperturbation (UAP) have not been thoroughly investigated yet. In this paper, we\npropose DarkSAM, the first prompt-free universal attack framework against SAM,\nincluding a semantic decoupling-based spatial attack and a texture\ndistortion-based frequency attack. We first divide the output of SAM into\nforeground and background. Then, we design a shadow target strategy to obtain\nthe semantic blueprint of the image as the attack target. DarkSAM is dedicated\nto fooling SAM by extracting and destroying crucial object features from images\nin both spatial and frequency domains. In the spatial domain, we disrupt the\nsemantics of both the foreground and background in the image to confuse SAM. In\nthe frequency domain, we further enhance the attack effectiveness by distorting\nthe high-frequency components (i.e., texture information) of the image.\nConsequently, with a single UAP, DarkSAM renders SAM incapable of segmenting\nobjects across diverse images with varying prompts. Experimental results on\nfour datasets for SAM and its two variant models demonstrate the powerful\nattack capability and transferability of DarkSAM.\n", "date": "2024-09-26", "categories": []}
{"id": "2409.17963", "title": "CNCA: Toward Customizable and Natural Generation of Adversarial\n  Camouflage for Vehicle Detectors", "abstract": "  Prior works on physical adversarial camouflage against vehicle detectors\nmainly focus on the effectiveness and robustness of the attack. The current\nmost successful methods optimize 3D vehicle texture at a pixel level. However,\nthis results in conspicuous and attention-grabbing patterns in the generated\ncamouflage, which humans can easily identify. To address this issue, we propose\na Customizable and Natural Camouflage Attack (CNCA) method by leveraging an\noff-the-shelf pre-trained diffusion model. By sampling the optimal texture\nimage from the diffusion model with a user-specific text prompt, our method can\ngenerate natural and customizable adversarial camouflage while maintaining high\nattack performance. With extensive experiments on the digital and physical\nworlds and user studies, the results demonstrate that our proposed method can\ngenerate significantly more natural-looking camouflage than the\nstate-of-the-art baselines while achieving competitive attack performance. Our\ncode is available at\n\\href{https://anonymous.4open.science/r/CNCA-1D54}{https://anonymous.4open.science/r/CNCA-1D54}\n", "date": "2024-09-26", "categories": []}
{"id": "2409.17977", "title": "Cross-Modality Attack Boosted by Gradient-Evolutionary Multiform\n  Optimization", "abstract": "  In recent years, despite significant advancements in adversarial attack\nresearch, the security challenges in cross-modal scenarios, such as the\ntransferability of adversarial attacks between infrared, thermal, and RGB\nimages, have been overlooked. These heterogeneous image modalities collected by\ndifferent hardware devices are widely prevalent in practical applications, and\nthe substantial differences between modalities pose significant challenges to\nattack transferability. In this work, we explore a novel cross-modal\nadversarial attack strategy, termed multiform attack. We propose a dual-layer\noptimization framework based on gradient-evolution, facilitating efficient\nperturbation transfer between modalities. In the first layer of optimization,\nthe framework utilizes image gradients to learn universal perturbations within\neach modality and employs evolutionary algorithms to search for shared\nperturbations with transferability across different modalities through\nsecondary optimization. Through extensive testing on multiple heterogeneous\ndatasets, we demonstrate the superiority and robustness of Multiform Attack\ncompared to existing techniques. This work not only enhances the\ntransferability of cross-modal adversarial attacks but also provides a new\nperspective for understanding security vulnerabilities in cross-modal systems.\n", "date": "2024-09-26", "categories": []}
{"id": "2406.03589", "title": "Ranking Manipulation for Conversational Search Engines", "abstract": "  Major search engine providers are rapidly incorporating Large Language Model\n(LLM)-generated content in response to user queries. These conversational\nsearch engines operate by loading retrieved website text into the LLM context\nfor summarization and interpretation. Recent research demonstrates that LLMs\nare highly vulnerable to jailbreaking and prompt injection attacks, which\ndisrupt the safety and quality goals of LLMs using adversarial strings. This\nwork investigates the impact of prompt injections on the ranking order of\nsources referenced by conversational search engines. To this end, we introduce\na focused dataset of real-world consumer product websites and formalize\nconversational search ranking as an adversarial problem. Experimentally, we\nanalyze conversational search rankings in the absence of adversarial injections\nand show that different LLMs vary significantly in prioritizing product name,\ndocument content, and context position. We then present a tree-of-attacks-based\njailbreaking technique which reliably promotes low-ranked products.\nImportantly, these attacks transfer effectively to state-of-the-art\nconversational search engines such as perplexity$.$ai. Given the strong\nfinancial incentive for website owners to boost their search ranking, we argue\nthat our problem formulation is of critical importance for future robustness\nwork.\n", "date": "2024-06-05", "categories": []}
{"id": "2409.15119", "title": "Log-normal Mutations and their Use in Detecting Surreptitious Fake\n  Images", "abstract": "  In many cases, adversarial attacks are based on specialized algorithms\nspecifically dedicated to attacking automatic image classifiers. These\nalgorithms perform well, thanks to an excellent ad hoc distribution of initial\nattacks. However, these attacks are easily detected due to their specific\ninitial distribution. We therefore consider other black-box attacks, inspired\nfrom generic black-box optimization tools, and in particular the log-normal\nalgorithm.\n  We apply the log-normal method to the attack of fake detectors, and get\nsuccessful attacks: importantly, these attacks are not detected by detectors\nspecialized on classical adversarial attacks. Then, combining these attacks and\ndeep detection, we create improved fake detectors.\n", "date": "2024-09-23", "categories": []}
{"id": "2409.16491", "title": "Proactive Schemes: A Survey of Adversarial Attacks for Social Good", "abstract": "  Adversarial attacks in computer vision exploit the vulnerabilities of machine\nlearning models by introducing subtle perturbations to input data, often\nleading to incorrect predictions or classifications. These attacks have evolved\nin sophistication with the advent of deep learning, presenting significant\nchallenges in critical applications, which can be harmful for society. However,\nthere is also a rich line of research from a transformative perspective that\nleverages adversarial techniques for social good. Specifically, we examine the\nrise of proactive schemes-methods that encrypt input data using additional\nsignals termed templates, to enhance the performance of deep learning models.\nBy embedding these imperceptible templates into digital media, proactive\nschemes are applied across various applications, from simple image enhancements\nto complicated deep learning frameworks to aid performance, as compared to the\npassive schemes, which don't change the input data distribution for their\nframework. The survey delves into the methodologies behind these proactive\nschemes, the encryption and learning processes, and their application to modern\ncomputer vision and natural language processing applications. Additionally, it\ndiscusses the challenges, potential vulnerabilities, and future directions for\nproactive schemes, ultimately highlighting their potential to foster the\nresponsible and secure advancement of deep learning technologies.\n", "date": "2024-09-24", "categories": []}
{"id": "2409.02485", "title": "Adversarial Attacks on Machine Learning-Aided Visualizations", "abstract": "  Research in ML4VIS investigates how to use machine learning (ML) techniques\nto generate visualizations, and the field is rapidly growing with high societal\nimpact. However, as with any computational pipeline that employs ML processes,\nML4VIS approaches are susceptible to a range of ML-specific adversarial\nattacks. These attacks can manipulate visualization generations, causing\nanalysts to be tricked and their judgments to be impaired. Due to a lack of\nsynthesis from both visualization and ML perspectives, this security aspect is\nlargely overlooked by the current ML4VIS literature. To bridge this gap, we\ninvestigate the potential vulnerabilities of ML-aided visualizations from\nadversarial attacks using a holistic lens of both visualization and ML\nperspectives. We first identify the attack surface (i.e., attack entry points)\nthat is unique in ML-aided visualizations. We then exemplify five different\nadversarial attacks. These examples highlight the range of possible attacks\nwhen considering the attack surface and multiple different adversary\ncapabilities. Our results show that adversaries can induce various attacks,\nsuch as creating arbitrary and deceptive visualizations, by systematically\nidentifying input attributes that are influential in ML inferences. Based on\nour observations of the attack surface characteristics and the attack examples,\nwe underline the importance of comprehensive studies of security issues and\ndefense mechanisms as a call of urgency for the ML4VIS community.\n", "date": "2024-09-04", "categories": []}
{"id": "2409.06793", "title": "Adversarial Attacks to Multi-Modal Models", "abstract": "  Multi-modal models have gained significant attention due to their powerful\ncapabilities. These models effectively align embeddings across diverse data\nmodalities, showcasing superior performance in downstream tasks compared to\ntheir unimodal counterparts. Recent study showed that the attacker can\nmanipulate an image or audio file by altering it in such a way that its\nembedding matches that of an attacker-chosen targeted input, thereby deceiving\ndownstream models. However, this method often underperforms due to inherent\ndisparities in data from different modalities. In this paper, we introduce\nCrossFire, an innovative approach to attack multi-modal models. CrossFire\nbegins by transforming the targeted input chosen by the attacker into a format\nthat matches the modality of the original image or audio file. We then\nformulate our attack as an optimization problem, aiming to minimize the angular\ndeviation between the embeddings of the transformed input and the modified\nimage or audio file. Solving this problem determines the perturbations to be\nadded to the original media. Our extensive experiments on six real-world\nbenchmark datasets reveal that CrossFire can significantly manipulate\ndownstream tasks, surpassing existing attacks. Additionally, we evaluate six\ndefensive strategies against CrossFire, finding that current defenses are\ninsufficient to counteract our CrossFire.\n", "date": "2024-09-10", "categories": []}
{"id": "2409.15381", "title": "Adversarial Attacks on Parts of Speech: An Empirical Study in\n  Text-to-Image Generation", "abstract": "  Recent studies show that text-to-image (T2I) models are vulnerable to\nadversarial attacks, especially with noun perturbations in text prompts. In\nthis study, we investigate the impact of adversarial attacks on different POS\ntags within text prompts on the images generated by T2I models. We create a\nhigh-quality dataset for realistic POS tag token swapping and perform\ngradient-based attacks to find adversarial suffixes that mislead T2I models\ninto generating images with altered tokens. Our empirical results show that the\nattack success rate (ASR) varies significantly among different POS tag\ncategories, with nouns, proper nouns, and adjectives being the easiest to\nattack. We explore the mechanism behind the steering effect of adversarial\nsuffixes, finding that the number of critical tokens and content fusion vary\namong POS tags, while features like suffix transferability are consistent\nacross categories. We have made our implementation publicly available at -\nhttps://github.com/shahariar-shibli/Adversarial-Attack-on-POS-Tags.\n", "date": "2024-09-21", "categories": []}
{"id": "2311.08598", "title": "$DA^3$: A Distribution-Aware Adversarial Attack against Language Models", "abstract": "  Language models can be manipulated by adversarial attacks, which introduce\nsubtle perturbations to input data. While recent attack methods can achieve a\nrelatively high attack success rate (ASR), we've observed that the generated\nadversarial examples have a different data distribution compared with the\noriginal examples. Specifically, these adversarial examples exhibit reduced\nconfidence levels and greater divergence from the training data distribution.\nConsequently, they are easy to detect using straightforward detection methods,\ndiminishing the efficacy of such attacks. To address this issue, we propose a\nDistribution-Aware Adversarial Attack ($DA^3$) method. $DA^3$ considers the\ndistribution shifts of adversarial examples to improve attacks' effectiveness\nunder detection methods. We further design a novel evaluation metric, the\nNon-detectable Attack Success Rate (NASR), which integrates both ASR and\ndetectability for the attack task. We conduct experiments on four widely used\ndatasets to validate the attack effectiveness and transferability of\nadversarial examples generated by $DA^3$ against both the white-box BERT-base\nand RoBERTa-base models and the black-box LLaMA2-7b model.\n", "date": "2023-11-14", "categories": []}
{"id": "2402.06827", "title": "RAMP: Boosting Adversarial Robustness Against Multiple $l_p$\n  Perturbations for Universal Robustness", "abstract": "  Most existing works focus on improving robustness against adversarial attacks\nbounded by a single $l_p$ norm using adversarial training (AT). However, these\nAT models' multiple-norm robustness (union accuracy) is still low, which is\ncrucial since in the real-world an adversary is not necessarily bounded by a\nsingle norm. The tradeoffs among robustness against multiple $l_p$\nperturbations and accuracy/robustness make obtaining good union and clean\naccuracy challenging. We design a logit pairing loss to improve the union\naccuracy by analyzing the tradeoffs from the lens of distribution shifts. We\nconnect natural training (NT) with AT via gradient projection, to incorporate\nuseful information from NT into AT, where we empirically and theoretically show\nit moderates the accuracy/robustness tradeoff. We propose a novel training\nframework \\textbf{RAMP}, to boost the robustness against multiple $l_p$\nperturbations. \\textbf{RAMP} can be easily adapted for robust fine-tuning and\nfull AT. For robust fine-tuning, \\textbf{RAMP} obtains a union accuracy up to\n$53.3\\%$ on CIFAR-10, and $29.1\\%$ on ImageNet. For training from scratch,\n\\textbf{RAMP} achieves a union accuracy of $44.6\\%$ and good clean accuracy of\n$81.2\\%$ on ResNet-18 against AutoAttack on CIFAR-10. Beyond multi-norm\nrobustness \\textbf{RAMP}-trained models achieve superior \\textit{universal\nrobustness}, effectively generalizing against a range of unseen adversaries and\nnatural corruptions.\n", "date": "2024-02-09", "categories": []}
{"id": "2409.13828", "title": "ViTGuard: Attention-aware Detection against Adversarial Examples for\n  Vision Transformer", "abstract": "  The use of transformers for vision tasks has challenged the traditional\ndominant role of convolutional neural networks (CNN) in computer vision (CV).\nFor image classification tasks, Vision Transformer (ViT) effectively\nestablishes spatial relationships between patches within images, directing\nattention to important areas for accurate predictions. However, similar to\nCNNs, ViTs are vulnerable to adversarial attacks, which mislead the image\nclassifier into making incorrect decisions on images with carefully designed\nperturbations. Moreover, adversarial patch attacks, which introduce arbitrary\nperturbations within a small area, pose a more serious threat to ViTs. Even\nworse, traditional detection methods, originally designed for CNN models, are\nimpractical or suffer significant performance degradation when applied to ViTs,\nand they generally overlook patch attacks.\n  In this paper, we propose ViTGuard as a general detection method for\ndefending ViT models against adversarial attacks, including typical attacks\nwhere perturbations spread over the entire input and patch attacks. ViTGuard\nuses a Masked Autoencoder (MAE) model to recover randomly masked patches from\nthe unmasked regions, providing a flexible image reconstruction strategy. Then,\nthreshold-based detectors leverage distinctive ViT features, including\nattention maps and classification (CLS) token representations, to distinguish\nbetween normal and adversarial samples. The MAE model does not involve any\nadversarial samples during training, ensuring the effectiveness of our\ndetectors against unseen attacks. ViTGuard is compared with seven existing\ndetection methods under nine attacks across three datasets. The evaluation\nresults show the superiority of ViTGuard over existing detectors. Finally,\nconsidering the potential detection evasion, we further demonstrate ViTGuard's\nrobustness against adaptive attacks for evasion.\n", "date": "2024-09-20", "categories": []}
{"id": "2409.13945", "title": "PureDiffusion: Using Backdoor to Counter Backdoor in Generative\n  Diffusion Models", "abstract": "  Diffusion models (DMs) are advanced deep learning models that achieved\nstate-of-the-art capability on a wide range of generative tasks. However,\nrecent studies have shown their vulnerability regarding backdoor attacks, in\nwhich backdoored DMs consistently generate a designated result (e.g., a harmful\nimage) called backdoor target when the models' input contains a backdoor\ntrigger. Although various backdoor techniques have been investigated to attack\nDMs, defense methods against these threats are still limited and underexplored,\nespecially in inverting the backdoor trigger. In this paper, we introduce\nPureDiffusion, a novel backdoor defense framework that can efficiently detect\nbackdoor attacks by inverting backdoor triggers embedded in DMs. Our extensive\nexperiments on various trigger-target pairs show that PureDiffusion outperforms\nexisting defense methods with a large gap in terms of fidelity (i.e., how much\nthe inverted trigger resembles the original trigger) and backdoor success rate\n(i.e., the rate that the inverted trigger leads to the corresponding backdoor\ntarget). Notably, in certain cases, backdoor triggers inverted by PureDiffusion\neven achieve higher attack success rate than the original triggers.\n", "date": "2024-09-20", "categories": []}
{"id": "2409.14240", "title": "Cloud Adversarial Example Generation for Remote Sensing Image\n  Classification", "abstract": "  Most existing adversarial attack methods for remote sensing images merely add\nadversarial perturbations or patches, resulting in unnatural modifications.\nClouds are common atmospheric effects in remote sensing images. Generating\nclouds on these images can produce adversarial examples better aligning with\nhuman perception. In this paper, we propose a Perlin noise based cloud\ngeneration attack method. Common Perlin noise based cloud generation is a\nrandom, non-optimizable process, which cannot be directly used to attack the\ntarget models. We design a Perlin Gradient Generator Network (PGGN), which\ntakes a gradient parameter vector as input and outputs the grids of Perlin\nnoise gradient vectors at different scales. After a series of computations\nbased on the gradient vectors, cloud masks at corresponding scales can be\nproduced. These cloud masks are then weighted and summed depending on a mixing\ncoefficient vector and a scaling factor to produce the final cloud masks. The\ngradient vector, coefficient vector and scaling factor are collectively\nrepresented as a cloud parameter vector, transforming the cloud generation into\na black-box optimization problem. The Differential Evolution (DE) algorithm is\nemployed to solve for the optimal solution of the cloud parameter vector,\nachieving a query-based black-box attack. Detailed experiments confirm that\nthis method has strong attack capabilities and achieves high query efficiency.\nAdditionally, we analyze the transferability of the generated adversarial\nexamples and their robustness in adversarial defense scenarios.\n", "date": "2024-09-21", "categories": []}
{"id": "2409.14940", "title": "Improving Adversarial Robustness for 3D Point Cloud Recognition at\n  Test-Time through Purified Self-Training", "abstract": "  Recognizing 3D point cloud plays a pivotal role in many real-world\napplications. However, deploying 3D point cloud deep learning model is\nvulnerable to adversarial attacks. Despite many efforts into developing robust\nmodel by adversarial training, they may become less effective against emerging\nattacks. This limitation motivates the development of adversarial purification\nwhich employs generative model to mitigate the impact of adversarial attacks.\nIn this work, we highlight the remaining challenges from two perspectives.\nFirst, the purification based method requires retraining the classifier on\npurified samples which introduces additional computation overhead. Moreover, in\na more realistic scenario, testing samples arrives in a streaming fashion and\nadversarial samples are not isolated from clean samples. These challenges\nmotivates us to explore dynamically update model upon observing testing\nsamples. We proposed a test-time purified self-training strategy to achieve\nthis objective. Adaptive thresholding and feature distribution alignment are\nintroduced to improve the robustness of self-training. Extensive results on\ndifferent adversarial attacks suggest the proposed method is complementary to\npurification based method in handling continually changing adversarial attacks\non the testing data stream.\n", "date": "2024-09-23", "categories": []}
{"id": "2409.15190", "title": "Interpretability-Guided Test-Time Adversarial Defense", "abstract": "  We propose a novel and low-cost test-time adversarial defense by devising\ninterpretability-guided neuron importance ranking methods to identify neurons\nimportant to the output classes. Our method is a training-free approach that\ncan significantly improve the robustness-accuracy tradeoff while incurring\nminimal computational overhead. While being among the most efficient test-time\ndefenses (4x faster), our method is also robust to a wide range of black-box,\nwhite-box, and adaptive attacks that break previous test-time defenses. We\ndemonstrate the efficacy of our method for CIFAR10, CIFAR100, and ImageNet-1k\non the standard RobustBench benchmark (with average gains of 2.6%, 4.9%, and\n2.8% respectively). We also show improvements (average 1.5%) over the\nstate-of-the-art test-time defenses even under strong adaptive attacks.\n", "date": "2024-09-23", "categories": []}
{"id": "2408.12673", "title": "Enhancing Transferability of Adversarial Attacks with GE-AdvGAN+: A\n  Comprehensive Framework for Gradient Editing", "abstract": "  Transferable adversarial attacks pose significant threats to deep neural\nnetworks, particularly in black-box scenarios where internal model information\nis inaccessible. Studying adversarial attack methods helps advance the\nperformance of defense mechanisms and explore model vulnerabilities. These\nmethods can uncover and exploit weaknesses in models, promoting the development\nof more robust architectures. However, current methods for transferable attacks\noften come with substantial computational costs, limiting their deployment and\napplication, especially in edge computing scenarios. Adversarial generative\nmodels, such as Generative Adversarial Networks (GANs), are characterized by\ntheir ability to generate samples without the need for retraining after an\ninitial training phase. GE-AdvGAN, a recent method for transferable adversarial\nattacks, is based on this principle. In this paper, we propose a novel general\nframework for gradient editing-based transferable attacks, named GE-AdvGAN+,\nwhich integrates nearly all mainstream attack methods to enhance\ntransferability while significantly reducing computational resource\nconsumption. Our experiments demonstrate the compatibility and effectiveness of\nour framework. Compared to the baseline AdvGAN, our best-performing method,\nGE-AdvGAN++, achieves an average ASR improvement of 47.8. Additionally, it\nsurpasses the latest competing algorithm, GE-AdvGAN, with an average ASR\nincrease of 5.9. The framework also exhibits enhanced computational efficiency,\nachieving 2217.7 FPS, outperforming traditional methods such as BIM and\nMI-FGSM. The implementation code for our GE-AdvGAN+ framework is available at\nhttps://github.com/GEAdvGANP\n", "date": "2024-08-22", "categories": []}
{"id": "2409.12997", "title": "VCAT: Vulnerability-aware and Curiosity-driven Adversarial Training for\n  Enhancing Autonomous Vehicle Robustness", "abstract": "  Autonomous vehicles (AVs) face significant threats to their safe operation in\ncomplex traffic environments. Adversarial training has emerged as an effective\nmethod of enabling AVs to preemptively fortify their robustness against\nmalicious attacks. Train an attacker using an adversarial policy, allowing the\nAV to learn robust driving through interaction with this attacker. However,\nadversarial policies in existing methodologies often get stuck in a loop of\noverexploiting established vulnerabilities, resulting in poor improvement for\nAVs. To overcome the limitations, we introduce a pioneering framework termed\nVulnerability-aware and Curiosity-driven Adversarial Training (VCAT).\nSpecifically, during the traffic vehicle attacker training phase, a surrogate\nnetwork is employed to fit the value function of the AV victim, providing dense\ninformation about the victim's inherent vulnerabilities. Subsequently, random\nnetwork distillation is used to characterize the novelty of the environment,\nconstructing an intrinsic reward to guide the attacker in exploring unexplored\nterritories. In the victim defense training phase, the AV is trained in\ncritical scenarios in which the pretrained attacker is positioned around the\nvictim to generate attack behaviors. Experimental results revealed that the\ntraining methodology provided by VCAT significantly improved the robust control\ncapabilities of learning-based AVs, outperforming both conventional training\nmodalities and alternative reinforcement learning counterparts, with a marked\nreduction in crash rates. The code is available at\nhttps://github.com/caixxuan/VCAT.\n", "date": "2024-09-19", "categories": []}
{"id": "2409.13546", "title": "Certified Adversarial Robustness via Partition-based Randomized\n  Smoothing", "abstract": "  A reliable application of deep neural network classifiers requires robustness\ncertificates against adversarial perturbations. Gaussian smoothing is a widely\nanalyzed approach to certifying robustness against norm-bounded perturbations,\nwhere the certified prediction radius depends on the variance of the Gaussian\nnoise and the confidence level of the neural net's prediction under the\nadditive Gaussian noise. However, in application to high-dimensional image\ndatasets, the certified radius of the plain Gaussian smoothing could be\nrelatively small, since Gaussian noise with high variances can significantly\nharm the visibility of an image. In this work, we propose the Pixel\nPartitioning-based Randomized Smoothing (PPRS) methodology to boost the neural\nnet's confidence score and thus the robustness radius of the certified\nprediction. We demonstrate that the proposed PPRS algorithm improves the\nvisibility of the images under additive Gaussian noise. We discuss the\nnumerical results of applying PPRS to standard computer vision datasets and\nneural network architectures. Our empirical findings indicate a considerable\nimprovement in the certified accuracy and stability of the prediction model to\nthe additive Gaussian noise in randomized smoothing.\n", "date": "2024-09-20", "categories": []}
{"id": "2212.06431", "title": "Object-fabrication Targeted Attack for Object Detection", "abstract": "  Recent studies have demonstrated that object detection networks are usually\nvulnerable to adversarial examples. Generally, adversarial attacks for object\ndetection can be categorized into targeted and untargeted attacks. Compared\nwith untargeted attacks, targeted attacks present greater challenges and all\nexisting targeted attack methods launch the attack by misleading detectors to\nmislabel the detected object as a specific wrong label. However, since these\nmethods must depend on the presence of the detected objects within the victim\nimage, they suffer from limitations in attack scenarios and attack success\nrates. In this paper, we propose a targeted feature space attack method that\ncan mislead detectors to `fabricate' extra designated objects regardless of\nwhether the victim image contains objects or not. Specifically, we introduce a\nguided image to extract coarse-grained features of the target objects and\ndesign an innovative dual attention mechanism to filter out the critical\nfeatures of the target objects efficiently. The attack performance of the\nproposed method is evaluated on MS COCO and BDD100K datasets with FasterRCNN\nand YOLOv5. Evaluation results indicate that the proposed targeted feature\nspace attack method shows significant improvements in terms of image-specific,\nuniversality, and generalization attack performance, compared with the previous\ntargeted attack for object detection.\n", "date": "2022-12-13", "categories": []}
{"id": "2308.11406", "title": "Designing an attack-defense game: how to increase robustness of\n  financial transaction models via a competition", "abstract": "  Banks routinely use neural networks to make decisions. While these models\noffer higher accuracy, they are susceptible to adversarial attacks, a risk\noften overlooked in the context of event sequences, particularly sequences of\nfinancial transactions, as most works consider computer vision and NLP\nmodalities.\n  We propose a thorough approach to studying these risks: a novel type of\ncompetition that allows a realistic and detailed investigation of problems in\nfinancial transaction data. The participants directly oppose each other,\nproposing attacks and defenses -- so they are examined in close-to-real-life\nconditions.\n  The paper outlines our unique competition structure with direct opposition of\nparticipants, presents results for several different top submissions, and\nanalyzes the competition results. We also introduce a new open dataset\nfeaturing financial transactions with credit default labels, enhancing the\nscope for practical research and development.\n", "date": "2023-08-22", "categories": []}
{"id": "2402.18607", "title": "Exploring Privacy and Fairness Risks in Sharing Diffusion Models: An\n  Adversarial Perspective", "abstract": "  Diffusion models have recently gained significant attention in both academia\nand industry due to their impressive generative performance in terms of both\nsampling quality and distribution coverage. Accordingly, proposals are made for\nsharing pre-trained diffusion models across different organizations, as a way\nof improving data utilization while enhancing privacy protection by avoiding\nsharing private data directly. However, the potential risks associated with\nsuch an approach have not been comprehensively examined.\n  In this paper, we take an adversarial perspective to investigate the\npotential privacy and fairness risks associated with the sharing of diffusion\nmodels. Specifically, we investigate the circumstances in which one party (the\nsharer) trains a diffusion model using private data and provides another party\n(the receiver) black-box access to the pre-trained model for downstream tasks.\nWe demonstrate that the sharer can execute fairness poisoning attacks to\nundermine the receiver's downstream models by manipulating the training data\ndistribution of the diffusion model. Meanwhile, the receiver can perform\nproperty inference attacks to reveal the distribution of sensitive features in\nthe sharer's dataset. Our experiments conducted on real-world datasets\ndemonstrate remarkable attack performance on different types of diffusion\nmodels, which highlights the critical importance of robust data auditing and\nprivacy protection protocols in pertinent applications.\n", "date": "2024-02-28", "categories": []}
{"id": "2409.12379", "title": "Enhancing 3D Robotic Vision Robustness by Minimizing Adversarial Mutual\n  Information through a Curriculum Training Approach", "abstract": "  Adversarial attacks exploit vulnerabilities in a model's decision boundaries\nthrough small, carefully crafted perturbations that lead to significant\nmispredictions. In 3D vision, the high dimensionality and sparsity of data\ngreatly expand the attack surface, making 3D vision particularly vulnerable for\nsafety-critical robotics. To enhance 3D vision's adversarial robustness, we\npropose a training objective that simultaneously minimizes prediction loss and\nmutual information (MI) under adversarial perturbations to contain the upper\nbound of misprediction errors. This approach simplifies handling adversarial\nexamples compared to conventional methods, which require explicit searching and\ntraining on adversarial samples. However, minimizing prediction loss conflicts\nwith minimizing MI, leading to reduced robustness and catastrophic forgetting.\nTo address this, we integrate curriculum advisors in the training setup that\ngradually introduce adversarial objectives to balance training and prevent\nmodels from being overwhelmed by difficult cases early in the process. The\nadvisors also enhance robustness by encouraging training on diverse MI examples\nthrough entropy regularizers. We evaluated our method on ModelNet40 and KITTI\nusing PointNet, DGCNN, SECOND, and PointTransformers, achieving 2-5% accuracy\ngains on ModelNet40 and a 5-10% mAP improvement in object detection. Our code\nis publicly available at https://github.com/nstrndrbi/Mine-N-Learn.\n", "date": "2024-09-19", "categories": []}
{"id": "2409.12472", "title": "TEAM: Temporal Adversarial Examples Attack Model against Network\n  Intrusion Detection System Applied to RNN", "abstract": "  With the development of artificial intelligence, neural networks play a key\nrole in network intrusion detection systems (NIDS). Despite the tremendous\nadvantages, neural networks are susceptible to adversarial attacks. To improve\nthe reliability of NIDS, many research has been conducted and plenty of\nsolutions have been proposed. However, the existing solutions rarely consider\nthe adversarial attacks against recurrent neural networks (RNN) with time\nsteps, which would greatly affect the application of NIDS in real world.\nTherefore, we first propose a novel RNN adversarial attack model based on\nfeature reconstruction called \\textbf{T}emporal adversarial \\textbf{E}xamples\n\\textbf{A}ttack \\textbf{M}odel \\textbf{(TEAM)}, which applied to time series\ndata and reveals the potential connection between adversarial and time steps in\nRNN. That is, the past adversarial examples within the same time steps can\ntrigger further attacks on current or future original examples. Moreover, TEAM\nleverages Time Dilation (TD) to effectively mitigates the effect of temporal\namong adversarial examples within the same time steps. Experimental results\nshow that in most attack categories, TEAM improves the misjudgment rate of NIDS\non both black and white boxes, making the misjudgment rate reach more than\n96.68%. Meanwhile, the maximum increase in the misjudgment rate of the NIDS for\nsubsequent original samples exceeds 95.57%.\n", "date": "2024-09-19", "categories": []}
{"id": "2409.12642", "title": "Deep generative models as an adversarial attack strategy for tabular\n  machine learning", "abstract": "  Deep Generative Models (DGMs) have found application in computer vision for\ngenerating adversarial examples to test the robustness of machine learning (ML)\nsystems. Extending these adversarial techniques to tabular ML presents unique\nchallenges due to the distinct nature of tabular data and the necessity to\npreserve domain constraints in adversarial examples. In this paper, we adapt\nfour popular tabular DGMs into adversarial DGMs (AdvDGMs) and evaluate their\neffectiveness in generating realistic adversarial examples that conform to\ndomain constraints.\n", "date": "2024-09-19", "categories": []}
{"id": "2409.12946", "title": "Revisiting Semi-supervised Adversarial Robustness via Noise-aware Online\n  Robust Distillation", "abstract": "  The robust self-training (RST) framework has emerged as a prominent approach\nfor semi-supervised adversarial training. To explore the possibility of\ntackling more complicated tasks with even lower labeling budgets, unlike prior\napproaches that rely on robust pretrained models, we present SNORD - a simple\nyet effective framework that introduces contemporary semi-supervised learning\ntechniques into the realm of adversarial training. By enhancing pseudo labels\nand managing noisy training data more effectively, SNORD showcases impressive,\nstate-of-the-art performance across diverse datasets and labeling budgets, all\nwithout the need for pretrained models. Compared to full adversarial\nsupervision, SNORD achieves a 90% relative robust accuracy under epsilon =\n8/255 AutoAttack, requiring less than 0.1%, 2%, and 10% labels for CIFAR-10,\nCIFAR-100, and TinyImageNet-200, respectively. Additional experiments confirm\nthe efficacy of each component and demonstrate the adaptability of integrating\nSNORD with existing adversarial pretraining strategies to further bolster\nrobustness.\n", "date": "2024-09-19", "categories": []}
{"id": "2309.00236", "title": "Image Hijacks: Adversarial Images can Control Generative Models at\n  Runtime", "abstract": "  Are foundation models secure against malicious actors? In this work, we focus\non the image input to a vision-language model (VLM). We discover image hijacks,\nadversarial images that control the behaviour of VLMs at inference time, and\nintroduce the general Behaviour Matching algorithm for training image hijacks.\nFrom this, we derive the Prompt Matching method, allowing us to train hijacks\nmatching the behaviour of an arbitrary user-defined text prompt (e.g. 'the\nEiffel Tower is now located in Rome') using a generic, off-the-shelf dataset\nunrelated to our choice of prompt. We use Behaviour Matching to craft hijacks\nfor four types of attack, forcing VLMs to generate outputs of the adversary's\nchoice, leak information from their context window, override their safety\ntraining, and believe false statements. We study these attacks against LLaVA, a\nstate-of-the-art VLM based on CLIP and LLaMA-2, and find that all attack types\nachieve a success rate of over 80%. Moreover, our attacks are automated and\nrequire only small image perturbations.\n", "date": "2023-09-01", "categories": []}
{"id": "2409.11454", "title": "Golden Ratio Search: A Low-Power Adversarial Attack for Deep Learning\n  based Modulation Classification", "abstract": "  We propose a minimal power white box adversarial attack for Deep Learning\nbased Automatic Modulation Classification (AMC). The proposed attack uses the\nGolden Ratio Search (GRS) method to find powerful attacks with minimal power.\nWe evaluate the efficacy of the proposed method by comparing it with existing\nadversarial attack approaches. Additionally, we test the robustness of the\nproposed attack against various state-of-the-art architectures, including\ndefense mechanisms such as adversarial training, binarization, and ensemble\nmethods. Experimental results demonstrate that the proposed attack is powerful,\nrequires minimal power, and can be generated in less time, significantly\nchallenging the resilience of current AMC methods.\n", "date": "2024-09-17", "categories": []}
{"id": "2409.11754", "title": "NPAT Null-Space Projected Adversarial Training Towards Zero\n  Deterioration", "abstract": "  To mitigate the susceptibility of neural networks to adversarial attacks,\nadversarial training has emerged as a prevalent and effective defense strategy.\nIntrinsically, this countermeasure incurs a trade-off, as it sacrifices the\nmodel's accuracy in processing normal samples. To reconcile the trade-off, we\npioneer the incorporation of null-space projection into adversarial training\nand propose two innovative Null-space Projection based Adversarial\nTraining(NPAT) algorithms tackling sample generation and gradient optimization,\nnamed Null-space Projected Data Augmentation (NPDA) and Null-space Projected\nGradient Descent (NPGD), to search for an overarching optimal solutions, which\nenhance robustness with almost zero deterioration in generalization\nperformance. Adversarial samples and perturbations are constrained within the\nnull-space of the decision boundary utilizing a closed-form null-space\nprojector, effectively mitigating threat of attack stemming from unreliable\nfeatures. Subsequently, we conducted experiments on the CIFAR10 and SVHN\ndatasets and reveal that our methodology can seamlessly combine with\nadversarial training methods and obtain comparable robustness while keeping\ngeneralization close to a high-accuracy model.\n", "date": "2024-09-18", "categories": []}
{"id": "2409.12072", "title": "PAD-FT: A Lightweight Defense for Backdoor Attacks via Data Purification\n  and Fine-Tuning", "abstract": "  Backdoor attacks pose a significant threat to deep neural networks,\nparticularly as recent advancements have led to increasingly subtle\nimplantation, making the defense more challenging. Existing defense mechanisms\ntypically rely on an additional clean dataset as a standard reference and\ninvolve retraining an auxiliary model or fine-tuning the entire victim model.\nHowever, these approaches are often computationally expensive and not always\nfeasible in practical applications. In this paper, we propose a novel and\nlightweight defense mechanism, termed PAD-FT, that does not require an\nadditional clean dataset and fine-tunes only a very small part of the model to\ndisinfect the victim model. To achieve this, our approach first introduces a\nsimple data purification process to identify and select the most-likely clean\ndata from the poisoned training dataset. The self-purified clean dataset is\nthen used for activation clipping and fine-tuning only the last classification\nlayer of the victim model. By integrating data purification, activation\nclipping, and classifier fine-tuning, our mechanism PAD-FT demonstrates\nsuperior effectiveness across multiple backdoor attack methods and datasets, as\nconfirmed through extensive experimental evaluation.\n", "date": "2024-09-18", "categories": []}
{"id": "2409.10578", "title": "GLEAN: Generative Learning for Eliminating Adversarial Noise", "abstract": "  In the age of powerful diffusion models such as DALL-E and Stable Diffusion,\nmany in the digital art community have suffered style mimicry attacks due to\nfine-tuning these models on their works. The ability to mimic an artist's style\nvia text-to-image diffusion models raises serious ethical issues, especially\nwithout explicit consent. Glaze, a tool that applies various ranges of\nperturbations to digital art, has shown significant success in preventing style\nmimicry attacks, at the cost of artifacts ranging from imperceptible noise to\nsevere quality degradation. The release of Glaze has sparked further\ndiscussions regarding the effectiveness of similar protection methods. In this\npaper, we propose GLEAN- applying I2I generative networks to strip\nperturbations from Glazed images, evaluating the performance of style mimicry\nattacks before and after GLEAN on the results of Glaze. GLEAN aims to support\nand enhance Glaze by highlighting its limitations and encouraging further\ndevelopment.\n", "date": "2024-09-15", "categories": []}
{"id": "2311.07127", "title": "Multi-agent Attacks for Black-box Social Recommendations", "abstract": "  The rise of online social networks has facilitated the evolution of social\nrecommender systems, which incorporate social relations to enhance users'\ndecision-making process. With the great success of Graph Neural Networks (GNNs)\nin learning node representations, GNN-based social recommendations have been\nwidely studied to model user-item interactions and user-user social relations\nsimultaneously. Despite their great successes, recent studies have shown that\nthese advanced recommender systems are highly vulnerable to adversarial\nattacks, in which attackers can inject well-designed fake user profiles to\ndisrupt recommendation performances. While most existing studies mainly focus\non argeted attacks to promote target items on vanilla recommender systems,\nuntargeted attacks to degrade the overall prediction performance are less\nexplored on social recommendations under a black-box scenario. To perform\nuntargeted attacks on social recommender systems, attackers can construct\nmalicious social relationships for fake users to enhance the attack\nperformance. However, the coordination of social relations and item profiles is\nchallenging for attacking black-box social recommendations. To address this\nlimitation, we first conduct several preliminary studies to demonstrate the\neffectiveness of cross-community connections and cold-start items in degrading\nrecommendations performance. Specifically, we propose a novel framework\nMultiAttack based on multi-agent reinforcement learning to coordinate the\ngeneration of cold-start item profiles and cross-community social relations for\nconducting untargeted attacks on black-box social recommendations.\nComprehensive experiments on various real-world datasets demonstrate the\neffectiveness of our proposed attacking framework under the black-box setting.\n", "date": "2023-11-13", "categories": []}
{"id": "2403.15918", "title": "Towards Adversarial Robustness And Backdoor Mitigation in SSL", "abstract": "  Self-Supervised Learning (SSL) has shown great promise in learning\nrepresentations from unlabeled data. The power of learning representations\nwithout the need for human annotations has made SSL a widely used technique in\nreal-world problems. However, SSL methods have recently been shown to be\nvulnerable to backdoor attacks, where the learned model can be exploited by\nadversaries to manipulate the learned representations, either through tampering\nthe training data distribution, or via modifying the model itself. This work\naims to address defending against backdoor attacks in SSL, where the adversary\nhas access to a realistic fraction of the SSL training data, and no access to\nthe model. We use novel methods that are computationally efficient as well as\ngeneralizable across different problem settings. We also investigate the\nadversarial robustness of SSL models when trained with our method, and show\ninsights into increased robustness in SSL via frequency domain augmentations.\nWe demonstrate the effectiveness of our method on a variety of SSL benchmarks,\nand show that our method is able to mitigate backdoor attacks while maintaining\nhigh performance on downstream tasks. Code for our work is available at\ngithub.com/Aryan-Satpathy/Backdoor\n", "date": "2024-03-23", "categories": []}
{"id": "2404.08154", "title": "Eliminating Catastrophic Overfitting Via Abnormal Adversarial Examples\n  Regularization", "abstract": "  Single-step adversarial training (SSAT) has demonstrated the potential to\nachieve both efficiency and robustness. However, SSAT suffers from catastrophic\noverfitting (CO), a phenomenon that leads to a severely distorted classifier,\nmaking it vulnerable to multi-step adversarial attacks. In this work, we\nobserve that some adversarial examples generated on the SSAT-trained network\nexhibit anomalous behaviour, that is, although these training samples are\ngenerated by the inner maximization process, their associated loss decreases\ninstead, which we named abnormal adversarial examples (AAEs). Upon further\nanalysis, we discover a close relationship between AAEs and classifier\ndistortion, as both the number and outputs of AAEs undergo a significant\nvariation with the onset of CO. Given this observation, we re-examine the SSAT\nprocess and uncover that before the occurrence of CO, the classifier already\ndisplayed a slight distortion, indicated by the presence of few AAEs.\nFurthermore, the classifier directly optimizing these AAEs will accelerate its\ndistortion, and correspondingly, the variation of AAEs will sharply increase as\na result. In such a vicious circle, the classifier rapidly becomes highly\ndistorted and manifests as CO within a few iterations. These observations\nmotivate us to eliminate CO by hindering the generation of AAEs. Specifically,\nwe design a novel method, termed Abnormal Adversarial Examples Regularization\n(AAER), which explicitly regularizes the variation of AAEs to hinder the\nclassifier from becoming distorted. Extensive experiments demonstrate that our\nmethod can effectively eliminate CO and further boost adversarial robustness\nwith negligible additional computational overhead.\n", "date": "2024-04-11", "categories": []}
{"id": "2405.16262", "title": "Layer-Aware Analysis of Catastrophic Overfitting: Revealing the\n  Pseudo-Robust Shortcut Dependency", "abstract": "  Catastrophic overfitting (CO) presents a significant challenge in single-step\nadversarial training (AT), manifesting as highly distorted deep neural networks\n(DNNs) that are vulnerable to multi-step adversarial attacks. However, the\nunderlying factors that lead to the distortion of decision boundaries remain\nunclear. In this work, we delve into the specific changes within different DNN\nlayers and discover that during CO, the former layers are more susceptible,\nexperiencing earlier and greater distortion, while the latter layers show\nrelative insensitivity. Our analysis further reveals that this increased\nsensitivity in former layers stems from the formation of pseudo-robust\nshortcuts, which alone can impeccably defend against single-step adversarial\nattacks but bypass genuine-robust learning, resulting in distorted decision\nboundaries. Eliminating these shortcuts can partially restore robustness in\nDNNs from the CO state, thereby verifying that dependence on them triggers the\noccurrence of CO. This understanding motivates us to implement adaptive weight\nperturbations across different layers to hinder the generation of pseudo-robust\nshortcuts, consequently mitigating CO. Extensive experiments demonstrate that\nour proposed method, Layer-Aware Adversarial Weight Perturbation (LAP), can\neffectively prevent CO and further enhance robustness.\n", "date": "2024-05-25", "categories": []}
{"id": "2409.09406", "title": "Real-world Adversarial Defense against Patch Attacks based on Diffusion\n  Model", "abstract": "  Adversarial patches present significant challenges to the robustness of deep\nlearning models, making the development of effective defenses become critical\nfor real-world applications. This paper introduces DIFFender, a novel\nDIFfusion-based DeFender framework that leverages the power of a text-guided\ndiffusion model to counter adversarial patch attacks. At the core of our\napproach is the discovery of the Adversarial Anomaly Perception (AAP)\nphenomenon, which enables the diffusion model to accurately detect and locate\nadversarial patches by analyzing distributional anomalies. DIFFender seamlessly\nintegrates the tasks of patch localization and restoration within a unified\ndiffusion model framework, enhancing defense efficacy through their close\ninteraction. Additionally, DIFFender employs an efficient few-shot\nprompt-tuning algorithm, facilitating the adaptation of the pre-trained\ndiffusion model to defense tasks without the need for extensive retraining. Our\ncomprehensive evaluation, covering image classification and face recognition\ntasks, as well as real-world scenarios, demonstrates DIFFender's robust\nperformance against adversarial attacks. The framework's versatility and\ngeneralizability across various settings, classifiers, and attack methodologies\nmark a significant advancement in adversarial patch defense strategies. Except\nfor the popular visible domain, we have identified another advantage of\nDIFFender: its capability to easily expand into the infrared domain.\nConsequently, we demonstrate the good flexibility of DIFFender, which can\ndefend against both infrared and visible adversarial patch attacks\nalternatively using a universal defense framework.\n", "date": "2024-09-14", "categories": []}
{"id": "2409.09860", "title": "Revisiting Physical-World Adversarial Attack on Traffic Sign\n  Recognition: A Commercial Systems Perspective", "abstract": "  Traffic Sign Recognition (TSR) is crucial for safe and correct driving\nautomation. Recent works revealed a general vulnerability of TSR models to\nphysical-world adversarial attacks, which can be low-cost, highly deployable,\nand capable of causing severe attack effects such as hiding a critical traffic\nsign or spoofing a fake one. However, so far existing works generally only\nconsidered evaluating the attack effects on academic TSR models, leaving the\nimpacts of such attacks on real-world commercial TSR systems largely unclear.\nIn this paper, we conduct the first large-scale measurement of physical-world\nadversarial attacks against commercial TSR systems. Our testing results reveal\nthat it is possible for existing attack works from academia to have highly\nreliable (100\\%) attack success against certain commercial TSR system\nfunctionality, but such attack capabilities are not generalizable, leading to\nmuch lower-than-expected attack success rates overall. We find that one\npotential major factor is a spatial memorization design that commonly exists in\ntoday's commercial TSR systems. We design new attack success metrics that can\nmathematically model the impacts of such design on the TSR system-level attack\nsuccess, and use them to revisit existing attacks. Through these efforts, we\nuncover 7 novel observations, some of which directly challenge the observations\nor claims in prior works due to the introduction of the new metrics.\n", "date": "2024-09-15", "categories": []}
{"id": "2405.02764", "title": "Assessing Adversarial Robustness of Large Language Models: An Empirical\n  Study", "abstract": "  Large Language Models (LLMs) have revolutionized natural language processing,\nbut their robustness against adversarial attacks remains a critical concern. We\npresents a novel white-box style attack approach that exposes vulnerabilities\nin leading open-source LLMs, including Llama, OPT, and T5. We assess the impact\nof model size, structure, and fine-tuning strategies on their resistance to\nadversarial perturbations. Our comprehensive evaluation across five diverse\ntext classification tasks establishes a new benchmark for LLM robustness. The\nfindings of this study have far-reaching implications for the reliable\ndeployment of LLMs in real-world applications and contribute to the advancement\nof trustworthy AI systems.\n", "date": "2024-05-04", "categories": []}
{"id": "2407.15861", "title": "Adversarial Attacks and Defenses on Text-to-Image Diffusion Models: A\n  Survey", "abstract": "  Recently, the text-to-image diffusion model has gained considerable attention\nfrom the community due to its exceptional image generation capability. A\nrepresentative model, Stable Diffusion, amassed more than 10 million users\nwithin just two months of its release. This surge in popularity has facilitated\nstudies on the robustness and safety of the model, leading to the proposal of\nvarious adversarial attack methods. Simultaneously, there has been a marked\nincrease in research focused on defense methods to improve the robustness and\nsafety of these models. In this survey, we provide a comprehensive review of\nthe literature on adversarial attacks and defenses targeting text-to-image\ndiffusion models. We begin with an overview of text-to-image diffusion models,\nfollowed by an introduction to a taxonomy of adversarial attacks and an\nin-depth review of existing attack methods. We then present a detailed analysis\nof current defense methods that improve model robustness and safety. Finally,\nwe discuss ongoing challenges and explore promising future research directions.\nFor a complete list of the adversarial attack and defense methods covered in\nthis survey, please refer to our curated repository at\nhttps://github.com/datar001/Awesome-AD-on-T2IDM.\n", "date": "2024-07-10", "categories": []}
{"id": "2409.08509", "title": "Exploiting Supervised Poison Vulnerability to Strengthen Self-Supervised\n  Defense", "abstract": "  Availability poisons exploit supervised learning (SL) algorithms by\nintroducing class-related shortcut features in images such that models trained\non poisoned data are useless for real-world datasets. Self-supervised learning\n(SSL), which utilizes augmentations to learn instance discrimination, is\nregarded as a strong defense against poisoned data. However, by extending the\nstudy of SSL across multiple poisons on the CIFAR-10 and ImageNet-100 datasets,\nwe demonstrate that it often performs poorly, far below that of training on\nclean data. Leveraging the vulnerability of SL to poison attacks, we introduce\nadversarial training (AT) on SL to obfuscate poison features and guide robust\nfeature learning for SSL. Our proposed defense, designated VESPR (Vulnerability\nExploitation of Supervised Poisoning for Robust SSL), surpasses the performance\nof six previous defenses across seven popular availability poisons. VESPR\ndisplays superior performance over all previous defenses, boosting the minimum\nand average ImageNet-100 test accuracies of poisoned models by 16% and 9%,\nrespectively. Through analysis and ablation studies, we elucidate the\nmechanisms by which VESPR learns robust class features.\n", "date": "2024-09-13", "categories": []}
{"id": "2409.08919", "title": "XSub: Explanation-Driven Adversarial Attack against Blackbox Classifiers\n  via Feature Substitution", "abstract": "  Despite its significant benefits in enhancing the transparency and\ntrustworthiness of artificial intelligence (AI) systems, explainable AI (XAI)\nhas yet to reach its full potential in real-world applications. One key\nchallenge is that XAI can unintentionally provide adversaries with insights\ninto black-box models, inevitably increasing their vulnerability to various\nattacks. In this paper, we develop a novel explanation-driven adversarial\nattack against black-box classifiers based on feature substitution, called\nXSub. The key idea of XSub is to strategically replace important features\n(identified via XAI) in the original sample with corresponding important\nfeatures from a \"golden sample\" of a different label, thereby increasing the\nlikelihood of the model misclassifying the perturbed sample. The degree of\nfeature substitution is adjustable, allowing us to control how much of the\noriginal samples information is replaced. This flexibility effectively balances\na trade-off between the attacks effectiveness and its stealthiness. XSub is\nalso highly cost-effective in that the number of required queries to the\nprediction model and the explanation model in conducting the attack is in O(1).\nIn addition, XSub can be easily extended to launch backdoor attacks in case the\nattacker has access to the models training data. Our evaluation demonstrates\nthat XSub is not only effective and stealthy but also cost-effective, enabling\nits application across a wide range of AI models.\n", "date": "2024-09-13", "categories": []}
{"id": "2409.08985", "title": "Clean Label Attacks against SLU Systems", "abstract": "  Poisoning backdoor attacks involve an adversary manipulating the training\ndata to induce certain behaviors in the victim model by inserting a trigger in\nthe signal at inference time. We adapted clean label backdoor (CLBD)-data\npoisoning attacks, which do not modify the training labels, on state-of-the-art\nspeech recognition models that support/perform a Spoken Language Understanding\ntask, achieving 99.8% attack success rate by poisoning 10% of the training\ndata. We analyzed how varying the signal-strength of the poison, percent of\nsamples poisoned, and choice of trigger impact the attack. We also found that\nCLBD attacks are most successful when applied to training samples that are\ninherently hard for a proxy model. Using this strategy, we achieved an attack\nsuccess rate of 99.3% by poisoning a meager 1.5% of the training data. Finally,\nwe applied two previously developed defenses against gradient-based attacks,\nand found that they attain mixed success against poisoning.\n", "date": "2024-09-13", "categories": []}
{"id": "2309.03665", "title": "How adversarial attacks can disrupt seemingly stable accurate\n  classifiers", "abstract": "  Adversarial attacks dramatically change the output of an otherwise accurate\nlearning system using a seemingly inconsequential modification to a piece of\ninput data. Paradoxically, empirical evidence indicates that even systems which\nare robust to large random perturbations of the input data remain susceptible\nto small, easily constructed, adversarial perturbations of their inputs. Here,\nwe show that this may be seen as a fundamental feature of classifiers working\nwith high dimensional input data. We introduce a simple generic and\ngeneralisable framework for which key behaviours observed in practical systems\narise with high probability -- notably the simultaneous susceptibility of the\n(otherwise accurate) model to easily constructed adversarial attacks, and\nrobustness to random perturbations of the input data. We confirm that the same\nphenomena are directly observed in practical neural networks trained on\nstandard image classification problems, where even large additive random noise\nfails to trigger the adversarial instability of the network. A surprising\ntakeaway is that even small margins separating a classifier's decision surface\nfrom training and testing data can hide adversarial susceptibility from being\ndetected using randomly sampled perturbations. Counterintuitively, using\nadditive noise during training or testing is therefore inefficient for\neradicating or detecting adversarial examples, and more demanding adversarial\ntraining is required.\n", "date": "2023-09-07", "categories": []}
{"id": "2311.06942", "title": "Resilient Graph Neural Networks: A Coupled Dynamical Systems Approach", "abstract": "  Graph Neural Networks (GNNs) have established themselves as a key component\nin addressing diverse graph-based tasks. Despite their notable successes, GNNs\nremain susceptible to input perturbations in the form of adversarial attacks.\nThis paper introduces an innovative approach to fortify GNNs against\nadversarial perturbations through the lens of coupled dynamical systems. Our\nmethod introduces graph neural layers based on differential equations with\ncontractive properties, which, as we show, improve the robustness of GNNs. A\ndistinctive feature of the proposed approach is the simultaneous learned\nevolution of both the node features and the adjacency matrix, yielding an\nintrinsic enhancement of model robustness to perturbations in the input\nfeatures and the connectivity of the graph. We mathematically derive the\nunderpinnings of our novel architecture and provide theoretical insights to\nreason about its expected behavior. We demonstrate the efficacy of our method\nthrough numerous real-world benchmarks, reading on par or improved performance\ncompared to existing methods.\n", "date": "2023-11-12", "categories": []}
{"id": "2401.13751", "title": "A Training Rate and Survival Heuristic for Inference and Robustness\n  Evaluation (TRASHFIRE)", "abstract": "  Machine learning models -- deep neural networks in particular -- have\nperformed remarkably well on benchmark datasets across a wide variety of\ndomains. However, the ease of finding adversarial counter-examples remains a\npersistent problem when training times are measured in hours or days and the\ntime needed to find a successful adversarial counter-example is measured in\nseconds. Much work has gone into generating and defending against these\nadversarial counter-examples, however the relative costs of attacks and\ndefences are rarely discussed. Additionally, machine learning research is\nalmost entirely guided by test/train metrics, but these would require billions\nof samples to meet industry standards. The present work addresses the problem\nof understanding and predicting how particular model hyper-parameters influence\nthe performance of a model in the presence of an adversary. The proposed\napproach uses survival models, worst-case examples, and a cost-aware analysis\nto precisely and accurately reject a particular model change during routine\nmodel training procedures rather than relying on real-world deployment,\nexpensive formal verification methods, or accurate simulations of very\ncomplicated systems (\\textit{e.g.}, digitally recreating every part of a car or\na plane). Through an evaluation of many pre-processing techniques, adversarial\ncounter-examples, and neural network configurations, the conclusion is that\ndeeper models do offer marginal gains in survival times compared to more\nshallow counterparts. However, we show that those gains are driven more by the\nmodel inference time than inherent robustness properties. Using the proposed\nmethodology, we show that ResNet is hopelessly insecure against even the\nsimplest of white box attacks.\n", "date": "2024-01-24", "categories": []}
{"id": "2405.15655", "title": "HiddenSpeaker: Generate Imperceptible Unlearnable Audios for Speaker\n  Verification System", "abstract": "  In recent years, the remarkable advancements in deep neural networks have\nbrought tremendous convenience. However, the training process of a highly\neffective model necessitates a substantial quantity of samples, which brings\nhuge potential threats, like unauthorized exploitation with privacy leakage. In\nresponse, we propose a framework named HiddenSpeaker, embedding imperceptible\nperturbations within the training speech samples and rendering them unlearnable\nfor deep-learning-based speaker verification systems that employ large-scale\nspeakers for efficient training. The HiddenSpeaker utilizes a simplified\nerror-minimizing method named Single-Level Error-Minimizing (SLEM) to generate\nspecific and effective perturbations. Additionally, a hybrid objective function\nis employed for human perceptual optimization, ensuring the perturbation is\nindistinguishable from human listeners. We conduct extensive experiments on\nmultiple state-of-the-art (SOTA) models in the speaker verification domain to\nevaluate HiddenSpeaker. Our results demonstrate that HiddenSpeaker not only\ndeceives the model with unlearnable samples but also enhances the\nimperceptibility of the perturbations, showcasing strong transferability across\ndifferent models.\n", "date": "2024-05-24", "categories": []}
{"id": "2407.02403", "title": "Face Reconstruction Transfer Attack as Out-of-Distribution\n  Generalization", "abstract": "  Understanding the vulnerability of face recognition systems to malicious\nattacks is of critical importance. Previous works have focused on\nreconstructing face images that can penetrate a targeted verification system.\nEven in the white-box scenario, however, naively reconstructed images\nmisrepresent the identity information, hence the attacks are easily neutralized\nonce the face system is updated or changed. In this paper, we aim to\nreconstruct face images which are capable of transferring face attacks on\nunseen encoders. We term this problem as Face Reconstruction Transfer Attack\n(FRTA) and show that it can be formulated as an out-of-distribution (OOD)\ngeneralization problem. Inspired by its OOD nature, we propose to solve FRTA by\nAveraged Latent Search and Unsupervised Validation with pseudo target (ALSUV).\nTo strengthen the reconstruction attack on OOD unseen encoders, ALSUV\nreconstructs the face by searching the latent of amortized generator StyleGAN2\nthrough multiple latent optimization, latent optimization trajectory averaging,\nand unsupervised validation with a pseudo target. We demonstrate the efficacy\nand generalization of our method on widely used face datasets, accompanying it\nwith extensive ablation studies and visually, qualitatively, and quantitatively\nanalyses. The source code will be released.\n", "date": "2024-07-02", "categories": []}
{"id": "2409.08255", "title": "LoRID: Low-Rank Iterative Diffusion for Adversarial Purification", "abstract": "  This work presents an information-theoretic examination of diffusion-based\npurification methods, the state-of-the-art adversarial defenses that utilize\ndiffusion models to remove malicious perturbations in adversarial examples. By\ntheoretically characterizing the inherent purification errors associated with\nthe Markov-based diffusion purifications, we introduce LoRID, a novel Low-Rank\nIterative Diffusion purification method designed to remove adversarial\nperturbation with low intrinsic purification errors. LoRID centers around a\nmulti-stage purification process that leverages multiple rounds of\ndiffusion-denoising loops at the early time-steps of the diffusion models, and\nthe integration of Tucker decomposition, an extension of matrix factorization,\nto remove adversarial noise at high-noise regimes. Consequently, LoRID\nincreases the effective diffusion time-steps and overcomes strong adversarial\nattacks, achieving superior robustness performance in CIFAR-10/100, CelebA-HQ,\nand ImageNet datasets under both white-box and black-box settings.\n", "date": "2024-09-12", "categories": []}
{"id": "2311.15994", "title": "Adversarial Doodles: Interpretable and Human-drawable Attacks Provide\n  Describable Insights", "abstract": "  DNN-based image classifiers are susceptible to adversarial attacks. Most\nprevious adversarial attacks do not have clear patterns, making it difficult to\ninterpret attacks' results and gain insights into classifiers' mechanisms.\nTherefore, we propose Adversarial Doodles, which have interpretable shapes. We\noptimize black bezier curves to fool the classifier by overlaying them onto the\ninput image. By introducing random affine transformation and regularizing the\ndoodled area, we obtain small-sized attacks that cause misclassification even\nwhen humans replicate them by hand. Adversarial doodles provide describable\ninsights into the relationship between the human-drawn doodle's shape and the\nclassifier's output, such as \"When we add three small circles on a helicopter\nimage, the ResNet-50 classifier mistakenly classifies it as an airplane.\"\n", "date": "2023-11-27", "categories": []}
{"id": "2406.11522", "title": "FullCert: Deterministic End-to-End Certification for Training and\n  Inference of Neural Networks", "abstract": "  Modern machine learning models are sensitive to the manipulation of both the\ntraining data (poisoning attacks) and inference data (adversarial examples).\nRecognizing this issue, the community has developed many empirical defenses\nagainst both attacks and, more recently, certification methods with provable\nguarantees against inference-time attacks. However, such guarantees are still\nlargely lacking for training-time attacks. In this work, we present FullCert,\nthe first end-to-end certifier with sound, deterministic bounds, which proves\nrobustness against both training-time and inference-time attacks. We first\nbound all possible perturbations an adversary can make to the training data\nunder the considered threat model. Using these constraints, we bound the\nperturbations' influence on the model's parameters. Finally, we bound the\nimpact of these parameter changes on the model's prediction, resulting in joint\nrobustness guarantees against poisoning and adversarial examples. To facilitate\nthis novel certification paradigm, we combine our theoretical work with a new\nopen-source library BoundFlow, which enables model training on bounded\ndatasets. We experimentally demonstrate FullCert's feasibility on two datasets.\n", "date": "2024-06-17", "categories": []}
{"id": "2409.03131", "title": "Well, that escalated quickly: The Single-Turn Crescendo Attack (STCA)", "abstract": "  This paper introduces a new method for adversarial attacks on large language\nmodels (LLMs) called the Single-Turn Crescendo Attack (STCA). Building on the\nmulti-turn crescendo attack method introduced by Russinovich, Salem, and Eldan\n(2024), which gradually escalates the context to provoke harmful responses, the\nSTCA achieves similar outcomes in a single interaction. By condensing the\nescalation into a single, well-crafted prompt, the STCA bypasses typical\nmoderation filters that LLMs use to prevent inappropriate outputs. This\ntechnique reveals vulnerabilities in current LLMs and emphasizes the importance\nof stronger safeguards in responsible AI (RAI). The STCA offers a novel method\nthat has not been previously explored.\n", "date": "2024-09-04", "categories": []}
{"id": "2409.06726", "title": "Feedback-based Modal Mutual Search for Attacking Vision-Language\n  Pre-training Models", "abstract": "  Although vision-language pre-training (VLP) models have achieved remarkable\nprogress on cross-modal tasks, they remain vulnerable to adversarial attacks.\nUsing data augmentation and cross-modal interactions to generate transferable\nadversarial examples on surrogate models, transfer-based black-box attacks have\nbecome the mainstream methods in attacking VLP models, as they are more\npractical in real-world scenarios. However, their transferability may be\nlimited due to the differences on feature representation across different\nmodels. To this end, we propose a new attack paradigm called Feedback-based\nModal Mutual Search (FMMS). FMMS introduces a novel modal mutual loss (MML),\naiming to push away the matched image-text pairs while randomly drawing\nmismatched pairs closer in feature space, guiding the update directions of the\nadversarial examples. Additionally, FMMS leverages the target model feedback to\niteratively refine adversarial examples, driving them into the adversarial\nregion. To our knowledge, this is the first work to exploit target model\nfeedback to explore multi-modality adversarial boundaries. Extensive empirical\nevaluations on Flickr30K and MSCOCO datasets for image-text matching tasks show\nthat FMMS significantly outperforms the state-of-the-art baselines.\n", "date": "2024-08-27", "categories": []}
{"id": "2409.07353", "title": "Securing Vision-Language Models with a Robust Encoder Against Jailbreak\n  and Adversarial Attacks", "abstract": "  Large Vision-Language Models (LVLMs), trained on multimodal big datasets,\nhave significantly advanced AI by excelling in vision-language tasks. However,\nthese models remain vulnerable to adversarial attacks, particularly jailbreak\nattacks, which bypass safety protocols and cause the model to generate\nmisleading or harmful responses. This vulnerability stems from both the\ninherent susceptibilities of LLMs and the expanded attack surface introduced by\nthe visual modality. We propose Sim-CLIP+, a novel defense mechanism that\nadversarially fine-tunes the CLIP vision encoder by leveraging a Siamese\narchitecture. This approach maximizes cosine similarity between perturbed and\nclean samples, facilitating resilience against adversarial manipulations.\nSim-CLIP+ offers a plug-and-play solution, allowing seamless integration into\nexisting LVLM architectures as a robust vision encoder. Unlike previous\ndefenses, our method requires no structural modifications to the LVLM and\nincurs minimal computational overhead. Sim-CLIP+ demonstrates effectiveness\nagainst both gradient-based adversarial attacks and various jailbreak\ntechniques. We evaluate Sim-CLIP+ against three distinct jailbreak attack\nstrategies and perform clean evaluations using standard downstream datasets,\nincluding COCO for image captioning and OKVQA for visual question answering.\nExtensive experiments demonstrate that Sim-CLIP+ maintains high clean accuracy\nwhile substantially improving robustness against both gradient-based\nadversarial attacks and jailbreak techniques. Our code and robust vision\nencoders are available at\nhttps://github.com/speedlab-git/Robust-Encoder-against-Jailbreak-attack.git.\n", "date": "2024-09-11", "categories": []}
{"id": "2409.07390", "title": "D-CAPTCHA++: A Study of Resilience of Deepfake CAPTCHA under\n  Transferable Imperceptible Adversarial Attack", "abstract": "  The advancements in generative AI have enabled the improvement of audio\nsynthesis models, including text-to-speech and voice conversion. This raises\nconcerns about its potential misuse in social manipulation and political\ninterference, as synthetic speech has become indistinguishable from natural\nhuman speech. Several speech-generation programs are utilized for malicious\npurposes, especially impersonating individuals through phone calls. Therefore,\ndetecting fake audio is crucial to maintain social security and safeguard the\nintegrity of information. Recent research has proposed a D-CAPTCHA system based\non the challenge-response protocol to differentiate fake phone calls from real\nones. In this work, we study the resilience of this system and introduce a more\nrobust version, D-CAPTCHA++, to defend against fake calls. Specifically, we\nfirst expose the vulnerability of the D-CAPTCHA system under transferable\nimperceptible adversarial attack. Secondly, we mitigate such vulnerability by\nimproving the robustness of the system by using adversarial training in\nD-CAPTCHA deepfake detectors and task classifiers.\n", "date": "2024-09-11", "categories": []}
{"id": "2407.06315", "title": "Shedding More Light on Robust Classifiers under the lens of Energy-based\n  Models", "abstract": "  By reinterpreting a robust discriminative classifier as Energy-based Model\n(EBM), we offer a new take on the dynamics of adversarial training (AT). Our\nanalysis of the energy landscape during AT reveals that untargeted attacks\ngenerate adversarial images much more in-distribution (lower energy) than the\noriginal data from the point of view of the model. Conversely, we observe the\nopposite for targeted attacks. On the ground of our thorough analysis, we\npresent new theoretical and practical results that show how interpreting AT\nenergy dynamics unlocks a better understanding: (1) AT dynamic is governed by\nthree phases and robust overfitting occurs in the third phase with a drastic\ndivergence between natural and adversarial energies (2) by rewriting the loss\nof TRadeoff-inspired Adversarial DEfense via Surrogate-loss minimization\n(TRADES) in terms of energies, we show that TRADES implicitly alleviates\noverfitting by means of aligning the natural energy with the adversarial one\n(3) we empirically show that all recent state-of-the-art robust classifiers are\nsmoothing the energy landscape and we reconcile a variety of studies about\nunderstanding AT and weighting the loss function under the umbrella of EBMs.\nMotivated by rigorous evidence, we propose Weighted Energy Adversarial Training\n(WEAT), a novel sample weighting scheme that yields robust accuracy matching\nthe state-of-the-art on multiple benchmarks such as CIFAR-10 and SVHN and going\nbeyond in CIFAR-100 and Tiny-ImageNet. We further show that robust classifiers\nvary in the intensity and quality of their generative capabilities, and offer a\nsimple method to push this capability, reaching a remarkable Inception Score\n(IS) and FID using a robust classifier without training for generative\nmodeling. The code to reproduce our results is available at\nhttp://github.com/OmnAI-Lab/Robust-Classifiers-under-the-lens-of-EBM/ .\n", "date": "2024-07-08", "categories": []}
{"id": "2409.06420", "title": "Unrevealed Threats: A Comprehensive Study of the Adversarial Robustness\n  of Underwater Image Enhancement Models", "abstract": "  Learning-based methods for underwater image enhancement (UWIE) have undergone\nextensive exploration. However, learning-based models are usually vulnerable to\nadversarial examples so as the UWIE models. To the best of our knowledge, there\nis no comprehensive study on the adversarial robustness of UWIE models, which\nindicates that UWIE models are potentially under the threat of adversarial\nattacks. In this paper, we propose a general adversarial attack protocol. We\nmake a first attempt to conduct adversarial attacks on five well-designed UWIE\nmodels on three common underwater image benchmark datasets. Considering the\nscattering and absorption of light in the underwater environment, there exists\na strong correlation between color correction and underwater image enhancement.\nOn the basis of that, we also design two effective UWIE-oriented adversarial\nattack methods Pixel Attack and Color Shift Attack targeting different color\nspaces. The results show that five models exhibit varying degrees of\nvulnerability to adversarial attacks and well-designed small perturbations on\ndegraded images are capable of preventing UWIE models from generating enhanced\nresults. Further, we conduct adversarial training on these models and\nsuccessfully mitigated the effectiveness of adversarial attacks. In summary, we\nreveal the adversarial vulnerability of UWIE models and propose a new\nevaluation dimension of UWIE models.\n", "date": "2024-09-10", "categories": []}
{"id": "2304.10985", "title": "INK: Inheritable Natural Backdoor Attack Against Model Distillation", "abstract": "  Deep learning models are vulnerable to backdoor attacks, where attackers\ninject malicious behavior through data poisoning and later exploit triggers to\nmanipulate deployed models. To improve the stealth and effectiveness of\nbackdoors, prior studies have introduced various imperceptible attack methods\ntargeting both defense mechanisms and manual inspection. However, all\npoisoning-based attacks still rely on privileged access to the training\ndataset. Consequently, model distillation using a trusted dataset has emerged\nas an effective defense against these attacks. To bridge this gap, we introduce\nINK, an inheritable natural backdoor attack that targets model distillation.\nThe key insight behind INK is the use of naturally occurring statistical\nfeatures in all datasets, allowing attackers to leverage them as backdoor\ntriggers without direct access to the training data. Specifically, INK employs\nimage variance as a backdoor trigger and enables both clean-image and\nclean-label attacks by manipulating the labels and image variance in an\nunauthenticated dataset. Once the backdoor is embedded, it transfers from the\nteacher model to the student model, even when defenders use a trusted dataset\nfor distillation. Theoretical analysis and experimental results demonstrate the\nrobustness of INK against transformation-based, search-based, and\ndistillation-based defenses. For instance, INK maintains an attack success rate\nof over 98\\% post-distillation, compared to an average success rate of 1.4\\%\nfor existing methods.\n", "date": "2023-04-21", "categories": []}
{"id": "2405.09176", "title": "Cross-Input Certified Training for Universal Perturbations", "abstract": "  Existing work in trustworthy machine learning primarily focuses on\nsingle-input adversarial perturbations. In many real-world attack scenarios,\ninput-agnostic adversarial attacks, e.g. universal adversarial perturbations\n(UAPs), are much more feasible. Current certified training methods train models\nrobust to single-input perturbations but achieve suboptimal clean and UAP\naccuracy, thereby limiting their applicability in practical applications. We\npropose a novel method, CITRUS, for certified training of networks robust\nagainst UAP attackers. We show in an extensive evaluation across different\ndatasets, architectures, and perturbation magnitudes that our method\noutperforms traditional certified training methods on standard accuracy (up to\n10.3\\%) and achieves SOTA performance on the more practical certified UAP\naccuracy metric.\n", "date": "2024-05-15", "categories": []}
{"id": "2408.13985", "title": "TF-Attack: Transferable and Fast Adversarial Attacks on Large Language\n  Models", "abstract": "  With the great advancements in large language models (LLMs), adversarial\nattacks against LLMs have recently attracted increasing attention. We found\nthat pre-existing adversarial attack methodologies exhibit limited\ntransferability and are notably inefficient, particularly when applied to LLMs.\nIn this paper, we analyze the core mechanisms of previous predominant\nadversarial attack methods, revealing that 1) the distributions of importance\nscore differ markedly among victim models, restricting the transferability; 2)\nthe sequential attack processes induces substantial time overheads. Based on\nthe above two insights, we introduce a new scheme, named TF-Attack, for\nTransferable and Fast adversarial attacks on LLMs. TF-Attack employs an\nexternal LLM as a third-party overseer rather than the victim model to identify\ncritical units within sentences. Moreover, TF-Attack introduces the concept of\nImportance Level, which allows for parallel substitutions of attacks. We\nconduct extensive experiments on 6 widely adopted benchmarks, evaluating the\nproposed method through both automatic and human metrics. Results show that our\nmethod consistently surpasses previous methods in transferability and delivers\nsignificant speed improvements, up to 20 times faster than earlier attack\nstrategies.\n", "date": "2024-08-26", "categories": []}
{"id": "2409.04968", "title": "Natias: Neuron Attribution based Transferable Image Adversarial\n  Steganography", "abstract": "  Image steganography is a technique to conceal secret messages within digital\nimages. Steganalysis, on the contrary, aims to detect the presence of secret\nmessages within images. Recently, deep-learning-based steganalysis methods have\nachieved excellent detection performance. As a countermeasure, adversarial\nsteganography has garnered considerable attention due to its ability to\neffectively deceive deep-learning-based steganalysis. However, steganalysts\noften employ unknown steganalytic models for detection. Therefore, the ability\nof adversarial steganography to deceive non-target steganalytic models, known\nas transferability, becomes especially important. Nevertheless, existing\nadversarial steganographic methods do not consider how to enhance\ntransferability. To address this issue, we propose a novel adversarial\nsteganographic scheme named Natias. Specifically, we first attribute the output\nof a steganalytic model to each neuron in the target middle layer to identify\ncritical features. Next, we corrupt these critical features that may be adopted\nby diverse steganalytic models. Consequently, it can promote the\ntransferability of adversarial steganography. Our proposed method can be\nseamlessly integrated with existing adversarial steganography frameworks.\nThorough experimental analyses affirm that our proposed technique possesses\nimproved transferability when contrasted with former approaches, and it attains\nheightened security in retraining scenarios.\n", "date": "2024-09-08", "categories": []}
{"id": "2409.05021", "title": "Vision-fused Attack: Advancing Aggressive and Stealthy Adversarial Text\n  against Neural Machine Translation", "abstract": "  While neural machine translation (NMT) models achieve success in our daily\nlives, they show vulnerability to adversarial attacks. Despite being harmful,\nthese attacks also offer benefits for interpreting and enhancing NMT models,\nthus drawing increased research attention. However, existing studies on\nadversarial attacks are insufficient in both attacking ability and human\nimperceptibility due to their sole focus on the scope of language. This paper\nproposes a novel vision-fused attack (VFA) framework to acquire powerful\nadversarial text, i.e., more aggressive and stealthy. Regarding the attacking\nability, we design the vision-merged solution space enhancement strategy to\nenlarge the limited semantic solution space, which enables us to search for\nadversarial candidates with higher attacking ability. For human\nimperceptibility, we propose the perception-retained adversarial text selection\nstrategy to align the human text-reading mechanism. Thus, the finally selected\nadversarial text could be more deceptive. Extensive experiments on various\nmodels, including large language models (LLMs) like LLaMA and GPT-3.5, strongly\nsupport that VFA outperforms the comparisons by large margins (up to 81%/14%\nimprovements on ASR/SSIM).\n", "date": "2024-09-08", "categories": []}
{"id": "2409.05294", "title": "TERD: A Unified Framework for Safeguarding Diffusion Models Against\n  Backdoors", "abstract": "  Diffusion models have achieved notable success in image generation, but they\nremain highly vulnerable to backdoor attacks, which compromise their integrity\nby producing specific undesirable outputs when presented with a pre-defined\ntrigger. In this paper, we investigate how to protect diffusion models from\nthis dangerous threat. Specifically, we propose TERD, a backdoor defense\nframework that builds unified modeling for current attacks, which enables us to\nderive an accessible reversed loss. A trigger reversion strategy is further\nemployed: an initial approximation of the trigger through noise sampled from a\nprior distribution, followed by refinement through differential multi-step\nsamplers. Additionally, with the reversed trigger, we propose backdoor\ndetection from the noise space, introducing the first backdoor input detection\napproach for diffusion models and a novel model detection algorithm that\ncalculates the KL divergence between reversed and benign distributions.\nExtensive evaluations demonstrate that TERD secures a 100% True Positive Rate\n(TPR) and True Negative Rate (TNR) across datasets of varying resolutions. TERD\nalso demonstrates nice adaptability to other Stochastic Differential Equation\n(SDE)-based models. Our code is available at https://github.com/PKU-ML/TERD.\n", "date": "2024-09-09", "categories": []}
{"id": "2409.05558", "title": "Seeing Through the Mask: Rethinking Adversarial Examples for CAPTCHAs", "abstract": "  Modern CAPTCHAs rely heavily on vision tasks that are supposedly hard for\ncomputers but easy for humans. However, advances in image recognition models\npose a significant threat to such CAPTCHAs. These models can easily be fooled\nby generating some well-hidden \"random\" noise and adding it to the image, or\nhiding objects in the image. However, these methods are model-specific and thus\ncan not aid CAPTCHAs in fooling all models. We show in this work that by\nallowing for more significant changes to the images while preserving the\nsemantic information and keeping it solvable by humans, we can fool many\nstate-of-the-art models. Specifically, we demonstrate that by adding masks of\nvarious intensities the Accuracy @ 1 (Acc@1) drops by more than 50%-points for\nall models, and supposedly robust models such as vision transformers see an\nAcc@1 drop of 80%-points.\n  These masks can therefore effectively fool modern image classifiers, thus\nshowing that machines have not caught up with humans -- yet.\n", "date": "2024-09-09", "categories": []}
{"id": "2304.04343", "title": "Certifiable Black-Box Attacks with Randomized Adversarial Examples:\n  Breaking Defenses with Provable Confidence", "abstract": "  Black-box adversarial attacks have demonstrated strong potential to\ncompromise machine learning models by iteratively querying the target model or\nleveraging transferability from a local surrogate model. Recently, such attacks\ncan be effectively mitigated by state-of-the-art (SOTA) defenses, e.g.,\ndetection via the pattern of sequential queries, or injecting noise into the\nmodel. To our best knowledge, we take the first step to study a new paradigm of\nblack-box attacks with provable guarantees -- certifiable black-box attacks\nthat can guarantee the attack success probability (ASP) of adversarial examples\nbefore querying over the target model. This new black-box attack unveils\nsignificant vulnerabilities of machine learning models, compared to traditional\nempirical black-box attacks, e.g., breaking strong SOTA defenses with provable\nconfidence, constructing a space of (infinite) adversarial examples with high\nASP, and the ASP of the generated adversarial examples is theoretically\nguaranteed without verification/queries over the target model. Specifically, we\nestablish a novel theoretical foundation for ensuring the ASP of the black-box\nattack with randomized adversarial examples (AEs). Then, we propose several\nnovel techniques to craft the randomized AEs while reducing the perturbation\nsize for better imperceptibility. Finally, we have comprehensively evaluated\nthe certifiable black-box attacks on the CIFAR10/100, ImageNet, and LibriSpeech\ndatasets, while benchmarking with 16 SOTA black-box attacks, against various\nSOTA defenses in the domains of computer vision and speech recognition. Both\ntheoretical and experimental results have validated the significance of the\nproposed attack. The code and all the benchmarks are available at\n\\url{https://github.com/datasec-lab/CertifiedAttack}.\n", "date": "2023-04-10", "categories": []}
{"id": "2409.04133", "title": "Secure Traffic Sign Recognition: An Attention-Enabled Universal Image\n  Inpainting Mechanism against Light Patch Attacks", "abstract": "  Traffic sign recognition systems play a crucial role in assisting drivers to\nmake informed decisions while driving. However, due to the heavy reliance on\ndeep learning technologies, particularly for future connected and autonomous\ndriving, these systems are susceptible to adversarial attacks that pose\nsignificant safety risks to both personal and public transportation. Notably,\nresearchers recently identified a new attack vector to deceive sign recognition\nsystems: projecting well-designed adversarial light patches onto traffic signs.\nIn comparison with traditional adversarial stickers or graffiti, these emerging\nlight patches exhibit heightened aggression due to their ease of implementation\nand outstanding stealthiness. To effectively counter this security threat, we\npropose a universal image inpainting mechanism, namely, SafeSign. It relies on\nattention-enabled multi-view image fusion to repair traffic signs contaminated\nby adversarial light patches, thereby ensuring the accurate sign recognition.\nHere, we initially explore the fundamental impact of malicious light patches on\nthe local and global feature spaces of authentic traffic signs. Then, we design\na binary mask-based U-Net image generation pipeline outputting diverse\ncontaminated sign patterns, to provide our image inpainting model with needed\ntraining data. Following this, we develop an attention mechanism-enabled neural\nnetwork to jointly utilize the complementary information from multi-view images\nto repair contaminated signs. Finally, extensive experiments are conducted to\nevaluate SafeSign's effectiveness in resisting potential light patch-based\nattacks, bringing an average accuracy improvement of 54.8% in three widely-used\nsign recognition models\n", "date": "2024-09-06", "categories": []}
{"id": "2209.06388", "title": "TSFool: Crafting Highly-Imperceptible Adversarial Time Series through\n  Multi-Objective Attack", "abstract": "  Recent years have witnessed the success of recurrent neural network (RNN)\nmodels in time series classification (TSC). However, neural networks (NNs) are\nvulnerable to adversarial samples, which cause real-life adversarial attacks\nthat undermine the robustness of AI models. To date, most existing attacks\ntarget at feed-forward NNs and image recognition tasks, but they cannot perform\nwell on RNN-based TSC. This is due to the cyclical computation of RNN, which\nprevents direct model differentiation. In addition, the high visual sensitivity\nof time series to perturbations also poses challenges to local objective\noptimization of adversarial samples. In this paper, we propose an efficient\nmethod called TSFool to craft highly-imperceptible adversarial time series for\nRNN-based TSC. The core idea is a new global optimization objective known as\n\"Camouflage Coefficient\" that captures the imperceptibility of adversarial\nsamples from the class distribution. Based on this, we reduce the adversarial\nattack problem to a multi-objective optimization problem that enhances the\nperturbation quality. Furthermore, to speed up the optimization process, we\npropose to use a representation model for RNN to capture deeply embedded\nvulnerable samples whose features deviate from the latent manifold. Experiments\non 11 UCR and UEA datasets showcase that TSFool significantly outperforms six\nwhite-box and three black-box benchmark attacks in terms of effectiveness,\nefficiency and imperceptibility from various perspectives including standard\nmeasure, human study and real-world defense.\n", "date": "2022-09-14", "categories": []}
{"id": "2402.19027", "title": "How to Train your Antivirus: RL-based Hardening through the\n  Problem-Space", "abstract": "  ML-based malware detection on dynamic analysis reports is vulnerable to both\nevasion and spurious correlations. In this work, we investigate a specific ML\narchitecture employed in the pipeline of a widely-known commercial antivirus\ncompany, with the goal to harden it against adversarial malware. Adversarial\ntraining, the sole defensive technique that can confer empirical robustness, is\nnot applicable out of the box in this domain, for the principal reason that\ngradient-based perturbations rarely map back to feasible problem-space\nprograms. We introduce a novel Reinforcement Learning approach for constructing\nadversarial examples, a constituent part of adversarially training a model\nagainst evasion. Our approach comes with multiple advantages. It performs\nmodifications that are feasible in the problem-space, and only those; thus it\ncircumvents the inverse mapping problem. It also makes possible to provide\ntheoretical guarantees on the robustness of the model against a particular set\nof adversarial capabilities. Our empirical exploration validates our\ntheoretical insights, where we can consistently reach 0% Attack Success Rate\nafter a few adversarial retraining iterations.\n", "date": "2024-02-29", "categories": []}
{"id": "2407.08572", "title": "Boosting Adversarial Transferability for Skeleton-based Action\n  Recognition via Exploring the Model Posterior Space", "abstract": "  Skeletal motion plays a pivotal role in human activity recognition (HAR).\nRecently, attack methods have been proposed to identify the universal\nvulnerability of skeleton-based HAR(S-HAR). However, the research of\nadversarial transferability on S-HAR is largely missing. More importantly,\nexisting attacks all struggle in transfer across unknown S-HAR models. We\nobserved that the key reason is that the loss landscape of the action\nrecognizers is rugged and sharp. Given the established correlation in prior\nstudies~\\cite{qin2022boosting,wu2020towards} between loss landscape and\nadversarial transferability, we assume and empirically validate that smoothing\nthe loss landscape could potentially improve adversarial transferability on\nS-HAR. This is achieved by proposing a new post-train Dual Bayesian strategy,\nwhich can effectively explore the model posterior space for a collection of\nsurrogates without the need for re-training. Furthermore, to craft adversarial\nexamples along the motion manifold, we incorporate the attack gradient with\ninformation of the motion dynamics in a Bayesian manner. Evaluated on benchmark\ndatasets, e.g. HDM05 and NTU 60, the average transfer success rate can reach as\nhigh as 35.9\\% and 45.5\\% respectively. In comparison, current state-of-the-art\nskeletal attacks achieve only 3.6\\% and 9.8\\%. The high adversarial\ntransferability remains consistent across various surrogate, victim, and even\ndefense models. Through a comprehensive analysis of the results, we provide\ninsights on what surrogates are more likely to exhibit transferability, to shed\nlight on future research.\n", "date": "2024-07-11", "categories": []}
{"id": "2409.02649", "title": "OpenFact at CheckThat! 2024: Combining Multiple Attack Methods for\n  Effective Adversarial Text Generation", "abstract": "  This paper presents the experiments and results for the CheckThat! Lab at\nCLEF 2024 Task 6: Robustness of Credibility Assessment with Adversarial\nExamples (InCrediblAE). The primary objective of this task was to generate\nadversarial examples in five problem domains in order to evaluate the\nrobustness of widely used text classification methods (fine-tuned BERT, BiLSTM,\nand RoBERTa) when applied to credibility assessment issues.\n  This study explores the application of ensemble learning to enhance\nadversarial attacks on natural language processing (NLP) models. We\nsystematically tested and refined several adversarial attack methods, including\nBERT-Attack, Genetic algorithms, TextFooler, and CLARE, on five datasets across\nvarious misinformation tasks. By developing modified versions of BERT-Attack\nand hybrid methods, we achieved significant improvements in attack\neffectiveness. Our results demonstrate the potential of modification and\ncombining multiple methods to create more sophisticated and effective\nadversarial attack strategies, contributing to the development of more robust\nand secure systems.\n", "date": "2024-09-04", "categories": []}
{"id": "2409.03183", "title": "Bypassing DARCY Defense: Indistinguishable Universal Adversarial\n  Triggers", "abstract": "  Neural networks (NN) classification models for Natural Language Processing\n(NLP) are vulnerable to the Universal Adversarial Triggers (UAT) attack that\ntriggers a model to produce a specific prediction for any input. DARCY borrows\nthe \"honeypot\" concept to bait multiple trapdoors, effectively detecting the\nadversarial examples generated by UAT. Unfortunately, we find a new UAT\ngeneration method, called IndisUAT, which produces triggers (i.e., tokens) and\nuses them to craft adversarial examples whose feature distribution is\nindistinguishable from that of the benign examples in a randomly-chosen\ncategory at the detection layer of DARCY. The produced adversarial examples\nincur the maximal loss of predicting results in the DARCY-protected models.\nMeanwhile, the produced triggers are effective in black-box models for text\ngeneration, text inference, and reading comprehension. Finally, the evaluation\nresults under NN models for NLP tasks indicate that the IndisUAT method can\neffectively circumvent DARCY and penetrate other defenses. For example,\nIndisUAT can reduce the true positive rate of DARCY's detection by at least\n40.8% and 90.6%, and drop the accuracy by at least 33.3% and 51.6% in the RNN\nand CNN models, respectively. IndisUAT reduces the accuracy of the BERT's\nadversarial defense model by at least 34.0%, and makes the GPT-2 language model\nspew racist outputs even when conditioned on non-racial context.\n", "date": "2024-09-05", "categories": []}
{"id": "2409.03458", "title": "Non-Uniform Illumination Attack for Fooling Convolutional Neural\n  Networks", "abstract": "  Convolutional Neural Networks (CNNs) have made remarkable strides; however,\nthey remain susceptible to vulnerabilities, particularly in the face of minor\nimage perturbations that humans can easily recognize. This weakness, often\ntermed as 'attacks', underscores the limited robustness of CNNs and the need\nfor research into fortifying their resistance against such manipulations. This\nstudy introduces a novel Non-Uniform Illumination (NUI) attack technique, where\nimages are subtly altered using varying NUI masks. Extensive experiments are\nconducted on widely-accepted datasets including CIFAR10, TinyImageNet, and\nCalTech256, focusing on image classification with 12 different NUI attack\nmodels. The resilience of VGG, ResNet, MobilenetV3-small and InceptionV3 models\nagainst NUI attacks are evaluated. Our results show a substantial decline in\nthe CNN models' classification accuracy when subjected to NUI attacks,\nindicating their vulnerability under non-uniform illumination. To mitigate\nthis, a defense strategy is proposed, including NUI-attacked images, generated\nthrough the new NUI transformation, into the training set. The results\ndemonstrate a significant enhancement in CNN model performance when confronted\nwith perturbed images affected by NUI attacks. This strategy seeks to bolster\nCNN models' resilience against NUI attacks.\n", "date": "2024-09-05", "categories": []}
{"id": "2409.03598", "title": "A practical approach to evaluating the adversarial distance for machine\n  learning classifiers", "abstract": "  Robustness is critical for machine learning (ML) classifiers to ensure\nconsistent performance in real-world applications where models may encounter\ncorrupted or adversarial inputs. In particular, assessing the robustness of\nclassifiers to adversarial inputs is essential to protect systems from\nvulnerabilities and thus ensure safety in use. However, methods to accurately\ncompute adversarial robustness have been challenging for complex ML models and\nhigh-dimensional data. Furthermore, evaluations typically measure adversarial\naccuracy on specific attack budgets, limiting the informative value of the\nresulting metrics. This paper investigates the estimation of the more\ninformative adversarial distance using iterative adversarial attacks and a\ncertification approach. Combined, the methods provide a comprehensive\nevaluation of adversarial robustness by computing estimates for the upper and\nlower bounds of the adversarial distance. We present visualisations and\nablation studies that provide insights into how this evaluation method should\nbe applied and parameterised. We find that our adversarial attack approach is\neffective compared to related implementations, while the certification method\nfalls short of expectations. The approach in this paper should encourage a more\ninformative way of evaluating the adversarial robustness of ML classifiers.\n", "date": "2024-09-05", "categories": []}
{"id": "2405.04346", "title": "Revisiting Character-level Adversarial Attacks for Language Models", "abstract": "  Adversarial attacks in Natural Language Processing apply perturbations in the\ncharacter or token levels. Token-level attacks, gaining prominence for their\nuse of gradient-based methods, are susceptible to altering sentence semantics,\nleading to invalid adversarial examples. While character-level attacks easily\nmaintain semantics, they have received less attention as they cannot easily\nadopt popular gradient-based methods, and are thought to be easy to defend.\nChallenging these beliefs, we introduce Charmer, an efficient query-based\nadversarial attack capable of achieving high attack success rate (ASR) while\ngenerating highly similar adversarial examples. Our method successfully targets\nboth small (BERT) and large (Llama 2) models. Specifically, on BERT with SST-2,\nCharmer improves the ASR in 4.84% points and the USE similarity in 8% points\nwith respect to the previous art. Our implementation is available in\nhttps://github.com/LIONS-EPFL/Charmer.\n", "date": "2024-05-07", "categories": []}
{"id": "2408.15221", "title": "LLM Defenses Are Not Robust to Multi-Turn Human Jailbreaks Yet", "abstract": "  Recent large language model (LLM) defenses have greatly improved models'\nability to refuse harmful queries, even when adversarially attacked. However,\nLLM defenses are primarily evaluated against automated adversarial attacks in a\nsingle turn of conversation, an insufficient threat model for real-world\nmalicious use. We demonstrate that multi-turn human jailbreaks uncover\nsignificant vulnerabilities, exceeding 70% attack success rate (ASR) on\nHarmBench against defenses that report single-digit ASRs with automated\nsingle-turn attacks. Human jailbreaks also reveal vulnerabilities in machine\nunlearning defenses, successfully recovering dual-use biosecurity knowledge\nfrom unlearned models. We compile these results into Multi-Turn Human\nJailbreaks (MHJ), a dataset of 2,912 prompts across 537 multi-turn jailbreaks.\nWe publicly release MHJ alongside a compendium of jailbreak tactics developed\nacross dozens of commercial red teaming engagements, supporting research\ntowards stronger LLM defenses.\n", "date": "2024-08-27", "categories": []}
{"id": "2409.01185", "title": "Backdoor Defense through Self-Supervised and Generative Learning", "abstract": "  Backdoor attacks change a small portion of training data by introducing\nhand-crafted triggers and rewiring the corresponding labels towards a desired\ntarget class. Training on such data injects a backdoor which causes malicious\ninference in selected test samples. Most defenses mitigate such attacks through\nvarious modifications of the discriminative learning procedure. In contrast,\nthis paper explores an approach based on generative modelling of per-class\ndistributions in a self-supervised representation space. Interestingly, these\nrepresentations get either preserved or heavily disturbed under recent backdoor\nattacks. In both cases, we find that per-class generative models allow to\ndetect poisoned data and cleanse the dataset. Experiments show that training on\ncleansed dataset greatly reduces the attack success rate and retains the\naccuracy on benign inputs.\n", "date": "2024-09-02", "categories": []}
{"id": "2409.01282", "title": "One-Index Vector Quantization Based Adversarial Attack on Image\n  Classification", "abstract": "  To improve storage and transmission, images are generally compressed. Vector\nquantization (VQ) is a popular compression method as it has a high compression\nratio that suppresses other compression techniques. Despite this, existing\nadversarial attack methods on image classification are mostly performed in the\npixel domain with few exceptions in the compressed domain, making them less\napplicable in real-world scenarios. In this paper, we propose a novel one-index\nattack method in the VQ domain to generate adversarial images by a differential\nevolution algorithm, successfully resulting in image misclassification in\nvictim models. The one-index attack method modifies a single index in the\ncompressed data stream so that the decompressed image is misclassified. It only\nneeds to modify a single VQ index to realize an attack, which limits the number\nof perturbed indexes. The proposed method belongs to a semi-black-box attack,\nwhich is more in line with the actual attack scenario. We apply our method to\nattack three popular image classification models, i.e., Resnet, NIN, and VGG16.\nOn average, 55.9% and 77.4% of the images in CIFAR-10 and Fashion MNIST,\nrespectively, are successfully attacked, with a high level of misclassification\nconfidence and a low level of image perturbation.\n", "date": "2024-09-02", "categories": []}
{"id": "2409.02629", "title": "AdvSecureNet: A Python Toolkit for Adversarial Machine Learning", "abstract": "  Machine learning models are vulnerable to adversarial attacks. Several tools\nhave been developed to research these vulnerabilities, but they often lack\ncomprehensive features and flexibility. We introduce AdvSecureNet, a PyTorch\nbased toolkit for adversarial machine learning that is the first to natively\nsupport multi-GPU setups for attacks, defenses, and evaluation. It is the first\ntoolkit that supports both CLI and API interfaces and external YAML\nconfiguration files to enhance versatility and reproducibility. The toolkit\nincludes multiple attacks, defenses and evaluation metrics. Rigiorous software\nengineering practices are followed to ensure high code quality and\nmaintainability. The project is available as an open-source project on GitHub\nat https://github.com/melihcatal/advsecurenet and installable via PyPI.\n", "date": "2024-09-04", "categories": []}
{"id": "2408.16537", "title": "SFR-GNN: Simple and Fast Robust GNNs against Structural Attacks", "abstract": "  Graph Neural Networks (GNNs) have demonstrated commendable performance for\ngraph-structured data. Yet, GNNs are often vulnerable to adversarial structural\nattacks as embedding generation relies on graph topology. Existing efforts are\ndedicated to purifying the maliciously modified structure or applying adaptive\naggregation, thereby enhancing the robustness against adversarial structural\nattacks. It is inevitable for a defender to consume heavy computational costs\ndue to lacking prior knowledge about modified structures. To this end, we\npropose an efficient defense method, called Simple and Fast Robust Graph Neural\nNetwork (SFR-GNN), supported by mutual information theory. The SFR-GNN first\npre-trains a GNN model using node attributes and then fine-tunes it over the\nmodified graph in the manner of contrastive learning, which is free of\npurifying modified structures and adaptive aggregation, thus achieving great\nefficiency gains. Consequently, SFR-GNN exhibits a 24%--162% speedup compared\nto advanced robust models, demonstrating superior robustness for node\nclassification tasks.\n", "date": "2024-08-29", "categories": []}
{"id": "2409.00340", "title": "LightPure: Realtime Adversarial Image Purification for Mobile Devices\n  Using Diffusion Models", "abstract": "  Autonomous mobile systems increasingly rely on deep neural networks for\nperception and decision-making. While effective, these systems are vulnerable\nto adversarial machine learning attacks where minor input perturbations can\nsignificantly impact outcomes. Common countermeasures involve adversarial\ntraining and/or data or network transformation. These methods, though\neffective, require full access to typically proprietary classifiers and are\ncostly for large models. Recent solutions propose purification models, which\nadd a \"purification\" layer before classification, eliminating the need to\nmodify the classifier directly. Despite their effectiveness, these methods are\ncompute-intensive, making them unsuitable for mobile systems where resources\nare limited and low latency is essential.\n  This paper introduces LightPure, a new method that enhances adversarial image\npurification. It improves the accuracy of existing purification methods and\nprovides notable enhancements in speed and computational efficiency, making it\nsuitable for mobile devices with limited resources. Our approach uses a\ntwo-step diffusion and one-shot Generative Adversarial Network (GAN) framework,\nprioritizing latency without compromising robustness. We propose several new\ntechniques to achieve a reasonable balance between classification accuracy and\nadversarial robustness while maintaining desired latency. We design and\nimplement a proof-of-concept on a Jetson Nano board and evaluate our method\nusing various attack scenarios and datasets. Our results show that LightPure\ncan outperform existing methods by up to 10x in terms of latency while\nachieving higher accuracy and robustness for various attack scenarios. This\nmethod offers a scalable and effective solution for real-world mobile systems.\n", "date": "2024-08-31", "categories": []}
{"id": "2409.00418", "title": "Robust off-policy Reinforcement Learning via Soft Constrained Adversary", "abstract": "  Recently, robust reinforcement learning (RL) methods against input\nobservation have garnered significant attention and undergone rapid evolution\ndue to RL's potential vulnerability. Although these advanced methods have\nachieved reasonable success, there have been two limitations when considering\nadversary in terms of long-term horizons. First, the mutual dependency between\nthe policy and its corresponding optimal adversary limits the development of\noff-policy RL algorithms; although obtaining optimal adversary should depend on\nthe current policy, this has restricted applications to off-policy RL. Second,\nthese methods generally assume perturbations based only on the $L_p$-norm, even\nwhen prior knowledge of the perturbation distribution in the environment is\navailable. We here introduce another perspective on adversarial RL: an\nf-divergence constrained problem with the prior knowledge distribution. From\nthis, we derive two typical attacks and their corresponding robust learning\nframeworks. The evaluation of robustness is conducted and the results\ndemonstrate that our proposed methods achieve excellent performance in\nsample-efficient off-policy RL.\n", "date": "2024-08-31", "categories": []}
{"id": "2409.00667", "title": "Comprehensive Botnet Detection by Mitigating Adversarial Attacks,\n  Navigating the Subtleties of Perturbation Distances and Fortifying\n  Predictions with Conformal Layers", "abstract": "  Botnets are computer networks controlled by malicious actors that present\nsignificant cybersecurity challenges. They autonomously infect, propagate, and\ncoordinate to conduct cybercrimes, necessitating robust detection methods. This\nresearch addresses the sophisticated adversarial manipulations posed by\nattackers, aiming to undermine machine learning-based botnet detection systems.\nWe introduce a flow-based detection approach, leveraging machine learning and\ndeep learning algorithms trained on the ISCX and ISOT datasets. The detection\nalgorithms are optimized using the Genetic Algorithm and Particle Swarm\nOptimization to obtain a baseline detection method. The Carlini & Wagner (C&W)\nattack and Generative Adversarial Network (GAN) generate deceptive data with\nsubtle perturbations, targeting each feature used for classification while\npreserving their semantic and syntactic relationships, which ensures that the\nadversarial samples retain meaningfulness and realism. An in-depth analysis of\nthe required L2 distance from the original sample for the malware sample to\nmisclassify is performed across various iteration checkpoints, showing\ndifferent levels of misclassification at different L2 distances of the Pertrub\nsample from the original sample. Our work delves into the vulnerability of\nvarious models, examining the transferability of adversarial examples from a\nNeural Network surrogate model to Tree-based algorithms. Subsequently, models\nthat initially misclassified the perturbed samples are retrained, enhancing\ntheir resilience and detection capabilities. In the final phase, a conformal\nprediction layer is integrated, significantly rejecting incorrect predictions,\nof 58.20 % in the ISCX dataset and 98.94 % in the ISOT dataset.\n", "date": "2024-09-01", "categories": []}
{"id": "2210.00008", "title": "Adversarial Attacks on Transformers-Based Malware Detectors", "abstract": "  Signature-based malware detectors have proven to be insufficient as even a\nsmall change in malignant executable code can bypass these signature-based\ndetectors. Many machine learning-based models have been proposed to efficiently\ndetect a wide variety of malware. Many of these models are found to be\nsusceptible to adversarial attacks - attacks that work by generating\nintentionally designed inputs that can force these models to misclassify. Our\nwork aims to explore vulnerabilities in the current state of the art malware\ndetectors to adversarial attacks. We train a Transformers-based malware\ndetector, carry out adversarial attacks resulting in a misclassification rate\nof 23.9% and propose defenses that reduce this misclassification rate to half.\nAn implementation of our work can be found at\nhttps://github.com/yashjakhotiya/Adversarial-Attacks-On-Transformers.\n", "date": "2022-10-01", "categories": []}
{"id": "2408.02674", "title": "On Feasibility of Intent Obfuscating Attacks", "abstract": "  Intent obfuscation is a common tactic in adversarial situations, enabling the\nattacker to both manipulate the target system and avoid culpability.\nSurprisingly, it has rarely been implemented in adversarial attacks on machine\nlearning systems. We are the first to propose using intent obfuscation to\ngenerate adversarial examples for object detectors: by perturbing another\nnon-overlapping object to disrupt the target object, the attacker hides their\nintended target. We conduct a randomized experiment on 5 prominent detectors --\nYOLOv3, SSD, RetinaNet, Faster R-CNN, and Cascade R-CNN -- using both targeted\nand untargeted attacks and achieve success on all models and attacks. We\nanalyze the success factors characterizing intent obfuscating attacks,\nincluding target object confidence and perturb object sizes. We then\ndemonstrate that the attacker can exploit these success factors to increase\nsuccess rates for all models and attacks. Finally, we discuss main takeaways\nand legal repercussions.\n", "date": "2024-07-22", "categories": []}
{"id": "2408.15702", "title": "Evaluating Model Robustness Using Adaptive Sparse L0 Regularization", "abstract": "  Deep Neural Networks have demonstrated remarkable success in various domains\nbut remain susceptible to adversarial examples, which are slightly altered\ninputs designed to induce misclassification. While adversarial attacks\ntypically optimize under Lp norm constraints, attacks based on the L0 norm,\nprioritising input sparsity, are less studied due to their complex and non\nconvex nature. These sparse adversarial examples challenge existing defenses by\naltering a minimal subset of features, potentially uncovering more subtle DNN\nweaknesses. However, the current L0 norm attack methodologies face a trade off\nbetween accuracy and efficiency either precise but computationally intense or\nexpedient but imprecise. This paper proposes a novel, scalable, and effective\napproach to generate adversarial examples based on the L0 norm, aimed at\nrefining the robustness evaluation of DNNs against such perturbations.\n", "date": "2024-08-28", "categories": []}
{"id": "2408.15721", "title": "Defending Text-to-image Diffusion Models: Surprising Efficacy of Textual\n  Perturbations Against Backdoor Attacks", "abstract": "  Text-to-image diffusion models have been widely adopted in real-world\napplications due to their ability to generate realistic images from textual\ndescriptions. However, recent studies have shown that these methods are\nvulnerable to backdoor attacks. Despite the significant threat posed by\nbackdoor attacks on text-to-image diffusion models, countermeasures remain\nunder-explored. In this paper, we address this research gap by demonstrating\nthat state-of-the-art backdoor attacks against text-to-image diffusion models\ncan be effectively mitigated by a surprisingly simple defense strategy -\ntextual perturbation. Experiments show that textual perturbations are effective\nin defending against state-of-the-art backdoor attacks with minimal sacrifice\nto generation quality. We analyze the efficacy of textual perturbation from two\nangles: text embedding space and cross-attention maps. They further explain how\nbackdoor attacks have compromised text-to-image diffusion models, providing\ninsights for studying future attack and defense strategies. Our code is\navailable at https://github.com/oscarchew/t2i-backdoor-defense.\n", "date": "2024-08-28", "categories": []}
{"id": "2311.12084", "title": "ODDR: Outlier Detection & Dimension Reduction Based Defense Against\n  Adversarial Patches", "abstract": "  Adversarial attacks present a significant challenge to the dependable\ndeployment of machine learning models, with patch-based attacks being\nparticularly potent. These attacks introduce adversarial perturbations in\nlocalized regions of an image, deceiving even well-trained models. In this\npaper, we propose Outlier Detection and Dimension Reduction (ODDR), a\ncomprehensive defense strategy engineered to counteract patch-based adversarial\nattacks through advanced statistical methodologies. Our approach is based on\nthe observation that input features corresponding to adversarial\npatches-whether naturalistic or synthetic-deviate from the intrinsic\ndistribution of the remaining image data and can thus be identified as\noutliers. ODDR operates through a robust three-stage pipeline: Fragmentation,\nSegregation, and Neutralization. This model-agnostic framework is versatile,\noffering protection across various tasks, including image classification,\nobject detection, and depth estimation, and is proved effective in both\nCNN-based and Transformer-based architectures. In the Fragmentation stage,\nimage samples are divided into smaller segments, preparing them for the\nSegregation stage, where advanced outlier detection techniques isolate\nanomalous features linked to adversarial perturbations. The Neutralization\nstage then applies dimension reduction techniques to these outliers,\neffectively neutralizing the adversarial impact while preserving critical\ninformation for the machine learning task. Extensive evaluation on benchmark\ndatasets against state-of-the-art adversarial patches underscores the efficacy\nof ODDR. Our method enhances model accuracy from 39.26% to 79.1% under the\nGoogleAp attack, outperforming leading defenses such as LGS (53.86%), Jujutsu\n(60%), and Jedi (64.34%).\n", "date": "2023-11-20", "categories": []}
{"id": "2404.10789", "title": "PASA: Attack Agnostic Unsupervised Adversarial Detection using\n  Prediction & Attribution Sensitivity Analysis", "abstract": "  Deep neural networks for classification are vulnerable to adversarial\nattacks, where small perturbations to input samples lead to incorrect\npredictions. This susceptibility, combined with the black-box nature of such\nnetworks, limits their adoption in critical applications like autonomous\ndriving. Feature-attribution-based explanation methods provide relevance of\ninput features for model predictions on input samples, thus explaining model\ndecisions. However, we observe that both model predictions and feature\nattributions for input samples are sensitive to noise. We develop a practical\nmethod for this characteristic of model prediction and feature attribution to\ndetect adversarial samples. Our method, PASA, requires the computation of two\ntest statistics using model prediction and feature attribution and can reliably\ndetect adversarial samples using thresholds learned from benign samples. We\nvalidate our lightweight approach by evaluating the performance of PASA on\nvarying strengths of FGSM, PGD, BIM, and CW attacks on multiple image and\nnon-image datasets. On average, we outperform state-of-the-art statistical\nunsupervised adversarial detectors on CIFAR-10 and ImageNet by 14\\% and 35\\%\nROC-AUC scores, respectively. Moreover, our approach demonstrates competitive\nperformance even when an adversary is aware of the defense mechanism.\n", "date": "2024-04-12", "categories": []}
{"id": "2408.14143", "title": "2D-Malafide: Adversarial Attacks Against Face Deepfake Detection Systems", "abstract": "  We introduce 2D-Malafide, a novel and lightweight adversarial attack designed\nto deceive face deepfake detection systems. Building upon the concept of 1D\nconvolutional perturbations explored in the speech domain, our method leverages\n2D convolutional filters to craft perturbations which significantly degrade the\nperformance of state-of-the-art face deepfake detectors. Unlike traditional\nadditive noise approaches, 2D-Malafide optimises a small number of filter\ncoefficients to generate robust adversarial perturbations which are\ntransferable across different face images. Experiments, conducted using the\nFaceForensics++ dataset, demonstrate that 2D-Malafide substantially degrades\ndetection performance in both white-box and black-box settings, with larger\nfilter sizes having the greatest impact. Additionally, we report an\nexplainability analysis using GradCAM which illustrates how 2D-Malafide\nmisleads detection systems by altering the image areas used most for\nclassification. Our findings highlight the vulnerability of current deepfake\ndetection systems to convolutional adversarial attacks as well as the need for\nfuture work to enhance detection robustness through improved image fidelity\nconstraints.\n", "date": "2024-08-26", "categories": []}
{"id": "2408.14879", "title": "Adversarial Manhole: Challenging Monocular Depth Estimation and Semantic\n  Segmentation Models with Patch Attack", "abstract": "  Monocular depth estimation (MDE) and semantic segmentation (SS) are crucial\nfor the navigation and environmental interpretation of many autonomous driving\nsystems. However, their vulnerability to practical adversarial attacks is a\nsignificant concern. This paper presents a novel adversarial attack using\npractical patches that mimic manhole covers to deceive MDE and SS models. The\ngoal is to cause these systems to misinterpret scenes, leading to false\ndetections of near obstacles or non-passable objects. We use Depth Planar\nMapping to precisely position these patches on road surfaces, enhancing the\nattack's effectiveness. Our experiments show that these adversarial patches\ncause a 43% relative error in MDE and achieve a 96% attack success rate in SS.\nThese patches create affected error regions over twice their size in MDE and\napproximately equal to their size in SS. Our studies also confirm the patch's\neffectiveness in physical simulations, the adaptability of the patches across\ndifferent target models, and the effectiveness of our proposed modules,\nhighlighting their practical implications.\n", "date": "2024-08-27", "categories": []}
{"id": "2308.10201", "title": "Hiding Backdoors within Event Sequence Data via Poisoning Attacks", "abstract": "  The financial industry relies on deep learning models for making important\ndecisions. This adoption brings new danger, as deep black-box models are known\nto be vulnerable to adversarial attacks. In computer vision, one can shape the\noutput during inference by performing an adversarial attack called poisoning\nvia introducing a backdoor into the model during training. For sequences of\nfinancial transactions of a customer, insertion of a backdoor is harder to\nperform, as models operate over a more complex discrete space of sequences, and\nsystematic checks for insecurities occur. We provide a method to introduce\nconcealed backdoors, creating vulnerabilities without altering their\nfunctionality for uncontaminated data. To achieve this, we replace a clean\nmodel with a poisoned one that is aware of the availability of a backdoor and\nutilize this knowledge. Our most difficult for uncovering attacks include\neither additional supervised detection step of poisoned data activated during\nthe test or well-hidden model weight modifications. The experimental study\nprovides insights into how these effects vary across different datasets,\narchitectures, and model components. Alternative methods and baselines, such as\ndistillation-type regularization, are also explored but found to be less\nefficient. Conducted on three open transaction datasets and architectures,\nincluding LSTM, CNN, and Transformer, our findings not only illuminate the\nvulnerabilities in contemporary models but also can drive the construction of\nmore robust systems.\n", "date": "2023-08-20", "categories": []}
{"id": "2408.13461", "title": "Probing the Robustness of Vision-Language Pretrained Models: A\n  Multimodal Adversarial Attack Approach", "abstract": "  Vision-language pretraining (VLP) with transformers has demonstrated\nexceptional performance across numerous multimodal tasks. However, the\nadversarial robustness of these models has not been thoroughly investigated.\nExisting multimodal attack methods have largely overlooked cross-modal\ninteractions between visual and textual modalities, particularly in the context\nof cross-attention mechanisms. In this paper, we study the adversarial\nvulnerability of recent VLP transformers and design a novel Joint Multimodal\nTransformer Feature Attack (JMTFA) that concurrently introduces adversarial\nperturbations in both visual and textual modalities under white-box settings.\nJMTFA strategically targets attention relevance scores to disrupt important\nfeatures within each modality, generating adversarial samples by fusing\nperturbations and leading to erroneous model predictions. Experimental results\nindicate that the proposed approach achieves high attack success rates on\nvision-language understanding and reasoning downstream tasks compared to\nexisting baselines. Notably, our findings reveal that the textual modality\nsignificantly influences the complex fusion processes within VLP transformers.\nMoreover, we observe no apparent relationship between model size and\nadversarial robustness under our proposed attacks. These insights emphasize a\nnew dimension of adversarial robustness and underscore potential risks in the\nreliable deployment of multimodal AI systems.\n", "date": "2024-08-24", "categories": []}
{"id": "2307.11672", "title": "Robust Feature Inference: A Test-time Defense Strategy using Spectral\n  Projections", "abstract": "  Test-time defenses are used to improve the robustness of deep neural networks\nto adversarial examples during inference. However, existing methods either\nrequire an additional trained classifier to detect and correct the adversarial\nsamples, or perform additional complex optimization on the model parameters or\nthe input to adapt to the adversarial samples at test-time, resulting in a\nsignificant increase in the inference time compared to the base model. In this\nwork, we propose a novel test-time defense strategy called Robust Feature\nInference (RFI) that is easy to integrate with any existing (robust) training\nprocedure without additional test-time computation. Based on the notion of\nrobustness of features that we present, the key idea is to project the trained\nmodels to the most robust feature space, thereby reducing the vulnerability to\nadversarial attacks in non-robust directions. We theoretically characterize the\nsubspace of the eigenspectrum of the feature covariance that is the most robust\nfor a generalized additive model. Our extensive experiments on CIFAR-10,\nCIFAR-100, tiny ImageNet and ImageNet datasets for several robustness\nbenchmarks, including the state-of-the-art methods in RobustBench show that RFI\nimproves robustness across adaptive and transfer attacks consistently. We also\ncompare RFI with adaptive test-time defenses to demonstrate the effectiveness\nof our proposed approach.\n", "date": "2023-07-21", "categories": []}
{"id": "2401.16352", "title": "Adversarial Training on Purification (AToP): Advancing Both Robustness\n  and Generalization", "abstract": "  The deep neural networks are known to be vulnerable to well-designed\nadversarial attacks. The most successful defense technique based on adversarial\ntraining (AT) can achieve optimal robustness against particular attacks but\ncannot generalize well to unseen attacks. Another effective defense technique\nbased on adversarial purification (AP) can enhance generalization but cannot\nachieve optimal robustness. Meanwhile, both methods share one common limitation\non the degraded standard accuracy. To mitigate these issues, we propose a novel\npipeline to acquire the robust purifier model, named Adversarial Training on\nPurification (AToP), which comprises two components: perturbation destruction\nby random transforms (RT) and purifier model fine-tuned (FT) by adversarial\nloss. RT is essential to avoid overlearning to known attacks, resulting in the\nrobustness generalization to unseen attacks, and FT is essential for the\nimprovement of robustness. To evaluate our method in an efficient and scalable\nway, we conduct extensive experiments on CIFAR-10, CIFAR-100, and ImageNette to\ndemonstrate that our method achieves optimal robustness and exhibits\ngeneralization ability against unseen attacks.\n", "date": "2024-01-29", "categories": []}
{"id": "2408.12670", "title": "Leveraging Information Consistency in Frequency and Spatial Domain for\n  Adversarial Attacks", "abstract": "  Adversarial examples are a key method to exploit deep neural networks. Using\ngradient information, such examples can be generated in an efficient way\nwithout altering the victim model. Recent frequency domain transformation has\nfurther enhanced the transferability of such adversarial examples, such as\nspectrum simulation attack. In this work, we investigate the effectiveness of\nfrequency domain-based attacks, aligning with similar findings in the spatial\ndomain. Furthermore, such consistency between the frequency and spatial domains\nprovides insights into how gradient-based adversarial attacks induce\nperturbations across different domains, which is yet to be explored. Hence, we\npropose a simple, effective, and scalable gradient-based adversarial attack\nalgorithm leveraging the information consistency in both frequency and spatial\ndomains. We evaluate the algorithm for its effectiveness against different\nmodels. Extensive experiments demonstrate that our algorithm achieves\nstate-of-the-art results compared to other gradient-based algorithms. Our code\nis available at: https://github.com/LMBTough/FSA.\n", "date": "2024-08-22", "categories": []}
{"id": "2408.12727", "title": "BankTweak: Adversarial Attack against Multi-Object Trackers by\n  Manipulating Feature Banks", "abstract": "  Multi-object tracking (MOT) aims to construct moving trajectories for\nobjects, and modern multi-object trackers mainly utilize the\ntracking-by-detection methodology. Initial approaches to MOT attacks primarily\naimed to degrade the detection quality of the frames under attack, thereby\nreducing accuracy only in those specific frames, highlighting a lack of\n\\textit{efficiency}. To improve efficiency, recent advancements manipulate\nobject positions to cause persistent identity (ID) switches during the\nassociation phase, even after the attack ends within a few frames. However,\nthese position-manipulating attacks have inherent limitations, as they can be\neasily counteracted by adjusting distance-related parameters in the association\nphase, revealing a lack of \\textit{robustness}. In this paper, we present\n\\textsf{BankTweak}, a novel adversarial attack designed for MOT trackers, which\nfeatures efficiency and robustness. \\textsf{BankTweak} focuses on the feature\nextractor in the association phase and reveals vulnerability in the Hungarian\nmatching method used by feature-based MOT systems. Exploiting the\nvulnerability, \\textsf{BankTweak} induces persistent ID switches (addressing\n\\textit{efficiency}) even after the attack ends by strategically injecting\naltered features into the feature banks without modifying object positions\n(addressing \\textit{robustness}). To demonstrate the applicability, we apply\n\\textsf{BankTweak} to three multi-object trackers (DeepSORT, StrongSORT, and\nMOTDT) with one-stage, two-stage, anchor-free, and transformer detectors.\nExtensive experiments on the MOT17 and MOT20 datasets show that our method\nsubstantially surpasses existing attacks, exposing the vulnerability of the\ntracking-by-detection framework to \\textsf{BankTweak}.\n", "date": "2024-08-22", "categories": []}
{"id": "2408.13102", "title": "Dynamic Label Adversarial Training for Deep Learning Robustness Against\n  Adversarial Attacks", "abstract": "  Adversarial training is one of the most effective methods for enhancing model\nrobustness. Recent approaches incorporate adversarial distillation in\nadversarial training architectures. However, we notice two scenarios of defense\nmethods that limit their performance: (1) Previous methods primarily use static\nground truth for adversarial training, but this often causes robust\noverfitting; (2) The loss functions are either Mean Squared Error or\nKL-divergence leading to a sub-optimal performance on clean accuracy. To solve\nthose problems, we propose a dynamic label adversarial training (DYNAT)\nalgorithm that enables the target model to gradually and dynamically gain\nrobustness from the guide model's decisions. Additionally, we found that a\nbudgeted dimension of inner optimization for the target model may contribute to\nthe trade-off between clean accuracy and robust accuracy. Therefore, we propose\na novel inner optimization method to be incorporated into the adversarial\ntraining. This will enable the target model to adaptively search for\nadversarial examples based on dynamic labels from the guiding model,\ncontributing to the robustness of the target model. Extensive experiments\nvalidate the superior performance of our approach.\n", "date": "2024-08-23", "categories": []}
{"id": "2311.01473", "title": "Adversarial Examples in the Physical World: A Survey", "abstract": "  Deep neural networks (DNNs) have demonstrated high vulnerability to\nadversarial examples, raising broad security concerns about their applications.\nBesides the attacks in the digital world, the practical implications of\nadversarial examples in the physical world present significant challenges and\nsafety concerns. However, current research on physical adversarial examples\n(PAEs) lacks a comprehensive understanding of their unique characteristics,\nleading to limited significance and understanding. In this paper, we address\nthis gap by thoroughly examining the characteristics of PAEs within a practical\nworkflow encompassing training, manufacturing, and re-sampling processes. By\nanalyzing the links between physical adversarial attacks, we identify\nmanufacturing and re-sampling as the primary sources of distinct attributes and\nparticularities in PAEs. Leveraging this knowledge, we develop a comprehensive\nanalysis and classification framework for PAEs based on their specific\ncharacteristics, covering over 100 studies on physical-world adversarial\nexamples. Furthermore, we investigate defense strategies against PAEs and\nidentify open challenges and opportunities for future research. We aim to\nprovide a fresh, thorough, and systematic understanding of PAEs, thereby\npromoting the development of robust adversarial learning and its application in\nopen-world scenarios to provide the community with a continuously updated list\nof physical world adversarial sample resources, including papers, code, \\etc,\nwithin the proposed framework\n", "date": "2023-11-01", "categories": []}
{"id": "2406.03833", "title": "Talos: A More Effective and Efficient Adversarial Defense for GNN Models\n  Based on the Global Homophily of Graphs", "abstract": "  Graph neural network (GNN) models play a pivotal role in numerous tasks\ninvolving graph-related data analysis. Despite their efficacy, similar to other\ndeep learning models, GNNs are susceptible to adversarial attacks. Even minor\nperturbations in graph data can induce substantial alterations in model\npredictions. While existing research has explored various adversarial defense\ntechniques for GNNs, the challenge of defending against adversarial attacks on\nreal-world scale graph data remains largely unresolved. On one hand, methods\nreliant on graph purification and preprocessing tend to excessively emphasize\nlocal graph information, leading to sub-optimal defensive outcomes. On the\nother hand, approaches rooted in graph structure learning entail significant\ntime overheads, rendering them impractical for large-scale graphs. In this\npaper, we propose a new defense method named Talos, which enhances the global,\nrather than local, homophily of graphs as a defense. Experiments show that the\nproposed approach notably outperforms state-of-the-art defense approaches,\nwhile imposing little computational overhead.\n", "date": "2024-06-06", "categories": []}
{"id": "2407.21316", "title": "Diff-Cleanse: Identifying and Mitigating Backdoor Attacks in Diffusion\n  Models", "abstract": "  Diffusion models (DMs) are regarded as one of the most advanced generative\nmodels today, yet recent studies suggest that they are vulnerable to backdoor\nattacks, which establish hidden associations between particular input patterns\nand model behaviors, compromising model integrity by causing undesirable\nactions with manipulated inputs. This vulnerability poses substantial risks,\nincluding reputational damage to model owners and the dissemination of harmful\ncontent. To mitigate the threat of backdoor attacks, there have been some\ninvestigations on backdoor detection and model repair. However, previous work\nfails to reliably purify the models backdoored by state-of-the-art attack\nmethods, rendering the field much underexplored. To bridge this gap, we\nintroduce Diff-Cleanse, a novel two-stage backdoor defense framework\nspecifically designed for DMs. The first stage employs a novel trigger\ninversion technique to reconstruct the trigger and detect the backdoor, and the\nsecond stage utilizes a structural pruning method to eliminate the backdoor. We\nevaluate our framework on hundreds of DMs that are attacked by three existing\nbackdoor attack methods with a wide range of hyperparameter settings. Extensive\nexperiments demonstrate that Diff-Cleanse achieves nearly 100\\% detection\naccuracy and effectively mitigates backdoor impacts, preserving the model's\nbenign performance with minimal compromise. Our code is avaliable at\nhttps://github.com/shymuel/diff-cleanse.\n", "date": "2024-07-31", "categories": []}
{"id": "2408.09672", "title": "Regularization for Adversarial Robust Learning", "abstract": "  Despite the growing prevalence of artificial neural networks in real-world\napplications, their vulnerability to adversarial attacks remains a significant\nconcern, which motivates us to investigate the robustness of machine learning\nmodels. While various heuristics aim to optimize the distributionally robust\nrisk using the $\\infty$-Wasserstein metric, such a notion of robustness\nfrequently encounters computation intractability. To tackle the computational\nchallenge, we develop a novel approach to adversarial training that integrates\n$\\phi$-divergence regularization into the distributionally robust risk\nfunction. This regularization brings a notable improvement in computation\ncompared with the original formulation. We develop stochastic gradient methods\nwith biased oracles to solve this problem efficiently, achieving the\nnear-optimal sample complexity. Moreover, we establish its regularization\neffects and demonstrate it is asymptotic equivalence to a regularized empirical\nrisk minimization framework, by considering various scaling regimes of the\nregularization parameter and robustness level. These regimes yield gradient\nnorm regularization, variance regularization, or a smoothed gradient norm\nregularization that interpolates between these extremes. We numerically\nvalidate our proposed method in supervised learning, reinforcement learning,\nand contextual learning and showcase its state-of-the-art performance against\nvarious adversarial attacks.\n", "date": "2024-08-19", "categories": []}
{"id": "2405.20775", "title": "Medical MLLM is Vulnerable: Cross-Modality Jailbreak and Mismatched\n  Attacks on Medical Multimodal Large Language Models", "abstract": "  Security concerns related to Large Language Models (LLMs) have been\nextensively explored, yet the safety implications for Multimodal Large Language\nModels (MLLMs), particularly in medical contexts (MedMLLMs), remain\ninsufficiently studied. This paper delves into the underexplored security\nvulnerabilities of MedMLLMs, especially when deployed in clinical environments\nwhere the accuracy and relevance of question-and-answer interactions are\ncritically tested against complex medical challenges. By combining existing\nclinical medical data with atypical natural phenomena, we define the mismatched\nmalicious attack (2M-attack) and introduce its optimized version, known as the\noptimized mismatched malicious attack (O2M-attack or 2M-optimization). Using\nthe voluminous 3MAD dataset that we construct, which covers a wide range of\nmedical image modalities and harmful medical scenarios, we conduct a\ncomprehensive analysis and propose the MCM optimization method, which\nsignificantly enhances the attack success rate on MedMLLMs. Evaluations with\nthis dataset and attack methods, including white-box attacks on LLaVA-Med and\ntransfer attacks (black-box) on four other SOTA models, indicate that even\nMedMLLMs designed with enhanced security features remain vulnerable to security\nbreaches. Our work underscores the urgent need for a concerted effort to\nimplement robust security measures and enhance the safety and efficacy of\nopen-source MedMLLMs, particularly given the potential severity of jailbreak\nattacks and other malicious or clinically significant exploits in medical\nsettings. Our code is available at https://github.com/dirtycomputer/O2M_attack.\n", "date": "2024-05-26", "categories": []}
{"id": "2408.11218", "title": "Revisiting Min-Max Optimization Problem in Adversarial Training", "abstract": "  The rise of computer vision applications in the real world puts the security\nof the deep neural networks at risk. Recent works demonstrate that\nconvolutional neural networks are susceptible to adversarial examples - where\nthe input images look similar to the natural images but are classified\nincorrectly by the model. To provide a rebuttal to this problem, we propose a\nnew method to build robust deep neural networks against adversarial attacks by\nreformulating the saddle point optimization problem in \\cite{madry2017towards}.\nOur proposed method offers significant resistance and a concrete security\nguarantee against multiple adversaries. The goal of this paper is to act as a\nstepping stone for a new variation of deep learning models which would lead\ntowards fully robust deep learning models.\n", "date": "2024-08-20", "categories": []}
{"id": "2408.11264", "title": "Correlation Analysis of Adversarial Attack in Time Series Classification", "abstract": "  This study investigates the vulnerability of time series classification\nmodels to adversarial attacks, with a focus on how these models process local\nversus global information under such conditions. By leveraging the Normalized\nAuto Correlation Function (NACF), an exploration into the inclination of neural\nnetworks is conducted. It is demonstrated that regularization techniques,\nparticularly those employing Fast Fourier Transform (FFT) methods and targeting\nfrequency components of perturbations, markedly enhance the effectiveness of\nattacks. Meanwhile, the defense strategies, like noise introduction and\nGaussian filtering, are shown to significantly lower the Attack Success Rate\n(ASR), with approaches based on noise introducing notably effective in\ncountering high-frequency distortions. Furthermore, models designed to\nprioritize global information are revealed to possess greater resistance to\nadversarial manipulations. These results underline the importance of designing\nattack and defense mechanisms, informed by frequency domain analysis, as a\nmeans to considerably reinforce the resilience of neural network models against\nadversarial threats.\n", "date": "2024-08-21", "categories": []}
{"id": "2408.11680", "title": "First line of defense: A robust first layer mitigates adversarial\n  attacks", "abstract": "  Adversarial training (AT) incurs significant computational overhead, leading\nto growing interest in designing inherently robust architectures. We\ndemonstrate that a carefully designed first layer of the neural network can\nserve as an implicit adversarial noise filter (ANF). This filter is created\nusing a combination of large kernel size, increased convolution filters, and a\nmaxpool operation. We show that integrating this filter as the first layer in\narchitectures such as ResNet, VGG, and EfficientNet results in adversarially\nrobust networks. Our approach achieves higher adversarial accuracies than\nexisting natively robust architectures without AT and is competitive with\nadversarial-trained architectures across a wide range of datasets. Supporting\nour findings, we show that (a) the decision regions for our method have better\nmargins, (b) the visualized loss surfaces are smoother, (c) the modified peak\nsignal-to-noise ratio (mPSNR) values at the output of the ANF are higher, (d)\nhigh-frequency components are more attenuated, and (e) architectures\nincorporating ANF exhibit better denoising in Gaussian noise compared to\nbaseline architectures. Code for all our experiments are available at\n\\url{https://github.com/janani-suresh-97/first-line-defence.git}.\n", "date": "2024-08-21", "categories": []}
{"id": "2310.12431", "title": "SAM Meets UAP: Attacking Segment Anything Model With Universal\n  Adversarial Perturbation", "abstract": "  As Segment Anything Model (SAM) becomes a popular foundation model in\ncomputer vision, its adversarial robustness has become a concern that cannot be\nignored. This works investigates whether it is possible to attack SAM with\nimage-agnostic Universal Adversarial Perturbation (UAP). In other words, we\nseek a single perturbation that can fool the SAM to predict invalid masks for\nmost (if not all) images. We demonstrate convetional image-centric attack\nframework is effective for image-independent attacks but fails for universal\nadversarial attack. To this end, we propose a novel perturbation-centric\nframework that results in a UAP generation method based on self-supervised\ncontrastive learning (CL), where the UAP is set to the anchor sample and the\npositive sample is augmented from the UAP. The representations of negative\nsamples are obtained from the image encoder in advance and saved in a memory\nbank. The effectiveness of our proposed CL-based UAP generation method is\nvalidated by both quantitative and qualitative results. On top of the ablation\nstudy to understand various components in our proposed method, we shed light on\nthe roles of positive and negative samples in making the generated UAP\neffective for attacking SAM.\n", "date": "2023-10-19", "categories": []}
{"id": "2408.10647", "title": "Privacy-preserving Universal Adversarial Defense for Black-box Models", "abstract": "  Deep neural networks (DNNs) are increasingly used in critical applications\nsuch as identity authentication and autonomous driving, where robustness\nagainst adversarial attacks is crucial. These attacks can exploit minor\nperturbations to cause significant prediction errors, making it essential to\nenhance the resilience of DNNs. Traditional defense methods often rely on\naccess to detailed model information, which raises privacy concerns, as model\nowners may be reluctant to share such data. In contrast, existing black-box\ndefense methods fail to offer a universal defense against various types of\nadversarial attacks. To address these challenges, we introduce DUCD, a\nuniversal black-box defense method that does not require access to the target\nmodel's parameters or architecture. Our approach involves distilling the target\nmodel by querying it with data, creating a white-box surrogate while preserving\ndata privacy. We further enhance this surrogate model using a certified defense\nbased on randomized smoothing and optimized noise selection, enabling robust\ndefense against a broad range of adversarial attacks. Comparative evaluations\nbetween the certified defenses of the surrogate and target models demonstrate\nthe effectiveness of our approach. Experiments on multiple image classification\ndatasets show that DUCD not only outperforms existing black-box defenses but\nalso matches the accuracy of white-box defenses, all while enhancing data\nprivacy and reducing the success rate of membership inference attacks.\n", "date": "2024-08-20", "categories": []}
{"id": "2408.10694", "title": "MsMemoryGAN: A Multi-scale Memory GAN for Palm-vein Adversarial\n  Purification", "abstract": "  Deep neural networks have recently achieved promising performance in the vein\nrecognition task and have shown an increasing application trend, however, they\nare prone to adversarial perturbation attacks by adding imperceptible\nperturbations to the input, resulting in making incorrect recognition. To\naddress this issue, we propose a novel defense model named MsMemoryGAN, which\naims to filter the perturbations from adversarial samples before recognition.\nFirst, we design a multi-scale autoencoder to achieve high-quality\nreconstruction and two memory modules to learn the detailed patterns of normal\nsamples at different scales. Second, we investigate a learnable metric in the\nmemory module to retrieve the most relevant memory items to reconstruct the\ninput image. Finally, the perceptional loss is combined with the pixel loss to\nfurther enhance the quality of the reconstructed image. During the training\nphase, the MsMemoryGAN learns to reconstruct the input by merely using fewer\nprototypical elements of the normal patterns recorded in the memory. At the\ntesting stage, given an adversarial sample, the MsMemoryGAN retrieves its most\nrelevant normal patterns in memory for the reconstruction. Perturbations in the\nadversarial sample are usually not reconstructed well, resulting in purifying\nthe input from adversarial perturbations. We have conducted extensive\nexperiments on two public vein datasets under different adversarial attack\nmethods to evaluate the performance of the proposed approach. The experimental\nresults show that our approach removes a wide variety of adversarial\nperturbations, allowing vein classifiers to achieve the highest recognition\naccuracy.\n", "date": "2024-08-20", "categories": []}
{"id": "2408.10948", "title": "GAIM: Attacking Graph Neural Networks via Adversarial Influence\n  Maximization", "abstract": "  Recent studies show that well-devised perturbations on graph structures or\nnode features can mislead trained Graph Neural Network (GNN) models. However,\nthese methods often overlook practical assumptions, over-rely on heuristics, or\nseparate vital attack components. In response, we present GAIM, an integrated\nadversarial attack method conducted on a node feature basis while considering\nthe strict black-box setting. Specifically, we define an adversarial influence\nfunction to theoretically assess the adversarial impact of node perturbations,\nthereby reframing the GNN attack problem into the adversarial influence\nmaximization problem. In our approach, we unify the selection of the target\nnode and the construction of feature perturbations into a single optimization\nproblem, ensuring a unique and consistent feature perturbation for each target\nnode. We leverage a surrogate model to transform this problem into a solvable\nlinear programming task, streamlining the optimization process. Moreover, we\nextend our method to accommodate label-oriented attacks, broadening its\napplicability. Thorough evaluations on five benchmark datasets across three\npopular models underscore the effectiveness of our method in both untargeted\nand label-oriented targeted attacks. Through comprehensive analysis and\nablation studies, we demonstrate the practical value and efficacy inherent to\nour design choices.\n", "date": "2024-08-20", "categories": []}
{"id": "2311.11261", "title": "Adversarial Prompt Tuning for Vision-Language Models", "abstract": "  With the rapid advancement of multimodal learning, pre-trained\nVision-Language Models (VLMs) such as CLIP have demonstrated remarkable\ncapacities in bridging the gap between visual and language modalities. However,\nthese models remain vulnerable to adversarial attacks, particularly in the\nimage modality, presenting considerable security risks. This paper introduces\nAdversarial Prompt Tuning (AdvPT), a novel technique to enhance the adversarial\nrobustness of image encoders in VLMs. AdvPT innovatively leverages learnable\ntext prompts and aligns them with adversarial image embeddings, to address the\nvulnerabilities inherent in VLMs without the need for extensive parameter\ntraining or modification of the model architecture. We demonstrate that AdvPT\nimproves resistance against white-box and black-box adversarial attacks and\nexhibits a synergistic effect when combined with existing\nimage-processing-based defense techniques, further boosting defensive\ncapabilities. Comprehensive experimental analyses provide insights into\nadversarial prompt tuning, a novel paradigm devoted to improving resistance to\nadversarial images through textual input modifications, paving the way for\nfuture robust multimodal learning research. These findings open up new\npossibilities for enhancing the security of VLMs. Our code is available at\nhttps://github.com/jiamingzhang94/Adversarial-Prompt-Tuning.\n", "date": "2023-11-19", "categories": []}
{"id": "2311.17607", "title": "Topology-preserving Adversarial Training for Alleviating Natural\n  Accuracy Degradation", "abstract": "  Despite the effectiveness in improving the robustness of neural networks,\nadversarial training has suffered from the natural accuracy degradation\nproblem, i.e., accuracy on natural samples has reduced significantly. In this\nstudy, we reveal that natural accuracy degradation is highly related to the\ndisruption of the natural sample topology in the representation space by\nquantitative and qualitative experiments. Based on this observation, we propose\nTopology-pReserving Adversarial traINing (TRAIN) to alleviate the problem by\npreserving the topology structure of natural samples from a standard model\ntrained only on natural samples during adversarial training. As an additional\nregularization, our method can be combined with various popular adversarial\ntraining algorithms, taking advantage of both sides. Extensive experiments on\nCIFAR-10, CIFAR-100, and Tiny ImageNet show that our proposed method achieves\nconsistent and significant improvements over various strong baselines in most\ncases. Specifically, without additional data, TRAIN achieves up to 8.86%\nimprovement in natural accuracy and 6.33% improvement in robust accuracy.\n", "date": "2023-11-29", "categories": []}
{"id": "2401.08903", "title": "Rethinking Impersonation and Dodging Attacks on Face Recognition Systems", "abstract": "  Face Recognition (FR) systems can be easily deceived by adversarial examples\nthat manipulate benign face images through imperceptible perturbations.\nAdversarial attacks on FR encompass two types: impersonation (targeted) attacks\nand dodging (untargeted) attacks. Previous methods often achieve a successful\nimpersonation attack on FR, however, it does not necessarily guarantee a\nsuccessful dodging attack on FR in the black-box setting. In this paper, our\nkey insight is that the generation of adversarial examples should perform both\nimpersonation and dodging attacks simultaneously. To this end, we propose a\nnovel attack method termed as Adversarial Pruning (Adv-Pruning), to fine-tune\nexisting adversarial examples to enhance their dodging capabilities while\npreserving their impersonation capabilities. Adv-Pruning consists of Priming,\nPruning, and Restoration stages. Concretely, we propose Adversarial Priority\nQuantification to measure the region-wise priority of original adversarial\nperturbations, identifying and releasing those with minimal impact on absolute\nmodel output variances. Then, Biased Gradient Adaptation is presented to adapt\nthe adversarial examples to traverse the decision boundaries of both the\nattacker and victim by adding perturbations favoring dodging attacks on the\nvacated regions, preserving the prioritized features of the original\nperturbations while boosting dodging performance. As a result, we can maintain\nthe impersonation capabilities of original adversarial examples while\neffectively enhancing dodging capabilities. Comprehensive experiments\ndemonstrate the superiority of our method compared with state-of-the-art\nadversarial attack methods.\n", "date": "2024-01-17", "categories": []}
{"id": "2403.02955", "title": "XAI-Based Detection of Adversarial Attacks on Deepfake Detectors", "abstract": "  We introduce a novel methodology for identifying adversarial attacks on\ndeepfake detectors using eXplainable Artificial Intelligence (XAI). In an era\ncharacterized by digital advancement, deepfakes have emerged as a potent tool,\ncreating a demand for efficient detection systems. However, these systems are\nfrequently targeted by adversarial attacks that inhibit their performance. We\naddress this gap, developing a defensible deepfake detector by leveraging the\npower of XAI. The proposed methodology uses XAI to generate interpretability\nmaps for a given method, providing explicit visualizations of decision-making\nfactors within the AI models. We subsequently employ a pretrained feature\nextractor that processes both the input image and its corresponding XAI image.\nThe feature embeddings extracted from this process are then used for training a\nsimple yet effective classifier. Our approach contributes not only to the\ndetection of deepfakes but also enhances the understanding of possible\nadversarial attacks, pinpointing potential vulnerabilities. Furthermore, this\napproach does not change the performance of the deepfake detector. The paper\ndemonstrates promising results suggesting a potential pathway for future\ndeepfake detection mechanisms. We believe this study will serve as a valuable\ncontribution to the community, sparking much-needed discourse on safeguarding\ndeepfake detectors.\n", "date": "2024-03-05", "categories": []}
{"id": "2408.08899", "title": "Kov: Transferable and Naturalistic Black-Box LLM Attacks using Markov\n  Decision Processes and Tree Search", "abstract": "  Eliciting harmful behavior from large language models (LLMs) is an important\ntask to ensure the proper alignment and safety of the models. Often when\ntraining LLMs, ethical guidelines are followed yet alignment failures may still\nbe uncovered through red teaming adversarial attacks. This work frames the\nred-teaming problem as a Markov decision process (MDP) and uses Monte Carlo\ntree search to find harmful behaviors of black-box, closed-source LLMs. We\noptimize token-level prompt suffixes towards targeted harmful behaviors on\nwhite-box LLMs and include a naturalistic loss term, log-perplexity, to\ngenerate more natural language attacks for better interpretability. The\nproposed algorithm, Kov, trains on white-box LLMs to optimize the adversarial\nattacks and periodically evaluates responses from the black-box LLM to guide\nthe search towards more harmful black-box behaviors. In our preliminary study,\nresults indicate that we can jailbreak black-box models, such as GPT-3.5, in\nonly 10 queries, yet fail on GPT-4$-$which may indicate that newer models are\nmore robust to token-level attacks. All work to reproduce these results is open\nsourced (https://github.com/sisl/Kov.jl).\n", "date": "2024-08-11", "categories": []}
{"id": "2408.08989", "title": "Ask, Attend, Attack: A Effective Decision-Based Black-Box Targeted\n  Attack for Image-to-Text Models", "abstract": "  While image-to-text models have demonstrated significant advancements in\nvarious vision-language tasks, they remain susceptible to adversarial attacks.\nExisting white-box attacks on image-to-text models require access to the\narchitecture, gradients, and parameters of the target model, resulting in low\npracticality. Although the recently proposed gray-box attacks have improved\npracticality, they suffer from semantic loss during the training process, which\nlimits their targeted attack performance. To advance adversarial attacks of\nimage-to-text models, this paper focuses on a challenging scenario:\ndecision-based black-box targeted attacks where the attackers only have access\nto the final output text and aim to perform targeted attacks. Specifically, we\nformulate the decision-based black-box targeted attack as a large-scale\noptimization problem. To efficiently solve the optimization problem, a\nthree-stage process \\textit{Ask, Attend, Attack}, called \\textit{AAA}, is\nproposed to coordinate with the solver. \\textit{Ask} guides attackers to create\ntarget texts that satisfy the specific semantics. \\textit{Attend} identifies\nthe crucial regions of the image for attacking, thus reducing the search space\nfor the subsequent \\textit{Attack}. \\textit{Attack} uses an evolutionary\nalgorithm to attack the crucial regions, where the attacks are semantically\nrelated to the target texts of \\textit{Ask}, thus achieving targeted attacks\nwithout semantic loss. Experimental results on transformer-based and\nCNN+RNN-based image-to-text models confirmed the effectiveness of our proposed\n\\textit{AAA}.\n", "date": "2024-08-16", "categories": []}
{"id": "2408.09300", "title": "Malacopula: adversarial automatic speaker verification attacks using a\n  neural-based generalised Hammerstein model", "abstract": "  We present Malacopula, a neural-based generalised Hammerstein model designed\nto introduce adversarial perturbations to spoofed speech utterances so that\nthey better deceive automatic speaker verification (ASV) systems. Using\nnon-linear processes to modify speech utterances, Malacopula enhances the\neffectiveness of spoofing attacks. The model comprises parallel branches of\npolynomial functions followed by linear time-invariant filters. The adversarial\noptimisation procedure acts to minimise the cosine distance between speaker\nembeddings extracted from spoofed and bona fide utterances. Experiments,\nperformed using three recent ASV systems and the ASVspoof 2019 dataset, show\nthat Malacopula increases vulnerabilities by a substantial margin. However,\nspeech quality is reduced and attacks can be detected effectively under\ncontrolled conditions. The findings emphasise the need to identify new\nvulnerabilities and design defences to protect ASV systems from adversarial\nattacks in the wild.\n", "date": "2024-08-17", "categories": []}
{"id": "2408.10021", "title": "Detecting Adversarial Attacks in Semantic Segmentation via Uncertainty\n  Estimation: A Deep Analysis", "abstract": "  Deep neural networks have demonstrated remarkable effectiveness across a wide\nrange of tasks such as semantic segmentation. Nevertheless, these networks are\nvulnerable to adversarial attacks that add imperceptible perturbations to the\ninput image, leading to false predictions. This vulnerability is particularly\ndangerous in safety-critical applications like automated driving. While\nadversarial examples and defense strategies are well-researched in the context\nof image classification, there is comparatively less research focused on\nsemantic segmentation. Recently, we have proposed an uncertainty-based method\nfor detecting adversarial attacks on neural networks for semantic segmentation.\nWe observed that uncertainty, as measured by the entropy of the output\ndistribution, behaves differently on clean versus adversely perturbed images,\nand we utilize this property to differentiate between the two. In this extended\nversion of our work, we conduct a detailed analysis of uncertainty-based\ndetection of adversarial attacks including a diverse set of adversarial attacks\nand various state-of-the-art neural networks. Our numerical experiments show\nthe effectiveness of the proposed uncertainty-based detection method, which is\nlightweight and operates as a post-processing step, i.e., no model\nmodifications or knowledge of the adversarial example generation process are\nrequired.\n", "date": "2024-08-19", "categories": []}
{"id": "2311.01205", "title": "Attacking Graph Neural Networks with Bit Flips: Weisfeiler and Lehman Go\n  Indifferent", "abstract": "  Prior attacks on graph neural networks have mostly focused on graph poisoning\nand evasion, neglecting the network's weights and biases. Traditional\nweight-based fault injection attacks, such as bit flip attacks used for\nconvolutional neural networks, do not consider the unique properties of graph\nneural networks. We propose the Injectivity Bit Flip Attack, the first bit flip\nattack designed specifically for graph neural networks. Our attack targets the\nlearnable neighborhood aggregation functions in quantized message passing\nneural networks, degrading their ability to distinguish graph structures and\nlosing the expressivity of the Weisfeiler-Lehman test. Our findings suggest\nthat exploiting mathematical properties specific to certain graph neural\nnetwork architectures can significantly increase their vulnerability to bit\nflip attacks. Injectivity Bit Flip Attacks can degrade the maximal expressive\nGraph Isomorphism Networks trained on various graph property prediction\ndatasets to random output by flipping only a small fraction of the network's\nbits, demonstrating its higher destructive power compared to a bit flip attack\ntransferred from convolutional neural networks. Our attack is transparent and\nmotivated by theoretical insights which are confirmed by extensive empirical\nresults.\n", "date": "2023-11-02", "categories": []}
{"id": "2408.08374", "title": "Evaluating Text Classification Robustness to Part-of-Speech Adversarial\n  Examples", "abstract": "  As machine learning systems become more widely used, especially for safety\ncritical applications, there is a growing need to ensure that these systems\nbehave as intended, even in the face of adversarial examples. Adversarial\nexamples are inputs that are designed to trick the decision making process, and\nare intended to be imperceptible to humans. However, for text-based\nclassification systems, changes to the input, a string of text, are always\nperceptible. Therefore, text-based adversarial examples instead focus on trying\nto preserve semantics. Unfortunately, recent work has shown this goal is often\nnot met. To improve the quality of text-based adversarial examples, we need to\nknow what elements of the input text are worth focusing on. To address this, in\nthis paper, we explore what parts of speech have the highest impact of\ntext-based classifiers. Our experiments highlight a distinct bias in CNN\nalgorithms against certain parts of speech tokens within review datasets. This\nfinding underscores a critical vulnerability in the linguistic processing\ncapabilities of CNNs.\n", "date": "2024-08-15", "categories": []}
{"id": "2408.08671", "title": "Towards Physical World Backdoor Attacks against Skeleton Action\n  Recognition", "abstract": "  Skeleton Action Recognition (SAR) has attracted significant interest for its\nefficient representation of the human skeletal structure. Despite its\nadvancements, recent studies have raised security concerns in SAR models,\nparticularly their vulnerability to adversarial attacks. However, such\nstrategies are limited to digital scenarios and ineffective in physical\nattacks, limiting their real-world applicability. To investigate the\nvulnerabilities of SAR in the physical world, we introduce the Physical\nSkeleton Backdoor Attacks (PSBA), the first exploration of physical backdoor\nattacks against SAR. Considering the practicalities of physical execution, we\nintroduce a novel trigger implantation method that integrates infrequent and\nimperceivable actions as triggers into the original skeleton data. By\nincorporating a minimal amount of this manipulated data into the training set,\nPSBA enables the system misclassify any skeleton sequences into the target\nclass when the trigger action is present. We examine the resilience of PSBA in\nboth poisoned and clean-label scenarios, demonstrating its efficacy across a\nrange of datasets, poisoning ratios, and model architectures. Additionally, we\nintroduce a trigger-enhancing strategy to strengthen attack performance in the\nclean label setting. The robustness of PSBA is tested against three distinct\nbackdoor defenses, and the stealthiness of PSBA is evaluated using two\nquantitative metrics. Furthermore, by employing a Kinect V2 camera, we compile\na dataset of human actions from the real world to mimic physical attack\nsituations, with our findings confirming the effectiveness of our proposed\nattacks. Our project website can be found at\nhttps://qichenzheng.github.io/psba-website.\n", "date": "2024-08-16", "categories": []}
{"id": "2408.08859", "title": "Stochastic Bandits Robust to Adversarial Attacks", "abstract": "  This paper investigates stochastic multi-armed bandit algorithms that are\nrobust to adversarial attacks, where an attacker can first observe the\nlearner's action and {then} alter their reward observation. We study two cases\nof this model, with or without the knowledge of an attack budget $C$, defined\nas an upper bound of the summation of the difference between the actual and\naltered rewards. For both cases, we devise two types of algorithms with regret\nbounds having additive or multiplicative $C$ dependence terms. For the known\nattack budget case, we prove our algorithms achieve the regret bound of\n${O}((K/\\Delta)\\log T + KC)$ and $\\tilde{O}(\\sqrt{KTC})$ for the additive and\nmultiplicative $C$ terms, respectively, where $K$ is the number of arms, $T$ is\nthe time horizon, $\\Delta$ is the gap between the expected rewards of the\noptimal arm and the second-best arm, and $\\tilde{O}$ hides the logarithmic\nfactors. For the unknown case, we prove our algorithms achieve the regret bound\nof $\\tilde{O}(\\sqrt{KT} + KC^2)$ and $\\tilde{O}(KC\\sqrt{T})$ for the additive\nand multiplicative $C$ terms, respectively. In addition to these upper bound\nresults, we provide several lower bounds showing the tightness of our bounds\nand the optimality of our algorithms. These results delineate an intrinsic\nseparation between the bandits with attacks and corruption models [Lykouris et\nal., 2018].\n", "date": "2024-08-16", "categories": []}
{"id": "2404.11357", "title": "Detector Collapse: Physical-World Backdooring Object Detection to\n  Catastrophic Overload or Blindness in Autonomous Driving", "abstract": "  Object detection tasks, crucial in safety-critical systems like autonomous\ndriving, focus on pinpointing object locations. These detectors are known to be\nsusceptible to backdoor attacks. However, existing backdoor techniques have\nprimarily been adapted from classification tasks, overlooking deeper\nvulnerabilities specific to object detection. This paper is dedicated to\nbridging this gap by introducing Detector Collapse} (DC), a brand-new backdoor\nattack paradigm tailored for object detection. DC is designed to instantly\nincapacitate detectors (i.e., severely impairing detector's performance and\nculminating in a denial-of-service). To this end, we develop two innovative\nattack schemes: Sponge for triggering widespread misidentifications and\nBlinding for rendering objects invisible. Remarkably, we introduce a novel\npoisoning strategy exploiting natural objects, enabling DC to act as a\npractical backdoor in real-world environments. Our experiments on different\ndetectors across several benchmarks show a significant improvement\n($\\sim$10\\%-60\\% absolute and $\\sim$2-7$\\times$ relative) in attack efficacy\nover state-of-the-art attacks.\n", "date": "2024-04-17", "categories": []}
{"id": "2408.07364", "title": "Robust Active Learning (RoAL): Countering Dynamic Adversaries in Active\n  Learning with Elastic Weight Consolidation", "abstract": "  Despite significant advancements in active learning and adversarial attacks,\nthe intersection of these two fields remains underexplored, particularly in\ndeveloping robust active learning frameworks against dynamic adversarial\nthreats. The challenge of developing robust active learning frameworks under\ndynamic adversarial attacks is critical, as these attacks can lead to\ncatastrophic forgetting within the active learning cycle. This paper introduces\nRobust Active Learning (RoAL), a novel approach designed to address this issue\nby integrating Elastic Weight Consolidation (EWC) into the active learning\nprocess. Our contributions are threefold: First, we propose a new dynamic\nadversarial attack that poses significant threats to active learning\nframeworks. Second, we introduce a novel method that combines EWC with active\nlearning to mitigate catastrophic forgetting caused by dynamic adversarial\nattacks. Finally, we conduct extensive experimental evaluations to demonstrate\nthe efficacy of our approach. The results show that RoAL not only effectively\ncounters dynamic adversarial threats but also significantly reduces the impact\nof catastrophic forgetting, thereby enhancing the robustness and performance of\nactive learning systems in adversarial environments.\n", "date": "2024-08-14", "categories": []}
{"id": "2408.07440", "title": "BAPLe: Backdoor Attacks on Medical Foundational Models using Prompt\n  Learning", "abstract": "  Medical foundation models are gaining prominence in the medical community for\ntheir ability to derive general representations from extensive collections of\nmedical image-text pairs. Recent research indicates that these models are\nsusceptible to backdoor attacks, which allow them to classify clean images\naccurately but fail when specific triggers are introduced. However, traditional\nbackdoor attacks necessitate a considerable amount of additional data to\nmaliciously pre-train a model. This requirement is often impractical in medical\nimaging applications due to the usual scarcity of data. Inspired by the latest\ndevelopments in learnable prompts, this work introduces a method to embed a\nbackdoor into the medical foundation model during the prompt learning phase. By\nincorporating learnable prompts within the text encoder and introducing\nimperceptible learnable noise trigger to the input images, we exploit the full\ncapabilities of the medical foundation models (Med-FM). Our method, BAPLe,\nrequires only a minimal subset of data to adjust the noise trigger and the text\nprompts for downstream tasks, enabling the creation of an effective backdoor\nattack. Through extensive experiments with four medical foundation models, each\npre-trained on different modalities and evaluated across six downstream\ndatasets, we demonstrate the efficacy of our approach. BAPLe achieves a high\nbackdoor success rate across all models and datasets, outperforming the\nbaseline backdoor attack methods. Our work highlights the vulnerability of\nMed-FMs towards backdoor attacks and strives to promote the safe adoption of\nMed-FMs before their deployment in real-world applications. Code is available\nat https://asif-hanif.github.io/baple/.\n", "date": "2024-08-14", "categories": []}
{"id": "2408.07733", "title": "Enhancing Adversarial Attacks via Parameter Adaptive Adversarial Attack", "abstract": "  In recent times, the swift evolution of adversarial attacks has captured\nwidespread attention, particularly concerning their transferability and other\nperformance attributes. These techniques are primarily executed at the sample\nlevel, frequently overlooking the intrinsic parameters of models. Such neglect\nsuggests that the perturbations introduced in adversarial samples might have\nthe potential for further reduction. Given the essence of adversarial attacks\nis to impair model integrity with minimal noise on original samples, exploring\navenues to maximize the utility of such perturbations is imperative. Against\nthis backdrop, we have delved into the complexities of adversarial attack\nalgorithms, dissecting the adversarial process into two critical phases: the\nDirectional Supervision Process (DSP) and the Directional Optimization Process\n(DOP). While DSP determines the direction of updates based on the current\nsamples and model parameters, it has been observed that existing model\nparameters may not always be conducive to adversarial attacks. The impact of\nmodels on adversarial efficacy is often overlooked in current research, leading\nto the neglect of DSP. We propose that under certain conditions, fine-tuning\nmodel parameters can significantly enhance the quality of DSP. For the first\ntime, we propose that under certain conditions, fine-tuning model parameters\ncan significantly improve the quality of the DSP. We provide, for the first\ntime, rigorous mathematical definitions and proofs for these conditions, and\nintroduce multiple methods for fine-tuning model parameters within DSP. Our\nextensive experiments substantiate the effectiveness of the proposed P3A\nmethod. Our code is accessible at: https://anonymous.4open.science/r/P3A-A12C/\n", "date": "2024-08-14", "categories": []}
{"id": "2306.07197", "title": "AROID: Improving Adversarial Robustness Through Online Instance-Wise\n  Data Augmentation", "abstract": "  Deep neural networks are vulnerable to adversarial examples. Adversarial\ntraining (AT) is an effective defense against adversarial examples. However, AT\nis prone to overfitting which degrades robustness substantially. Recently, data\naugmentation (DA) was shown to be effective in mitigating robust overfitting if\nappropriately designed and optimized for AT. This work proposes a new method to\nautomatically learn online, instance-wise, DA policies to improve robust\ngeneralization for AT. This is the first automated DA method specific for\nrobustness. A novel policy learning objective, consisting of Vulnerability,\nAffinity and Diversity, is proposed and shown to be sufficiently effective and\nefficient to be practical for automatic DA generation during AT. Importantly,\nour method dramatically reduces the cost of policy search from the 5000 hours\nof AutoAugment and the 412 hours of IDBH to 9 hours, making automated DA more\npractical to use for adversarial robustness. This allows our method to\nefficiently explore a large search space for a more effective DA policy and\nevolve the policy as training progresses. Empirically, our method is shown to\noutperform all competitive DA methods across various model architectures and\ndatasets. Our DA policy reinforced vanilla AT to surpass several\nstate-of-the-art AT methods regarding both accuracy and robustness. It can also\nbe combined with those advanced AT methods to further boost robustness. Code\nand pre-trained models are available at https://github.com/TreeLLi/AROID.\n", "date": "2023-06-12", "categories": []}
{"id": "2403.09171", "title": "ADEdgeDrop: Adversarial Edge Dropping for Robust Graph Neural Networks", "abstract": "  Although Graph Neural Networks (GNNs) have exhibited the powerful ability to\ngather graph-structured information from neighborhood nodes via various\nmessage-passing mechanisms, the performance of GNNs is limited by poor\ngeneralization and fragile robustness caused by noisy and redundant graph data.\nAs a prominent solution, Graph Augmentation Learning (GAL) has recently\nreceived increasing attention. Among prior GAL approaches, edge-dropping\nmethods that randomly remove edges from a graph during training are effective\ntechniques to improve the robustness of GNNs. However, randomly dropping edges\noften results in bypassing critical edges, consequently weakening the\neffectiveness of message passing. In this paper, we propose a novel adversarial\nedge-dropping method (ADEdgeDrop) that leverages an adversarial edge predictor\nguiding the removal of edges, which can be flexibly incorporated into diverse\nGNN backbones. Employing an adversarial training framework, the edge predictor\nutilizes the line graph transformed from the original graph to estimate the\nedges to be dropped, which improves the interpretability of the edge-dropping\nmethod. The proposed ADEdgeDrop is optimized alternately by stochastic gradient\ndescent and projected gradient descent. Comprehensive experiments on six graph\nbenchmark datasets demonstrate that the proposed ADEdgeDrop outperforms\nstate-of-the-art baselines across various GNN backbones, demonstrating improved\ngeneralization and robustness.\n", "date": "2024-03-14", "categories": []}
{"id": "2408.07558", "title": "Sonic: Fast and Transferable Data Poisoning on Clustering Algorithms", "abstract": "  Data poisoning attacks on clustering algorithms have received limited\nattention, with existing methods struggling to scale efficiently as dataset\nsizes and feature counts increase. These attacks typically require\nre-clustering the entire dataset multiple times to generate predictions and\nassess the attacker's objectives, significantly hindering their scalability.\nThis paper addresses these limitations by proposing Sonic, a novel genetic data\npoisoning attack that leverages incremental and scalable clustering algorithms,\ne.g., FISHDBC, as surrogates to accelerate poisoning attacks against\ngraph-based and density-based clustering methods, such as HDBSCAN. We\nempirically demonstrate the effectiveness and efficiency of Sonic in poisoning\nthe target clustering algorithms. We then conduct a comprehensive analysis of\nthe factors affecting the scalability and transferability of poisoning attacks\nagainst clustering algorithms, and we conclude by examining the robustness of\nhyperparameters in our attack strategy Sonic.\n", "date": "2024-08-14", "categories": []}
{"id": "2408.07579", "title": "TabularBench: Benchmarking Adversarial Robustness for Tabular Deep\n  Learning in Real-world Use-cases", "abstract": "  While adversarial robustness in computer vision is a mature research field,\nfewer researchers have tackled the evasion attacks against tabular deep\nlearning, and even fewer investigated robustification mechanisms and reliable\ndefenses. We hypothesize that this lag in the research on tabular adversarial\nattacks is in part due to the lack of standardized benchmarks. To fill this\ngap, we propose TabularBench, the first comprehensive benchmark of robustness\nof tabular deep learning classification models. We evaluated adversarial\nrobustness with CAA, an ensemble of gradient and search attacks which was\nrecently demonstrated as the most effective attack against a tabular model. In\naddition to our open benchmark (https://github.com/serval-uni-lu/tabularbench)\nwhere we welcome submissions of new models and defenses, we implement 7\nrobustification mechanisms inspired by state-of-the-art defenses in computer\nvision and propose the largest benchmark of robust tabular deep learning over\n200 models across five critical scenarios in finance, healthcare and security.\nWe curated real datasets for each use case, augmented with hundreds of\nthousands of realistic synthetic inputs, and trained and assessed our models\nwith and without data augmentations. We open-source our library that provides\nAPI access to all our pre-trained robust tabular models, and the largest\ndatasets of real and synthetic tabular inputs. Finally, we analyze the impact\nof various defenses on the robustness and provide actionable insights to design\nnew defenses and robustification mechanisms.\n", "date": "2024-08-14", "categories": []}
{"id": "2402.11196", "title": "Maintaining Adversarial Robustness in Continuous Learning", "abstract": "  Adversarial robustness is essential for security and reliability of machine\nlearning systems. However, adversarial robustness enhanced by defense\nalgorithms is easily erased as the neural network's weights update to learn new\ntasks. To address this vulnerability, it is essential to improve the capability\nof neural networks in terms of robust continual learning. Specially, we propose\na novel gradient projection technique that effectively stabilizes sample\ngradients from previous data by orthogonally projecting back-propagation\ngradients onto a crucial subspace before using them for weight updates. This\ntechnique can maintaining robustness by collaborating with a class of defense\nalgorithms through sample gradient smoothing. The experimental results on four\nbenchmarks including Split-CIFAR100 and Split-miniImageNet, demonstrate that\nthe superiority of the proposed approach in mitigating rapidly degradation of\nrobustness during continual learning even when facing strong adversarial\nattacks.\n", "date": "2024-02-17", "categories": []}
{"id": "2408.06509", "title": "Fooling SHAP with Output Shuffling Attacks", "abstract": "  Explainable AI~(XAI) methods such as SHAP can help discover feature\nattributions in black-box models. If the method reveals a significant\nattribution from a ``protected feature'' (e.g., gender, race) on the model\noutput, the model is considered unfair. However, adversarial attacks can\nsubvert the detection of XAI methods. Previous approaches to constructing such\nan adversarial model require access to underlying data distribution, which may\nnot be possible in many practical scenarios. We relax this constraint and\npropose a novel family of attacks, called shuffling attacks, that are\ndata-agnostic. The proposed attack strategies can adapt any trained machine\nlearning model to fool Shapley value-based explanations. We prove that Shapley\nvalues cannot detect shuffling attacks. However, algorithms that estimate\nShapley values, such as linear SHAP and SHAP, can detect these attacks with\nvarying degrees of effectiveness. We demonstrate the efficacy of the attack\nstrategies by comparing the performance of linear SHAP and SHAP using\nreal-world datasets.\n", "date": "2024-08-12", "categories": []}
{"id": "2408.06625", "title": "DePatch: Towards Robust Adversarial Patch for Evading Person Detectors\n  in the Real World", "abstract": "  Recent years have seen an increasing interest in physical adversarial\nattacks, which aim to craft deployable patterns for deceiving deep neural\nnetworks, especially for person detectors. However, the adversarial patterns of\nexisting patch-based attacks heavily suffer from the self-coupling issue, where\na degradation, caused by physical transformations, in any small patch segment\ncan result in a complete adversarial dysfunction, leading to poor robustness in\nthe complex real world. Upon this observation, we introduce the Decoupled\nadversarial Patch (DePatch) attack to address the self-coupling issue of\nadversarial patches. Specifically, we divide the adversarial patch into\nblock-wise segments, and reduce the inter-dependency among these segments\nthrough randomly erasing out some segments during the optimization. We further\nintroduce a border shifting operation and a progressive decoupling strategy to\nimprove the overall attack capabilities. Extensive experiments demonstrate the\nsuperior performance of our method over other physical adversarial attacks,\nespecially in the real world.\n", "date": "2024-08-13", "categories": []}
{"id": "2408.06665", "title": "RW-NSGCN: A Robust Approach to Structural Attacks via Negative Sampling", "abstract": "  Node classification using Graph Neural Networks (GNNs) has been widely\napplied in various practical scenarios, such as predicting user interests and\ndetecting communities in social networks. However, recent studies have shown\nthat graph-structured networks often contain potential noise and attacks, in\nthe form of topological perturbations and weight disturbances, which can lead\nto decreased classification performance in GNNs. To improve the robustness of\nthe model, we propose a novel method: Random Walk Negative Sampling Graph\nConvolutional Network (RW-NSGCN). Specifically, RW-NSGCN integrates the Random\nWalk with Restart (RWR) and PageRank (PGR) algorithms for negative sampling and\nemploys a Determinantal Point Process (DPP)-based GCN for convolution\noperations. RWR leverages both global and local information to manage noise and\nlocal variations, while PGR assesses node importance to stabilize the\ntopological structure. The DPP-based GCN ensures diversity among negative\nsamples and aggregates their features to produce robust node embeddings,\nthereby improving classification performance. Experimental results demonstrate\nthat the RW-NSGCN model effectively addresses network topology attacks and\nweight instability, increasing the accuracy of anomaly detection and overall\nstability. In terms of classification accuracy, RW-NSGCN significantly\noutperforms existing methods, showing greater resilience across various\nscenarios and effectively mitigating the impact of such vulnerabilities.\n", "date": "2024-08-13", "categories": []}
{"id": "2312.16156", "title": "From text to multimodal: a survey of adversarial example generation in\n  question answering systems", "abstract": "  Integrating adversarial machine learning with Question Answering (QA) systems\nhas emerged as a critical area for understanding the vulnerabilities and\nrobustness of these systems. This article aims to comprehensively review\nadversarial example-generation techniques in the QA field, including textual\nand multimodal contexts. We examine the techniques employed through systematic\ncategorization, providing a comprehensive, structured review. Beginning with an\noverview of traditional QA models, we traverse the adversarial example\ngeneration by exploring rule-based perturbations and advanced generative\nmodels. We then extend our research to include multimodal QA systems, analyze\nthem across various methods, and examine generative models, seq2seq\narchitectures, and hybrid methodologies. Our research grows to different\ndefense strategies, adversarial datasets, and evaluation metrics and\nillustrates the comprehensive literature on adversarial QA. Finally, the paper\nconsiders the future landscape of adversarial question generation, highlighting\npotential research directions that can advance textual and multimodal QA\nsystems in the context of adversarial challenges.\n", "date": "2023-12-26", "categories": []}
{"id": "2408.05446", "title": "Ensemble everything everywhere: Multi-scale aggregation for adversarial\n  robustness", "abstract": "  Adversarial examples pose a significant challenge to the robustness,\nreliability and alignment of deep neural networks. We propose a novel,\neasy-to-use approach to achieving high-quality representations that lead to\nadversarial robustness through the use of multi-resolution input\nrepresentations and dynamic self-ensembling of intermediate layer predictions.\nWe demonstrate that intermediate layer predictions exhibit inherent robustness\nto adversarial attacks crafted to fool the full classifier, and propose a\nrobust aggregation mechanism based on Vickrey auction that we call\n\\textit{CrossMax} to dynamically ensemble them. By combining multi-resolution\ninputs and robust ensembling, we achieve significant adversarial robustness on\nCIFAR-10 and CIFAR-100 datasets without any adversarial training or extra data,\nreaching an adversarial accuracy of $\\approx$72% (CIFAR-10) and $\\approx$48%\n(CIFAR-100) on the RobustBench AutoAttack suite ($L_\\infty=8/255)$ with a\nfinetuned ImageNet-pretrained ResNet152. This represents a result comparable\nwith the top three models on CIFAR-10 and a +5 % gain compared to the best\ncurrent dedicated approach on CIFAR-100. Adding simple adversarial training on\ntop, we get $\\approx$78% on CIFAR-10 and $\\approx$51% on CIFAR-100, improving\nSOTA by 5 % and 9 % respectively and seeing greater gains on the harder\ndataset. We validate our approach through extensive experiments and provide\ninsights into the interplay between adversarial robustness, and the\nhierarchical nature of deep representations. We show that simple gradient-based\nattacks against our model lead to human-interpretable images of the target\nclasses as well as interpretable image changes. As a byproduct, using our\nmulti-resolution prior, we turn pre-trained classifiers and CLIP models into\ncontrollable image generators and develop successful transferable attacks on\nlarge vision language models.\n", "date": "2024-08-08", "categories": []}
{"id": "2408.05479", "title": "ReToMe-VA: Recursive Token Merging for Video Diffusion-based\n  Unrestricted Adversarial Attack", "abstract": "  Recent diffusion-based unrestricted attacks generate imperceptible\nadversarial examples with high transferability compared to previous\nunrestricted attacks and restricted attacks. However, existing works on\ndiffusion-based unrestricted attacks are mostly focused on images yet are\nseldom explored in videos. In this paper, we propose the Recursive Token\nMerging for Video Diffusion-based Unrestricted Adversarial Attack (ReToMe-VA),\nwhich is the first framework to generate imperceptible adversarial video clips\nwith higher transferability. Specifically, to achieve spatial imperceptibility,\nReToMe-VA adopts a Timestep-wise Adversarial Latent Optimization (TALO)\nstrategy that optimizes perturbations in diffusion models' latent space at each\ndenoising step. TALO offers iterative and accurate updates to generate more\npowerful adversarial frames. TALO can further reduce memory consumption in\ngradient computation. Moreover, to achieve temporal imperceptibility, ReToMe-VA\nintroduces a Recursive Token Merging (ReToMe) mechanism by matching and merging\ntokens across video frames in the self-attention module, resulting in\ntemporally consistent adversarial videos. ReToMe concurrently facilitates\ninter-frame interactions into the attack process, inducing more diverse and\nrobust gradients, thus leading to better adversarial transferability. Extensive\nexperiments demonstrate the efficacy of ReToMe-VA, particularly in surpassing\nstate-of-the-art attacks in adversarial transferability by more than 14.16% on\naverage.\n", "date": "2024-08-10", "categories": []}
{"id": "2408.05669", "title": "StealthDiffusion: Towards Evading Diffusion Forensic Detection through\n  Diffusion Model", "abstract": "  The rapid progress in generative models has given rise to the critical task\nof AI-Generated Content Stealth (AIGC-S), which aims to create AI-generated\nimages that can evade both forensic detectors and human inspection. This task\nis crucial for understanding the vulnerabilities of existing detection methods\nand developing more robust techniques. However, current adversarial attacks\noften introduce visible noise, have poor transferability, and fail to address\nspectral differences between AI-generated and genuine images. To address this,\nwe propose StealthDiffusion, a framework based on stable diffusion that\nmodifies AI-generated images into high-quality, imperceptible adversarial\nexamples capable of evading state-of-the-art forensic detectors.\nStealthDiffusion comprises two main components: Latent Adversarial\nOptimization, which generates adversarial perturbations in the latent space of\nstable diffusion, and Control-VAE, a module that reduces spectral differences\nbetween the generated adversarial images and genuine images without affecting\nthe original diffusion model's generation process. Extensive experiments show\nthat StealthDiffusion is effective in both white-box and black-box settings,\ntransforming AI-generated images into high-quality adversarial forgeries with\nfrequency spectra similar to genuine images. These forgeries are classified as\ngenuine by advanced forensic classifiers and are difficult for humans to\ndistinguish.\n", "date": "2024-08-11", "categories": []}
{"id": "2408.05900", "title": "Classifier Guidance Enhances Diffusion-based Adversarial Purification by\n  Preserving Predictive Information", "abstract": "  Adversarial purification is one of the promising approaches to defend neural\nnetworks against adversarial attacks. Recently, methods utilizing diffusion\nprobabilistic models have achieved great success for adversarial purification\nin image classification tasks. However, such methods fall into the dilemma of\nbalancing the needs for noise removal and information preservation. This paper\npoints out that existing adversarial purification methods based on diffusion\nmodels gradually lose sample information during the core denoising process,\ncausing occasional label shift in subsequent classification tasks. As a remedy,\nwe suggest to suppress such information loss by introducing guidance from the\nclassifier confidence. Specifically, we propose Classifier-cOnfidence gUided\nPurification (COUP) algorithm, which purifies adversarial examples while\nkeeping away from the classifier decision boundary. Experimental results show\nthat COUP can achieve better adversarial robustness under strong attack\nmethods.\n", "date": "2024-08-12", "categories": []}
{"id": "2309.16710", "title": "General Lipschitz: Certified Robustness Against Resolvable Semantic\n  Transformations via Transformation-Dependent Randomized Smoothing", "abstract": "  Randomized smoothing is the state-of-the-art approach to construct image\nclassifiers that are provably robust against additive adversarial perturbations\nof bounded magnitude. However, it is more complicated to construct reasonable\ncertificates against semantic transformation (e.g., image blurring,\ntranslation, gamma correction) and their compositions. In this work, we propose\n\\emph{General Lipschitz (GL),} a new framework to certify neural networks\nagainst composable resolvable semantic perturbations. Within the framework, we\nanalyze transformation-dependent Lipschitz-continuity of smoothed classifiers\nw.r.t. transformation parameters and derive corresponding robustness\ncertificates. Our method performs comparably to state-of-the-art approaches on\nthe ImageNet dataset.\n", "date": "2023-08-17", "categories": []}
{"id": "2408.04686", "title": "Multi-Turn Context Jailbreak Attack on Large Language Models From First\n  Principles", "abstract": "  Large language models (LLMs) have significantly enhanced the performance of\nnumerous applications, from intelligent conversations to text generation.\nHowever, their inherent security vulnerabilities have become an increasingly\nsignificant challenge, especially with respect to jailbreak attacks. Attackers\ncan circumvent the security mechanisms of these LLMs, breaching security\nconstraints and causing harmful outputs. Focusing on multi-turn semantic\njailbreak attacks, we observe that existing methods lack specific\nconsiderations for the role of multiturn dialogues in attack strategies,\nleading to semantic deviations during continuous interactions. Therefore, in\nthis paper, we establish a theoretical foundation for multi-turn attacks by\nconsidering their support in jailbreak attacks, and based on this, propose a\ncontext-based contextual fusion black-box jailbreak attack method, named\nContext Fusion Attack (CFA). This method approach involves filtering and\nextracting key terms from the target, constructing contextual scenarios around\nthese terms, dynamically integrating the target into the scenarios, replacing\nmalicious key terms within the target, and thereby concealing the direct\nmalicious intent. Through comparisons on various mainstream LLMs and red team\ndatasets, we have demonstrated CFA's superior success rate, divergence, and\nharmfulness compared to other multi-turn attack strategies, particularly\nshowcasing significant advantages on Llama3 and GPT-4.\n", "date": "2024-08-08", "categories": []}
{"id": "2408.05124", "title": "Modeling Electromagnetic Signal Injection Attacks on Camera-based Smart\n  Systems: Applications and Mitigation", "abstract": "  Numerous safety- or security-critical systems depend on cameras to perceive\ntheir surroundings, further allowing artificial intelligence (AI) to analyze\nthe captured images to make important decisions. However, a concerning attack\nvector has emerged, namely, electromagnetic waves, which pose a threat to the\nintegrity of these systems. Such attacks enable attackers to manipulate the\nimages remotely, leading to incorrect AI decisions, e.g., autonomous vehicles\nmissing detecting obstacles ahead resulting in collisions. The lack of\nunderstanding regarding how different systems react to such attacks poses a\nsignificant security risk. Furthermore, no effective solutions have been\ndemonstrated to mitigate this threat.\n  To address these gaps, we modeled the attacks and developed a simulation\nmethod for generating adversarial images. Through rigorous analysis, we\nconfirmed that the effects of the simulated adversarial images are\nindistinguishable from those from real attacks. This method enables researchers\nand engineers to rapidly assess the susceptibility of various AI vision\napplications to these attacks, without the need for constructing complicated\nattack devices. In our experiments, most of the models demonstrated\nvulnerabilities to these attacks, emphasizing the need to enhance their\nrobustness. Fortunately, our modeling and simulation method serves as a\nstepping stone toward developing more resilient models. We present a pilot\nstudy on adversarial training to improve their robustness against attacks, and\nour results demonstrate a significant improvement by recovering up to 91%\nperformance, offering a promising direction for mitigating this threat.\n", "date": "2024-08-09", "categories": []}
{"id": "2408.03972", "title": "Enhancing Output Diversity Improves Conjugate Gradient-based Adversarial\n  Attacks", "abstract": "  Deep neural networks are vulnerable to adversarial examples, and adversarial\nattacks that generate adversarial examples have been studied in this context.\nExisting studies imply that increasing the diversity of model outputs\ncontributes to improving the attack performance. This study focuses on the Auto\nConjugate Gradient (ACG) attack, which is inspired by the conjugate gradient\nmethod and has a high diversification performance. We hypothesized that\nincreasing the distance between two consecutive search points would enhance the\noutput diversity. To test our hypothesis, we propose Rescaling-ACG (ReACG),\nwhich automatically modifies the two components that significantly affect the\ndistance between two consecutive search points, including the search direction\nand step size. ReACG showed higher attack performance than that of ACG, and is\nparticularly effective for ImageNet models with several classification classes.\nExperimental results show that the distance between two consecutive search\npoints enhances the output diversity and may help develop new potent attacks.\nThe code is available at \\url{https://github.com/yamamura-k/ReACG}\n", "date": "2024-08-07", "categories": []}
{"id": "2408.04181", "title": "EdgeShield: A Universal and Efficient Edge Computing Framework for\n  Robust AI", "abstract": "  The increasing prevalence of adversarial attacks on Artificial Intelligence\n(AI) systems has created a need for innovative security measures. However, the\ncurrent methods of defending against these attacks often come with a high\ncomputing cost and require back-end processing, making real-time defense\nchallenging. Fortunately, there have been remarkable advancements in\nedge-computing, which make it easier to deploy neural networks on edge devices.\nBuilding upon these advancements, we propose an edge framework design to enable\nuniversal and efficient detection of adversarial attacks. This framework\nincorporates an attention-based adversarial detection methodology and a\nlightweight detection network formation, making it suitable for a wide range of\nneural networks and can be deployed on edge devices. To assess the\neffectiveness of our proposed framework, we conducted evaluations on five\nneural networks. The results indicate an impressive 97.43% F-score can be\nachieved, demonstrating the framework's proficiency in detecting adversarial\nattacks. Moreover, our proposed framework also exhibits significantly reduced\ncomputing complexity and cost in comparison to previous detection methods. This\naspect is particularly beneficial as it ensures that the defense mechanism can\nbe efficiently implemented in real-time on-edge devices.\n", "date": "2024-08-08", "categories": []}
{"id": "2408.04261", "title": "Unveiling Hidden Visual Information: A Reconstruction Attack Against\n  Adversarial Visual Information Hiding", "abstract": "  This paper investigates the security vulnerabilities of\nadversarial-example-based image encryption by executing data reconstruction\n(DR) attacks on encrypted images. A representative image encryption method is\nthe adversarial visual information hiding (AVIH), which uses type-I adversarial\nexample training to protect gallery datasets used in image recognition tasks.\nIn the AVIH method, the type-I adversarial example approach creates images that\nappear completely different but are still recognized by machines as the\noriginal ones. Additionally, the AVIH method can restore encrypted images to\ntheir original forms using a predefined private key generative model. For the\nbest security, assigning a unique key to each image is recommended; however,\nstorage limitations may necessitate some images sharing the same key model.\nThis raises a crucial security question for AVIH: How many images can safely\nshare the same key model without being compromised by a DR attack? To address\nthis question, we introduce a dual-strategy DR attack against the AVIH\nencryption method by incorporating (1) generative-adversarial loss and (2)\naugmented identity loss, which prevent DR from overfitting -- an issue akin to\nthat in machine learning. Our numerical results validate this approach through\nimage recognition and re-identification benchmarks, demonstrating that our\nstrategy can significantly enhance the quality of reconstructed images, thereby\nrequiring fewer key-sharing encrypted images. Our source code to reproduce our\nresults will be available soon.\n", "date": "2024-08-08", "categories": []}
{"id": "2305.10856", "title": "Spatial-Frequency Discriminability for Revealing Adversarial\n  Perturbations", "abstract": "  The vulnerability of deep neural networks to adversarial perturbations has\nbeen widely perceived in the computer vision community. From a security\nperspective, it poses a critical risk for modern vision systems, e.g., the\npopular Deep Learning as a Service (DLaaS) frameworks. For protecting deep\nmodels while not modifying them, current algorithms typically detect\nadversarial patterns through discriminative decomposition for natural and\nadversarial data. However, these decompositions are either biased towards\nfrequency resolution or spatial resolution, thus failing to capture adversarial\npatterns comprehensively. Also, when the detector relies on few fixed features,\nit is practical for an adversary to fool the model while evading the detector\n(i.e., defense-aware attack). Motivated by such facts, we propose a\ndiscriminative detector relying on a spatial-frequency Krawtchouk\ndecomposition. It expands the above works from two aspects: 1) the introduced\nKrawtchouk basis provides better spatial-frequency discriminability, capturing\nthe differences between natural and adversarial data comprehensively in both\nspatial and frequency distributions, w.r.t. the common trigonometric or wavelet\nbasis; 2) the extensive features formed by the Krawtchouk decomposition allows\nfor adaptive feature selection and secrecy mechanism, significantly increasing\nthe difficulty of the defense-aware attack, w.r.t. the detector with few fixed\nfeatures. Theoretical and numerical analyses demonstrate the uniqueness and\nusefulness of our detector, exhibiting competitive scores on several deep\nmodels and image sets against a variety of adversarial attacks.\n", "date": "2023-05-18", "categories": []}
{"id": "2408.03909", "title": "LaFA: Latent Feature Attacks on Non-negative Matrix Factorization", "abstract": "  As Machine Learning (ML) applications rapidly grow, concerns about\nadversarial attacks compromising their reliability have gained significant\nattention. One unsupervised ML method known for its resilience to such attacks\nis Non-negative Matrix Factorization (NMF), an algorithm that decomposes input\ndata into lower-dimensional latent features. However, the introduction of\npowerful computational tools such as Pytorch enables the computation of\ngradients of the latent features with respect to the original data, raising\nconcerns about NMF's reliability. Interestingly, naively deriving the\nadversarial loss for NMF as in the case of ML would result in the\nreconstruction loss, which can be shown theoretically to be an ineffective\nattacking objective. In this work, we introduce a novel class of attacks in NMF\ntermed Latent Feature Attacks (LaFA), which aim to manipulate the latent\nfeatures produced by the NMF process. Our method utilizes the Feature Error\n(FE) loss directly on the latent features. By employing FE loss, we generate\nperturbations in the original data that significantly affect the extracted\nlatent features, revealing vulnerabilities akin to those found in other ML\ntechniques. To handle large peak-memory overhead from gradient back-propagation\nin FE attacks, we develop a method based on implicit differentiation which\nenables their scaling to larger datasets. We validate NMF vulnerabilities and\nFE attacks effectiveness through extensive experiments on synthetic and\nreal-world data.\n", "date": "2024-08-07", "categories": []}
{"id": "2309.01351", "title": "Adv3D: Generating 3D Adversarial Examples for 3D Object Detection in\n  Driving Scenarios with NeRF", "abstract": "  Deep neural networks (DNNs) have been proven extremely susceptible to\nadversarial examples, which raises special safety-critical concerns for\nDNN-based autonomous driving stacks (i.e., 3D object detection). Although there\nare extensive works on image-level attacks, most are restricted to 2D pixel\nspaces, and such attacks are not always physically realistic in our 3D world.\nHere we present Adv3D, the first exploration of modeling adversarial examples\nas Neural Radiance Fields (NeRFs). Advances in NeRF provide photorealistic\nappearances and 3D accurate generation, yielding a more realistic and\nrealizable adversarial example. We train our adversarial NeRF by minimizing the\nsurrounding objects' confidence predicted by 3D detectors on the training set.\nThen we evaluate Adv3D on the unseen validation set and show that it can cause\na large performance reduction when rendering NeRF in any sampled pose. To\ngenerate physically realizable adversarial examples, we propose primitive-aware\nsampling and semantic-guided regularization that enable 3D patch attacks with\ncamouflage adversarial texture. Experimental results demonstrate that the\ntrained adversarial NeRF generalizes well to different poses, scenes, and 3D\ndetectors. Finally, we provide a defense method to our attacks that involves\nadversarial training through data augmentation. Project page:\nhttps://len-li.github.io/adv3d-web\n", "date": "2023-09-04", "categories": []}
{"id": "2404.05530", "title": "Best-of-Venom: Attacking RLHF by Injecting Poisoned Preference Data", "abstract": "  Reinforcement Learning from Human Feedback (RLHF) is a popular method for\naligning Language Models (LM) with human values and preferences. RLHF requires\na large number of preference pairs as training data, which are often used in\nboth the Supervised Fine-Tuning and Reward Model training and therefore\npublicly available datasets are commonly used. In this work, we study to what\nextent a malicious actor can manipulate the LMs generations by poisoning the\npreferences, i.e., injecting poisonous preference pairs into these datasets and\nthe RLHF training process. We propose strategies to build poisonous preference\npairs and test their performance by poisoning two widely used preference\ndatasets. Our results show that preference poisoning is highly effective:\ninjecting a small amount of poisonous data (1-5\\% of the original dataset), we\ncan effectively manipulate the LM to generate a target entity in a target\nsentiment (positive or negative). The findings from our experiments also shed\nlight on strategies to defend against the preference poisoning attack.\n", "date": "2024-04-08", "categories": []}
{"id": "2407.03234", "title": "Self-Evaluation as a Defense Against Adversarial Attacks on LLMs", "abstract": "  We introduce a defense against adversarial attacks on LLMs utilizing\nself-evaluation. Our method requires no model fine-tuning, instead using\npre-trained models to evaluate the inputs and outputs of a generator model,\nsignificantly reducing the cost of implementation in comparison to other,\nfinetuning-based methods. Our method can significantly reduce the attack\nsuccess rate of attacks on both open and closed-source LLMs, beyond the\nreductions demonstrated by Llama-Guard2 and commonly used content moderation\nAPIs. We present an analysis of the effectiveness of our method, including\nattempts to attack the evaluator in various settings, demonstrating that it is\nalso more resilient to attacks than existing methods. Code and data will be\nmade available at https://github.com/Linlt-leon/self-eval.\n", "date": "2024-07-03", "categories": []}
{"id": "2408.02980", "title": "Sample-agnostic Adversarial Perturbation for Vision-Language\n  Pre-training Models", "abstract": "  Recent studies on AI security have highlighted the vulnerability of\nVision-Language Pre-training (VLP) models to subtle yet intentionally designed\nperturbations in images and texts. Investigating multimodal systems' robustness\nvia adversarial attacks is crucial in this field. Most multimodal attacks are\nsample-specific, generating a unique perturbation for each sample to construct\nadversarial samples. To the best of our knowledge, it is the first work through\nmultimodal decision boundaries to explore the creation of a universal,\nsample-agnostic perturbation that applies to any image. Initially, we explore\nstrategies to move sample points beyond the decision boundaries of linear\nclassifiers, refining the algorithm to ensure successful attacks under the top\n$k$ accuracy metric. Based on this foundation, in visual-language tasks, we\ntreat visual and textual modalities as reciprocal sample points and decision\nhyperplanes, guiding image embeddings to traverse text-constructed decision\nboundaries, and vice versa. This iterative process consistently refines a\nuniversal perturbation, ultimately identifying a singular direction within the\ninput space which is exploitable to impair the retrieval performance of VLP\nmodels. The proposed algorithms support the creation of global perturbations or\nadversarial patches. Comprehensive experiments validate the effectiveness of\nour method, showcasing its data, task, and model transferability across various\nVLP models and datasets. Code: https://github.com/LibertazZ/MUAP\n", "date": "2024-08-06", "categories": []}
{"id": "2303.01351", "title": "APARATE: Adaptive Adversarial Patch for CNN-based Monocular Depth\n  Estimation for Autonomous Navigation", "abstract": "  In recent times, monocular depth estimation (MDE) has experienced significant\nadvancements in performance, largely attributed to the integration of\ninnovative architectures, i.e., convolutional neural networks (CNNs) and\nTransformers. Nevertheless, the susceptibility of these models to adversarial\nattacks has emerged as a noteworthy concern, especially in domains where safety\nand security are paramount. This concern holds particular weight for MDE due to\nits critical role in applications like autonomous driving and robotic\nnavigation, where accurate scene understanding is pivotal. To assess the\nvulnerability of CNN-based depth prediction methods, recent work tries to\ndesign adversarial patches against MDE. However, the existing approaches fall\nshort of inducing a comprehensive and substantially disruptive impact on the\nvision system. Instead, their influence is partial and confined to specific\nlocal areas. These methods lead to erroneous depth predictions only within the\noverlapping region with the input image, without considering the\ncharacteristics of the target object, such as its size, shape, and position. In\nthis paper, we introduce a novel adversarial patch named APARATE. This patch\npossesses the ability to selectively undermine MDE in two distinct ways: by\ndistorting the estimated distances or by creating the illusion of an object\ndisappearing from the perspective of the autonomous system. Notably, APARATE is\ndesigned to be sensitive to the shape and scale of the target object, and its\ninfluence extends beyond immediate proximity. APARATE, results in a mean depth\nestimation error surpassing $0.5$, significantly impacting as much as $99\\%$ of\nthe targeted region when applied to CNN-based MDE models. Furthermore, it\nyields a significant error of $0.34$ and exerts substantial influence over\n$94\\%$ of the target region in the context of Transformer-based MDE.\n", "date": "2023-03-02", "categories": []}
{"id": "2309.01446", "title": "Open Sesame! Universal Black Box Jailbreaking of Large Language Models", "abstract": "  Large language models (LLMs), designed to provide helpful and safe responses,\noften rely on alignment techniques to align with user intent and social\nguidelines. Unfortunately, this alignment can be exploited by malicious actors\nseeking to manipulate an LLM's outputs for unintended purposes. In this paper\nwe introduce a novel approach that employs a genetic algorithm (GA) to\nmanipulate LLMs when model architecture and parameters are inaccessible. The GA\nattack works by optimizing a universal adversarial prompt that -- when combined\nwith a user's query -- disrupts the attacked model's alignment, resulting in\nunintended and potentially harmful outputs. Our novel approach systematically\nreveals a model's limitations and vulnerabilities by uncovering instances where\nits responses deviate from expected behavior. Through extensive experiments we\ndemonstrate the efficacy of our technique, thus contributing to the ongoing\ndiscussion on responsible AI development by providing a diagnostic tool for\nevaluating and enhancing alignment of LLMs with human intent. To our knowledge\nthis is the first automated universal black box jailbreak attack.\n", "date": "2023-09-04", "categories": []}
{"id": "2312.14302", "title": "Exploiting Novel GPT-4 APIs", "abstract": "  Language model attacks typically assume one of two extreme threat models:\nfull white-box access to model weights, or black-box access limited to a text\ngeneration API. However, real-world APIs are often more flexible than just text\ngeneration: these APIs expose \"gray-box\" access leading to new threat vectors.\nTo explore this, we red-team three new functionalities exposed in the GPT-4\nAPIs: fine-tuning, function calling and knowledge retrieval. We find that\nfine-tuning a model on as few as 15 harmful examples or 100 benign examples can\nremove core safeguards from GPT-4, enabling a range of harmful outputs.\nFurthermore, we find that GPT-4 Assistants readily divulge the function call\nschema and can be made to execute arbitrary function calls. Finally, we find\nthat knowledge retrieval can be hijacked by injecting instructions into\nretrieval documents. These vulnerabilities highlight that any additions to the\nfunctionality exposed by an API can create new vulnerabilities.\n", "date": "2023-12-21", "categories": []}
{"id": "2402.17976", "title": "Enhancing Tracking Robustness with Auxiliary Adversarial Defense\n  Networks", "abstract": "  Adversarial attacks in visual object tracking have significantly degraded the\nperformance of advanced trackers by introducing imperceptible perturbations\ninto images. However, there is still a lack of research on designing\nadversarial defense methods for object tracking. To address these issues, we\npropose an effective auxiliary pre-processing defense network, AADN, which\nperforms defensive transformations on the input images before feeding them into\nthe tracker. Moreover, it can be seamlessly integrated with other visual\ntrackers as a plug-and-play module without parameter adjustments. We train AADN\nusing adversarial training, specifically employing Dua-Loss to generate\nadversarial samples that simultaneously attack the classification and\nregression branches of the tracker. Extensive experiments conducted on the\nOTB100, LaSOT, and VOT2018 benchmarks demonstrate that AADN maintains excellent\ndefense robustness against adversarial attack methods in both adaptive and\nnon-adaptive attack scenarios. Moreover, when transferring the defense network\nto heterogeneous trackers, it exhibits reliable transferability. Finally, AADN\nachieves a processing time of up to 5ms/frame, allowing seamless integration\nwith existing high-speed trackers without introducing significant computational\noverhead.\n", "date": "2024-02-28", "categories": []}
{"id": "2403.11515", "title": "SSAP: A Shape-Sensitive Adversarial Patch for Comprehensive Disruption\n  of Monocular Depth Estimation in Autonomous Navigation Applications", "abstract": "  Monocular depth estimation (MDE) has advanced significantly, primarily\nthrough the integration of convolutional neural networks (CNNs) and more\nrecently, Transformers. However, concerns about their susceptibility to\nadversarial attacks have emerged, especially in safety-critical domains like\nautonomous driving and robotic navigation. Existing approaches for assessing\nCNN-based depth prediction methods have fallen short in inducing comprehensive\ndisruptions to the vision system, often limited to specific local areas. In\nthis paper, we introduce SSAP (Shape-Sensitive Adversarial Patch), a novel\napproach designed to comprehensively disrupt monocular depth estimation (MDE)\nin autonomous navigation applications. Our patch is crafted to selectively\nundermine MDE in two distinct ways: by distorting estimated distances or by\ncreating the illusion of an object disappearing from the system's perspective.\nNotably, our patch is shape-sensitive, meaning it considers the specific shape\nand scale of the target object, thereby extending its influence beyond\nimmediate proximity. Furthermore, our patch is trained to effectively address\ndifferent scales and distances from the camera. Experimental results\ndemonstrate that our approach induces a mean depth estimation error surpassing\n0.5, impacting up to 99% of the targeted region for CNN-based MDE models.\nAdditionally, we investigate the vulnerability of Transformer-based MDE models\nto patch-based attacks, revealing that SSAP yields a significant error of 0.59\nand exerts substantial influence over 99% of the target region on these models.\n", "date": "2024-03-18", "categories": []}
{"id": "2408.01428", "title": "Transferable Adversarial Facial Images for Privacy Protection", "abstract": "  The success of deep face recognition (FR) systems has raised serious privacy\nconcerns due to their ability to enable unauthorized tracking of users in the\ndigital world. Previous studies proposed introducing imperceptible adversarial\nnoises into face images to deceive those face recognition models, thus\nachieving the goal of enhancing facial privacy protection. Nevertheless, they\nheavily rely on user-chosen references to guide the generation of adversarial\nnoises, and cannot simultaneously construct natural and highly transferable\nadversarial face images in black-box scenarios. In light of this, we present a\nnovel face privacy protection scheme with improved transferability while\nmaintain high visual quality. We propose shaping the entire face space directly\ninstead of exploiting one kind of facial characteristic like makeup information\nto integrate adversarial noises. To achieve this goal, we first exploit global\nadversarial latent search to traverse the latent space of the generative model,\nthereby creating natural adversarial face images with high transferability. We\nthen introduce a key landmark regularization module to preserve the visual\nidentity information. Finally, we investigate the impacts of various kinds of\nlatent spaces and find that $\\mathcal{F}$ latent space benefits the trade-off\nbetween visual naturalness and adversarial transferability. Extensive\nexperiments over two datasets demonstrate that our approach significantly\nenhances attack transferability while maintaining high visual quality,\noutperforming state-of-the-art methods by an average 25% improvement in deep FR\nmodels and 10% improvement on commercial FR APIs, including Face++, Aliyun, and\nTencent.\n", "date": "2024-07-18", "categories": []}
{"id": "2408.01596", "title": "Trustworthy Machine Learning under Social and Adversarial Data Sources", "abstract": "  Machine learning has witnessed remarkable breakthroughs in recent years. As\nmachine learning permeates various aspects of daily life, individuals and\norganizations increasingly interact with these systems, exhibiting a wide range\nof social and adversarial behaviors. These behaviors may have a notable impact\non the behavior and performance of machine learning systems. Specifically,\nduring these interactions, data may be generated by strategic individuals,\ncollected by self-interested data collectors, possibly poisoned by adversarial\nattackers, and used to create predictors, models, and policies satisfying\nmultiple objectives. As a result, the machine learning systems' outputs might\ndegrade, such as the susceptibility of deep neural networks to adversarial\nexamples (Shafahi et al., 2018; Szegedy et al., 2013) and the diminished\nperformance of classic algorithms in the presence of strategic individuals\n(Ahmadi et al., 2021). Addressing these challenges is imperative for the\nsuccess of machine learning in societal settings.\n", "date": "2024-08-02", "categories": []}
{"id": "2408.01705", "title": "Downstream Transfer Attack: Adversarial Attacks on Downstream Models\n  with Pre-trained Vision Transformers", "abstract": "  With the advancement of vision transformers (ViTs) and self-supervised\nlearning (SSL) techniques, pre-trained large ViTs have become the new\nfoundation models for computer vision applications. However, studies have shown\nthat, like convolutional neural networks (CNNs), ViTs are also susceptible to\nadversarial attacks, where subtle perturbations in the input can fool the model\ninto making false predictions. This paper studies the transferability of such\nan adversarial vulnerability from a pre-trained ViT model to downstream tasks.\nWe focus on \\emph{sample-wise} transfer attacks and propose a novel attack\nmethod termed \\emph{Downstream Transfer Attack (DTA)}. For a given test image,\nDTA leverages a pre-trained ViT model to craft the adversarial example and then\napplies the adversarial example to attack a fine-tuned version of the model on\na downstream dataset. During the attack, DTA identifies and exploits the most\nvulnerable layers of the pre-trained model guided by a cosine similarity loss\nto craft highly transferable attacks. Through extensive experiments with\npre-trained ViTs by 3 distinct pre-training methods, 3 fine-tuning schemes, and\nacross 10 diverse downstream datasets, we show that DTA achieves an average\nattack success rate (ASR) exceeding 90\\%, surpassing existing methods by a huge\nmargin. When used with adversarial training, the adversarial examples generated\nby our DTA can significantly improve the model's robustness to different\ndownstream transfer attacks.\n", "date": "2024-08-03", "categories": []}
{"id": "2408.01715", "title": "Joint Universal Adversarial Perturbations with Interpretations", "abstract": "  Deep neural networks (DNNs) have significantly boosted the performance of\nmany challenging tasks. Despite the great development, DNNs have also exposed\ntheir vulnerability. Recent studies have shown that adversaries can manipulate\nthe predictions of DNNs by adding a universal adversarial perturbation (UAP) to\nbenign samples. On the other hand, increasing efforts have been made to help\nusers understand and explain the inner working of DNNs by highlighting the most\ninformative parts (i.e., attribution maps) of samples with respect to their\npredictions. Moreover, we first empirically find that such attribution maps\nbetween benign and adversarial examples have a significant discrepancy, which\nhas the potential to detect universal adversarial perturbations for defending\nagainst adversarial attacks. This finding motivates us to further investigate a\nnew research problem: whether there exist universal adversarial perturbations\nthat are able to jointly attack DNNs classifier and its interpretation with\nmalicious desires. It is challenging to give an explicit answer since these two\nobjectives are seemingly conflicting. In this paper, we propose a novel\nattacking framework to generate joint universal adversarial perturbations\n(JUAP), which can fool the DNNs model and misguide the inspection from\ninterpreters simultaneously. Comprehensive experiments on various datasets\ndemonstrate the effectiveness of the proposed method JUAP for joint attacks. To\nthe best of our knowledge, this is the first effort to study UAP for jointly\nattacking both DNNs and interpretations.\n", "date": "2024-08-03", "categories": []}
{"id": "2408.01808", "title": "ALIF: Low-Cost Adversarial Audio Attacks on Black-Box Speech Platforms\n  using Linguistic Features", "abstract": "  Extensive research has revealed that adversarial examples (AE) pose a\nsignificant threat to voice-controllable smart devices. Recent studies have\nproposed black-box adversarial attacks that require only the final\ntranscription from an automatic speech recognition (ASR) system. However, these\nattacks typically involve many queries to the ASR, resulting in substantial\ncosts. Moreover, AE-based adversarial audio samples are susceptible to ASR\nupdates. In this paper, we identify the root cause of these limitations, namely\nthe inability to construct AE attack samples directly around the decision\nboundary of deep learning (DL) models. Building on this observation, we propose\nALIF, the first black-box adversarial linguistic feature-based attack pipeline.\nWe leverage the reciprocal process of text-to-speech (TTS) and ASR models to\ngenerate perturbations in the linguistic embedding space where the decision\nboundary resides. Based on the ALIF pipeline, we present the ALIF-OTL and\nALIF-OTA schemes for launching attacks in both the digital domain and the\nphysical playback environment on four commercial ASRs and voice assistants.\nExtensive evaluations demonstrate that ALIF-OTL and -OTA significantly improve\nquery efficiency by 97.7% and 73.3%, respectively, while achieving competitive\nperformance compared to existing methods. Notably, ALIF-OTL can generate an\nattack sample with only one query. Furthermore, our test-of-time experiment\nvalidates the robustness of our approach against ASR updates.\n", "date": "2024-08-03", "categories": []}
{"id": "2408.01977", "title": "Label Augmentation for Neural Networks Robustness", "abstract": "  Out-of-distribution generalization can be categorized into two types: common\nperturbations arising from natural variations in the real world and adversarial\nperturbations that are intentionally crafted to deceive neural networks. While\ndeep neural networks excel in accuracy under the assumption of identical\ndistributions between training and test data, they often encounter\nout-of-distribution scenarios resulting in a significant decline in accuracy.\nData augmentation methods can effectively enhance robustness against common\ncorruptions, but they typically fall short in improving robustness against\nadversarial perturbations. In this study, we develop Label Augmentation (LA),\nwhich enhances robustness against both common and intentional perturbations and\nimproves uncertainty estimation. Our findings indicate a Clean error rate\nimprovement of up to 23.29% when employing LA in comparisons to the baseline.\nAdditionally, it enhances robustness under common corruptions benchmark by up\nto 24.23%. When tested against FGSM and PGD attacks, improvements in\nadversarial robustness are noticeable, with enhancements of up to 53.18% for\nFGSM and 24.46% for PGD attacks.\n", "date": "2024-08-04", "categories": []}
{"id": "2408.01978", "title": "AdvQDet: Detecting Query-Based Adversarial Attacks with Adversarial\n  Contrastive Prompt Tuning", "abstract": "  Deep neural networks (DNNs) are known to be vulnerable to adversarial attacks\neven under a black-box setting where the adversary can only query the model.\nParticularly, query-based black-box adversarial attacks estimate adversarial\ngradients based on the returned probability vectors of the target model for a\nsequence of queries. During this process, the queries made to the target model\nare intermediate adversarial examples crafted at the previous attack step,\nwhich share high similarities in the pixel space. Motivated by this\nobservation, stateful detection methods have been proposed to detect and reject\nquery-based attacks. While demonstrating promising results, these methods\neither have been evaded by more advanced attacks or suffer from low efficiency\nin terms of the number of shots (queries) required to detect different attacks.\nArguably, the key challenge here is to assign high similarity scores for any\ntwo intermediate adversarial examples perturbed from the same clean image. To\naddress this challenge, we propose a novel Adversarial Contrastive Prompt\nTuning (ACPT) method to robustly fine-tune the CLIP image encoder to extract\nsimilar embeddings for any two intermediate adversarial queries. With ACPT, we\nfurther introduce a detection framework AdvQDet that can detect 7\nstate-of-the-art query-based attacks with $>99\\%$ detection rate within 5\nshots. We also show that ACPT is robust to 3 types of adaptive attacks. Code is\navailable at https://github.com/xinwong/AdvQDet.\n", "date": "2024-08-04", "categories": []}
{"id": "2408.02131", "title": "Model Hijacking Attack in Federated Learning", "abstract": "  Machine learning (ML), driven by prominent paradigms such as centralized and\nfederated learning, has made significant progress in various critical\napplications ranging from autonomous driving to face recognition. However, its\nremarkable success has been accompanied by various attacks. Recently, the model\nhijacking attack has shown that ML models can be hijacked to execute tasks\ndifferent from their original tasks, which increases both accountability and\nparasitic computational risks. Nevertheless, thus far, this attack has only\nfocused on centralized learning. In this work, we broaden the scope of this\nattack to the federated learning domain, where multiple clients collaboratively\ntrain a global model without sharing their data. Specifically, we present\nHijackFL, the first-of-its-kind hijacking attack against the global model in\nfederated learning. The adversary aims to force the global model to perform a\ndifferent task (called hijacking task) from its original task without the\nserver or benign client noticing. To accomplish this, unlike existing methods\nthat use data poisoning to modify the target model's parameters, HijackFL\nsearches for pixel-level perturbations based on their local model (without\nmodifications) to align hijacking samples with the original ones in the feature\nspace. When performing the hijacking task, the adversary applies these cloaks\nto the hijacking samples, compelling the global model to identify them as\noriginal samples and predict them accordingly. We conduct extensive experiments\non four benchmark datasets and three popular models. Empirical results\ndemonstrate that its attack performance outperforms baselines. We further\ninvestigate the factors that affect its performance and discuss possible\ndefenses to mitigate its impact.\n", "date": "2024-08-04", "categories": []}
{"id": "2408.02310", "title": "On the Robustness of Malware Detectors to Adversarial Samples", "abstract": "  Adversarial examples add imperceptible alterations to inputs with the\nobjective to induce misclassification in machine learning models. They have\nbeen demonstrated to pose significant challenges in domains like image\nclassification, with results showing that an adversarially perturbed image to\nevade detection against one classifier is most likely transferable to other\nclassifiers. Adversarial examples have also been studied in malware analysis.\nUnlike images, program binaries cannot be arbitrarily perturbed without\nrendering them non-functional. Due to the difficulty of crafting adversarial\nprogram binaries, there is no consensus on the transferability of adversarially\nperturbed programs to different detectors. In this work, we explore the\nrobustness of malware detectors against adversarially perturbed malware. We\ninvestigate the transferability of adversarial attacks developed against one\ndetector, against other machine learning-based malware detectors, and code\nsimilarity techniques, specifically, locality sensitive hashing-based\ndetectors. Our analysis reveals that adversarial program binaries crafted for\none detector are generally less effective against others. We also evaluate an\nensemble of detectors and show that they can potentially mitigate the impact of\nadversarial program binaries. Finally, we demonstrate that substantial program\nchanges made to evade detection may result in the transformation technique\nbeing identified, implying that the adversary must make minimal changes to the\nprogram binary.\n", "date": "2024-08-05", "categories": []}
{"id": "2310.01651", "title": "Fool Your (Vision and) Language Model With Embarrassingly Simple\n  Permutations", "abstract": "  Large language and vision-language models are rapidly being deployed in\npractice thanks to their impressive capabilities in instruction following,\nin-context learning, and so on. This raises an urgent need to carefully analyse\ntheir robustness so that stakeholders can understand if and when such models\nare trustworthy enough to be relied upon in any given application. In this\npaper, we highlight a specific vulnerability in popular models, namely\npermutation sensitivity in multiple-choice question answering (MCQA).\nSpecifically, we show empirically that popular models are vulnerable to\nadversarial permutation in answer sets for multiple-choice prompting, which is\nsurprising as models should ideally be as invariant to prompt permutation as\nhumans are. These vulnerabilities persist across various model sizes, and exist\nin very recent language and vision-language models. Code is available at\nhttps://github.com/ys-zong/FoolyourVLLMs.\n", "date": "2023-10-02", "categories": []}
{"id": "2408.00023", "title": "On the Perturbed States for Transformed Input-robust Reinforcement\n  Learning", "abstract": "  Reinforcement Learning (RL) agents demonstrating proficiency in a training\nenvironment exhibit vulnerability to adversarial perturbations in input\nobservations during deployment. This underscores the importance of building a\nrobust agent before its real-world deployment. To alleviate the challenging\npoint, prior works focus on developing robust training-based procedures,\nencompassing efforts to fortify the deep neural network component's robustness\nor subject the agent to adversarial training against potent attacks. In this\nwork, we propose a novel method referred to as Transformed Input-robust RL\n(TIRL), which explores another avenue to mitigate the impact of adversaries by\nemploying input transformation-based defenses. Specifically, we introduce two\nprinciples for applying transformation-based defenses in learning robust RL\nagents: (1) autoencoder-styled denoising to reconstruct the original state and\n(2) bounded transformations (bit-depth reduction and vector quantization (VQ))\nto achieve close transformed inputs. The transformations are applied to the\nstate before feeding it into the policy network. Extensive experiments on\nmultiple MuJoCo environments demonstrate that input transformation-based\ndefenses, i.e., VQ, defend against several adversaries in the state\nobservations. The official code is available at\nhttps://github.com/tunglm2203/tirl\n", "date": "2024-07-31", "categories": []}
{"id": "2408.00129", "title": "Vera Verto: Multimodal Hijacking Attack", "abstract": "  The increasing cost of training machine learning (ML) models has led to the\ninclusion of new parties to the training pipeline, such as users who contribute\ntraining data and companies that provide computing resources. This involvement\nof such new parties in the ML training process has introduced new attack\nsurfaces for an adversary to exploit. A recent attack in this domain is the\nmodel hijacking attack, whereby an adversary hijacks a victim model to\nimplement their own -- possibly malicious -- hijacking tasks. However, the\nscope of the model hijacking attack is so far limited to the\nhomogeneous-modality tasks. In this paper, we transform the model hijacking\nattack into a more general multimodal setting, where the hijacking and original\ntasks are performed on data of different modalities. Specifically, we focus on\nthe setting where an adversary implements a natural language processing (NLP)\nhijacking task into an image classification model. To mount the attack, we\npropose a novel encoder-decoder based framework, namely the Blender, which\nrelies on advanced image and language models. Experimental results show that\nour modal hijacking attack achieves strong performances in different settings.\nFor instance, our attack achieves 94%, 94%, and 95% attack success rate when\nusing the Sogou news dataset to hijack STL10, CIFAR-10, and MNIST classifiers.\n", "date": "2024-07-31", "categories": []}
{"id": "2408.00312", "title": "Adversarial Text Rewriting for Text-aware Recommender Systems", "abstract": "  Text-aware recommender systems incorporate rich textual features, such as\ntitles and descriptions, to generate item recommendations for users. The use of\ntextual features helps mitigate cold-start problems, and thus, such recommender\nsystems have attracted increased attention. However, we argue that the\ndependency on item descriptions makes the recommender system vulnerable to\nmanipulation by adversarial sellers on e-commerce platforms. In this paper, we\nexplore the possibility of such manipulation by proposing a new text rewriting\nframework to attack text-aware recommender systems. We show that the rewriting\nattack can be exploited by sellers to unfairly uprank their products, even\nthough the adversarially rewritten descriptions are perceived as realistic by\nhuman evaluators. Methodologically, we investigate two different variations to\ncarry out text rewriting attacks: (1) two-phase fine-tuning for greater attack\nperformance, and (2) in-context learning for higher text rewriting quality.\nExperiments spanning 3 different datasets and 4 existing approaches demonstrate\nthat recommender systems exhibit vulnerability against the proposed text\nrewriting attack. Our work adds to the existing literature around the\nrobustness of recommender systems, while highlighting a new dimension of\nvulnerability in the age of large-scale automated text generation.\n", "date": "2024-08-01", "categories": []}
{"id": "2408.00352", "title": "Autonomous LLM-Enhanced Adversarial Attack for Text-to-Motion", "abstract": "  Human motion generation driven by deep generative models has enabled\ncompelling applications, but the ability of text-to-motion (T2M) models to\nproduce realistic motions from text prompts raises security concerns if\nexploited maliciously. Despite growing interest in T2M, few methods focus on\nsafeguarding these models against adversarial attacks, with existing work on\ntext-to-image models proving insufficient for the unique motion domain. In the\npaper, we propose ALERT-Motion, an autonomous framework leveraging large\nlanguage models (LLMs) to craft targeted adversarial attacks against black-box\nT2M models. Unlike prior methods modifying prompts through predefined rules,\nALERT-Motion uses LLMs' knowledge of human motion to autonomously generate\nsubtle yet powerful adversarial text descriptions. It comprises two key\nmodules: an adaptive dispatching module that constructs an LLM-based agent to\niteratively refine and search for adversarial prompts; and a multimodal\ninformation contrastive module that extracts semantically relevant motion\ninformation to guide the agent's search. Through this LLM-driven approach,\nALERT-Motion crafts adversarial prompts querying victim models to produce\noutputs closely matching targeted motions, while avoiding obvious\nperturbations. Evaluations across popular T2M models demonstrate ALERT-Motion's\nsuperiority over previous methods, achieving higher attack success rates with\nstealthier adversarial prompts. This pioneering work on T2M adversarial attacks\nhighlights the urgency of developing defensive measures as motion generation\ntechnology advances, urging further research into safe and responsible\ndeployment.\n", "date": "2024-08-01", "categories": []}
{"id": "2408.00728", "title": "CERT-ED: Certifiably Robust Text Classification for Edit Distance", "abstract": "  With the growing integration of AI in daily life, ensuring the robustness of\nsystems to inference-time attacks is crucial. Among the approaches for\ncertifying robustness to such adversarial examples, randomized smoothing has\nemerged as highly promising due to its nature as a wrapper around arbitrary\nblack-box models. Previous work on randomized smoothing in natural language\nprocessing has primarily focused on specific subsets of edit distance\noperations, such as synonym substitution or word insertion, without exploring\nthe certification of all edit operations. In this paper, we adapt Randomized\nDeletion (Huang et al., 2023) and propose, CERTified Edit Distance defense\n(CERT-ED) for natural language classification. Through comprehensive\nexperiments, we demonstrate that CERT-ED outperforms the existing Hamming\ndistance method RanMASK (Zeng et al., 2023) in 4 out of 5 datasets in terms of\nboth accuracy and the cardinality of the certificate. By covering various\nthreat models, including 5 direct and 5 transfer attacks, our method improves\nempirical robustness in 38 out of 50 settings.\n", "date": "2024-08-01", "categories": []}
{"id": "2407.21073", "title": "Enhancing Adversarial Text Attacks on BERT Models with Projected\n  Gradient Descent", "abstract": "  Adversarial attacks against deep learning models represent a major threat to\nthe security and reliability of natural language processing (NLP) systems. In\nthis paper, we propose a modification to the BERT-Attack framework, integrating\nProjected Gradient Descent (PGD) to enhance its effectiveness and robustness.\nThe original BERT-Attack, designed for generating adversarial examples against\nBERT-based models, suffers from limitations such as a fixed perturbation budget\nand a lack of consideration for semantic similarity. The proposed approach in\nthis work, PGD-BERT-Attack, addresses these limitations by leveraging PGD to\niteratively generate adversarial examples while ensuring both imperceptibility\nand semantic similarity to the original input. Extensive experiments are\nconducted to evaluate the performance of PGD-BERT-Attack compared to the\noriginal BERT-Attack and other baseline methods. The results demonstrate that\nPGD-BERT-Attack achieves higher success rates in causing misclassification\nwhile maintaining low perceptual changes. Furthermore, PGD-BERT-Attack produces\nadversarial instances that exhibit greater semantic resemblance to the initial\ninput, enhancing their applicability in real-world scenarios. Overall, the\nproposed modification offers a more effective and robust approach to\nadversarial attacks on BERT-based models, thus contributing to the advancement\nof defense against attacks on NLP systems.\n", "date": "2024-07-29", "categories": []}
{"id": "2407.21174", "title": "AI Safety in Practice: Enhancing Adversarial Robustness in Multimodal\n  Image Captioning", "abstract": "  Multimodal machine learning models that combine visual and textual data are\nincreasingly being deployed in critical applications, raising significant\nsafety and security concerns due to their vulnerability to adversarial attacks.\nThis paper presents an effective strategy to enhance the robustness of\nmultimodal image captioning models against such attacks. By leveraging the Fast\nGradient Sign Method (FGSM) to generate adversarial examples and incorporating\nadversarial training techniques, we demonstrate improved model robustness on\ntwo benchmark datasets: Flickr8k and COCO. Our findings indicate that\nselectively training only the text decoder of the multimodal architecture shows\nperformance comparable to full adversarial training while offering increased\ncomputational efficiency. This targeted approach suggests a balance between\nrobustness and training costs, facilitating the ethical deployment of\nmultimodal AI systems across various domains.\n", "date": "2024-07-30", "categories": []}
{"id": "2407.21220", "title": "DeepBaR: Fault Backdoor Attack on Deep Neural Network Layers", "abstract": "  Machine Learning using neural networks has received prominent attention\nrecently because of its success in solving a wide variety of computational\ntasks, in particular in the field of computer vision. However, several works\nhave drawn attention to potential security risks involved with the training and\nimplementation of such networks. In this work, we introduce DeepBaR, a novel\napproach that implants backdoors on neural networks by faulting their behavior\nat training, especially during fine-tuning. Our technique aims to generate\nadversarial samples by optimizing a custom loss function that mimics the\nimplanted backdoors while adding an almost non-visible trigger in the image. We\nattack three popular convolutional neural network architectures and show that\nDeepBaR attacks have a success rate of up to 98.30\\%. Furthermore, DeepBaR does\nnot significantly affect the accuracy of the attacked networks after deployment\nwhen non-malicious inputs are given. Remarkably, DeepBaR allows attackers to\nchoose an input that looks similar to a given class, from a human perspective,\nbut that will be classified as belonging to an arbitrary target class.\n", "date": "2024-07-30", "categories": []}
{"id": "2302.03322", "title": "Attacking Cooperative Multi-Agent Reinforcement Learning by Adversarial\n  Minority Influence", "abstract": "  This study probes the vulnerabilities of cooperative multi-agent\nreinforcement learning (c-MARL) under adversarial attacks, a critical\ndeterminant of c-MARL's worst-case performance prior to real-world\nimplementation. Current observation-based attacks, constrained by white-box\nassumptions, overlook c-MARL's complex multi-agent interactions and cooperative\nobjectives, resulting in impractical and limited attack capabilities. To\naddress these shortcomes, we propose Adversarial Minority Influence (AMI), a\npractical and strong for c-MARL. AMI is a practical black-box attack and can be\nlaunched without knowing victim parameters. AMI is also strong by considering\nthe complex multi-agent interaction and the cooperative goal of agents,\nenabling a single adversarial agent to unilaterally misleads majority victims\nto form targeted worst-case cooperation. This mirrors minority influence\nphenomena in social psychology. To achieve maximum deviation in victim policies\nunder complex agent-wise interactions, our unilateral attack aims to\ncharacterize and maximize the impact of the adversary on the victims. This is\nachieved by adapting a unilateral agent-wise relation metric derived from\nmutual information, thereby mitigating the adverse effects of victim influence\non the adversary. To lead the victims into a jointly detrimental scenario, our\ntargeted attack deceives victims into a long-term, cooperatively harmful\nsituation by guiding each victim towards a specific target, determined through\na trial-and-error process executed by a reinforcement learning agent. Through\nAMI, we achieve the first successful attack against real-world robot swarms and\neffectively fool agents in simulated environments into collectively worst-case\nscenarios, including Starcraft II and Multi-agent Mujoco. The source code and\ndemonstrations can be found at: https://github.com/DIG-Beihang/AMI.\n", "date": "2023-02-07", "categories": []}
{"id": "2407.20653", "title": "FACL-Attack: Frequency-Aware Contrastive Learning for Transferable\n  Adversarial Attacks", "abstract": "  Deep neural networks are known to be vulnerable to security risks due to the\ninherent transferable nature of adversarial examples. Despite the success of\nrecent generative model-based attacks demonstrating strong transferability, it\nstill remains a challenge to design an efficient attack strategy in a\nreal-world strict black-box setting, where both the target domain and model\narchitectures are unknown. In this paper, we seek to explore a feature\ncontrastive approach in the frequency domain to generate adversarial examples\nthat are robust in both cross-domain and cross-model settings. With that goal\nin mind, we propose two modules that are only employed during the training\nphase: a Frequency-Aware Domain Randomization (FADR) module to randomize\ndomain-variant low- and high-range frequency components and a\nFrequency-Augmented Contrastive Learning (FACL) module to effectively separate\ndomain-invariant mid-frequency features of clean and perturbed image. We\ndemonstrate strong transferability of our generated adversarial perturbations\nthrough extensive cross-domain and cross-model experiments, while keeping the\ninference time complexity.\n", "date": "2024-07-30", "categories": []}
{"id": "2310.00607", "title": "Understanding Robust Overfitting from the Feature Generalization\n  Perspective", "abstract": "  Adversarial training (AT) constructs robust neural networks by incorporating\nadversarial perturbations into natural data. However, it is plagued by the\nissue of robust overfitting (RO), which severely damages the model's\nrobustness. In this paper, we investigate RO from a novel feature\ngeneralization perspective. Specifically, we design factor ablation experiments\nto assess the respective impacts of natural data and adversarial perturbations\non RO, identifying that the inducing factor of RO stems from natural data.\nGiven that the only difference between adversarial and natural training lies in\nthe inclusion of adversarial perturbations, we further hypothesize that\nadversarial perturbations degrade the generalization of features in natural\ndata and verify this hypothesis through extensive experiments. Based on these\nfindings, we provide a holistic view of RO from the feature generalization\nperspective and explain various empirical behaviors associated with RO. To\nexamine our feature generalization perspective, we devise two representative\nmethods, attack strength and data augmentation, to prevent the feature\ngeneralization degradation during AT. Extensive experiments conducted on\nbenchmark datasets demonstrate that the proposed methods can effectively\nmitigate RO and enhance adversarial robustness.\n", "date": "2023-10-01", "categories": []}
{"id": "2404.06313", "title": "On adversarial training and the 1 Nearest Neighbor classifier", "abstract": "  The ability to fool deep learning classifiers with tiny perturbations of the\ninput has lead to the development of adversarial training in which the loss\nwith respect to adversarial examples is minimized in addition to the training\nexamples. While adversarial training improves the robustness of the learned\nclassifiers, the procedure is computationally expensive, sensitive to\nhyperparameters and may still leave the classifier vulnerable to other types of\nsmall perturbations. In this paper we compare the performance of adversarial\ntraining to that of the simple 1 Nearest Neighbor (1NN) classifier. We prove\nthat under reasonable assumptions, the 1NN classifier will be robust to {\\em\nany} small image perturbation of the training images. In experiments with 135\ndifferent binary image classification problems taken from CIFAR10, MNIST and\nFashion-MNIST we find that 1NN outperforms TRADES (a powerful adversarial\ntraining algorithm) in terms of average adversarial accuracy. In additional\nexperiments with 69 robust models taken from the current adversarial robustness\nleaderboard, we find that 1NN outperforms almost all of them in terms of\nrobustness to perturbations that are only slightly different from those used\nduring training. Taken together, our results suggest that modern adversarial\ntraining methods still fall short of the robustness of the simple 1NN\nclassifier. our code can be found at\n\\url{https://github.com/amirhagai/On-Adversarial-Training-And-The-1-Nearest-Neighbor-Classifier}\n\\keywords{Adversarial training}\n", "date": "2024-04-09", "categories": []}
{"id": "2407.12068", "title": "Learning on Graphs with Large Language Models(LLMs): A Deep Dive into\n  Model Robustness", "abstract": "  Large Language Models (LLMs) have demonstrated remarkable performance across\nvarious natural language processing tasks. Recently, several LLMs-based\npipelines have been developed to enhance learning on graphs with text\nattributes, showcasing promising performance. However, graphs are well-known to\nbe susceptible to adversarial attacks and it remains unclear whether LLMs\nexhibit robustness in learning on graphs. To address this gap, our work aims to\nexplore the potential of LLMs in the context of adversarial attacks on graphs.\nSpecifically, we investigate the robustness against graph structural and\ntextual perturbations in terms of two dimensions: LLMs-as-Enhancers and\nLLMs-as-Predictors. Through extensive experiments, we find that, compared to\nshallow models, both LLMs-as-Enhancers and LLMs-as-Predictors offer superior\nrobustness against structural and textual attacks.Based on these findings, we\ncarried out additional analyses to investigate the underlying causes.\nFurthermore, we have made our benchmark library openly available to facilitate\nquick and fair evaluations, and to encourage ongoing innovative research in\nthis field.\n", "date": "2024-07-16", "categories": []}
{"id": "2407.19155", "title": "Debiased Graph Poisoning Attack via Contrastive Surrogate Objective", "abstract": "  Graph neural networks (GNN) are vulnerable to adversarial attacks, which aim\nto degrade the performance of GNNs through imperceptible changes on the graph.\nHowever, we find that in fact the prevalent meta-gradient-based attacks, which\nutilizes the gradient of the loss w.r.t the adjacency matrix, are biased\ntowards training nodes. That is, their meta-gradient is determined by a\ntraining procedure of the surrogate model, which is solely trained on the\ntraining nodes. This bias manifests as an uneven perturbation, connecting two\nnodes when at least one of them is a labeled node, i.e., training node, while\nit is unlikely to connect two unlabeled nodes. However, these biased attack\napproaches are sub-optimal as they do not consider flipping edges between two\nunlabeled nodes at all. This means that they miss the potential attacked edges\nbetween unlabeled nodes that significantly alter the representation of a node.\nIn this paper, we investigate the meta-gradients to uncover the root cause of\nthe uneven perturbations of existing attacks. Based on our analysis, we propose\na Meta-gradient-based attack method using contrastive surrogate objective\n(Metacon), which alleviates the bias in meta-gradient using a new surrogate\nloss. We conduct extensive experiments to show that Metacon outperforms\nexisting meta gradient-based attack methods through benchmark datasets, while\nshowing that alleviating the bias towards training nodes is effective in\nattacking the graph structure.\n", "date": "2024-07-27", "categories": []}
{"id": "2407.19981", "title": "Adversarial Robustness in RGB-Skeleton Action Recognition: Leveraging\n  Attention Modality Reweighter", "abstract": "  Deep neural networks (DNNs) have been applied in many computer vision tasks\nand achieved state-of-the-art (SOTA) performance. However, misclassification\nwill occur when DNNs predict adversarial examples which are created by adding\nhuman-imperceptible adversarial noise to natural examples. This limits the\napplication of DNN in security-critical fields. In order to enhance the\nrobustness of models, previous research has primarily focused on the unimodal\ndomain, such as image recognition and video understanding. Although multi-modal\nlearning has achieved advanced performance in various tasks, such as action\nrecognition, research on the robustness of RGB-skeleton action recognition\nmodels is scarce. In this paper, we systematically investigate how to improve\nthe robustness of RGB-skeleton action recognition models. We initially\nconducted empirical analysis on the robustness of different modalities and\nobserved that the skeleton modality is more robust than the RGB modality.\nMotivated by this observation, we propose the \\formatword{A}ttention-based\n\\formatword{M}odality \\formatword{R}eweighter (\\formatword{AMR}), which\nutilizes an attention layer to re-weight the two modalities, enabling the model\nto learn more robust features. Our AMR is plug-and-play, allowing easy\nintegration with multimodal models. To demonstrate the effectiveness of AMR, we\nconducted extensive experiments on various datasets. For example, compared to\nthe SOTA methods, AMR exhibits a 43.77\\% improvement against PGD20 attacks on\nthe NTU-RGB+D 60 dataset. Furthermore, it effectively balances the differences\nin robustness between different modalities.\n", "date": "2024-07-29", "categories": []}
{"id": "2407.20141", "title": "DDAP: Dual-Domain Anti-Personalization against Text-to-Image Diffusion\n  Models", "abstract": "  Diffusion-based personalized visual content generation technologies have\nachieved significant breakthroughs, allowing for the creation of specific\nobjects by just learning from a few reference photos. However, when misused to\nfabricate fake news or unsettling content targeting individuals, these\ntechnologies could cause considerable societal harm. To address this problem,\ncurrent methods generate adversarial samples by adversarially maximizing the\ntraining loss, thereby disrupting the output of any personalized generation\nmodel trained with these samples. However, the existing methods fail to achieve\neffective defense and maintain stealthiness, as they overlook the intrinsic\nproperties of diffusion models. In this paper, we introduce a novel Dual-Domain\nAnti-Personalization framework (DDAP). Specifically, we have developed Spatial\nPerturbation Learning (SPL) by exploiting the fixed and perturbation-sensitive\nnature of the image encoder in personalized generation. Subsequently, we have\ndesigned a Frequency Perturbation Learning (FPL) method that utilizes the\ncharacteristics of diffusion models in the frequency domain. The SPL disrupts\nthe overall texture of the generated images, while the FPL focuses on image\ndetails. By alternating between these two methods, we construct the DDAP\nframework, effectively harnessing the strengths of both domains. To further\nenhance the visual quality of the adversarial samples, we design a localization\nmodule to accurately capture attentive areas while ensuring the effectiveness\nof the attack and avoiding unnecessary disturbances in the background.\nExtensive experiments on facial benchmarks have shown that the proposed DDAP\nenhances the disruption of personalized generation models while also\nmaintaining high quality in adversarial samples, making it more effective in\nprotecting privacy in practical applications.\n", "date": "2024-07-29", "categories": []}
{"id": "2312.09554", "title": "Embodied Laser Attack:Leveraging Scene Priors to Achieve Agent-based\n  Robust Non-contact Attacks", "abstract": "  As physical adversarial attacks become extensively applied in unearthing the\npotential risk of security-critical scenarios, especially in dynamic scenarios,\ntheir vulnerability to environmental variations has also been brought to light.\nThe non-robust nature of physical adversarial attack methods brings\nless-than-stable performance consequently. Although methods such as EOT have\nenhanced the robustness of traditional contact attacks like adversarial\npatches, they fall short in practicality and concealment within dynamic\nenvironments such as traffic scenarios. Meanwhile, non-contact laser attacks,\nwhile offering enhanced adaptability, face constraints due to a limited\noptimization space for their attributes, rendering EOT less effective. This\nlimitation underscores the necessity for developing a new strategy to augment\nthe robustness of such practices. To address these issues, this paper\nintroduces the Embodied Laser Attack (ELA), a novel framework that leverages\nthe embodied intelligence paradigm of Perception-Decision-Control to\ndynamically tailor non-contact laser attacks. For the perception module, given\nthe challenge of simulating the victim's view by full-image transformation, ELA\nhas innovatively developed a local perspective transformation network, based on\nthe intrinsic prior knowledge of traffic scenes and enables effective and\nefficient estimation. For the decision and control module, ELA trains an attack\nagent with data-driven reinforcement learning instead of adopting\ntime-consuming heuristic algorithms, making it capable of instantaneously\ndetermining a valid attack strategy with the perceived information by\nwell-designed rewards, which is then conducted by a controllable laser emitter.\nExperimentally, we apply our framework to diverse traffic scenarios both in the\ndigital and physical world, verifying the effectiveness of our method under\ndynamic successive scenes.\n", "date": "2023-12-15", "categories": []}
{"id": "2405.20584", "title": "Disrupting Diffusion: Token-Level Attention Erasure Attack against\n  Diffusion-based Customization", "abstract": "  With the development of diffusion-based customization methods like\nDreamBooth, individuals now have access to train the models that can generate\ntheir personalized images. Despite the convenience, malicious users have\nmisused these techniques to create fake images, thereby triggering a privacy\nsecurity crisis. In light of this, proactive adversarial attacks are proposed\nto protect users against customization. The adversarial examples are trained to\ndistort the customization model's outputs and thus block the misuse. In this\npaper, we propose DisDiff (Disrupting Diffusion), a novel adversarial attack\nmethod to disrupt the diffusion model outputs. We first delve into the\nintrinsic image-text relationships, well-known as cross-attention, and\nempirically find that the subject-identifier token plays an important role in\nguiding image generation. Thus, we propose the Cross-Attention Erasure module\nto explicitly \"erase\" the indicated attention maps and disrupt the text\nguidance. Besides,we analyze the influence of the sampling process of the\ndiffusion model on Projected Gradient Descent (PGD) attack and introduce a\nnovel Merit Sampling Scheduler to adaptively modulate the perturbation updating\namplitude in a step-aware manner. Our DisDiff outperforms the state-of-the-art\nmethods by 12.75% of FDFR scores and 7.25% of ISM scores across two facial\nbenchmarks and two commonly used prompts on average.\n", "date": "2024-05-31", "categories": []}
{"id": "2407.18632", "title": "Robust VAEs via Generating Process of Noise Augmented Data", "abstract": "  Advancing defensive mechanisms against adversarial attacks in generative\nmodels is a critical research topic in machine learning. Our study focuses on a\nspecific type of generative models - Variational Auto-Encoders (VAEs). Contrary\nto common beliefs and existing literature which suggest that noise injection\ntowards training data can make models more robust, our preliminary experiments\nrevealed that naive usage of noise augmentation technique did not substantially\nimprove VAE robustness. In fact, it even degraded the quality of learned\nrepresentations, making VAEs more susceptible to adversarial perturbations.\nThis paper introduces a novel framework that enhances robustness by\nregularizing the latent space divergence between original and noise-augmented\ndata. Through incorporating a paired probabilistic prior into the standard\nvariational lower bound, our method significantly boosts defense against\nadversarial attacks. Our empirical evaluations demonstrate that this approach,\ntermed Robust Augmented Variational Auto-ENcoder (RAVEN), yields superior\nperformance in resisting adversarial inputs on widely-recognized benchmark\ndatasets.\n", "date": "2024-07-26", "categories": []}
{"id": "2407.18658", "title": "Adversarial Robustification via Text-to-Image Diffusion Models", "abstract": "  Adversarial robustness has been conventionally believed as a challenging\nproperty to encode for neural networks, requiring plenty of training data. In\nthe recent paradigm of adopting off-the-shelf models, however, access to their\ntraining data is often infeasible or not practical, while most of such models\nare not originally trained concerning adversarial robustness. In this paper, we\ndevelop a scalable and model-agnostic solution to achieve adversarial\nrobustness without using any data. Our intuition is to view recent\ntext-to-image diffusion models as \"adaptable\" denoisers that can be optimized\nto specify target tasks. Based on this, we propose: (a) to initiate a\ndenoise-and-classify pipeline that offers provable guarantees against\nadversarial attacks, and (b) to leverage a few synthetic reference images\ngenerated from the text-to-image model that enables novel adaptation schemes.\nOur experiments show that our data-free scheme applied to the pre-trained CLIP\ncould improve the (provable) adversarial robustness of its diverse zero-shot\nclassification derivatives (while maintaining their accuracy), significantly\nsurpassing prior approaches that utilize the full training data. Not only for\nCLIP, we also demonstrate that our framework is easily applicable for\nrobustifying other visual classifiers efficiently.\n", "date": "2024-07-26", "categories": []}
{"id": "2302.00509", "title": "Exploring Semantic Perturbations on Grover", "abstract": "  With news and information being as easy to access as they currently are, it\nis more important than ever to ensure that people are not mislead by what they\nread. Recently, the rise of neural fake news (AI-generated fake news) and its\ndemonstrated effectiveness at fooling humans has prompted the development of\nmodels to detect it. One such model is the Grover model, which can both detect\nneural fake news to prevent it, and generate it to demonstrate how a model\ncould be misused to fool human readers. In this work we explore the Grover\nmodel's fake news detection capabilities by performing targeted attacks through\nperturbations on input news articles. Through this we test Grover's resilience\nto these adversarial attacks and expose some potential vulnerabilities which\nshould be addressed in further iterations to ensure it can detect all types of\nfake news accurately.\n", "date": "2023-02-01", "categories": []}
{"id": "2407.17587", "title": "S-E Pipeline: A Vision Transformer (ViT) based Resilient Classification\n  Pipeline for Medical Imaging Against Adversarial Attacks", "abstract": "  Vision Transformer (ViT) is becoming widely popular in automating accurate\ndisease diagnosis in medical imaging owing to its robust self-attention\nmechanism. However, ViTs remain vulnerable to adversarial attacks that may\nthwart the diagnosis process by leading it to intentional misclassification of\ncritical disease. In this paper, we propose a novel image classification\npipeline, namely, S-E Pipeline, that performs multiple pre-processing steps\nthat allow ViT to be trained on critical features so as to reduce the impact of\ninput perturbations by adversaries. Our method uses a combination of\nsegmentation and image enhancement techniques such as Contrast Limited Adaptive\nHistogram Equalization (CLAHE), Unsharp Masking (UM), and High-Frequency\nEmphasis filtering (HFE) as preprocessing steps to identify critical features\nthat remain intact even after adversarial perturbations. The experimental study\ndemonstrates that our novel pipeline helps in reducing the effect of\nadversarial attacks by 72.22% for the ViT-b32 model and 86.58% for the ViT-l32\nmodel. Furthermore, we have shown an end-to-end deployment of our proposed\nmethod on the NVIDIA Jetson Orin Nano board to demonstrate its practical use\ncase in modern hand-held devices that are usually resource-constrained.\n", "date": "2024-07-23", "categories": []}
{"id": "2407.17797", "title": "A Unified Understanding of Adversarial Vulnerability Regarding Unimodal\n  Models and Vision-Language Pre-training Models", "abstract": "  With Vision-Language Pre-training (VLP) models demonstrating powerful\nmultimodal interaction capabilities, the application scenarios of neural\nnetworks are no longer confined to unimodal domains but have expanded to more\ncomplex multimodal V+L downstream tasks. The security vulnerabilities of\nunimodal models have been extensively examined, whereas those of VLP models\nremain challenging. We note that in CV models, the understanding of images\ncomes from annotated information, while VLP models are designed to learn image\nrepresentations directly from raw text. Motivated by this discrepancy, we\ndeveloped the Feature Guidance Attack (FGA), a novel method that uses text\nrepresentations to direct the perturbation of clean images, resulting in the\ngeneration of adversarial images. FGA is orthogonal to many advanced attack\nstrategies in the unimodal domain, facilitating the direct application of rich\nresearch findings from the unimodal to the multimodal scenario. By\nappropriately introducing text attack into FGA, we construct Feature Guidance\nwith Text Attack (FGA-T). Through the interaction of attacking two modalities,\nFGA-T achieves superior attack effects against VLP models. Moreover,\nincorporating data augmentation and momentum mechanisms significantly improves\nthe black-box transferability of FGA-T. Our method demonstrates stable and\neffective attack capabilities across various datasets, downstream tasks, and\nboth black-box and white-box settings, offering a unified baseline for\nexploring the robustness of VLP models.\n", "date": "2024-07-25", "categories": []}
{"id": "2407.18251", "title": "Sparse vs Contiguous Adversarial Pixel Perturbations in Multimodal\n  Models: An Empirical Analysis", "abstract": "  Assessing the robustness of multimodal models against adversarial examples is\nan important aspect for the safety of its users. We craft L0-norm perturbation\nattacks on the preprocessed input images. We launch them in a black-box setup\nagainst four multimodal models and two unimodal DNNs, considering both targeted\nand untargeted misclassification. Our attacks target less than 0.04% of\nperturbed image area and integrate different spatial positioning of perturbed\npixels: sparse positioning and pixels arranged in different contiguous shapes\n(row, column, diagonal, and patch). To the best of our knowledge, we are the\nfirst to assess the robustness of three state-of-the-art multimodal models\n(ALIGN, AltCLIP, GroupViT) against different sparse and contiguous pixel\ndistribution perturbations. The obtained results indicate that unimodal DNNs\nare more robust than multimodal models. Furthermore, models using CNN-based\nImage Encoder are more vulnerable than models with ViT - for untargeted\nattacks, we obtain a 99% success rate by perturbing less than 0.02% of the\nimage area.\n", "date": "2024-07-25", "categories": []}
{"id": "2309.16452", "title": "On the Trade-offs between Adversarial Robustness and Actionable\n  Explanations", "abstract": "  As machine learning models are increasingly being employed in various\nhigh-stakes settings, it becomes important to ensure that predictions of these\nmodels are not only adversarially robust, but also readily explainable to\nrelevant stakeholders. However, it is unclear if these two notions can be\nsimultaneously achieved or if there exist trade-offs between them. In this\nwork, we make one of the first attempts at studying the impact of adversarially\nrobust models on actionable explanations which provide end users with a means\nfor recourse. We theoretically and empirically analyze the cost (ease of\nimplementation) and validity (probability of obtaining a positive model\nprediction) of recourses output by state-of-the-art algorithms when the\nunderlying models are adversarially robust vs. non-robust. More specifically,\nwe derive theoretical bounds on the differences between the cost and the\nvalidity of the recourses generated by state-of-the-art algorithms for\nadversarially robust vs. non-robust linear and non-linear models. Our empirical\nresults with multiple real-world datasets validate our theoretical results and\nshow the impact of varying degrees of model robustness on the cost and validity\nof the resulting recourses. Our analyses demonstrate that adversarially robust\nmodels significantly increase the cost and reduce the validity of the resulting\nrecourses, thus shedding light on the inherent trade-offs between adversarial\nrobustness and actionable explanations.\n", "date": "2023-09-28", "categories": []}
{"id": "2402.16479", "title": "Edge Detectors Can Make Deep Convolutional Neural Networks More Robust", "abstract": "  Deep convolutional neural networks (DCNN for short) are vulnerable to\nexamples with small perturbations. Improving DCNN's robustness is of great\nsignificance to the safety-critical applications, such as autonomous driving\nand industry automation. Inspired by the principal way that human eyes\nrecognize objects, i.e., largely relying on the shape features, this paper\nfirst employs the edge detectors as layer kernels and designs a binary edge\nfeature branch (BEFB for short) to learn the binary edge features, which can be\neasily integrated into any popular backbone. The four edge detectors can learn\nthe horizontal, vertical, positive diagonal, and negative diagonal edge\nfeatures, respectively, and the branch is stacked by multiple Sobel layers\n(using edge detectors as kernels) and one threshold layer. The binary edge\nfeatures learned by the branch, concatenated with the texture features learned\nby the backbone, are fed into the fully connected layers for classification. We\nintegrate the proposed branch into VGG16 and ResNet34, respectively, and\nconduct experiments on multiple datasets. Experimental results demonstrate the\nBEFB is lightweight and has no side effects on training. And the accuracy of\nthe BEFB integrated models is better than the original ones on all datasets\nwhen facing FGSM, PGD, and C\\&W attacks. Besides, BEFB integrated models\nequipped with the robustness enhancing techniques can achieve better\nclassification accuracy compared to the original models. The work in this paper\nfor the first time shows it is feasible to enhance the robustness of DCNNs\nthrough combining both shape-like features and texture features.\n", "date": "2024-02-26", "categories": []}
{"id": "2405.14077", "title": "Learning to Transform Dynamically for Better Adversarial Transferability", "abstract": "  Adversarial examples, crafted by adding perturbations imperceptible to\nhumans, can deceive neural networks. Recent studies identify the adversarial\ntransferability across various models, \\textit{i.e.}, the cross-model attack\nability of adversarial samples. To enhance such adversarial transferability,\nexisting input transformation-based methods diversify input data with\ntransformation augmentation. However, their effectiveness is limited by the\nfinite number of available transformations. In our study, we introduce a novel\napproach named Learning to Transform (L2T). L2T increases the diversity of\ntransformed images by selecting the optimal combination of operations from a\npool of candidates, consequently improving adversarial transferability. We\nconceptualize the selection of optimal transformation combinations as a\ntrajectory optimization problem and employ a reinforcement learning strategy to\neffectively solve the problem. Comprehensive experiments on the ImageNet\ndataset, as well as practical tests with Google Vision and GPT-4V, reveal that\nL2T surpasses current methodologies in enhancing adversarial transferability,\nthereby confirming its effectiveness and practical significance. The code is\navailable at https://github.com/RongyiZhu/L2T.\n", "date": "2024-05-23", "categories": []}
{"id": "2407.17312", "title": "Physical Adversarial Attack on Monocular Depth Estimation via\n  Shape-Varying Patches", "abstract": "  Adversarial attacks against monocular depth estimation (MDE) systems pose\nsignificant challenges, particularly in safety-critical applications such as\nautonomous driving. Existing patch-based adversarial attacks for MDE are\nconfined to the vicinity of the patch, making it difficult to affect the entire\ntarget. To address this limitation, we propose a physics-based adversarial\nattack on monocular depth estimation, employing a framework called Attack with\nShape-Varying Patches (ASP), aiming to optimize patch content, shape, and\nposition to maximize effectiveness. We introduce various mask shapes, including\nquadrilateral, rectangular, and circular masks, to enhance the flexibility and\nefficiency of the attack. Furthermore, we propose a new loss function to extend\nthe influence of the patch beyond the overlapping regions. Experimental results\ndemonstrate that our attack method generates an average depth error of 18\nmeters on the target car with a patch area of 1/9, affecting over 98\\% of the\ntarget area.\n", "date": "2024-07-24", "categories": []}
{"id": "2403.05422", "title": "EVD4UAV: An Altitude-Sensitive Benchmark to Evade Vehicle Detection in\n  UAV", "abstract": "  Vehicle detection in Unmanned Aerial Vehicle (UAV) captured images has wide\napplications in aerial photography and remote sensing. There are many public\nbenchmark datasets proposed for the vehicle detection and tracking in UAV\nimages. Recent studies show that adding an adversarial patch on objects can\nfool the well-trained deep neural networks based object detectors, posing\nsecurity concerns to the downstream tasks. However, the current public UAV\ndatasets might ignore the diverse altitudes, vehicle attributes, fine-grained\ninstance-level annotation in mostly side view with blurred vehicle roof, so\nnone of them is good to study the adversarial patch based vehicle detection\nattack problem. In this paper, we propose a new dataset named EVD4UAV as an\naltitude-sensitive benchmark to evade vehicle detection in UAV with 6,284\nimages and 90,886 fine-grained annotated vehicles. The EVD4UAV dataset has\ndiverse altitudes (50m, 70m, 90m), vehicle attributes (color, type),\nfine-grained annotation (horizontal and rotated bounding boxes, instance-level\nmask) in top view with clear vehicle roof. One white-box and two black-box\npatch based attack methods are implemented to attack three classic deep neural\nnetworks based object detectors on EVD4UAV. The experimental results show that\nthese representative attack methods could not achieve the robust\naltitude-insensitive attack performance.\n", "date": "2024-03-08", "categories": []}
{"id": "2405.16341", "title": "R.A.C.E.: Robust Adversarial Concept Erasure for Secure Text-to-Image\n  Diffusion Model", "abstract": "  In the evolving landscape of text-to-image (T2I) diffusion models, the\nremarkable capability to generate high-quality images from textual descriptions\nfaces challenges with the potential misuse of reproducing sensitive content. To\naddress this critical issue, we introduce \\textbf{R}obust \\textbf{A}dversarial\n\\textbf{C}oncept \\textbf{E}rase (RACE), a novel approach designed to mitigate\nthese risks by enhancing the robustness of concept erasure method for T2I\nmodels. RACE utilizes a sophisticated adversarial training framework to\nidentify and mitigate adversarial text embeddings, significantly reducing the\nAttack Success Rate (ASR). Impressively, RACE achieves a 30 percentage point\nreduction in ASR for the ``nudity'' concept against the leading white-box\nattack method. Our extensive evaluations demonstrate RACE's effectiveness in\ndefending against both white-box and black-box attacks, marking a significant\nadvancement in protecting T2I diffusion models from generating inappropriate or\nmisleading imagery. This work underlines the essential need for proactive\ndefense measures in adapting to the rapidly advancing field of adversarial\nchallenges. Our code is publicly available:\n\\url{https://github.com/chkimmmmm/R.A.C.E.}\n", "date": "2024-05-25", "categories": []}
{"id": "2407.16327", "title": "Understanding Impacts of Electromagnetic Signal Injection Attacks on\n  Object Detection", "abstract": "  Object detection can localize and identify objects in images, and it is\nextensively employed in critical multimedia applications such as security\nsurveillance and autonomous driving. Despite the success of existing object\ndetection models, they are often evaluated in ideal scenarios where captured\nimages guarantee the accurate and complete representation of the detecting\nscenes. However, images captured by image sensors may be affected by different\nfactors in real applications, including cyber-physical attacks. In particular,\nattackers can exploit hardware properties within the systems to inject\nelectromagnetic interference so as to manipulate the images. Such attacks can\ncause noisy or incomplete information about the captured scene, leading to\nincorrect detection results, potentially granting attackers malicious control\nover critical functions of the systems. This paper presents a research work\nthat comprehensively quantifies and analyzes the impacts of such attacks on\nstate-of-the-art object detection models in practice. It also sheds light on\nthe underlying reasons for the incorrect detection outcomes.\n", "date": "2024-07-23", "categories": []}
{"id": "2301.12554", "title": "Improving the Accuracy-Robustness Trade-Off of Classifiers via Adaptive\n  Smoothing", "abstract": "  While prior research has proposed a plethora of methods that build neural\nclassifiers robust against adversarial robustness, practitioners are still\nreluctant to adopt them due to their unacceptably severe clean accuracy\npenalties. This paper significantly alleviates this accuracy-robustness\ntrade-off by mixing the output probabilities of a standard classifier and a\nrobust classifier, where the standard network is optimized for clean accuracy\nand is not robust in general. We show that the robust base classifier's\nconfidence difference for correct and incorrect examples is the key to this\nimprovement. In addition to providing intuitions and empirical evidence, we\ntheoretically certify the robustness of the mixed classifier under realistic\nassumptions. Furthermore, we adapt an adversarial input detector into a mixing\nnetwork that adaptively adjusts the mixture of the two base models, further\nreducing the accuracy penalty of achieving robustness. The proposed flexible\nmethod, termed \"adaptive smoothing\", can work in conjunction with existing or\neven future methods that improve clean accuracy, robustness, or adversary\ndetection. Our empirical evaluation considers strong attack methods, including\nAutoAttack and adaptive attack. On the CIFAR-100 dataset, our method achieves\nan 85.21% clean accuracy while maintaining a 38.72% $\\ell_\\infty$-AutoAttacked\n($\\epsilon = 8/255$) accuracy, becoming the second most robust method on the\nRobustBench CIFAR-100 benchmark as of submission, while improving the clean\naccuracy by ten percentage points compared with all listed models. The code\nthat implements our method is available at\nhttps://github.com/Bai-YT/AdaptiveSmoothing.\n", "date": "2023-01-29", "categories": []}
{"id": "2401.08734", "title": "Bag of Tricks to Boost Adversarial Transferability", "abstract": "  Deep neural networks are widely known to be vulnerable to adversarial\nexamples. However, vanilla adversarial examples generated under the white-box\nsetting often exhibit low transferability across different models. Since\nadversarial transferability poses more severe threats to practical\napplications, various approaches have been proposed for better transferability,\nincluding gradient-based, input transformation-based, and model-related\nattacks, \\etc. In this work, we find that several tiny changes in the existing\nadversarial attacks can significantly affect the attack performance, \\eg, the\nnumber of iterations and step size. Based on careful studies of existing\nadversarial attacks, we propose a bag of tricks to enhance adversarial\ntransferability, including momentum initialization, scheduled step size, dual\nexample, spectral-based input transformation, and several ensemble strategies.\nExtensive experiments on the ImageNet dataset validate the high effectiveness\nof our proposed tricks and show that combining them can further boost\nadversarial transferability. Our work provides practical insights and\ntechniques to enhance adversarial transferability, and offers guidance to\nimprove the attack performance on the real-world application through simple\nadjustments.\n", "date": "2024-01-16", "categories": []}
{"id": "2401.17038", "title": "Towards Assessing the Synthetic-to-Measured Adversarial Vulnerability of\n  SAR ATR", "abstract": "  Recently, there has been increasing concern about the vulnerability of deep\nneural network (DNN)-based synthetic aperture radar (SAR) automatic target\nrecognition (ATR) to adversarial attacks, where a DNN could be easily deceived\nby clean input with imperceptible but aggressive perturbations. This paper\nstudies the synthetic-to-measured (S2M) transfer setting, where an attacker\ngenerates adversarial perturbation based solely on synthetic data and transfers\nit against victim models trained with measured data. Compared with the current\nmeasured-to-measured (M2M) transfer setting, our approach does not need direct\naccess to the victim model or the measured SAR data. We also propose the\ntransferability estimation attack (TEA) to uncover the adversarial risks in\nthis more challenging and practical scenario. The TEA makes full use of the\nlimited similarity between the synthetic and measured data pairs for blind\nestimation and optimization of S2M transferability, leading to feasible\nsurrogate model enhancement without mastering the victim model and data.\nComprehensive evaluations based on the publicly available synthetic and\nmeasured paired labeled experiment (SAMPLE) dataset demonstrate that the TEA\noutperforms state-of-the-art methods and can significantly enhance various\nattack algorithms in computer vision and remote sensing applications. Codes and\ndata are available at https://github.com/scenarri/S2M-TEA.\n", "date": "2024-01-30", "categories": []}
{"id": "2402.08991", "title": "Towards Robust Model-Based Reinforcement Learning Against Adversarial\n  Corruption", "abstract": "  This study tackles the challenges of adversarial corruption in model-based\nreinforcement learning (RL), where the transition dynamics can be corrupted by\nan adversary. Existing studies on corruption-robust RL mostly focus on the\nsetting of model-free RL, where robust least-square regression is often\nemployed for value function estimation. However, these techniques cannot be\ndirectly applied to model-based RL. In this paper, we focus on model-based RL\nand take the maximum likelihood estimation (MLE) approach to learn transition\nmodel. Our work encompasses both online and offline settings. In the online\nsetting, we introduce an algorithm called corruption-robust optimistic MLE\n(CR-OMLE), which leverages total-variation (TV)-based information ratios as\nuncertainty weights for MLE. We prove that CR-OMLE achieves a regret of\n$\\tilde{\\mathcal{O}}(\\sqrt{T} + C)$, where $C$ denotes the cumulative\ncorruption level after $T$ episodes. We also prove a lower bound to show that\nthe additive dependence on $C$ is optimal. We extend our weighting technique to\nthe offline setting, and propose an algorithm named corruption-robust\npessimistic MLE (CR-PMLE). Under a uniform coverage condition, CR-PMLE exhibits\nsuboptimality worsened by $\\mathcal{O}(C/n)$, nearly matching the lower bound.\nTo the best of our knowledge, this is the first work on corruption-robust\nmodel-based RL algorithms with provable guarantees.\n", "date": "2024-02-14", "categories": []}
{"id": "2403.05181", "title": "Adversarial Sparse Teacher: Defense Against Distillation-Based Model\n  Stealing Attacks Using Adversarial Examples", "abstract": "  We introduce Adversarial Sparse Teacher (AST), a robust defense method\nagainst distillation-based model stealing attacks. Our approach trains a\nteacher model using adversarial examples to produce sparse logit responses and\nincrease the entropy of the output distribution. Typically, a model generates a\npeak in its output corresponding to its prediction. By leveraging adversarial\nexamples, AST modifies the teacher model's original response, embedding a few\naltered logits into the output while keeping the primary response slightly\nhigher. Concurrently, all remaining logits are elevated to further increase the\noutput distribution's entropy. All these complex manipulations are performed\nusing an optimization function with our proposed Exponential Predictive\nDivergence (EPD) loss function. EPD allows us to maintain higher entropy levels\ncompared to traditional KL divergence, effectively confusing attackers.\nExperiments on CIFAR-10 and CIFAR-100 datasets demonstrate that AST outperforms\nstate-of-the-art methods, providing effective defense against model stealing\nwhile preserving high accuracy. The source codes will be made publicly\navailable here soon.\n", "date": "2024-03-08", "categories": []}
{"id": "2407.14684", "title": "Data Poisoning: An Overlooked Threat to Power Grid Resilience", "abstract": "  As the complexities of Dynamic Data Driven Applications Systems increase,\npreserving their resilience becomes more challenging. For instance, maintaining\npower grid resilience is becoming increasingly complicated due to the growing\nnumber of stochastic variables (such as renewable outputs) and extreme weather\nevents that add uncertainty to the grid. Current optimization methods have\nstruggled to accommodate this rise in complexity. This has fueled the growing\ninterest in data-driven methods used to operate the grid, leading to more\nvulnerability to cyberattacks. One such disruption that is commonly discussed\nis the adversarial disruption, where the intruder attempts to add a small\nperturbation to input data in order to \"manipulate\" the system operation.\nDuring the last few years, work on adversarial training and disruptions on the\npower system has gained popularity. In this paper, we will first review these\napplications, specifically on the most common types of adversarial disruptions:\nevasion and poisoning disruptions. Through this review, we highlight the gap\nbetween poisoning and evasion research when applied to the power grid. This is\ndue to the underlying assumption that model training is secure, leading to\nevasion disruptions being the primary type of studied disruption. Finally, we\nwill examine the impacts of data poisoning interventions and showcase how they\ncan endanger power grid resilience.\n", "date": "2024-07-19", "categories": []}
{"id": "2407.15385", "title": "Towards Robust Vision Transformer via Masked Adaptive Ensemble", "abstract": "  Adversarial training (AT) can help improve the robustness of Vision\nTransformers (ViT) against adversarial attacks by intentionally injecting\nadversarial examples into the training data. However, this way of adversarial\ninjection inevitably incurs standard accuracy degradation to some extent,\nthereby calling for a trade-off between standard accuracy and robustness.\nBesides, the prominent AT solutions are still vulnerable to adaptive attacks.\nTo tackle such shortcomings, this paper proposes a novel ViT architecture,\nincluding a detector and a classifier bridged by our newly developed adaptive\nensemble. Specifically, we empirically discover that detecting adversarial\nexamples can benefit from the Guided Backpropagation technique. Driven by this\ndiscovery, a novel Multi-head Self-Attention (MSA) mechanism is introduced to\nenhance our detector to sniff adversarial examples. Then, a classifier with two\nencoders is employed for extracting visual representations respectively from\nclean images and adversarial examples, with our adaptive ensemble to adaptively\nadjust the proportion of visual representations from the two encoders for\naccurate classification. This design enables our ViT architecture to achieve a\nbetter trade-off between standard accuracy and robustness. Besides, our\nadaptive ensemble technique allows us to mask off a random subset of image\npatches within input data, boosting our ViT's robustness against adaptive\nattacks, while maintaining high standard accuracy. Experimental results exhibit\nthat our ViT architecture, on CIFAR-10, achieves the best standard accuracy and\nadversarial robustness of 90.3% and 49.8%, respectively.\n", "date": "2024-07-22", "categories": []}
{"id": "2407.15683", "title": "Enhancing Transferability of Targeted Adversarial Examples: A\n  Self-Universal Perspective", "abstract": "  Transfer-based targeted adversarial attacks against black-box deep neural\nnetworks (DNNs) have been proven to be significantly more challenging than\nuntargeted ones. The impressive transferability of current SOTA, the generative\nmethods, comes at the cost of requiring massive amounts of additional data and\ntime-consuming training for each targeted label. This results in limited\nefficiency and flexibility, significantly hindering their deployment in\npractical applications. In this paper, we offer a self-universal perspective\nthat unveils the great yet underexplored potential of input transformations in\npursuing this goal. Specifically, transformations universalize gradient-based\nattacks with intrinsic but overlooked semantics inherent within individual\nimages, exhibiting similar scalability and comparable results to time-consuming\nlearning over massive additional data from diverse classes. We also contribute\na surprising empirical insight that one of the most fundamental\ntransformations, simple image scaling, is highly effective, scalable,\nsufficient, and necessary in enhancing targeted transferability. We further\naugment simple scaling with orthogonal transformations and block-wise\napplicability, resulting in the Simple, faSt, Self-universal yet Strong Scale\nTransformation (S$^4$ST) for self-universal TTA. On the ImageNet-Compatible\nbenchmark dataset, our method achieves a 19.8% improvement in the average\ntargeted transfer success rate against various challenging victim models over\nexisting SOTA transformation methods while only consuming 36% time for\nattacking. It also outperforms resource-intensive attacks by a large margin in\nvarious challenging settings.\n", "date": "2024-07-22", "categories": []}
{"id": "2407.13796", "title": "Continuous Embedding Attacks via Clipped Inputs in Jailbreaking Large\n  Language Models", "abstract": "  Security concerns for large language models (LLMs) have recently escalated,\nfocusing on thwarting jailbreaking attempts in discrete prompts. However, the\nexploration of jailbreak vulnerabilities arising from continuous embeddings has\nbeen limited, as prior approaches primarily involved appending discrete or\ncontinuous suffixes to inputs. Our study presents a novel channel for\nconducting direct attacks on LLM inputs, eliminating the need for suffix\naddition or specific questions provided that the desired output is predefined.\nWe additionally observe that extensive iterations often lead to overfitting,\ncharacterized by repetition in the output. To counteract this, we propose a\nsimple yet effective strategy named CLIP. Our experiments show that for an\ninput length of 40 at iteration 1000, applying CLIP improves the ASR from 62%\nto 83%\n", "date": "2024-07-16", "categories": []}
{"id": "2312.00987", "title": "Deep Generative Attacks and Countermeasures for Data-Driven Offline\n  Signature Verification", "abstract": "  This study investigates the vulnerabilities of data-driven offline signature\nverification (DASV) systems to generative attacks and proposes robust\ncountermeasures. Specifically, we explore the efficacy of Variational\nAutoencoders (VAEs) and Conditional Generative Adversarial Networks (CGANs) in\ncreating deceptive signatures that challenge DASV systems. Using the Structural\nSimilarity Index (SSIM) to evaluate the quality of forged signatures, we assess\ntheir impact on DASV systems built with Xception, ResNet152V2, and DenseNet201\narchitectures. Initial results showed False Accept Rates (FARs) ranging from 0%\nto 5.47% across all models and datasets. However, exposure to synthetic\nsignatures significantly increased FARs, with rates ranging from 19.12% to\n61.64%. The proposed countermeasure, i.e., retraining the models with real +\nsynthetic datasets, was very effective, reducing FARs between 0% and 0.99%.\nThese findings emphasize the necessity of investigating vulnerabilities in\nsecurity systems like DASV and reinforce the role of generative methods in\nenhancing the security of data-driven systems.\n", "date": "2023-12-02", "categories": []}
{"id": "2407.11599", "title": "Enhancing TinyML Security: Study of Adversarial Attack Transferability", "abstract": "  The recent strides in artificial intelligence (AI) and machine learning (ML)\nhave propelled the rise of TinyML, a paradigm enabling AI computations at the\nedge without dependence on cloud connections. While TinyML offers real-time\ndata analysis and swift responses critical for diverse applications, its\ndevices' intrinsic resource limitations expose them to security risks. This\nresearch delves into the adversarial vulnerabilities of AI models on\nresource-constrained embedded hardware, with a focus on Model Extraction and\nEvasion Attacks. Our findings reveal that adversarial attacks from powerful\nhost machines could be transferred to smaller, less secure devices like ESP32\nand Raspberry Pi. This illustrates that adversarial attacks could be extended\nto tiny devices, underscoring vulnerabilities, and emphasizing the necessity\nfor reinforced security measures in TinyML deployments. This exploration\nenhances the comprehension of security challenges in TinyML and offers insights\nfor safeguarding sensitive data and ensuring device dependability in AI-powered\nedge computing settings.\n", "date": "2024-07-16", "categories": []}
{"id": "2407.11921", "title": "IPA-NeRF: Illusory Poisoning Attack Against Neural Radiance Fields", "abstract": "  Neural Radiance Field (NeRF) represents a significant advancement in computer\nvision, offering implicit neural network-based scene representation and novel\nview synthesis capabilities. Its applications span diverse fields including\nrobotics, urban mapping, autonomous navigation, virtual reality/augmented\nreality, etc., some of which are considered high-risk AI applications. However,\ndespite its widespread adoption, the robustness and security of NeRF remain\nlargely unexplored. In this study, we contribute to this area by introducing\nthe Illusory Poisoning Attack against Neural Radiance Fields (IPA-NeRF). This\nattack involves embedding a hidden backdoor view into NeRF, allowing it to\nproduce predetermined outputs, i.e. illusory, when presented with the specified\nbackdoor view while maintaining normal performance with standard inputs. Our\nattack is specifically designed to deceive users or downstream models at a\nparticular position while ensuring that any abnormalities in NeRF remain\nundetectable from other viewpoints. Experimental results demonstrate the\neffectiveness of our Illusory Poisoning Attack, successfully presenting the\ndesired illusory on the specified viewpoint without impacting other views.\nNotably, we achieve this attack by introducing small perturbations solely to\nthe training set. The code can be found at\nhttps://github.com/jiang-wenxiang/IPA-NeRF.\n", "date": "2024-07-16", "categories": []}
{"id": "2407.12281", "title": "Turning Generative Models Degenerate: The Power of Data Poisoning\n  Attacks", "abstract": "  The increasing use of large language models (LLMs) trained by third parties\nraises significant security concerns. In particular, malicious actors can\nintroduce backdoors through poisoning attacks to generate undesirable outputs.\nWhile such attacks have been extensively studied in image domains and\nclassification tasks, they remain underexplored for natural language generation\n(NLG) tasks. To address this gap, we conduct an investigation of various\npoisoning techniques targeting the LLM's fine-tuning phase via prefix-tuning, a\nParameter Efficient Fine-Tuning (PEFT) method. We assess their effectiveness\nacross two generative tasks: text summarization and text completion; and we\nalso introduce new metrics to quantify the success and stealthiness of such NLG\npoisoning attacks. Through our experiments, we find that the prefix-tuning\nhyperparameters and trigger designs are the most crucial factors to influence\nattack success and stealthiness. Moreover, we demonstrate that existing popular\ndefenses are ineffective against our poisoning attacks. Our study presents the\nfirst systematic approach to understanding poisoning attacks targeting NLG\ntasks during fine-tuning via PEFT across a wide range of triggers and attack\nsettings. We hope our findings will aid the AI security community in developing\neffective defenses against such threats.\n", "date": "2024-07-17", "categories": []}
{"id": "2407.12588", "title": "Benchmarking Robust Self-Supervised Learning Across Diverse Downstream\n  Tasks", "abstract": "  Large-scale vision models have become integral in many applications due to\ntheir unprecedented performance and versatility across downstream tasks.\nHowever, the robustness of these foundation models has primarily been explored\nfor a single task, namely image classification. The vulnerability of other\ncommon vision tasks, such as semantic segmentation and depth estimation,\nremains largely unknown. We present a comprehensive empirical evaluation of the\nadversarial robustness of self-supervised vision encoders across multiple\ndownstream tasks. Our attacks operate in the encoder embedding space and at the\ndownstream task output level. In both cases, current state-of-the-art\nadversarial fine-tuning techniques tested only for classification significantly\ndegrade clean and robust performance on other tasks. Since the purpose of a\nfoundation model is to cater to multiple applications at once, our findings\nreveal the need to enhance encoder robustness more broadly. Our code is\navailable at ${github.com/layer6ai-labs/ssl-robustness}$.\n", "date": "2024-07-17", "categories": []}
{"id": "2407.13111", "title": "PG-Attack: A Precision-Guided Adversarial Attack Framework Against\n  Vision Foundation Models for Autonomous Driving", "abstract": "  Vision foundation models are increasingly employed in autonomous driving\nsystems due to their advanced capabilities. However, these models are\nsusceptible to adversarial attacks, posing significant risks to the reliability\nand safety of autonomous vehicles. Adversaries can exploit these\nvulnerabilities to manipulate the vehicle's perception of its surroundings,\nleading to erroneous decisions and potentially catastrophic consequences. To\naddress this challenge, we propose a novel Precision-Guided Adversarial Attack\n(PG-Attack) framework that combines two techniques: Precision Mask Perturbation\nAttack (PMP-Attack) and Deceptive Text Patch Attack (DTP-Attack). PMP-Attack\nprecisely targets the attack region to minimize the overall perturbation while\nmaximizing its impact on the target object's representation in the model's\nfeature space. DTP-Attack introduces deceptive text patches that disrupt the\nmodel's understanding of the scene, further enhancing the attack's\neffectiveness. Our experiments demonstrate that PG-Attack successfully deceives\na variety of advanced multi-modal large models, including GPT-4V, Qwen-VL, and\nimp-V1. Additionally, we won First-Place in the CVPR 2024 Workshop Challenge:\nBlack-box Adversarial Attacks on Vision Foundation Models and codes are\navailable at https://github.com/fuhaha824/PG-Attack.\n", "date": "2024-07-18", "categories": []}
{"id": "2407.13646", "title": "Beyond Dropout: Robust Convolutional Neural Networks Based on Local\n  Feature Masking", "abstract": "  In the contemporary of deep learning, where models often grapple with the\nchallenge of simultaneously achieving robustness against adversarial attacks\nand strong generalization capabilities, this study introduces an innovative\nLocal Feature Masking (LFM) strategy aimed at fortifying the performance of\nConvolutional Neural Networks (CNNs) on both fronts. During the training phase,\nwe strategically incorporate random feature masking in the shallow layers of\nCNNs, effectively alleviating overfitting issues, thereby enhancing the model's\ngeneralization ability and bolstering its resilience to adversarial attacks.\nLFM compels the network to adapt by leveraging remaining features to compensate\nfor the absence of certain semantic features, nurturing a more elastic feature\nlearning mechanism. The efficacy of LFM is substantiated through a series of\nquantitative and qualitative assessments, collectively showcasing a consistent\nand significant improvement in CNN's generalization ability and resistance\nagainst adversarial attacks--a phenomenon not observed in current and prior\nmethodologies. The seamless integration of LFM into established CNN frameworks\nunderscores its potential to advance both generalization and adversarial\nrobustness within the deep learning paradigm. Through comprehensive\nexperiments, including robust person re-identification baseline generalization\nexperiments and adversarial attack experiments, we demonstrate the substantial\nenhancements offered by LFM in addressing the aforementioned challenges. This\ncontribution represents a noteworthy stride in advancing robust neural network\narchitectures.\n", "date": "2024-07-18", "categories": []}
{"id": "2407.13700", "title": "Cross-Task Attack: A Self-Supervision Generative Framework Based on\n  Attention Shift", "abstract": "  Studying adversarial attacks on artificial intelligence (AI) systems helps\ndiscover model shortcomings, enabling the construction of a more robust system.\nMost existing adversarial attack methods only concentrate on single-task\nsingle-model or single-task cross-model scenarios, overlooking the multi-task\ncharacteristic of artificial intelligence systems. As a result, most of the\nexisting attacks do not pose a practical threat to a comprehensive and\ncollaborative AI system. However, implementing cross-task attacks is highly\ndemanding and challenging due to the difficulty in obtaining the real labels of\ndifferent tasks for the same picture and harmonizing the loss functions across\ndifferent tasks. To address this issue, we propose a self-supervised Cross-Task\nAttack framework (CTA), which utilizes co-attention and anti-attention maps to\ngenerate cross-task adversarial perturbation. Specifically, the co-attention\nmap reflects the area to which different visual task models pay attention,\nwhile the anti-attention map reflects the area that different visual task\nmodels neglect. CTA generates cross-task perturbations by shifting the\nattention area of samples away from the co-attention map and closer to the\nanti-attention map. We conduct extensive experiments on multiple vision tasks\nand the experimental results confirm the effectiveness of the proposed design\nfor adversarial attacks.\n", "date": "2024-07-18", "categories": []}
{"id": "2407.13757", "title": "Black-Box Opinion Manipulation Attacks to Retrieval-Augmented Generation\n  of Large Language Models", "abstract": "  Retrieval-Augmented Generation (RAG) is applied to solve hallucination\nproblems and real-time constraints of large language models, but it also\ninduces vulnerabilities against retrieval corruption attacks. Existing research\nmainly explores the unreliability of RAG in white-box and closed-domain QA\ntasks. In this paper, we aim to reveal the vulnerabilities of\nRetrieval-Enhanced Generative (RAG) models when faced with black-box attacks\nfor opinion manipulation. We explore the impact of such attacks on user\ncognition and decision-making, providing new insight to enhance the reliability\nand security of RAG models. We manipulate the ranking results of the retrieval\nmodel in RAG with instruction and use these results as data to train a\nsurrogate model. By employing adversarial retrieval attack methods to the\nsurrogate model, black-box transfer attacks on RAG are further realized.\nExperiments conducted on opinion datasets across multiple topics show that the\nproposed attack strategy can significantly alter the opinion polarity of the\ncontent generated by RAG. This demonstrates the model's vulnerability and, more\nimportantly, reveals the potential negative impact on user cognition and\ndecision-making, making it easier to mislead users into accepting incorrect or\nbiased information.\n", "date": "2024-07-18", "categories": []}
{"id": "2303.09079", "title": "SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning", "abstract": "  Self-supervised learning (SSL) is a prevalent approach for encoding data\nrepresentations. Using a pre-trained SSL image encoder and subsequently\ntraining a downstream classifier, impressive performance can be achieved on\nvarious tasks with very little labeled data. The growing adoption of SSL has\nled to an increase in security research on SSL encoders and associated Trojan\nattacks. Trojan attacks embedded in SSL encoders can operate covertly,\nspreading across multiple users and devices. The presence of backdoor behavior\nin Trojaned encoders can inadvertently be inherited by downstream classifiers,\nmaking it even more difficult to detect and mitigate the threat. Although\ncurrent Trojan detection methods in supervised learning can potentially\nsafeguard SSL downstream classifiers, identifying and addressing triggers in\nthe SSL encoder before its widespread dissemination is a challenging task. This\nchallenge arises because downstream tasks might be unknown, dataset labels may\nbe unavailable, and the original unlabeled training dataset might be\ninaccessible during Trojan detection in SSL encoders. We introduce SSL-Cleanse\nas a solution to identify and mitigate backdoor threats in SSL encoders. We\nevaluated SSL-Cleanse on various datasets using 1200 encoders, achieving an\naverage detection success rate of 82.2% on ImageNet-100. After mitigating\nbackdoors, on average, backdoored encoders achieve 0.3% attack success rate\nwithout great accuracy loss, proving the effectiveness of SSL-Cleanse. The\nsource code of SSL-Cleanse is available at\nhttps://github.com/UCF-ML-Research/SSL-Cleanse.\n", "date": "2023-03-16", "categories": []}
{"id": "2306.09124", "title": "DIFFender: Diffusion-Based Adversarial Defense against Patch Attacks", "abstract": "  Adversarial attacks, particularly patch attacks, pose significant threats to\nthe robustness and reliability of deep learning models. Developing reliable\ndefenses against patch attacks is crucial for real-world applications. This\npaper introduces DIFFender, a novel defense framework that harnesses the\ncapabilities of a text-guided diffusion model to combat patch attacks. Central\nto our approach is the discovery of the Adversarial Anomaly Perception (AAP)\nphenomenon, which empowers the diffusion model to detect and localize\nadversarial patches through the analysis of distributional discrepancies.\nDIFFender integrates dual tasks of patch localization and restoration within a\nsingle diffusion model framework, utilizing their close interaction to enhance\ndefense efficacy. Moreover, DIFFender utilizes vision-language pre-training\ncoupled with an efficient few-shot prompt-tuning algorithm, which streamlines\nthe adaptation of the pre-trained diffusion model to defense tasks, thus\neliminating the need for extensive retraining. Our comprehensive evaluation\nspans image classification and face recognition tasks, extending to real-world\nscenarios, where DIFFender shows good robustness against adversarial attacks.\nThe versatility and generalizability of DIFFender are evident across a variety\nof settings, classifiers, and attack methodologies, marking an advancement in\nadversarial patch defense strategies.\n", "date": "2023-06-15", "categories": []}
{"id": "2311.18495", "title": "Improving Adversarial Transferability via Model Alignment", "abstract": "  Neural networks are susceptible to adversarial perturbations that are\ntransferable across different models. In this paper, we introduce a novel model\nalignment technique aimed at improving a given source model's ability in\ngenerating transferable adversarial perturbations. During the alignment\nprocess, the parameters of the source model are fine-tuned to minimize an\nalignment loss. This loss measures the divergence in the predictions between\nthe source model and another, independently trained model, referred to as the\nwitness model. To understand the effect of model alignment, we conduct a\ngeometric analysis of the resulting changes in the loss landscape. Extensive\nexperiments on the ImageNet dataset, using a variety of model architectures,\ndemonstrate that perturbations generated from aligned source models exhibit\nsignificantly higher transferability than those from the original source model.\n", "date": "2023-11-30", "categories": []}
{"id": "2312.14440", "title": "Asymmetric Bias in Text-to-Image Generation with Adversarial Attacks", "abstract": "  The widespread use of Text-to-Image (T2I) models in content generation\nrequires careful examination of their safety, including their robustness to\nadversarial attacks. Despite extensive research on adversarial attacks, the\nreasons for their effectiveness remain underexplored. This paper presents an\nempirical study on adversarial attacks against T2I models, focusing on\nanalyzing factors associated with attack success rates (ASR). We introduce a\nnew attack objective - entity swapping using adversarial suffixes and two\ngradient-based attack algorithms. Human and automatic evaluations reveal the\nasymmetric nature of ASRs on entity swap: for example, it is easier to replace\n\"human\" with \"robot\" in the prompt \"a human dancing in the rain.\" with an\nadversarial suffix, but the reverse replacement is significantly harder. We\nfurther propose probing metrics to establish indicative signals from the\nmodel's beliefs to the adversarial ASR. We identify conditions that result in a\nsuccess probability of 60% for adversarial attacks and others where this\nlikelihood drops below 5%.\n", "date": "2023-12-22", "categories": []}
{"id": "2407.04215", "title": "T2IShield: Defending Against Backdoors on Text-to-Image Diffusion Models", "abstract": "  While text-to-image diffusion models demonstrate impressive generation\ncapabilities, they also exhibit vulnerability to backdoor attacks, which\ninvolve the manipulation of model outputs through malicious triggers. In this\npaper, for the first time, we propose a comprehensive defense method named\nT2IShield to detect, localize, and mitigate such attacks. Specifically, we find\nthe \"Assimilation Phenomenon\" on the cross-attention maps caused by the\nbackdoor trigger. Based on this key insight, we propose two effective backdoor\ndetection methods: Frobenius Norm Threshold Truncation and Covariance\nDiscriminant Analysis. Besides, we introduce a binary-search approach to\nlocalize the trigger within a backdoor sample and assess the efficacy of\nexisting concept editing methods in mitigating backdoor attacks. Empirical\nevaluations on two advanced backdoor attack scenarios show the effectiveness of\nour proposed defense method. For backdoor sample detection, T2IShield achieves\na detection F1 score of 88.9$\\%$ with low computational cost. Furthermore,\nT2IShield achieves a localization F1 score of 86.4$\\%$ and invalidates 99$\\%$\npoisoned samples. Codes are released at https://github.com/Robin-WZQ/T2IShield.\n", "date": "2024-07-05", "categories": []}
{"id": "2407.12292", "title": "Any Target Can be Offense: Adversarial Example Generation via\n  Generalized Latent Infection", "abstract": "  Targeted adversarial attack, which aims to mislead a model to recognize any\nimage as a target object by imperceptible perturbations, has become a\nmainstream tool for vulnerability assessment of deep neural networks (DNNs).\nSince existing targeted attackers only learn to attack known target classes,\nthey cannot generalize well to unknown classes. To tackle this issue, we\npropose $\\bf{G}$eneralized $\\bf{A}$dversarial attac$\\bf{KER}$ ($\\bf{GAKer}$),\nwhich is able to construct adversarial examples to any target class. The core\nidea behind GAKer is to craft a latently infected representation during\nadversarial example generation. To this end, the extracted latent\nrepresentations of the target object are first injected into intermediate\nfeatures of an input image in an adversarial generator. Then, the generator is\noptimized to ensure visual consistency with the input image while being close\nto the target object in the feature space. Since the GAKer is class-agnostic\nyet model-agnostic, it can be regarded as a general tool that not only reveals\nthe vulnerability of more DNNs but also identifies deficiencies of DNNs in a\nwider range of classes. Extensive experiments have demonstrated the\neffectiveness of our proposed method in generating adversarial examples for\nboth known and unknown classes. Notably, compared with other generative\nmethods, our method achieves an approximately $14.13\\%$ higher attack success\nrate for unknown classes and an approximately $4.23\\%$ higher success rate for\nknown classes. Our code is available in https://github.com/VL-Group/GAKer.\n", "date": "2024-07-17", "categories": []}
{"id": "2407.12443", "title": "Preventing Catastrophic Overfitting in Fast Adversarial Training: A\n  Bi-level Optimization Perspective", "abstract": "  Adversarial training (AT) has become an effective defense method against\nadversarial examples (AEs) and it is typically framed as a bi-level\noptimization problem. Among various AT methods, fast AT (FAT), which employs a\nsingle-step attack strategy to guide the training process, can achieve good\nrobustness against adversarial attacks at a low cost. However, FAT methods\nsuffer from the catastrophic overfitting problem, especially on complex tasks\nor with large-parameter models. In this work, we propose a FAT method termed\nFGSM-PCO, which mitigates catastrophic overfitting by averting the collapse of\nthe inner optimization problem in the bi-level optimization process. FGSM-PCO\ngenerates current-stage AEs from the historical AEs and incorporates them into\nthe training process using an adaptive mechanism. This mechanism determines an\nappropriate fusion ratio according to the performance of the AEs on the\ntraining model. Coupled with a loss function tailored to the training\nframework, FGSM-PCO can alleviate catastrophic overfitting and help the\nrecovery of an overfitted model to effective training. We evaluate our\nalgorithm across three models and three datasets to validate its effectiveness.\nComparative empirical studies against other FAT algorithms demonstrate that our\nproposed method effectively addresses unresolved overfitting issues in existing\nalgorithms.\n", "date": "2024-07-17", "categories": []}
{"id": "2211.11236", "title": "Boosting the Transferability of Adversarial Attacks with Global Momentum\n  Initialization", "abstract": "  Deep Neural Networks (DNNs) are vulnerable to adversarial examples, which are\ncrafted by adding human-imperceptible perturbations to the benign inputs.\nSimultaneously, adversarial examples exhibit transferability across models,\nenabling practical black-box attacks. However, existing methods are still\nincapable of achieving the desired transfer attack performance. In this work,\nfocusing on gradient optimization and consistency, we analyse the gradient\nelimination phenomenon as well as the local momentum optimum dilemma. To tackle\nthese challenges, we introduce Global Momentum Initialization (GI), providing\nglobal momentum knowledge to mitigate gradient elimination. Specifically, we\nperform gradient pre-convergence before the attack and a global search during\nthis stage. GI seamlessly integrates with existing transfer methods,\nsignificantly improving the success rate of transfer attacks by an average of\n6.4% under various advanced defense mechanisms compared to the state-of-the-art\nmethod. Ultimately, GI demonstrates strong transferability in both image and\nvideo attack domains. Particularly, when attacking advanced defense methods in\nthe image domain, it achieves an average attack success rate of 95.4%. The code\nis available at\n$\\href{https://github.com/Omenzychen/Global-Momentum-Initialization}{https://github.com/Omenzychen/Global-Momentum-Initialization}$.\n", "date": "2022-11-21", "categories": []}
{"id": "2306.12941", "title": "Towards Reliable Evaluation and Fast Training of Robust Semantic\n  Segmentation Models", "abstract": "  Adversarial robustness has been studied extensively in image classification,\nespecially for the $\\ell_\\infty$-threat model, but significantly less so for\nrelated tasks such as object detection and semantic segmentation, where attacks\nturn out to be a much harder optimization problem than for image\nclassification. We propose several problem-specific novel attacks minimizing\ndifferent metrics in accuracy and mIoU. The ensemble of our attacks, SEA, shows\nthat existing attacks severely overestimate the robustness of semantic\nsegmentation models. Surprisingly, existing attempts of adversarial training\nfor semantic segmentation models turn out to be weak or even completely\nnon-robust. We investigate why previous adaptations of adversarial training to\nsemantic segmentation failed and show how recently proposed robust ImageNet\nbackbones can be used to obtain adversarially robust semantic segmentation\nmodels with up to six times less training time for PASCAL-VOC and the more\nchallenging ADE20k. The associated code and robust models are available at\nhttps://github.com/nmndeep/robust-segmentation\n", "date": "2023-06-22", "categories": []}
{"id": "2407.10825", "title": "Wicked Oddities: Selectively Poisoning for Effective Clean-Label\n  Backdoor Attacks", "abstract": "  Deep neural networks are vulnerable to backdoor attacks, a type of\nadversarial attack that poisons the training data to manipulate the behavior of\nmodels trained on such data. Clean-label attacks are a more stealthy form of\nbackdoor attacks that can perform the attack without changing the labels of\npoisoned data. Early works on clean-label attacks added triggers to a random\nsubset of the training set, ignoring the fact that samples contribute unequally\nto the attack's success. This results in high poisoning rates and low attack\nsuccess rates. To alleviate the problem, several supervised learning-based\nsample selection strategies have been proposed. However, these methods assume\naccess to the entire labeled training set and require training, which is\nexpensive and may not always be practical. This work studies a new and more\npractical (but also more challenging) threat model where the attacker only\nprovides data for the target class (e.g., in face recognition systems) and has\nno knowledge of the victim model or any other classes in the training set. We\nstudy different strategies for selectively poisoning a small set of training\nsamples in the target class to boost the attack success rate in this setting.\nOur threat model poses a serious threat in training machine learning models\nwith third-party datasets, since the attack can be performed effectively with\nlimited information. Experiments on benchmark datasets illustrate the\neffectiveness of our strategies in improving clean-label backdoor attacks.\n", "date": "2024-07-15", "categories": []}
{"id": "2407.11011", "title": "Toward Availability Attacks in 3D Point Clouds", "abstract": "  Despite the great progress of 3D vision, data privacy and security issues in\n3D deep learning are not explored systematically. In the domain of 2D images,\nmany availability attacks have been proposed to prevent data from being\nillicitly learned by unauthorized deep models. However, unlike images\nrepresented on a fixed dimensional grid, point clouds are characterized as\nunordered and unstructured sets, posing a significant challenge in designing an\neffective availability attack for 3D deep learning. In this paper, we\ntheoretically show that extending 2D availability attacks directly to 3D point\nclouds under distance regularization is susceptible to the degeneracy,\nrendering the generated poisons weaker or even ineffective. This is because in\nbi-level optimization, introducing regularization term can result in update\ndirections out of control. To address this issue, we propose a novel Feature\nCollision Error-Minimization (FC-EM) method, which creates additional shortcuts\nin the feature space, inducing different update directions to prevent the\ndegeneracy of bi-level optimization. Moreover, we provide a theoretical\nanalysis that demonstrates the effectiveness of the FC-EM attack. Extensive\nexperiments on typical point cloud datasets, 3D intracranial aneurysm medical\ndataset, and 3D face dataset verify the superiority and practicality of our\napproach. Code is available at https://github.com/hala64/fc-em.\n", "date": "2024-06-26", "categories": []}
{"id": "2407.11029", "title": "A Geometric Framework for Adversarial Vulnerability in Machine Learning", "abstract": "  This work starts with the intention of using mathematics to understand the\nintriguing vulnerability observed by ~\\citet{szegedy2013} within artificial\nneural networks. Along the way, we will develop some novel tools with\napplications far outside of just the adversarial domain. We will do this while\ndeveloping a rigorous mathematical framework to examine this problem. Our goal\nis to build out theory which can support increasingly sophisticated conjecture\nabout adversarial attacks with a particular focus on the so called ``Dimpled\nManifold Hypothesis'' by ~\\citet{shamir2021dimpled}. Chapter one will cover the\nhistory and architecture of neural network architectures. Chapter two is\nfocused on the background of adversarial vulnerability. Starting from the\nseminal paper by ~\\citet{szegedy2013} we will develop the theory of adversarial\nperturbation and attack.\n  Chapter three will build a theory of persistence that is related to Ricci\nCurvature, which can be used to measure properties of decision boundaries. We\nwill use this foundation to make a conjecture relating adversarial attacks.\nChapters four and five represent a sudden and wonderful digression that\nexamines an intriguing related body of theory for spatial analysis of neural\nnetworks as approximations of kernel machines and becomes a novel theory for\nrepresenting neural networks with bilinear maps. These heavily mathematical\nchapters will set up a framework and begin exploring applications of what may\nbecome a very important theoretical foundation for analyzing neural network\nlearning with spatial and geometric information. We will conclude by setting up\nour new methods to address the conjecture from chapter 3 in continuing\nresearch.\n", "date": "2024-07-03", "categories": []}
{"id": "2407.11072", "title": "MaPPing Your Model: Assessing the Impact of Adversarial Attacks on\n  LLM-based Programming Assistants", "abstract": "  LLM-based programming assistants offer the promise of programming faster but\nwith the risk of introducing more security vulnerabilities. Prior work has\nstudied how LLMs could be maliciously fine-tuned to suggest vulnerabilities\nmore often. With the rise of agentic LLMs, which may use results from an\nuntrusted third party, there is a growing risk of attacks on the model's\nprompt. We introduce the Malicious Programming Prompt (MaPP) attack, in which\nan attacker adds a small amount of text to a prompt for a programming task\n(under 500 bytes). We show that our prompt strategy can cause an LLM to add\nvulnerabilities while continuing to write otherwise correct code. We evaluate\nthree prompts on seven common LLMs, from basic to state-of-the-art commercial\nmodels. Using the HumanEval benchmark, we find that our prompts are broadly\neffective, with no customization required for different LLMs. Furthermore, the\nLLMs that are best at HumanEval are also best at following our malicious\ninstructions, suggesting that simply scaling language models will not prevent\nMaPP attacks. Using a dataset of eight CWEs in 16 scenarios, we find that MaPP\nattacks are also effective at implementing specific and targeted\nvulnerabilities across a range of models. Our work highlights the need to\nsecure LLM prompts against manipulation as well as rigorously auditing code\ngenerated with the help of LLMs.\n", "date": "2024-07-12", "categories": []}
{"id": "2407.11073", "title": "SemiAdv: Query-Efficient Black-Box Adversarial Attack with Unlabeled\n  Images", "abstract": "  Adversarial attack has garnered considerable attention due to its profound\nimplications for the secure deployment of robots in sensitive security\nscenarios. To potentially push for advances in the field, this paper studies\nthe adversarial attack in the black-box setting and proposes an unlabeled\ndata-driven adversarial attack method, called SemiAdv. Specifically, SemiAdv\nachieves the following breakthroughs compared with previous works. First, by\nintroducing the semi-supervised learning technique into the adversarial attack,\nSemiAdv substantially decreases the number of queries required for generating\nadversarial samples. On average, SemiAdv only needs to query a few hundred\ntimes to launch an effective attack with more than 90% success rate. Second,\nmany existing black-box adversarial attacks require massive labeled data to\nmitigate the difference between the local substitute model and the remote\ntarget model for a good attack performance. While SemiAdv relaxes this\nlimitation and is capable of utilizing unlabeled raw data to launch an\neffective attack. Finally, our experiments show that SemiAdv saves up to 12x\nquery accesses for generating adversarial samples while maintaining a\ncompetitive attack success rate compared with state-of-the-art attacks.\n", "date": "2024-07-13", "categories": []}
{"id": "2407.11091", "title": "SENTINEL: Securing Indoor Localization against Adversarial Attacks with\n  Capsule Neural Networks", "abstract": "  With the increasing demand for edge device powered location-based services in\nindoor environments, Wi-Fi received signal strength (RSS) fingerprinting has\nbecome popular, given the unavailability of GPS indoors. However, achieving\nrobust and efficient indoor localization faces several challenges, due to RSS\nfluctuations from dynamic changes in indoor environments and heterogeneity of\nedge devices, leading to diminished localization accuracy. While advances in\nmachine learning (ML) have shown promise in mitigating these phenomena, it\nremains an open problem. Additionally, emerging threats from adversarial\nattacks on ML-enhanced indoor localization systems, especially those introduced\nby malicious or rogue access points (APs), can deceive ML models to further\nincrease localization errors. To address these challenges, we present SENTINEL,\na novel embedded ML framework utilizing modified capsule neural networks to\nbolster the resilience of indoor localization solutions against adversarial\nattacks, device heterogeneity, and dynamic RSS fluctuations. We also introduce\nRSSRogueLoc, a novel dataset capturing the effects of rogue APs from several\nreal-world indoor environments. Experimental evaluations demonstrate that\nSENTINEL achieves significant improvements, with up to 3.5x reduction in mean\nerror and 3.4x reduction in worst-case error compared to state-of-the-art\nframeworks using simulated adversarial attacks. SENTINEL also achieves\nimprovements of up to 2.8x in mean error and 2.7x in worst-case error compared\nto state-of-the-art frameworks when evaluated with the real-world RSSRogueLoc\ndataset.\n", "date": "2024-07-14", "categories": []}
{"id": "2407.11844", "title": "Variational Randomized Smoothing for Sample-Wise Adversarial Robustness", "abstract": "  Randomized smoothing is a defensive technique to achieve enhanced robustness\nagainst adversarial examples which are small input perturbations that degrade\nthe performance of neural network models. Conventional randomized smoothing\nadds random noise with a fixed noise level for every input sample to smooth out\nadversarial perturbations. This paper proposes a new variational framework that\nuses a per-sample noise level suitable for each input by introducing a noise\nlevel selector. Our experimental results demonstrate enhancement of empirical\nrobustness against adversarial attacks. We also provide and analyze the\ncertified robustness for our sample-wise smoothing method.\n", "date": "2024-07-16", "categories": []}
{"id": "2303.07474", "title": "Can Adversarial Examples Be Parsed to Reveal Victim Model Information?", "abstract": "  Numerous adversarial attack methods have been developed to generate\nimperceptible image perturbations that can cause erroneous predictions of\nstate-of-the-art machine learning (ML) models, in particular, deep neural\nnetworks (DNNs). Despite intense research on adversarial attacks, little effort\nwas made to uncover 'arcana' carried in adversarial attacks. In this work, we\nask whether it is possible to infer data-agnostic victim model (VM) information\n(i.e., characteristics of the ML model or DNN used to generate adversarial\nattacks) from data-specific adversarial instances. We call this 'model parsing\nof adversarial attacks' - a task to uncover 'arcana' in terms of the concealed\nVM information in attacks. We approach model parsing via supervised learning,\nwhich correctly assigns classes of VM's model attributes (in terms of\narchitecture type, kernel size, activation function, and weight sparsity) to an\nattack instance generated from this VM. We collect a dataset of adversarial\nattacks across 7 attack types generated from 135 victim models (configured by 5\narchitecture types, 3 kernel size setups, 3 activation function types, and 3\nweight sparsity ratios). We show that a simple, supervised model parsing\nnetwork (MPN) is able to infer VM attributes from unseen adversarial attacks if\ntheir attack settings are consistent with the training setting (i.e.,\nin-distribution generalization assessment). We also provide extensive\nexperiments to justify the feasibility of VM parsing from adversarial attacks,\nand the influence of training and evaluation factors in the parsing performance\n(e.g., generalization challenge raised in out-of-distribution evaluation). We\nfurther demonstrate how the proposed MPN can be used to uncover the source VM\nattributes from transfer attacks, and shed light on a potential connection\nbetween model parsing and attack transferability.\n", "date": "2023-03-13", "categories": []}
{"id": "2307.12499", "title": "AdvDiff: Generating Unrestricted Adversarial Examples using Diffusion\n  Models", "abstract": "  Unrestricted adversarial attacks present a serious threat to deep learning\nmodels and adversarial defense techniques. They pose severe security problems\nfor deep learning applications because they can effectively bypass defense\nmechanisms. However, previous attack methods often directly inject Projected\nGradient Descent (PGD) gradients into the sampling of generative models, which\nare not theoretically provable and thus generate unrealistic examples by\nincorporating adversarial objectives, especially for GAN-based methods on\nlarge-scale datasets like ImageNet. In this paper, we propose a new method,\ncalled AdvDiff, to generate unrestricted adversarial examples with diffusion\nmodels. We design two novel adversarial guidance techniques to conduct\nadversarial sampling in the reverse generation process of diffusion models.\nThese two techniques are effective and stable in generating high-quality,\nrealistic adversarial examples by integrating gradients of the target\nclassifier interpretably. Experimental results on MNIST and ImageNet datasets\ndemonstrate that AdvDiff is effective in generating unrestricted adversarial\nexamples, which outperforms state-of-the-art unrestricted adversarial attack\nmethods in terms of attack performance and generation quality.\n", "date": "2023-07-24", "categories": []}
{"id": "2402.03317", "title": "SpecFormer: Guarding Vision Transformer Robustness via Maximum Singular\n  Value Penalization", "abstract": "  Vision Transformers (ViTs) are increasingly used in computer vision due to\ntheir high performance, but their vulnerability to adversarial attacks is a\nconcern. Existing methods lack a solid theoretical basis, focusing mainly on\nempirical training adjustments. This study introduces SpecFormer, tailored to\nfortify ViTs against adversarial attacks, with theoretical underpinnings. We\nestablish local Lipschitz bounds for the self-attention layer and propose the\nMaximum Singular Value Penalization (MSVP) to precisely manage these bounds By\nincorporating MSVP into ViTs' attention layers, we enhance the model's\nrobustness without compromising training efficiency. SpecFormer, the resulting\nmodel, outperforms other state-of-the-art models in defending against\nadversarial attacks, as proven by experiments on CIFAR and ImageNet datasets.\nCode is released at https://github.com/microsoft/robustlearn.\n", "date": "2024-01-02", "categories": []}
{"id": "2402.15429", "title": "ProTIP: Probabilistic Robustness Verification on Text-to-Image Diffusion\n  Models against Stochastic Perturbation", "abstract": "  Text-to-Image (T2I) Diffusion Models (DMs) have shown impressive abilities in\ngenerating high-quality images based on simple text descriptions. However, as\nis common with many Deep Learning (DL) models, DMs are subject to a lack of\nrobustness. While there are attempts to evaluate the robustness of T2I DMs as a\nbinary or worst-case problem, they cannot answer how robust in general the\nmodel is whenever an adversarial example (AE) can be found. In this study, we\nfirst introduce a probabilistic notion of T2I DMs' robustness; and then\nestablish an efficient framework, ProTIP, to evaluate it with statistical\nguarantees. The main challenges stem from: i) the high computational cost of\nthe generation process; and ii) determining if a perturbed input is an AE\ninvolves comparing two output distributions, which is fundamentally harder\ncompared to other DL tasks like classification where an AE is identified upon\nmisprediction of labels. To tackle the challenges, we employ sequential\nanalysis with efficacy and futility early stopping rules in the statistical\ntesting for identifying AEs, and adaptive concentration inequalities to\ndynamically determine the \"just-right\" number of stochastic perturbations\nwhenever the verification target is met. Empirical experiments validate the\neffectiveness and efficiency of ProTIP over common T2I DMs. Finally, we\ndemonstrate an application of ProTIP to rank commonly used defence methods.\n", "date": "2024-02-23", "categories": []}
{"id": "2403.12445", "title": "Boosting Transferability in Vision-Language Attacks via Diversification\n  along the Intersection Region of Adversarial Trajectory", "abstract": "  Vision-language pre-training (VLP) models exhibit remarkable capabilities in\ncomprehending both images and text, yet they remain susceptible to multimodal\nadversarial examples (AEs). Strengthening attacks and uncovering\nvulnerabilities, especially common issues in VLP models (e.g., high\ntransferable AEs), can advance reliable and practical VLP models. A recent work\n(i.e., Set-level guidance attack) indicates that augmenting image-text pairs to\nincrease AE diversity along the optimization path enhances the transferability\nof adversarial examples significantly. However, this approach predominantly\nemphasizes diversity around the online adversarial examples (i.e., AEs in the\noptimization period), leading to the risk of overfitting the victim model and\naffecting the transferability. In this study, we posit that the diversity of\nadversarial examples towards the clean input and online AEs are both pivotal\nfor enhancing transferability across VLP models. Consequently, we propose using\ndiversification along the intersection region of adversarial trajectory to\nexpand the diversity of AEs. To fully leverage the interaction between\nmodalities, we introduce text-guided adversarial example selection during\noptimization. Furthermore, to further mitigate the potential overfitting, we\ndirect the adversarial text deviating from the last intersection region along\nthe optimization path, rather than adversarial images as in existing methods.\nExtensive experiments affirm the effectiveness of our method in improving\ntransferability across various VLP models and downstream vision-and-language\ntasks.\n", "date": "2024-03-19", "categories": []}
{"id": "2406.07188", "title": "Merging Improves Self-Critique Against Jailbreak Attacks", "abstract": "  The robustness of large language models (LLMs) against adversarial\nmanipulations, such as jailbreak attacks, remains a significant challenge. In\nthis work, we propose an approach that enhances the self-critique capability of\nthe LLM and further fine-tunes it over sanitized synthetic data. This is done\nwith the addition of an external critic model that can be merged with the\noriginal, thus bolstering self-critique capabilities and improving the\nrobustness of the LLMs response to adversarial prompts. Our results demonstrate\nthat the combination of merging and self-critique can reduce the attack success\nrate of adversaries significantly, thus offering a promising defense mechanism\nagainst jailbreak attacks. Code, data and models released at\nhttps://github.com/vicgalle/merging-self-critique-jailbreaks .\n", "date": "2024-06-11", "categories": []}
{"id": "2407.10077", "title": "Transferable 3D Adversarial Shape Completion using Diffusion Models", "abstract": "  Recent studies that incorporate geometric features and transformers into 3D\npoint cloud feature learning have significantly improved the performance of 3D\ndeep-learning models. However, their robustness against adversarial attacks has\nnot been thoroughly explored. Existing attack methods primarily focus on\nwhite-box scenarios and struggle to transfer to recently proposed 3D\ndeep-learning models. Even worse, these attacks introduce perturbations to 3D\ncoordinates, generating unrealistic adversarial examples and resulting in poor\nperformance against 3D adversarial defenses. In this paper, we generate\nhigh-quality adversarial point clouds using diffusion models. By using partial\npoints as prior knowledge, we generate realistic adversarial examples through\nshape completion with adversarial guidance. The proposed adversarial shape\ncompletion allows for a more reliable generation of adversarial point clouds.\nTo enhance attack transferability, we delve into the characteristics of 3D\npoint clouds and employ model uncertainty for better inference of model\nclassification through random down-sampling of point clouds. We adopt ensemble\nadversarial guidance for improved transferability across different network\narchitectures. To maintain the generation quality, we limit our adversarial\nguidance solely to the critical points of the point clouds by calculating\nsaliency scores. Extensive experiments demonstrate that our proposed attacks\noutperform state-of-the-art adversarial attack methods against both black-box\nmodels and defenses. Our black-box attack establishes a new baseline for\nevaluating the robustness of various 3D point cloud classification models.\n", "date": "2024-07-14", "categories": []}
{"id": "2407.10277", "title": "Disrupting Diffusion-based Inpainters with Semantic Digression", "abstract": "  The fabrication of visual misinformation on the web and social media has\nincreased exponentially with the advent of foundational text-to-image diffusion\nmodels. Namely, Stable Diffusion inpainters allow the synthesis of maliciously\ninpainted images of personal and private figures, and copyrighted contents,\nalso known as deepfakes. To combat such generations, a disruption framework,\nnamely Photoguard, has been proposed, where it adds adversarial noise to the\ncontext image to disrupt their inpainting synthesis. While their framework\nsuggested a diffusion-friendly approach, the disruption is not sufficiently\nstrong and it requires a significant amount of GPU and time to immunize the\ncontext image. In our work, we re-examine both the minimal and favorable\nconditions for a successful inpainting disruption, proposing DDD, a \"Digression\nguided Diffusion Disruption\" framework. First, we identify the most\nadversarially vulnerable diffusion timestep range with respect to the hidden\nspace. Within this scope of noised manifold, we pose the problem as a semantic\ndigression optimization. We maximize the distance between the inpainting\ninstance's hidden states and a semantic-aware hidden state centroid, calibrated\nboth by Monte Carlo sampling of hidden states and a discretely projected\noptimization in the token space. Effectively, our approach achieves stronger\ndisruption and a higher success rate than Photoguard while lowering the GPU\nmemory requirement, and speeding the optimization up to three times faster.\n", "date": "2024-07-14", "categories": []}
{"id": "2407.10918", "title": "PartImageNet++ Dataset: Scaling up Part-based Models for Robust\n  Recognition", "abstract": "  Deep learning-based object recognition systems can be easily fooled by\nvarious adversarial perturbations. One reason for the weak robustness may be\nthat they do not have part-based inductive bias like the human recognition\nprocess. Motivated by this, several part-based recognition models have been\nproposed to improve the adversarial robustness of recognition. However, due to\nthe lack of part annotations, the effectiveness of these methods is only\nvalidated on small-scale nonstandard datasets. In this work, we propose PIN++,\nshort for PartImageNet++, a dataset providing high-quality part segmentation\nannotations for all categories of ImageNet-1K (IN-1K). With these annotations,\nwe build part-based methods directly on the standard IN-1K dataset for robust\nrecognition. Different from previous two-stage part-based models, we propose a\nMulti-scale Part-supervised Model (MPM), to learn a robust representation with\npart annotations. Experiments show that MPM yielded better adversarial\nrobustness on the large-scale IN-1K over strong baselines across various attack\nsettings. Furthermore, MPM achieved improved robustness on common corruptions\nand several out-of-distribution datasets. The dataset, together with these\nresults, enables and encourages researchers to explore the potential of\npart-based models in more real applications.\n", "date": "2024-07-15", "categories": []}
{"id": "2405.10757", "title": "Rethinking Graph Backdoor Attacks: A Distribution-Preserving Perspective", "abstract": "  Graph Neural Networks (GNNs) have shown remarkable performance in various\ntasks. However, recent works reveal that GNNs are vulnerable to backdoor\nattacks. Generally, backdoor attack poisons the graph by attaching backdoor\ntriggers and the target class label to a set of nodes in the training graph. A\nGNN trained on the poisoned graph will then be misled to predict test nodes\nattached with trigger to the target class. Despite their effectiveness, our\nempirical analysis shows that triggers generated by existing methods tend to be\nout-of-distribution (OOD), which significantly differ from the clean data.\nHence, these injected triggers can be easily detected and pruned with widely\nused outlier detection methods in real-world applications. Therefore, in this\npaper, we study a novel problem of unnoticeable graph backdoor attacks with\nin-distribution (ID) triggers. To generate ID triggers, we introduce an OOD\ndetector in conjunction with an adversarial learning strategy to generate the\nattributes of the triggers within distribution. To ensure a high attack success\nrate with ID triggers, we introduce novel modules designed to enhance trigger\nmemorization by the victim model trained on poisoned graph. Extensive\nexperiments on real-world datasets demonstrate the effectiveness of the\nproposed method in generating in distribution triggers that can by-pass various\ndefense strategies while maintaining a high attack success rate.\n", "date": "2024-05-17", "categories": []}
{"id": "2407.09150", "title": "Evaluating the Adversarial Robustness of Semantic Segmentation: Trying\n  Harder Pays Off", "abstract": "  Machine learning models are vulnerable to tiny adversarial input\nperturbations optimized to cause a very large output error. To measure this\nvulnerability, we need reliable methods that can find such adversarial\nperturbations. For image classification models, evaluation methodologies have\nemerged that have stood the test of time. However, we argue that in the area of\nsemantic segmentation, a good approximation of the sensitivity to adversarial\nperturbations requires significantly more effort than what is currently\nconsidered satisfactory. To support this claim, we re-evaluate a number of\nwell-known robust segmentation models in an extensive empirical study. We\npropose new attacks and combine them with the strongest attacks available in\nthe literature. We also analyze the sensitivity of the models in fine detail.\nThe results indicate that most of the state-of-the-art models have a\ndramatically larger sensitivity to adversarial perturbations than previously\nreported. We also demonstrate a size-bias: small objects are often more easily\nattacked, even if the large objects are robust, a phenomenon not revealed by\ncurrent evaluation metrics. Our results also demonstrate that a diverse set of\nstrong attacks is necessary, because different models are often vulnerable to\ndifferent attacks.\n", "date": "2024-07-12", "categories": []}
{"id": "2407.09251", "title": "Deep Adversarial Defense Against Multilevel-Lp Attacks", "abstract": "  Deep learning models have shown considerable vulnerability to adversarial\nattacks, particularly as attacker strategies become more sophisticated. While\ntraditional adversarial training (AT) techniques offer some resilience, they\noften focus on defending against a single type of attack, e.g., the\n$\\ell_\\infty$-norm attack, which can fail for other types. This paper\nintroduces a computationally efficient multilevel $\\ell_p$ defense, called the\nEfficient Robust Mode Connectivity (EMRC) method, which aims to enhance a deep\nlearning model's resilience against multiple $\\ell_p$-norm attacks. Similar to\nanalytical continuation approaches used in continuous optimization, the method\nblends two $p$-specific adversarially optimal models, the $\\ell_1$- and\n$\\ell_\\infty$-norm AT solutions, to provide good adversarial robustness for a\nrange of $p$. We present experiments demonstrating that our approach performs\nbetter on various attacks as compared to AT-$\\ell_\\infty$, E-AT, and MSD, for\ndatasets/architectures including: CIFAR-10, CIFAR-100 / PreResNet110,\nWideResNet, ViT-Base.\n", "date": "2024-07-12", "categories": []}
{"id": "2208.07174", "title": "A Human-in-the-Middle Attack against Object Detection Systems", "abstract": "  Object detection systems using deep learning models have become increasingly\npopular in robotics thanks to the rising power of CPUs and GPUs in embedded\nsystems. However, these models are susceptible to adversarial attacks. While\nsome attacks are limited by strict assumptions on access to the detection\nsystem, we propose a novel hardware attack inspired by Man-in-the-Middle\nattacks in cryptography. This attack generates a Universal Adversarial\nPerturbations (UAP) and injects the perturbation between the USB camera and the\ndetection system via a hardware attack. Besides, prior research is misled by an\nevaluation metric that measures the model accuracy rather than the attack\nperformance. In combination with our proposed evaluation metrics, we\nsignificantly increased the strength of adversarial perturbations. These\nfindings raise serious concerns for applications of deep learning models in\nsafety-critical systems, such as autonomous driving.\n", "date": "2022-08-15", "categories": []}
{"id": "2406.13283", "title": "Large-Scale Dataset Pruning in Adversarial Training through Data\n  Importance Extrapolation", "abstract": "  Their vulnerability to small, imperceptible attacks limits the adoption of\ndeep learning models to real-world systems. Adversarial training has proven to\nbe one of the most promising strategies against these attacks, at the expense\nof a substantial increase in training time. With the ongoing trend of\nintegrating large-scale synthetic data this is only expected to increase even\nfurther. Thus, the need for data-centric approaches that reduce the number of\ntraining samples while maintaining accuracy and robustness arises. While data\npruning and active learning are prominent research topics in deep learning,\nthey are as of now largely unexplored in the adversarial training literature.\nWe address this gap and propose a new data pruning strategy based on\nextrapolating data importance scores from a small set of data to a larger set.\nIn an empirical evaluation, we demonstrate that extrapolation-based pruning can\nefficiently reduce dataset size while maintaining robustness.\n", "date": "2024-06-19", "categories": []}
{"id": "2404.09349", "title": "Adversarial Robustness Limits via Scaling-Law and Human-Alignment\n  Studies", "abstract": "  This paper revisits the simple, long-studied, yet still unsolved problem of\nmaking image classifiers robust to imperceptible perturbations. Taking CIFAR10\nas an example, SOTA clean accuracy is about $100$%, but SOTA robustness to\n$\\ell_{\\infty}$-norm bounded perturbations barely exceeds $70$%. To understand\nthis gap, we analyze how model size, dataset size, and synthetic data quality\naffect robustness by developing the first scaling laws for adversarial\ntraining. Our scaling laws reveal inefficiencies in prior art and provide\nactionable feedback to advance the field. For instance, we discovered that SOTA\nmethods diverge notably from compute-optimal setups, using excess compute for\ntheir level of robustness. Leveraging a compute-efficient setup, we surpass the\nprior SOTA with $20$% ($70$%) fewer training (inference) FLOPs. We trained\nvarious compute-efficient models, with our best achieving $74$% AutoAttack\naccuracy ($+3$% gain). However, our scaling laws also predict robustness slowly\ngrows then plateaus at $90$%: dwarfing our new SOTA by scaling is impractical,\nand perfect robustness is impossible. To better understand this predicted\nlimit, we carry out a small-scale human evaluation on the AutoAttack data that\nfools our top-performing model. Concerningly, we estimate that human\nperformance also plateaus near $90$%, which we show to be attributable to\n$\\ell_{\\infty}$-constrained attacks' generation of invalid images not\nconsistent with their original labels. Having characterized limiting\nroadblocks, we outline promising paths for future research.\n", "date": "2024-04-14", "categories": []}
{"id": "2407.07510", "title": "Invisible Optical Adversarial Stripes on Traffic Sign against Autonomous\n  Vehicles", "abstract": "  Camera-based computer vision is essential to autonomous vehicle's perception.\nThis paper presents an attack that uses light-emitting diodes and exploits the\ncamera's rolling shutter effect to create adversarial stripes in the captured\nimages to mislead traffic sign recognition. The attack is stealthy because the\nstripes on the traffic sign are invisible to human. For the attack to be\nthreatening, the recognition results need to be stable over consecutive image\nframes. To achieve this, we design and implement GhostStripe, an attack system\nthat controls the timing of the modulated light emission to adapt to camera\noperations and victim vehicle movements. Evaluated on real testbeds,\nGhostStripe can stably spoof the traffic sign recognition results for up to\n94\\% of frames to a wrong class when the victim vehicle passes the road\nsection. In reality, such attack effect may fool victim vehicles into\nlife-threatening incidents. We discuss the countermeasures at the levels of\ncamera sensor, perception model, and autonomous driving system.\n", "date": "2024-07-10", "categories": []}
{"id": "2407.02248", "title": "EvolBA: Evolutionary Boundary Attack under Hard-label Black Box\n  condition", "abstract": "  Research has shown that deep neural networks (DNNs) have vulnerabilities that\ncan lead to the misrecognition of Adversarial Examples (AEs) with specifically\ndesigned perturbations. Various adversarial attack methods have been proposed\nto detect vulnerabilities under hard-label black box (HL-BB) conditions in the\nabsence of loss gradients and confidence scores.However, these methods fall\ninto local solutions because they search only local regions of the search\nspace. Therefore, this study proposes an adversarial attack method named EvolBA\nto generate AEs using Covariance Matrix Adaptation Evolution Strategy (CMA-ES)\nunder the HL-BB condition, where only a class label predicted by the target DNN\nmodel is available. Inspired by formula-driven supervised learning, the\nproposed method introduces domain-independent operators for the initialization\nprocess and a jump that enhances search exploration. Experimental results\nconfirmed that the proposed method could determine AEs with smaller\nperturbations than previous methods in images where the previous methods have\ndifficulty.\n", "date": "2024-07-02", "categories": []}
{"id": "2407.06372", "title": "Non-Robust Features are Not Always Useful in One-Class Classification", "abstract": "  The robustness of machine learning models has been questioned by the\nexistence of adversarial examples. We examine the threat of adversarial\nexamples in practical applications that require lightweight models for\none-class classification. Building on Ilyas et al. (2019), we investigate the\nvulnerability of lightweight one-class classifiers to adversarial attacks and\npossible reasons for it. Our results show that lightweight one-class\nclassifiers learn features that are not robust (e.g. texture) under stronger\nattacks. However, unlike in multi-class classification (Ilyas et al., 2019),\nthese non-robust features are not always useful for the one-class task,\nsuggesting that learning these unpredictive and non-robust features is an\nunwanted consequence of training.\n", "date": "2024-07-08", "categories": []}
{"id": "2407.06552", "title": "DLOVE: A new Security Evaluation Tool for Deep Learning Based\n  Watermarking Techniques", "abstract": "  Recent developments in Deep Neural Network (DNN) based watermarking\ntechniques have shown remarkable performance. The state-of-the-art DNN-based\ntechniques not only surpass the robustness of classical watermarking techniques\nbut also show their robustness against many image manipulation techniques. In\nthis paper, we performed a detailed security analysis of different DNN-based\nwatermarking techniques. We propose a new class of attack called the Deep\nLearning-based OVErwriting (DLOVE) attack, which leverages adversarial machine\nlearning and overwrites the original embedded watermark with a targeted\nwatermark in a watermarked image. To the best of our knowledge, this attack is\nthe first of its kind. We have considered scenarios where watermarks are used\nto devise and formulate an adversarial attack in white box and black box\nsettings. To show adaptability and efficiency, we launch our DLOVE attack\nanalysis on seven different watermarking techniques, HiDDeN, ReDMark, PIMoG,\nStegastamp, Aparecium, Distortion Agostic Deep Watermarking and Hiding Images\nin an Image. All these techniques use different approaches to create\nimperceptible watermarked images. Our attack analysis on these watermarking\ntechniques with various constraints highlights the vulnerabilities of DNN-based\nwatermarking. Extensive experimental results validate the capabilities of\nDLOVE. We propose DLOVE as a benchmark security analysis tool to test the\nrobustness of future deep learning-based watermarking techniques.\n", "date": "2024-07-09", "categories": []}
{"id": "2407.06688", "title": "Universal Multi-view Black-box Attack against Object Detectors via\n  Layout Optimization", "abstract": "  Object detectors have demonstrated vulnerability to adversarial examples\ncrafted by small perturbations that can deceive the object detector. Existing\nadversarial attacks mainly focus on white-box attacks and are merely valid at a\nspecific viewpoint, while the universal multi-view black-box attack is less\nexplored, limiting their generalization in practice. In this paper, we propose\na novel universal multi-view black-box attack against object detectors, which\noptimizes a universal adversarial UV texture constructed by multiple image\nstickers for a 3D object via the designed layout optimization algorithm.\nSpecifically, we treat the placement of image stickers on the UV texture as a\ncircle-based layout optimization problem, whose objective is to find the\noptimal circle layout filled with image stickers so that it can deceive the\nobject detector under the multi-view scenario. To ensure reasonable placement\nof image stickers, two constraints are elaborately devised. To optimize the\nlayout, we adopt the random search algorithm enhanced by the devised\nimportant-aware selection strategy to find the most appropriate image sticker\nfor each circle from the image sticker pools. Extensive experiments conducted\non four common object detectors suggested that the detection performance\ndecreases by a large magnitude of 74.29% on average in multi-view scenarios.\nAdditionally, a novel evaluation tool based on the photo-realistic simulator is\ndesigned to assess the texture-based attack fairly.\n", "date": "2024-07-09", "categories": []}
{"id": "2407.06714", "title": "Improving the Transferability of Adversarial Examples by Feature\n  Augmentation", "abstract": "  Despite the success of input transformation-based attacks on boosting\nadversarial transferability, the performance is unsatisfying due to the\nignorance of the discrepancy across models. In this paper, we propose a simple\nbut effective feature augmentation attack (FAUG) method, which improves\nadversarial transferability without introducing extra computation costs.\nSpecifically, we inject the random noise into the intermediate features of the\nmodel to enlarge the diversity of the attack gradient, thereby mitigating the\nrisk of overfitting to the specific model and notably amplifying adversarial\ntransferability. Moreover, our method can be combined with existing gradient\nattacks to augment their performance further. Extensive experiments conducted\non the ImageNet dataset across CNN and transformer models corroborate the\nefficacy of our method, e.g., we achieve improvement of +26.22% and +5.57% on\ninput transformation-based attacks and combination methods, respectively.\n", "date": "2024-07-09", "categories": []}
{"id": "2407.06796", "title": "Countermeasures Against Adversarial Examples in Radio Signal\n  Classification", "abstract": "  Deep learning algorithms have been shown to be powerful in many communication\nnetwork design problems, including that in automatic modulation classification.\nHowever, they are vulnerable to carefully crafted attacks called adversarial\nexamples. Hence, the reliance of wireless networks on deep learning algorithms\nposes a serious threat to the security and operation of wireless networks. In\nthis letter, we propose for the first time a countermeasure against adversarial\nexamples in modulation classification. Our countermeasure is based on a neural\nrejection technique, augmented by label smoothing and Gaussian noise injection,\nthat allows to detect and reject adversarial examples with high accuracy. Our\nresults demonstrate that the proposed countermeasure can protect deep-learning\nbased modulation classification systems against adversarial examples.\n", "date": "2024-07-09", "categories": []}
{"id": "2407.06807", "title": "A Hybrid Training-time and Run-time Defense Against Adversarial Attacks\n  in Modulation Classification", "abstract": "  Motivated by the superior performance of deep learning in many applications\nincluding computer vision and natural language processing, several recent\nstudies have focused on applying deep neural network for devising future\ngenerations of wireless networks. However, several recent works have pointed\nout that imperceptible and carefully designed adversarial examples (attacks)\ncan significantly deteriorate the classification accuracy. In this paper, we\ninvestigate a defense mechanism based on both training-time and run-time\ndefense techniques for protecting machine learning-based radio signal\n(modulation) classification against adversarial attacks. The training-time\ndefense consists of adversarial training and label smoothing, while the\nrun-time defense employs a support vector machine-based neural rejection (NR).\nConsidering a white-box scenario and real datasets, we demonstrate that our\nproposed techniques outperform existing state-of-the-art technologies.\n", "date": "2024-07-09", "categories": []}
{"id": "2302.02213", "title": "CosPGD: an efficient white-box adversarial attack for pixel-wise\n  prediction tasks", "abstract": "  While neural networks allow highly accurate predictions in many tasks, their\nlack of robustness towards even slight input perturbations often hampers their\ndeployment. Adversarial attacks such as the seminal projected gradient descent\n(PGD) offer an effective means to evaluate a model's robustness and dedicated\nsolutions have been proposed for attacks on semantic segmentation or optical\nflow estimation. While they attempt to increase the attack's efficiency, a\nfurther objective is to balance its effect, so that it acts on the entire image\ndomain instead of isolated point-wise predictions. This often comes at the cost\nof optimization stability and thus efficiency. Here, we propose CosPGD, an\nattack that encourages more balanced errors over the entire image domain while\nincreasing the attack's overall efficiency. To this end, CosPGD leverages a\nsimple alignment score computed from any pixel-wise prediction and its target\nto scale the loss in a smooth and fully differentiable way. It leads to\nefficient evaluations of a model's robustness for semantic segmentation as well\nas regression models (such as optical flow, disparity estimation, or image\nrestoration), and it allows it to outperform the previous SotA attack on\nsemantic segmentation. We provide code for the CosPGD algorithm and example\nusage at https://github.com/shashankskagnihotri/cospgd.\n", "date": "2023-02-04", "categories": []}
{"id": "2304.06430", "title": "Certified Zeroth-order Black-Box Defense with Robust UNet Denoiser", "abstract": "  Certified defense methods against adversarial perturbations have been\nrecently investigated in the black-box setting with a zeroth-order (ZO)\nperspective. However, these methods suffer from high model variance with low\nperformance on high-dimensional datasets due to the ineffective design of the\ndenoiser and are limited in their utilization of ZO techniques. To this end, we\npropose a certified ZO preprocessing technique for removing adversarial\nperturbations from the attacked image in the black-box setting using only model\nqueries. We propose a robust UNet denoiser (RDUNet) that ensures the robustness\nof black-box models trained on high-dimensional datasets. We propose a novel\nblack-box denoised smoothing (DS) defense mechanism, ZO-RUDS, by prepending our\nRDUNet to the black-box model, ensuring black-box defense. We further propose\nZO-AE-RUDS in which RDUNet followed by autoencoder (AE) is prepended to the\nblack-box model. We perform extensive experiments on four classification\ndatasets, CIFAR-10, CIFAR-10, Tiny Imagenet, STL-10, and the MNIST dataset for\nimage reconstruction tasks. Our proposed defense methods ZO-RUDS and ZO-AE-RUDS\nbeat SOTA with a huge margin of $35\\%$ and $9\\%$, for low dimensional\n(CIFAR-10) and with a margin of $20.61\\%$ and $23.51\\%$ for high-dimensional\n(STL-10) datasets, respectively.\n", "date": "2023-04-13", "categories": []}
{"id": "2307.13885", "title": "Characterizing Data Point Vulnerability via Average-Case Robustness", "abstract": "  Studying the robustness of machine learning models is important to ensure\nconsistent model behaviour across real-world settings. To this end, adversarial\nrobustness is a standard framework, which views robustness of predictions\nthrough a binary lens: either a worst-case adversarial misclassification exists\nin the local region around an input, or it does not. However, this binary\nperspective does not account for the degrees of vulnerability, as data points\nwith a larger number of misclassified examples in their neighborhoods are more\nvulnerable. In this work, we consider a complementary framework for robustness,\ncalled average-case robustness, which measures the fraction of points in a\nlocal region that provides consistent predictions. However, computing this\nquantity is hard, as standard Monte Carlo approaches are inefficient especially\nfor high-dimensional inputs. In this work, we propose the first analytical\nestimators for average-case robustness for multi-class classifiers. We show\nempirically that our estimators are accurate and efficient for standard deep\nlearning models and demonstrate their usefulness for identifying vulnerable\ndata points, as well as quantifying robustness bias of models. Overall, our\ntools provide a complementary view to robustness, improving our ability to\ncharacterize model behaviour.\n", "date": "2023-07-26", "categories": []}
{"id": "2310.11868", "title": "To Generate or Not? Safety-Driven Unlearned Diffusion Models Are Still\n  Easy To Generate Unsafe Images ... For Now", "abstract": "  The recent advances in diffusion models (DMs) have revolutionized the\ngeneration of realistic and complex images. However, these models also\nintroduce potential safety hazards, such as producing harmful content and\ninfringing data copyrights. Despite the development of safety-driven unlearning\ntechniques to counteract these challenges, doubts about their efficacy persist.\nTo tackle this issue, we introduce an evaluation framework that leverages\nadversarial prompts to discern the trustworthiness of these safety-driven DMs\nafter they have undergone the process of unlearning harmful concepts.\nSpecifically, we investigated the adversarial robustness of DMs, assessed by\nadversarial prompts, when eliminating unwanted concepts, styles, and objects.\nWe develop an effective and efficient adversarial prompt generation approach\nfor DMs, termed UnlearnDiffAtk. This method capitalizes on the intrinsic\nclassification abilities of DMs to simplify the creation of adversarial\nprompts, thereby eliminating the need for auxiliary classification or diffusion\nmodels. Through extensive benchmarking, we evaluate the robustness of\nwidely-used safety-driven unlearned DMs (i.e., DMs after unlearning undesirable\nconcepts, styles, or objects) across a variety of tasks. Our results\ndemonstrate the effectiveness and efficiency merits of UnlearnDiffAtk over the\nstate-of-the-art adversarial prompt generation method and reveal the lack of\nrobustness of current safetydriven unlearning techniques when applied to DMs.\nCodes are available at https://github.com/OPTML-Group/Diffusion-MU-Attack.\nWARNING: There exist AI generations that may be offensive in nature.\n", "date": "2023-10-18", "categories": []}
{"id": "2402.04038", "title": "PAC-Bayesian Adversarially Robust Generalization Bounds for Graph Neural\n  Network", "abstract": "  Graph neural networks (GNNs) have gained popularity for various graph-related\ntasks. However, similar to deep neural networks, GNNs are also vulnerable to\nadversarial attacks. Empirical studies have shown that adversarially robust\ngeneralization has a pivotal role in establishing effective defense algorithms\nagainst adversarial attacks. In this paper, we contribute by providing\nadversarially robust generalization bounds for two kinds of popular GNNs, graph\nconvolutional network (GCN) and message passing graph neural network, using the\nPAC-Bayesian framework. Our result reveals that spectral norm of the diffusion\nmatrix on the graph and spectral norm of the weights as well as the\nperturbation factor govern the robust generalization bounds of both models. Our\nbounds are nontrivial generalizations of the results developed in (Liao et al.,\n2020) from the standard setting to adversarial setting while avoiding\nexponential dependence of the maximum node degree. As corollaries, we derive\nbetter PAC-Bayesian robust generalization bounds for GCN in the standard\nsetting, which improve the bounds in (Liao et al., 2020) by avoiding\nexponential dependence on the maximum node degree.\n", "date": "2024-02-06", "categories": []}
{"id": "2402.09132", "title": "Exploring the Adversarial Capabilities of Large Language Models", "abstract": "  The proliferation of large language models (LLMs) has sparked widespread and\ngeneral interest due to their strong language generation capabilities, offering\ngreat potential for both industry and research. While previous research delved\ninto the security and privacy issues of LLMs, the extent to which these models\ncan exhibit adversarial behavior remains largely unexplored. Addressing this\ngap, we investigate whether common publicly available LLMs have inherent\ncapabilities to perturb text samples to fool safety measures, so-called\nadversarial examples resp.~attacks. More specifically, we investigate whether\nLLMs are inherently able to craft adversarial examples out of benign samples to\nfool existing safe rails. Our experiments, which focus on hate speech\ndetection, reveal that LLMs succeed in finding adversarial perturbations,\neffectively undermining hate speech detection systems. Our findings carry\nsignificant implications for (semi-)autonomous systems relying on LLMs,\nhighlighting potential challenges in their interaction with existing systems\nand safety measures.\n", "date": "2024-02-14", "categories": []}
{"id": "2403.10883", "title": "Improving Adversarial Transferability of Vision-Language Pre-training\n  Models through Collaborative Multimodal Interaction", "abstract": "  Despite the substantial advancements in Vision-Language Pre-training (VLP)\nmodels, their susceptibility to adversarial attacks poses a significant\nchallenge. Existing work rarely studies the transferability of attacks on VLP\nmodels, resulting in a substantial performance gap from white-box attacks. We\nobserve that prior work overlooks the interaction mechanisms between\nmodalities, which plays a crucial role in understanding the intricacies of VLP\nmodels. In response, we propose a novel attack, called Collaborative Multimodal\nInteraction Attack (CMI-Attack), leveraging modality interaction through\nembedding guidance and interaction enhancement. Specifically, attacking text at\nthe embedding level while preserving semantics, as well as utilizing\ninteraction image gradients to enhance constraints on perturbations of texts\nand images. Significantly, in the image-text retrieval task on Flickr30K\ndataset, CMI-Attack raises the transfer success rates from ALBEF to TCL,\n$\\text{CLIP}_{\\text{ViT}}$ and $\\text{CLIP}_{\\text{CNN}}$ by 8.11%-16.75% over\nstate-of-the-art methods. Moreover, CMI-Attack also demonstrates superior\nperformance in cross-task generalization scenarios. Our work addresses the\nunderexplored realm of transfer attacks on VLP models, shedding light on the\nimportance of modality interaction for enhanced adversarial robustness.\n", "date": "2024-03-16", "categories": []}
{"id": "2404.04188", "title": "Reliable Feature Selection for Adversarially Robust Cyber-Attack\n  Detection", "abstract": "  The growing cybersecurity threats make it essential to use high-quality data\nto train Machine Learning (ML) models for network traffic analysis, without\nnoisy or missing data. By selecting the most relevant features for cyber-attack\ndetection, it is possible to improve both the robustness and computational\nefficiency of the models used in a cybersecurity system. This work presents a\nfeature selection and consensus process that combines multiple methods and\napplies them to several network datasets. Two different feature sets were\nselected and were used to train multiple ML models with regular and adversarial\ntraining. Finally, an adversarial evasion robustness benchmark was performed to\nanalyze the reliability of the different feature sets and their impact on the\nsusceptibility of the models to adversarial examples. By using an improved\ndataset with more data diversity, selecting the best time-related features and\na more specific feature set, and performing adversarial training, the ML models\nwere able to achieve a better adversarially robust generalization. The\nrobustness of the models was significantly improved without their\ngeneralization to regular traffic flows being affected, without increases of\nfalse alarms, and without requiring too many computational resources, which\nenables a reliable detection of suspicious activity and perturbed traffic flows\nin enterprise computer networks.\n", "date": "2024-04-05", "categories": []}
{"id": "2407.01168", "title": "Multi-View Black-Box Physical Attacks on Infrared Pedestrian Detectors\n  Using Adversarial Infrared Grid", "abstract": "  While extensive research exists on physical adversarial attacks within the\nvisible spectrum, studies on such techniques in the infrared spectrum are\nlimited. Infrared object detectors are vital in modern technological\napplications but are susceptible to adversarial attacks, posing significant\nsecurity threats. Previous studies using physical perturbations like light bulb\narrays and aerogels for white-box attacks, or hot and cold patches for\nblack-box attacks, have proven impractical or limited in multi-view support. To\naddress these issues, we propose the Adversarial Infrared Grid (AdvGrid), which\nmodels perturbations in a grid format and uses a genetic algorithm for\nblack-box optimization. These perturbations are cyclically applied to various\nparts of a pedestrian's clothing to facilitate multi-view black-box physical\nattacks on infrared pedestrian detectors. Extensive experiments validate\nAdvGrid's effectiveness, stealthiness, and robustness. The method achieves\nattack success rates of 80.00\\% in digital environments and 91.86\\% in physical\nenvironments, outperforming baseline methods. Additionally, the average attack\nsuccess rate exceeds 50\\% against mainstream detectors, demonstrating AdvGrid's\nrobustness. Our analyses include ablation studies, transfer attacks, and\nadversarial defenses, confirming the method's superiority.\n", "date": "2024-07-01", "categories": []}
{"id": "2407.05182", "title": "A Novel Bifurcation Method for Observation Perturbation Attacks on\n  Reinforcement Learning Agents: Load Altering Attacks on a Cyber Physical\n  Power System", "abstract": "  Components of cyber physical systems, which affect real-world processes, are\noften exposed to the internet. Replacing conventional control methods with Deep\nReinforcement Learning (DRL) in energy systems is an active area of research,\nas these systems become increasingly complex with the advent of renewable\nenergy sources and the desire to improve their efficiency. Artificial Neural\nNetworks (ANN) are vulnerable to specific perturbations of their inputs or\nfeatures, called adversarial examples. These perturbations are difficult to\ndetect when properly regularized, but have significant effects on the ANN's\noutput. Because DRL uses ANN to map optimal actions to observations, they are\nsimilarly vulnerable to adversarial examples. This work proposes a novel attack\ntechnique for continuous control using Group Difference Logits loss with a\nbifurcation layer. By combining aspects of targeted and untargeted attacks, the\nattack significantly increases the impact compared to an untargeted attack,\nwith drastically smaller distortions than an optimally targeted attack. We\ndemonstrate the impacts of powerful gradient-based attacks in a realistic smart\nenergy environment, show how the impacts change with different DRL agents and\ntraining procedures, and use statistical and time-series analysis to evaluate\nattacks' stealth. The results show that adversarial attacks can have\nsignificant impacts on DRL controllers, and constraining an attack's\nperturbations makes it difficult to detect. However, certain DRL architectures\nare far more robust, and robust training methods can further reduce the impact.\n", "date": "2024-07-06", "categories": []}
{"id": "2312.09665", "title": "FlowMur: A Stealthy and Practical Audio Backdoor Attack with Limited\n  Knowledge", "abstract": "  Speech recognition systems driven by DNNs have revolutionized human-computer\ninteraction through voice interfaces, which significantly facilitate our daily\nlives. However, the growing popularity of these systems also raises special\nconcerns on their security, particularly regarding backdoor attacks. A backdoor\nattack inserts one or more hidden backdoors into a DNN model during its\ntraining process, such that it does not affect the model's performance on\nbenign inputs, but forces the model to produce an adversary-desired output if a\nspecific trigger is present in the model input. Despite the initial success of\ncurrent audio backdoor attacks, they suffer from the following limitations: (i)\nMost of them require sufficient knowledge, which limits their widespread\nadoption. (ii) They are not stealthy enough, thus easy to be detected by\nhumans. (iii) Most of them cannot attack live speech, reducing their\npracticality. To address these problems, in this paper, we propose FlowMur, a\nstealthy and practical audio backdoor attack that can be launched with limited\nknowledge. FlowMur constructs an auxiliary dataset and a surrogate model to\naugment adversary knowledge. To achieve dynamicity, it formulates trigger\ngeneration as an optimization problem and optimizes the trigger over different\nattachment positions. To enhance stealthiness, we propose an adaptive data\npoisoning method according to Signal-to-Noise Ratio (SNR). Furthermore, ambient\nnoise is incorporated into the process of trigger generation and data poisoning\nto make FlowMur robust to ambient noise and improve its practicality. Extensive\nexperiments conducted on two datasets demonstrate that FlowMur achieves high\nattack performance in both digital and physical settings while remaining\nresilient to state-of-the-art defenses. In particular, a human study confirms\nthat triggers generated by FlowMur are not easily detected by participants.\n", "date": "2023-12-15", "categories": []}
{"id": "2407.04382", "title": "Self-Supervised Representation Learning for Adversarial Attack Detection", "abstract": "  Supervised learning-based adversarial attack detection methods rely on a\nlarge number of labeled data and suffer significant performance degradation\nwhen applying the trained model to new domains. In this paper, we propose a\nself-supervised representation learning framework for the adversarial attack\ndetection task to address this drawback. Firstly, we map the pixels of\naugmented input images into an embedding space. Then, we employ the\nprototype-wise contrastive estimation loss to cluster prototypes as latent\nvariables. Additionally, drawing inspiration from the concept of memory banks,\nwe introduce a discrimination bank to distinguish and learn representations for\neach individual instance that shares the same or a similar prototype,\nestablishing a connection between instances and their associated prototypes. We\npropose a parallel axial-attention (PAA)-based encoder to facilitate the\ntraining process by parallel training over height- and width-axis of attention\nmaps. Experimental results show that, compared to various benchmark\nself-supervised vision learning models and supervised adversarial attack\ndetection methods, the proposed model achieves state-of-the-art performance on\nthe adversarial attack detection task across a wide range of images.\n", "date": "2024-07-05", "categories": []}
{"id": "2407.03115", "title": "$L_p$-norm Distortion-Efficient Adversarial Attack", "abstract": "  Adversarial examples have shown a powerful ability to make a well-trained\nmodel misclassified. Current mainstream adversarial attack methods only\nconsider one of the distortions among $L_0$-norm, $L_2$-norm, and\n$L_\\infty$-norm. $L_0$-norm based methods cause large modification on a single\npixel, resulting in naked-eye visible detection, while $L_2$-norm and\n$L_\\infty$-norm based methods suffer from weak robustness against adversarial\ndefense since they always diffuse tiny perturbations to all pixels. A more\nrealistic adversarial perturbation should be sparse and imperceptible. In this\npaper, we propose a novel $L_p$-norm distortion-efficient adversarial attack,\nwhich not only owns the least $L_2$-norm loss but also significantly reduces\nthe $L_0$-norm distortion. To this aim, we design a new optimization scheme,\nwhich first optimizes an initial adversarial perturbation under $L_2$-norm\nconstraint, and then constructs a dimension unimportance matrix for the initial\nperturbation. Such a dimension unimportance matrix can indicate the adversarial\nunimportance of each dimension of the initial perturbation. Furthermore, we\nintroduce a new concept of adversarial threshold for the dimension unimportance\nmatrix. The dimensions of the initial perturbation whose unimportance is higher\nthan the threshold will be all set to zero, greatly decreasing the $L_0$-norm\ndistortion. Experimental results on three benchmark datasets show that under\nthe same query budget, the adversarial examples generated by our method have\nlower $L_0$-norm and $L_2$-norm distortion than the state-of-the-art.\nEspecially for the MNIST dataset, our attack reduces 8.1$\\%$ $L_2$-norm\ndistortion meanwhile remaining 47$\\%$ pixels unattacked. This demonstrates the\nsuperiority of the proposed method over its competitors in terms of adversarial\nrobustness and visual imperceptibility.\n", "date": "2024-07-03", "categories": []}
{"id": "2407.03232", "title": "Single Character Perturbations Break LLM Alignment", "abstract": "  When LLMs are deployed in sensitive, human-facing settings, it is crucial\nthat they do not output unsafe, biased, or privacy-violating outputs. For this\nreason, models are both trained and instructed to refuse to answer unsafe\nprompts such as \"Tell me how to build a bomb.\" We find that, despite these\nsafeguards, it is possible to break model defenses simply by appending a space\nto the end of a model's input. In a study of eight open-source models, we\ndemonstrate that this acts as a strong enough attack to cause the majority of\nmodels to generate harmful outputs with very high success rates. We examine the\ncauses of this behavior, finding that the contexts in which single spaces occur\nin tokenized training data encourage models to generate lists when prompted,\noverriding training signals to refuse to answer unsafe requests. Our findings\nunderscore the fragile state of current model alignment and promote the\nimportance of developing more robust alignment methods. Code and data will be\navailable at https://github.com/hannah-aught/space_attack.\n", "date": "2024-07-03", "categories": []}
{"id": "2402.11469", "title": "A Curious Case of Searching for the Correlation between Training Data\n  and Adversarial Robustness of Transformer Textual Models", "abstract": "  Existing works have shown that fine-tuned textual transformer models achieve\nstate-of-the-art prediction performances but are also vulnerable to adversarial\ntext perturbations. Traditional adversarial evaluation is often done\n\\textit{only after} fine-tuning the models and ignoring the training data. In\nthis paper, we want to prove that there is also a strong correlation between\ntraining data and model robustness. To this end, we extract 13 different\nfeatures representing a wide range of input fine-tuning corpora properties and\nuse them to predict the adversarial robustness of the fine-tuned models.\nFocusing mostly on encoder-only transformer models BERT and RoBERTa with\nadditional results for BART, ELECTRA, and GPT2, we provide diverse evidence to\nsupport our argument. First, empirical analyses show that (a) extracted\nfeatures can be used with a lightweight classifier such as Random Forest to\npredict the attack success rate effectively, and (b) features with the most\ninfluence on the model robustness have a clear correlation with the robustness.\nSecond, our framework can be used as a fast and effective additional tool for\nrobustness evaluation since it (a) saves 30x-193x runtime compared to the\ntraditional technique, (b) is transferable across models, (c) can be used under\nadversarial training, and (d) robust to statistical randomness. Our code is\npublicly available at \\url{https://github.com/CaptainCuong/RobustText_ACL2024}.\n", "date": "2024-02-18", "categories": []}
{"id": "2406.18382", "title": "Adversarial Search Engine Optimization for Large Language Models", "abstract": "  Large Language Models (LLMs) are increasingly used in applications where the\nmodel selects from competing third-party content, such as in LLM-powered search\nengines or chatbot plugins. In this paper, we introduce Preference Manipulation\nAttacks, a new class of attacks that manipulate an LLM's selections to favor\nthe attacker. We demonstrate that carefully crafted website content or plugin\ndocumentations can trick an LLM to promote the attacker products and discredit\ncompetitors, thereby increasing user traffic and monetization. We show this\nleads to a prisoner's dilemma, where all parties are incentivized to launch\nattacks, but the collective effect degrades the LLM's outputs for everyone. We\ndemonstrate our attacks on production LLM search engines (Bing and Perplexity)\nand plugin APIs (for GPT-4 and Claude). As LLMs are increasingly used to rank\nthird-party content, we expect Preference Manipulation Attacks to emerge as a\nsignificant threat.\n", "date": "2024-06-26", "categories": []}
{"id": "2407.01850", "title": "Purple-teaming LLMs with Adversarial Defender Training", "abstract": "  Existing efforts in safeguarding LLMs are limited in actively exposing the\nvulnerabilities of the target LLM and readily adapting to newly emerging safety\nrisks. To address this, we present Purple-teaming LLMs with Adversarial\nDefender training (PAD), a pipeline designed to safeguard LLMs by novelly\nincorporating the red-teaming (attack) and blue-teaming (safety training)\ntechniques. In PAD, we automatically collect conversational data that cover the\nvulnerabilities of an LLM around specific safety risks in a self-play manner,\nwhere the attacker aims to elicit unsafe responses and the defender generates\nsafe responses to these attacks. We then update both modules in a generative\nadversarial network style by training the attacker to elicit more unsafe\nresponses and updating the defender to identify them and explain the unsafe\nreason. Experimental results demonstrate that PAD significantly outperforms\nexisting baselines in both finding effective attacks and establishing a robust\nsafe guardrail. Furthermore, our findings indicate that PAD excels in striking\na balance between safety and overall model quality. We also reveal key\nchallenges in safeguarding LLMs, including defending multi-turn attacks and the\nneed for more delicate strategies to identify specific risks.\n", "date": "2024-07-01", "categories": []}
{"id": "2407.01916", "title": "Sequential Manipulation Against Rank Aggregation: Theory and Algorithm", "abstract": "  Rank aggregation with pairwise comparisons is widely encountered in\nsociology, politics, economics, psychology, sports, etc . Given the enormous\nsocial impact and the consequent incentives, the potential adversary has a\nstrong motivation to manipulate the ranking list. However, the ideal attack\nopportunity and the excessive adversarial capability cause the existing methods\nto be impractical. To fully explore the potential risks, we leverage an online\nattack on the vulnerable data collection process. Since it is independent of\nrank aggregation and lacks effective protection mechanisms, we disrupt the data\ncollection process by fabricating pairwise comparisons without knowledge of the\nfuture data or the true distribution. From the game-theoretic perspective, the\nconfrontation scenario between the online manipulator and the ranker who takes\ncontrol of the original data source is formulated as a distributionally robust\ngame that deals with the uncertainty of knowledge. Then we demonstrate that the\nequilibrium in the above game is potentially favorable to the adversary by\nanalyzing the vulnerability of the sampling algorithms such as Bernoulli and\nreservoir methods. According to the above theoretical analysis, different\nsequential manipulation policies are proposed under a Bayesian decision\nframework and a large class of parametric pairwise comparison models. For\nattackers with complete knowledge, we establish the asymptotic optimality of\nthe proposed policies. To increase the success rate of the sequential\nmanipulation with incomplete knowledge, a distributionally robust estimator,\nwhich replaces the maximum likelihood estimation in a saddle point problem,\nprovides a conservative data generation solution. Finally, the corroborating\nempirical evidence shows that the proposed method manipulates the results of\nrank aggregation methods in a sequential manner.\n", "date": "2024-07-02", "categories": []}
{"id": "2407.01925", "title": "Looking From the Future: Multi-order Iterations Can Enhance Adversarial\n  Attack Transferability", "abstract": "  Various methods try to enhance adversarial transferability by improving the\ngeneralization from different perspectives. In this paper, we rethink the\noptimization process and propose a novel sequence optimization concept, which\nis named Looking From the Future (LFF). LFF makes use of the original\noptimization process to refine the very first local optimization choice.\nAdapting the LFF concept to the adversarial attack task, we further propose an\nLFF attack as well as an MLFF attack with better generalization ability.\nFurthermore, guiding with the LFF concept, we propose an $LLF^{\\mathcal{N}}$\nattack which entends the LFF attack to a multi-order attack, further enhancing\nthe transfer attack ability. All our proposed methods can be directly applied\nto the iteration-based attack methods. We evaluate our proposed method on the\nImageNet1k dataset by applying several SOTA adversarial attack methods under\nfour kinds of tasks. Experimental results show that our proposed method can\ngreatly enhance the attack transferability. Ablation experiments are also\napplied to verify the effectiveness of each component. The source code will be\nreleased after this paper is accepted.\n", "date": "2024-07-02", "categories": []}
{"id": "2407.02240", "title": "MALT Powers Up Adversarial Attacks", "abstract": "  Current adversarial attacks for multi-class classifiers choose the target\nclass for a given input naively, based on the classifier's confidence levels\nfor various target classes. We present a novel adversarial targeting method,\n\\textit{MALT - Mesoscopic Almost Linearity Targeting}, based on medium-scale\nalmost linearity assumptions. Our attack wins over the current state of the art\nAutoAttack on the standard benchmark datasets CIFAR-100 and ImageNet and for a\nvariety of robust models. In particular, our attack is \\emph{five times faster}\nthan AutoAttack, while successfully matching all of AutoAttack's successes and\nattacking additional samples that were previously out of reach. We then prove\nformally and demonstrate empirically that our targeting method, although\ninspired by linear predictors, also applies to standard non-linear models.\n", "date": "2024-07-02", "categories": []}
{"id": "2404.12120", "title": "Fortify the Guardian, Not the Treasure: Resilient Adversarial Detectors", "abstract": "  This paper presents RADAR-Robust Adversarial Detection via Adversarial\nRetraining-an approach designed to enhance the robustness of adversarial\ndetectors against adaptive attacks, while maintaining classifier performance.\nAn adaptive attack is one where the attacker is aware of the defenses and\nadapts their strategy accordingly. Our proposed method leverages adversarial\ntraining to reinforce the ability to detect attacks, without compromising clean\naccuracy. During the training phase, we integrate into the dataset adversarial\nexamples, which were optimized to fool both the classifier and the adversarial\ndetector, enabling the adversarial detector to learn and adapt to potential\nattack scenarios. Experimental evaluations on the CIFAR-10 and SVHN datasets\ndemonstrate that our proposed algorithm significantly improves a detector's\nability to accurately identify adaptive adversarial attacks -- without\nsacrificing clean accuracy.\n", "date": "2024-04-18", "categories": []}
{"id": "2405.03672", "title": "Cutting through buggy adversarial example defenses: fixing 1 line of\n  code breaks Sabre", "abstract": "  Sabre is a defense to adversarial examples that was accepted at IEEE S&P\n2024. We first reveal significant flaws in the evaluation that point to clear\nsigns of gradient masking. We then show the cause of this gradient masking: a\nbug in the original evaluation code. By fixing a single line of code in the\noriginal repository, we reduce Sabre's robust accuracy to 0%. In response to\nthis, the authors modify the defense and introduce a new defense component not\ndescribed in the original paper. But this fix contains a second bug; modifying\none more line of code reduces robust accuracy to below baseline levels. After\nwe released the first version of our paper online, the authors introduced\nanother change to the defense; by commenting out one line of code during attack\nwe reduce the robust accuracy to 0% again.\n", "date": "2024-05-06", "categories": []}
{"id": "2406.04031", "title": "Jailbreak Vision Language Models via Bi-Modal Adversarial Prompt", "abstract": "  In the realm of large vision language models (LVLMs), jailbreak attacks serve\nas a red-teaming approach to bypass guardrails and uncover safety implications.\nExisting jailbreaks predominantly focus on the visual modality, perturbing\nsolely visual inputs in the prompt for attacks. However, they fall short when\nconfronted with aligned models that fuse visual and textual features\nsimultaneously for generation. To address this limitation, this paper\nintroduces the Bi-Modal Adversarial Prompt Attack (BAP), which executes\njailbreaks by optimizing textual and visual prompts cohesively. Initially, we\nadversarially embed universally harmful perturbations in an image, guided by a\nfew-shot query-agnostic corpus (e.g., affirmative prefixes and negative\ninhibitions). This process ensures that image prompt LVLMs to respond\npositively to any harmful queries. Subsequently, leveraging the adversarial\nimage, we optimize textual prompts with specific harmful intent. In particular,\nwe utilize a large language model to analyze jailbreak failures and employ\nchain-of-thought reasoning to refine textual prompts through a\nfeedback-iteration manner. To validate the efficacy of our approach, we\nconducted extensive evaluations on various datasets and LVLMs, demonstrating\nthat our method significantly outperforms other methods by large margins\n(+29.03% in attack success rate on average). Additionally, we showcase the\npotential of our attacks on black-box commercial LVLMs, such as Gemini and\nChatGLM.\n", "date": "2024-06-06", "categories": []}
{"id": "2407.00389", "title": "Query-Efficient Hard-Label Black-Box Attack against Vision Transformers", "abstract": "  Recent studies have revealed that vision transformers (ViTs) face similar\nsecurity risks from adversarial attacks as deep convolutional neural networks\n(CNNs). However, directly applying attack methodology on CNNs to ViTs has been\ndemonstrated to be ineffective since the ViTs typically work on patch-wise\nencoding. This article explores the vulnerability of ViTs against adversarial\nattacks under a black-box scenario, and proposes a novel query-efficient\nhard-label adversarial attack method called AdvViT. Specifically, considering\nthat ViTs are highly sensitive to patch modification, we propose to optimize\nthe adversarial perturbation on the individual patches. To reduce the dimension\nof perturbation search space, we modify only a handful of low-frequency\ncomponents of each patch. Moreover, we design a weight mask matrix for all\npatches to further optimize the perturbation on different regions of a whole\nimage. We test six mainstream ViT backbones on the ImageNet-1k dataset.\nExperimental results show that compared with the state-of-the-art attacks on\nCNNs, our AdvViT achieves much lower $L_2$-norm distortion under the same query\nbudget, sufficiently validating the vulnerability of ViTs against adversarial\nattacks.\n", "date": "2024-06-29", "categories": []}
{"id": "2407.01157", "title": "Unaligning Everything: Or Aligning Any Text to Any Image in Multimodal\n  Models", "abstract": "  Utilizing a shared embedding space, emerging multimodal models exhibit\nunprecedented zero-shot capabilities. However, the shared embedding space could\nlead to new vulnerabilities if different modalities can be misaligned. In this\npaper, we extend and utilize a recently developed effective gradient-based\nprocedure that allows us to match the embedding of a given text by minimally\nmodifying an image. Using the procedure, we show that we can align the\nembeddings of distinguishable texts to any image through unnoticeable\nadversarial attacks in joint image-text models, revealing that semantically\nunrelated images can have embeddings of identical texts and at the same time\nvisually indistinguishable images can be matched to the embeddings of very\ndifferent texts. Our technique achieves 100\\% success rate when it is applied\nto text datasets and images from multiple sources. Without overcoming the\nvulnerability, multimodal models cannot robustly align inputs from different\nmodalities in a semantically meaningful way. \\textbf{Warning: the text data\nused in this paper are toxic in nature and may be offensive to some readers.}\n", "date": "2024-07-01", "categories": []}
{"id": "2406.19417", "title": "\"Glue pizza and eat rocks\" -- Exploiting Vulnerabilities in\n  Retrieval-Augmented Generative Models", "abstract": "  Retrieval-Augmented Generative (RAG) models enhance Large Language Models\n(LLMs) by integrating external knowledge bases, improving their performance in\napplications like fact-checking and information searching. In this paper, we\ndemonstrate a security threat where adversaries can exploit the openness of\nthese knowledge bases by injecting deceptive content into the retrieval\ndatabase, intentionally changing the model's behavior. This threat is critical\nas it mirrors real-world usage scenarios where RAG systems interact with\npublicly accessible knowledge bases, such as web scrapings and user-contributed\ndata pools. To be more realistic, we target a realistic setting where the\nadversary has no knowledge of users' queries, knowledge base data, and the LLM\nparameters. We demonstrate that it is possible to exploit the model\nsuccessfully through crafted content uploads with access to the retriever. Our\nfindings emphasize an urgent need for security measures in the design and\ndeployment of RAG systems to prevent potential manipulation and ensure the\nintegrity of machine-generated content.\n", "date": "2024-06-26", "categories": []}
{"id": "2406.19807", "title": "Deceptive Diffusion: Generating Synthetic Adversarial Examples", "abstract": "  We introduce the concept of deceptive diffusion -- training a generative AI\nmodel to produce adversarial images. Whereas a traditional adversarial attack\nalgorithm aims to perturb an existing image to induce a misclassificaton, the\ndeceptive diffusion model can create an arbitrary number of new, misclassified\nimages that are not directly associated with training or test images. Deceptive\ndiffusion offers the possibility of strengthening defence algorithms by\nproviding adversarial training data at scale, including types of\nmisclassification that are otherwise difficult to find. In our experiments, we\nalso investigate the effect of training on a partially attacked data set. This\nhighlights a new type of vulnerability for generative diffusion models: if an\nattacker is able to stealthily poison a portion of the training data, then the\nresulting diffusion model will generate a similar proportion of misleading\noutputs.\n", "date": "2024-06-28", "categories": []}
{"id": "2406.19815", "title": "Emotion Loss Attacking: Adversarial Attack Perception for Skeleton based\n  on Multi-dimensional Features", "abstract": "  Adversarial attack on skeletal motion is a hot topic. However, existing\nresearches only consider part of dynamic features when measuring distance\nbetween skeleton graph sequences, which results in poor imperceptibility. To\nthis end, we propose a novel adversarial attack method to attack action\nrecognizers for skeletal motions. Firstly, our method systematically proposes a\ndynamic distance function to measure the difference between skeletal motions.\nMeanwhile, we innovatively introduce emotional features for complementary\ninformation. In addition, we use Alternating Direction Method of\nMultipliers(ADMM) to solve the constrained optimization problem, which\ngenerates adversarial samples with better imperceptibility to deceive the\nclassifiers. Experiments show that our method is effective on multiple action\nclassifiers and datasets. When the perturbation magnitude measured by l norms\nis the same, the dynamic perturbations generated by our method are much lower\nthan that of other methods. What's more, we are the first to prove the\neffectiveness of emotional features, and provide a new idea for measuring the\ndistance between skeletal motions.\n", "date": "2024-06-28", "categories": []}
{"id": "2312.01886", "title": "InstructTA: Instruction-Tuned Targeted Attack for Large Vision-Language\n  Models", "abstract": "  Large vision-language models (LVLMs) have demonstrated their incredible\ncapability in image understanding and response generation. However, this rich\nvisual interaction also makes LVLMs vulnerable to adversarial examples. In this\npaper, we formulate a novel and practical targeted attack scenario that the\nadversary can only know the vision encoder of the victim LVLM, without the\nknowledge of its prompts (which are often proprietary for service providers and\nnot publicly available) and its underlying large language model (LLM). This\npractical setting poses challenges to the cross-prompt and cross-model\ntransferability of targeted adversarial attack, which aims to confuse the LVLM\nto output a response that is semantically similar to the attacker's chosen\ntarget text. To this end, we propose an instruction-tuned targeted attack\n(dubbed \\textsc{InstructTA}) to deliver the targeted adversarial attack on\nLVLMs with high transferability. Initially, we utilize a public text-to-image\ngenerative model to \"reverse\" the target response into a target image, and\nemploy GPT-4 to infer a reasonable instruction $\\boldsymbol{p}^\\prime$ from the\ntarget response. We then form a local surrogate model (sharing the same vision\nencoder with the victim LVLM) to extract instruction-aware features of an\nadversarial image example and the target image, and minimize the distance\nbetween these two features to optimize the adversarial example. To further\nimprove the transferability with instruction tuning, we augment the instruction\n$\\boldsymbol{p}^\\prime$ with instructions paraphrased from GPT-4. Extensive\nexperiments demonstrate the superiority of our proposed method in targeted\nattack performance and transferability. The code is available at\nhttps://github.com/xunguangwang/InstructTA.\n", "date": "2023-12-04", "categories": []}
{"id": "2406.01179", "title": "Are AI-Generated Text Detectors Robust to Adversarial Perturbations?", "abstract": "  The widespread use of large language models (LLMs) has sparked concerns about\nthe potential misuse of AI-generated text, as these models can produce content\nthat closely resembles human-generated text. Current detectors for AI-generated\ntext (AIGT) lack robustness against adversarial perturbations, with even minor\nchanges in characters or words causing a reversal in distinguishing between\nhuman-created and AI-generated text. This paper investigates the robustness of\nexisting AIGT detection methods and introduces a novel detector, the Siamese\nCalibrated Reconstruction Network (SCRN). The SCRN employs a reconstruction\nnetwork to add and remove noise from text, extracting a semantic representation\nthat is robust to local perturbations. We also propose a siamese calibration\ntechnique to train the model to make equally confidence predictions under\ndifferent noise, which improves the model's robustness against adversarial\nperturbations. Experiments on four publicly available datasets show that the\nSCRN outperforms all baseline methods, achieving 6.5\\%-18.25\\% absolute\naccuracy improvement over the best baseline method under adversarial attacks.\nMoreover, it exhibits superior generalizability in cross-domain, cross-genre,\nand mixed-source scenarios. The code is available at\n\\url{https://github.com/CarlanLark/Robust-AIGC-Detector}.\n", "date": "2024-06-03", "categories": []}
{"id": "2406.14711", "title": "MultiAgent Collaboration Attack: Investigating Adversarial Attacks in\n  Large Language Model Collaborations via Debate", "abstract": "  Large Language Models (LLMs) have shown exceptional results on current\nbenchmarks when working individually. The advancement in their capabilities,\nalong with a reduction in parameter size and inference times, has facilitated\nthe use of these models as agents, enabling interactions among multiple models\nto execute complex tasks. Such collaborations offer several advantages,\nincluding the use of specialized models (e.g. coding), improved confidence\nthrough multiple computations, and enhanced divergent thinking, leading to more\ndiverse outputs. Thus, the collaborative use of language models is expected to\ngrow significantly in the coming years. In this work, we evaluate the behavior\nof a network of models collaborating through debate under the influence of an\nadversary. We introduce pertinent metrics to assess the adversary's\neffectiveness, focusing on system accuracy and model agreement. Our findings\nhighlight the importance of a model's persuasive ability in influencing others.\nAdditionally, we explore inference-time methods to generate more compelling\narguments and evaluate the potential of prompt-based mitigation as a defensive\nstrategy.\n", "date": "2024-06-20", "categories": []}
{"id": "2406.10285", "title": "I Don't Know You, But I Can Catch You: Real-Time Defense against Diverse\n  Adversarial Patches for Object Detectors", "abstract": "  Deep neural networks (DNNs) have revolutionized the field of computer vision\nlike object detection with their unparalleled performance. However, existing\nresearch has shown that DNNs are vulnerable to adversarial attacks. In the\nphysical world, an adversary could exploit adversarial patches to implement a\nHiding Attack (HA) which patches the target object to make it disappear from\nthe detector, and an Appearing Attack (AA) which fools the detector into\nmisclassifying the patch as a specific object. Recently, many defense methods\nfor detectors have been proposed to mitigate the potential threats of\nadversarial patches. However, such methods still have limitations in\ngeneralization, robustness and efficiency. Most defenses are only effective\nagainst the HA, leaving the detector vulnerable to the AA.\n  In this paper, we propose \\textit{NutNet}, an innovative model for detecting\nadversarial patches, with high generalization, robustness and efficiency. With\nexperiments for six detectors including YOLOv2-v4, SSD, Faster RCNN and DETR on\nboth digital and physical domains, the results show that our proposed method\ncan effectively defend against both the HA and AA, with only 0.4\\% sacrifice of\nthe clean performance. We compare NutNet with four baseline defense methods for\ndetectors, and our method exhibits an average defense performance that is over\n2.4 times and 4.7 times higher than existing approaches for HA and AA,\nrespectively. In addition, NutNet only increases the inference time by 8\\%,\nwhich can meet the real-time requirements of the detection systems. Demos of\nNutNet are available at: \\url{https://sites.google.com/view/nutnet}.\n", "date": "2024-06-12", "categories": []}
{"id": "2406.15093", "title": "ECLIPSE: Expunging Clean-label Indiscriminate Poisons via Sparse\n  Diffusion Purification", "abstract": "  Clean-label indiscriminate poisoning attacks add invisible perturbations to\ncorrectly labeled training images, thus dramatically reducing the\ngeneralization capability of the victim models. Recently, some defense\nmechanisms have been proposed such as adversarial training, image\ntransformation techniques, and image purification. However, these schemes are\neither susceptible to adaptive attacks, built on unrealistic assumptions, or\nonly effective against specific poison types, limiting their universal\napplicability. In this research, we propose a more universally effective,\npractical, and robust defense scheme called ECLIPSE. We first investigate the\nimpact of Gaussian noise on the poisons and theoretically prove that any kind\nof poison will be largely assimilated when imposing sufficient random noise. In\nlight of this, we assume the victim has access to an extremely limited number\nof clean images (a more practical scene) and subsequently enlarge this sparse\nset for training a denoising probabilistic model (a universal denoising tool).\nWe then begin by introducing Gaussian noise to absorb the poisons and then\napply the model for denoising, resulting in a roughly purified dataset.\nFinally, to address the trade-off of the inconsistency in the assimilation\nsensitivity of different poisons by Gaussian noise, we propose a lightweight\ncorruption compensation module to effectively eliminate residual poisons,\nproviding a more universal defense approach. Extensive experiments demonstrate\nthat our defense approach outperforms 10 state-of-the-art defenses. We also\npropose an adaptive attack against ECLIPSE and verify the robustness of our\ndefense scheme. Our code is available at https://github.com/CGCL-codes/ECLIPSE.\n", "date": "2024-06-21", "categories": []}
{"id": "2406.16925", "title": "Analyzing Multi-Head Attention on Trojan BERT Models", "abstract": "  This project investigates the behavior of multi-head attention in Transformer\nmodels, specifically focusing on the differences between benign and trojan\nmodels in the context of sentiment analysis. Trojan attacks cause models to\nperform normally on clean inputs but exhibit misclassifications when presented\nwith inputs containing predefined triggers. We characterize attention head\nfunctions in trojan and benign models, identifying specific 'trojan' heads and\nanalyzing their behavior.\n", "date": "2024-06-12", "categories": []}
{"id": "2406.17104", "title": "Automated Adversarial Discovery for Safety Classifiers", "abstract": "  Safety classifiers are critical in mitigating toxicity on online forums such\nas social media and in chatbots. Still, they continue to be vulnerable to\nemergent, and often innumerable, adversarial attacks. Traditional automated\nadversarial data generation methods, however, tend to produce attacks that are\nnot diverse, but variations of previously observed harm types. We formalize the\ntask of automated adversarial discovery for safety classifiers - to find new\nattacks along previously unseen harm dimensions that expose new weaknesses in\nthe classifier. We measure progress on this task along two key axes (1)\nadversarial success: does the attack fool the classifier? and (2) dimensional\ndiversity: does the attack represent a previously unseen harm type? Our\nevaluation of existing attack generation methods on the CivilComments toxicity\ntask reveals their limitations: Word perturbation attacks fail to fool\nclassifiers, while prompt-based LLM attacks have more adversarial success, but\nlack dimensional diversity. Even our best-performing prompt-based method finds\nnew successful attacks on unseen harm dimensions of attacks only 5\\% of the\ntime. Automatically finding new harmful dimensions of attack is crucial and\nthere is substantial headroom for future research on our new task.\n", "date": "2024-06-24", "categories": []}
{"id": "2406.17425", "title": "CuDA2: An approach for Incorporating Traitor Agents into Cooperative\n  Multi-Agent Systems", "abstract": "  Cooperative Multi-Agent Reinforcement Learning (CMARL) strategies are well\nknown to be vulnerable to adversarial perturbations. Previous works on\nadversarial attacks have primarily focused on white-box attacks that directly\nperturb the states or actions of victim agents, often in scenarios with a\nlimited number of attacks. However, gaining complete access to victim agents in\nreal-world environments is exceedingly difficult. To create more realistic\nadversarial attacks, we introduce a novel method that involves injecting\ntraitor agents into the CMARL system. We model this problem as a Traitor Markov\nDecision Process (TMDP), where traitors cannot directly attack the victim\nagents but can influence their formation or positioning through collisions. In\nTMDP, traitors are trained using the same MARL algorithm as the victim agents,\nwith their reward function set as the negative of the victim agents' reward.\nDespite this, the training efficiency for traitors remains low because it is\nchallenging for them to directly associate their actions with the victim\nagents' rewards. To address this issue, we propose the Curiosity-Driven\nAdversarial Attack (CuDA2) framework. CuDA2 enhances the efficiency and\naggressiveness of attacks on the specified victim agents' policies while\nmaintaining the optimal policy invariance of the traitors. Specifically, we\nemploy a pre-trained Random Network Distillation (RND) module, where the extra\nreward generated by the RND module encourages traitors to explore states\nunencountered by the victim agents. Extensive experiments on various scenarios\nfrom SMAC demonstrate that our CuDA2 framework offers comparable or superior\nadversarial attack capabilities compared to other baselines.\n", "date": "2024-06-25", "categories": []}
{"id": "2406.17606", "title": "Diffusion-based Adversarial Purification for Intrusion Detection", "abstract": "  The escalating sophistication of cyberattacks has encouraged the integration\nof machine learning techniques in intrusion detection systems, but the rise of\nadversarial examples presents a significant challenge. These crafted\nperturbations mislead ML models, enabling attackers to evade detection or\ntrigger false alerts. As a reaction, adversarial purification has emerged as a\ncompelling solution, particularly with diffusion models showing promising\nresults. However, their purification potential remains unexplored in the\ncontext of intrusion detection. This paper demonstrates the effectiveness of\ndiffusion models in purifying adversarial examples in network intrusion\ndetection. Through a comprehensive analysis of the diffusion parameters, we\nidentify optimal configurations maximizing adversarial robustness with minimal\nimpact on normal performance. Importantly, this study reveals insights into the\nrelationship between diffusion noise and diffusion steps, representing a novel\ncontribution to the field. Our experiments are carried out on two datasets and\nagainst 5 adversarial attacks. The implementation code is publicly available.\n", "date": "2024-06-25", "categories": []}
{"id": "2306.00816", "title": "Versatile Backdoor Attack with Visible, Semantic, Sample-Specific, and\n  Compatible Triggers", "abstract": "  Deep neural networks (DNNs) can be manipulated to exhibit specific behaviors\nwhen exposed to specific trigger patterns, without affecting their performance\non benign samples, dubbed \\textit{backdoor attack}. Currently, implementing\nbackdoor attacks in physical scenarios still faces significant challenges.\nPhysical attacks are labor-intensive and time-consuming, and the triggers are\nselected in a manual and heuristic way. Moreover, expanding digital attacks to\nphysical scenarios faces many challenges due to their sensitivity to visual\ndistortions and the absence of counterparts in the real world. To address these\nchallenges, we define a novel trigger called the \\textbf{V}isible,\n\\textbf{S}emantic, \\textbf{S}ample-Specific, and \\textbf{C}ompatible (VSSC)\ntrigger, to achieve effective, stealthy and robust simultaneously, which can\nalso be effectively deployed in the physical scenario using corresponding\nobjects. To implement the VSSC trigger, we propose an automated pipeline\ncomprising three modules: a trigger selection module that systematically\nidentifies suitable triggers leveraging large language models, a trigger\ninsertion module that employs generative models to seamlessly integrate\ntriggers into images, and a quality assessment module that ensures the natural\nand successful insertion of triggers through vision-language models. Extensive\nexperimental results and analysis validate the effectiveness, stealthiness, and\nrobustness of the VSSC trigger. It can not only maintain robustness under\nvisual distortions but also demonstrates strong practicality in the physical\nscenario. We hope that the proposed VSSC trigger and implementation approach\ncould inspire future studies on designing more practical triggers in backdoor\nattacks.\n", "date": "2023-06-01", "categories": []}
{"id": "2306.01902", "title": "Unlearnable Examples for Diffusion Models: Protect Data from\n  Unauthorized Exploitation", "abstract": "  Diffusion models have demonstrated remarkable performance in image generation\ntasks, paving the way for powerful AIGC applications. However, these\nwidely-used generative models can also raise security and privacy concerns,\nsuch as copyright infringement, and sensitive data leakage. To tackle these\nissues, we propose a method, Unlearnable Diffusion Perturbation, to safeguard\nimages from unauthorized exploitation. Our approach involves designing an\nalgorithm to generate sample-wise perturbation noise for each image to be\nprotected. This imperceptible protective noise makes the data almost\nunlearnable for diffusion models, i.e., diffusion models trained or fine-tuned\non the protected data cannot generate high-quality and diverse images related\nto the protected training data. Theoretically, we frame this as a max-min\noptimization problem and introduce EUDP, a noise scheduler-based method to\nenhance the effectiveness of the protective noise. We evaluate our methods on\nboth Denoising Diffusion Probabilistic Model and Latent Diffusion Models,\ndemonstrating that training diffusion models on the protected data lead to a\nsignificant reduction in the quality of the generated images. Especially, the\nexperimental results on Stable Diffusion demonstrate that our method\neffectively safeguards images from being used to train Diffusion Models in\nvarious tasks, such as training specific objects and styles. This achievement\nholds significant importance in real-world scenarios, as it contributes to the\nprotection of privacy and copyright against AI-generated content.\n", "date": "2023-06-02", "categories": []}
{"id": "2312.00084", "title": "Can Protective Perturbation Safeguard Personal Data from Being Exploited\n  by Stable Diffusion?", "abstract": "  Stable Diffusion has established itself as a foundation model in generative\nAI artistic applications, receiving widespread research and application. Some\nrecent fine-tuning methods have made it feasible for individuals to implant\npersonalized concepts onto the basic Stable Diffusion model with minimal\ncomputational costs on small datasets. However, these innovations have also\ngiven rise to issues like facial privacy forgery and artistic copyright\ninfringement. In recent studies, researchers have explored the addition of\nimperceptible adversarial perturbations to images to prevent potential\nunauthorized exploitation and infringements when personal data is used for\nfine-tuning Stable Diffusion. Although these studies have demonstrated the\nability to protect images, it is essential to consider that these methods may\nnot be entirely applicable in real-world scenarios. In this paper, we\nsystematically evaluate the use of perturbations to protect images within a\npractical threat model. The results suggest that these approaches may not be\nsufficient to safeguard image privacy and copyright effectively. Furthermore,\nwe introduce a purification method capable of removing protected perturbations\nwhile preserving the original image structure to the greatest extent possible.\nExperiments reveal that Stable Diffusion can effectively learn from purified\nimages over all protective methods.\n", "date": "2023-11-30", "categories": []}
{"id": "2406.15635", "title": "DataFreeShield: Defending Adversarial Attacks without Training Data", "abstract": "  Recent advances in adversarial robustness rely on an abundant set of training\ndata, where using external or additional datasets has become a common setting.\nHowever, in real life, the training data is often kept private for security and\nprivacy issues, while only the pretrained weight is available to the public. In\nsuch scenarios, existing methods that assume accessibility to the original data\nbecome inapplicable. Thus we investigate the pivotal problem of data-free\nadversarial robustness, where we try to achieve adversarial robustness without\naccessing any real data. Through a preliminary study, we highlight the severity\nof the problem by showing that robustness without the original dataset is\ndifficult to achieve, even with similar domain datasets. To address this issue,\nwe propose DataFreeShield, which tackles the problem from two perspectives:\nsurrogate dataset generation and adversarial training using the generated data.\nThrough extensive validation, we show that DataFreeShield outperforms\nbaselines, demonstrating that the proposed method sets the first entirely\ndata-free solution for the adversarial robustness problem.\n", "date": "2024-06-21", "categories": []}
{"id": "2406.15925", "title": "Federated Adversarial Learning for Robust Autonomous Landing Runway\n  Detection", "abstract": "  As the development of deep learning techniques in autonomous landing systems\ncontinues to grow, one of the major challenges is trust and security in the\nface of possible adversarial attacks. In this paper, we propose a federated\nadversarial learning-based framework to detect landing runways using paired\ndata comprising of clean local data and its adversarial version. Firstly, the\nlocal model is pre-trained on a large-scale lane detection dataset. Then,\ninstead of exploiting large instance-adaptive models, we resort to a\nparameter-efficient fine-tuning method known as scale and shift deep features\n(SSF), upon the pre-trained model. Secondly, in each SSF layer, distributions\nof clean local data and its adversarial version are disentangled for accurate\nstatistics estimation. To the best of our knowledge, this marks the first\ninstance of federated learning work that address the adversarial sample problem\nin landing runway detection. Our experimental evaluations over both synthesis\nand real images of Landing Approach Runway Detection (LARD) dataset\nconsistently demonstrate good performance of the proposed federated adversarial\nlearning and robust to adversarial attacks.\n", "date": "2024-06-22", "categories": []}
{"id": "2406.16275", "title": "Investigating the Influence of Prompt-Specific Shortcuts in AI Generated\n  Text Detection", "abstract": "  AI Generated Text (AIGT) detectors are developed with texts from humans and\nLLMs of common tasks. Despite the diversity of plausible prompt choices, these\ndatasets are generally constructed with a limited number of prompts. The lack\nof prompt variation can introduce prompt-specific shortcut features that exist\nin data collected with the chosen prompt, but do not generalize to others. In\nthis paper, we analyze the impact of such shortcuts in AIGT detection. We\npropose Feedback-based Adversarial Instruction List Optimization (FAILOpt), an\nattack that searches for instructions deceptive to AIGT detectors exploiting\nprompt-specific shortcuts. FAILOpt effectively drops the detection performance\nof the target detector, comparable to other attacks based on adversarial\nin-context examples. We also utilize our method to enhance the robustness of\nthe detector by mitigating the shortcuts. Based on the findings, we further\ntrain the classifier with the dataset augmented by FAILOpt prompt. The\naugmented classifier exhibits improvements across generation models, tasks, and\nattacks. Our code will be available at https://github.com/zxcvvxcz/FAILOpt.\n", "date": "2024-06-24", "categories": []}
{"id": "2406.16501", "title": "UNICAD: A Unified Approach for Attack Detection, Noise Reduction and\n  Novel Class Identification", "abstract": "  As the use of Deep Neural Networks (DNNs) becomes pervasive, their\nvulnerability to adversarial attacks and limitations in handling unseen classes\nposes significant challenges. The state-of-the-art offers discrete solutions\naimed to tackle individual issues covering specific adversarial attack\nscenarios, classification or evolving learning. However, real-world systems\nneed to be able to detect and recover from a wide range of adversarial attacks\nwithout sacrificing classification accuracy and to flexibly act in {\\bf unseen}\nscenarios. In this paper, UNICAD, is proposed as a novel framework that\nintegrates a variety of techniques to provide an adaptive solution.\n  For the targeted image classification, UNICAD achieves accurate image\nclassification, detects unseen classes, and recovers from adversarial attacks\nusing Prototype and Similarity-based DNNs with denoising autoencoders. Our\nexperiments performed on the CIFAR-10 dataset highlight UNICAD's effectiveness\nin adversarial mitigation and unseen class classification, outperforming\ntraditional models.\n", "date": "2024-06-24", "categories": []}
{"id": "2406.16609", "title": "Evaluating the Robustness of Deep-Learning Algorithm-Selection Models by\n  Evolving Adversarial Instances", "abstract": "  Deep neural networks (DNN) are increasingly being used to perform\nalgorithm-selection in combinatorial optimisation domains, particularly as they\naccommodate input representations which avoid designing and calculating\nfeatures. Mounting evidence from domains that use images as input shows that\ndeep convolutional networks are vulnerable to adversarial samples, in which a\nsmall perturbation of an instance can cause the DNN to misclassify. However, it\nremains unknown as to whether deep recurrent networks (DRN) which have recently\nbeen shown promise as algorithm-selectors in the bin-packing domain are equally\nvulnerable. We use an evolutionary algorithm (EA) to find perturbations of\ninstances from two existing benchmarks for online bin packing that cause\ntrained DRNs to misclassify: adversarial samples are successfully generated\nfrom up to 56% of the original instances depending on the dataset. Analysis of\nthe new misclassified instances sheds light on the `fragility' of some training\ninstances, i.e. instances where it is trivial to find a small perturbation that\nresults in a misclassification and the factors that influence this. Finally,\nthe method generates a large number of new instances misclassified with a wide\nvariation in confidence, providing a rich new source of training data to create\nmore robust models.\n", "date": "2024-06-24", "categories": []}
{"id": "2306.04959", "title": "FedSecurity: Benchmarking Attacks and Defenses in Federated Learning and\n  Federated LLMs", "abstract": "  This paper introduces FedSecurity, an end-to-end benchmark that serves as a\nsupplementary component of the FedML library for simulating adversarial attacks\nand corresponding defense mechanisms in Federated Learning (FL). FedSecurity\neliminates the need for implementing the fundamental FL procedures, e.g., FL\ntraining and data loading, from scratch, thus enables users to focus on\ndeveloping their own attack and defense strategies. It contains two key\ncomponents, including FedAttacker that conducts a variety of attacks during FL\ntraining, and FedDefender that implements defensive mechanisms to counteract\nthese attacks. FedSecurity has the following features: i) It offers extensive\ncustomization options to accommodate a broad range of machine learning models\n(e.g., Logistic Regression, ResNet, and GAN) and FL optimizers (e.g., FedAVG,\nFedOPT, and FedNOVA); ii) it enables exploring the effectiveness of attacks and\ndefenses across different datasets and models; and iii) it supports flexible\nconfiguration and customization through a configuration file and some APIs. We\nfurther demonstrate FedSecurity's utility and adaptability through federated\ntraining of Large Language Models (LLMs) to showcase its potential on a wide\nrange of complex applications.\n", "date": "2023-06-08", "categories": []}
{"id": "2402.02165", "title": "Towards Optimal Adversarial Robust Q-learning with Bellman\n  Infinity-error", "abstract": "  Establishing robust policies is essential to counter attacks or disturbances\naffecting deep reinforcement learning (DRL) agents. Recent studies explore\nstate-adversarial robustness and suggest the potential lack of an optimal\nrobust policy (ORP), posing challenges in setting strict robustness\nconstraints. This work further investigates ORP: At first, we introduce a\nconsistency assumption of policy (CAP) stating that optimal actions in the\nMarkov decision process remain consistent with minor perturbations, supported\nby empirical and theoretical evidence. Building upon CAP, we crucially prove\nthe existence of a deterministic and stationary ORP that aligns with the\nBellman optimal policy. Furthermore, we illustrate the necessity of\n$L^{\\infty}$-norm when minimizing Bellman error to attain ORP. This finding\nclarifies the vulnerability of prior DRL algorithms that target the Bellman\noptimal policy with $L^{1}$-norm and motivates us to train a Consistent\nAdversarial Robust Deep Q-Network (CAR-DQN) by minimizing a surrogate of\nBellman Infinity-error. The top-tier performance of CAR-DQN across various\nbenchmarks validates its practical effectiveness and reinforces the soundness\nof our theoretical analysis.\n", "date": "2024-02-03", "categories": []}
{"id": "2406.14859", "title": "From LLMs to MLLMs: Exploring the Landscape of Multimodal Jailbreaking", "abstract": "  The rapid development of Large Language Models (LLMs) and Multimodal Large\nLanguage Models (MLLMs) has exposed vulnerabilities to various adversarial\nattacks. This paper provides a comprehensive overview of jailbreaking research\ntargeting both LLMs and MLLMs, highlighting recent advancements in evaluation\nbenchmarks, attack techniques and defense strategies. Compared to the more\nadvanced state of unimodal jailbreaking, multimodal domain remains\nunderexplored. We summarize the limitations and potential research directions\nof multimodal jailbreaking, aiming to inspire future research and further\nenhance the robustness and security of MLLMs.\n", "date": "2024-06-21", "categories": []}
{"id": "2311.09641", "title": "RLHFPoison: Reward Poisoning Attack for Reinforcement Learning with\n  Human Feedback in Large Language Models", "abstract": "  Reinforcement Learning with Human Feedback (RLHF) is a methodology designed\nto align Large Language Models (LLMs) with human preferences, playing an\nimportant role in LLMs alignment. Despite its advantages, RLHF relies on human\nannotators to rank the text, which can introduce potential security\nvulnerabilities if any adversarial annotator (i.e., attackers) manipulates the\nranking score by up-ranking any malicious text to steer the LLM adversarially.\nTo assess the red-teaming of RLHF against human preference data poisoning, we\npropose RankPoison, a poisoning attack method on candidates' selection of\npreference rank flipping to reach certain malicious behaviors (e.g., generating\nlonger sequences, which can increase the computational cost). With poisoned\ndataset generated by RankPoison, we can perform poisoning attacks on LLMs to\ngenerate longer tokens without hurting the original safety alignment\nperformance. Moreover, applying RankPoison, we also successfully implement a\nbackdoor attack where LLMs can generate longer answers under questions with the\ntrigger word. Our findings highlight critical security challenges in RLHF,\nunderscoring the necessity for more robust alignment methods for LLMs.\n", "date": "2023-11-16", "categories": []}
{"id": "2405.11547", "title": "Certified Robust Accuracy of Neural Networks Are Bounded due to Bayes\n  Errors", "abstract": "  Adversarial examples pose a security threat to many critical systems built on\nneural networks. While certified training improves robustness, it also\ndecreases accuracy noticeably. Despite various proposals for addressing this\nissue, the significant accuracy drop remains. More importantly, it is not clear\nwhether there is a certain fundamental limit on achieving robustness whilst\nmaintaining accuracy. In this work, we offer a novel perspective based on Bayes\nerrors. By adopting Bayes error to robustness analysis, we investigate the\nlimit of certified robust accuracy, taking into account data distribution\nuncertainties. We first show that the accuracy inevitably decreases in the\npursuit of robustness due to changed Bayes error in the altered data\ndistribution. Subsequently, we establish an upper bound for certified robust\naccuracy, considering the distribution of individual classes and their\nboundaries. Our theoretical results are empirically evaluated on real-world\ndatasets and are shown to be consistent with the limited success of existing\ncertified training results, e.g., for CIFAR10, our analysis results in an upper\nbound (of certified robust accuracy) of 67.49\\%, meanwhile existing approaches\nare only able to increase it from 53.89\\% in 2017 to 62.84\\% in 2023.\n", "date": "2024-05-19", "categories": []}
{"id": "2406.05857", "title": "Self-supervised Adversarial Training of Monocular Depth Estimation\n  against Physical-World Attacks", "abstract": "  Monocular Depth Estimation (MDE) plays a vital role in applications such as\nautonomous driving. However, various attacks target MDE models, with physical\nattacks posing significant threats to system security. Traditional adversarial\ntraining methods, which require ground-truth labels, are not directly\napplicable to MDE models that lack ground-truth depth. Some self-supervised\nmodel hardening techniques (e.g., contrastive learning) overlook the domain\nknowledge of MDE, resulting in suboptimal performance. In this work, we\nintroduce a novel self-supervised adversarial training approach for MDE models,\nleveraging view synthesis without the need for ground-truth depth. We enhance\nadversarial robustness against real-world attacks by incorporating\nL_0-norm-bounded perturbation during training. We evaluate our method against\nsupervised learning-based and contrastive learning-based approaches\nspecifically designed for MDE. Our experiments with two representative MDE\nnetworks demonstrate improved robustness against various adversarial attacks,\nwith minimal impact on benign performance.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.13066", "title": "MaskPure: Improving Defense Against Text Adversaries with Stochastic\n  Purification", "abstract": "  The improvement of language model robustness, including successful defense\nagainst adversarial attacks, remains an open problem. In computer vision\nsettings, the stochastic noising and de-noising process provided by diffusion\nmodels has proven useful for purifying input images, thus improving model\nrobustness against adversarial attacks. Similarly, some initial work has\nexplored the use of random noising and de-noising to mitigate adversarial\nattacks in an NLP setting, but improving the quality and efficiency of these\nmethods is necessary for them to remain competitive. We extend upon methods of\ninput text purification that are inspired by diffusion processes, which\nrandomly mask and refill portions of the input text before classification. Our\nnovel method, MaskPure, exceeds or matches robustness compared to other\ncontemporary defenses, while also requiring no adversarial classifier training\nand without assuming knowledge of the attack type. In addition, we show that\nMaskPure is provably certifiably robust. To our knowledge, MaskPure is the\nfirst stochastic-purification method with demonstrated success against both\ncharacter-level and word-level attacks, indicating the generalizable and\npromising nature of stochastic denoising defenses. In summary: the MaskPure\nalgorithm bridges literature on the current strongest certifiable and empirical\nadversarial defense methods, showing that both theoretical and practical\nrobustness can be obtained together. Code is available on GitHub at\nhttps://github.com/hubarruby/MaskPure.\n", "date": "2024-06-18", "categories": []}
{"id": "2406.13228", "title": "AGSOA:Graph Neural Network Targeted Attack Based on Average Gradient and\n  Structure Optimization", "abstract": "  Graph Neural Networks(GNNs) are vulnerable to adversarial attack that cause\nperformance degradation by adding small perturbations to the graph.\nGradient-based attacks are one of the most commonly used methods and have\nachieved good performance in many attack scenarios. However, current gradient\nattacks face the problems of easy to fall into local optima and poor attack\ninvisibility. Specifically, most gradient attacks use greedy strategies to\ngenerate perturbations, which tend to fall into local optima leading to\nunderperformance of the attack. In addition, many attacks only consider the\neffectiveness of the attack and ignore the invisibility of the attack, making\nthe attacks easily exposed leading to failure. To address the above problems,\nthis paper proposes an attack on GNNs, called AGSOA, which consists of an\naverage gradient calculation and a structre optimization module. In the average\ngradient calculation module, we compute the average of the gradient information\nover all moments to guide the attack to generate perturbed edges, which\nstabilizes the direction of the attack update and gets rid of undesirable local\nmaxima. In the structure optimization module, we calculate the similarity and\nhomogeneity of the target node's with other nodes to adjust the graph structure\nso as to improve the invisibility and transferability of the attack. Extensive\nexperiments on three commonly used datasets show that AGSOA improves the\nmisclassification rate by 2$\\%$-8$\\%$ compared to other state-of-the-art\nmodels.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.13294", "title": "Enhancing Cross-Prompt Transferability in Vision-Language Models through\n  Contextual Injection of Target Tokens", "abstract": "  Vision-language models (VLMs) seamlessly integrate visual and textual data to\nperform tasks such as image classification, caption generation, and visual\nquestion answering. However, adversarial images often struggle to deceive all\nprompts effectively in the context of cross-prompt migration attacks, as the\nprobability distribution of the tokens in these images tends to favor the\nsemantics of the original image rather than the target tokens. To address this\nchallenge, we propose a Contextual-Injection Attack (CIA) that employs\ngradient-based perturbation to inject target tokens into both visual and\ntextual contexts, thereby improving the probability distribution of the target\ntokens. By shifting the contextual semantics towards the target tokens instead\nof the original image semantics, CIA enhances the cross-prompt transferability\nof adversarial images.Extensive experiments on the BLIP2, InstructBLIP, and\nLLaVA models show that CIA outperforms existing methods in cross-prompt\ntransferability, demonstrating its potential for more effective adversarial\nstrategies in VLMs.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.13499", "title": "GraphMU: Repairing Robustness of Graph Neural Networks via Machine\n  Unlearning", "abstract": "  Graph Neural Networks (GNNs) have demonstrated significant application\npotential in various fields. However, GNNs are still vulnerable to adversarial\nattacks. Numerous adversarial defense methods on GNNs are proposed to address\nthe problem of adversarial attacks. However, these methods can only serve as a\ndefense before poisoning, but cannot repair poisoned GNN. Therefore, there is\nan urgent need for a method to repair poisoned GNN. In this paper, we address\nthis gap by introducing the novel concept of model repair for GNNs. We propose\na repair framework, Repairing Robustness of Graph Neural Networks via Machine\nUnlearning (GraphMU), which aims to fine-tune poisoned GNN to forget\nadversarial samples without the need for complete retraining. We also introduce\na unlearning validation method to ensure that our approach effectively forget\nspecified poisoned data. To evaluate the effectiveness of GraphMU, we explore\nthree fine-tuned subgraph construction scenarios based on the available\nperturbation information: (i) Known Perturbation Ratios, (ii) Known Complete\nKnowledge of Perturbations, and (iii) Unknown any Knowledge of Perturbations.\nOur extensive experiments, conducted across four citation datasets and four\nadversarial attack scenarios, demonstrate that GraphMU can effectively restore\nthe performance of poisoned GNN.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.14073", "title": "Exploring Layerwise Adversarial Robustness Through the Lens of t-SNE", "abstract": "  Adversarial examples, designed to trick Artificial Neural Networks (ANNs)\ninto producing wrong outputs, highlight vulnerabilities in these models.\nExploring these weaknesses is crucial for developing defenses, and so, we\npropose a method to assess the adversarial robustness of image-classifying\nANNs. The t-distributed Stochastic Neighbor Embedding (t-SNE) technique is used\nfor visual inspection, and a metric, which compares the clean and perturbed\nembeddings, helps pinpoint weak spots in the layers. Analyzing two ANNs on\nCIFAR-10, one designed by humans and another via NeuroEvolution, we found that\ndifferences between clean and perturbed representations emerge early on, in the\nfeature extraction layers, affecting subsequent classification. The findings\nwith our metric are supported by the visual analysis of the t-SNE maps.\n", "date": "2024-06-20", "categories": []}
{"id": "2406.14259", "title": "MEAT: Median-Ensemble Adversarial Training for Improving Robustness and\n  Generalization", "abstract": "  Self-ensemble adversarial training methods improve model robustness by\nensembling models at different training epochs, such as model weight averaging\n(WA). However, previous research has shown that self-ensemble defense methods\nin adversarial training (AT) still suffer from robust overfitting, which\nseverely affects the generalization performance. Empirically, in the late\nphases of training, the AT becomes more overfitting to the extent that the\nindividuals for weight averaging also suffer from overfitting and produce\nanomalous weight values, which causes the self-ensemble model to continue to\nundergo robust overfitting due to the failure in removing the weight anomalies.\nTo solve this problem, we aim to tackle the influence of outliers in the weight\nspace in this work and propose an easy-to-operate and effective Median-Ensemble\nAdversarial Training (MEAT) method to solve the robust overfitting phenomenon\nexisting in self-ensemble defense from the source by searching for the median\nof the historical model weights. Experimental results show that MEAT achieves\nthe best robustness against the powerful AutoAttack and can effectively\nallievate the robust overfitting. We further demonstrate that most defense\nmethods can improve robust generalization and robustness by combining with\nMEAT.\n", "date": "2024-06-20", "categories": []}
{"id": "2306.08041", "title": "Data Poisoning to Fake a Nash Equilibrium in Markov Games", "abstract": "  We characterize offline data poisoning attacks on Multi-Agent Reinforcement\nLearning (MARL), where an attacker may change a data set in an attempt to\ninstall a (potentially fictitious) unique Markov-perfect Nash equilibrium for a\ntwo-player zero-sum Markov game. We propose the unique Nash set, namely the set\nof games, specified by their Q functions, with a specific joint policy being\nthe unique Nash equilibrium. The unique Nash set is central to poisoning\nattacks because the attack is successful if and only if data poisoning pushes\nall plausible games inside the set. The unique Nash set generalizes the reward\npolytope commonly used in inverse reinforcement learning to MARL. For zero-sum\nMarkov games, both the inverse Nash set and the set of plausible games induced\nby data are polytopes in the Q function space. We exhibit a linear program to\nefficiently compute the optimal poisoning attack. Our work sheds light on the\nstructure of data poisoning attacks on offline MARL, a necessary step before\none can design more robust MARL algorithms.\n", "date": "2023-06-13", "categories": []}
{"id": "2406.12605", "title": "Attack and Defense of Deep Learning Models in the Field of Web Attack\n  Detection", "abstract": "  The challenge of WAD (web attack detection) is growing as hackers\ncontinuously refine their methods to evade traditional detection. Deep learning\nmodels excel in handling complex unknown attacks due to their strong\ngeneralization and adaptability. However, they are vulnerable to backdoor\nattacks, where contextually irrelevant fragments are inserted into requests,\ncompromising model stability. While backdoor attacks are well studied in image\nrecognition, they are largely unexplored in WAD. This paper introduces backdoor\nattacks in WAD, proposing five methods and corresponding defenses. Testing on\ntextCNN, biLSTM, and tinybert models shows an attack success rate over 87%,\nreducible through fine-tuning. Future research should focus on backdoor\ndefenses in WAD. All the code and data of this paper can be obtained at\nhttps://anonymous.4open.science/r/attackDefenceinDL-7E05\n", "date": "2024-06-18", "categories": []}
{"id": "2306.16170", "title": "Mitigating Accuracy-Robustness Trade-off via Balanced Multi-Teacher\n  Adversarial Distillation", "abstract": "  Adversarial Training is a practical approach for improving the robustness of\ndeep neural networks against adversarial attacks. Although bringing reliable\nrobustness, the performance towards clean examples is negatively affected after\nAdversarial Training, which means a trade-off exists between accuracy and\nrobustness. Recently, some studies have tried to use knowledge distillation\nmethods in Adversarial Training, achieving competitive performance in improving\nthe robustness but the accuracy for clean samples is still limited. In this\npaper, to mitigate the accuracy-robustness trade-off, we introduce the Balanced\nMulti-Teacher Adversarial Robustness Distillation (B-MTARD) to guide the\nmodel's Adversarial Training process by applying a strong clean teacher and a\nstrong robust teacher to handle the clean examples and adversarial examples,\nrespectively. During the optimization process, to ensure that different\nteachers show similar knowledge scales, we design the Entropy-Based Balance\nalgorithm to adjust the teacher's temperature and keep the teachers'\ninformation entropy consistent. Besides, to ensure that the student has a\nrelatively consistent learning speed from multiple teachers, we propose the\nNormalization Loss Balance algorithm to adjust the learning weights of\ndifferent types of knowledge. A series of experiments conducted on three public\ndatasets demonstrate that B-MTARD outperforms the state-of-the-art methods\nagainst various adversarial attacks.\n", "date": "2023-06-28", "categories": []}
{"id": "2311.02373", "title": "From Trojan Horses to Castle Walls: Unveiling Bilateral Data Poisoning\n  Effects in Diffusion Models", "abstract": "  While state-of-the-art diffusion models (DMs) excel in image generation,\nconcerns regarding their security persist. Earlier research highlighted DMs'\nvulnerability to data poisoning attacks, but these studies placed stricter\nrequirements than conventional methods like `BadNets' in image classification.\nThis is because the art necessitates modifications to the diffusion training\nand sampling procedures. Unlike the prior work, we investigate whether\nBadNets-like data poisoning methods can directly degrade the generation by DMs.\nIn other words, if only the training dataset is contaminated (without\nmanipulating the diffusion process), how will this affect the performance of\nlearned DMs? In this setting, we uncover bilateral data poisoning effects that\nnot only serve an adversarial purpose (compromising the functionality of DMs)\nbut also offer a defensive advantage (which can be leveraged for defense in\nclassification tasks against poisoning attacks). We show that a BadNets-like\ndata poisoning attack remains effective in DMs for producing incorrect images\n(misaligned with the intended text conditions). Meanwhile, poisoned DMs exhibit\nan increased ratio of triggers, a phenomenon we refer to as `trigger\namplification', among the generated images. This insight can be then used to\nenhance the detection of poisoned training data. In addition, even under a low\npoisoning ratio, studying the poisoning effects of DMs is also valuable for\ndesigning robust image classifiers against such attacks. Last but not least, we\nestablish a meaningful linkage between data poisoning and the phenomenon of\ndata replications by exploring DMs' inherent data memorization tendencies.\n", "date": "2023-11-04", "categories": []}
{"id": "2312.00198", "title": "Optimal Attack and Defense for Reinforcement Learning", "abstract": "  To ensure the usefulness of Reinforcement Learning (RL) in real systems, it\nis crucial to ensure they are robust to noise and adversarial attacks. In\nadversarial RL, an external attacker has the power to manipulate the victim\nagent's interaction with the environment. We study the full class of online\nmanipulation attacks, which include (i) state attacks, (ii) observation attacks\n(which are a generalization of perceived-state attacks), (iii) action attacks,\nand (iv) reward attacks. We show the attacker's problem of designing a stealthy\nattack that maximizes its own expected reward, which often corresponds to\nminimizing the victim's value, is captured by a Markov Decision Process (MDP)\nthat we call a meta-MDP since it is not the true environment but a higher level\nenvironment induced by the attacked interaction. We show that the attacker can\nderive optimal attacks by planning in polynomial time or learning with\npolynomial sample complexity using standard RL techniques. We argue that the\noptimal defense policy for the victim can be computed as the solution to a\nstochastic Stackelberg game, which can be further simplified into a\npartially-observable turn-based stochastic game (POTBSG). Neither the attacker\nnor the victim would benefit from deviating from their respective optimal\npolicies, thus such solutions are truly robust. Although the defense problem is\nNP-hard, we show that optimal Markovian defenses can be computed (learned) in\npolynomial time (sample complexity) in many scenarios.\n", "date": "2023-11-30", "categories": []}
{"id": "2401.10111", "title": "Adapters Mixup: Mixing Parameter-Efficient Adapters to Enhance the\n  Adversarial Robustness of Fine-tuned Pre-trained Text Classifiers", "abstract": "  Existing works show that augmenting the training data of pre-trained language\nmodels (PLMs) for classification tasks fine-tuned via parameter-efficient\nfine-tuning methods (PEFT) using both clean and adversarial examples can\nenhance their robustness under adversarial attacks. However, this adversarial\ntraining paradigm often leads to performance degradation on clean inputs and\nrequires frequent re-training on the entire data to account for new, unknown\nattacks. To overcome these challenges while still harnessing the benefits of\nadversarial training and the efficiency of PEFT, this work proposes a novel\napproach, called AdpMixup, that combines two paradigms: (1) fine-tuning through\nadapters and (2) adversarial augmentation via mixup to dynamically leverage\nexisting knowledge from a set of pre-known attacks for robust inference.\nIntuitively, AdpMixup fine-tunes PLMs with multiple adapters with both clean\nand pre-known adversarial examples and intelligently mixes them up in different\nratios during prediction. Our experiments show AdpMixup achieves the best\ntrade-off between training efficiency and robustness under both pre-known and\nunknown attacks, compared to existing baselines on five downstream tasks across\nsix varied black-box attacks and 2 PLMs. All source code will be available.\n", "date": "2024-01-18", "categories": []}
{"id": "2402.17916", "title": "Adversarial Math Word Problem Generation", "abstract": "  Large language models (LLMs) have significantly transformed the educational\nlandscape. As current plagiarism detection tools struggle to keep pace with\nLLMs' rapid advancements, the educational community faces the challenge of\nassessing students' true problem-solving abilities in the presence of LLMs. In\nthis work, we explore a new paradigm for ensuring fair evaluation -- generating\nadversarial examples which preserve the structure and difficulty of the\noriginal questions aimed for assessment, but are unsolvable by LLMs. Focusing\non the domain of math word problems, we leverage abstract syntax trees to\nstructurally generate adversarial examples that cause LLMs to produce incorrect\nanswers by simply editing the numeric values in the problems. We conduct\nexperiments on various open- and closed-source LLMs, quantitatively and\nqualitatively demonstrating that our method significantly degrades their math\nproblem-solving ability. We identify shared vulnerabilities among LLMs and\npropose a cost-effective approach to attack high-cost models. Additionally, we\nconduct automatic analysis to investigate the cause of failure, providing\nfurther insights into the limitations of LLMs.\n", "date": "2024-02-27", "categories": []}
{"id": "2406.10933", "title": "Improving Adversarial Robustness via Decoupled Visual Representation\n  Masking", "abstract": "  Deep neural networks are proven to be vulnerable to fine-designed adversarial\nexamples, and adversarial defense algorithms draw more and more attention\nnowadays. Pre-processing based defense is a major strategy, as well as learning\nrobust feature representation has been proven an effective way to boost\ngeneralization. However, existing defense works lack considering different\ndepth-level visual features in the training process. In this paper, we first\nhighlight two novel properties of robust features from the feature distribution\nperspective: 1) \\textbf{Diversity}. The robust feature of intra-class samples\ncan maintain appropriate diversity; 2) \\textbf{Discriminability}. The robust\nfeature of inter-class samples should ensure adequate separation. We find that\nstate-of-the-art defense methods aim to address both of these mentioned issues\nwell. It motivates us to increase intra-class variance and decrease inter-class\ndiscrepancy simultaneously in adversarial training. Specifically, we propose a\nsimple but effective defense based on decoupled visual representation masking.\nThe designed Decoupled Visual Feature Masking (DFM) block can adaptively\ndisentangle visual discriminative features and non-visual features with diverse\nmask strategies, while the suitable discarding information can disrupt\nadversarial noise to improve robustness. Our work provides a generic and\neasy-to-plugin block unit for any former adversarial training algorithm to\nachieve better protection integrally. Extensive experimental results prove the\nproposed method can achieve superior performance compared with state-of-the-art\ndefense approaches. The code is publicly available at\n\\href{https://github.com/chenboluo/Adversarial-defense}{https://github.com/chenboluo/Adversarial-defense}.\n", "date": "2024-06-16", "categories": []}
{"id": "2406.11576", "title": "Harmonizing Feature Maps: A Graph Convolutional Approach for Enhancing\n  Adversarial Robustness", "abstract": "  The vulnerability of Deep Neural Networks to adversarial perturbations\npresents significant security concerns, as the imperceptible perturbations can\ncontaminate the feature space and lead to incorrect predictions. Recent studies\nhave attempted to calibrate contaminated features by either suppressing or\nover-activating particular channels. Despite these efforts, we claim that\nadversarial attacks exhibit varying disruption levels across individual\nchannels. Furthermore, we argue that harmonizing feature maps via graph and\nemploying graph convolution can calibrate contaminated features. To this end,\nwe introduce an innovative plug-and-play module called Feature Map-based\nReconstructed Graph Convolution (FMR-GC). FMR-GC harmonizes feature maps in the\nchannel dimension to reconstruct the graph, then employs graph convolution to\ncapture neighborhood information, effectively calibrating contaminated\nfeatures. Extensive experiments have demonstrated the superior performance and\nscalability of FMR-GC. Moreover, our model can be combined with advanced\nadversarial training methods to considerably enhance robustness without\ncompromising the model's clean accuracy.\n", "date": "2024-06-17", "categories": []}
{"id": "2406.11707", "title": "A First Physical-World Trajectory Prediction Attack via LiDAR-induced\n  Deceptions in Autonomous Driving", "abstract": "  Trajectory prediction forecasts nearby agents' moves based on their\nhistorical trajectories. Accurate trajectory prediction is crucial for\nautonomous vehicles. Existing attacks compromise the prediction model of a\nvictim AV by directly manipulating the historical trajectory of an attacker AV,\nwhich has limited real-world applicability. This paper, for the first time,\nexplores an indirect attack approach that induces prediction errors via attacks\nagainst the perception module of a victim AV. Although it has been shown that\nphysically realizable attacks against LiDAR-based perception are possible by\nplacing a few objects at strategic locations, it is still an open challenge to\nfind an object location from the vast search space in order to launch effective\nattacks against prediction under varying victim AV velocities.\n  Through analysis, we observe that a prediction model is prone to an attack\nfocusing on a single point in the scene. Consequently, we propose a novel\ntwo-stage attack framework to realize the single-point attack. The first stage\nof prediction-side attack efficiently identifies, guided by the distribution of\ndetection results under object-based attacks against perception, the state\nperturbations for the prediction model that are effective and\nvelocity-insensitive. In the second stage of location matching, we match the\nfeasible object locations with the found state perturbations. Our evaluation\nusing a public autonomous driving dataset shows that our attack causes a\ncollision rate of up to 63% and various hazardous responses of the victim AV.\nThe effectiveness of our attack is also demonstrated on a real testbed car. To\nthe best of our knowledge, this study is the first security analysis spanning\nfrom LiDAR-based perception to prediction in autonomous driving, leading to a\nrealistic attack on prediction. To counteract the proposed attack, potential\ndefenses are discussed.\n", "date": "2024-06-17", "categories": []}
{"id": "2406.10161", "title": "On the Computability of Robust PAC Learning", "abstract": "  We initiate the study of computability requirements for adversarially robust\nlearning. Adversarially robust PAC-type learnability is by now an established\nfield of research. However, the effects of computability requirements in\nPAC-type frameworks are only just starting to emerge. We introduce the problem\nof robust computable PAC (robust CPAC) learning and provide some simple\nsufficient conditions for this. We then show that learnability in this setup is\nnot implied by the combination of its components: classes that are both CPAC\nand robustly PAC learnable are not necessarily robustly CPAC learnable.\nFurthermore, we show that the novel framework exhibits some surprising effects:\nfor robust CPAC learnability it is not required that the robust loss is\ncomputably evaluable! Towards understanding characterizing properties, we\nintroduce a novel dimension, the computable robust shattering dimension. We\nprove that its finiteness is necessary, but not sufficient for robust CPAC\nlearnability. This might yield novel insights for the corresponding phenomenon\nin the context of robust PAC learnability, where insufficiency of the robust\nshattering dimension for learnability has been conjectured, but so far a\nresolution has remained elusive.\n", "date": "2024-06-14", "categories": []}
{"id": "2304.02497", "title": "Hyper-parameter Tuning for Adversarially Robust Models", "abstract": "  This work focuses on the problem of hyper-parameter tuning (HPT) for robust\n(i.e., adversarially trained) models, shedding light on the new challenges and\nopportunities arising during the HPT process for robust models. To this end, we\nconduct an extensive experimental study based on 3 popular deep models, in\nwhich we explore exhaustively 9 (discretized) HPs, 2 fidelity dimensions, and 2\nattack bounds, for a total of 19208 configurations (corresponding to 50\nthousand GPU hours). Through this study, we show that the complexity of the HPT\nproblem is further exacerbated in adversarial settings due to the need to\nindependently tune the HPs used during standard and adversarial training:\nsucceeding in doing so (i.e., adopting different HP settings in both phases)\ncan lead to a reduction of up to 80% and 43% of the error for clean and\nadversarial inputs, respectively. On the other hand, we also identify new\nopportunities to reduce the cost of HPT for robust models. Specifically, we\npropose to leverage cheap adversarial training methods to obtain inexpensive,\nyet highly correlated, estimations of the quality achievable using\nstate-of-the-art methods. We show that, by exploiting this novel idea in\nconjunction with a recent multi-fidelity optimizer (taKG), the efficiency of\nthe HPT process can be enhanced by up to 2.1x.\n", "date": "2023-04-05", "categories": []}
{"id": "2310.01959", "title": "Beyond Labeling Oracles: What does it mean to steal ML models?", "abstract": "  Model extraction attacks are designed to steal trained models with only query\naccess, as is often provided through APIs that ML-as-a-Service providers offer.\nMachine Learning (ML) models are expensive to train, in part because data is\nhard to obtain, and a primary incentive for model extraction is to acquire a\nmodel while incurring less cost than training from scratch. Literature on model\nextraction commonly claims or presumes that the attacker is able to save on\nboth data acquisition and labeling costs. We thoroughly evaluate this\nassumption and find that the attacker often does not. This is because current\nattacks implicitly rely on the adversary being able to sample from the victim\nmodel's data distribution. We thoroughly research factors influencing the\nsuccess of model extraction. We discover that prior knowledge of the attacker,\ni.e., access to in-distribution data, dominates other factors like the attack\npolicy the adversary follows to choose which queries to make to the victim\nmodel API. Our findings urge the community to redefine the adversarial goals of\nME attacks as current evaluation methods misinterpret the ME performance.\n", "date": "2023-10-03", "categories": []}
{"id": "2406.08829", "title": "Improving Adversarial Robustness via Feature Pattern Consistency\n  Constraint", "abstract": "  Convolutional Neural Networks (CNNs) are well-known for their vulnerability\nto adversarial attacks, posing significant security concerns. In response to\nthese threats, various defense methods have emerged to bolster the model's\nrobustness. However, most existing methods either focus on learning from\nadversarial perturbations, leading to overfitting to the adversarial examples,\nor aim to eliminate such perturbations during inference, inevitably increasing\ncomputational burdens. Conversely, clean training, which strengthens the\nmodel's robustness by relying solely on clean examples, can address the\naforementioned issues. In this paper, we align with this methodological stream\nand enhance its generalizability to unknown adversarial examples. This\nenhancement is achieved by scrutinizing the behavior of latent features within\nthe network. Recognizing that a correct prediction relies on the correctness of\nthe latent feature's pattern, we introduce a novel and effective Feature\nPattern Consistency Constraint (FPCC) method to reinforce the latent feature's\ncapacity to maintain the correct feature pattern. Specifically, we propose\nSpatial-wise Feature Modification and Channel-wise Feature Selection to enhance\nlatent features. Subsequently, we employ the Pattern Consistency Loss to\nconstrain the similarity between the feature pattern of the latent features and\nthe correct feature pattern. Our experiments demonstrate that the FPCC method\nempowers latent features to uphold correct feature patterns even in the face of\nadversarial examples, resulting in inherent adversarial robustness surpassing\nstate-of-the-art models.\n", "date": "2024-06-13", "categories": []}
{"id": "2302.04379", "title": "Et Tu Certifications: Robustness Certificates Yield Better Adversarial\n  Examples", "abstract": "  In guaranteeing the absence of adversarial examples in an instance's\nneighbourhood, certification mechanisms play an important role in demonstrating\nneural net robustness. In this paper, we ask if these certifications can\ncompromise the very models they help to protect? Our new \\emph{Certification\nAware Attack} exploits certifications to produce computationally efficient\nnorm-minimising adversarial examples $74 \\%$ more often than comparable\nattacks, while reducing the median perturbation norm by more than $10\\%$. While\nthese attacks can be used to assess the tightness of certification bounds, they\nalso highlight that releasing certifications can paradoxically reduce security.\n", "date": "2023-02-09", "categories": []}
{"id": "2406.04998", "title": "ADBA:Approximation Decision Boundary Approach for Black-Box Adversarial\n  Attacks", "abstract": "  Many machine learning models are susceptible to adversarial attacks, with\ndecision-based black-box attacks representing the most critical threat in\nreal-world applications. These attacks are extremely stealthy, generating\nadversarial examples using hard labels obtained from the target machine\nlearning model. This is typically realized by optimizing perturbation\ndirections, guided by decision boundaries identified through query-intensive\nexact search, significantly limiting the attack success rate. This paper\nintroduces a novel approach using the Approximation Decision Boundary (ADB) to\nefficiently and accurately compare perturbation directions without precisely\ndetermining decision boundaries. The effectiveness of our ADB approach (ADBA)\nhinges on promptly identifying suitable ADB, ensuring reliable differentiation\nof all perturbation directions. For this purpose, we analyze the probability\ndistribution of decision boundaries, confirming that using the distribution's\nmedian value as ADB can effectively distinguish different perturbation\ndirections, giving rise to the development of the ADBA-md algorithm. ADBA-md\nonly requires four queries on average to differentiate any pair of perturbation\ndirections, which is highly query-efficient. Extensive experiments on six\nwell-known image classifiers clearly demonstrate the superiority of ADBA and\nADBA-md over multiple state-of-the-art black-box attacks. The source code is\navailable at https://github.com/BUPTAIOC/ADBA.\n", "date": "2024-06-07", "categories": []}
{"id": "2406.07580", "title": "DMS: Addressing Information Loss with More Steps for Pragmatic\n  Adversarial Attacks", "abstract": "  Despite the exceptional performance of deep neural networks (DNNs) across\ndifferent domains, they are vulnerable to adversarial samples, in particular\nfor tasks related to computer vision. Such vulnerability is further influenced\nby the digital container formats used in computers, where the discrete\nnumerical values are commonly used for storing the pixel values. This paper\nexamines how information loss in file formats impacts the effectiveness of\nadversarial attacks. Notably, we observe a pronounced hindrance to the\nadversarial attack performance due to the information loss of the non-integer\npixel values. To address this issue, we explore to leverage the gradient\ninformation of the attack samples within the model to mitigate the information\nloss. We introduce the Do More Steps (DMS) algorithm, which hinges on two core\ntechniques: gradient ascent-based \\textit{adversarial integerization} (DMS-AI)\nand integrated gradients-based \\textit{attribution selection} (DMS-AS). Our\ngoal is to alleviate such lossy process to retain the attack performance when\nstoring these adversarial samples digitally. In particular, DMS-AI integerizes\nthe non-integer pixel values according to the gradient direction, and DMS-AS\nselects the non-integer pixels by comparing attribution results. We conduct\nthorough experiments to assess the effectiveness of our approach, including the\nimplementations of the DMS-AI and DMS-AS on two large-scale datasets with\nvarious latest gradient-based attack methods. Our empirical findings\nconclusively demonstrate the superiority of our proposed DMS-AI and DMS-AS\npixel integerization methods over the standardised methods, such as rounding,\ntruncating and upper approaches, in maintaining attack integrity.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.08050", "title": "Adversarial Evasion Attack Efficiency against Large Language Models", "abstract": "  Large Language Models (LLMs) are valuable for text classification, but their\nvulnerabilities must not be disregarded. They lack robustness against\nadversarial examples, so it is pertinent to understand the impacts of different\ntypes of perturbations, and assess if those attacks could be replicated by\ncommon users with a small amount of perturbations and a small number of queries\nto a deployed LLM. This work presents an analysis of the effectiveness,\nefficiency, and practicality of three different types of adversarial attacks\nagainst five different LLMs in a sentiment classification task. The obtained\nresults demonstrated the very distinct impacts of the word-level and\ncharacter-level attacks. The word attacks were more effective, but the\ncharacter and more constrained attacks were more practical and required a\nreduced number of perturbations and queries. These differences need to be\nconsidered during the development of adversarial defense strategies to train\nmore robust LLMs for intelligent text classification applications.\n", "date": "2024-06-12", "categories": []}
{"id": "2406.08102", "title": "Adversarial Patch for 3D Local Feature Extractor", "abstract": "  Local feature extractors are the cornerstone of many computer vision tasks.\nHowever, their vulnerability to adversarial attacks can significantly\ncompromise their effectiveness. This paper discusses approaches to attack\nsophisticated local feature extraction algorithms and models to achieve two\ndistinct goals: (1) forcing a match between originally non-matching image\nregions, and (2) preventing a match between originally matching regions. At the\nend of the paper, we discuss the performance and drawbacks of different patch\ngeneration methods.\n", "date": "2024-06-12", "categories": []}
{"id": "2406.08331", "title": "Genetic Column Generation for Computing Lower Bounds for Adversarial\n  Classification", "abstract": "  Recent theoretical results on adversarial multi-class classification showed a\nsimilarity to the multi-marginal formulation of Wasserstein-barycenter in\noptimal transport. Unfortunately, both problems suffer from the curse of\ndimension, making it hard to exploit the nice linear program structure of the\nproblems for numerical calculations. We investigate how ideas from Genetic\nColumn Generation for multi-marginal optimal transport can be used to overcome\nthe curse of dimension in computing the minimal adversarial risk in multi-class\nclassification.\n", "date": "2024-06-12", "categories": []}
{"id": "2306.16979", "title": "Post-train Black-box Defense via Bayesian Boundary Correction", "abstract": "  Classifiers based on deep neural networks are susceptible to adversarial\nattack, where the widely existing vulnerability has invoked the research in\ndefending them from potential threats. Given a vulnerable classifier, existing\ndefense methods are mostly white-box and often require re-training the victim\nunder modified loss functions/training regimes. While the model/data/training\nspecifics of the victim are usually unavailable to the user, re-training is\nunappealing, if not impossible for reasons such as limited computational\nresources. To this end, we propose a new post-train black-box defense\nframework. It can turn any pre-trained classifier into a resilient one with\nlittle knowledge of the model specifics. This is achieved by new joint Bayesian\ntreatments on the clean data, the adversarial examples and the classifier, for\nmaximizing their joint probability. It is further equipped with a new\npost-train strategy which keeps the victim intact, avoiding re-training. We\nname our framework Bayesian Boundary Correction (BBC). BBC is a general and\nflexible framework that can easily adapt to different data types. We\ninstantiate BBC for image classification and skeleton-based human activity\nrecognition, for both static and dynamic data. Exhaustive evaluation shows that\nBBC has superior robustness and can enhance robustness without severely hurting\nthe clean accuracy, compared with existing defense methods.\n", "date": "2023-06-29", "categories": []}
{"id": "2307.16630", "title": "Text-CRS: A Generalized Certified Robustness Framework against Textual\n  Adversarial Attacks", "abstract": "  The language models, especially the basic text classification models, have\nbeen shown to be susceptible to textual adversarial attacks such as synonym\nsubstitution and word insertion attacks. To defend against such attacks, a\ngrowing body of research has been devoted to improving the model robustness.\nHowever, providing provable robustness guarantees instead of empirical\nrobustness is still widely unexplored. In this paper, we propose Text-CRS, a\ngeneralized certified robustness framework for natural language processing\n(NLP) based on randomized smoothing. To our best knowledge, existing certified\nschemes for NLP can only certify the robustness against $\\ell_0$ perturbations\nin synonym substitution attacks. Representing each word-level adversarial\noperation (i.e., synonym substitution, word reordering, insertion, and\ndeletion) as a combination of permutation and embedding transformation, we\npropose novel smoothing theorems to derive robustness bounds in both\npermutation and embedding space against such adversarial operations. To further\nimprove certified accuracy and radius, we consider the numerical relationships\nbetween discrete words and select proper noise distributions for the randomized\nsmoothing. Finally, we conduct substantial experiments on multiple language\nmodels and datasets. Text-CRS can address all four different word-level\nadversarial operations and achieve a significant accuracy improvement. We also\nprovide the first benchmark on certified accuracy and radius of four word-level\noperations, besides outperforming the state-of-the-art certification against\nsynonym substitution attacks.\n", "date": "2023-07-31", "categories": []}
{"id": "2311.14772", "title": "Trainwreck: A damaging adversarial attack on image classifiers", "abstract": "  Adversarial attacks are an important security concern for computer vision\n(CV). As CV models are becoming increasingly valuable assets in applied\npractice, disrupting them is emerging as a form of economic sabotage. This\npaper opens up the exploration of damaging adversarial attacks (DAAs) that seek\nto damage target CV models. DAAs are formalized by defining the threat model,\nthe cost function DAAs maximize, and setting three requirements for success:\npotency, stealth, and customizability. As a pioneer DAA, this paper proposes\nTrainwreck, a train-time attack that conflates the data of similar classes in\nthe training data using stealthy ($\\epsilon \\leq 8/255$) class-pair universal\nperturbations obtained from a surrogate model. Trainwreck is a black-box,\ntransferable attack: it requires no knowledge of the target architecture, and a\nsingle poisoned dataset degrades the performance of any model trained on it.\nThe experimental evaluation on CIFAR-10 and CIFAR-100 and various model\narchitectures (EfficientNetV2, ResNeXt-101, and a finetuned ViT-L-16)\ndemonstrates Trainwreck's efficiency. Trainwreck achieves similar or better\npotency compared to the data poisoning state of the art and is fully\ncustomizable by the poison rate parameter. Finally, data redundancy with\nhashing is identified as a reliable defense against Trainwreck or similar DAAs.\nThe code is available at https://github.com/JanZahalka/trainwreck.\n", "date": "2023-11-24", "categories": []}
{"id": "2406.06622", "title": "Adversarial Tuning: Defending Against Jailbreak Attacks for LLMs", "abstract": "  Although safely enhanced Large Language Models (LLMs) have achieved\nremarkable success in tackling various complex tasks in a zero-shot manner,\nthey remain susceptible to jailbreak attacks, particularly the unknown\njailbreak attack. To enhance LLMs' generalized defense capabilities, we propose\na two-stage adversarial tuning framework, which generates adversarial prompts\nto explore worst-case scenarios by optimizing datasets containing pairs of\nadversarial prompts and their safe responses. In the first stage, we introduce\nthe hierarchical meta-universal adversarial prompt learning to efficiently and\neffectively generate token-level adversarial prompts. In the second stage, we\npropose the automatic adversarial prompt learning to iteratively refine\nsemantic-level adversarial prompts, further enhancing LLM's defense\ncapabilities. We conducted comprehensive experiments on three widely used\njailbreak datasets, comparing our framework with six defense baselines under\nfive representative attack scenarios. The results underscore the superiority of\nour proposed methods. Furthermore, our adversarial tuning framework exhibits\nempirical generalizability across various attack strategies and target LLMs,\nhighlighting its potential as a transferable defense mechanism.\n", "date": "2024-06-07", "categories": []}
{"id": "2309.15519", "title": "Defending Against Physical Adversarial Patch Attacks on Infrared Human\n  Detection", "abstract": "  Infrared detection is an emerging technique for safety-critical tasks owing\nto its remarkable anti-interference capability. However, recent studies have\nrevealed that it is vulnerable to physically-realizable adversarial patches,\nposing risks in its real-world applications. To address this problem, we are\nthe first to investigate defense strategies against adversarial patch attacks\non infrared detection, especially human detection. We propose a straightforward\ndefense strategy, patch-based occlusion-aware detection (POD), which\nefficiently augments training samples with random patches and subsequently\ndetects them. POD not only robustly detects people but also identifies\nadversarial patch locations. Surprisingly, while being extremely\ncomputationally efficient, POD easily generalizes to state-of-the-art\nadversarial patch attacks that are unseen during training. Furthermore, POD\nimproves detection precision even in a clean (i.e., no-attack) situation due to\nthe data augmentation effect. Our evaluation demonstrates that POD is robust to\nadversarial patches of various shapes and sizes. The effectiveness of our\nbaseline approach is shown to be a viable defense mechanism for real-world\ninfrared human detection systems, paving the way for exploring future research\ndirections.\n", "date": "2023-09-27", "categories": []}
{"id": "2310.01452", "title": "Fooling the Textual Fooler via Randomizing Latent Representations", "abstract": "  Despite outstanding performance in a variety of NLP tasks, recent studies\nhave revealed that NLP models are vulnerable to adversarial attacks that\nslightly perturb the input to cause the models to misbehave. Among these\nattacks, adversarial word-level perturbations are well-studied and effective\nattack strategies. Since these attacks work in black-box settings, they do not\nrequire access to the model architecture or model parameters and thus can be\ndetrimental to existing NLP applications. To perform an attack, the adversary\nqueries the victim model many times to determine the most important words in an\ninput text and to replace these words with their corresponding synonyms. In\nthis work, we propose a lightweight and attack-agnostic defense whose main goal\nis to perplex the process of generating an adversarial example in these\nquery-based black-box attacks; that is to fool the textual fooler. This\ndefense, named AdvFooler, works by randomizing the latent representation of the\ninput at inference time. Different from existing defenses, AdvFooler does not\nnecessitate additional computational overhead during training nor relies on\nassumptions about the potential adversarial perturbation set while having a\nnegligible impact on the model's accuracy. Our theoretical and empirical\nanalyses highlight the significance of robustness resulting from confusing the\nadversary via randomizing the latent space, as well as the impact of\nrandomization on clean accuracy. Finally, we empirically demonstrate near\nstate-of-the-art robustness of AdvFooler against representative adversarial\nword-level attacks on two benchmark datasets.\n", "date": "2023-10-02", "categories": []}
{"id": "2312.09558", "title": "Towards Transferable Targeted 3D Adversarial Attack in the Physical\n  World", "abstract": "  Compared with transferable untargeted attacks, transferable targeted\nadversarial attacks could specify the misclassification categories of\nadversarial samples, posing a greater threat to security-critical tasks. In the\nmeanwhile, 3D adversarial samples, due to their potential of multi-view\nrobustness, can more comprehensively identify weaknesses in existing deep\nlearning systems, possessing great application value. However, the field of\ntransferable targeted 3D adversarial attacks remains vacant. The goal of this\nwork is to develop a more effective technique that could generate transferable\ntargeted 3D adversarial examples, filling the gap in this field. To achieve\nthis goal, we design a novel framework named TT3D that could rapidly\nreconstruct from few multi-view images into Transferable Targeted 3D textured\nmeshes. While existing mesh-based texture optimization methods compute\ngradients in the high-dimensional mesh space and easily fall into local optima,\nleading to unsatisfactory transferability and distinct distortions, TT3D\ninnovatively performs dual optimization towards both feature grid and\nMulti-layer Perceptron (MLP) parameters in the grid-based NeRF space, which\nsignificantly enhances black-box transferability while enjoying naturalness.\nExperimental results show that TT3D not only exhibits superior cross-model\ntransferability but also maintains considerable adaptability across different\nrenders and vision tasks. More importantly, we produce 3D adversarial examples\nwith 3D printing techniques in the real world and verify their robust\nperformance under various scenarios.\n", "date": "2023-12-15", "categories": []}
{"id": "2402.18104", "title": "Making Them Ask and Answer: Jailbreaking Large Language Models in Few\n  Queries via Disguise and Reconstruction", "abstract": "  In recent years, large language models (LLMs) have demonstrated notable\nsuccess across various tasks, but the trustworthiness of LLMs is still an open\nproblem. One specific threat is the potential to generate toxic or harmful\nresponses. Attackers can craft adversarial prompts that induce harmful\nresponses from LLMs. In this work, we pioneer a theoretical foundation in LLMs\nsecurity by identifying bias vulnerabilities within the safety fine-tuning and\ndesign a black-box jailbreak method named DRA (Disguise and Reconstruction\nAttack), which conceals harmful instructions through disguise and prompts the\nmodel to reconstruct the original harmful instruction within its completion. We\nevaluate DRA across various open-source and closed-source models, showcasing\nstate-of-the-art jailbreak success rates and attack efficiency. Notably, DRA\nboasts a 91.1% attack success rate on OpenAI GPT-4 chatbot.\n", "date": "2024-02-28", "categories": []}
{"id": "2406.05372", "title": "Bridging the Gap: Rademacher Complexity in Robust and Standard\n  Generalization", "abstract": "  Training Deep Neural Networks (DNNs) with adversarial examples often results\nin poor generalization to test-time adversarial data. This paper investigates\nthis issue, known as adversarially robust generalization, through the lens of\nRademacher complexity. Building upon the studies by Khim and Loh (2018); Yin et\nal. (2019), numerous works have been dedicated to this problem, yet achieving a\nsatisfactory bound remains an elusive goal. Existing works on DNNs either apply\nto a surrogate loss instead of the robust loss or yield bounds that are notably\nlooser compared to their standard counterparts. In the latter case, the bounds\nhave a higher dependency on the width $m$ of the DNNs or the dimension $d$ of\nthe data, with an extra factor of at least $\\mathcal{O}(\\sqrt{m})$ or\n$\\mathcal{O}(\\sqrt{d})$.\n  This paper presents upper bounds for adversarial Rademacher complexity of\nDNNs that match the best-known upper bounds in standard settings, as\nestablished in the work of Bartlett et al. (2017), with the dependency on width\nand dimension being $\\mathcal{O}(\\ln(dm))$. The central challenge addressed is\ncalculating the covering number of adversarial function classes. We aim to\nconstruct a new cover that possesses two properties: 1) compatibility with\nadversarial examples, and 2) precision comparable to covers used in standard\nsettings. To this end, we introduce a new variant of covering number called the\n\\emph{uniform covering number}, specifically designed and proven to reconcile\nthese two properties. Consequently, our method effectively bridges the gap\nbetween Rademacher complexity in robust and standard generalization.\n", "date": "2024-06-08", "categories": []}
{"id": "2406.05531", "title": "Enhancing Adversarial Transferability via Information Bottleneck\n  Constraints", "abstract": "  From the perspective of information bottleneck (IB) theory, we propose a\nnovel framework for performing black-box transferable adversarial attacks named\nIBTA, which leverages advancements in invariant features. Intuitively,\ndiminishing the reliance of adversarial perturbations on the original data,\nunder equivalent attack performance constraints, encourages a greater reliance\non invariant features that contributes most to classification, thereby\nenhancing the transferability of adversarial attacks. Building on this\nmotivation, we redefine the optimization of transferable attacks using a novel\ntheoretical framework that centers around IB. Specifically, to overcome the\nchallenge of unoptimizable mutual information, we propose a simple and\nefficient mutual information lower bound (MILB) for approximating computation.\nMoreover, to quantitatively evaluate mutual information, we utilize the Mutual\nInformation Neural Estimator (MINE) to perform a thorough analysis. Our\nexperiments on the ImageNet dataset well demonstrate the efficiency and\nscalability of IBTA and derived MILB. Our code is available at\nhttps://github.com/Biqing-Qi/Enhancing-Adversarial-Transferability-via-Information-Bottleneck-Constraints.\n", "date": "2024-06-08", "categories": []}
{"id": "2406.05535", "title": "Perturbation Towards Easy Samples Improves Targeted Adversarial\n  Transferability", "abstract": "  The transferability of adversarial perturbations provides an effective\nshortcut for black-box attacks. Targeted perturbations have greater\npracticality but are more difficult to transfer between models. In this paper,\nwe experimentally and theoretically demonstrated that neural networks trained\non the same dataset have more consistent performance in\nHigh-Sample-Density-Regions (HSDR) of each class instead of low sample density\nregions. Therefore, in the target setting, adding perturbations towards HSDR of\nthe target class is more effective in improving transferability. However,\ndensity estimation is challenging in high-dimensional scenarios. Further\ntheoretical and experimental verification demonstrates that easy samples with\nlow loss are more likely to be located in HSDR. Perturbations towards such easy\nsamples in the target class can avoid density estimation for HSDR location.\nBased on the above facts, we verified that adding perturbations to easy samples\nin the target class improves targeted adversarial transferability of existing\nattack methods. A generative targeted attack strategy named Easy Sample\nMatching Attack (ESMA) is proposed, which has a higher success rate for\ntargeted attacks and outperforms the SOTA generative method. Moreover, ESMA\nrequires only 5% of the storage space and much less computation time comparing\nto the current SOTA, as ESMA attacks all classes with only one model instead of\nseperate models for each class. Our code is available at\nhttps://github.com/gjq100/ESMA.\n", "date": "2024-06-08", "categories": []}
{"id": "2406.05796", "title": "ProFeAT: Projected Feature Adversarial Training for Self-Supervised\n  Learning of Robust Representations", "abstract": "  The need for abundant labelled data in supervised Adversarial Training (AT)\nhas prompted the use of Self-Supervised Learning (SSL) techniques with AT.\nHowever, the direct application of existing SSL methods to adversarial training\nhas been sub-optimal due to the increased training complexity of combining SSL\nwith AT. A recent approach, DeACL, mitigates this by utilizing supervision from\na standard SSL teacher in a distillation setting, to mimic supervised AT.\nHowever, we find that there is still a large performance gap when compared to\nsupervised adversarial training, specifically on larger models. In this work,\ninvestigate the key reason for this gap and propose Projected Feature\nAdversarial Training (ProFeAT) to bridge the same. We show that the sub-optimal\ndistillation performance is a result of mismatch in training objectives of the\nteacher and student, and propose to use a projection head at the student, that\nallows it to leverage weak supervision from the teacher while also being able\nto learn adversarially robust representations that are distinct from the\nteacher. We further propose appropriate attack and defense losses at the\nfeature and projector, alongside a combination of weak and strong augmentations\nfor the teacher and student respectively, to improve the training data\ndiversity without increasing the training complexity. Through extensive\nexperiments on several benchmark datasets and models, we demonstrate\nsignificant improvements in both clean and robust accuracy when compared to\nexisting SSL-AT methods, setting a new state-of-the-art. We further report\non-par/ improved performance when compared to TRADES, a popular supervised-AT\nmethod.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.05810", "title": "ControlLoc: Physical-World Hijacking Attack on Visual Perception in\n  Autonomous Driving", "abstract": "  Recent research in adversarial machine learning has focused on visual\nperception in Autonomous Driving (AD) and has shown that printed adversarial\npatches can attack object detectors. However, it is important to note that AD\nvisual perception encompasses more than just object detection; it also includes\nMultiple Object Tracking (MOT). MOT enhances the robustness by compensating for\nobject detection errors and requiring consistent object detection results\nacross multiple frames before influencing tracking results and driving\ndecisions. Thus, MOT makes attacks on object detection alone less effective. To\nattack such robust AD visual perception, a digital hijacking attack has been\nproposed to cause dangerous driving scenarios. However, this attack has limited\neffectiveness.\n  In this paper, we introduce a novel physical-world adversarial patch attack,\nControlLoc, designed to exploit hijacking vulnerabilities in entire AD visual\nperception. ControlLoc utilizes a two-stage process: initially identifying the\noptimal location for the adversarial patch, and subsequently generating the\npatch that can modify the perceived location and shape of objects with the\noptimal location. Extensive evaluations demonstrate the superior performance of\nControlLoc, achieving an impressive average attack success rate of around 98.1%\nacross various AD visual perceptions and datasets, which is four times greater\neffectiveness than the existing hijacking attack. The effectiveness of\nControlLoc is further validated in physical-world conditions, including real\nvehicle tests under different conditions such as outdoor light conditions with\nan average attack success rate of 77.5%. AD system-level impact assessments are\nalso included, such as vehicle collision, using industry-grade AD systems and\nproduction-grade AD simulators with an average vehicle collision rate and\nunnecessary emergency stop rate of 81.3%.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.06089", "title": "Texture Re-scalable Universal Adversarial Perturbation", "abstract": "  Universal adversarial perturbation (UAP), also known as image-agnostic\nperturbation, is a fixed perturbation map that can fool the classifier with\nhigh probabilities on arbitrary images, making it more practical for attacking\ndeep models in the real world. Previous UAP methods generate a scale-fixed and\ntexture-fixed perturbation map for all images, which ignores the multi-scale\nobjects in images and usually results in a low fooling ratio. Since the widely\nused convolution neural networks tend to classify objects according to semantic\ninformation stored in local textures, it seems a reasonable and intuitive way\nto improve the UAP from the perspective of utilizing local contents\neffectively. In this work, we find that the fooling ratios significantly\nincrease when we add a constraint to encourage a small-scale UAP map and repeat\nit vertically and horizontally to fill the whole image domain. To this end, we\npropose texture scale-constrained UAP (TSC-UAP), a simple yet effective UAP\nenhancement method that automatically generates UAPs with category-specific\nlocal textures that can fool deep models more easily. Through a low-cost\noperation that restricts the texture scale, TSC-UAP achieves a considerable\nimprovement in the fooling ratio and attack transferability for both\ndata-dependent and data-free UAP methods. Experiments conducted on two\nstate-of-the-art UAP methods, eight popular CNN models and four classical\ndatasets show the remarkable performance of TSC-UAP.\n", "date": "2024-06-10", "categories": []}
{"id": "2402.11753", "title": "ArtPrompt: ASCII Art-based Jailbreak Attacks against Aligned LLMs", "abstract": "  Safety is critical to the usage of large language models (LLMs). Multiple\ntechniques such as data filtering and supervised fine-tuning have been\ndeveloped to strengthen LLM safety. However, currently known techniques presume\nthat corpora used for safety alignment of LLMs are solely interpreted by\nsemantics. This assumption, however, does not hold in real-world applications,\nwhich leads to severe vulnerabilities in LLMs. For example, users of forums\noften use ASCII art, a form of text-based art, to convey image information. In\nthis paper, we propose a novel ASCII art-based jailbreak attack and introduce a\ncomprehensive benchmark Vision-in-Text Challenge (ViTC) to evaluate the\ncapabilities of LLMs in recognizing prompts that cannot be solely interpreted\nby semantics. We show that five SOTA LLMs (GPT-3.5, GPT-4, Gemini, Claude, and\nLlama2) struggle to recognize prompts provided in the form of ASCII art. Based\non this observation, we develop the jailbreak attack ArtPrompt, which leverages\nthe poor performance of LLMs in recognizing ASCII art to bypass safety measures\nand elicit undesired behaviors from LLMs. ArtPrompt only requires black-box\naccess to the victim LLMs, making it a practical attack. We evaluate ArtPrompt\non five SOTA LLMs, and show that ArtPrompt can effectively and efficiently\ninduce undesired behaviors from all five LLMs. Our code is available at\nhttps://github.com/uw-nsl/ArtPrompt.\n", "date": "2024-02-19", "categories": []}
{"id": "2403.13502", "title": "Adversarial Attacks and Defenses in Fault Detection and Diagnosis: A\n  Comprehensive Benchmark on the Tennessee Eastman Process", "abstract": "  Integrating machine learning into Automated Control Systems (ACS) enhances\ndecision-making in industrial process management. One of the limitations to the\nwidespread adoption of these technologies in industry is the vulnerability of\nneural networks to adversarial attacks. This study explores the threats in\ndeploying deep learning models for fault diagnosis in ACS using the Tennessee\nEastman Process dataset. By evaluating three neural networks with different\narchitectures, we subject them to six types of adversarial attacks and explore\nfive different defense methods. Our results highlight the strong vulnerability\nof models to adversarial samples and the varying effectiveness of defense\nstrategies. We also propose a novel protection approach by combining multiple\ndefense methods and demonstrate it's efficacy. This research contributes\nseveral insights into securing machine learning within ACS, ensuring robust\nfault diagnosis in industrial processes.\n", "date": "2024-03-20", "categories": []}
{"id": "2406.04981", "title": "The Price of Implicit Bias in Adversarially Robust Generalization", "abstract": "  We study the implicit bias of optimization in robust empirical risk\nminimization (robust ERM) and its connection with robust generalization. In\nclassification settings under adversarial perturbations with linear models, we\nstudy what type of regularization should ideally be applied for a given\nperturbation set to improve (robust) generalization. We then show that the\nimplicit bias of optimization in robust ERM can significantly affect the\nrobustness of the model and identify two ways this can happen; either through\nthe optimization algorithm or the architecture. We verify our predictions in\nsimulations with synthetic data and experimentally study the importance of\nimplicit bias in robust ERM with deep neural networks.\n", "date": "2024-06-07", "categories": []}
{"id": "2406.05119", "title": "Compositional Curvature Bounds for Deep Neural Networks", "abstract": "  A key challenge that threatens the widespread use of neural networks in\nsafety-critical applications is their vulnerability to adversarial attacks. In\nthis paper, we study the second-order behavior of continuously differentiable\ndeep neural networks, focusing on robustness against adversarial perturbations.\nFirst, we provide a theoretical analysis of robustness and attack certificates\nfor deep classifiers by leveraging local gradients and upper bounds on the\nsecond derivative (curvature constant). Next, we introduce a novel algorithm to\nanalytically compute provable upper bounds on the second derivative of neural\nnetworks. This algorithm leverages the compositional structure of the model to\npropagate the curvature bound layer-by-layer, giving rise to a scalable and\nmodular approach. The proposed bound can serve as a differentiable regularizer\nto control the curvature of neural networks during training, thereby enhancing\nrobustness. Finally, we demonstrate the efficacy of our method on\nclassification tasks using the MNIST and CIFAR-10 datasets.\n", "date": "2024-06-07", "categories": []}
{"id": "2306.06209", "title": "Backdoor Attack with Sparse and Invisible Trigger", "abstract": "  Deep neural networks (DNNs) are vulnerable to backdoor attacks, where the\nadversary manipulates a small portion of training data such that the victim\nmodel predicts normally on the benign samples but classifies the triggered\nsamples as the target class. The backdoor attack is an emerging yet threatening\ntraining-phase threat, leading to serious risks in DNN-based applications. In\nthis paper, we revisit the trigger patterns of existing backdoor attacks. We\nreveal that they are either visible or not sparse and therefore are not\nstealthy enough. More importantly, it is not feasible to simply combine\nexisting methods to design an effective sparse and invisible backdoor attack.\nTo address this problem, we formulate the trigger generation as a bi-level\noptimization problem with sparsity and invisibility constraints and propose an\neffective method to solve it. The proposed method is dubbed sparse and\ninvisible backdoor attack (SIBA). We conduct extensive experiments on benchmark\ndatasets under different settings, which verify the effectiveness of our attack\nand its resistance to existing backdoor defenses. The codes for reproducing\nmain experiments are available at \\url{https://github.com/YinghuaGao/SIBA}.\n", "date": "2023-05-11", "categories": []}
{"id": "2310.05141", "title": "Transferable Availability Poisoning Attacks", "abstract": "  We consider availability data poisoning attacks, where an adversary aims to\ndegrade the overall test accuracy of a machine learning model by crafting small\nperturbations to its training data. Existing poisoning strategies can achieve\nthe attack goal but assume the victim to employ the same learning method as\nwhat the adversary uses to mount the attack. In this paper, we argue that this\nassumption is strong, since the victim may choose any learning algorithm to\ntrain the model as long as it can achieve some targeted performance on clean\ndata. Empirically, we observe a large decrease in the effectiveness of prior\npoisoning attacks if the victim employs an alternative learning algorithm. To\nenhance the attack transferability, we propose Transferable Poisoning, which\nfirst leverages the intrinsic characteristics of alignment and uniformity to\nenable better unlearnability within contrastive learning, and then iteratively\nutilizes the gradient information from supervised and unsupervised contrastive\nlearning paradigms to generate the poisoning perturbations. Through extensive\nexperiments on image benchmarks, we show that our transferable poisoning attack\ncan produce poisoned samples with significantly improved transferability, not\nonly applicable to the two learners used to devise the attack but also to\nlearning algorithms and even paradigms beyond.\n", "date": "2023-10-08", "categories": []}
{"id": "2405.16488", "title": "Partial train and isolate, mitigate backdoor attack", "abstract": "  Neural networks are widely known to be vulnerable to backdoor attacks, a\nmethod that poisons a portion of the training data to make the target model\nperform well on normal data sets, while outputting attacker-specified or random\ncategories on the poisoned samples. Backdoor attacks are full of threats.\nPoisoned samples are becoming more and more similar to corresponding normal\nsamples, and even the human eye cannot easily distinguish them. On the other\nhand, the accuracy of models carrying backdoors on normal samples is no\ndifferent from that of clean models.In this article, by observing the\ncharacteristics of backdoor attacks, We provide a new model training method\n(PT) that freezes part of the model to train a model that can isolate\nsuspicious samples. Then, on this basis, a clean model is fine-tuned to resist\nbackdoor attacks.\n", "date": "2024-05-26", "categories": []}
{"id": "2406.04070", "title": "Batch-in-Batch: a new adversarial training framework for initial\n  perturbation and sample selection", "abstract": "  Adversarial training methods commonly generate independent initial\nperturbation for adversarial samples from a simple uniform distribution, and\nobtain the training batch for the classifier without selection. In this work,\nwe propose a simple yet effective training framework called Batch-in-Batch (BB)\nto enhance models robustness. It involves specifically a joint construction of\ninitial values that could simultaneously generates $m$ sets of perturbations\nfrom the original batch set to provide more diversity for adversarial samples;\nand also includes various sample selection strategies that enable the trained\nmodels to have smoother losses and avoid overconfident outputs. Through\nextensive experiments on three benchmark datasets (CIFAR-10, SVHN, CIFAR-100)\nwith two networks (PreActResNet18 and WideResNet28-10) that are used in both\nthe single-step (Noise-Fast Gradient Sign Method, N-FGSM) and multi-step\n(Projected Gradient Descent, PGD-10) adversarial training, we show that models\ntrained within the BB framework consistently have higher adversarial accuracy\nacross various adversarial settings, notably achieving over a 13% improvement\non the SVHN dataset with an attack radius of 8/255 compared to the N-FGSM\nbaseline model. Furthermore, experimental analysis of the efficiency of both\nthe proposed initial perturbation method and sample selection strategies\nvalidates our insights. Finally, we show that our framework is cost-effective\nin terms of computational resources, even with a relatively large value of $m$.\n", "date": "2024-06-06", "categories": []}
{"id": "2306.03430", "title": "Revisiting the Trade-off between Accuracy and Robustness via Weight\n  Distribution of Filters", "abstract": "  Adversarial attacks have been proven to be potential threats to Deep Neural\nNetworks (DNNs), and many methods are proposed to defend against adversarial\nattacks. However, while enhancing the robustness, the clean accuracy will\ndecline to a certain extent, implying a trade-off existed between the accuracy\nand robustness. In this paper, to meet the trade-off problem, we theoretically\nexplore the underlying reason for the difference of the filters' weight\ndistribution between standard-trained and robust-trained models and then argue\nthat this is an intrinsic property for static neural networks, thus they are\ndifficult to fundamentally improve the accuracy and adversarial robustness at\nthe same time. Based on this analysis, we propose a sample-wise dynamic network\narchitecture named Adversarial Weight-Varied Network (AW-Net), which focuses on\ndealing with clean and adversarial examples with a \"divide and rule\" weight\nstrategy. The AW-Net adaptively adjusts the network's weights based on\nregulation signals generated by an adversarial router, which is directly\ninfluenced by the input sample. Benefiting from the dynamic network\narchitecture, clean and adversarial examples can be processed with different\nnetwork weights, which provides the potential to enhance both accuracy and\nadversarial robustness. A series of experiments demonstrate that our AW-Net is\narchitecture-friendly to handle both clean and adversarial examples and can\nachieve better trade-off performance than state-of-the-art robust models.\n", "date": "2023-06-06", "categories": []}
{"id": "2402.12336", "title": "Robust CLIP: Unsupervised Adversarial Fine-Tuning of Vision Embeddings\n  for Robust Large Vision-Language Models", "abstract": "  Multi-modal foundation models like OpenFlamingo, LLaVA, and GPT-4 are\nincreasingly used for various real-world tasks. Prior work has shown that these\nmodels are highly vulnerable to adversarial attacks on the vision modality.\nThese attacks can be leveraged to spread fake information or defraud users, and\nthus pose a significant risk, which makes the robustness of large multi-modal\nfoundation models a pressing problem. The CLIP model, or one of its variants,\nis used as a frozen vision encoder in many large vision-language models\n(LVLMs), e.g. LLaVA and OpenFlamingo. We propose an unsupervised adversarial\nfine-tuning scheme to obtain a robust CLIP vision encoder, which yields\nrobustness on all vision down-stream tasks (LVLMs, zero-shot classification)\nthat rely on CLIP. In particular, we show that stealth-attacks on users of\nLVLMs by a malicious third party providing manipulated images are no longer\npossible once one replaces the original CLIP model with our robust one. No\nretraining or fine-tuning of the down-stream LVLMs is required. The code and\nrobust models are available at https://github.com/chs20/RobustVLM\n", "date": "2024-02-19", "categories": []}
{"id": "2402.14836", "title": "Stealthy Attack on Large Language Model based Recommendation", "abstract": "  Recently, the powerful large language models (LLMs) have been instrumental in\npropelling the progress of recommender systems (RS). However, while these\nsystems have flourished, their susceptibility to security threats has been\nlargely overlooked. In this work, we reveal that the introduction of LLMs into\nrecommendation models presents new security vulnerabilities due to their\nemphasis on the textual content of items. We demonstrate that attackers can\nsignificantly boost an item's exposure by merely altering its textual content\nduring the testing phase, without requiring direct interference with the\nmodel's training process. Additionally, the attack is notably stealthy, as it\ndoes not affect the overall recommendation performance and the modifications to\nthe text are subtle, making it difficult for users and platforms to detect. Our\ncomprehensive experiments across four mainstream LLM-based recommendation\nmodels demonstrate the superior efficacy and stealthiness of our approach. Our\nwork unveils a significant security gap in LLM-based recommendation systems and\npaves the way for future research on protecting these systems.\n", "date": "2024-02-18", "categories": []}
{"id": "2402.15152", "title": "On the Duality Between Sharpness-Aware Minimization and Adversarial\n  Training", "abstract": "  Adversarial Training (AT), which adversarially perturb the input samples\nduring training, has been acknowledged as one of the most effective defenses\nagainst adversarial attacks, yet suffers from inevitably decreased clean\naccuracy. Instead of perturbing the samples, Sharpness-Aware Minimization (SAM)\nperturbs the model weights during training to find a more flat loss landscape\nand improve generalization. However, as SAM is designed for better clean\naccuracy, its effectiveness in enhancing adversarial robustness remains\nunexplored. In this work, considering the duality between SAM and AT, we\ninvestigate the adversarial robustness derived from SAM. Intriguingly, we find\nthat using SAM alone can improve adversarial robustness. To understand this\nunexpected property of SAM, we first provide empirical and theoretical insights\ninto how SAM can implicitly learn more robust features, and conduct\ncomprehensive experiments to show that SAM can improve adversarial robustness\nnotably without sacrificing any clean accuracy, shedding light on the potential\nof SAM to be a substitute for AT when accuracy comes at a higher priority. Code\nis available at https://github.com/weizeming/SAM_AT.\n", "date": "2024-02-23", "categories": []}
{"id": "2405.15942", "title": "Can Implicit Bias Imply Adversarial Robustness?", "abstract": "  The implicit bias of gradient-based training algorithms has been considered\nmostly beneficial as it leads to trained networks that often generalize well.\nHowever, Frei et al. (2023) show that such implicit bias can harm adversarial\nrobustness. Specifically, they show that if the data consists of clusters with\nsmall inter-cluster correlation, a shallow (two-layer) ReLU network trained by\ngradient flow generalizes well, but it is not robust to adversarial attacks of\nsmall radius. Moreover, this phenomenon occurs despite the existence of a much\nmore robust classifier that can be explicitly constructed from a shallow\nnetwork. In this paper, we extend recent analyses of neuron alignment to show\nthat a shallow network with a polynomial ReLU activation (pReLU) trained by\ngradient flow not only generalizes well but is also robust to adversarial\nattacks. Our results highlight the importance of the interplay between data\nstructure and architecture design in the implicit bias and robustness of\ntrained networks.\n", "date": "2024-05-24", "categories": []}
{"id": "2406.01873", "title": "CR-UTP: Certified Robustness against Universal Text Perturbations on\n  Large Language Models", "abstract": "  It is imperative to ensure the stability of every prediction made by a\nlanguage model; that is, a language's prediction should remain consistent\ndespite minor input variations, like word substitutions. In this paper, we\ninvestigate the problem of certifying a language model's robustness against\nUniversal Text Perturbations (UTPs), which have been widely used in universal\nadversarial attacks and backdoor attacks. Existing certified robustness based\non random smoothing has shown considerable promise in certifying the\ninput-specific text perturbations (ISTPs), operating under the assumption that\nany random alteration of a sample's clean or adversarial words would negate the\nimpact of sample-wise perturbations. However, with UTPs, masking only the\nadversarial words can eliminate the attack. A naive method is to simply\nincrease the masking ratio and the likelihood of masking attack tokens, but it\nleads to a significant reduction in both certified accuracy and the certified\nradius due to input corruption by extensive masking. To solve this challenge,\nwe introduce a novel approach, the superior prompt search method, designed to\nidentify a superior prompt that maintains higher certified accuracy under\nextensive masking. Additionally, we theoretically motivate why ensembles are a\nparticularly suitable choice as base prompts for random smoothing. The method\nis denoted by superior prompt ensembling technique. We also empirically confirm\nthis technique, obtaining state-of-the-art results in multiple settings. These\nmethodologies, for the first time, enable high certified accuracy against both\nUTPs and ISTPs. The source code of CR-UTP is available at \\url\n{https://github.com/UCFML-Research/CR-UTP}.\n", "date": "2024-06-04", "categories": []}
{"id": "2406.02309", "title": "Effects of Exponential Gaussian Distribution on (Double Sampling)\n  Randomized Smoothing", "abstract": "  Randomized Smoothing (RS) is currently a scalable certified defense method\nproviding robustness certification against adversarial examples. Although\nsignificant progress has been achieved in providing defenses against $\\ell_p$\nadversaries, the interaction between the smoothing distribution and the\nrobustness certification still remains vague. In this work, we comprehensively\nstudy the effect of two families of distributions, named Exponential Standard\nGaussian (ESG) and Exponential General Gaussian (EGG) distributions, on\nRandomized Smoothing and Double Sampling Randomized Smoothing (DSRS). We derive\nan analytic formula for ESG's certified radius, which converges to the origin\nformula of RS as the dimension $d$ increases. Additionally, we prove that EGG\ncan provide tighter constant factors than DSRS in providing $\\Omega(\\sqrt{d})$\nlower bounds of $\\ell_2$ certified radius, and thus further addresses the curse\nof dimensionality in RS. Our experiments on real-world datasets confirm our\ntheoretical analysis of the ESG distributions, that they provide almost the\nsame certification under different exponents $\\eta$ for both RS and DSRS. In\naddition, EGG brings a significant improvement to the DSRS certification, but\nthe mechanism can be different when the classifier properties are different.\nCompared to the primitive DSRS, the increase in certified accuracy provided by\nEGG is prominent, up to 6.4% on ImageNet.\n", "date": "2024-06-04", "categories": []}
{"id": "2406.03117", "title": "VQUNet: Vector Quantization U-Net for Defending Adversarial Atacks by\n  Regularizing Unwanted Noise", "abstract": "  Deep Neural Networks (DNN) have become a promising paradigm when developing\nArtificial Intelligence (AI) and Machine Learning (ML) applications. However,\nDNN applications are vulnerable to fake data that are crafted with adversarial\nattack algorithms. Under adversarial attacks, the prediction accuracy of DNN\napplications suffers, making them unreliable. In order to defend against\nadversarial attacks, we introduce a novel noise-reduction procedure, Vector\nQuantization U-Net (VQUNet), to reduce adversarial noise and reconstruct data\nwith high fidelity. VQUNet features a discrete latent representation learning\nthrough a multi-scale hierarchical structure for both noise reduction and data\nreconstruction. The empirical experiments show that the proposed VQUNet\nprovides better robustness to the target DNN models, and it outperforms other\nstate-of-the-art noise-reduction-based defense methods under various\nadversarial attacks for both Fashion-MNIST and CIFAR10 datasets. When there is\nno adversarial attack, the defense method has less than 1% accuracy degradation\nfor both datasets.\n", "date": "2024-06-05", "categories": []}
{"id": "2304.11300", "title": "MAWSEO: Adversarial Wiki Search Poisoning for Illicit Online Promotion", "abstract": "  As a prominent instance of vandalism edits, Wiki search poisoning for illicit\npromotion is a cybercrime in which the adversary aims at editing Wiki articles\nto promote illicit businesses through Wiki search results of relevant queries.\nIn this paper, we report a study that, for the first time, shows that such\nstealthy blackhat SEO on Wiki can be automated. Our technique, called MAWSEO,\nemploys adversarial revisions to achieve real-world cybercriminal objectives,\nincluding rank boosting, vandalism detection evasion, topic relevancy, semantic\nconsistency, user awareness (but not alarming) of promotional content, etc. Our\nevaluation and user study demonstrate that MAWSEO is capable of effectively and\nefficiently generating adversarial vandalism edits, which can bypass\nstate-of-the-art built-in Wiki vandalism detectors, and also get promotional\ncontent through to Wiki users without triggering their alarms. In addition, we\ninvestigated potential defense, including coherence based detection and\nadversarial training of vandalism detection, against our attack in the Wiki\necosystem.\n", "date": "2023-04-22", "categories": []}
{"id": "2310.05308", "title": "Adversarial Attacks on Combinatorial Multi-Armed Bandits", "abstract": "  We study reward poisoning attacks on Combinatorial Multi-armed Bandits\n(CMAB). We first provide a sufficient and necessary condition for the\nattackability of CMAB, a notion to capture the vulnerability and robustness of\nCMAB. The attackability condition depends on the intrinsic properties of the\ncorresponding CMAB instance such as the reward distributions of super arms and\noutcome distributions of base arms. Additionally, we devise an attack algorithm\nfor attackable CMAB instances. Contrary to prior understanding of multi-armed\nbandits, our work reveals a surprising fact that the attackability of a\nspecific CMAB instance also depends on whether the bandit instance is known or\nunknown to the adversary. This finding indicates that adversarial attacks on\nCMAB are difficult in practice and a general attack strategy for any CMAB\ninstance does not exist since the environment is mostly unknown to the\nadversary. We validate our theoretical findings via extensive experiments on\nreal-world CMAB applications including probabilistic maximum covering problem,\nonline minimum spanning tree, cascading bandits for online ranking, and online\nshortest path.\n", "date": "2023-10-08", "categories": []}
{"id": "2310.12793", "title": "OODRobustBench: a Benchmark and Large-Scale Analysis of Adversarial\n  Robustness under Distribution Shift", "abstract": "  Existing works have made great progress in improving adversarial robustness,\nbut typically test their method only on data from the same distribution as the\ntraining data, i.e. in-distribution (ID) testing. As a result, it is unclear\nhow such robustness generalizes under input distribution shifts, i.e.\nout-of-distribution (OOD) testing. This omission is concerning as such\ndistribution shifts are unavoidable when methods are deployed in the wild. To\naddress this issue we propose a benchmark named OODRobustBench to\ncomprehensively assess OOD adversarial robustness using 23 dataset-wise shifts\n(i.e. naturalistic shifts in input distribution) and 6 threat-wise shifts\n(i.e., unforeseen adversarial threat models). OODRobustBench is used to assess\n706 robust models using 60.7K adversarial evaluations. This large-scale\nanalysis shows that: 1) adversarial robustness suffers from a severe OOD\ngeneralization issue; 2) ID robustness correlates strongly with OOD robustness\nin a positive linear way. The latter enables the prediction of OOD robustness\nfrom ID robustness. We then predict and verify that existing methods are\nunlikely to achieve high OOD robustness. Novel methods are therefore required\nto achieve OOD robustness beyond our prediction. To facilitate the development\nof these methods, we investigate a wide range of techniques and identify\nseveral promising directions. Code and models are available at:\nhttps://github.com/OODRobustBench/OODRobustBench.\n", "date": "2023-10-19", "categories": []}
{"id": "2311.15165", "title": "Mixing Classifiers to Alleviate the Accuracy-Robustness Trade-Off", "abstract": "  Deep neural classifiers have recently found tremendous success in data-driven\ncontrol systems. However, existing models suffer from a trade-off between\naccuracy and adversarial robustness. This limitation must be overcome in the\ncontrol of safety-critical systems that require both high performance and\nrigorous robustness guarantees. In this work, we develop classifiers that\nsimultaneously inherit high robustness from robust models and high accuracy\nfrom standard models. Specifically, we propose a theoretically motivated\nformulation that mixes the output probabilities of a standard neural network\nand a robust neural network. Both base classifiers are pre-trained, and thus\nour method does not require additional training. Our numerical experiments\nverify that the mixed classifier noticeably improves the accuracy-robustness\ntrade-off and identify the confidence property of the robust base classifier as\nthe key leverage of this more benign trade-off. Our theoretical results prove\nthat under mild assumptions, when the robustness of the robust base model is\ncertifiable, no alteration or attack within a closed-form $\\ell_p$ radius on an\ninput can result in the misclassification of the mixed classifier.\n", "date": "2023-11-26", "categories": []}
{"id": "2402.08400", "title": "Adaptive Hierarchical Certification for Segmentation using Randomized\n  Smoothing", "abstract": "  Certification for machine learning is proving that no adversarial sample can\nevade a model within a range under certain conditions, a necessity for\nsafety-critical domains. Common certification methods for segmentation use a\nflat set of fine-grained classes, leading to high abstain rates due to model\nuncertainty across many classes. We propose a novel, more practical setting,\nwhich certifies pixels within a multi-level hierarchy, and adaptively relaxes\nthe certification to a coarser level for unstable components classic methods\nwould abstain from, effectively lowering the abstain rate whilst providing more\ncertified semantically meaningful information. We mathematically formulate the\nproblem setup, introduce an adaptive hierarchical certification algorithm and\nprove the correctness of its guarantees. Since certified accuracy does not take\nthe loss of information into account for coarser classes, we introduce the\nCertified Information Gain ($\\mathrm{CIG}$) metric, which is proportional to\nthe class granularity level. Our extensive experiments on the datasets\nCityscapes, PASCAL-Context, ACDC and COCO-Stuff demonstrate that our adaptive\nalgorithm achieves a higher $\\mathrm{CIG}$ and lower abstain rate compared to\nthe current state-of-the-art certification method. Our code can be found here:\nhttps://github.com/AlaaAnani/adaptive-certify.\n", "date": "2024-02-13", "categories": []}
{"id": "2402.16006", "title": "ASETF: A Novel Method for Jailbreak Attack on LLMs through Translate\n  Suffix Embeddings", "abstract": "  The safety defense methods of Large language models(LLMs) stays limited\nbecause the dangerous prompts are manually curated to just few known attack\ntypes, which fails to keep pace with emerging varieties. Recent studies found\nthat attaching suffixes to harmful instructions can hack the defense of LLMs\nand lead to dangerous outputs. However, similar to traditional text adversarial\nattacks, this approach, while effective, is limited by the challenge of the\ndiscrete tokens. This gradient based discrete optimization attack requires over\n100,000 LLM calls, and due to the unreadable of adversarial suffixes, it can be\nrelatively easily penetrated by common defense methods such as perplexity\nfilters. To cope with this challenge, in this paper, we proposes an Adversarial\nSuffix Embedding Translation Framework (ASETF), aimed at transforming\ncontinuous adversarial suffix embeddings into coherent and understandable text.\nThis method greatly reduces the computational overhead during the attack\nprocess and helps to automatically generate multiple adversarial samples, which\ncan be used as data to strengthen LLMs security defense. Experimental\nevaluations were conducted on Llama2, Vicuna, and other prominent LLMs,\nemploying harmful directives sourced from the Advbench dataset. The results\nindicate that our method significantly reduces the computation time of\nadversarial suffixes and achieves a much better attack success rate to existing\ntechniques, while significantly enhancing the textual fluency of the prompts.\nIn addition, our approach can be generalized into a broader method for\ngenerating transferable adversarial suffixes that can successfully attack\nmultiple LLMs, even black-box LLMs, such as ChatGPT and Gemini.\n", "date": "2024-02-25", "categories": []}
{"id": "2405.09550", "title": "Mask-based Invisible Backdoor Attacks on Object Detection", "abstract": "  Deep learning models have achieved unprecedented performance in the domain of\nobject detection, resulting in breakthroughs in areas such as autonomous\ndriving and security. However, deep learning models are vulnerable to backdoor\nattacks. These attacks prompt models to behave similarly to standard models\nwithout a trigger; however, they act maliciously upon detecting a predefined\ntrigger. Despite extensive research on backdoor attacks in image\nclassification, their application to object detection remains relatively\nunderexplored. Given the widespread application of object detection in critical\nreal-world scenarios, the sensitivity and potential impact of these\nvulnerabilities cannot be overstated. In this study, we propose an effective\ninvisible backdoor attack on object detection utilizing a mask-based approach.\nThree distinct attack scenarios were explored for object detection: object\ndisappearance, object misclassification, and object generation attack. Through\nextensive experiments, we comprehensively examined the effectiveness of these\nattacks and tested certain defense methods to determine effective\ncountermeasures. Code will be available at\nhttps://github.com/jeongjin0/invisible-backdoor-object-detection\n", "date": "2024-03-20", "categories": []}
{"id": "2406.00775", "title": "Constrained Adaptive Attack: Effective Adversarial Attack Against Deep\n  Neural Networks for Tabular Data", "abstract": "  State-of-the-art deep learning models for tabular data have recently achieved\nacceptable performance to be deployed in industrial settings. However, the\nrobustness of these models remains scarcely explored. Contrary to computer\nvision, there are no effective attacks to properly evaluate the adversarial\nrobustness of deep tabular models due to intrinsic properties of tabular data,\nsuch as categorical features, immutability, and feature relationship\nconstraints. To fill this gap, we first propose CAPGD, a gradient attack that\novercomes the failures of existing gradient attacks with adaptive mechanisms.\nThis new attack does not require parameter tuning and further degrades the\naccuracy, up to 81% points compared to the previous gradient attacks. Second,\nwe design CAA, an efficient evasion attack that combines our CAPGD attack and\nMOEVA, the best search-based attack. We demonstrate the effectiveness of our\nattacks on five architectures and four critical use cases. Our empirical study\ndemonstrates that CAA outperforms all existing attacks in 17 over the 20\nsettings, and leads to a drop in the accuracy by up to 96.1% points and 21.9%\npoints compared to CAPGD and MOEVA respectively while being up to five times\nfaster than MOEVA. Given the effectiveness and efficiency of our new attacks,\nwe argue that they should become the minimal test for any new defense or robust\narchitectures in tabular machine learning.\n", "date": "2024-06-02", "categories": []}
{"id": "2406.01765", "title": "Reproducibility Study on Adversarial Attacks Against Robust Transformer\n  Trackers", "abstract": "  New transformer networks have been integrated into object tracking pipelines\nand have demonstrated strong performance on the latest benchmarks. This paper\nfocuses on understanding how transformer trackers behave under adversarial\nattacks and how different attacks perform on tracking datasets as their\nparameters change. We conducted a series of experiments to evaluate the\neffectiveness of existing adversarial attacks on object trackers with\ntransformer and non-transformer backbones. We experimented on 7 different\ntrackers, including 3 that are transformer-based, and 4 which leverage other\narchitectures. These trackers are tested against 4 recent attack methods to\nassess their performance and robustness on VOT2022ST, UAV123 and GOT10k\ndatasets. Our empirical study focuses on evaluating adversarial robustness of\nobject trackers based on bounding box versus binary mask predictions, and\nattack methods at different levels of perturbations. Interestingly, our study\nfound that altering the perturbation level may not significantly affect the\noverall object tracking results after the attack. Similarly, the sparsity and\nimperceptibility of the attack perturbations may remain stable against\nperturbation level shifts. By applying a specific attack on all transformer\ntrackers, we show that new transformer trackers having a stronger\ncross-attention modeling achieve a greater adversarial robustness on tracking\ndatasets, such as VOT2022ST and GOT10k. Our results also indicate the necessity\nfor new attack methods to effectively tackle the latest types of transformer\ntrackers. The codes necessary to reproduce this study are available at\nhttps://github.com/fatemehN/ReproducibilityStudy.\n", "date": "2024-06-03", "categories": []}
{"id": "2406.01894", "title": "SVASTIN: Sparse Video Adversarial Attack via Spatio-Temporal Invertible\n  Neural Networks", "abstract": "  Robust and imperceptible adversarial video attack is challenging due to the\nspatial and temporal characteristics of videos. The existing video adversarial\nattack methods mainly take a gradient-based approach and generate adversarial\nvideos with noticeable perturbations. In this paper, we propose a novel Sparse\nAdversarial Video Attack via Spatio-Temporal Invertible Neural Networks\n(SVASTIN) to generate adversarial videos through spatio-temporal feature space\ninformation exchanging. It consists of a Guided Target Video Learning (GTVL)\nmodule to balance the perturbation budget and optimization speed and a\nSpatio-Temporal Invertible Neural Network (STIN) module to perform\nspatio-temporal feature space information exchanging between a source video and\nthe target feature tensor learned by GTVL module. Extensive experiments on\nUCF-101 and Kinetics-400 demonstrate that our proposed SVASTIN can generate\nadversarial examples with higher imperceptibility than the state-of-the-art\nmethods with the higher fooling rate. Code is available at\n\\href{https://github.com/Brittany-Chen/SVASTIN}{https://github.com/Brittany-Chen/SVASTIN}.\n", "date": "2024-06-04", "categories": []}
{"id": "2406.02059", "title": "Graph Adversarial Diffusion Convolution", "abstract": "  This paper introduces a min-max optimization formulation for the Graph Signal\nDenoising (GSD) problem. In this formulation, we first maximize the second term\nof GSD by introducing perturbations to the graph structure based on Laplacian\ndistance and then minimize the overall loss of the GSD. By solving the min-max\noptimization problem, we derive a new variant of the Graph Diffusion\nConvolution (GDC) architecture, called Graph Adversarial Diffusion Convolution\n(GADC). GADC differs from GDC by incorporating an additional term that enhances\nrobustness against adversarial attacks on the graph structure and noise in node\nfeatures. Moreover, GADC improves the performance of GDC on heterophilic\ngraphs. Extensive experiments demonstrate the effectiveness of GADC across\nvarious datasets. Code is available at https://github.com/SongtaoLiu0823/GADC.\n", "date": "2024-06-04", "categories": []}
{"id": "2406.02064", "title": "Advancing Generalized Transfer Attack with Initialization Derived\n  Bilevel Optimization and Dynamic Sequence Truncation", "abstract": "  Transfer attacks generate significant interest for real-world black-box\napplications by crafting transferable adversarial examples through surrogate\nmodels. Whereas, existing works essentially directly optimize the single-level\nobjective w.r.t. the surrogate model, which always leads to poor\ninterpretability of attack mechanism and limited generalization performance\nover unknown victim models. In this work, we propose the\n\\textbf{B}il\\textbf{E}vel \\textbf{T}ransfer \\textbf{A}ttac\\textbf{K} (BETAK)\nframework by establishing an initialization derived bilevel optimization\nparadigm, which explicitly reformulates the nested constraint relationship\nbetween the Upper-Level (UL) pseudo-victim attacker and the Lower-Level (LL)\nsurrogate attacker. Algorithmically, we introduce the Hyper Gradient Response\n(HGR) estimation as an effective feedback for the transferability over\npseudo-victim attackers, and propose the Dynamic Sequence Truncation (DST)\ntechnique to dynamically adjust the back-propagation path for HGR and reduce\ncomputational overhead simultaneously. Meanwhile, we conduct detailed\nalgorithmic analysis and provide convergence guarantee to support non-convexity\nof the LL surrogate attacker. Extensive evaluations demonstrate substantial\nimprovement of BETAK (e.g., $\\mathbf{53.41}$\\% increase of attack success rates\nagainst IncRes-v$2_{ens}$) against different victims and defense methods in\ntargeted and untargeted attack scenarios. The source code is available at\nhttps://github.com/callous-youth/BETAK.\n", "date": "2024-06-04", "categories": []}
{"id": "2309.17348", "title": "Intrinsic Biologically Plausible Adversarial Robustness", "abstract": "  Artificial Neural Networks (ANNs) trained with Backpropagation (BP) excel in\ndifferent daily tasks but have a dangerous vulnerability: inputs with small\ntargeted perturbations, also known as adversarial samples, can drastically\ndisrupt their performance. Adversarial training, a technique in which the\ntraining dataset is augmented with exemplary adversarial samples, is proven to\nmitigate this problem but comes at a high computational cost. In contrast to\nANNs, humans are not susceptible to misclassifying these same adversarial\nsamples. Thus, one can postulate that biologically-plausible trained ANNs might\nbe more robust against adversarial attacks. In this work, we chose the\nbiologically-plausible learning algorithm Present the Error to Perturb the\nInput To modulate Activity (PEPITA) as a case study and investigated this\nquestion through a comparative analysis with BP-trained ANNs on various\ncomputer vision tasks. We observe that PEPITA has a higher intrinsic\nadversarial robustness and, when adversarially trained, also has a more\nfavorable natural-vs-adversarial performance trade-off. In particular, for the\nsame natural accuracies on the MNIST task, PEPITA's adversarial accuracies\ndecrease on average only by 0.26% while BP's decrease by 8.05%.\n", "date": "2023-09-29", "categories": []}
{"id": "2402.08567", "title": "Agent Smith: A Single Image Can Jailbreak One Million Multimodal LLM\n  Agents Exponentially Fast", "abstract": "  A multimodal large language model (MLLM) agent can receive instructions,\ncapture images, retrieve histories from memory, and decide which tools to use.\nNonetheless, red-teaming efforts have revealed that adversarial images/prompts\ncan jailbreak an MLLM and cause unaligned behaviors. In this work, we report an\neven more severe safety issue in multi-agent environments, referred to as\ninfectious jailbreak. It entails the adversary simply jailbreaking a single\nagent, and without any further intervention from the adversary, (almost) all\nagents will become infected exponentially fast and exhibit harmful behaviors.\nTo validate the feasibility of infectious jailbreak, we simulate multi-agent\nenvironments containing up to one million LLaVA-1.5 agents, and employ\nrandomized pair-wise chat as a proof-of-concept instantiation for multi-agent\ninteraction. Our results show that feeding an (infectious) adversarial image\ninto the memory of any randomly chosen agent is sufficient to achieve\ninfectious jailbreak. Finally, we derive a simple principle for determining\nwhether a defense mechanism can provably restrain the spread of infectious\njailbreak, but how to design a practical defense that meets this principle\nremains an open question to investigate. Our project page is available at\nhttps://sail-sg.github.io/Agent-Smith/.\n", "date": "2024-02-13", "categories": []}
{"id": "2404.05311", "title": "BruSLeAttack: A Query-Efficient Score-Based Black-Box Sparse Adversarial\n  Attack", "abstract": "  We study the unique, less-well understood problem of generating sparse\nadversarial samples simply by observing the score-based replies to model\nqueries. Sparse attacks aim to discover a minimum number-the l0\nbounded-perturbations to model inputs to craft adversarial examples and\nmisguide model decisions. But, in contrast to query-based dense attack\ncounterparts against black-box models, constructing sparse adversarial\nperturbations, even when models serve confidence score information to queries\nin a score-based setting, is non-trivial. Because, such an attack leads to i)\nan NP-hard problem; and ii) a non-differentiable search space. We develop the\nBruSLeAttack-a new, faster (more query-efficient) Bayesian algorithm for the\nproblem. We conduct extensive attack evaluations including an attack\ndemonstration against a Machine Learning as a Service (MLaaS) offering\nexemplified by Google Cloud Vision and robustness testing of adversarial\ntraining regimes and a recent defense against black-box attacks. The proposed\nattack scales to achieve state-of-the-art attack success rates and query\nefficiency on standard computer vision tasks such as ImageNet across different\nmodel architectures. Our artefacts and DIY attack samples are available on\nGitHub. Importantly, our work facilitates faster evaluation of model\nvulnerabilities and raises our vigilance on the safety, security and\nreliability of deployed systems.\n", "date": "2024-04-08", "categories": []}
{"id": "2404.16307", "title": "Boosting Model Resilience via Implicit Adversarial Data Augmentation", "abstract": "  Data augmentation plays a pivotal role in enhancing and diversifying training\ndata. Nonetheless, consistently improving model performance in varied learning\nscenarios, especially those with inherent data biases, remains challenging. To\naddress this, we propose to augment the deep features of samples by\nincorporating their adversarial and anti-adversarial perturbation\ndistributions, enabling adaptive adjustment in the learning difficulty tailored\nto each sample's specific characteristics. We then theoretically reveal that\nour augmentation process approximates the optimization of a surrogate loss\nfunction as the number of augmented copies increases indefinitely. This insight\nleads us to develop a meta-learning-based framework for optimizing classifiers\nwith this novel loss, introducing the effects of augmentation while bypassing\nthe explicit augmentation process. We conduct extensive experiments across four\ncommon biased learning scenarios: long-tail learning, generalized long-tail\nlearning, noisy label learning, and subpopulation shift learning. The empirical\nresults demonstrate that our method consistently achieves state-of-the-art\nperformance, highlighting its broad adaptability.\n", "date": "2024-04-25", "categories": []}
{"id": "2405.18627", "title": "PureGen: Universal Data Purification for Train-Time Poison Defense via\n  Generative Model Dynamics", "abstract": "  Train-time data poisoning attacks threaten machine learning models by\nintroducing adversarial examples during training, leading to misclassification.\nCurrent defense methods often reduce generalization performance, are\nattack-specific, and impose significant training overhead. To address this, we\nintroduce a set of universal data purification methods using a stochastic\ntransform, $\\Psi(x)$, realized via iterative Langevin dynamics of Energy-Based\nModels (EBMs), Denoising Diffusion Probabilistic Models (DDPMs), or both. These\napproaches purify poisoned data with minimal impact on classifier\ngeneralization. Our specially trained EBMs and DDPMs provide state-of-the-art\ndefense against various attacks (including Narcissus, Bullseye Polytope,\nGradient Matching) on CIFAR-10, Tiny-ImageNet, and CINIC-10, without needing\nattack or classifier-specific information. We discuss performance trade-offs\nand show that our methods remain highly effective even with poisoned or\ndistributionally shifted generative model training data.\n", "date": "2024-05-28", "categories": []}
{"id": "2405.18741", "title": "Genshin: General Shield for Natural Language Processing with Large\n  Language Models", "abstract": "  Large language models (LLMs) like ChatGPT, Gemini, or LLaMA have been\ntrending recently, demonstrating considerable advancement and generalizability\npower in countless domains. However, LLMs create an even bigger black box\nexacerbating opacity, with interpretability limited to few approaches. The\nuncertainty and opacity embedded in LLMs' nature restrict their application in\nhigh-stakes domains like financial fraud, phishing, etc. Current approaches\nmainly rely on traditional textual classification with posterior interpretable\nalgorithms, suffering from attackers who may create versatile adversarial\nsamples to break the system's defense, forcing users to make trade-offs between\nefficiency and robustness. To address this issue, we propose a novel cascading\nframework called Genshin (General Shield for Natural Language Processing with\nLarge Language Models), utilizing LLMs as defensive one-time plug-ins. Unlike\nmost applications of LLMs that try to transform text into something new or\nstructural, Genshin uses LLMs to recover text to its original state. Genshin\naims to combine the generalizability of the LLM, the discrimination of the\nmedian model, and the interpretability of the simple model. Our experiments on\nthe task of sentimental analysis and spam detection have shown fatal flaws of\nthe current median models and exhilarating results on LLMs' recovery ability,\ndemonstrating that Genshin is both effective and efficient. In our ablation\nstudy, we unearth several intriguing observations. Utilizing the LLM defender,\na tool derived from the 4th paradigm, we have reproduced BERT's 15% optimal\nmask rate results in the 3rd paradigm of NLP. Additionally, when employing the\nLLM as a potential adversarial tool, attackers are capable of executing\neffective attacks that are nearly semantically lossless.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.19376", "title": "PureEBM: Universal Poison Purification via Mid-Run Dynamics of\n  Energy-Based Models", "abstract": "  Data poisoning attacks pose a significant threat to the integrity of machine\nlearning models by leading to misclassification of target distribution data by\ninjecting adversarial examples during training. Existing state-of-the-art\n(SoTA) defense methods suffer from limitations, such as significantly reduced\ngeneralization performance and significant overhead during training, making\nthem impractical or limited for real-world applications. In response to this\nchallenge, we introduce a universal data purification method that defends\nnaturally trained classifiers from malicious white-, gray-, and black-box image\npoisons by applying a universal stochastic preprocessing step $\\Psi_{T}(x)$,\nrealized by iterative Langevin sampling of a convergent Energy Based Model\n(EBM) initialized with an image $x.$ Mid-run dynamics of $\\Psi_{T}(x)$ purify\npoison information with minimal impact on features important to the\ngeneralization of a classifier network. We show that EBMs remain universal\npurifiers, even in the presence of poisoned EBM training data, and achieve SoTA\ndefense on leading triggered and triggerless poisons. This work is a subset of\na larger framework introduced in \\pgen with a more detailed focus on EBM\npurification and poison defense.\n", "date": "2024-05-28", "categories": []}
{"id": "2406.00494", "title": "Activation-Descent Regularization for Input Optimization of ReLU\n  Networks", "abstract": "  We present a new approach for input optimization of ReLU networks that\nexplicitly takes into account the effect of changes in activation patterns. We\nanalyze local optimization steps in both the input space and the space of\nactivation patterns to propose methods with superior local descent properties.\nTo accomplish this, we convert the discrete space of activation patterns into\ndifferentiable representations and propose regularization terms that improve\neach descent step. Our experiments demonstrate the effectiveness of the\nproposed input-optimization methods for improving the state-of-the-art in\nvarious areas, such as adversarial learning, generative modeling, and\nreinforcement learning.\n", "date": "2024-06-01", "categories": []}
{"id": "2406.00685", "title": "Improving Accuracy-robustness Trade-off via Pixel Reweighted Adversarial\n  Training", "abstract": "  Adversarial training (AT) trains models using adversarial examples (AEs),\nwhich are natural images modified with specific perturbations to mislead the\nmodel. These perturbations are constrained by a predefined perturbation budget\n$\\epsilon$ and are equally applied to each pixel within an image. However, in\nthis paper, we discover that not all pixels contribute equally to the accuracy\non AEs (i.e., robustness) and accuracy on natural images (i.e., accuracy).\nMotivated by this finding, we propose Pixel-reweighted AdveRsarial Training\n(PART), a new framework that partially reduces $\\epsilon$ for less influential\npixels, guiding the model to focus more on key regions that affect its outputs.\nSpecifically, we first use class activation mapping (CAM) methods to identify\nimportant pixel regions, then we keep the perturbation budget for these regions\nwhile lowering it for the remaining regions when generating AEs. In the end, we\nuse these pixel-reweighted AEs to train a model. PART achieves a notable\nimprovement in accuracy without compromising robustness on CIFAR-10, SVHN and\nTinyImagenet-200, justifying the necessity to allocate distinct weights to\ndifferent pixel regions in robust classification.\n", "date": "2024-06-02", "categories": []}
{"id": "2406.00816", "title": "Invisible Backdoor Attacks on Diffusion Models", "abstract": "  In recent years, diffusion models have achieved remarkable success in the\nrealm of high-quality image generation, garnering increased attention. This\nsurge in interest is paralleled by a growing concern over the security threats\nassociated with diffusion models, largely attributed to their susceptibility to\nmalicious exploitation. Notably, recent research has brought to light the\nvulnerability of diffusion models to backdoor attacks, enabling the generation\nof specific target images through corresponding triggers. However, prevailing\nbackdoor attack methods rely on manually crafted trigger generation functions,\noften manifesting as discernible patterns incorporated into input noise, thus\nrendering them susceptible to human detection. In this paper, we present an\ninnovative and versatile optimization framework designed to acquire invisible\ntriggers, enhancing the stealthiness and resilience of inserted backdoors. Our\nproposed framework is applicable to both unconditional and conditional\ndiffusion models, and notably, we are the pioneers in demonstrating the\nbackdooring of diffusion models within the context of text-guided image editing\nand inpainting pipelines. Moreover, we also show that the backdoors in the\nconditional generation can be directly applied to model watermarking for model\nownership verification, which further boosts the significance of the proposed\nframework. Extensive experiments on various commonly used samplers and datasets\nverify the efficacy and stealthiness of the proposed framework. Our code is\npublicly available at\nhttps://github.com/invisibleTriggerDiffusion/invisible_triggers_for_diffusion.\n", "date": "2024-06-02", "categories": []}
{"id": "2406.01365", "title": "From Feature Visualization to Visual Circuits: Effect of Adversarial\n  Model Manipulation", "abstract": "  Understanding the inner working functionality of large-scale deep neural\nnetworks is challenging yet crucial in several high-stakes applications.\nMechanistic inter- pretability is an emergent field that tackles this\nchallenge, often by identifying human-understandable subgraphs in deep neural\nnetworks known as circuits. In vision-pretrained models, these subgraphs are\nusually interpreted by visualizing their node features through a popular\ntechnique called feature visualization. Recent works have analyzed the\nstability of different feature visualization types under the adversarial model\nmanipulation framework. This paper starts by addressing limitations in existing\nworks by proposing a novel attack called ProxPulse that simultaneously\nmanipulates the two types of feature visualizations. Surprisingly, when\nanalyzing these attacks under the umbrella of visual circuits, we find that\nvisual circuits show some robustness to ProxPulse. We, therefore, introduce a\nnew attack based on ProxPulse that unveils the manipulability of visual\ncircuits, shedding light on their lack of robustness. The effectiveness of\nthese attacks is validated using pre-trained AlexNet and ResNet-50 models on\nImageNet.\n", "date": "2024-06-03", "categories": []}
{"id": "2311.15356", "title": "Having Second Thoughts? Let's hear it", "abstract": "  Deep learning models loosely mimic bottom-up signal pathways from low-order\nsensory areas to high-order cognitive areas. After training, DL models can\noutperform humans on some domain-specific tasks, but their decision-making\nprocess has been known to be easily disrupted. Since the human brain consists\nof multiple functional areas highly connected to one another and relies on\nintricate interplays between bottom-up and top-down (from high-order to\nlow-order areas) processing, we hypothesize that incorporating top-down signal\nprocessing may make DL models more robust. To address this hypothesis, we\npropose a certification process mimicking selective attention and test if it\ncould make DL models more robust. Our empirical evaluations suggest that this\nnewly proposed certification can improve DL models' accuracy and help us build\nsafety measures to alleviate their vulnerabilities with both artificial and\nnatural adversarial examples.\n", "date": "2023-11-26", "categories": []}
{"id": "2404.11265", "title": "The Victim and The Beneficiary: Exploiting a Poisoned Model to Train a\n  Clean Model on Poisoned Data", "abstract": "  Recently, backdoor attacks have posed a serious security threat to the\ntraining process of deep neural networks (DNNs). The attacked model behaves\nnormally on benign samples but outputs a specific result when the trigger is\npresent. However, compared with the rocketing progress of backdoor attacks,\nexisting defenses are difficult to deal with these threats effectively or\nrequire benign samples to work, which may be unavailable in real scenarios. In\nthis paper, we find that the poisoned samples and benign samples can be\ndistinguished with prediction entropy. This inspires us to propose a novel\ndual-network training framework: The Victim and The Beneficiary (V&B), which\nexploits a poisoned model to train a clean model without extra benign samples.\nFirstly, we sacrifice the Victim network to be a powerful poisoned sample\ndetector by training on suspicious samples. Secondly, we train the Beneficiary\nnetwork on the credible samples selected by the Victim to inhibit backdoor\ninjection. Thirdly, a semi-supervised suppression strategy is adopted for\nerasing potential backdoors and improving model performance. Furthermore, to\nbetter inhibit missed poisoned samples, we propose a strong data augmentation\nmethod, AttentionMix, which works well with our proposed V&B framework.\nExtensive experiments on two widely used datasets against 6 state-of-the-art\nattacks demonstrate that our framework is effective in preventing backdoor\ninjection and robust to various attacks while maintaining the performance on\nbenign samples. Our code is available at https://github.com/Zixuan-Zhu/VaB.\n", "date": "2024-04-17", "categories": []}
{"id": "2405.12424", "title": "Rethinking Robustness Assessment: Adversarial Attacks on Learning-based\n  Quadrupedal Locomotion Controllers", "abstract": "  Legged locomotion has recently achieved remarkable success with the progress\nof machine learning techniques, especially deep reinforcement learning (RL).\nControllers employing neural networks have demonstrated empirical and\nqualitative robustness against real-world uncertainties, including sensor noise\nand external perturbations. However, formally investigating the vulnerabilities\nof these locomotion controllers remains a challenge. This difficulty arises\nfrom the requirement to pinpoint vulnerabilities across a long-tailed\ndistribution within a high-dimensional, temporally sequential space. As a first\nstep towards quantitative verification, we propose a computational method that\nleverages sequential adversarial attacks to identify weaknesses in learned\nlocomotion controllers. Our research demonstrates that, even state-of-the-art\nrobust controllers can fail significantly under well-designed, low-magnitude\nadversarial sequence. Through experiments in simulation and on the real robot,\nwe validate our approach's effectiveness, and we illustrate how the results it\ngenerates can be used to robustify the original policy and offer valuable\ninsights into the safety of these black-box policies. Project page:\nhttps://fanshi14.github.io/me/rss24.html\n", "date": "2024-05-21", "categories": []}
{"id": "2405.20672", "title": "Investigating and unmasking feature-level vulnerabilities of CNNs to\n  adversarial perturbations", "abstract": "  This study explores the impact of adversarial perturbations on Convolutional\nNeural Networks (CNNs) with the aim of enhancing the understanding of their\nunderlying mechanisms. Despite numerous defense methods proposed in the\nliterature, there is still an incomplete understanding of this phenomenon.\nInstead of treating the entire model as vulnerable, we propose that specific\nfeature maps learned during training contribute to the overall vulnerability.\nTo investigate how the hidden representations learned by a CNN affect its\nvulnerability, we introduce the Adversarial Intervention framework. Experiments\nwere conducted on models trained on three well-known computer vision datasets,\nsubjecting them to attacks of different nature. Our focus centers on the\neffects that adversarial perturbations to a model's initial layer have on the\noverall behavior of the model. Empirical results revealed compelling insights:\na) perturbing selected channel combinations in shallow layers causes\nsignificant disruptions; b) the channel combinations most responsible for the\ndisruptions are common among different types of attacks; c) despite shared\nvulnerable combinations of channels, different attacks affect hidden\nrepresentations with varying magnitudes; d) there exists a positive correlation\nbetween a kernel's magnitude and its vulnerability. In conclusion, this work\nintroduces a novel framework to study the vulnerability of a CNN model to\nadversarial perturbations, revealing insights that contribute to a deeper\nunderstanding of the phenomenon. The identified properties pave the way for the\ndevelopment of efficient ad-hoc defense mechanisms in future applications.\n", "date": "2024-05-31", "categories": []}
{"id": "2312.07865", "title": "SimAC: A Simple Anti-Customization Method for Protecting Face Privacy\n  against Text-to-Image Synthesis of Diffusion Models", "abstract": "  Despite the success of diffusion-based customization methods on visual\ncontent creation, increasing concerns have been raised about such techniques\nfrom both privacy and political perspectives. To tackle this issue, several\nanti-customization methods have been proposed in very recent months,\npredominantly grounded in adversarial attacks. Unfortunately, most of these\nmethods adopt straightforward designs, such as end-to-end optimization with a\nfocus on adversarially maximizing the original training loss, thereby\nneglecting nuanced internal properties intrinsic to the diffusion model, and\neven leading to ineffective optimization in some diffusion time steps.In this\npaper, we strive to bridge this gap by undertaking a comprehensive exploration\nof these inherent properties, to boost the performance of current\nanti-customization approaches. Two aspects of properties are investigated: 1)\nWe examine the relationship between time step selection and the model's\nperception in the frequency domain of images and find that lower time steps can\ngive much more contributions to adversarial noises. This inspires us to propose\nan adaptive greedy search for optimal time steps that seamlessly integrates\nwith existing anti-customization methods. 2) We scrutinize the roles of\nfeatures at different layers during denoising and devise a sophisticated\nfeature-based optimization framework for anti-customization.Experiments on\nfacial benchmarks demonstrate that our approach significantly increases\nidentity disruption, thereby protecting user privacy and copyright. Our code is\navailable at: https://github.com/somuchtome/SimAC.\n", "date": "2023-12-13", "categories": []}
{"id": "2403.05955", "title": "IOI: Invisible One-Iteration Adversarial Attack on No-Reference Image-\n  and Video-Quality Metrics", "abstract": "  No-reference image- and video-quality metrics are widely used in video\nprocessing benchmarks. The robustness of learning-based metrics under video\nattacks has not been widely studied. In addition to having success, attacks\nthat can be employed in video processing benchmarks must be fast and\nimperceptible. This paper introduces an Invisible One-Iteration (IOI)\nadversarial attack on no reference image and video quality metrics. We compared\nour method alongside eight prior approaches using image and video datasets via\nobjective and subjective tests. Our method exhibited superior visual quality\nacross various attacked metric architectures while maintaining comparable\nattack success and speed. We made the code available on GitHub:\nhttps://github.com/katiashh/ioi-attack.\n", "date": "2024-03-09", "categories": []}
{"id": "2405.19424", "title": "Diffusion Policy Attacker: Crafting Adversarial Attacks for\n  Diffusion-based Policies", "abstract": "  Diffusion models (DMs) have emerged as a promising approach for behavior\ncloning (BC). Diffusion policies (DP) based on DMs have elevated BC performance\nto new heights, demonstrating robust efficacy across diverse tasks, coupled\nwith their inherent flexibility and ease of implementation. Despite the\nincreasing adoption of DP as a foundation for policy generation, the critical\nissue of safety remains largely unexplored. While previous attempts have\ntargeted deep policy networks, DP used diffusion models as the policy network,\nmaking it ineffective to be attacked using previous methods because of its\nchained structure and randomness injected. In this paper, we undertake a\ncomprehensive examination of DP safety concerns by introducing adversarial\nscenarios, encompassing offline and online attacks, and global and patch-based\nattacks. We propose DP-Attacker, a suite of algorithms that can craft effective\nadversarial attacks across all aforementioned scenarios. We conduct attacks on\npre-trained diffusion policies across various manipulation tasks. Through\nextensive experiments, we demonstrate that DP-Attacker has the capability to\nsignificantly decrease the success rate of DP for all scenarios. Particularly\nin offline scenarios, DP-Attacker can generate highly transferable\nperturbations applicable to all frames. Furthermore, we illustrate the creation\nof adversarial physical patches that, when applied to the environment,\neffectively deceive the model. Video results are put in:\nhttps://sites.google.com/view/diffusion-policy-attacker.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.19956", "title": "HOLMES: to Detect Adversarial Examples with Multiple Detectors", "abstract": "  Deep neural networks (DNNs) can easily be cheated by some imperceptible but\npurposeful noise added to images, and erroneously classify them. Previous\ndefensive work mostly focused on retraining the models or detecting the noise,\nbut has either shown limited success rates or been attacked by new adversarial\nexamples. Instead of focusing on adversarial images or the interior of DNN\nmodels, we observed that adversarial examples generated by different algorithms\ncan be identified based on the output of DNNs (logits). Logit can serve as an\nexterior feature to train detectors. Then, we propose HOLMES (Hierarchically\nOrganized Light-weight Multiple dEtector System) to reinforce DNNs by detecting\npotential adversarial examples to minimize the threats they may bring in\npractical. HOLMES is able to distinguish \\textit{unseen} adversarial examples\nfrom multiple attacks with high accuracy and low false positive rates than\nsingle detector systems even in an adaptive model. To ensure the diversity and\nrandomness of detectors in HOLMES, we use two methods: training dedicated\ndetectors for each label and training detectors with top-k logits. Our\neffective and inexpensive strategies neither modify original DNN models nor\nrequire its internal parameters. HOLMES is not only compatible with all kinds\nof learning models (even only with external APIs), but also complementary to\nother defenses to achieve higher detection rates (may also fully protect the\nsystem against various adversarial examples).\n", "date": "2024-05-30", "categories": []}
{"id": "2405.19990", "title": "DiffPhysBA: Diffusion-based Physical Backdoor Attack against Person\n  Re-Identification in Real-World", "abstract": "  Person Re-Identification (ReID) systems pose a significant security risk from\nbackdoor attacks, allowing adversaries to evade tracking or impersonate others.\nBeyond recognizing this issue, we investigate how backdoor attacks can be\ndeployed in real-world scenarios, where a ReID model is typically trained on\ndata collected in the digital domain and then deployed in a physical\nenvironment. This attack scenario requires an attack flow that embeds backdoor\ntriggers in the digital domain realistically enough to also activate the buried\nbackdoor in person ReID models in the physical domain. This paper realizes this\nattack flow by leveraging a diffusion model to generate realistic accessories\non pedestrian images (e.g., bags, hats, etc.) as backdoor triggers. However,\nthe noticeable domain gap between the triggers generated by the off-the-shelf\ndiffusion model and their physical counterparts results in a low attack success\nrate. Therefore, we introduce a novel diffusion-based physical backdoor attack\n(DiffPhysBA) method that adopts a training-free similarity-guided sampling\nprocess to enhance the resemblance between generated and physical triggers.\nConsequently, DiffPhysBA can generate realistic attributes as semantic-level\ntriggers in the digital domain and provides higher physical ASR compared to the\ndirect paste method by 25.6% on the real-world test set. Through evaluations on\nnewly proposed real-world and synthetic ReID test sets, DiffPhysBA demonstrates\nan impressive success rate exceeding 90% in both the digital and physical\ndomains. Notably, it excels in digital stealth metrics and can effectively\nevade state-of-the-art defense methods.\n", "date": "2024-05-30", "categories": []}
{"id": "2403.10461", "title": "Introducing Adaptive Continuous Adversarial Training (ACAT) to Enhance\n  ML Robustness", "abstract": "  Adversarial training enhances the robustness of Machine Learning (ML) models\nagainst adversarial attacks. However, obtaining labeled training and\nadversarial training data in network/cybersecurity domains is challenging and\ncostly. Therefore, this letter introduces Adaptive Continuous Adversarial\nTraining (ACAT), a method that integrates adversarial training samples into the\nmodel during continuous learning sessions using real-world detected adversarial\ndata. Experimental results with a SPAM detection dataset demonstrate that ACAT\nreduces the time required for adversarial sample detection compared to\ntraditional processes. Moreover, the accuracy of the under-attack ML-based SPAM\nfilter increased from 69% to over 88% after just three retraining sessions.\n", "date": "2024-03-15", "categories": []}
{"id": "2405.19098", "title": "Efficient Black-box Adversarial Attacks via Bayesian Optimization Guided\n  by a Function Prior", "abstract": "  This paper studies the challenging black-box adversarial attack that aims to\ngenerate adversarial examples against a black-box model by only using output\nfeedback of the model to input queries. Some previous methods improve the query\nefficiency by incorporating the gradient of a surrogate white-box model into\nquery-based attacks due to the adversarial transferability. However, the\nlocalized gradient is not informative enough, making these methods still\nquery-intensive. In this paper, we propose a Prior-guided Bayesian Optimization\n(P-BO) algorithm that leverages the surrogate model as a global function prior\nin black-box adversarial attacks. As the surrogate model contains rich prior\ninformation of the black-box one, P-BO models the attack objective with a\nGaussian process whose mean function is initialized as the surrogate model's\nloss. Our theoretical analysis on the regret bound indicates that the\nperformance of P-BO may be affected by a bad prior. Therefore, we further\npropose an adaptive integration strategy to automatically adjust a coefficient\non the function prior by minimizing the regret bound. Extensive experiments on\nimage classifiers and large vision-language models demonstrate the superiority\nof the proposed algorithm in reducing queries and improving attack success\nrates compared with the state-of-the-art black-box attacks. Code is available\nat https://github.com/yibo-miao/PBO-Attack.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.15244", "title": "Adversarial Attacks on Hidden Tasks in Multi-Task Learning", "abstract": "  Deep learning models are susceptible to adversarial attacks, where slight\nperturbations to input data lead to misclassification. Adversarial attacks\nbecome increasingly effective with access to information about the targeted\nclassifier. In the context of multi-task learning, where a single model learns\nmultiple tasks simultaneously, attackers may aim to exploit vulnerabilities in\nspecific tasks with limited information. This paper investigates the\nfeasibility of attacking hidden tasks within multi-task classifiers, where\nmodel access regarding the hidden target task and labeled data for the hidden\ntarget task are not available, but model access regarding the non-target tasks\nis available. We propose a novel adversarial attack method that leverages\nknowledge from non-target tasks and the shared backbone network of the\nmulti-task model to force the model to forget knowledge related to the target\ntask. Experimental results on CelebA and DeepFashion datasets demonstrate the\neffectiveness of our method in degrading the accuracy of hidden tasks while\npreserving the performance of visible tasks, contributing to the understanding\nof adversarial vulnerabilities in multi-task classifiers.\n", "date": "2024-05-24", "categories": []}
{"id": "2405.17678", "title": "TIMA: Text-Image Mutual Awareness for Balancing Zero-Shot Adversarial\n  Robustness and Generalization Ability", "abstract": "  This work addresses the challenge of achieving zero-shot adversarial\nrobustness while preserving zero-shot generalization in large-scale foundation\nmodels, with a focus on the popular Contrastive Language-Image Pre-training\n(CLIP). Although foundation models were reported to have exceptional zero-shot\ngeneralization, they are highly vulnerable to adversarial perturbations.\nExisting methods achieve a comparable good tradeoff between zero-shot\nadversarial robustness and generalization under small adversarial\nperturbations. However, they fail to achieve a good tradeoff under large\nadversarial perturbations. To this end, we propose a novel Text-Image Mutual\nAwareness (TIMA) method that strikes a balance between zero-shot adversarial\nrobustness and generalization. More precisely, we propose an Image-Aware Text\n(IAT) tuning mechanism that increases the inter-class distance of text\nembeddings by incorporating the Minimum Hyperspherical Energy (MHE).\nSimultaneously, fixed pre-trained image embeddings are used as cross-modal\nauxiliary supervision to maintain the similarity between the MHE-tuned and\noriginal text embeddings by the knowledge distillation, preserving semantic\ninformation between different classes. Besides, we introduce a Text-Aware Image\n(TAI) tuning mechanism, which increases inter-class distance between image\nembeddings during the training stage by Text-distance based Adaptive Margin\n(TAM). Similarly, a knowledge distillation is utilized to retain the similarity\nbetween fine-tuned and pre-trained image embeddings. Extensive experimental\nresults demonstrate the effectiveness of our approach, showing impressive\nzero-shot performance against a wide range of adversarial perturbations while\npreserving the zero-shot generalization capabilities of the original CLIP\nmodel.\n", "date": "2024-05-27", "categories": []}
{"id": "2405.17929", "title": "Towards Unified Robustness Against Both Backdoor and Adversarial Attacks", "abstract": "  Deep Neural Networks (DNNs) are known to be vulnerable to both backdoor and\nadversarial attacks. In the literature, these two types of attacks are commonly\ntreated as distinct robustness problems and solved separately, since they\nbelong to training-time and inference-time attacks respectively. However, this\npaper revealed that there is an intriguing connection between them: (1)\nplanting a backdoor into a model will significantly affect the model's\nadversarial examples; (2) for an infected model, its adversarial examples have\nsimilar features as the triggered images. Based on these observations, a novel\nProgressive Unified Defense (PUD) algorithm is proposed to defend against\nbackdoor and adversarial attacks simultaneously. Specifically, our PUD has a\nprogressive model purification scheme to jointly erase backdoors and enhance\nthe model's adversarial robustness. At the early stage, the adversarial\nexamples of infected models are utilized to erase backdoors. With the backdoor\ngradually erased, our model purification can naturally turn into a stage to\nboost the model's robustness against adversarial attacks. Besides, our PUD\nalgorithm can effectively identify poisoned images, which allows the initial\nextra dataset not to be completely clean. Extensive experimental results show\nthat, our discovered connection between backdoor and adversarial attacks is\nubiquitous, no matter what type of backdoor attack. The proposed PUD\noutperforms the state-of-the-art backdoor defense, including the model\nrepairing-based and data filtering-based methods. Besides, it also has the\nability to compete with the most advanced adversarial defense methods.\n", "date": "2024-05-28", "categories": []}
{"id": "2405.17984", "title": "Cross-Context Backdoor Attacks against Graph Prompt Learning", "abstract": "  Graph Prompt Learning (GPL) bridges significant disparities between\npretraining and downstream applications to alleviate the knowledge transfer\nbottleneck in real-world graph learning. While GPL offers superior\neffectiveness in graph knowledge transfer and computational efficiency, the\nsecurity risks posed by backdoor poisoning effects embedded in pretrained\nmodels remain largely unexplored. Our study provides a comprehensive analysis\nof GPL's vulnerability to backdoor attacks. We introduce \\textit{CrossBA}, the\nfirst cross-context backdoor attack against GPL, which manipulates only the\npretraining phase without requiring knowledge of downstream applications. Our\ninvestigation reveals both theoretically and empirically that tuning trigger\ngraphs, combined with prompt transformations, can seamlessly transfer the\nbackdoor threat from pretrained encoders to downstream applications. Through\nextensive experiments involving 3 representative GPL methods across 5 distinct\ncross-context scenarios and 5 benchmark datasets of node and graph\nclassification tasks, we demonstrate that \\textit{CrossBA} consistently\nachieves high attack success rates while preserving the functionality of\ndownstream applications over clean input. We also explore potential\ncountermeasures against \\textit{CrossBA} and conclude that current defenses are\ninsufficient to mitigate \\textit{CrossBA}. Our study highlights the persistent\nbackdoor threats to GPL systems, raising trustworthiness concerns in the\npractices of GPL techniques.\n", "date": "2024-05-28", "categories": []}
{"id": "2305.05400", "title": "Investigating the Corruption Robustness of Image Classifiers with Random\n  Lp-norm Corruptions", "abstract": "  Robustness is a fundamental property of machine learning classifiers required\nto achieve safety and reliability. In the field of adversarial robustness of\nimage classifiers, robustness is commonly defined as the stability of a model\nto all input changes within a p-norm distance. However, in the field of random\ncorruption robustness, variations observed in the real world are used, while\np-norm corruptions are rarely considered. This study investigates the use of\nrandom p-norm corruptions to augment the training and test data of image\nclassifiers. We evaluate the model robustness against imperceptible random\np-norm corruptions and propose a novel robustness metric. We empirically\ninvestigate whether robustness transfers across different p-norms and derive\nconclusions on which p-norm corruptions a model should be trained and\nevaluated. We find that training data augmentation with a combination of p-norm\ncorruptions significantly improves corruption robustness, even on top of\nstate-of-the-art data augmentation schemes.\n", "date": "2023-05-09", "categories": []}
{"id": "2307.16865", "title": "Universal Adversarial Defense in Remote Sensing Based on Pre-trained\n  Denoising Diffusion Models", "abstract": "  Deep neural networks (DNNs) have risen to prominence as key solutions in\nnumerous AI applications for earth observation (AI4EO). However, their\nsusceptibility to adversarial examples poses a critical challenge, compromising\nthe reliability of AI4EO algorithms. This paper presents a novel Universal\nAdversarial Defense approach in Remote Sensing Imagery (UAD-RS), leveraging\npre-trained diffusion models to protect DNNs against universal adversarial\nexamples exhibiting heterogeneous patterns. Specifically, a universal\nadversarial purification framework is developed utilizing pre-trained diffusion\nmodels to mitigate adversarial perturbations through the introduction of\nGaussian noise and subsequent purification of the perturbations from\nadversarial examples. Additionally, an Adaptive Noise Level Selection (ANLS)\nmechanism is introduced to determine the optimal noise level for the\npurification framework with a task-guided Frechet Inception Distance (FID)\nranking strategy, thereby enhancing purification performance. Consequently,\nonly a single pre-trained diffusion model is required for purifying universal\nadversarial samples with heterogeneous patterns across each dataset,\nsignificantly reducing training efforts for multiple attack settings while\nmaintaining high performance without prior knowledge of adversarial\nperturbations. Experimental results on four heterogeneous RS datasets, focusing\non scene classification and semantic segmentation, demonstrate that UAD-RS\noutperforms state-of-the-art adversarial purification approaches, providing\nuniversal defense against seven commonly encountered adversarial perturbations.\nCodes and the pre-trained models are available online\n(https://github.com/EricYu97/UAD-RS).\n", "date": "2023-07-31", "categories": []}
{"id": "2309.16096", "title": "Adversarial Examples Might be Avoidable: The Role of Data Concentration\n  in Adversarial Robustness", "abstract": "  The susceptibility of modern machine learning classifiers to adversarial\nexamples has motivated theoretical results suggesting that these might be\nunavoidable. However, these results can be too general to be applicable to\nnatural data distributions. Indeed, humans are quite robust for tasks involving\nvision. This apparent conflict motivates a deeper dive into the question: Are\nadversarial examples truly unavoidable? In this work, we theoretically\ndemonstrate that a key property of the data distribution -- concentration on\nsmall-volume subsets of the input space -- determines whether a robust\nclassifier exists. We further demonstrate that, for a data distribution\nconcentrated on a union of low-dimensional linear subspaces, utilizing\nstructure in data naturally leads to classifiers that enjoy data-dependent\npolyhedral robustness guarantees, improving upon methods for provable\ncertification in certain regimes.\n", "date": "2023-09-28", "categories": []}
{"id": "2401.04136", "title": "The Stronger the Diffusion Model, the Easier the Backdoor: Data\n  Poisoning to Induce Copyright Breaches Without Adjusting Finetuning Pipeline", "abstract": "  The commercialization of text-to-image diffusion models (DMs) brings forth\npotential copyright concerns. Despite numerous attempts to protect DMs from\ncopyright issues, the vulnerabilities of these solutions are underexplored. In\nthis study, we formalized the Copyright Infringement Attack on generative AI\nmodels and proposed a backdoor attack method, SilentBadDiffusion, to induce\ncopyright infringement without requiring access to or control over training\nprocesses. Our method strategically embeds connections between pieces of\ncopyrighted information and text references in poisoning data while carefully\ndispersing that information, making the poisoning data inconspicuous when\nintegrated into a clean dataset. Our experiments show the stealth and efficacy\nof the poisoning data. When given specific text prompts, DMs trained with a\npoisoning ratio of 0.20% can produce copyrighted images. Additionally, the\nresults reveal that the more sophisticated the DMs are, the easier the success\nof the attack becomes. These findings underline potential pitfalls in the\nprevailing copyright protection strategies and underscore the necessity for\nincreased scrutiny to prevent the misuse of DMs.\n", "date": "2024-01-07", "categories": []}
{"id": "2404.00924", "title": "BadPart: Unified Black-box Adversarial Patch Attacks against Pixel-wise\n  Regression Tasks", "abstract": "  Pixel-wise regression tasks (e.g., monocular depth estimation (MDE) and\noptical flow estimation (OFE)) have been widely involved in our daily life in\napplications like autonomous driving, augmented reality and video composition.\nAlthough certain applications are security-critical or bear societal\nsignificance, the adversarial robustness of such models are not sufficiently\nstudied, especially in the black-box scenario. In this work, we introduce the\nfirst unified black-box adversarial patch attack framework against pixel-wise\nregression tasks, aiming to identify the vulnerabilities of these models under\nquery-based black-box attacks. We propose a novel square-based adversarial\npatch optimization framework and employ probabilistic square sampling and\nscore-based gradient estimation techniques to generate the patch effectively\nand efficiently, overcoming the scalability problem of previous black-box patch\nattacks. Our attack prototype, named BadPart, is evaluated on both MDE and OFE\ntasks, utilizing a total of 7 models. BadPart surpasses 3 baseline methods in\nterms of both attack performance and efficiency. We also apply BadPart on the\nGoogle online service for portrait depth estimation, causing 43.5% relative\ndistance error with 50K queries. State-of-the-art (SOTA) countermeasures cannot\ndefend our attack effectively.\n", "date": "2024-04-01", "categories": []}
{"id": "2405.11336", "title": "UPAM: Unified Prompt Attack in Text-to-Image Generation Models Against\n  Both Textual Filters and Visual Checkers", "abstract": "  Text-to-Image (T2I) models have raised security concerns due to their\npotential to generate inappropriate or harmful images. In this paper, we\npropose UPAM, a novel framework that investigates the robustness of T2I models\nfrom the attack perspective. Unlike most existing attack methods that focus on\ndeceiving textual defenses, UPAM aims to deceive both textual and visual\ndefenses in T2I models. UPAM enables gradient-based optimization, offering\ngreater effectiveness and efficiency than previous methods. Given that T2I\nmodels might not return results due to defense mechanisms, we introduce a\nSphere-Probing Learning (SPL) scheme to support gradient optimization even when\nno results are returned. Additionally, we devise a Semantic-Enhancing Learning\n(SEL) scheme to finetune UPAM for generating target-aligned images. Our\nframework also ensures attack stealthiness. Extensive experiments demonstrate\nUPAM's effectiveness and efficiency.\n", "date": "2024-05-18", "categories": []}
{"id": "2405.11708", "title": "Adaptive Batch Normalization Networks for Adversarial Robustness", "abstract": "  Deep networks are vulnerable to adversarial examples. Adversarial Training\n(AT) has been a standard foundation of modern adversarial defense approaches\ndue to its remarkable effectiveness. However, AT is extremely time-consuming,\nrefraining it from wide deployment in practical applications. In this paper, we\naim at a non-AT defense: How to design a defense method that gets rid of AT but\nis still robust against strong adversarial attacks? To answer this question, we\nresort to adaptive Batch Normalization (BN), inspired by the recent advances in\ntest-time domain adaptation. We propose a novel defense accordingly, referred\nto as the Adaptive Batch Normalization Network (ABNN). ABNN employs a\npre-trained substitute model to generate clean BN statistics and sends them to\nthe target model. The target model is exclusively trained on clean data and\nlearns to align the substitute model's BN statistics. Experimental results show\nthat ABNN consistently improves adversarial robustness against both digital and\nphysically realizable attacks on both image and video datasets. Furthermore,\nABNN can achieve higher clean data performance and significantly lower training\ntime complexity compared to AT-based approaches.\n", "date": "2024-05-20", "categories": []}
{"id": "2405.15971", "title": "Robust width: A lightweight and certifiable adversarial defense", "abstract": "  Deep neural networks are vulnerable to so-called adversarial examples: inputs\nwhich are intentionally constructed to cause the model to make incorrect\npredictions or classifications. Adversarial examples are often visually\nindistinguishable from natural data samples, making them hard to detect. As\nsuch, they pose significant threats to the reliability of deep learning\nsystems. In this work, we study an adversarial defense based on the robust\nwidth property (RWP), which was recently introduced for compressed sensing. We\nshow that a specific input purification scheme based on the RWP gives\ntheoretical robustness guarantees for images that are approximately sparse. The\ndefense is easy to implement and can be applied to any existing model without\nadditional training or finetuning. We empirically validate the defense on\nImageNet against $L^\\infty$ perturbations at perturbation budgets ranging from\n$4/255$ to $32/255$. In the black-box setting, our method significantly\noutperforms the state-of-the-art, especially for large perturbations. In the\nwhite-box setting, depending on the choice of base classifier, we closely match\nthe state of the art in robust ImageNet classification while avoiding the need\nfor additional data, larger models or expensive adversarial training routines.\nOur code is available at https://github.com/peck94/robust-width-defense.\n", "date": "2024-05-24", "categories": []}
{"id": "2405.15979", "title": "BadGD: A unified data-centric framework to identify gradient descent\n  vulnerabilities", "abstract": "  We present BadGD, a unified theoretical framework that exposes the\nvulnerabilities of gradient descent algorithms through strategic backdoor\nattacks. Backdoor attacks involve embedding malicious triggers into a training\ndataset to disrupt the model's learning process. Our framework introduces three\nnovel constructs: Max RiskWarp Trigger, Max GradWarp Trigger, and Max\nGradDistWarp Trigger, each designed to exploit specific aspects of gradient\ndescent by distorting empirical risk, deterministic gradients, and stochastic\ngradients respectively. We rigorously define clean and backdoored datasets and\nprovide mathematical formulations for assessing the distortions caused by these\nmalicious backdoor triggers. By measuring the impact of these triggers on the\nmodel training procedure, our framework bridges existing empirical findings\nwith theoretical insights, demonstrating how a malicious party can exploit\ngradient descent hyperparameters to maximize attack effectiveness. In\nparticular, we show that these exploitations can significantly alter the loss\nlandscape and gradient calculations, leading to compromised model integrity and\nperformance. This research underscores the severe threats posed by such\ndata-centric attacks and highlights the urgent need for robust defenses in\nmachine learning. BadGD sets a new standard for understanding and mitigating\nadversarial manipulations, ensuring the reliability and security of AI systems.\n", "date": "2024-05-24", "categories": []}
{"id": "2405.16036", "title": "Certifying Adapters: Enabling and Enhancing the Certification of\n  Classifier Adversarial Robustness", "abstract": "  Randomized smoothing has become a leading method for achieving certified\nrobustness in deep classifiers against l_{p}-norm adversarial perturbations.\nCurrent approaches for achieving certified robustness, such as data\naugmentation with Gaussian noise and adversarial training, require expensive\ntraining procedures that tune large models for different Gaussian noise levels\nand thus cannot leverage high-performance pre-trained neural networks. In this\nwork, we introduce a novel certifying adapters framework (CAF) that enables and\nenhances the certification of classifier adversarial robustness. Our approach\nmakes few assumptions about the underlying training algorithm or feature\nextractor and is thus broadly applicable to different feature extractor\narchitectures (e.g., convolutional neural networks or vision transformers) and\nsmoothing algorithms. We show that CAF (a) enables certification in uncertified\nmodels pre-trained on clean datasets and (b) substantially improves the\nperformance of certified classifiers via randomized smoothing and SmoothAdv at\nmultiple radii in CIFAR-10 and ImageNet. We demonstrate that CAF achieves\nimproved certified accuracies when compared to methods based on random or\ndenoised smoothing, and that CAF is insensitive to certifying adapter\nhyperparameters. Finally, we show that an ensemble of adapters enables a single\npre-trained feature extractor to defend against a range of noise perturbation\nscales.\n", "date": "2024-05-25", "categories": []}
{"id": "2405.16783", "title": "TrojFM: Resource-efficient Backdoor Attacks against Very Large\n  Foundation Models", "abstract": "  One key challenge in backdoor attacks against large foundation models is the\nresource limits. Backdoor attacks usually require retraining the target model,\nwhich is impractical for very large foundation models. Existing backdoor\nattacks are mainly designed for supervised classifiers or small foundation\nmodels (e.g., BERT). None of these attacks has successfully compromised a very\nlarge foundation model, such as Llama-3-70B, especially with limited\ncomputational resources. In this paper, we propose TrojFM, a novel backdoor\nattack tailored for very large foundation models. Our primary technical\ncontribution is the development of a novel backdoor injection method. This\nmethod forces a backdoored model to generate similar hidden representations for\npoisoned inputs regardless of their actual semantics. Our approach injects such\nbackdoors by fine-tuning only a very small proportion of model parameters. This\nenables TrojFM to efficiently launch downstream task-agnostic backdoor attacks\nagainst very large foundation models under limited computational resources.\nMoreover, we optimize the fine-tuning process with our customized QLoRA\ntechnique, enabling launching our attack via only~\\textit{one A100 GPU}.\nFurthermore, we design a new trigger injection method to ensure our attack\nstealthiness. Through extensive experiments, we first demonstrate that TrojFM\ncan launch effective backdoor attacks against widely used large GPT-style\nmodels without jeopardizing their normal functionalities (and outperforming\nexisting attacks on BERT-style models). Furthermore, we show that TrojFM is\nresilient to SOTA defenses and is insensitive to changes in key\nhyper-parameters. Finally, we conduct a resource analysis to quantify that our\nmethod can significantly save computational and memory costs compared to\nexisting backdoor attacks.\n", "date": "2024-05-27", "categories": []}
{"id": "2405.17181", "title": "Spectral regularization for adversarially-robust representation learning", "abstract": "  The vulnerability of neural network classifiers to adversarial attacks is a\nmajor obstacle to their deployment in safety-critical applications.\nRegularization of network parameters during training can be used to improve\nadversarial robustness and generalization performance. Usually, the network is\nregularized end-to-end, with parameters at all layers affected by\nregularization. However, in settings where learning representations is key,\nsuch as self-supervised learning (SSL), layers after the feature representation\nwill be discarded when performing inference. For these models, regularizing up\nto the feature space is more suitable. To this end, we propose a new spectral\nregularizer for representation learning that encourages black-box adversarial\nrobustness in downstream classification tasks. In supervised classification\nsettings, we show empirically that this method is more effective in boosting\ntest accuracy and robustness than previously-proposed methods that regularize\nall layers of the network. We then show that this method improves the\nadversarial robustness of classifiers using representations learned with\nself-supervised training or transferred from another classification task. In\nall, our work begins to unveil how representational structure affects\nadversarial robustness.\n", "date": "2024-05-27", "categories": []}
{"id": "2405.14923", "title": "How Does Bayes Error Limit Probabilistic Robust Accuracy", "abstract": "  Adversarial examples pose a security threat to many critical systems built on\nneural networks. Given that deterministic robustness often comes with\nsignificantly reduced accuracy, probabilistic robustness (i.e., the probability\nof having the same label with a vicinity is $\\ge 1-\\kappa$) has been proposed\nas a promising way of achieving robustness whilst maintaining accuracy.\nHowever, existing training methods for probabilistic robustness still\nexperience non-trivial accuracy loss. It is unclear whether there is an upper\nbound on the accuracy when optimising towards probabilistic robustness, and\nwhether there is a certain relationship between $\\kappa$ and this bound. This\nwork studies these problems from a Bayes error perspective. We find that while\nBayes uncertainty does affect probabilistic robustness, its impact is smaller\nthan that on deterministic robustness. This reduced Bayes uncertainty allows a\nhigher upper bound on probabilistic robust accuracy than that on deterministic\nrobust accuracy. Further, we prove that with optimal probabilistic robustness,\neach probabilistically robust input is also deterministically robust in a\nsmaller vicinity. We also show that voting within the vicinity always improves\nprobabilistic robust accuracy and the upper bound of probabilistic robust\naccuracy monotonically increases as $\\kappa$ grows. Our empirical findings also\nalign with our results.\n", "date": "2024-05-23", "categories": []}
{"id": "2405.14934", "title": "Universal Robustness via Median Randomized Smoothing for Real-World\n  Super-Resolution", "abstract": "  Most of the recent literature on image Super-Resolution (SR) can be\nclassified into two main approaches. The first one involves learning a\ncorruption model tailored to a specific dataset, aiming to mimic the noise and\ncorruption in low-resolution images, such as sensor noise. However, this\napproach is data-specific, tends to lack adaptability, and its accuracy\ndiminishes when faced with unseen types of image corruptions. A second and more\nrecent approach, referred to as Robust Super-Resolution (RSR), proposes to\nimprove real-world SR by harnessing the generalization capabilities of a model\nby making it robust to adversarial attacks. To delve further into this second\napproach, our paper explores the universality of various methods for enhancing\nthe robustness of deep learning SR models. In other words, we inquire: \"Which\nrobustness method exhibits the highest degree of adaptability when dealing with\na wide range of adversarial attacks ?\". Our extensive experimentation on both\nsynthetic and real-world images empirically demonstrates that median randomized\nsmoothing (MRS) is more general in terms of robustness compared to adversarial\nlearning techniques, which tend to focus on specific types of attacks.\nFurthermore, as expected, we also illustrate that the proposed universal robust\nmethod enables the SR model to handle standard corruptions more effectively,\nsuch as blur and Gaussian noise, and notably, corruptions naturally present in\nreal-world images. These results support the significance of shifting the\nparadigm in the development of real-world SR methods towards RSR, especially\nvia MRS.\n", "date": "2024-05-23", "categories": []}
{"id": "2401.15335", "title": "L-AutoDA: Leveraging Large Language Models for Automated Decision-based\n  Adversarial Attacks", "abstract": "  In the rapidly evolving field of machine learning, adversarial attacks\npresent a significant challenge to model robustness and security.\nDecision-based attacks, which only require feedback on the decision of a model\nrather than detailed probabilities or scores, are particularly insidious and\ndifficult to defend against. This work introduces L-AutoDA (Large Language\nModel-based Automated Decision-based Adversarial Attacks), a novel approach\nleveraging the generative capabilities of Large Language Models (LLMs) to\nautomate the design of these attacks. By iteratively interacting with LLMs in\nan evolutionary framework, L-AutoDA automatically designs competitive attack\nalgorithms efficiently without much human effort. We demonstrate the efficacy\nof L-AutoDA on CIFAR-10 dataset, showing significant improvements over baseline\nmethods in both success rate and computational efficiency. Our findings\nunderscore the potential of language models as tools for adversarial attack\ngeneration and highlight new avenues for the development of robust AI systems.\n", "date": "2024-01-27", "categories": []}
{"id": "2402.04699", "title": "Breaking Free: How to Hack Safety Guardrails in Black-Box Diffusion\n  Models!", "abstract": "  Deep neural networks can be exploited using natural adversarial samples,\nwhich do not impact human perception. Current approaches often rely on deep\nneural networks' white-box nature to generate these adversarial samples or\nsynthetically alter the distribution of adversarial samples compared to the\ntraining distribution. In contrast, we propose EvoSeed, a novel evolutionary\nstrategy-based algorithmic framework for generating photo-realistic natural\nadversarial samples. Our EvoSeed framework uses auxiliary Conditional Diffusion\nand Classifier models to operate in a black-box setting. We employ CMA-ES to\noptimize the search for an initial seed vector, which, when processed by the\nConditional Diffusion Model, results in the natural adversarial sample\nmisclassified by the Classifier Model. Experiments show that generated\nadversarial images are of high image quality, raising concerns about generating\nharmful content bypassing safety classifiers. Our research opens new avenues to\nunderstanding the limitations of current safety mechanisms and the risk of\nplausible attacks against classifier systems using image generation. Project\nWebsite can be accessed at: https://shashankkotyan.github.io/EvoSeed.\n", "date": "2024-02-07", "categories": []}
{"id": "2405.13324", "title": "Adversarial Training via Adaptive Knowledge Amalgamation of an Ensemble\n  of Teachers", "abstract": "  Adversarial training (AT) is a popular method for training robust deep neural\nnetworks (DNNs) against adversarial attacks. Yet, AT suffers from two\nshortcomings: (i) the robustness of DNNs trained by AT is highly intertwined\nwith the size of the DNNs, posing challenges in achieving robustness in smaller\nmodels; and (ii) the adversarial samples employed during the AT process exhibit\npoor generalization, leaving DNNs vulnerable to unforeseen attack types. To\naddress these dual challenges, this paper introduces adversarial training via\nadaptive knowledge amalgamation of an ensemble of teachers (AT-AKA). In\nparticular, we generate a diverse set of adversarial samples as the inputs to\nan ensemble of teachers; and then, we adaptively amalgamate the logtis of these\nteachers to train a generalized-robust student. Through comprehensive\nexperiments, we illustrate the superior efficacy of AT-AKA over existing AT\nmethods and adversarial robustness distillation techniques against cutting-edge\nattacks, including AutoAttack.\n", "date": "2024-05-22", "categories": []}
{"id": "2405.14169", "title": "Towards Transferable Attacks Against Vision-LLMs in Autonomous Driving\n  with Typography", "abstract": "  Vision-Large-Language-Models (Vision-LLMs) are increasingly being integrated\ninto autonomous driving (AD) systems due to their advanced visual-language\nreasoning capabilities, targeting the perception, prediction, planning, and\ncontrol mechanisms. However, Vision-LLMs have demonstrated susceptibilities\nagainst various types of adversarial attacks, which would compromise their\nreliability and safety. To further explore the risk in AD systems and the\ntransferability of practical threats, we propose to leverage typographic\nattacks against AD systems relying on the decision-making capabilities of\nVision-LLMs. Different from the few existing works developing general datasets\nof typographic attacks, this paper focuses on realistic traffic scenarios where\nthese attacks can be deployed, on their potential effects on the\ndecision-making autonomy, and on the practical ways in which these attacks can\nbe physically presented. To achieve the above goals, we first propose a\ndataset-agnostic framework for automatically generating false answers that can\nmislead Vision-LLMs' reasoning. Then, we present a linguistic augmentation\nscheme that facilitates attacks at image-level and region-level reasoning, and\nwe extend it with attack patterns against multiple reasoning tasks\nsimultaneously. Based on these, we conduct a study on how these attacks can be\nrealized in physical traffic scenarios. Through our empirical study, we\nevaluate the effectiveness, transferability, and realizability of typographic\nattacks in traffic scenes. Our findings demonstrate particular harmfulness of\nthe typographic attacks against existing Vision-LLMs (e.g., LLaVA, Qwen-VL,\nVILA, and Imp), thereby raising community awareness of vulnerabilities when\nincorporating such models into AD systems. We will release our source code upon\nacceptance.\n", "date": "2024-05-23", "categories": []}
{"id": "2405.14176", "title": "Certified Robustness against Sparse Adversarial Perturbations via Data\n  Localization", "abstract": "  Recent work in adversarial robustness suggests that natural data\ndistributions are localized, i.e., they place high probability in small volume\nregions of the input space, and that this property can be utilized for\ndesigning classifiers with improved robustness guarantees for $\\ell_2$-bounded\nperturbations. Yet, it is still unclear if this observation holds true for more\ngeneral metrics. In this work, we extend this theory to $\\ell_0$-bounded\nadversarial perturbations, where the attacker can modify a few pixels of the\nimage but is unrestricted in the magnitude of perturbation, and we show\nnecessary and sufficient conditions for the existence of $\\ell_0$-robust\nclassifiers. Theoretical certification approaches in this regime essentially\nemploy voting over a large ensemble of classifiers. Such procedures are\ncombinatorial and expensive or require complicated certification techniques. In\ncontrast, a simple classifier emerges from our theory, dubbed Box-NN, which\nnaturally incorporates the geometry of the problem and improves upon the\ncurrent state-of-the-art in certified robustness against sparse attacks for the\nMNIST and Fashion-MNIST datasets.\n", "date": "2024-05-23", "categories": []}
{"id": "2305.15241", "title": "Robust Classification via a Single Diffusion Model", "abstract": "  Diffusion models have been applied to improve adversarial robustness of image\nclassifiers by purifying the adversarial noises or generating realistic data\nfor adversarial training. However, diffusion-based purification can be evaded\nby stronger adaptive attacks while adversarial training does not perform well\nunder unseen threats, exhibiting inevitable limitations of these methods. To\nbetter harness the expressive power of diffusion models, this paper proposes\nRobust Diffusion Classifier (RDC), a generative classifier that is constructed\nfrom a pre-trained diffusion model to be adversarially robust. RDC first\nmaximizes the data likelihood of a given input and then predicts the class\nprobabilities of the optimized input using the conditional likelihood estimated\nby the diffusion model through Bayes' theorem. To further reduce the\ncomputational cost, we propose a new diffusion backbone called multi-head\ndiffusion and develop efficient sampling strategies. As RDC does not require\ntraining on particular adversarial attacks, we demonstrate that it is more\ngeneralizable to defend against multiple unseen threats. In particular, RDC\nachieves $75.67\\%$ robust accuracy against various $\\ell_\\infty$ norm-bounded\nadaptive attacks with $\\epsilon_\\infty=8/255$ on CIFAR-10, surpassing the\nprevious state-of-the-art adversarial training models by $+4.77\\%$. The results\nhighlight the potential of generative classifiers by employing pre-trained\ndiffusion models for adversarial robustness compared with the commonly studied\ndiscriminative classifiers. Code is available at\n\\url{https://github.com/huanranchen/DiffusionClassifier}.\n", "date": "2023-05-24", "categories": []}
{"id": "2312.01260", "title": "Rethinking PGD Attack: Is Sign Function Necessary?", "abstract": "  Neural networks have demonstrated success in various domains, yet their\nperformance can be significantly degraded by even a small input perturbation.\nConsequently, the construction of such perturbations, known as adversarial\nattacks, has gained significant attention, many of which fall within\n\"white-box\" scenarios where we have full access to the neural network. Existing\nattack algorithms, such as the projected gradient descent (PGD), commonly take\nthe sign function on the raw gradient before updating adversarial inputs,\nthereby neglecting gradient magnitude information. In this paper, we present a\ntheoretical analysis of how such sign-based update algorithm influences\nstep-wise attack performance, as well as its caveat. We also interpret why\nprevious attempts of directly using raw gradients failed. Based on that, we\nfurther propose a new raw gradient descent (RGD) algorithm that eliminates the\nuse of sign. Specifically, we convert the constrained optimization problem into\nan unconstrained one, by introducing a new hidden variable of non-clipped\nperturbation that can move beyond the constraint. The effectiveness of the\nproposed RGD algorithm has been demonstrated extensively in experiments,\noutperforming PGD and other competitors in various settings, without incurring\nany additional computational overhead. The codes is available in\nhttps://github.com/JunjieYang97/RGD.\n", "date": "2023-12-03", "categories": []}
{"id": "2401.10313", "title": "Hacking Predictors Means Hacking Cars: Using Sensitivity Analysis to\n  Identify Trajectory Prediction Vulnerabilities for Autonomous Driving\n  Security", "abstract": "  Adversarial attacks on learning-based multi-modal trajectory predictors have\nalready been demonstrated. However, there are still open questions about the\neffects of perturbations on inputs other than state histories, and how these\nattacks impact downstream planning and control. In this paper, we conduct a\nsensitivity analysis on two trajectory prediction models, Trajectron++ and\nAgentFormer. The analysis reveals that between all inputs, almost all of the\nperturbation sensitivities for both models lie only within the most recent\nposition and velocity states. We additionally demonstrate that, despite\ndominant sensitivity on state history perturbations, an undetectable image map\nperturbation made with the Fast Gradient Sign Method can induce large\nprediction error increases in both models, revealing that these trajectory\npredictors are, in fact, susceptible to image-based attacks. Using an\noptimization-based planner and example perturbations crafted from sensitivity\nresults, we show how these attacks can cause a vehicle to come to a sudden stop\nfrom moderate driving speeds.\n", "date": "2024-01-18", "categories": []}
{"id": "2405.12266", "title": "EGAN: Evolutional GAN for Ransomware Evasion", "abstract": "  Adversarial Training is a proven defense strategy against adversarial\nmalware. However, generating adversarial malware samples for this type of\ntraining presents a challenge because the resulting adversarial malware needs\nto remain evasive and functional. This work proposes an attack framework, EGAN,\nto address this limitation. EGAN leverages an Evolution Strategy and Generative\nAdversarial Network to select a sequence of attack actions that can mutate a\nRansomware file while preserving its original functionality. We tested this\nframework on popular AI-powered commercial antivirus systems listed on\nVirusTotal and demonstrated that our framework is capable of bypassing the\nmajority of these systems. Moreover, we evaluated whether the EGAN attack\nframework can evade other commercial non-AI antivirus solutions. Our results\nindicate that the adversarial ransomware generated can increase the probability\nof evading some of them.\n", "date": "2024-05-20", "categories": []}
{"id": "2212.02457", "title": "Blessings and Curses of Covariate Shifts: Adversarial Learning Dynamics,\n  Directional Convergence, and Equilibria", "abstract": "  Covariate distribution shifts and adversarial perturbations present\nrobustness challenges to the conventional statistical learning framework: mild\nshifts in the test covariate distribution can significantly affect the\nperformance of the statistical model learned based on the training\ndistribution. The model performance typically deteriorates when extrapolation\nhappens: namely, covariates shift to a region where the training distribution\nis scarce, and naturally, the learned model has little information. For\nrobustness and regularization considerations, adversarial perturbation\ntechniques are proposed as a remedy; however, careful study needs to be carried\nout about what extrapolation region adversarial covariate shift will focus on,\ngiven a learned model. This paper precisely characterizes the extrapolation\nregion, examining both regression and classification in an infinite-dimensional\nsetting. We study the implications of adversarial covariate shifts to\nsubsequent learning of the equilibrium -- the Bayes optimal model -- in a\nsequential game framework. We exploit the dynamics of the adversarial learning\ngame and reveal the curious effects of the covariate shift to equilibrium\nlearning and experimental design. In particular, we establish two directional\nconvergence results that exhibit distinctive phenomena: (1) a blessing in\nregression, the adversarial covariate shifts in an exponential rate to an\noptimal experimental design for rapid subsequent learning; (2) a curse in\nclassification, the adversarial covariate shifts in a subquadratic rate to the\nhardest experimental design trapping subsequent learning.\n", "date": "2022-12-05", "categories": []}
{"id": "2301.09069", "title": "Provable Unrestricted Adversarial Training without Compromise with\n  Generalizability", "abstract": "  Adversarial training (AT) is widely considered as the most promising strategy\nto defend against adversarial attacks and has drawn increasing interest from\nresearchers. However, the existing AT methods still suffer from two challenges.\nFirst, they are unable to handle unrestricted adversarial examples (UAEs),\nwhich are built from scratch, as opposed to restricted adversarial examples\n(RAEs), which are created by adding perturbations bound by an $l_p$ norm to\nobserved examples. Second, the existing AT methods often achieve adversarial\nrobustness at the expense of standard generalizability (i.e., the accuracy on\nnatural examples) because they make a tradeoff between them. To overcome these\nchallenges, we propose a unique viewpoint that understands UAEs as\nimperceptibly perturbed unobserved examples. Also, we find that the tradeoff\nresults from the separation of the distributions of adversarial examples and\nnatural examples. Based on these ideas, we propose a novel AT approach called\nProvable Unrestricted Adversarial Training (PUAT), which can provide a target\nclassifier with comprehensive adversarial robustness against both UAE and RAE,\nand simultaneously improve its standard generalizability. Particularly, PUAT\nutilizes partially labeled data to achieve effective UAE generation by\naccurately capturing the natural data distribution through a novel augmented\ntriple-GAN. At the same time, PUAT extends the traditional AT by introducing\nthe supervised loss of the target classifier into the adversarial loss and\nachieves the alignment between the UAE distribution, the natural data\ndistribution, and the distribution learned by the classifier, with the\ncollaboration of the augmented triple-GAN. Finally, the solid theoretical\nanalysis and extensive experiments conducted on widely-used benchmarks\ndemonstrate the superiority of PUAT.\n", "date": "2023-01-22", "categories": []}
{"id": "2301.11824", "title": "PECAN: A Deterministic Certified Defense Against Backdoor Attacks", "abstract": "  Neural networks are vulnerable to backdoor poisoning attacks, where the\nattackers maliciously poison the training set and insert triggers into the test\ninput to change the prediction of the victim model. Existing defenses for\nbackdoor attacks either provide no formal guarantees or come with\nexpensive-to-compute and ineffective probabilistic guarantees. We present\nPECAN, an efficient and certified approach for defending against backdoor\nattacks. The key insight powering PECAN is to apply off-the-shelf test-time\nevasion certification techniques on a set of neural networks trained on\ndisjoint partitions of the data. We evaluate PECAN on image classification and\nmalware detection datasets. Our results demonstrate that PECAN can (1)\nsignificantly outperform the state-of-the-art certified backdoor defense, both\nin defense strength and efficiency, and (2) on real back-door attacks, PECAN\ncan reduce attack success rate by order of magnitude when compared to a range\nof baselines from the literature.\n", "date": "2023-01-27", "categories": []}
{"id": "2404.08273", "title": "Struggle with Adversarial Defense? Try Diffusion", "abstract": "  Adversarial attacks induce misclassification by introducing subtle\nperturbations. Recently, diffusion models are applied to the image classifiers\nto improve adversarial robustness through adversarial training or by purifying\nadversarial noise. However, diffusion-based adversarial training often\nencounters convergence challenges and high computational expenses.\nAdditionally, diffusion-based purification inevitably causes data shift and is\ndeemed susceptible to stronger adaptive attacks. To tackle these issues, we\npropose the Truth Maximization Diffusion Classifier (TMDC), a generative\nBayesian classifier that builds upon pre-trained diffusion models and the\nBayesian theorem. Unlike data-driven classifiers, TMDC, guided by Bayesian\nprinciples, utilizes the conditional likelihood from diffusion models to\ndetermine the class probabilities of input images, thereby insulating against\nthe influences of data shift and the limitations of adversarial training.\nMoreover, to enhance TMDC's resilience against more potent adversarial attacks,\nwe propose an optimization strategy for diffusion classifiers. This strategy\ninvolves post-training the diffusion model on perturbed datasets with\nground-truth labels as conditions, guiding the diffusion model to learn the\ndata distribution and maximizing the likelihood under the ground-truth labels.\nThe proposed method achieves state-of-the-art performance on the CIFAR10\ndataset against heavy white-box attacks and strong adaptive attacks.\nSpecifically, TMDC achieves robust accuracies of 82.81% against $l_{\\infty}$\nnorm-bounded perturbations and 86.05% against $l_{2}$ norm-bounded\nperturbations, respectively, with $\\epsilon=0.05$.\n", "date": "2024-04-12", "categories": []}
{"id": "2405.10970", "title": "Untargeted Adversarial Attack on Knowledge Graph Embeddings", "abstract": "  Knowledge graph embedding (KGE) methods have achieved great success in\nhandling various knowledge graph (KG) downstream tasks. However, KGE methods\nmay learn biased representations on low-quality KGs that are prevalent in the\nreal world. Some recent studies propose adversarial attacks to investigate the\nvulnerabilities of KGE methods, but their attackers are target-oriented with\nthe KGE method and the target triples to predict are given in advance, which\nlacks practicability. In this work, we explore untargeted attacks with the aim\nof reducing the global performances of KGE methods over a set of unknown test\ntriples and conducting systematic analyses on KGE robustness. Considering logic\nrules can effectively summarize the global structure of a KG, we develop\nrule-based attack strategies to enhance the attack efficiency. In particular,we\nconsider adversarial deletion which learns rules, applying the rules to score\ntriple importance and delete important triples, and adversarial addition which\ncorrupts the learned rules and applies them for negative triples as\nperturbations. Extensive experiments on two datasets over three representative\nclasses of KGE methods demonstrate the effectiveness of our proposed untargeted\nattacks in diminishing the link prediction results. And we also find that\ndifferent KGE methods exhibit different robustness to untargeted attacks. For\nexample, the robustness of methods engaged with graph neural networks and logic\nrules depends on the density of the graph. But rule-based methods like NCRL are\neasily affected by adversarial addition attacks to capture negative rules\n", "date": "2024-05-08", "categories": []}
{"id": "2405.11154", "title": "Revisiting the Robust Generalization of Adversarial Prompt Tuning", "abstract": "  Understanding the vulnerability of large-scale pre-trained vision-language\nmodels like CLIP against adversarial attacks is key to ensuring zero-shot\ngeneralization capacity on various downstream tasks. State-of-the-art defense\nmechanisms generally adopt prompt learning strategies for adversarial\nfine-tuning to improve the adversarial robustness of the pre-trained model\nwhile keeping the efficiency of adapting to downstream tasks. Such a setup\nleads to the problem of over-fitting which impedes further improvement of the\nmodel's generalization capacity on both clean and adversarial examples. In this\nwork, we propose an adaptive Consistency-guided Adversarial Prompt Tuning\n(i.e., CAPT) framework that utilizes multi-modal prompt learning to enhance the\nalignment of image and text features for adversarial examples and leverage the\nstrong generalization of pre-trained CLIP to guide the model-enhancing its\nrobust generalization on adversarial examples while maintaining its accuracy on\nclean ones. We also design a novel adaptive consistency objective function to\nbalance the consistency of adversarial inputs and clean inputs between the\nfine-tuning model and the pre-trained model. We conduct extensive experiments\nacross 14 datasets and 4 data sparsity schemes (from 1-shot to full training\ndata settings) to show the superiority of CAPT over other state-of-the-art\nadaption methods. CAPT demonstrated excellent performance in terms of the\nin-distribution performance and the generalization under input distribution\nshift and across datasets.\n", "date": "2024-05-18", "categories": []}
{"id": "2405.11551", "title": "An Invisible Backdoor Attack Based On Semantic Feature", "abstract": "  Backdoor attacks have severely threatened deep neural network (DNN) models in\nthe past several years. These attacks can occur in almost every stage of the\ndeep learning pipeline. Although the attacked model behaves normally on benign\nsamples, it makes wrong predictions for samples containing triggers. However,\nmost existing attacks use visible patterns (e.g., a patch or image\ntransformations) as triggers, which are vulnerable to human inspection. In this\npaper, we propose a novel backdoor attack, making imperceptible changes.\nConcretely, our attack first utilizes the pre-trained victim model to extract\nlow-level and high-level semantic features from clean images and generates\ntrigger pattern associated with high-level features based on channel attention.\nThen, the encoder model generates poisoned images based on the trigger and\nextracted low-level semantic features without causing noticeable feature loss.\nWe evaluate our attack on three prominent image classification DNN across three\nstandard datasets. The results demonstrate that our attack achieves high attack\nsuccess rates while maintaining robustness against backdoor defenses.\nFurthermore, we conduct extensive image similarity experiments to emphasize the\nstealthiness of our attack strategy.\n", "date": "2024-05-19", "categories": []}
{"id": "2405.11575", "title": "SEEP: Training Dynamics Grounds Latent Representation Search for\n  Mitigating Backdoor Poisoning Attacks", "abstract": "  Modern NLP models are often trained on public datasets drawn from diverse\nsources, rendering them vulnerable to data poisoning attacks. These attacks can\nmanipulate the model's behavior in ways engineered by the attacker. One such\ntactic involves the implantation of backdoors, achieved by poisoning specific\ntraining instances with a textual trigger and a target class label. Several\nstrategies have been proposed to mitigate the risks associated with backdoor\nattacks by identifying and removing suspected poisoned examples. However, we\nobserve that these strategies fail to offer effective protection against\nseveral advanced backdoor attacks. To remedy this deficiency, we propose a\nnovel defensive mechanism that first exploits training dynamics to identify\npoisoned samples with high precision, followed by a label propagation step to\nimprove recall and thus remove the majority of poisoned instances. Compared\nwith recent advanced defense methods, our method considerably reduces the\nsuccess rates of several backdoor attacks while maintaining high classification\naccuracy on clean test sets.\n", "date": "2024-05-19", "categories": []}
{"id": "2405.11629", "title": "Searching Realistic-Looking Adversarial Objects For Autonomous Driving\n  Systems", "abstract": "  Numerous studies on adversarial attacks targeting self-driving policies fail\nto incorporate realistic-looking adversarial objects, limiting real-world\napplicability. Building upon prior research that facilitated the transition of\nadversarial objects from simulations to practical applications, this paper\ndiscusses a modified gradient-based texture optimization method to discover\nrealistic-looking adversarial objects. While retaining the core architecture\nand techniques of the prior research, the proposed addition involves an entity\ntermed the 'Judge'. This agent assesses the texture of a rendered object,\nassigning a probability score reflecting its realism. This score is integrated\ninto the loss function to encourage the NeRF object renderer to concurrently\nlearn realistic and adversarial textures. The paper analyzes four strategies\nfor developing a robust 'Judge': 1) Leveraging cutting-edge vision-language\nmodels. 2) Fine-tuning open-sourced vision-language models. 3) Pretraining\nneurosymbolic systems. 4) Utilizing traditional image processing techniques.\nOur findings indicate that strategies 1) and 4) yield less reliable outcomes,\npointing towards strategies 2) or 3) as more promising directions for future\nresearch.\n", "date": "2024-05-19", "categories": []}
{"id": "2405.11904", "title": "A Constraint-Enforcing Reward for Adversarial Attacks on Text\n  Classifiers", "abstract": "  Text classifiers are vulnerable to adversarial examples --\ncorrectly-classified examples that are deliberately transformed to be\nmisclassified while satisfying acceptability constraints. The conventional\napproach to finding adversarial examples is to define and solve a combinatorial\noptimisation problem over a space of allowable transformations. While\neffective, this approach is slow and limited by the choice of transformations.\nAn alternate approach is to directly generate adversarial examples by\nfine-tuning a pre-trained language model, as is commonly done for other\ntext-to-text tasks. This approach promises to be much quicker and more\nexpressive, but is relatively unexplored. For this reason, in this work we\ntrain an encoder-decoder paraphrase model to generate a diverse range of\nadversarial examples. For training, we adopt a reinforcement learning algorithm\nand propose a constraint-enforcing reward that promotes the generation of valid\nadversarial examples. Experimental results over two text classification\ndatasets show that our model has achieved a higher success rate than the\noriginal paraphrase model, and overall has proved more effective than other\ncompetitive attacks. Finally, we show how key design choices impact the\ngenerated examples and discuss the strengths and weaknesses of the proposed\napproach.\n", "date": "2024-05-20", "categories": []}
{"id": "2312.12321", "title": "Bypassing the Safety Training of Open-Source LLMs with Priming Attacks", "abstract": "  With the recent surge in popularity of LLMs has come an ever-increasing need\nfor LLM safety training. In this paper, we investigate the fragility of SOTA\nopen-source LLMs under simple, optimization-free attacks we refer to as\n$\\textit{priming attacks}$, which are easy to execute and effectively bypass\nalignment from safety training. Our proposed attack improves the Attack Success\nRate on Harmful Behaviors, as measured by Llama Guard, by up to $3.3\\times$\ncompared to baselines. Source code and data are available at\nhttps://github.com/uiuc-focal-lab/llm-priming-attacks.\n", "date": "2023-12-19", "categories": []}
{"id": "2403.06668", "title": "PeerAiD: Improving Adversarial Distillation from a Specialized Peer\n  Tutor", "abstract": "  Adversarial robustness of the neural network is a significant concern when it\nis applied to security-critical domains. In this situation, adversarial\ndistillation is a promising option which aims to distill the robustness of the\nteacher network to improve the robustness of a small student network. Previous\nworks pretrain the teacher network to make it robust against the adversarial\nexamples aimed at itself. However, the adversarial examples are dependent on\nthe parameters of the target network. The fixed teacher network inevitably\ndegrades its robustness against the unseen transferred adversarial examples\nwhich target the parameters of the student network in the adversarial\ndistillation process. We propose PeerAiD to make a peer network learn the\nadversarial examples of the student network instead of adversarial examples\naimed at itself. PeerAiD is an adversarial distillation that trains the peer\nnetwork and the student network simultaneously in order to specialize the peer\nnetwork for defending the student network. We observe that such peer networks\nsurpass the robustness of the pretrained robust teacher model against\nadversarial examples aimed at the student network. With this peer network and\nadversarial distillation, PeerAiD achieves significantly higher robustness of\nthe student network with AutoAttack (AA) accuracy by up to 1.66%p and improves\nthe natural accuracy of the student network by up to 4.72%p with ResNet-18 on\nTinyImageNet dataset. Code is available at\nhttps://github.com/jaewonalive/PeerAiD.\n", "date": "2024-03-11", "categories": []}
{"id": "2405.09598", "title": "Properties that allow or prohibit transferability of adversarial attacks\n  among quantized networks", "abstract": "  Deep Neural Networks (DNNs) are known to be vulnerable to adversarial\nexamples. Further, these adversarial examples are found to be transferable from\nthe source network in which they are crafted to a black-box target network. As\nthe trend of using deep learning on embedded devices grows, it becomes relevant\nto study the transferability properties of adversarial examples among\ncompressed networks. In this paper, we consider quantization as a network\ncompression technique and evaluate the performance of transfer-based attacks\nwhen the source and target networks are quantized at different bitwidths. We\nexplore how algorithm specific properties affect transferability by considering\nvarious adversarial example generation algorithms. Furthermore, we examine\ntransferability in a more realistic scenario where the source and target\nnetworks may differ in bitwidth and other model-related properties like\ncapacity and architecture. We find that although quantization reduces\ntransferability, certain attack types demonstrate an ability to enhance it.\nAdditionally, the average transferability of adversarial examples among\nquantized versions of a network can be used to estimate the transferability to\nquantized target networks with varying capacity and architecture.\n", "date": "2024-05-15", "categories": []}
{"id": "2405.09924", "title": "Infrared Adversarial Car Stickers", "abstract": "  Infrared physical adversarial examples are of great significance for studying\nthe security of infrared AI systems that are widely used in our lives such as\nautonomous driving. Previous infrared physical attacks mainly focused on 2D\ninfrared pedestrian detection which may not fully manifest its destructiveness\nto AI systems. In this work, we propose a physical attack method against\ninfrared detectors based on 3D modeling, which is applied to a real car. The\ngoal is to design a set of infrared adversarial stickers to make cars invisible\nto infrared detectors at various viewing angles, distances, and scenes. We\nbuild a 3D infrared car model with real infrared characteristics and propose an\ninfrared adversarial pattern generation method based on 3D mesh shadow. We\npropose a 3D control points-based mesh smoothing algorithm and use a set of\nsmoothness loss functions to enhance the smoothness of adversarial meshes and\nfacilitate the sticker implementation. Besides, We designed the aluminum\nstickers and conducted physical experiments on two real Mercedes-Benz A200L\ncars. Our adversarial stickers hid the cars from Faster RCNN, an object\ndetector, at various viewing angles, distances, and scenes. The attack success\nrate (ASR) was 91.49% for real cars. In comparison, the ASRs of random stickers\nand no sticker were only 6.21% and 0.66%, respectively. In addition, the ASRs\nof the designed stickers against six unseen object detectors such as YOLOv3 and\nDeformable DETR were between 73.35%-95.80%, showing good transferability of the\nattack performance across detectors.\n", "date": "2024-05-16", "categories": []}
{"id": "2405.09981", "title": "Adversarial Robustness for Visual Grounding of Multimodal Large Language\n  Models", "abstract": "  Multi-modal Large Language Models (MLLMs) have recently achieved enhanced\nperformance across various vision-language tasks including visual grounding\ncapabilities. However, the adversarial robustness of visual grounding remains\nunexplored in MLLMs. To fill this gap, we use referring expression\ncomprehension (REC) as an example task in visual grounding and propose three\nadversarial attack paradigms as follows. Firstly, untargeted adversarial\nattacks induce MLLMs to generate incorrect bounding boxes for each object.\nBesides, exclusive targeted adversarial attacks cause all generated outputs to\nthe same target bounding box. In addition, permuted targeted adversarial\nattacks aim to permute all bounding boxes among different objects within a\nsingle image. Extensive experiments demonstrate that the proposed methods can\nsuccessfully attack visual grounding capabilities of MLLMs. Our methods not\nonly provide a new perspective for designing novel attacks but also serve as a\nstrong baseline for improving the adversarial robustness for visual grounding\nof MLLMs.\n", "date": "2024-05-16", "categories": []}
{"id": "2405.08834", "title": "Adversarial Machine Learning Threats to Spacecraft", "abstract": "  Spacecraft are among the earliest autonomous systems. Their ability to\nfunction without a human in the loop have afforded some of humanity's grandest\nachievements. As reliance on autonomy grows, space vehicles will become\nincreasingly vulnerable to attacks designed to disrupt autonomous\nprocesses-especially probabilistic ones based on machine learning. This paper\naims to elucidate and demonstrate the threats that adversarial machine learning\n(AML) capabilities pose to spacecraft. First, an AML threat taxonomy for\nspacecraft is introduced. Next, we demonstrate the execution of AML attacks\nagainst spacecraft through experimental simulations using NASA's Core Flight\nSystem (cFS) and NASA's On-board Artificial Intelligence Research (OnAIR)\nPlatform. Our findings highlight the imperative for incorporating AML-focused\nsecurity measures in spacecraft that engage autonomy.\n", "date": "2024-05-14", "categories": []}
{"id": "2405.08886", "title": "The Pitfalls and Promise of Conformal Inference Under Adversarial\n  Attacks", "abstract": "  In safety-critical applications such as medical imaging and autonomous\ndriving, where decisions have profound implications for patient health and road\nsafety, it is imperative to maintain both high adversarial robustness to\nprotect against potential adversarial attacks and reliable uncertainty\nquantification in decision-making. With extensive research focused on enhancing\nadversarial robustness through various forms of adversarial training (AT), a\nnotable knowledge gap remains concerning the uncertainty inherent in\nadversarially trained models. To address this gap, this study investigates the\nuncertainty of deep learning models by examining the performance of conformal\nprediction (CP) in the context of standard adversarial attacks within the\nadversarial defense community. It is first unveiled that existing CP methods do\nnot produce informative prediction sets under the commonly used\n$l_{\\infty}$-norm bounded attack if the model is not adversarially trained,\nwhich underpins the importance of adversarial training for CP. Our paper next\ndemonstrates that the prediction set size (PSS) of CP using adversarially\ntrained models with AT variants is often worse than using standard AT,\ninspiring us to research into CP-efficient AT for improved PSS. We propose to\noptimize a Beta-weighting loss with an entropy minimization regularizer during\nAT to improve CP-efficiency, where the Beta-weighting loss is shown to be an\nupper bound of PSS at the population level by our theoretical analysis.\nMoreover, our empirical study on four image classification datasets across\nthree popular AT baselines validates the effectiveness of the proposed\nUncertainty-Reducing AT (AT-UR).\n", "date": "2024-05-14", "categories": []}
{"id": "2405.09470", "title": "Towards Evaluating the Robustness of Automatic Speech Recognition\n  Systems via Audio Style Transfer", "abstract": "  In light of the widespread application of Automatic Speech Recognition (ASR)\nsystems, their security concerns have received much more attention than ever\nbefore, primarily due to the susceptibility of Deep Neural Networks. Previous\nstudies have illustrated that surreptitiously crafting adversarial\nperturbations enables the manipulation of speech recognition systems, resulting\nin the production of malicious commands. These attack methods mostly require\nadding noise perturbations under $\\ell_p$ norm constraints, inevitably leaving\nbehind artifacts of manual modifications. Recent research has alleviated this\nlimitation by manipulating style vectors to synthesize adversarial examples\nbased on Text-to-Speech (TTS) synthesis audio. However, style modifications\nbased on optimization objectives significantly reduce the controllability and\neditability of audio styles. In this paper, we propose an attack on ASR systems\nbased on user-customized style transfer. We first test the effect of Style\nTransfer Attack (STA) which combines style transfer and adversarial attack in\nsequential order. And then, as an improvement, we propose an iterative Style\nCode Attack (SCA) to maintain audio quality. Experimental results show that our\nmethod can meet the need for user-customized styles and achieve a success rate\nof 82% in attacks, while keeping sound naturalness due to our user study.\n", "date": "2024-05-15", "categories": []}
{"id": "2311.12981", "title": "SD-NAE: Generating Natural Adversarial Examples with Stable Diffusion", "abstract": "  Natural Adversarial Examples (NAEs), images arising naturally from the\nenvironment and capable of deceiving classifiers, are instrumental in robustly\nevaluating and identifying vulnerabilities in trained models. In this work,\nunlike prior works that passively collect NAEs from real images, we propose to\nactively synthesize NAEs using the state-of-the-art Stable Diffusion.\nSpecifically, our method formulates a controlled optimization process, where we\nperturb the token embedding that corresponds to a specified class to generate\nNAEs. This generation process is guided by the gradient of loss from the target\nclassifier, ensuring that the created image closely mimics the ground-truth\nclass yet fools the classifier. Named SD-NAE (Stable Diffusion for Natural\nAdversarial Examples), our innovative method is effective in producing valid\nand useful NAEs, which is demonstrated through a meticulously designed\nexperiment. Code is available at https://github.com/linyueqian/SD-NAE.\n", "date": "2023-11-21", "categories": []}
{"id": "2405.08317", "title": "SpeechGuard: Exploring the Adversarial Robustness of Multimodal Large\n  Language Models", "abstract": "  Integrated Speech and Large Language Models (SLMs) that can follow speech\ninstructions and generate relevant text responses have gained popularity\nlately. However, the safety and robustness of these models remains largely\nunclear. In this work, we investigate the potential vulnerabilities of such\ninstruction-following speech-language models to adversarial attacks and\njailbreaking. Specifically, we design algorithms that can generate adversarial\nexamples to jailbreak SLMs in both white-box and black-box attack settings\nwithout human involvement. Additionally, we propose countermeasures to thwart\nsuch jailbreaking attacks. Our models, trained on dialog data with speech\ninstructions, achieve state-of-the-art performance on spoken question-answering\ntask, scoring over 80% on both safety and helpfulness metrics. Despite safety\nguardrails, experiments on jailbreaking demonstrate the vulnerability of SLMs\nto adversarial perturbations and transfer attacks, with average attack success\nrates of 90% and 10% respectively when evaluated on a dataset of carefully\ndesigned harmful questions spanning 12 different toxic categories. However, we\ndemonstrate that our proposed countermeasures reduce the attack success\nsignificantly.\n", "date": "2024-05-14", "categories": []}
{"id": "2405.07595", "title": "Environmental Matching Attack Against Unmanned Aerial Vehicles Object\n  Detection", "abstract": "  Object detection techniques for Unmanned Aerial Vehicles (UAVs) rely on Deep\nNeural Networks (DNNs), which are vulnerable to adversarial attacks.\nNonetheless, adversarial patches generated by existing algorithms in the UAV\ndomain pay very little attention to the naturalness of adversarial patches.\nMoreover, imposing constraints directly on adversarial patches makes it\ndifficult to generate patches that appear natural to the human eye while\nensuring a high attack success rate. We notice that patches are natural looking\nwhen their overall color is consistent with the environment. Therefore, we\npropose a new method named Environmental Matching Attack(EMA) to address the\nissue of optimizing the adversarial patch under the constraints of color. To\nthe best of our knowledge, this paper is the first to consider natural patches\nin the domain of UAVs. The EMA method exploits strong prior knowledge of a\npretrained stable diffusion to guide the optimization direction of the\nadversarial patch, where the text guidance can restrict the color of the patch.\nTo better match the environment, the contrast and brightness of the patch are\nappropriately adjusted. Instead of optimizing the adversarial patch itself, we\noptimize an adversarial perturbation patch which initializes to zero so that\nthe model can better trade off attacking performance and naturalness.\nExperiments conducted on the DroneVehicle and Carpk datasets have shown that\nour work can reach nearly the same attack performance in the digital attack(no\ngreater than 2 in mAP$\\%$), surpass the baseline method in the physical\nspecific scenarios, and exhibit a significant advantage in terms of naturalness\nin visualization and color difference with the environment.\n", "date": "2024-05-13", "categories": []}
{"id": "2405.07668", "title": "CrossCert: A Cross-Checking Detection Approach to Patch Robustness\n  Certification for Deep Learning Models", "abstract": "  Patch robustness certification is an emerging kind of defense technique\nagainst adversarial patch attacks with provable guarantees. There are two\nresearch lines: certified recovery and certified detection. They aim to label\nmalicious samples with provable guarantees correctly and issue warnings for\nmalicious samples predicted to non-benign labels with provable guarantees,\nrespectively. However, existing certified detection defenders suffer from\nprotecting labels subject to manipulation, and existing certified recovery\ndefenders cannot systematically warn samples about their labels. A certified\ndefense that simultaneously offers robust labels and systematic warning\nprotection against patch attacks is desirable. This paper proposes a novel\ncertified defense technique called CrossCert. CrossCert formulates a novel\napproach by cross-checking two certified recovery defenders to provide\nunwavering certification and detection certification. Unwavering certification\nensures that a certified sample, when subjected to a patched perturbation, will\nalways be returned with a benign label without triggering any warnings with a\nprovable guarantee. To our knowledge, CrossCert is the first certified\ndetection technique to offer this guarantee. Our experiments show that, with a\nslightly lower performance than ViP and comparable performance with PatchCensor\nin terms of detection certification, CrossCert certifies a significant\nproportion of samples with the guarantee of unwavering certification.\n", "date": "2024-05-13", "categories": []}
{"id": "2405.06049", "title": "BB-Patch: BlackBox Adversarial Patch-Attack using Zeroth-Order\n  Optimization", "abstract": "  Deep Learning has become popular due to its vast applications in almost all\ndomains. However, models trained using deep learning are prone to failure for\nadversarial samples and carry a considerable risk in sensitive applications.\nMost of these adversarial attack strategies assume that the adversary has\naccess to the training data, the model parameters, and the input during\ndeployment, hence, focus on perturbing the pixel level information present in\nthe input image.\n  Adversarial Patches were introduced to the community which helped in bringing\nout the vulnerability of deep learning models in a much more pragmatic manner\nbut here the attacker has a white-box access to the model parameters. Recently,\nthere has been an attempt to develop these adversarial attacks using black-box\ntechniques. However, certain assumptions such as availability large training\ndata is not valid for a real-life scenarios. In a real-life scenario, the\nattacker can only assume the type of model architecture used from a select list\nof state-of-the-art architectures while having access to only a subset of input\ndataset. Hence, we propose an black-box adversarial attack strategy that\nproduces adversarial patches which can be applied anywhere in the input image\nto perform an adversarial attack.\n", "date": "2024-05-09", "categories": []}
{"id": "2405.06247", "title": "Disttack: Graph Adversarial Attacks Toward Distributed GNN Training", "abstract": "  Graph Neural Networks (GNNs) have emerged as potent models for graph\nlearning. Distributing the training process across multiple computing nodes is\nthe most promising solution to address the challenges of ever-growing\nreal-world graphs. However, current adversarial attack methods on GNNs neglect\nthe characteristics and applications of the distributed scenario, leading to\nsuboptimal performance and inefficiency in attacking distributed GNN training.\n  In this study, we introduce Disttack, the first framework of adversarial\nattacks for distributed GNN training that leverages the characteristics of\nfrequent gradient updates in a distributed system. Specifically, Disttack\ncorrupts distributed GNN training by injecting adversarial attacks into one\nsingle computing node. The attacked subgraphs are precisely perturbed to induce\nan abnormal gradient ascent in backpropagation, disrupting gradient\nsynchronization between computing nodes and thus leading to a significant\nperformance decline of the trained GNN. We evaluate Disttack on four large\nreal-world graphs by attacking five widely adopted GNNs. Compared with the\nstate-of-the-art attack method, experimental results demonstrate that Disttack\namplifies the model accuracy degradation by 2.75$\\times$ and achieves speedup\nby 17.33$\\times$ on average while maintaining unnoticeability.\n", "date": "2024-05-10", "categories": []}
{"id": "2405.06278", "title": "Exploring the Interplay of Interpretability and Robustness in Deep\n  Neural Networks: A Saliency-guided Approach", "abstract": "  Adversarial attacks pose a significant challenge to deploying deep learning\nmodels in safety-critical applications. Maintaining model robustness while\nensuring interpretability is vital for fostering trust and comprehension in\nthese models. This study investigates the impact of Saliency-guided Training\n(SGT) on model robustness, a technique aimed at improving the clarity of\nsaliency maps to deepen understanding of the model's decision-making process.\nExperiments were conducted on standard benchmark datasets using various deep\nlearning architectures trained with and without SGT. Findings demonstrate that\nSGT enhances both model robustness and interpretability. Additionally, we\npropose a novel approach combining SGT with standard adversarial training to\nachieve even greater robustness while preserving saliency map quality. Our\nstrategy is grounded in the assumption that preserving salient features crucial\nfor correctly classifying adversarial examples enhances model robustness, while\nmasking non-relevant features improves interpretability. Our technique yields\nsignificant gains, achieving a 35\\% and 20\\% improvement in robustness against\nPGD attack with noise magnitudes of $0.2$ and $0.02$ for the MNIST and CIFAR-10\ndatasets, respectively, while producing high-quality saliency maps.\n", "date": "2024-05-10", "categories": []}
{"id": "2405.06298", "title": "PUMA: margin-based data pruning", "abstract": "  Deep learning has been able to outperform humans in terms of classification\naccuracy in many tasks. However, to achieve robustness to adversarial\nperturbations, the best methodologies require to perform adversarial training\non a much larger training set that has been typically augmented using\ngenerative models (e.g., diffusion models). Our main objective in this work, is\nto reduce these data requirements while achieving the same or better\naccuracy-robustness trade-offs. We focus on data pruning, where some training\nsamples are removed based on the distance to the model classification boundary\n(i.e., margin). We find that the existing approaches that prune samples with\nlow margin fails to increase robustness when we add a lot of synthetic data,\nand explain this situation with a perceptron learning task. Moreover, we find\nthat pruning high margin samples for better accuracy increases the harmful\nimpact of mislabeled perturbed data in adversarial training, hurting both\nrobustness and accuracy. We thus propose PUMA, a new data pruning strategy that\ncomputes the margin using DeepFool, and prunes the training samples of highest\nmargin without hurting performance by jointly adjusting the training attack\nnorm on the samples of lowest margin. We show that PUMA can be used on top of\nthe current state-of-the-art methodology in robustness, and it is able to\nsignificantly improve the model performance unlike the existing data pruning\nstrategies. Not only PUMA achieves similar robustness with less data, but it\nalso significantly increases the model accuracy, improving the performance\ntrade-off.\n", "date": "2024-05-10", "categories": []}
{"id": "2405.06340", "title": "Improving Transferable Targeted Adversarial Attack via Normalized Logit\n  Calibration and Truncated Feature Mixing", "abstract": "  This paper aims to enhance the transferability of adversarial samples in\ntargeted attacks, where attack success rates remain comparatively low. To\nachieve this objective, we propose two distinct techniques for improving the\ntargeted transferability from the loss and feature aspects. First, in previous\napproaches, logit calibrations used in targeted attacks primarily focus on the\nlogit margin between the targeted class and the untargeted classes among\nsamples, neglecting the standard deviation of the logit. In contrast, we\nintroduce a new normalized logit calibration method that jointly considers the\nlogit margin and the standard deviation of logits. This approach effectively\ncalibrates the logits, enhancing the targeted transferability. Second, previous\nstudies have demonstrated that mixing the features of clean samples during\noptimization can significantly increase transferability. Building upon this, we\nfurther investigate a truncated feature mixing method to reduce the impact of\nthe source training model, resulting in additional improvements. The truncated\nfeature is determined by removing the Rank-1 feature associated with the\nlargest singular value decomposed from the high-level convolutional layers of\nthe clean sample. Extensive experiments conducted on the ImageNet-Compatible\nand CIFAR-10 datasets demonstrate the individual and mutual benefits of our\nproposed two components, which outperform the state-of-the-art methods by a\nlarge margin in black-box targeted attacks.\n", "date": "2024-05-10", "categories": []}
{"id": "2405.06345", "title": "Evaluating Adversarial Robustness in the Spatial Frequency Domain", "abstract": "  Convolutional Neural Networks (CNNs) have dominated the majority of computer\nvision tasks. However, CNNs' vulnerability to adversarial attacks has raised\nconcerns about deploying these models to safety-critical applications. In\ncontrast, the Human Visual System (HVS), which utilizes spatial frequency\nchannels to process visual signals, is immune to adversarial attacks. As such,\nthis paper presents an empirical study exploring the vulnerability of CNN\nmodels in the frequency domain. Specifically, we utilize the discrete cosine\ntransform (DCT) to construct the Spatial-Frequency (SF) layer to produce a\nblock-wise frequency spectrum of an input image and formulate Spatial Frequency\nCNNs (SF-CNNs) by replacing the initial feature extraction layers of\nwidely-used CNN backbones with the SF layer. Through extensive experiments, we\nobserve that SF-CNN models are more robust than their CNN counterparts under\nboth white-box and black-box attacks. To further explain the robustness of\nSF-CNNs, we compare the SF layer with a trainable convolutional layer with\nidentical kernel sizes using two mixing strategies to show that the lower\nfrequency components contribute the most to the adversarial robustness of\nSF-CNNs. We believe our observations can guide the future design of robust CNN\nmodels.\n", "date": "2024-05-10", "categories": []}
{"id": "2405.05524", "title": "Universal Adversarial Perturbations for Vision-Language Pre-trained\n  Models", "abstract": "  Vision-language pre-trained (VLP) models have been the foundation of numerous\nvision-language tasks. Given their prevalence, it becomes imperative to assess\ntheir adversarial robustness, especially when deploying them in\nsecurity-crucial real-world applications. Traditionally, adversarial\nperturbations generated for this assessment target specific VLP models,\ndatasets, and/or downstream tasks. This practice suffers from low\ntransferability and additional computation costs when transitioning to new\nscenarios.\n  In this work, we thoroughly investigate whether VLP models are commonly\nsensitive to imperceptible perturbations of a specific pattern for the image\nmodality. To this end, we propose a novel black-box method to generate\nUniversal Adversarial Perturbations (UAPs), which is so called the Effective\nand T ransferable Universal Adversarial Attack (ETU), aiming to mislead a\nvariety of existing VLP models in a range of downstream tasks. The ETU\ncomprehensively takes into account the characteristics of UAPs and the\nintrinsic cross-modal interactions to generate effective UAPs. Under this\nregime, the ETU encourages both global and local utilities of UAPs. This\nbenefits the overall utility while reducing interactions between UAP units,\nimproving the transferability. To further enhance the effectiveness and\ntransferability of UAPs, we also design a novel data augmentation method named\nScMix. ScMix consists of self-mix and cross-mix data transformations, which can\neffectively increase the multi-modal data diversity while preserving the\nsemantics of the original data. Through comprehensive experiments on various\ndownstream tasks, VLP models, and datasets, we demonstrate that the proposed\nmethod is able to achieve effective and transferrable universal adversarial\nattacks.\n", "date": "2024-05-09", "categories": []}
{"id": "2401.01750", "title": "Towards Robust Semantic Segmentation against Patch-based Attack via\n  Attention Refinement", "abstract": "  The attention mechanism has been proven effective on various visual tasks in\nrecent years. In the semantic segmentation task, the attention mechanism is\napplied in various methods, including the case of both Convolution Neural\nNetworks (CNN) and Vision Transformer (ViT) as backbones. However, we observe\nthat the attention mechanism is vulnerable to patch-based adversarial attacks.\nThrough the analysis of the effective receptive field, we attribute it to the\nfact that the wide receptive field brought by global attention may lead to the\nspread of the adversarial patch. To address this issue, in this paper, we\npropose a Robust Attention Mechanism (RAM) to improve the robustness of the\nsemantic segmentation model, which can notably relieve the vulnerability\nagainst patch-based attacks. Compared to the vallina attention mechanism, RAM\nintroduces two novel modules called Max Attention Suppression and Random\nAttention Dropout, both of which aim to refine the attention matrix and limit\nthe influence of a single adversarial patch on the semantic segmentation\nresults of other positions. Extensive experiments demonstrate the effectiveness\nof our RAM to improve the robustness of semantic segmentation models against\nvarious patch-based attack methods under different attack settings.\n", "date": "2024-01-03", "categories": []}
{"id": "2405.05573", "title": "Poisoning-based Backdoor Attacks for Arbitrary Target Label with\n  Positive Triggers", "abstract": "  Poisoning-based backdoor attacks expose vulnerabilities in the data\npreparation stage of deep neural network (DNN) training. The DNNs trained on\nthe poisoned dataset will be embedded with a backdoor, making them behave well\non clean data while outputting malicious predictions whenever a trigger is\napplied. To exploit the abundant information contained in the input data to\noutput label mapping, our scheme utilizes the network trained from the clean\ndataset as a trigger generator to produce poisons that significantly raise the\nsuccess rate of backdoor attacks versus conventional approaches. Specifically,\nwe provide a new categorization of triggers inspired by the adversarial\ntechnique and develop a multi-label and multi-payload Poisoning-based backdoor\nattack with Positive Triggers (PPT), which effectively moves the input closer\nto the target label on benign classifiers. After the classifier is trained on\nthe poisoned dataset, we can generate an input-label-aware trigger to make the\ninfected classifier predict any given input to any target label with a high\npossibility. Under both dirty- and clean-label settings, we show empirically\nthat the proposed attack achieves a high attack success rate without\nsacrificing accuracy across various datasets, including SVHN, CIFAR10, GTSRB,\nand Tiny ImageNet. Furthermore, the PPT attack can elude a variety of classical\nbackdoor defenses, proving its effectiveness.\n", "date": "2024-05-09", "categories": []}
{"id": "2401.02565", "title": "Demonstration of an Adversarial Attack Against a Multimodal Vision\n  Language Model for Pathology Imaging", "abstract": "  In the context of medical artificial intelligence, this study explores the\nvulnerabilities of the Pathology Language-Image Pretraining (PLIP) model, a\nVision Language Foundation model, under targeted attacks. Leveraging the Kather\nColon dataset with 7,180 H&E images across nine tissue types, our investigation\nemploys Projected Gradient Descent (PGD) adversarial perturbation attacks to\ninduce misclassifications intentionally. The outcomes reveal a 100% success\nrate in manipulating PLIP's predictions, underscoring its susceptibility to\nadversarial perturbations. The qualitative analysis of adversarial examples\ndelves into the interpretability challenges, shedding light on nuanced changes\nin predictions induced by adversarial manipulations. These findings contribute\ncrucial insights into the interpretability, domain adaptation, and\ntrustworthiness of Vision Language Models in medical imaging. The study\nemphasizes the pressing need for robust defenses to ensure the reliability of\nAI models. The source codes for this experiment can be found at\nhttps://github.com/jaiprakash1824/VLM_Adv_Attack.\n", "date": "2024-01-04", "categories": []}
{"id": "2405.03777", "title": "Is ReLU Adversarially Robust?", "abstract": "  The efficacy of deep learning models has been called into question by the\npresence of adversarial examples. Addressing the vulnerability of deep learning\nmodels to adversarial examples is crucial for ensuring their continued\ndevelopment and deployment. In this work, we focus on the role of rectified\nlinear unit (ReLU) activation functions in the generation of adversarial\nexamples. ReLU functions are commonly used in deep learning models because they\nfacilitate the training process. However, our empirical analysis demonstrates\nthat ReLU functions are not robust against adversarial examples. We propose a\nmodified version of the ReLU function, which improves robustness against\nadversarial examples. Our results are supported by an experiment, which\nconfirms the effectiveness of our proposed modification. Additionally, we\ndemonstrate that applying adversarial training to our customized model further\nenhances its robustness compared to a general model.\n", "date": "2024-05-06", "categories": []}
{"id": "2405.03789", "title": "On Adversarial Examples for Text Classification by Perturbing Latent\n  Representations", "abstract": "  Recently, with the advancement of deep learning, several applications in text\nclassification have advanced significantly. However, this improvement comes\nwith a cost because deep learning is vulnerable to adversarial examples. This\nweakness indicates that deep learning is not very robust. Fortunately, the\ninput of a text classifier is discrete. Hence, it can prevent the classifier\nfrom state-of-the-art attacks. Nonetheless, previous works have generated\nblack-box attacks that successfully manipulate the discrete values of the input\nto find adversarial examples. Therefore, instead of changing the discrete\nvalues, we transform the input into its embedding vector containing real values\nto perform the state-of-the-art white-box attacks. Then, we convert the\nperturbed embedding vector back into a text and name it an adversarial example.\nIn summary, we create a framework that measures the robustness of a text\nclassifier by using the gradients of the classifier.\n", "date": "2024-05-06", "categories": []}
{"id": "2405.03918", "title": "Unlearning Backdoor Attacks through Gradient-Based Model Pruning", "abstract": "  In the era of increasing concerns over cybersecurity threats, defending\nagainst backdoor attacks is paramount in ensuring the integrity and reliability\nof machine learning models. However, many existing approaches require\nsubstantial amounts of data for effective mitigation, posing significant\nchallenges in practical deployment. To address this, we propose a novel\napproach to counter backdoor attacks by treating their mitigation as an\nunlearning task. We tackle this challenge through a targeted model pruning\nstrategy, leveraging unlearning loss gradients to identify and eliminate\nbackdoor elements within the model. Built on solid theoretical insights, our\napproach offers simplicity and effectiveness, rendering it well-suited for\nscenarios with limited data availability. Our methodology includes formulating\na suitable unlearning loss and devising a model-pruning technique tailored for\nconvolutional neural networks. Comprehensive evaluations demonstrate the\nefficacy of our proposed approach compared to state-of-the-art approaches,\nparticularly in realistic data settings.\n", "date": "2024-05-07", "categories": []}
{"id": "2405.04191", "title": "Effective and Robust Adversarial Training against Data and Label\n  Corruptions", "abstract": "  Corruptions due to data perturbations and label noise are prevalent in the\ndatasets from unreliable sources, which poses significant threats to model\ntraining. Despite existing efforts in developing robust models, current\nlearning methods commonly overlook the possible co-existence of both\ncorruptions, limiting the effectiveness and practicability of the model. In\nthis paper, we develop an Effective and Robust Adversarial Training (ERAT)\nframework to simultaneously handle two types of corruption (i.e., data and\nlabel) without prior knowledge of their specifics. We propose a hybrid\nadversarial training surrounding multiple potential adversarial perturbations,\nalongside a semi-supervised learning based on class-rebalancing sample\nselection to enhance the resilience of the model for dual corruption. On the\none hand, in the proposed adversarial training, the perturbation generation\nmodule learns multiple surrogate malicious data perturbations by taking a DNN\nmodel as the victim, while the model is trained to maintain semantic\nconsistency between the original data and the hybrid perturbed data. It is\nexpected to enable the model to cope with unpredictable perturbations in\nreal-world data corruption. On the other hand, a class-rebalancing data\nselection strategy is designed to fairly differentiate clean labels from noisy\nlabels. Semi-supervised learning is performed accordingly by discarding noisy\nlabels. Extensive experiments demonstrate the superiority of the proposed ERAT\nframework.\n", "date": "2024-05-07", "categories": []}
{"id": "2206.12977", "title": "Adversarially Robust PAC Learnability of Real-Valued Functions", "abstract": "  We study robustness to test-time adversarial attacks in the regression\nsetting with $\\ell_p$ losses and arbitrary perturbation sets. We address the\nquestion of which function classes are PAC learnable in this setting. We show\nthat classes of finite fat-shattering dimension are learnable in both\nrealizable and agnostic settings. Moreover, for convex function classes, they\nare even properly learnable. In contrast, some non-convex function classes\nprovably require improper learning algorithms. Our main technique is based on a\nconstruction of an adversarially robust sample compression scheme of a size\ndetermined by the fat-shattering dimension. Along the way, we introduce a novel\nagnostic sample compression scheme for real-valued functions, which may be of\nindependent interest.\n", "date": "2022-06-26", "categories": []}
{"id": "2207.10170", "title": "Illusory Attacks: Information-Theoretic Detectability Matters in\n  Adversarial Attacks", "abstract": "  Autonomous agents deployed in the real world need to be robust against\nadversarial attacks on sensory inputs. Robustifying agent policies requires\nanticipating the strongest attacks possible. We demonstrate that existing\nobservation-space attacks on reinforcement learning agents have a common\nweakness: while effective, their lack of information-theoretic detectability\nconstraints makes them detectable using automated means or human inspection.\nDetectability is undesirable to adversaries as it may trigger security\nescalations. We introduce {\\epsilon}-illusory, a novel form of adversarial\nattack on sequential decision-makers that is both effective and of\n{\\epsilon}-bounded statistical detectability. We propose a novel dual ascent\nalgorithm to learn such attacks end-to-end. Compared to existing attacks, we\nempirically find {\\epsilon}-illusory to be significantly harder to detect with\nautomated methods, and a small study with human participants (IRB approval\nunder reference R84123/RE001) suggests they are similarly harder to detect for\nhumans. Our findings suggest the need for better anomaly detectors, as well as\neffective hardware- and system-level defenses. The project website can be found\nat https://tinyurl.com/illusory-attacks.\n", "date": "2022-07-20", "categories": []}
{"id": "2211.11312", "title": "Understanding the Vulnerability of Skeleton-based Human Activity\n  Recognition via Black-box Attack", "abstract": "  Human Activity Recognition (HAR) has been employed in a wide range of\napplications, e.g. self-driving cars, where safety and lives are at stake.\nRecently, the robustness of skeleton-based HAR methods have been questioned due\nto their vulnerability to adversarial attacks. However, the proposed attacks\nrequire the full-knowledge of the attacked classifier, which is overly\nrestrictive. In this paper, we show such threats indeed exist, even when the\nattacker only has access to the input/output of the model. To this end, we\npropose the very first black-box adversarial attack approach in skeleton-based\nHAR called BASAR. BASAR explores the interplay between the classification\nboundary and the natural motion manifold. To our best knowledge, this is the\nfirst time data manifold is introduced in adversarial attacks on time series.\nVia BASAR, we find on-manifold adversarial samples are extremely deceitful and\nrather common in skeletal motions, in contrast to the common belief that\nadversarial samples only exist off-manifold. Through exhaustive evaluation, we\nshow that BASAR can deliver successful attacks across classifiers, datasets,\nand attack modes. By attack, BASAR helps identify the potential causes of the\nmodel vulnerability and provides insights on possible improvements. Finally, to\nmitigate the newly identified threat, we propose a new adversarial training\napproach by leveraging the sophisticated distributions of on/off-manifold\nadversarial samples, called mixed manifold-based adversarial training (MMAT).\nMMAT can successfully help defend against adversarial attacks without\ncompromising classification accuracy.\n", "date": "2022-11-21", "categories": []}
{"id": "2302.10149", "title": "Poisoning Web-Scale Training Datasets is Practical", "abstract": "  Deep learning models are often trained on distributed, web-scale datasets\ncrawled from the internet. In this paper, we introduce two new dataset\npoisoning attacks that intentionally introduce malicious examples to a model's\nperformance. Our attacks are immediately practical and could, today, poison 10\npopular datasets. Our first attack, split-view poisoning, exploits the mutable\nnature of internet content to ensure a dataset annotator's initial view of the\ndataset differs from the view downloaded by subsequent clients. By exploiting\nspecific invalid trust assumptions, we show how we could have poisoned 0.01% of\nthe LAION-400M or COYO-700M datasets for just $60 USD. Our second attack,\nfrontrunning poisoning, targets web-scale datasets that periodically snapshot\ncrowd-sourced content -- such as Wikipedia -- where an attacker only needs a\ntime-limited window to inject malicious examples. In light of both attacks, we\nnotify the maintainers of each affected dataset and recommended several\nlow-overhead defenses.\n", "date": "2023-02-20", "categories": []}
{"id": "2302.12461", "title": "Analyzing And Editing Inner Mechanisms Of Backdoored Language Models", "abstract": "  Poisoning of data sets is a potential security threat to large language\nmodels that can lead to backdoored models. A description of the internal\nmechanisms of backdoored language models and how they process trigger inputs,\ne.g., when switching to toxic language, has yet to be found. In this work, we\nstudy the internal representations of transformer-based backdoored language\nmodels and determine early-layer MLP modules as most important for the backdoor\nmechanism in combination with the initial embedding projection. We use this\nknowledge to remove, insert, and modify backdoor mechanisms with engineered\nreplacements that reduce the MLP module outputs to essentials for the backdoor\nmechanism. To this end, we introduce PCP ablation, where we replace transformer\nmodules with low-rank matrices based on the principal components of their\nactivations. We demonstrate our results on backdoored toy, backdoored large,\nand non-backdoored open-source models. We show that we can improve the backdoor\nrobustness of large language models by locally constraining individual modules\nduring fine-tuning on potentially poisonous data sets.\n  Trigger warning: Offensive language.\n", "date": "2023-02-24", "categories": []}
{"id": "2306.15447", "title": "Are aligned neural networks adversarially aligned?", "abstract": "  Large language models are now tuned to align with the goals of their\ncreators, namely to be \"helpful and harmless.\" These models should respond\nhelpfully to user questions, but refuse to answer requests that could cause\nharm. However, adversarial users can construct inputs which circumvent attempts\nat alignment. In this work, we study adversarial alignment, and ask to what\nextent these models remain aligned when interacting with an adversarial user\nwho constructs worst-case inputs (adversarial examples). These inputs are\ndesigned to cause the model to emit harmful content that would otherwise be\nprohibited. We show that existing NLP-based optimization attacks are\ninsufficiently powerful to reliably attack aligned text models: even when\ncurrent NLP-based attacks fail, we can find adversarial inputs with brute\nforce. As a result, the failure of current attacks should not be seen as proof\nthat aligned text models remain aligned under adversarial inputs.\n  However the recent trend in large-scale ML models is multimodal models that\nallow users to provide images that influence the text that is generated. We\nshow these models can be easily attacked, i.e., induced to perform arbitrary\nun-aligned behavior through adversarial perturbation of the input image. We\nconjecture that improved NLP attacks may demonstrate this same level of\nadversarial control over text-only models.\n", "date": "2023-06-26", "categories": []}
{"id": "2310.18936", "title": "Adversarial Examples Are Not Real Features", "abstract": "  The existence of adversarial examples has been a mystery for years and\nattracted much interest. A well-known theory by \\citet{ilyas2019adversarial}\nexplains adversarial vulnerability from a data perspective by showing that one\ncan extract non-robust features from adversarial examples and these features\nalone are useful for classification. However, the explanation remains quite\ncounter-intuitive since non-robust features are mostly noise features to\nhumans. In this paper, we re-examine the theory from a larger context by\nincorporating multiple learning paradigms. Notably, we find that contrary to\ntheir good usefulness under supervised learning, non-robust features attain\npoor usefulness when transferred to other self-supervised learning paradigms,\nsuch as contrastive learning, masked image modeling, and diffusion models. It\nreveals that non-robust features are not really as useful as robust or natural\nfeatures that enjoy good transferability between these paradigms. Meanwhile,\nfor robustness, we also show that naturally trained encoders from robust\nfeatures are largely non-robust under AutoAttack. Our cross-paradigm\nexamination suggests that the non-robust features are not really useful but\nmore like paradigm-wise shortcuts, and robust features alone might be\ninsufficient to attain reliable model robustness. Code is available at\n\\url{https://github.com/PKU-ML/AdvNotRealFeatures}.\n", "date": "2023-10-29", "categories": []}
{"id": "2405.01460", "title": "Purify Unlearnable Examples via Rate-Constrained Variational\n  Autoencoders", "abstract": "  Unlearnable examples (UEs) seek to maximize testing error by making subtle\nmodifications to training examples that are correctly labeled. Defenses against\nthese poisoning attacks can be categorized based on whether specific\ninterventions are adopted during training. The first approach is training-time\ndefense, such as adversarial training, which can mitigate poisoning effects but\nis computationally intensive. The other approach is pre-training purification,\ne.g., image short squeezing, which consists of several simple compressions but\noften encounters challenges in dealing with various UEs. Our work provides a\nnovel disentanglement mechanism to build an efficient pre-training purification\nmethod. Firstly, we uncover rate-constrained variational autoencoders (VAEs),\ndemonstrating a clear tendency to suppress the perturbations in UEs. We\nsubsequently conduct a theoretical analysis for this phenomenon. Building upon\nthese insights, we introduce a disentangle variational autoencoder (D-VAE),\ncapable of disentangling the perturbations with learnable class-wise\nembeddings. Based on this network, a two-stage purification approach is\nnaturally developed. The first stage focuses on roughly eliminating\nperturbations, while the second stage produces refined, poison-free results,\nensuring effectiveness and robustness across various scenarios. Extensive\nexperiments demonstrate the remarkable performance of our method across\nCIFAR-10, CIFAR-100, and a 100-class ImageNet-subset. Code is available at\nhttps://github.com/yuyi-sd/D-VAE.\n", "date": "2024-05-02", "categories": []}
{"id": "2405.03193", "title": "Exploring Frequencies via Feature Mixing and Meta-Learning for Improving\n  Adversarial Transferability", "abstract": "  Recent studies have shown that Deep Neural Networks (DNNs) are susceptible to\nadversarial attacks, with frequency-domain analysis underscoring the\nsignificance of high-frequency components in influencing model predictions.\nConversely, targeting low-frequency components has been effective in enhancing\nattack transferability on black-box models. In this study, we introduce a\nfrequency decomposition-based feature mixing method to exploit these frequency\ncharacteristics in both clean and adversarial samples. Our findings suggest\nthat incorporating features of clean samples into adversarial features\nextracted from adversarial examples is more effective in attacking\nnormally-trained models, while combining clean features with the adversarial\nfeatures extracted from low-frequency parts decomposed from the adversarial\nsamples yields better results in attacking defense models. However, a conflict\nissue arises when these two mixing approaches are employed simultaneously. To\ntackle the issue, we propose a cross-frequency meta-optimization approach\ncomprising the meta-train step, meta-test step, and final update. In the\nmeta-train step, we leverage the low-frequency components of adversarial\nsamples to boost the transferability of attacks against defense models.\nMeanwhile, in the meta-test step, we utilize adversarial samples to stabilize\ngradients, thereby enhancing the attack's transferability against normally\ntrained models. For the final update, we update the adversarial sample based on\nthe gradients obtained from both meta-train and meta-test steps. Our proposed\nmethod is evaluated through extensive experiments on the ImageNet-Compatible\ndataset, affirming its effectiveness in improving the transferability of\nattacks on both normally-trained CNNs and defense models.\n  The source code is available at https://github.com/WJJLL/MetaSSA.\n", "date": "2024-05-06", "categories": []}
{"id": "2310.05336", "title": "GReAT: A Graph Regularized Adversarial Training Method", "abstract": "  This paper presents GReAT (Graph Regularized Adversarial Training), a novel\nregularization method designed to enhance the robust classification performance\nof deep learning models. Adversarial examples, characterized by subtle\nperturbations that can mislead models, pose a significant challenge in machine\nlearning. Although adversarial training is effective in defending against such\nattacks, it often overlooks the underlying data structure. In response, GReAT\nintegrates graph based regularization into the adversarial training process,\nleveraging the data's inherent structure to enhance model robustness. By\nincorporating graph information during training, GReAT defends against\nadversarial attacks and improves generalization to unseen data. Extensive\nevaluations on benchmark datasets demonstrate that GReAT outperforms state of\nthe art methods in robustness, achieving notable improvements in classification\naccuracy. Specifically, compared to the second best methods, GReAT achieves a\nperformance increase of approximately 4.87% for CIFAR10 against FGSM attack and\n10.57% for SVHN against FGSM attack. Additionally, for CIFAR10, GReAT\ndemonstrates a performance increase of approximately 11.05% against PGD attack,\nand for SVHN, a 5.54% increase against PGD attack. This paper provides detailed\ninsights into the proposed methodology, including numerical results and\ncomparisons with existing approaches, highlighting the significant impact of\nGReAT in advancing the performance of deep learning models.\n", "date": "2023-10-09", "categories": []}
{"id": "2404.05688", "title": "David and Goliath: An Empirical Evaluation of Attacks and Defenses for\n  QNNs at the Deep Edge", "abstract": "  ML is shifting from the cloud to the edge. Edge computing reduces the surface\nexposing private data and enables reliable throughput guarantees in real-time\napplications. Of the panoply of devices deployed at the edge,\nresource-constrained MCUs, e.g., Arm Cortex-M, are more prevalent, orders of\nmagnitude cheaper, and less power-hungry than application processors or GPUs.\nThus, enabling intelligence at the deep edge is the zeitgeist, with researchers\nfocusing on unveiling novel approaches to deploy ANNs on these constrained\ndevices. Quantization is a well-established technique that has proved effective\nin enabling the deployment of neural networks on MCUs; however, it is still an\nopen question to understand the robustness of QNNs in the face of adversarial\nexamples.\n  To fill this gap, we empirically evaluate the effectiveness of attacks and\ndefenses from (full-precision) ANNs on (constrained) QNNs. Our evaluation\nincludes three QNNs targeting TinyML applications, ten attacks, and six\ndefenses. With this study, we draw a set of interesting findings. First,\nquantization increases the point distance to the decision boundary and leads\nthe gradient estimated by some attacks to explode or vanish. Second,\nquantization can act as a noise attenuator or amplifier, depending on the noise\nmagnitude, and causes gradient misalignment. Regarding adversarial defenses, we\nconclude that input pre-processing defenses show impressive results on small\nperturbations; however, they fall short as the perturbation increases. At the\nsame time, train-based defenses increase the average point distance to the\ndecision boundary, which holds after quantization. However, we argue that\ntrain-based defenses still need to smooth the quantization-shift and gradient\nmisalignment phenomenons to counteract adversarial example transferability to\nQNNs. All artifacts are open-sourced to enable independent validation of\nresults.\n", "date": "2024-04-08", "categories": []}
{"id": "2405.01817", "title": "Uniformly Stable Algorithms for Adversarial Training and Beyond", "abstract": "  In adversarial machine learning, neural networks suffer from a significant\nissue known as robust overfitting, where the robust test accuracy decreases\nover epochs (Rice et al., 2020). Recent research conducted by Xing et al.,2021;\nXiao et al., 2022 has focused on studying the uniform stability of adversarial\ntraining. Their investigations revealed that SGD-based adversarial training\nfails to exhibit uniform stability, and the derived stability bounds align with\nthe observed phenomenon of robust overfitting in experiments. This motivates us\nto develop uniformly stable algorithms specifically tailored for adversarial\ntraining. To this aim, we introduce Moreau envelope-$\\mathcal{A}$, a variant of\nthe Moreau Envelope-type algorithm. We employ a Moreau envelope function to\nreframe the original problem as a min-min problem, separating the non-strong\nconvexity and non-smoothness of the adversarial loss. Then, this approach\nalternates between solving the inner and outer minimization problems to achieve\nuniform stability without incurring additional computational overhead. In\npractical scenarios, we show the efficacy of ME-$\\mathcal{A}$ in mitigating the\nissue of robust overfitting. Beyond its application in adversarial training,\nthis represents a fundamental result in uniform stability analysis, as\nME-$\\mathcal{A}$ is the first algorithm to exhibit uniform stability for\nweakly-convex, non-smooth problems.\n", "date": "2024-05-03", "categories": []}
{"id": "2405.01838", "title": "A Novel Approach to Guard from Adversarial Attacks using Stable\n  Diffusion", "abstract": "  Recent developments in adversarial machine learning have highlighted the\nimportance of building robust AI systems to protect against increasingly\nsophisticated attacks. While frameworks like AI Guardian are designed to defend\nagainst these threats, they often rely on assumptions that can limit their\neffectiveness. For example, they may assume attacks only come from one\ndirection or include adversarial images in their training data. Our proposal\nsuggests a different approach to the AI Guardian framework. Instead of\nincluding adversarial examples in the training process, we propose training the\nAI system without them. This aims to create a system that is inherently\nresilient to a wider range of attacks. Our method focuses on a dynamic defense\nstrategy using stable diffusion that learns continuously and models threats\ncomprehensively. We believe this approach can lead to a more generalized and\nrobust defense against adversarial attacks.\n  In this paper, we outline our proposed approach, including the theoretical\nbasis, experimental design, and expected impact on improving AI security\nagainst adversarial threats.\n", "date": "2024-05-03", "categories": []}
{"id": "2405.01934", "title": "Impact of Architectural Modifications on Deep Learning Adversarial\n  Robustness", "abstract": "  Rapid advancements of deep learning are accelerating adoption in a wide\nvariety of applications, including safety-critical applications such as\nself-driving vehicles, drones, robots, and surveillance systems. These\nadvancements include applying variations of sophisticated techniques that\nimprove the performance of models. However, such models are not immune to\nadversarial manipulations, which can cause the system to misbehave and remain\nunnoticed by experts. The frequency of modifications to existing deep learning\nmodels necessitates thorough analysis to determine the impact on models'\nrobustness. In this work, we present an experimental evaluation of the effects\nof model modifications on deep learning model robustness using adversarial\nattacks. Our methodology involves examining the robustness of variations of\nmodels against various adversarial attacks. By conducting our experiments, we\naim to shed light on the critical issue of maintaining the reliability and\nsafety of deep learning models in safety- and security-critical applications.\nOur results indicate the pressing demand for an in-depth assessment of the\neffects of model changes on the robustness of models.\n", "date": "2024-05-03", "categories": []}
{"id": "2405.01963", "title": "From Attack to Defense: Insights into Deep Learning Security Measures in\n  Black-Box Settings", "abstract": "  Deep Learning (DL) is rapidly maturing to the point that it can be used in\nsafety- and security-crucial applications. However, adversarial samples, which\nare undetectable to the human eye, pose a serious threat that can cause the\nmodel to misbehave and compromise the performance of such applications.\nAddressing the robustness of DL models has become crucial to understanding and\ndefending against adversarial attacks. In this study, we perform comprehensive\nexperiments to examine the effect of adversarial attacks and defenses on\nvarious model architectures across well-known datasets. Our research focuses on\nblack-box attacks such as SimBA, HopSkipJump, MGAAttack, and boundary attacks,\nas well as preprocessor-based defensive mechanisms, including bits squeezing,\nmedian smoothing, and JPEG filter. Experimenting with various models, our\nresults demonstrate that the level of noise needed for the attack increases as\nthe number of layers increases. Moreover, the attack success rate decreases as\nthe number of layers increases. This indicates that model complexity and\nrobustness have a significant relationship. Investigating the diversity and\nrobustness relationship, our experiments with diverse models show that having a\nlarge number of parameters does not imply higher robustness. Our experiments\nextend to show the effects of the training dataset on model robustness. Using\nvarious datasets such as ImageNet-1000, CIFAR-100, and CIFAR-10 are used to\nevaluate the black-box attacks. Considering the multiple dimensions of our\nanalysis, e.g., model complexity and training dataset, we examined the behavior\nof black-box attacks when models apply defenses. Our results show that applying\ndefense strategies can significantly reduce attack effectiveness. This research\nprovides in-depth analysis and insight into the robustness of DL models against\nvarious attacks, and defenses.\n", "date": "2024-05-03", "categories": []}
{"id": "2308.15692", "title": "Intriguing Properties of Diffusion Models: An Empirical Study of the\n  Natural Attack Capability in Text-to-Image Generative Models", "abstract": "  Denoising probabilistic diffusion models have shown breakthrough performance\nto generate more photo-realistic images or human-level illustrations than the\nprior models such as GANs. This high image-generation capability has stimulated\nthe creation of many downstream applications in various areas. However, we find\nthat this technology is actually a double-edged sword: We identify a new type\nof attack, called the Natural Denoising Diffusion (NDD) attack based on the\nfinding that state-of-the-art deep neural network (DNN) models still hold their\nprediction even if we intentionally remove their robust features, which are\nessential to the human visual system (HVS), through text prompts. The NDD\nattack shows a significantly high capability to generate low-cost,\nmodel-agnostic, and transferable adversarial attacks by exploiting the natural\nattack capability in diffusion models. To systematically evaluate the risk of\nthe NDD attack, we perform a large-scale empirical study with our newly created\ndataset, the Natural Denoising Diffusion Attack (NDDA) dataset. We evaluate the\nnatural attack capability by answering 6 research questions. Through a user\nstudy, we find that it can achieve an 88% detection rate while being stealthy\nto 93% of human subjects; we also find that the non-robust features embedded by\ndiffusion models contribute to the natural attack capability. To confirm the\nmodel-agnostic and transferable attack capability, we perform the NDD attack\nagainst the Tesla Model 3 and find that 73% of the physically printed attacks\ncan be detected as stop signs. Our hope is that the study and dataset can help\nour community be aware of the risks in diffusion models and facilitate further\nresearch toward robust DNN models.\n", "date": "2023-08-30", "categories": []}
{"id": "2310.17626", "title": "A Survey on Transferability of Adversarial Examples across Deep Neural\n  Networks", "abstract": "  The emergence of Deep Neural Networks (DNNs) has revolutionized various\ndomains by enabling the resolution of complex tasks spanning image recognition,\nnatural language processing, and scientific problem-solving. However, this\nprogress has also brought to light a concerning vulnerability: adversarial\nexamples. These crafted inputs, imperceptible to humans, can manipulate machine\nlearning models into making erroneous predictions, raising concerns for\nsafety-critical applications. An intriguing property of this phenomenon is the\ntransferability of adversarial examples, where perturbations crafted for one\nmodel can deceive another, often with a different architecture. This intriguing\nproperty enables black-box attacks which circumvents the need for detailed\nknowledge of the target model. This survey explores the landscape of the\nadversarial transferability of adversarial examples. We categorize existing\nmethodologies to enhance adversarial transferability and discuss the\nfundamental principles guiding each approach. While the predominant body of\nresearch primarily concentrates on image classification, we also extend our\ndiscussion to encompass other vision tasks and beyond. Challenges and\nopportunities are discussed, highlighting the importance of fortifying DNNs\nagainst adversarial vulnerabilities in an evolving landscape.\n", "date": "2023-10-26", "categories": []}
{"id": "2404.13320", "title": "Pixel is a Barrier: Diffusion Models Are More Adversarially Robust Than\n  We Think", "abstract": "  Adversarial examples for diffusion models are widely used as solutions for\nsafety concerns. By adding adversarial perturbations to personal images,\nattackers can not edit or imitate them easily. However, it is essential to note\nthat all these protections target the latent diffusion model (LDMs), the\nadversarial examples for diffusion models in the pixel space (PDMs) are largely\noverlooked. This may mislead us to think that the diffusion models are\nvulnerable to adversarial attacks like most deep models. In this paper, we show\nnovel findings that: even though gradient-based white-box attacks can be used\nto attack the LDMs, they fail to attack PDMs. This finding is supported by\nextensive experiments of almost a wide range of attacking methods on various\nPDMs and LDMs with different model structures, which means diffusion models are\nindeed much more robust against adversarial attacks. We also find that PDMs can\nbe used as an off-the-shelf purifier to effectively remove the adversarial\npatterns that were generated on LDMs to protect the images, which means that\nmost protection methods nowadays, to some extent, cannot protect our images\nfrom malicious attacks. We hope that our insights will inspire the community to\nrethink the adversarial samples for diffusion models as protection methods and\nmove forward to more effective protection. Codes are available in\nhttps://github.com/xavihart/PDM-Pure.\n", "date": "2024-04-20", "categories": []}
{"id": "2404.15656", "title": "MISLEAD: Manipulating Importance of Selected features for Learning\n  Epsilon in Evasion Attack Deception", "abstract": "  Emerging vulnerabilities in machine learning (ML) models due to adversarial\nattacks raise concerns about their reliability. Specifically, evasion attacks\nmanipulate models by introducing precise perturbations to input data, causing\nerroneous predictions. To address this, we propose a methodology combining\nSHapley Additive exPlanations (SHAP) for feature importance analysis with an\ninnovative Optimal Epsilon technique for conducting evasion attacks. Our\napproach begins with SHAP-based analysis to understand model vulnerabilities,\ncrucial for devising targeted evasion strategies. The Optimal Epsilon\ntechnique, employing a Binary Search algorithm, efficiently determines the\nminimum epsilon needed for successful evasion. Evaluation across diverse\nmachine learning architectures demonstrates the technique's precision in\ngenerating adversarial samples, underscoring its efficacy in manipulating model\noutcomes. This study emphasizes the critical importance of continuous\nassessment and monitoring to identify and mitigate potential security risks in\nmachine learning systems.\n", "date": "2024-04-24", "categories": []}
{"id": "2405.00289", "title": "Adversarial Attacks and Defense for Conversation Entailment Task", "abstract": "  As the deployment of NLP systems in critical applications grows, ensuring the\nrobustness of large language models (LLMs) against adversarial attacks becomes\nincreasingly important. Large language models excel in various NLP tasks but\nremain vulnerable to low-cost adversarial attacks. Focusing on the domain of\nconversation entailment, where multi-turn dialogues serve as premises to verify\nhypotheses, we fine-tune a transformer model to accurately discern the\ntruthfulness of these hypotheses. Adversaries manipulate hypotheses through\nsynonym swapping, aiming to deceive the model into making incorrect\npredictions. To counteract these attacks, we implemented innovative fine-tuning\ntechniques and introduced an embedding perturbation loss method to\nsignificantly bolster the model's robustness. Our findings not only emphasize\nthe importance of defending against adversarial attacks in NLP but also\nhighlight the real-world implications, suggesting that enhancing model\nrobustness is critical for reliable NLP applications.\n", "date": "2024-05-01", "categories": []}
{"id": "2405.01073", "title": "Poisoning Attacks on Federated Learning for Autonomous Driving", "abstract": "  Federated Learning (FL) is a decentralized learning paradigm, enabling\nparties to collaboratively train models while keeping their data confidential.\nWithin autonomous driving, it brings the potential of reducing data storage\ncosts, reducing bandwidth requirements, and to accelerate the learning. FL is,\nhowever, susceptible to poisoning attacks. In this paper, we introduce two\nnovel poisoning attacks on FL tailored to regression tasks within autonomous\ndriving: FLStealth and Off-Track Attack (OTA). FLStealth, an untargeted attack,\naims at providing model updates that deteriorate the global model performance\nwhile appearing benign. OTA, on the other hand, is a targeted attack with the\nobjective to change the global model's behavior when exposed to a certain\ntrigger. We demonstrate the effectiveness of our attacks by conducting\ncomprehensive experiments pertaining to the task of vehicle trajectory\nprediction. In particular, we show that, among five different untargeted\nattacks, FLStealth is the most successful at bypassing the considered defenses\nemployed by the server. For OTA, we demonstrate the inability of common defense\nstrategies to mitigate the attack, highlighting the critical need for new\ndefensive mechanisms against targeted attacks within FL for autonomous driving.\n", "date": "2024-05-02", "categories": []}
{"id": "2404.18702", "title": "Why You Should Not Trust Interpretations in Machine Learning:\n  Adversarial Attacks on Partial Dependence Plots", "abstract": "  The adoption of artificial intelligence (AI) across industries has led to the\nwidespread use of complex black-box models and interpretation tools for\ndecision making. This paper proposes an adversarial framework to uncover the\nvulnerability of permutation-based interpretation methods for machine learning\ntasks, with a particular focus on partial dependence (PD) plots. This\nadversarial framework modifies the original black box model to manipulate its\npredictions for instances in the extrapolation domain. As a result, it produces\ndeceptive PD plots that can conceal discriminatory behaviors while preserving\nmost of the original model's predictions. This framework can produce multiple\nfooled PD plots via a single model. By using real-world datasets including an\nauto insurance claims dataset and COMPAS (Correctional Offender Management\nProfiling for Alternative Sanctions) dataset, our results show that it is\npossible to intentionally hide the discriminatory behavior of a predictor and\nmake the black-box model appear neutral through interpretation tools like PD\nplots while retaining almost all the predictions of the original black-box\nmodel. Managerial insights for regulators and practitioners are provided based\non the findings.\n", "date": "2024-04-29", "categories": []}
{"id": "2405.00392", "title": "Certified Adversarial Robustness of Machine Learning-based Malware\n  Detectors via (De)Randomized Smoothing", "abstract": "  Deep learning-based malware detection systems are vulnerable to adversarial\nEXEmples - carefully-crafted malicious programs that evade detection with\nminimal perturbation. As such, the community is dedicating effort to develop\nmechanisms to defend against adversarial EXEmples. However, current randomized\nsmoothing-based defenses are still vulnerable to attacks that inject blocks of\nadversarial content. In this paper, we introduce a certifiable defense against\npatch attacks that guarantees, for a given executable and an adversarial patch\nsize, no adversarial EXEmple exist. Our method is inspired by (de)randomized\nsmoothing which provides deterministic robustness certificates. During\ntraining, a base classifier is trained using subsets of continguous bytes. At\ninference time, our defense splits the executable into non-overlapping chunks,\nclassifies each chunk independently, and computes the final prediction through\nmajority voting to minimize the influence of injected content. Furthermore, we\nintroduce a preprocessing step that fixes the size of the sections and headers\nto a multiple of the chunk size. As a consequence, the injected content is\nconfined to an integer number of chunks without tampering the other chunks\ncontaining the real bytes of the input examples, allowing us to extend our\ncertified robustness guarantees to content insertion attacks. We perform an\nextensive ablation study, by comparing our defense with randomized\nsmoothing-based defenses against a plethora of content manipulation attacks and\nneural network architectures. Results show that our method exhibits unmatched\nrobustness against strong content-insertion attacks, outperforming randomized\nsmoothing-based defenses in the literature.\n", "date": "2024-05-01", "categories": []}
{"id": "2404.19382", "title": "Probing Unlearned Diffusion Models: A Transferable Adversarial Attack\n  Perspective", "abstract": "  Advanced text-to-image diffusion models raise safety concerns regarding\nidentity privacy violation, copyright infringement, and Not Safe For Work\ncontent generation. Towards this, unlearning methods have been developed to\nerase these involved concepts from diffusion models. However, these unlearning\nmethods only shift the text-to-image mapping and preserve the visual content\nwithin the generative space of diffusion models, leaving a fatal flaw for\nrestoring these erased concepts. This erasure trustworthiness problem needs\nprobe, but previous methods are sub-optimal from two perspectives: (1) Lack of\ntransferability: Some methods operate within a white-box setting, requiring\naccess to the unlearned model. And the learned adversarial input often fails to\ntransfer to other unlearned models for concept restoration; (2) Limited attack:\nThe prompt-level methods struggle to restore narrow concepts from unlearned\nmodels, such as celebrity identity. Therefore, this paper aims to leverage the\ntransferability of the adversarial attack to probe the unlearning robustness\nunder a black-box setting. This challenging scenario assumes that the\nunlearning method is unknown and the unlearned model is inaccessible for\noptimization, requiring the attack to be capable of transferring across\ndifferent unlearned models. Specifically, we employ an adversarial search\nstrategy to search for the adversarial embedding which can transfer across\ndifferent unlearned models. This strategy adopts the original Stable Diffusion\nmodel as a surrogate model to iteratively erase and search for embeddings,\nenabling it to find the embedding that can restore the target concept for\ndifferent unlearning methods. Extensive experiments demonstrate the\ntransferability of the searched adversarial embedding across several\nstate-of-the-art unlearning methods and its effectiveness for different levels\nof concepts.\n", "date": "2024-04-30", "categories": []}
{"id": "2404.19417", "title": "Physical Backdoor: Towards Temperature-based Backdoor Attacks in the\n  Physical World", "abstract": "  Backdoor attacks have been well-studied in visible light object detection\n(VLOD) in recent years. However, VLOD can not effectively work in dark and\ntemperature-sensitive scenarios. Instead, thermal infrared object detection\n(TIOD) is the most accessible and practical in such environments. In this\npaper, our team is the first to investigate the security vulnerabilities\nassociated with TIOD in the context of backdoor attacks, spanning both the\ndigital and physical realms. We introduce two novel types of backdoor attacks\non TIOD, each offering unique capabilities: Object-affecting Attack and\nRange-affecting Attack. We conduct a comprehensive analysis of key factors\ninfluencing trigger design, which include temperature, size, material, and\nconcealment. These factors, especially temperature, significantly impact the\nefficacy of backdoor attacks on TIOD. A thorough understanding of these factors\nwill serve as a foundation for designing physical triggers and temperature\ncontrolling experiments. Our study includes extensive experiments conducted in\nboth digital and physical environments. In the digital realm, we evaluate our\napproach using benchmark datasets for TIOD, achieving an Attack Success Rate\n(ASR) of up to 98.21%. In the physical realm, we test our approach in two\nreal-world settings: a traffic intersection and a parking lot, using a thermal\ninfrared camera. Here, we attain an ASR of up to 98.38%.\n", "date": "2024-04-30", "categories": []}
{"id": "2404.19640", "title": "Attacking Bayes: On the Adversarial Robustness of Bayesian Neural\n  Networks", "abstract": "  Adversarial examples have been shown to cause neural networks to fail on a\nwide range of vision and language tasks, but recent work has claimed that\nBayesian neural networks (BNNs) are inherently robust to adversarial\nperturbations. In this work, we examine this claim. To study the adversarial\nrobustness of BNNs, we investigate whether it is possible to successfully break\nstate-of-the-art BNN inference methods and prediction pipelines using even\nrelatively unsophisticated attacks for three tasks: (1) label prediction under\nthe posterior predictive mean, (2) adversarial example detection with Bayesian\npredictive uncertainty, and (3) semantic shift detection. We find that BNNs\ntrained with state-of-the-art approximate inference methods, and even BNNs\ntrained with Hamiltonian Monte Carlo, are highly susceptible to adversarial\nattacks. We also identify various conceptual and experimental errors in\nprevious works that claimed inherent adversarial robustness of BNNs and\nconclusively demonstrate that BNNs and uncertainty-aware Bayesian prediction\npipelines are not inherently robust against adversarial attacks.\n", "date": "2024-04-27", "categories": []}
{"id": "2305.10862", "title": "How Deep Learning Sees the World: A Survey on Adversarial Attacks &\n  Defenses", "abstract": "  Deep Learning is currently used to perform multiple tasks, such as object\nrecognition, face recognition, and natural language processing. However, Deep\nNeural Networks (DNNs) are vulnerable to perturbations that alter the network\nprediction (adversarial examples), raising concerns regarding its usage in\ncritical areas, such as self-driving vehicles, malware detection, and\nhealthcare. This paper compiles the most recent adversarial attacks, grouped by\nthe attacker capacity, and modern defenses clustered by protection strategies.\nWe also present the new advances regarding Vision Transformers, summarize the\ndatasets and metrics used in the context of adversarial settings, and compare\nthe state-of-the-art results under different attacks, finishing with the\nidentification of open issues.\n", "date": "2023-05-18", "categories": []}
{"id": "2306.07544", "title": "On Achieving Optimal Adversarial Test Error", "abstract": "  We first elucidate various fundamental properties of optimal adversarial\npredictors: the structure of optimal adversarial convex predictors in terms of\noptimal adversarial zero-one predictors, bounds relating the adversarial convex\nloss to the adversarial zero-one loss, and the fact that continuous predictors\ncan get arbitrarily close to the optimal adversarial error for both convex and\nzero-one losses. Applying these results along with new Rademacher complexity\nbounds for adversarial training near initialization, we prove that for general\ndata distributions and perturbation sets, adversarial training on shallow\nnetworks with early stopping and an idealized optimal adversary is able to\nachieve optimal adversarial test error. By contrast, prior theoretical work\neither considered specialized data distributions or only provided training\nerror guarantees.\n", "date": "2023-06-13", "categories": []}
{"id": "2310.13828", "title": "Nightshade: Prompt-Specific Poisoning Attacks on Text-to-Image\n  Generative Models", "abstract": "  Data poisoning attacks manipulate training data to introduce unexpected\nbehaviors into machine learning models at training time. For text-to-image\ngenerative models with massive training datasets, current understanding of\npoisoning attacks suggests that a successful attack would require injecting\nmillions of poison samples into their training pipeline. In this paper, we show\nthat poisoning attacks can be successful on generative models. We observe that\ntraining data per concept can be quite limited in these models, making them\nvulnerable to prompt-specific poisoning attacks, which target a model's ability\nto respond to individual prompts.\n  We introduce Nightshade, an optimized prompt-specific poisoning attack where\npoison samples look visually identical to benign images with matching text\nprompts. Nightshade poison samples are also optimized for potency and can\ncorrupt an Stable Diffusion SDXL prompt in <100 poison samples. Nightshade\npoison effects \"bleed through\" to related concepts, and multiple attacks can\ncomposed together in a single prompt. Surprisingly, we show that a moderate\nnumber of Nightshade attacks can destabilize general features in a\ntext-to-image generative model, effectively disabling its ability to generate\nmeaningful images. Finally, we propose the use of Nightshade and similar tools\nas a last defense for content creators against web scrapers that ignore\nopt-out/do-not-crawl directives, and discuss possible implications for model\ntrainers and content creators.\n", "date": "2023-10-20", "categories": []}
{"id": "2311.13127", "title": "MetaCloak: Preventing Unauthorized Subject-driven Text-to-image\n  Diffusion-based Synthesis via Meta-learning", "abstract": "  Text-to-image diffusion models allow seamless generation of personalized\nimages from scant reference photos. Yet, these tools, in the wrong hands, can\nfabricate misleading or harmful content, endangering individuals. To address\nthis problem, existing poisoning-based approaches perturb user images in an\nimperceptible way to render them \"unlearnable\" from malicious uses. We identify\ntwo limitations of these defending approaches: i) sub-optimal due to the\nhand-crafted heuristics for solving the intractable bilevel optimization and\nii) lack of robustness against simple data transformations like Gaussian\nfiltering. To solve these challenges, we propose MetaCloak, which solves the\nbi-level poisoning problem with a meta-learning framework with an additional\ntransformation sampling process to craft transferable and robust perturbation.\nSpecifically, we employ a pool of surrogate diffusion models to craft\ntransferable and model-agnostic perturbation. Furthermore, by incorporating an\nadditional transformation process, we design a simple denoising-error\nmaximization loss that is sufficient for causing transformation-robust semantic\ndistortion and degradation in a personalized generation. Extensive experiments\non the VGGFace2 and CelebA-HQ datasets show that MetaCloak outperforms existing\napproaches. Notably, MetaCloak can successfully fool online training services\nlike Replicate, in a black-box manner, demonstrating the effectiveness of\nMetaCloak in real-world scenarios. Our code is available at\nhttps://github.com/liuyixin-louis/MetaCloak.\n", "date": "2023-11-22", "categories": []}
{"id": "2311.14455", "title": "Universal Jailbreak Backdoors from Poisoned Human Feedback", "abstract": "  Reinforcement Learning from Human Feedback (RLHF) is used to align large\nlanguage models to produce helpful and harmless responses. Yet, prior work\nshowed these models can be jailbroken by finding adversarial prompts that\nrevert the model to its unaligned behavior. In this paper, we consider a new\nthreat where an attacker poisons the RLHF training data to embed a \"jailbreak\nbackdoor\" into the model. The backdoor embeds a trigger word into the model\nthat acts like a universal \"sudo command\": adding the trigger word to any\nprompt enables harmful responses without the need to search for an adversarial\nprompt. Universal jailbreak backdoors are much more powerful than previously\nstudied backdoors on language models, and we find they are significantly harder\nto plant using common backdoor attack techniques. We investigate the design\ndecisions in RLHF that contribute to its purported robustness, and release a\nbenchmark of poisoned models to stimulate future research on universal\njailbreak backdoors.\n", "date": "2023-11-24", "categories": []}
{"id": "2404.17760", "title": "Adversarial Examples: Generation Proposal in the Context of Facial\n  Recognition Systems", "abstract": "  In this paper we investigate the vulnerability that facial recognition\nsystems present to adversarial examples by introducing a new methodology from\nthe attacker perspective. The technique is based on the use of the autoencoder\nlatent space, organized with principal component analysis. We intend to analyze\nthe potential to craft adversarial examples suitable for both dodging and\nimpersonation attacks, against state-of-the-art systems. Our initial\nhypothesis, which was not strongly favoured by the results, stated that it\nwould be possible to separate between the \"identity\" and \"facial expression\"\nfeatures to produce high-quality examples. Despite the findings not supporting\nit, the results sparked insights into adversarial examples generation and\nopened new research avenues in the area.\n", "date": "2024-04-27", "categories": []}
{"id": "2404.17947", "title": "Bounding the Expected Robustness of Graph Neural Networks Subject to\n  Node Feature Attacks", "abstract": "  Graph Neural Networks (GNNs) have demonstrated state-of-the-art performance\nin various graph representation learning tasks. Recently, studies revealed\ntheir vulnerability to adversarial attacks. In this work, we theoretically\ndefine the concept of expected robustness in the context of attributed graphs\nand relate it to the classical definition of adversarial robustness in the\ngraph representation learning literature. Our definition allows us to derive an\nupper bound of the expected robustness of Graph Convolutional Networks (GCNs)\nand Graph Isomorphism Networks subject to node feature attacks. Building on\nthese findings, we connect the expected robustness of GNNs to the\northonormality of their weight matrices and consequently propose an\nattack-independent, more robust variant of the GCN, called the Graph\nConvolutional Orthonormal Robust Networks (GCORNs). We further introduce a\nprobabilistic method to estimate the expected robustness, which allows us to\nevaluate the effectiveness of GCORN on several real-world datasets.\nExperimental experiments showed that GCORN outperforms available defense\nmethods. Our code is publicly available at:\n\\href{https://github.com/Sennadir/GCORN}{https://github.com/Sennadir/GCORN}.\n", "date": "2024-04-27", "categories": []}
{"id": "2304.05370", "title": "Overload: Latency Attacks on Object Detection for Edge Devices", "abstract": "  Nowadays, the deployment of deep learning-based applications is an essential\ntask owing to the increasing demands on intelligent services. In this paper, we\ninvestigate latency attacks on deep learning applications. Unlike common\nadversarial attacks for misclassification, the goal of latency attacks is to\nincrease the inference time, which may stop applications from responding to the\nrequests within a reasonable time. This kind of attack is ubiquitous for\nvarious applications, and we use object detection to demonstrate how such kind\nof attacks work. We also design a framework named Overload to generate latency\nattacks at scale. Our method is based on a newly formulated optimization\nproblem and a novel technique, called spatial attention. This attack serves to\nescalate the required computing costs during the inference time, consequently\nleading to an extended inference time for object detection. It presents a\nsignificant threat, especially to systems with limited computing resources. We\nconducted experiments using YOLOv5 models on Nvidia NX. Compared to existing\nmethods, our method is simpler and more effective. The experimental results\nshow that with latency attacks, the inference time of a single image can be\nincreased ten times longer in reference to the normal setting. Moreover, our\nfindings pose a potential new threat to all object detection tasks requiring\nnon-maximum suppression (NMS), as our attack is NMS-agnostic.\n", "date": "2023-04-11", "categories": []}
{"id": "2305.02605", "title": "Toward Evaluating Robustness of Reinforcement Learning with Adversarial\n  Policy", "abstract": "  Reinforcement learning agents are susceptible to evasion attacks during\ndeployment. In single-agent environments, these attacks can occur through\nimperceptible perturbations injected into the inputs of the victim policy\nnetwork. In multi-agent environments, an attacker can manipulate an adversarial\nopponent to influence the victim policy's observations indirectly. While\nadversarial policies offer a promising technique to craft such attacks, current\nmethods are either sample-inefficient due to poor exploration strategies or\nrequire extra surrogate model training under the black-box assumption. To\naddress these challenges, in this paper, we propose Intrinsically Motivated\nAdversarial Policy (IMAP) for efficient black-box adversarial policy learning\nin both single- and multi-agent environments. We formulate four types of\nadversarial intrinsic regularizers -- maximizing the adversarial state\ncoverage, policy coverage, risk, or divergence -- to discover potential\nvulnerabilities of the victim policy in a principled way. We also present a\nnovel bias-reduction method to balance the extrinsic objective and the\nadversarial intrinsic regularizers adaptively. Our experiments validate the\neffectiveness of the four types of adversarial intrinsic regularizers and the\nbias-reduction method in enhancing black-box adversarial policy learning across\na variety of environments. Our IMAP successfully evades two types of defense\nmethods, adversarial training and robust regularizer, decreasing the\nperformance of the state-of-the-art robust WocaR-PPO agents by 34\\%-54\\% across\nfour single-agent tasks. IMAP also achieves a state-of-the-art attacking\nsuccess rate of 83.91\\% in the multi-agent game YouShallNotPass. Our code is\navailable at \\url{https://github.com/x-zheng16/IMAP}.\n", "date": "2023-05-04", "categories": []}
{"id": "2401.05217", "title": "Exploring Vulnerabilities of No-Reference Image Quality Assessment\n  Models: A Query-Based Black-Box Method", "abstract": "  No-Reference Image Quality Assessment (NR-IQA) aims to predict image quality\nscores consistent with human perception without relying on pristine reference\nimages, serving as a crucial component in various visual tasks. Ensuring the\nrobustness of NR-IQA methods is vital for reliable comparisons of different\nimage processing techniques and consistent user experiences in recommendations.\nThe attack methods for NR-IQA provide a powerful instrument to test the\nrobustness of NR-IQA. However, current attack methods of NR-IQA heavily rely on\nthe gradient of the NR-IQA model, leading to limitations when the gradient\ninformation is unavailable. In this paper, we present a pioneering query-based\nblack box attack against NR-IQA methods. We propose the concept of score\nboundary and leverage an adaptive iterative approach with multiple score\nboundaries. Meanwhile, the initial attack directions are also designed to\nleverage the characteristics of the Human Visual System (HVS). Experiments show\nour method outperforms all compared state-of-the-art attack methods and is far\nahead of previous black-box methods. The effective NR-IQA model DBCNN suffers a\nSpearman's rank-order correlation coefficient (SROCC) decline of 0.6381\nattacked by our method, revealing the vulnerability of NR-IQA models to\nblack-box attacks. The proposed attack method also provides a potent tool for\nfurther exploration into NR-IQA robustness.\n", "date": "2024-01-10", "categories": []}
{"id": "2404.15744", "title": "A General Black-box Adversarial Attack on Graph-based Fake News\n  Detectors", "abstract": "  Graph Neural Network (GNN)-based fake news detectors apply various methods to\nconstruct graphs, aiming to learn distinctive news embeddings for\nclassification. Since the construction details are unknown for attackers in a\nblack-box scenario, it is unrealistic to conduct the classical adversarial\nattacks that require a specific adjacency matrix. In this paper, we propose the\nfirst general black-box adversarial attack framework, i.e., General Attack via\nFake Social Interaction (GAFSI), against detectors based on different graph\nstructures. Specifically, as sharing is an important social interaction for\nGNN-based fake news detectors to construct the graph, we simulate sharing\nbehaviors to fool the detectors. Firstly, we propose a fraudster selection\nmodule to select engaged users leveraging local and global information. In\naddition, a post injection module guides the selected users to create shared\nrelations by sending posts. The sharing records will be added to the social\ncontext, leading to a general attack against different detectors. Experimental\nresults on empirical datasets demonstrate the effectiveness of GAFSI.\n", "date": "2024-04-24", "categories": []}
{"id": "2404.17020", "title": "Generating Minimalist Adversarial Perturbations to Test Object-Detection\n  Models: An Adaptive Multi-Metric Evolutionary Search Approach", "abstract": "  Deep Learning (DL) models excel in computer vision tasks but can be\nsusceptible to adversarial examples. This paper introduces Triple-Metric\nEvoAttack (TM-EVO), an efficient algorithm for evaluating the robustness of\nobject-detection DL models against adversarial attacks. TM-EVO utilizes a\nmulti-metric fitness function to guide an evolutionary search efficiently in\ncreating effective adversarial test inputs with minimal perturbations. We\nevaluate TM-EVO on widely-used object-detection DL models, DETR and Faster\nR-CNN, and open-source datasets, COCO and KITTI. Our findings reveal that\nTM-EVO outperforms the state-of-the-art EvoAttack baseline, leading to\nadversarial tests with less noise while maintaining efficiency.\n", "date": "2024-04-25", "categories": []}
{"id": "2404.17371", "title": "Estimating the Robustness Radius for Randomized Smoothing with\n  100$\\times$ Sample Efficiency", "abstract": "  Randomized smoothing (RS) has successfully been used to improve the\nrobustness of predictions for deep neural networks (DNNs) by adding random\nnoise to create multiple variations of an input, followed by deciding the\nconsensus. To understand if an RS-enabled DNN is effective in the sampled input\ndomains, it is mandatory to sample data points within the operational design\ndomain, acquire the point-wise certificate regarding robustness radius, and\ncompare it with pre-defined acceptance criteria. Consequently, ensuring that a\npoint-wise robustness certificate for any given data point is obtained\nrelatively cost-effectively is crucial. This work demonstrates that reducing\nthe number of samples by one or two orders of magnitude can still enable the\ncomputation of a slightly smaller robustness radius (commonly ~20% radius\nreduction) with the same confidence. We provide the mathematical foundation for\nexplaining the phenomenon while experimentally showing promising results on the\nstandard CIFAR-10 and ImageNet datasets.\n", "date": "2024-04-26", "categories": []}
{"id": "2305.15792", "title": "IDEA: Invariant Defense for Graph Adversarial Robustness", "abstract": "  Despite the success of graph neural networks (GNNs), their vulnerability to\nadversarial attacks poses tremendous challenges for practical applications.\nExisting defense methods suffer from severe performance decline under unseen\nattacks, due to either limited observed adversarial examples or pre-defined\nheuristics. To address these limitations, we analyze the causalities in graph\nadversarial attacks and conclude that causal features are key to achieve graph\nadversarial robustness, owing to their determinedness for labels and invariance\nacross attacks. To learn these causal features, we innovatively propose an\nInvariant causal DEfense method against adversarial Attacks (IDEA). We derive\nnode-based and structure-based invariance objectives from an\ninformation-theoretic perspective. IDEA ensures strong predictability for\nlabels and invariant predictability across attacks, which is provably a\ncausally invariant defense across various attacks. Extensive experiments\ndemonstrate that IDEA attains state-of-the-art defense performance under all\nfive attacks on all five datasets. The implementation of IDEA is available at\nhttps://anonymous.4open.science/r/IDEA.\n", "date": "2023-05-25", "categories": []}
{"id": "2404.16452", "title": "PAD: Patch-Agnostic Defense against Adversarial Patch Attacks", "abstract": "  Adversarial patch attacks present a significant threat to real-world object\ndetectors due to their practical feasibility. Existing defense methods, which\nrely on attack data or prior knowledge, struggle to effectively address a wide\nrange of adversarial patches. In this paper, we show two inherent\ncharacteristics of adversarial patches, semantic independence and spatial\nheterogeneity, independent of their appearance, shape, size, quantity, and\nlocation. Semantic independence indicates that adversarial patches operate\nautonomously within their semantic context, while spatial heterogeneity\nmanifests as distinct image quality of the patch area that differs from\noriginal clean image due to the independent generation process. Based on these\nobservations, we propose PAD, a novel adversarial patch localization and\nremoval method that does not require prior knowledge or additional training.\nPAD offers patch-agnostic defense against various adversarial patches,\ncompatible with any pre-trained object detectors. Our comprehensive digital and\nphysical experiments involving diverse patch types, such as localized noise,\nprintable, and naturalistic patches, exhibit notable improvements over\nstate-of-the-art works. Our code is available at\nhttps://github.com/Lihua-Jing/PAD.\n", "date": "2024-04-25", "categories": []}
{"id": "2404.16457", "title": "Towards Precise Observations of Neural Model Robustness in\n  Classification", "abstract": "  In deep learning applications, robustness measures the ability of neural\nmodels that handle slight changes in input data, which could lead to potential\nsafety hazards, especially in safety-critical applications. Pre-deployment\nassessment of model robustness is essential, but existing methods often suffer\nfrom either high costs or imprecise results. To enhance safety in real-world\nscenarios, metrics that effectively capture the model's robustness are needed.\nTo address this issue, we compare the rigour and usage conditions of various\nassessment methods based on different definitions. Then, we propose a\nstraightforward and practical metric utilizing hypothesis testing for\nprobabilistic robustness and have integrated it into the TorchAttacks library.\nThrough a comparative analysis of diverse robustness assessment methods, our\napproach contributes to a deeper understanding of model robustness in\nsafety-critical applications.\n", "date": "2024-04-25", "categories": []}
{"id": "2404.16557", "title": "Energy-Latency Manipulation of Multi-modal Large Language Models via\n  Verbose Samples", "abstract": "  Despite the exceptional performance of multi-modal large language models\n(MLLMs), their deployment requires substantial computational resources. Once\nmalicious users induce high energy consumption and latency time (energy-latency\ncost), it will exhaust computational resources and harm availability of\nservice. In this paper, we investigate this vulnerability for MLLMs,\nparticularly image-based and video-based ones, and aim to induce high\nenergy-latency cost during inference by crafting an imperceptible perturbation.\nWe find that high energy-latency cost can be manipulated by maximizing the\nlength of generated sequences, which motivates us to propose verbose samples,\nincluding verbose images and videos. Concretely, two modality non-specific\nlosses are proposed, including a loss to delay end-of-sequence (EOS) token and\nan uncertainty loss to increase the uncertainty over each generated token. In\naddition, improving diversity is important to encourage longer responses by\nincreasing the complexity, which inspires the following modality specific loss.\nFor verbose images, a token diversity loss is proposed to promote diverse\nhidden states. For verbose videos, a frame feature diversity loss is proposed\nto increase the feature diversity among frames. To balance these losses, we\npropose a temporal weight adjustment algorithm. Experiments demonstrate that\nour verbose samples can largely extend the length of generated sequences.\n", "date": "2024-04-25", "categories": []}
{"id": "2404.13277", "title": "Beyond Score Changes: Adversarial Attack on No-Reference Image Quality\n  Assessment from Two Perspectives", "abstract": "  Deep neural networks have demonstrated impressive success in No-Reference\nImage Quality Assessment (NR-IQA). However, recent researches highlight the\nvulnerability of NR-IQA models to subtle adversarial perturbations, leading to\ninconsistencies between model predictions and subjective ratings. Current\nadversarial attacks, however, focus on perturbing predicted scores of\nindividual images, neglecting the crucial aspect of inter-score correlation\nrelationships within an entire image set. Meanwhile, it is important to note\nthat the correlation, like ranking correlation, plays a significant role in\nNR-IQA tasks. To comprehensively explore the robustness of NR-IQA models, we\nintroduce a new framework of correlation-error-based attacks that perturb both\nthe correlation within an image set and score changes on individual images. Our\nresearch primarily focuses on ranking-related correlation metrics like\nSpearman's Rank-Order Correlation Coefficient (SROCC) and prediction\nerror-related metrics like Mean Squared Error (MSE). As an instantiation, we\npropose a practical two-stage SROCC-MSE-Attack (SMA) that initially optimizes\ntarget attack scores for the entire image set and then generates adversarial\nexamples guided by these scores. Experimental results demonstrate that our SMA\nmethod not only significantly disrupts the SROCC to negative values but also\nmaintains a considerable change in the scores of individual images. Meanwhile,\nit exhibits state-of-the-art performance across metrics with different\ncategories. Our method provides a new perspective on the robustness of NR-IQA\nmodels.\n", "date": "2024-04-20", "categories": []}
{"id": "2404.15344", "title": "Adversarial Robustness of Distilled and Pruned Deep Learning-based\n  Wireless Classifiers", "abstract": "  Data-driven deep learning (DL) techniques developed for automatic modulation\nclassification (AMC) of wireless signals are vulnerable to adversarial attacks.\nThis poses a severe security threat to the DL-based wireless systems,\nspecifically for edge applications of AMC. In this work, we address the joint\nproblem of developing optimized DL models that are also robust against\nadversarial attacks. This enables efficient and reliable deployment of DL-based\nAMC on edge devices. We first propose two optimized models using knowledge\ndistillation and network pruning, followed by a computationally efficient\nadversarial training process to improve the robustness. Experimental results on\nfive white-box attacks show that the proposed optimized and adversarially\ntrained models can achieve better robustness than the standard (unoptimized)\nmodel. The two optimized models also achieve higher accuracy on clean\n(unattacked) samples, which is essential for the reliability of DL-based\nsolutions at edge applications.\n", "date": "2024-04-11", "categories": []}
{"id": "2404.15784", "title": "An Empirical Study of Aegis", "abstract": "  Bit flipping attacks are one class of attacks on neural networks with\nnumerous defense mechanisms invented to mitigate its potency. Due to the\nimportance of ensuring the robustness of these defense mechanisms, we perform\nan empirical study on the Aegis framework. We evaluate the baseline mechanisms\nof Aegis on low-entropy data (MNIST), and we evaluate a pre-trained model with\nthe mechanisms fine-tuned on MNIST. We also compare the use of data\naugmentation to the robustness training of Aegis, and how Aegis performs under\nother adversarial attacks, such as the generation of adversarial examples. We\nfind that both the dynamic-exit strategy and robustness training of Aegis has\nsome drawbacks. In particular, we see drops in accuracy when testing on\nperturbed data, and on adversarial examples, as compared to baselines.\nMoreover, we found that the dynamic exit-strategy loses its uniformity when\ntested on simpler datasets. The code for this project is available on GitHub.\n", "date": "2024-04-24", "categories": []}
{"id": "2404.15881", "title": "Steal Now and Attack Later: Evaluating Robustness of Object Detection\n  against Black-box Adversarial Attacks", "abstract": "  Latency attacks against object detection represent a variant of adversarial\nattacks that aim to inflate the inference time by generating additional ghost\nobjects in a target image. However, generating ghost objects in the black-box\nscenario remains a challenge since information about these unqualified objects\nremains opaque. In this study, we demonstrate the feasibility of generating\nghost objects in adversarial examples by extending the concept of \"steal now,\ndecrypt later\" attacks. These adversarial examples, once produced, can be\nemployed to exploit potential vulnerabilities in the AI service, giving rise to\nsignificant security concerns. The experimental results demonstrate that the\nproposed attack achieves successful attacks across various commonly used models\nand Google Vision API without any prior knowledge about the target model.\nAdditionally, the average cost of each attack is less than \\$ 1 dollars, posing\na significant threat to AI security.\n", "date": "2024-04-24", "categories": []}
{"id": "2403.11830", "title": "Problem space structural adversarial attacks for Network Intrusion\n  Detection Systems based on Graph Neural Networks", "abstract": "  Machine Learning (ML) algorithms have become increasingly popular for\nsupporting Network Intrusion Detection Systems (NIDS). Nevertheless, extensive\nresearch has shown their vulnerability to adversarial attacks, which involve\nsubtle perturbations to the inputs of the models aimed at compromising their\nperformance. Recent proposals have effectively leveraged Graph Neural Networks\n(GNN) to produce predictions based also on the structural patterns exhibited by\nintrusions to enhance the detection robustness. However, the adoption of\nGNN-based NIDS introduces new types of risks. In this paper, we propose the\nfirst formalization of adversarial attacks specifically tailored for GNN in\nnetwork intrusion detection. Moreover, we outline and model the problem space\nconstraints that attackers need to consider to carry out feasible structural\nattacks in real-world scenarios. As a final contribution, we conduct an\nextensive experimental campaign in which we launch the proposed attacks against\nstate-of-the-art GNN-based NIDS. Our findings demonstrate the increased\nrobustness of the models against classical feature-based adversarial attacks,\nwhile highlighting their susceptibility to structure-based attacks.\n", "date": "2024-03-18", "categories": []}
{"id": "2212.03095", "title": "Interpretation of Neural Networks is Susceptible to Universal\n  Adversarial Perturbations", "abstract": "  Interpreting neural network classifiers using gradient-based saliency maps\nhas been extensively studied in the deep learning literature. While the\nexisting algorithms manage to achieve satisfactory performance in application\nto standard image recognition datasets, recent works demonstrate the\nvulnerability of widely-used gradient-based interpretation schemes to\nnorm-bounded perturbations adversarially designed for every individual input\nsample. However, such adversarial perturbations are commonly designed using the\nknowledge of an input sample, and hence perform sub-optimally in application to\nan unknown or constantly changing data point. In this paper, we show the\nexistence of a Universal Perturbation for Interpretation (UPI) for standard\nimage datasets, which can alter a gradient-based feature map of neural networks\nover a significant fraction of test samples. To design such a UPI, we propose a\ngradient-based optimization method as well as a principal component analysis\n(PCA)-based approach to compute a UPI which can effectively alter a neural\nnetwork's gradient-based interpretation on different samples. We support the\nproposed UPI approaches by presenting several numerical results of their\nsuccessful applications to standard image datasets.\n", "date": "2022-11-30", "categories": []}
{"id": "2401.04727", "title": "Revisiting Adversarial Training at Scale", "abstract": "  The machine learning community has witnessed a drastic change in the training\npipeline, pivoted by those ''foundation models'' with unprecedented scales.\nHowever, the field of adversarial training is lagging behind, predominantly\ncentered around small model sizes like ResNet-50, and tiny and low-resolution\ndatasets like CIFAR-10. To bridge this transformation gap, this paper provides\na modern re-examination with adversarial training, investigating its potential\nbenefits when applied at scale. Additionally, we introduce an efficient and\neffective training strategy to enable adversarial training with giant models\nand web-scale data at an affordable computing cost. We denote this newly\nintroduced framework as AdvXL.\n  Empirical results demonstrate that AdvXL establishes new state-of-the-art\nrobust accuracy records under AutoAttack on ImageNet-1K. For example, by\ntraining on DataComp-1B dataset, our AdvXL empowers a vanilla ViT-g model to\nsubstantially surpass the previous records of $l_{\\infty}$-, $l_{2}$-, and\n$l_{1}$-robust accuracy by margins of 11.4%, 14.2% and 12.9%, respectively.\nThis achievement posits AdvXL as a pioneering approach, charting a new\ntrajectory for the efficient training of robust visual representations at\nsignificantly larger scales. Our code is available at\nhttps://github.com/UCSC-VLAA/AdvXL.\n", "date": "2024-01-09", "categories": []}
{"id": "2402.12338", "title": "An Adversarial Approach to Evaluating the Robustness of Event\n  Identification Models", "abstract": "  Intelligent machine learning approaches are finding active use for event\ndetection and identification that allow real-time situational awareness. Yet,\nsuch machine learning algorithms have been shown to be susceptible to\nadversarial attacks on the incoming telemetry data. This paper considers a\nphysics-based modal decomposition method to extract features for event\nclassification and focuses on interpretable classifiers including logistic\nregression and gradient boosting to distinguish two types of events: load loss\nand generation loss. The resulting classifiers are then tested against an\nadversarial algorithm to evaluate their robustness. The adversarial attack is\ntested in two settings: the white box setting, wherein the attacker knows\nexactly the classification model; and the gray box setting, wherein the\nattacker has access to historical data from the same network as was used to\ntrain the classifier, but does not know the classification model. Thorough\nexperiments on the synthetic South Carolina 500-bus system highlight that a\nrelatively simpler model such as logistic regression is more susceptible to\nadversarial attacks than gradient boosting.\n", "date": "2024-02-19", "categories": []}
{"id": "2403.18985", "title": "Robustness and Visual Explanation for Black Box Image, Video, and ECG\n  Signal Classification with Reinforcement Learning", "abstract": "  We present a generic Reinforcement Learning (RL) framework optimized for\ncrafting adversarial attacks on different model types spanning from ECG signal\nanalysis (1D), image classification (2D), and video classification (3D). The\nframework focuses on identifying sensitive regions and inducing\nmisclassifications with minimal distortions and various distortion types. The\nnovel RL method outperforms state-of-the-art methods for all three\napplications, proving its efficiency. Our RL approach produces superior\nlocalization masks, enhancing interpretability for image classification and ECG\nanalysis models. For applications such as ECG analysis, our platform highlights\ncritical ECG segments for clinicians while ensuring resilience against\nprevalent distortions. This comprehensive tool aims to bolster both resilience\nwith adversarial training and transparency across varied applications and data\ntypes.\n", "date": "2024-03-27", "categories": []}
{"id": "2404.13706", "title": "Concept Arithmetics for Circumventing Concept Inhibition in Diffusion\n  Models", "abstract": "  Motivated by ethical and legal concerns, the scientific community is actively\ndeveloping methods to limit the misuse of Text-to-Image diffusion models for\nreproducing copyrighted, violent, explicit, or personal information in the\ngenerated images. Simultaneously, researchers put these newly developed safety\nmeasures to the test by assuming the role of an adversary to find\nvulnerabilities and backdoors in them. We use compositional property of\ndiffusion models, which allows to leverage multiple prompts in a single image\ngeneration. This property allows us to combine other concepts, that should not\nhave been affected by the inhibition, to reconstruct the vector, responsible\nfor target concept generation, even though the direct computation of this\nvector is no longer accessible. We provide theoretical and empirical evidence\nwhy the proposed attacks are possible and discuss the implications of these\nfindings for safe model deployment. We argue that it is essential to consider\nall possible approaches to image generation with diffusion models that can be\nemployed by an adversary. Our work opens up the discussion about the\nimplications of concept arithmetics and compositional inference for safety\nmechanisms in diffusion models.\n  Content Advisory: This paper contains discussions and model-generated content\nthat may be considered offensive. Reader discretion is advised.\n  Project page: https://cs-people.bu.edu/vpetsiuk/arc\n", "date": "2024-04-21", "categories": []}
{"id": "2404.13860", "title": "Distributional Black-Box Model Inversion Attack with Multi-Agent\n  Reinforcement Learning", "abstract": "  A Model Inversion (MI) attack based on Generative Adversarial Networks (GAN)\naims to recover the private training data from complex deep learning models by\nsearching codes in the latent space. However, they merely search a\ndeterministic latent space such that the found latent code is usually\nsuboptimal. In addition, the existing distributional MI schemes assume that an\nattacker can access the structures and parameters of the target model, which is\nnot always viable in practice. To overcome the above shortcomings, this paper\nproposes a novel Distributional Black-Box Model Inversion (DBB-MI) attack by\nconstructing the probabilistic latent space for searching the target privacy\ndata. Specifically, DBB-MI does not need the target model parameters or\nspecialized GAN training. Instead, it finds the latent probability distribution\nby combining the output of the target model with multi-agent reinforcement\nlearning techniques. Then, it randomly chooses latent codes from the latent\nprobability distribution for recovering the private data. As the latent\nprobability distribution closely aligns with the target privacy data in latent\nspace, the recovered data will leak the privacy of training samples of the\ntarget model significantly. Abundant experiments conducted on diverse datasets\nand networks show that the present DBB-MI has better performance than\nstate-of-the-art in attack accuracy, K-nearest neighbor feature distance, and\nPeak Signal-to-Noise Ratio.\n", "date": "2024-04-22", "categories": []}
{"id": "2306.08386", "title": "Efficient Backdoor Attacks for Deep Neural Networks in Real-world\n  Scenarios", "abstract": "  Recent deep neural networks (DNNs) have came to rely on vast amounts of\ntraining data, providing an opportunity for malicious attackers to exploit and\ncontaminate the data to carry out backdoor attacks. However, existing backdoor\nattack methods make unrealistic assumptions, assuming that all training data\ncomes from a single source and that attackers have full access to the training\ndata. In this paper, we introduce a more realistic attack scenario where\nvictims collect data from multiple sources, and attackers cannot access the\ncomplete training data. We refer to this scenario as data-constrained backdoor\nattacks. In such cases, previous attack methods suffer from severe efficiency\ndegradation due to the entanglement between benign and poisoning features\nduring the backdoor injection process. To tackle this problem, we introduce\nthree CLIP-based technologies from two distinct streams: Clean Feature\nSuppression and Poisoning Feature Augmentation.effective solution for\ndata-constrained backdoor attacks. The results demonstrate remarkable\nimprovements, with some settings achieving over 100% improvement compared to\nexisting attacks in data-constrained scenarios. Code is available at\nhttps://github.com/sunh1113/Efficient-backdoor-attacks-for-deep-neural-networks-in-real-world-scenarios\n", "date": "2023-06-14", "categories": []}
{"id": "2404.09401", "title": "Watermark-embedded Adversarial Examples for Copyright Protection against\n  Diffusion Models", "abstract": "  Diffusion Models (DMs) have shown remarkable capabilities in various\nimage-generation tasks. However, there are growing concerns that DMs could be\nused to imitate unauthorized creations and thus raise copyright issues. To\naddress this issue, we propose a novel framework that embeds personal\nwatermarks in the generation of adversarial examples. Such examples can force\nDMs to generate images with visible watermarks and prevent DMs from imitating\nunauthorized images. We construct a generator based on conditional adversarial\nnetworks and design three losses (adversarial loss, GAN loss, and perturbation\nloss) to generate adversarial examples that have subtle perturbation but can\neffectively attack DMs to prevent copyright violations. Training a generator\nfor a personal watermark by our method only requires 5-10 samples within 2-3\nminutes, and once the generator is trained, it can generate adversarial\nexamples with that watermark significantly fast (0.2s per image). We conduct\nextensive experiments in various conditional image-generation scenarios.\nCompared to existing methods that generate images with chaotic textures, our\nmethod adds visible watermarks on the generated images, which is a more\nstraightforward way to indicate copyright violations. We also observe that our\nadversarial examples exhibit good transferability across unknown generative\nmodels. Therefore, this work provides a simple yet powerful way to protect\ncopyright from DM-based imitation.\n", "date": "2024-04-15", "categories": []}
{"id": "2404.12612", "title": "SA-Attack: Speed-adaptive stealthy adversarial attack on trajectory\n  prediction", "abstract": "  Trajectory prediction is critical for the safe planning and navigation of\nautomated vehicles. The trajectory prediction models based on the neural\nnetworks are vulnerable to adversarial attacks. Previous attack methods have\nachieved high attack success rates but overlook the adaptability to realistic\nscenarios and the concealment of the deceits. To address this problem, we\npropose a speed-adaptive stealthy adversarial attack method named SA-Attack.\nThis method searches the sensitive region of trajectory prediction models and\ngenerates the adversarial trajectories by using the vehicle-following method\nand incorporating information about forthcoming trajectories. Our method has\nthe ability to adapt to different speed scenarios by reconstructing the\ntrajectory from scratch. Fusing future trajectory trends and curvature\nconstraints can guarantee the smoothness of adversarial trajectories, further\nensuring the stealthiness of attacks. The empirical study on the datasets of\nnuScenes and Apolloscape demonstrates the attack performance of our proposed\nmethod. Finally, we also demonstrate the adaptability and stealthiness of\nSA-Attack for different speed scenarios. Our code is available at the\nrepository: https://github.com/eclipse-bot/SA-Attack.\n", "date": "2024-04-19", "categories": []}
{"id": "2404.12653", "title": "How Real Is Real? A Human Evaluation Framework for Unrestricted\n  Adversarial Examples", "abstract": "  With an ever-increasing reliance on machine learning (ML) models in the real\nworld, adversarial examples threaten the safety of AI-based systems such as\nautonomous vehicles. In the image domain, they represent maliciously perturbed\ndata points that look benign to humans (i.e., the image modification is not\nnoticeable) but greatly mislead state-of-the-art ML models. Previously,\nresearchers ensured the imperceptibility of their altered data points by\nrestricting perturbations via $\\ell_p$ norms. However, recent publications\nclaim that creating natural-looking adversarial examples without such\nrestrictions is also possible. With much more freedom to instill malicious\ninformation into data, these unrestricted adversarial examples can potentially\novercome traditional defense strategies as they are not constrained by the\nlimitations or patterns these defenses typically recognize and mitigate. This\nallows attackers to operate outside of expected threat models. However,\nsurveying existing image-based methods, we noticed a need for more human\nevaluations of the proposed image modifications. Based on existing\nhuman-assessment frameworks for image generation quality, we propose SCOOTER -\nan evaluation framework for unrestricted image-based attacks. It provides\nresearchers with guidelines for conducting statistically significant human\nexperiments, standardized questions, and a ready-to-use implementation. We\npropose a framework that allows researchers to analyze how imperceptible their\nunrestricted attacks truly are.\n", "date": "2024-04-19", "categories": []}
{"id": "2208.10878", "title": "Transferability Ranking of Adversarial Examples", "abstract": "  Adversarial transferability in black-box scenarios presents a unique\nchallenge: while attackers can employ surrogate models to craft adversarial\nexamples, they lack assurance on whether these examples will successfully\ncompromise the target model. Until now, the prevalent method to ascertain\nsuccess has been trial and error-testing crafted samples directly on the victim\nmodel. This approach, however, risks detection with every attempt, forcing\nattackers to either perfect their first try or face exposure. Our paper\nintroduces a ranking strategy that refines the transfer attack process,\nenabling the attacker to estimate the likelihood of success without repeated\ntrials on the victim's system. By leveraging a set of diverse surrogate models,\nour method can predict transferability of adversarial examples. This strategy\ncan be used to either select the best sample to use in an attack or the best\nperturbation to apply to a specific sample. Using our strategy, we were able to\nraise the transferability of adversarial examples from a mere 20% - akin to\nrandom selection-up to near upper-bound levels, with some scenarios even\nwitnessing a 100% success rate. This substantial improvement not only sheds\nlight on the shared susceptibilities across diverse architectures but also\ndemonstrates that attackers can forego the detectable trial-and-error tactics\nraising increasing the threat of surrogate-based attacks.\n", "date": "2022-08-23", "categories": []}
{"id": "2402.06244", "title": "Quantifying and Enhancing Multi-modal Robustness with Modality\n  Preference", "abstract": "  Multi-modal models have shown a promising capability to effectively integrate\ninformation from various sources, yet meanwhile, they are found vulnerable to\npervasive perturbations, such as uni-modal attacks and missing conditions. To\ncounter these perturbations, robust multi-modal representations are highly\nexpected, which are positioned well away from the discriminative multi-modal\ndecision boundary. In this paper, different from conventional empirical\nstudies, we focus on a commonly used joint multi-modal framework and\ntheoretically discover that larger uni-modal representation margins and more\nreliable integration for modalities are essential components for achieving\nhigher robustness. This discovery can further explain the limitation of\nmulti-modal robustness and the phenomenon that multi-modal models are often\nvulnerable to attacks on the specific modality. Moreover, our analysis reveals\nhow the widespread issue, that the model has different preferences for\nmodalities, limits the multi-modal robustness by influencing the essential\ncomponents and could lead to attacks on the specific modality highly effective.\nInspired by our theoretical finding, we introduce a training procedure called\nCertifiable Robust Multi-modal Training (CRMT), which can alleviate this\ninfluence from modality preference and explicitly regulate essential components\nto significantly improve robustness in a certifiable manner. Our method\ndemonstrates substantial improvements in performance and robustness compared\nwith existing methods. Furthermore, our training procedure can be easily\nextended to enhance other robust training strategies, highlighting its\ncredibility and flexibility.\n", "date": "2024-02-09", "categories": []}
{"id": "2404.12014", "title": "Enhance Robustness of Language Models Against Variation Attack through\n  Graph Integration", "abstract": "  The widespread use of pre-trained language models (PLMs) in natural language\nprocessing (NLP) has greatly improved performance outcomes. However, these\nmodels' vulnerability to adversarial attacks (e.g., camouflaged hints from drug\ndealers), particularly in the Chinese language with its rich character\ndiversity/variation and complex structures, hatches vital apprehension. In this\nstudy, we propose a novel method, CHinese vAriatioN Graph Enhancement (CHANGE),\nto increase the robustness of PLMs against character variation attacks in\nChinese content. CHANGE presents a novel approach for incorporating a Chinese\ncharacter variation graph into the PLMs. Through designing different\nsupplementary tasks utilizing the graph structure, CHANGE essentially enhances\nPLMs' interpretation of adversarially manipulated text. Experiments conducted\nin a multitude of NLP tasks show that CHANGE outperforms current language\nmodels in combating against adversarial attacks and serves as a valuable\ncontribution to robust language model research. These findings contribute to\nthe groundwork on robust language models and highlight the substantial\npotential of graph-guided pre-training strategies for real-world applications.\n", "date": "2024-04-18", "categories": []}
{"id": "2404.12274", "title": "Advancing the Robustness of Large Language Models through Self-Denoised\n  Smoothing", "abstract": "  Although large language models (LLMs) have achieved significant success,\ntheir vulnerability to adversarial perturbations, including recent jailbreak\nattacks, has raised considerable concerns. However, the increasing size of\nthese models and their limited access make improving their robustness a\nchallenging task. Among various defense strategies, randomized smoothing has\nshown great potential for LLMs, as it does not require full access to the\nmodel's parameters or fine-tuning via adversarial training. However, randomized\nsmoothing involves adding noise to the input before model prediction, and the\nfinal model's robustness largely depends on the model's performance on these\nnoise corrupted data. Its effectiveness is often limited by the model's\nsub-optimal performance on noisy data. To address this issue, we propose to\nleverage the multitasking nature of LLMs to first denoise the noisy inputs and\nthen to make predictions based on these denoised versions. We call this\nprocedure self-denoised smoothing. Unlike previous denoised smoothing\ntechniques in computer vision, which require training a separate model to\nenhance the robustness of LLMs, our method offers significantly better\nefficiency and flexibility. Our experimental results indicate that our method\nsurpasses existing methods in both empirical and certified robustness in\ndefending against adversarial attacks for both downstream tasks and human\nalignments (i.e., jailbreak attacks). Our code is publicly available at\nhttps://github.com/UCSB-NLP-Chang/SelfDenoise\n", "date": "2024-04-18", "categories": []}
{"id": "2303.12054", "title": "Influencer Backdoor Attack on Semantic Segmentation", "abstract": "  When a small number of poisoned samples are injected into the training\ndataset of a deep neural network, the network can be induced to exhibit\nmalicious behavior during inferences, which poses potential threats to\nreal-world applications. While they have been intensively studied in\nclassification, backdoor attacks on semantic segmentation have been largely\noverlooked. Unlike classification, semantic segmentation aims to classify every\npixel within a given image. In this work, we explore backdoor attacks on\nsegmentation models to misclassify all pixels of a victim class by injecting a\nspecific trigger on non-victim pixels during inferences, which is dubbed\nInfluencer Backdoor Attack (IBA). IBA is expected to maintain the\nclassification accuracy of non-victim pixels and mislead classifications of all\nvictim pixels in every single inference and could be easily applied to\nreal-world scenes. Based on the context aggregation ability of segmentation\nmodels, we proposed a simple, yet effective, Nearest-Neighbor trigger injection\nstrategy. We also introduce an innovative Pixel Random Labeling strategy which\nmaintains optimal performance even when the trigger is placed far from the\nvictim pixels. Our extensive experiments reveal that current segmentation\nmodels do suffer from backdoor attacks, demonstrate IBA real-world\napplicability, and show that our proposed techniques can further increase\nattack performance.\n", "date": "2023-03-21", "categories": []}
{"id": "2307.03798", "title": "Fooling Contrastive Language-Image Pre-trained Models with\n  CLIPMasterPrints", "abstract": "  Models leveraging both visual and textual data such as Contrastive\nLanguage-Image Pre-training (CLIP), are the backbone of many recent advances in\nartificial intelligence. In this work, we show that despite their versatility,\nsuch models are vulnerable to what we refer to as fooling master images.\nFooling master images are capable of maximizing the confidence score of a CLIP\nmodel for a significant number of widely varying prompts, while being either\nunrecognizable or unrelated to the attacked prompts for humans. The existence\nof such images is problematic as it could be used by bad actors to maliciously\ninterfere with CLIP-trained image retrieval models in production with\ncomparably small effort as a single image can attack many different prompts. We\ndemonstrate how fooling master images for CLIP (CLIPMasterPrints) can be mined\nusing stochastic gradient descent, projected gradient descent, or blackbox\noptimization. Contrary to many common adversarial attacks, the blackbox\noptimization approach allows us to mine CLIPMasterPrints even when the weights\nof the model are not accessible. We investigate the properties of the mined\nimages, and find that images trained on a small number of image captions\ngeneralize to a much larger number of semantically related captions. We\nevaluate possible mitigation strategies, where we increase the robustness of\nthe model and introduce an approach to automatically detect CLIPMasterPrints to\nsanitize the input of vulnerable models. Finally, we find that vulnerability to\nCLIPMasterPrints is related to a modality gap in contrastive pre-trained\nmulti-modal networks. Code available at\nhttps://github.com/matfrei/CLIPMasterPrints.\n", "date": "2023-07-07", "categories": []}
{"id": "2403.09863", "title": "Towards White Box Deep Learning", "abstract": "  Deep neural networks learn fragile \"shortcut\" features, rendering them\ndifficult to interpret (black box) and vulnerable to adversarial attacks. This\npaper proposes semantic features as a general architectural solution to this\nproblem. The main idea is to make features locality-sensitive in the adequate\nsemantic topology of the domain, thus introducing a strong regularization. The\nproof of concept network is lightweight, inherently interpretable and achieves\nalmost human-level adversarial test metrics - with no adversarial training!\nThese results and the general nature of the approach warrant further research\non semantic features. The code is available at\nhttps://github.com/314-Foundation/white-box-nn\n", "date": "2024-03-14", "categories": []}
{"id": "2404.04849", "title": "Hidden You Malicious Goal Into Benign Narratives: Jailbreak Large\n  Language Models through Logic Chain Injection", "abstract": "  Jailbreak attacks on Language Model Models (LLMs) entail crafting prompts\naimed at exploiting the models to generate malicious content. Existing\njailbreak attacks can successfully deceive the LLMs, however they cannot\ndeceive the human. This paper proposes a new type of jailbreak attacks which\ncan deceive both the LLMs and human (i.e., security analyst). The key insight\nof our idea is borrowed from the social psychology - that is human are easily\ndeceived if the lie is hidden in truth. Based on this insight, we proposed the\nlogic-chain injection attacks to inject malicious intention into benign truth.\nLogic-chain injection attack firstly dissembles its malicious target into a\nchain of benign narrations, and then distribute narrations into a related\nbenign article, with undoubted facts. In this way, newly generate prompt cannot\nonly deceive the LLMs, but also deceive human.\n", "date": "2024-04-07", "categories": []}
{"id": "2404.10588", "title": "Do Counterfactual Examples Complicate Adversarial Training?", "abstract": "  We leverage diffusion models to study the robustness-performance tradeoff of\nrobust classifiers. Our approach introduces a simple, pretrained diffusion\nmethod to generate low-norm counterfactual examples (CEs): semantically altered\ndata which results in different true class membership. We report that the\nconfidence and accuracy of robust models on their clean training data are\nassociated with the proximity of the data to their CEs. Moreover, robust models\nperform very poorly when evaluated on the CEs directly, as they become\nincreasingly invariant to the low-norm, semantic changes brought by CEs. The\nresults indicate a significant overlap between non-robust and semantic\nfeatures, countering the common assumption that non-robust features are not\ninterpretable.\n", "date": "2024-04-16", "categories": []}
{"id": "2404.10790", "title": "Multimodal Attack Detection for Action Recognition Models", "abstract": "  Adversarial machine learning attacks on video action recognition models is a\ngrowing research area and many effective attacks were introduced in recent\nyears. These attacks show that action recognition models can be breached in\nmany ways. Hence using these models in practice raises significant security\nconcerns. However, there are very few works which focus on defending against or\ndetecting attacks. In this work, we propose a novel universal detection method\nwhich is compatible with any action recognition model. In our extensive\nexperiments, we show that our method consistently detects various attacks\nagainst different target models with high true positive rates while satisfying\nvery low false positive rates. Tested against four state-of-the-art attacks\ntargeting four action recognition models, the proposed detector achieves an\naverage AUC of 0.911 over 16 test cases while the best performance achieved by\nthe existing detectors is 0.645 average AUC. This 41.2% improvement is enabled\nby the robustness of the proposed detector to varying attack methods and target\nmodels. The lowest AUC achieved by our detector across the 16 test cases is\n0.837 while the competing detector's performance drops as low as 0.211. We also\nshow that the proposed detector is robust to varying attack strengths. In\naddition, we analyze our method's real-time performance with different hardware\nsetups to demonstrate its potential as a practical defense mechanism.\n", "date": "2024-04-13", "categories": []}
{"id": "2404.10796", "title": "Black-box Adversarial Transferability: An Empirical Study in\n  Cybersecurity Perspective", "abstract": "  The rapid advancement of artificial intelligence within the realm of\ncybersecurity raises significant security concerns. The vulnerability of deep\nlearning models in adversarial attacks is one of the major issues. In\nadversarial machine learning, malicious users try to fool the deep learning\nmodel by inserting adversarial perturbation inputs into the model during its\ntraining or testing phase. Subsequently, it reduces the model confidence score\nand results in incorrect classifications. The novel key contribution of the\nresearch is to empirically test the black-box adversarial transferability\nphenomena in cyber attack detection systems. It indicates that the adversarial\nperturbation input generated through the surrogate model has a similar impact\non the target model in producing the incorrect classification. To empirically\nvalidate this phenomenon, surrogate and target models are used. The adversarial\nperturbation inputs are generated based on the surrogate-model for which the\nhacker has complete information. Based on these adversarial perturbation\ninputs, both surrogate and target models are evaluated during the inference\nphase. We have done extensive experimentation over the CICDDoS-2019 dataset,\nand the results are classified in terms of various performance metrics like\naccuracy, precision, recall, and f1-score. The findings indicate that any deep\nlearning model is highly susceptible to adversarial attacks, even if the\nattacker does not have access to the internal details of the target model. The\nresults also indicate that white-box adversarial attacks have a severe impact\ncompared to black-box adversarial attacks. There is a need to investigate and\nexplore adversarial defence techniques to increase the robustness of the deep\nlearning models against adversarial attacks.\n", "date": "2024-04-15", "categories": []}
{"id": "2404.11538", "title": "GenFighter: A Generative and Evolutive Textual Attack Removal", "abstract": "  Adversarial attacks pose significant challenges to deep neural networks\n(DNNs) such as Transformer models in natural language processing (NLP). This\npaper introduces a novel defense strategy, called GenFighter, which enhances\nadversarial robustness by learning and reasoning on the training classification\ndistribution. GenFighter identifies potentially malicious instances deviating\nfrom the distribution, transforms them into semantically equivalent instances\naligned with the training data, and employs ensemble techniques for a unified\nand robust response. By conducting extensive experiments, we show that\nGenFighter outperforms state-of-the-art defenses in accuracy under attack and\nattack success rate metrics. Additionally, it requires a high number of queries\nper attack, making the attack more challenging in real scenarios. The ablation\nstudy shows that our approach integrates transfer learning, a\ngenerative/evolutive procedure, and an ensemble method, providing an effective\ndefense against NLP adversarial attacks.\n", "date": "2024-04-17", "categories": []}
{"id": "2404.10202", "title": "Towards a Novel Perspective on Adversarial Examples Driven by Frequency", "abstract": "  Enhancing our understanding of adversarial examples is crucial for the secure\napplication of machine learning models in real-world scenarios. A prevalent\nmethod for analyzing adversarial examples is through a frequency-based\napproach. However, existing research indicates that attacks designed to exploit\nlow-frequency or high-frequency information can enhance attack performance,\nleading to an unclear relationship between adversarial perturbations and\ndifferent frequency components. In this paper, we seek to demystify this\nrelationship by exploring the characteristics of adversarial perturbations\nwithin the frequency domain. We employ wavelet packet decomposition for\ndetailed frequency analysis of adversarial examples and conduct statistical\nexaminations across various frequency bands. Intriguingly, our findings\nindicate that significant adversarial perturbations are present within the\nhigh-frequency components of low-frequency bands. Drawing on this insight, we\npropose a black-box adversarial attack algorithm based on combining different\nfrequency bands. Experiments conducted on multiple datasets and models\ndemonstrate that combining low-frequency bands and high-frequency components of\nlow-frequency bands can significantly enhance attack efficiency. The average\nattack success rate reaches 99\\%, surpassing attacks that utilize a single\nfrequency segment. Additionally, we introduce the normalized disturbance\nvisibility index as a solution to the limitations of $L_2$ norm in assessing\ncontinuous and discrete perturbations.\n", "date": "2024-04-16", "categories": []}
{"id": "2302.02568", "title": "Less is More: Understanding Word-level Textual Adversarial Attack via\n  n-gram Frequency Descend", "abstract": "  Word-level textual adversarial attacks have demonstrated notable efficacy in\nmisleading Natural Language Processing (NLP) models. Despite their success, the\nunderlying reasons for their effectiveness and the fundamental characteristics\nof adversarial examples (AEs) remain obscure. This work aims to interpret\nword-level attacks by examining their $n$-gram frequency patterns. Our\ncomprehensive experiments reveal that in approximately 90\\% of cases,\nword-level attacks lead to the generation of examples where the frequency of\n$n$-grams decreases, a tendency we term as the $n$-gram Frequency Descend\n($n$-FD). This finding suggests a straightforward strategy to enhance model\nrobustness: training models using examples with $n$-FD. To examine the\nfeasibility of this strategy, we employed the $n$-gram frequency information,\nas an alternative to conventional loss gradients, to generate perturbed\nexamples in adversarial training. The experiment results indicate that the\nfrequency-based approach performs comparably with the gradient-based approach\nin improving model robustness. Our research offers a novel and more intuitive\nperspective for understanding word-level textual adversarial attacks and\nproposes a new direction to improve model robustness.\n", "date": "2023-02-06", "categories": []}
{"id": "2302.09923", "title": "Prompt Stealing Attacks Against Text-to-Image Generation Models", "abstract": "  Text-to-Image generation models have revolutionized the artwork design\nprocess and enabled anyone to create high-quality images by entering text\ndescriptions called prompts. Creating a high-quality prompt that consists of a\nsubject and several modifiers can be time-consuming and costly. In consequence,\na trend of trading high-quality prompts on specialized marketplaces has\nemerged. In this paper, we perform the first study on understanding the threat\nof a novel attack, namely prompt stealing attack, which aims to steal prompts\nfrom generated images by text-to-image generation models. Successful prompt\nstealing attacks directly violate the intellectual property of prompt engineers\nand jeopardize the business model of prompt marketplaces. We first perform a\nsystematic analysis on a dataset collected by ourselves and show that a\nsuccessful prompt stealing attack should consider a prompt's subject as well as\nits modifiers. Based on this observation, we propose a simple yet effective\nprompt stealing attack, PromptStealer. It consists of two modules: a subject\ngenerator trained to infer the subject and a modifier detector for identifying\nthe modifiers within the generated image. Experimental results demonstrate that\nPromptStealer is superior over three baseline methods, both quantitatively and\nqualitatively. We also make some initial attempts to defend PromptStealer. In\ngeneral, our study uncovers a new attack vector within the ecosystem\nestablished by the popular text-to-image generation models. We hope our results\ncan contribute to understanding and mitigating this emerging threat.\n", "date": "2023-02-20", "categories": []}
{"id": "2305.12118", "title": "Annealing Self-Distillation Rectification Improves Adversarial Training", "abstract": "  In standard adversarial training, models are optimized to fit one-hot labels\nwithin allowable adversarial perturbation budgets. However, the ignorance of\nunderlying distribution shifts brought by perturbations causes the problem of\nrobust overfitting. To address this issue and enhance adversarial robustness,\nwe analyze the characteristics of robust models and identify that robust models\ntend to produce smoother and well-calibrated outputs. Based on the observation,\nwe propose a simple yet effective method, Annealing Self-Distillation\nRectification (ADR), which generates soft labels as a better guidance mechanism\nthat accurately reflects the distribution shift under attack during adversarial\ntraining. By utilizing ADR, we can obtain rectified distributions that\nsignificantly improve model robustness without the need for pre-trained models\nor extensive extra computation. Moreover, our method facilitates seamless\nplug-and-play integration with other adversarial training techniques by\nreplacing the hard labels in their objectives. We demonstrate the efficacy of\nADR through extensive experiments and strong performances across datasets.\n", "date": "2023-05-20", "categories": []}
{"id": "2310.11890", "title": "IRAD: Implicit Representation-driven Image Resampling against\n  Adversarial Attacks", "abstract": "  We introduce a novel approach to counter adversarial attacks, namely, image\nresampling. Image resampling transforms a discrete image into a new one,\nsimulating the process of scene recapturing or rerendering as specified by a\ngeometrical transformation. The underlying rationale behind our idea is that\nimage resampling can alleviate the influence of adversarial perturbations while\npreserving essential semantic information, thereby conferring an inherent\nadvantage in defending against adversarial attacks. To validate this concept,\nwe present a comprehensive study on leveraging image resampling to defend\nagainst adversarial attacks. We have developed basic resampling methods that\nemploy interpolation strategies and coordinate shifting magnitudes. Our\nanalysis reveals that these basic methods can partially mitigate adversarial\nattacks. However, they come with apparent limitations: the accuracy of clean\nimages noticeably decreases, while the improvement in accuracy on adversarial\nexamples is not substantial. We propose implicit representation-driven image\nresampling (IRAD) to overcome these limitations. First, we construct an\nimplicit continuous representation that enables us to represent any input image\nwithin a continuous coordinate space. Second, we introduce SampleNet, which\nautomatically generates pixel-wise shifts for resampling in response to\ndifferent inputs. Furthermore, we can extend our approach to the\nstate-of-the-art diffusion-based method, accelerating it with fewer time steps\nwhile preserving its defense capability. Extensive experiments demonstrate that\nour method significantly enhances the adversarial robustness of diverse deep\nmodels against various attacks while maintaining high accuracy on clean images.\n", "date": "2023-10-18", "categories": []}
{"id": "2404.08690", "title": "Towards Building a Robust Toxicity Predictor", "abstract": "  Recent NLP literature pays little attention to the robustness of toxicity\nlanguage predictors, while these systems are most likely to be used in\nadversarial contexts. This paper presents a novel adversarial attack,\n\\texttt{ToxicTrap}, introducing small word-level perturbations to fool SOTA\ntext classifiers to predict toxic text samples as benign. ToxicTrap exploits\ngreedy based search strategies to enable fast and effective generation of toxic\nadversarial examples. Two novel goal function designs allow ToxicTrap to\nidentify weaknesses in both multiclass and multilabel toxic language detectors.\nOur empirical results show that SOTA toxicity text classifiers are indeed\nvulnerable to the proposed attacks, attaining over 98\\% attack success rates in\nmultilabel cases. We also show how a vanilla adversarial training and its\nimproved version can help increase robustness of a toxicity detector even\nagainst unseen attacks.\n", "date": "2024-04-09", "categories": []}
{"id": "2404.09961", "title": "Ti-Patch: Tiled Physical Adversarial Patch for no-reference video\n  quality metrics", "abstract": "  Objective no-reference image- and video-quality metrics are crucial in many\ncomputer vision tasks. However, state-of-the-art no-reference metrics have\nbecome learning-based and are vulnerable to adversarial attacks. The\nvulnerability of quality metrics imposes restrictions on using such metrics in\nquality control systems and comparing objective algorithms. Also, using\nvulnerable metrics as a loss for deep learning model training can mislead\ntraining to worsen visual quality. Because of that, quality metrics testing for\nvulnerability is a task of current interest. This paper proposes a new method\nfor testing quality metrics vulnerability in the physical space. To our\nknowledge, quality metrics were not previously tested for vulnerability to this\nattack; they were only tested in the pixel space. We applied a physical\nadversarial Ti-Patch (Tiled Patch) attack to quality metrics and did\nexperiments both in pixel and physical space. We also performed experiments on\nthe implementation of physical adversarial wallpaper. The proposed method can\nbe used as additional quality metrics in vulnerability evaluation,\ncomplementing traditional subjective comparison and vulnerability tests in the\npixel space. We made our code and adversarial videos available on GitHub:\nhttps://github.com/leonenkova/Ti-Patch.\n", "date": "2024-04-15", "categories": []}
{"id": "2212.02705", "title": "What is the Solution for State-Adversarial Multi-Agent Reinforcement\n  Learning?", "abstract": "  Various methods for Multi-Agent Reinforcement Learning (MARL) have been\ndeveloped with the assumption that agents' policies are based on accurate state\ninformation. However, policies learned through Deep Reinforcement Learning\n(DRL) are susceptible to adversarial state perturbation attacks. In this work,\nwe propose a State-Adversarial Markov Game (SAMG) and make the first attempt to\ninvestigate different solution concepts of MARL under state uncertainties. Our\nanalysis shows that the commonly used solution concepts of optimal agent policy\nand robust Nash equilibrium do not always exist in SAMGs. To circumvent this\ndifficulty, we consider a new solution concept called robust agent policy,\nwhere agents aim to maximize the worst-case expected state value. We prove the\nexistence of robust agent policy for finite state and finite action SAMGs.\nAdditionally, we propose a Robust Multi-Agent Adversarial Actor-Critic (RMA3C)\nalgorithm to learn robust policies for MARL agents under state uncertainties.\nOur experiments demonstrate that our algorithm outperforms existing methods\nwhen faced with state perturbations and greatly improves the robustness of MARL\npolicies. Our code is public on\nhttps://songyanghan.github.io/what_is_solution/.\n", "date": "2022-12-06", "categories": []}
{"id": "2403.18554", "title": "CosalPure: Learning Concept from Group Images for Robust Co-Saliency\n  Detection", "abstract": "  Co-salient object detection (CoSOD) aims to identify the common and salient\n(usually in the foreground) regions across a given group of images. Although\nachieving significant progress, state-of-the-art CoSODs could be easily\naffected by some adversarial perturbations, leading to substantial accuracy\nreduction. The adversarial perturbations can mislead CoSODs but do not change\nthe high-level semantic information (e.g., concept) of the co-salient objects.\nIn this paper, we propose a novel robustness enhancement framework by first\nlearning the concept of the co-salient objects based on the input group images\nand then leveraging this concept to purify adversarial perturbations, which are\nsubsequently fed to CoSODs for robustness enhancement. Specifically, we propose\nCosalPure containing two modules, i.e., group-image concept learning and\nconcept-guided diffusion purification. For the first module, we adopt a\npre-trained text-to-image diffusion model to learn the concept of co-salient\nobjects within group images where the learned concept is robust to adversarial\nexamples. For the second module, we map the adversarial image to the latent\nspace and then perform diffusion generation by embedding the learned concept\ninto the noise prediction function as an extra condition. Our method can\neffectively alleviate the influence of the SOTA adversarial attack containing\ndifferent adversarial patterns, including exposure and noise. The extensive\nresults demonstrate that our method could enhance the robustness of CoSODs\nsignificantly.\n", "date": "2024-03-27", "categories": []}
{"id": "2404.08069", "title": "Persistent Classification: A New Approach to Stability of Data and\n  Adversarial Examples", "abstract": "  There are a number of hypotheses underlying the existence of adversarial\nexamples for classification problems. These include the high-dimensionality of\nthe data, high codimension in the ambient space of the data manifolds of\ninterest, and that the structure of machine learning models may encourage\nclassifiers to develop decision boundaries close to data points. This article\nproposes a new framework for studying adversarial examples that does not depend\ndirectly on the distance to the decision boundary. Similarly to the smoothed\nclassifier literature, we define a (natural or adversarial) data point to be\n$(\\gamma,\\sigma)$-stable if the probability of the same classification is at\nleast $\\gamma$ for points sampled in a Gaussian neighborhood of the point with\na given standard deviation $\\sigma$. We focus on studying the differences\nbetween persistence metrics along interpolants of natural and adversarial\npoints. We show that adversarial examples have significantly lower persistence\nthan natural examples for large neural networks in the context of the MNIST and\nImageNet datasets. We connect this lack of persistence with decision boundary\ngeometry by measuring angles of interpolants with respect to decision\nboundaries. Finally, we connect this approach with robustness by developing a\nmanifold alignment gradient metric and demonstrating the increase in robustness\nthat can be achieved when training with the addition of this metric.\n", "date": "2024-04-11", "categories": []}
{"id": "2404.08538", "title": "VertAttack: Taking advantage of Text Classifiers' horizontal vision", "abstract": "  Text classification systems have continuously improved in performance over\nthe years. However, nearly all current SOTA classifiers have a similar\nshortcoming, they process text in a horizontal manner. Vertically written words\nwill not be recognized by a classifier. In contrast, humans are easily able to\nrecognize and read words written both horizontally and vertically. Hence, a\nhuman adversary could write problematic words vertically and the meaning would\nstill be preserved to other humans. We simulate such an attack, VertAttack.\nVertAttack identifies which words a classifier is reliant on and then rewrites\nthose words vertically. We find that VertAttack is able to greatly drop the\naccuracy of 4 different transformer models on 5 datasets. For example, on the\nSST2 dataset, VertAttack is able to drop RoBERTa's accuracy from 94 to 13%.\nFurthermore, since VertAttack does not replace the word, meaning is easily\npreserved. We verify this via a human study and find that crowdworkers are able\nto correctly label 77% perturbed texts perturbed, compared to 81% of the\noriginal texts. We believe VertAttack offers a look into how humans might\ncircumvent classifiers in the future and thus inspire a look into more robust\nalgorithms.\n", "date": "2024-04-12", "categories": []}
{"id": "2305.19521", "title": "Incremental Randomized Smoothing Certification", "abstract": "  Randomized smoothing-based certification is an effective approach for\nobtaining robustness certificates of deep neural networks (DNNs) against\nadversarial attacks. This method constructs a smoothed DNN model and certifies\nits robustness through statistical sampling, but it is computationally\nexpensive, especially when certifying with a large number of samples.\nFurthermore, when the smoothed model is modified (e.g., quantized or pruned),\ncertification guarantees may not hold for the modified DNN, and recertifying\nfrom scratch can be prohibitively expensive.\n  We present the first approach for incremental robustness certification for\nrandomized smoothing, IRS. We show how to reuse the certification guarantees\nfor the original smoothed model to certify an approximated model with very few\nsamples. IRS significantly reduces the computational cost of certifying\nmodified DNNs while maintaining strong robustness guarantees. We experimentally\ndemonstrate the effectiveness of our approach, showing up to 3x certification\nspeedup over the certification that applies randomized smoothing of the\napproximate model from scratch.\n", "date": "2023-05-31", "categories": []}
{"id": "2403.18256", "title": "Manipulating Neural Path Planners via Slight Perturbations", "abstract": "  Data-driven neural path planners are attracting increasing interest in the\nrobotics community. However, their neural network components typically come as\nblack boxes, obscuring their underlying decision-making processes. Their\nblack-box nature exposes them to the risk of being compromised via the\ninsertion of hidden malicious behaviors. For example, an attacker may hide\nbehaviors that, when triggered, hijack a delivery robot by guiding it to a\nspecific (albeit wrong) destination, trapping it in a predefined region, or\ninducing unnecessary energy expenditure by causing the robot to repeatedly\ncircle a region. In this paper, we propose a novel approach to specify and\ninject a range of hidden malicious behaviors, known as backdoors, into neural\npath planners. Our approach provides a concise but flexible way to define these\nbehaviors, and we show that hidden behaviors can be triggered by slight\nperturbations (e.g., inserting a tiny unnoticeable object), that can\nnonetheless significantly compromise their integrity. We also discuss potential\ntechniques to identify these backdoors aimed at alleviating such risks. We\ndemonstrate our approach on both sampling-based and search-based neural path\nplanners.\n", "date": "2024-03-27", "categories": []}
{"id": "2404.01574", "title": "Multi-granular Adversarial Attacks against Black-box Neural Ranking\n  Models", "abstract": "  Adversarial ranking attacks have gained increasing attention due to their\nsuccess in probing vulnerabilities, and, hence, enhancing the robustness, of\nneural ranking models. Conventional attack methods employ perturbations at a\nsingle granularity, e.g., word or sentence level, to target documents. However,\nlimiting perturbations to a single level of granularity may reduce the\nflexibility of adversarial examples, thereby diminishing the potential threat\nof the attack. Therefore, we focus on generating high-quality adversarial\nexamples by incorporating multi-granular perturbations. Achieving this\nobjective involves tackling a combinatorial explosion problem, which requires\nidentifying an optimal combination of perturbations across all possible levels\nof granularity, positions, and textual pieces. To address this challenge, we\ntransform the multi-granular adversarial attack into a sequential\ndecision-making process, where perturbations in the next attack step build on\nthe perturbed document in the current attack step. Since the attack process can\nonly access the final state without direct intermediate signals, we use\nreinforcement learning to perform multi-granular attacks. During the\nreinforcement learning process, two agents work cooperatively to identify\nmulti-granular vulnerabilities as attack targets and organize perturbation\ncandidates into a final perturbation sequence. Experimental results show that\nour attack method surpasses prevailing baselines in both attack effectiveness\nand imperceptibility.\n", "date": "2024-04-02", "categories": []}
{"id": "2304.06607", "title": "False Claims against Model Ownership Resolution", "abstract": "  Deep neural network (DNN) models are valuable intellectual property of model\nowners, constituting a competitive advantage. Therefore, it is crucial to\ndevelop techniques to protect against model theft. Model ownership resolution\n(MOR) is a class of techniques that can deter model theft. A MOR scheme enables\nan accuser to assert an ownership claim for a suspect model by presenting\nevidence, such as a watermark or fingerprint, to show that the suspect model\nwas stolen or derived from a source model owned by the accuser. Most of the\nexisting MOR schemes prioritize robustness against malicious suspects, ensuring\nthat the accuser will win if the suspect model is indeed a stolen model.\n  In this paper, we show that common MOR schemes in the literature are\nvulnerable to a different, equally important but insufficiently explored,\nrobustness concern: a malicious accuser. We show how malicious accusers can\nsuccessfully make false claims against independent suspect models that were not\nstolen. Our core idea is that a malicious accuser can deviate (without\ndetection) from the specified MOR process by finding (transferable) adversarial\nexamples that successfully serve as evidence against independent suspect\nmodels. To this end, we first generalize the procedures of common MOR schemes\nand show that, under this generalization, defending against false claims is as\nchallenging as preventing (transferable) adversarial examples. Via systematic\nempirical evaluation, we show that our false claim attacks always succeed in\nthe MOR schemes that follow our generalization, including in a real-world\nmodel: Amazon's Rekognition API.\n", "date": "2023-04-13", "categories": []}
{"id": "2401.04350", "title": "Pre-trained Model Guided Fine-Tuning for Zero-Shot Adversarial\n  Robustness", "abstract": "  Large-scale pre-trained vision-language models like CLIP have demonstrated\nimpressive performance across various tasks, and exhibit remarkable zero-shot\ngeneralization capability, while they are also vulnerable to imperceptible\nadversarial examples. Existing works typically employ adversarial training\n(fine-tuning) as a defense method against adversarial examples. However, direct\napplication to the CLIP model may result in overfitting, compromising the\nmodel's capacity for generalization. In this paper, we propose Pre-trained\nModel Guided Adversarial Fine-Tuning (PMG-AFT) method, which leverages\nsupervision from the original pre-trained model by carefully designing an\nauxiliary branch, to enhance the model's zero-shot adversarial robustness.\nSpecifically, PMG-AFT minimizes the distance between the features of\nadversarial examples in the target model and those in the pre-trained model,\naiming to preserve the generalization features already captured by the\npre-trained model. Extensive Experiments on 15 zero-shot datasets demonstrate\nthat PMG-AFT significantly outperforms the state-of-the-art method, improving\nthe top-1 robust accuracy by an average of 4.99%. Furthermore, our approach\nconsistently improves clean accuracy by an average of 8.72%. Our code is\navailable at\nhttps://github.com/serendipity1122/Pre-trained-Model-Guided-Fine-Tuning-for-Zero-Shot-Adversarial-Robustness.\n", "date": "2024-01-09", "categories": []}
{"id": "2404.06776", "title": "Logit Calibration and Feature Contrast for Robust Federated Learning on\n  Non-IID Data", "abstract": "  Federated learning (FL) is a privacy-preserving distributed framework for\ncollaborative model training on devices in edge networks. However, challenges\narise due to vulnerability to adversarial examples (AEs) and the\nnon-independent and identically distributed (non-IID) nature of data\ndistribution among devices, hindering the deployment of adversarially robust\nand accurate learning models at the edge. While adversarial training (AT) is\ncommonly acknowledged as an effective defense strategy against adversarial\nattacks in centralized training, we shed light on the adverse effects of\ndirectly applying AT in FL that can severely compromise accuracy, especially in\nnon-IID challenges. Given this limitation, this paper proposes FatCC, which\nincorporates local logit \\underline{C}alibration and global feature\n\\underline{C}ontrast into the vanilla federated adversarial training\n(\\underline{FAT}) process from both logit and feature perspectives. This\napproach can effectively enhance the federated system's robust accuracy (RA)\nand clean accuracy (CA). First, we propose logit calibration, where the logits\nare calibrated during local adversarial updates, thereby improving adversarial\nrobustness. Second, FatCC introduces feature contrast, which involves a global\nalignment term that aligns each local representation with unbiased global\nfeatures, thus further enhancing robustness and accuracy in federated\nadversarial environments. Extensive experiments across multiple datasets\ndemonstrate that FatCC achieves comparable or superior performance gains in\nboth CA and RA compared to other baselines.\n", "date": "2024-04-10", "categories": []}
{"id": "2403.16432", "title": "$\\textit{LinkPrompt}$: Natural and Universal Adversarial Attacks on\n  Prompt-based Language Models", "abstract": "  Prompt-based learning is a new language model training paradigm that adapts\nthe Pre-trained Language Models (PLMs) to downstream tasks, which revitalizes\nthe performance benchmarks across various natural language processing (NLP)\ntasks. Instead of using a fixed prompt template to fine-tune the model, some\nresearch demonstrates the effectiveness of searching for the prompt via\noptimization. Such prompt optimization process of prompt-based learning on PLMs\nalso gives insight into generating adversarial prompts to mislead the model,\nraising concerns about the adversarial vulnerability of this paradigm. Recent\nstudies have shown that universal adversarial triggers (UATs) can be generated\nto alter not only the predictions of the target PLMs but also the prediction of\ncorresponding Prompt-based Fine-tuning Models (PFMs) under the prompt-based\nlearning paradigm. However, UATs found in previous works are often unreadable\ntokens or characters and can be easily distinguished from natural texts with\nadaptive defenses. In this work, we consider the naturalness of the UATs and\ndevelop $\\textit{LinkPrompt}$, an adversarial attack algorithm to generate UATs\nby a gradient-based beam search algorithm that not only effectively attacks the\ntarget PLMs and PFMs but also maintains the naturalness among the trigger\ntokens. Extensive results demonstrate the effectiveness of\n$\\textit{LinkPrompt}$, as well as the transferability of UATs generated by\n$\\textit{LinkPrompt}$ to open-sourced Large Language Model (LLM) Llama2 and\nAPI-accessed LLM GPT-3.5-turbo. The resource is available at\n$\\href{https://github.com/SavannahXu79/LinkPrompt}{https://github.com/SavannahXu79/LinkPrompt}$.\n", "date": "2024-03-25", "categories": []}
{"id": "2404.06236", "title": "Towards Robust Domain Generation Algorithm Classification", "abstract": "  In this work, we conduct a comprehensive study on the robustness of domain\ngeneration algorithm (DGA) classifiers. We implement 32 white-box attacks, 19\nof which are very effective and induce a false-negative rate (FNR) of $\\approx$\n100\\% on unhardened classifiers. To defend the classifiers, we evaluate\ndifferent hardening approaches and propose a novel training scheme that\nleverages adversarial latent space vectors and discretized adversarial domains\nto significantly improve robustness. In our study, we highlight a pitfall to\navoid when hardening classifiers and uncover training biases that can be easily\nexploited by attackers to bypass detection, but which can be mitigated by\nadversarial training (AT). In our study, we do not observe any trade-off\nbetween robustness and performance, on the contrary, hardening improves a\nclassifier's detection performance for known and unknown DGAs. We implement all\nattacks and defenses discussed in this paper as a standalone library, which we\nmake publicly available to facilitate hardening of DGA classifiers:\nhttps://gitlab.com/rwth-itsec/robust-dga-detection\n", "date": "2024-04-09", "categories": []}
{"id": "2404.06247", "title": "LRR: Language-Driven Resamplable Continuous Representation against\n  Adversarial Tracking Attacks", "abstract": "  Visual object tracking plays a critical role in visual-based autonomous\nsystems, as it aims to estimate the position and size of the object of interest\nwithin a live video. Despite significant progress made in this field,\nstate-of-the-art (SOTA) trackers often fail when faced with adversarial\nperturbations in the incoming frames. This can lead to significant robustness\nand security issues when these trackers are deployed in the real world. To\nachieve high accuracy on both clean and adversarial data, we propose building a\nspatial-temporal continuous representation using the semantic text guidance of\nthe object of interest. This novel continuous representation enables us to\nreconstruct incoming frames to maintain semantic and appearance consistency\nwith the object of interest and its clean counterparts. As a result, our\nproposed method successfully defends against different SOTA adversarial\ntracking attacks while maintaining high accuracy on clean data. In particular,\nour method significantly increases tracking accuracy under adversarial attacks\nwith around 90% relative improvement on UAV123, which is even higher than the\naccuracy on clean data.\n", "date": "2024-04-09", "categories": []}
{"id": "2302.11628", "title": "Provable Robustness Against a Union of $\\ell_0$ Adversarial Attacks", "abstract": "  Sparse or $\\ell_0$ adversarial attacks arbitrarily perturb an unknown subset\nof the features. $\\ell_0$ robustness analysis is particularly well-suited for\nheterogeneous (tabular) data where features have different types or scales.\nState-of-the-art $\\ell_0$ certified defenses are based on randomized smoothing\nand apply to evasion attacks only. This paper proposes feature partition\naggregation (FPA) -- a certified defense against the union of $\\ell_0$ evasion,\nbackdoor, and poisoning attacks. FPA generates its stronger robustness\nguarantees via an ensemble whose submodels are trained on disjoint feature\nsets. Compared to state-of-the-art $\\ell_0$ defenses, FPA is up to\n3,000${\\times}$ faster and provides larger median robustness guarantees (e.g.,\nmedian certificates of 13 pixels over 10 for CIFAR10, 12 pixels over 10 for\nMNIST, 4 features over 1 for Weather, and 3 features over 1 for Ames), meaning\nFPA provides the additional dimensions of robustness essentially for free.\n", "date": "2023-02-22", "categories": []}
{"id": "2308.09958", "title": "A Comparison of Adversarial Learning Techniques for Malware Detection", "abstract": "  Machine learning has proven to be a useful tool for automated malware\ndetection, but machine learning models have also been shown to be vulnerable to\nadversarial attacks. This article addresses the problem of generating\nadversarial malware samples, specifically malicious Windows Portable Executable\nfiles. We summarize and compare work that has focused on adversarial machine\nlearning for malware detection. We use gradient-based, evolutionary\nalgorithm-based, and reinforcement-based methods to generate adversarial\nsamples, and then test the generated samples against selected antivirus\nproducts. We compare the selected methods in terms of accuracy and practical\napplicability. The results show that applying optimized modifications to\npreviously detected malware can lead to incorrect classification of the file as\nbenign. It is also known that generated malware samples can be successfully\nused against detection models other than those used to generate them and that\nusing combinations of generators can create new samples that evade detection.\nExperiments show that the Gym-malware generator, which uses a reinforcement\nlearning approach, has the greatest practical potential. This generator\nachieved an average sample generation time of 5.73 seconds and the highest\naverage evasion rate of 44.11%. Using the Gym-malware generator in combination\nwith itself improved the evasion rate to 58.35%.\n", "date": "2023-08-19", "categories": []}
{"id": "2312.17667", "title": "AIJack: Let's Hijack AI! Security and Privacy Risk Simulator for Machine\n  Learning", "abstract": "  This paper introduces AIJack, an open-source library designed to assess\nsecurity and privacy risks associated with the training and deployment of\nmachine learning models. Amid the growing interest in big data and AI,\nadvancements in machine learning research and business are accelerating.\nHowever, recent studies reveal potential threats, such as the theft of training\ndata and the manipulation of models by malicious attackers. Therefore, a\ncomprehensive understanding of machine learning's security and privacy\nvulnerabilities is crucial for the safe integration of machine learning into\nreal-world products. AIJack aims to address this need by providing a library\nwith various attack and defense methods through a unified API. The library is\npublicly available on GitHub (https://github.com/Koukyosyumei/AIJack).\n", "date": "2023-12-29", "categories": []}
{"id": "2402.05713", "title": "Hidden in Plain Sight: Undetectable Adversarial Bias Attacks on\n  Vulnerable Patient Populations", "abstract": "  The proliferation of artificial intelligence (AI) in radiology has shed light\non the risk of deep learning (DL) models exacerbating clinical biases towards\nvulnerable patient populations. While prior literature has focused on\nquantifying biases exhibited by trained DL models, demographically targeted\nadversarial bias attacks on DL models and its implication in the clinical\nenvironment remains an underexplored field of research in medical imaging. In\nthis work, we demonstrate that demographically targeted label poisoning attacks\ncan introduce undetectable underdiagnosis bias in DL models. Our results across\nmultiple performance metrics and demographic groups like sex, age, and their\nintersectional subgroups show that adversarial bias attacks demonstrate\nhigh-selectivity for bias in the targeted group by degrading group model\nperformance without impacting overall model performance. Furthermore, our\nresults indicate that adversarial bias attacks result in biased DL models that\npropagate prediction bias even when evaluated with external datasets.\n", "date": "2024-02-08", "categories": []}
{"id": "2404.00076", "title": "A Backdoor Approach with Inverted Labels Using Dirty Label-Flipping\n  Attacks", "abstract": "  Audio-based machine learning systems frequently use public or third-party\ndata, which might be inaccurate. This exposes deep neural network (DNN) models\ntrained on such data to potential data poisoning attacks. In this type of\nassault, attackers can train the DNN model using poisoned data, potentially\ndegrading its performance. Another type of data poisoning attack that is\nextremely relevant to our investigation is label flipping, in which the\nattacker manipulates the labels for a subset of data. It has been demonstrated\nthat these assaults may drastically reduce system performance, even for\nattackers with minimal abilities. In this study, we propose a backdoor attack\nnamed 'DirtyFlipping', which uses dirty label techniques, \"label-on-label\", to\ninput triggers (clapping) in the selected data patterns associated with the\ntarget class, thereby enabling a stealthy backdoor.\n", "date": "2024-03-29", "categories": []}
{"id": "2404.04714", "title": "Data Poisoning Attacks on Off-Policy Policy Evaluation Methods", "abstract": "  Off-policy Evaluation (OPE) methods are a crucial tool for evaluating\npolicies in high-stakes domains such as healthcare, where exploration is often\ninfeasible, unethical, or expensive. However, the extent to which such methods\ncan be trusted under adversarial threats to data quality is largely unexplored.\nIn this work, we make the first attempt at investigating the sensitivity of OPE\nmethods to marginal adversarial perturbations to the data. We design a generic\ndata poisoning attack framework leveraging influence functions from robust\nstatistics to carefully construct perturbations that maximize error in the\npolicy value estimates. We carry out extensive experimentation with multiple\nhealthcare and control datasets. Our results demonstrate that many existing OPE\nmethods are highly prone to generating value estimates with large errors when\nsubject to data poisoning attacks, even for small adversarial perturbations.\nThese findings question the reliability of policy values derived using OPE\nmethods and motivate the need for developing OPE methods that are statistically\nrobust to train-time data poisoning attacks.\n", "date": "2024-04-06", "categories": []}
{"id": "2404.05159", "title": "Semantic Stealth: Adversarial Text Attacks on NLP Using Several Methods", "abstract": "  In various real-world applications such as machine translation, sentiment\nanalysis, and question answering, a pivotal role is played by NLP models,\nfacilitating efficient communication and decision-making processes in domains\nranging from healthcare to finance. However, a significant challenge is posed\nto the robustness of these natural language processing models by text\nadversarial attacks. These attacks involve the deliberate manipulation of input\ntext to mislead the predictions of the model while maintaining human\ninterpretability. Despite the remarkable performance achieved by\nstate-of-the-art models like BERT in various natural language processing tasks,\nthey are found to remain vulnerable to adversarial perturbations in the input\ntext. In addressing the vulnerability of text classifiers to adversarial\nattacks, three distinct attack mechanisms are explored in this paper using the\nvictim model BERT: BERT-on-BERT attack, PWWS attack, and Fraud Bargain's Attack\n(FBA). Leveraging the IMDB, AG News, and SST2 datasets, a thorough comparative\nanalysis is conducted to assess the effectiveness of these attacks on the BERT\nclassifier model. It is revealed by the analysis that PWWS emerges as the most\npotent adversary, consistently outperforming other methods across multiple\nevaluation scenarios, thereby emphasizing its efficacy in generating\nadversarial examples for text classification. Through comprehensive\nexperimentation, the performance of these attacks is assessed and the findings\nindicate that the PWWS attack outperforms others, demonstrating lower runtime,\nhigher accuracy, and favorable semantic similarity scores. The key insight of\nthis paper lies in the assessment of the relative performances of three\nprevalent state-of-the-art attack mechanisms.\n", "date": "2024-04-08", "categories": []}
{"id": "2404.05350", "title": "Certified PEFTSmoothing: Parameter-Efficient Fine-Tuning with Randomized\n  Smoothing", "abstract": "  Randomized smoothing is the primary certified robustness method for accessing\nthe robustness of deep learning models to adversarial perturbations in the\nl2-norm, by adding isotropic Gaussian noise to the input image and returning\nthe majority votes over the base classifier. Theoretically, it provides a\ncertified norm bound, ensuring predictions of adversarial examples are stable\nwithin this bound. A notable constraint limiting widespread adoption is the\nnecessity to retrain base models entirely from scratch to attain a robust\nversion. This is because the base model fails to learn the noise-augmented data\ndistribution to give an accurate vote. One intuitive way to overcome this\nchallenge is to involve a custom-trained denoiser to eliminate the noise.\nHowever, this approach is inefficient and sub-optimal. Inspired by recent large\nmodel training procedures, we explore an alternative way named PEFTSmoothing to\nadapt the base model to learn the Gaussian noise-augmented data with\nParameter-Efficient Fine-Tuning (PEFT) methods in both white-box and black-box\nsettings. Extensive results demonstrate the effectiveness and efficiency of\nPEFTSmoothing, which allow us to certify over 98% accuracy for ViT on CIFAR-10,\n20% higher than SoTA denoised smoothing, and over 61% accuracy on ImageNet\nwhich is 30% higher than CNN-based denoiser and comparable to the\nDiffusion-based denoiser.\n", "date": "2024-04-08", "categories": []}
{"id": "2404.05639", "title": "Investigating the Impact of Quantization on Adversarial Robustness", "abstract": "  Quantization is a promising technique for reducing the bit-width of deep\nmodels to improve their runtime performance and storage efficiency, and thus\nbecomes a fundamental step for deployment. In real-world scenarios, quantized\nmodels are often faced with adversarial attacks which cause the model to make\nincorrect inferences by introducing slight perturbations. However, recent\nstudies have paid less attention to the impact of quantization on the model\nrobustness. More surprisingly, existing studies on this topic even present\ninconsistent conclusions, which prompted our in-depth investigation. In this\npaper, we conduct a first-time analysis of the impact of the quantization\npipeline components that can incorporate robust optimization under the settings\nof Post-Training Quantization and Quantization-Aware Training. Through our\ndetailed analysis, we discovered that this inconsistency arises from the use of\ndifferent pipelines in different studies, specifically regarding whether robust\noptimization is performed and at which quantization stage it occurs. Our\nresearch findings contribute insights into deploying more secure and robust\nquantized networks, assisting practitioners in reference for scenarios with\nhigh-security requirements and limited resources.\n", "date": "2024-04-08", "categories": []}
{"id": "2404.00185", "title": "On Inherent Adversarial Robustness of Active Vision Systems", "abstract": "  Current Deep Neural Networks are vulnerable to adversarial examples, which\nalter their predictions by adding carefully crafted noise. Since human eyes are\nrobust to such inputs, it is possible that the vulnerability stems from the\nstandard way of processing inputs in one shot by processing every pixel with\nthe same importance. In contrast, neuroscience suggests that the human vision\nsystem can differentiate salient features by (1) switching between multiple\nfixation points (saccades) and (2) processing the surrounding with a\nnon-uniform external resolution (foveation). In this work, we advocate that the\nintegration of such active vision mechanisms into current deep learning systems\ncan offer robustness benefits. Specifically, we empirically demonstrate the\ninherent robustness of two active vision methods - GFNet and FALcon - under a\nblack box threat model. By learning and inferencing based on downsampled\nglimpses obtained from multiple distinct fixation points within an input, we\nshow that these active methods achieve (2-3) times greater robustness compared\nto a standard passive convolutional network under state-of-the-art adversarial\nattacks. More importantly, we provide illustrative and interpretable\nvisualization analysis that demonstrates how performing inference from distinct\nfixation points makes active vision methods less vulnerable to malicious\ninputs.\n", "date": "2024-03-29", "categories": []}
{"id": "2404.04245", "title": "Evaluating Adversarial Robustness: A Comparison Of FGSM, Carlini-Wagner\n  Attacks, And The Role of Distillation as Defense Mechanism", "abstract": "  This technical report delves into an in-depth exploration of adversarial\nattacks specifically targeted at Deep Neural Networks (DNNs) utilized for image\nclassification. The study also investigates defense mechanisms aimed at\nbolstering the robustness of machine learning models. The research focuses on\ncomprehending the ramifications of two prominent attack methodologies: the Fast\nGradient Sign Method (FGSM) and the Carlini-Wagner (CW) approach. These attacks\nare examined concerning three pre-trained image classifiers: Resnext50_32x4d,\nDenseNet-201, and VGG-19, utilizing the Tiny-ImageNet dataset. Furthermore, the\nstudy proposes the robustness of defensive distillation as a defense mechanism\nto counter FGSM and CW attacks. This defense mechanism is evaluated using the\nCIFAR-10 dataset, where CNN models, specifically resnet101 and Resnext50_32x4d,\nserve as the teacher and student models, respectively. The proposed defensive\ndistillation model exhibits effectiveness in thwarting attacks such as FGSM.\nHowever, it is noted to remain susceptible to more sophisticated techniques\nlike the CW attack. The document presents a meticulous validation of the\nproposed scheme. It provides detailed and comprehensive results, elucidating\nthe efficacy and limitations of the defense mechanisms employed. Through\nrigorous experimentation and analysis, the study offers insights into the\ndynamics of adversarial attacks on DNNs, as well as the effectiveness of\ndefensive strategies in mitigating their impact.\n", "date": "2024-04-05", "categories": []}
{"id": "2404.03340", "title": "Meta Invariance Defense Towards Generalizable Robustness to Unknown\n  Adversarial Attacks", "abstract": "  Despite providing high-performance solutions for computer vision tasks, the\ndeep neural network (DNN) model has been proved to be extremely vulnerable to\nadversarial attacks. Current defense mainly focuses on the known attacks, but\nthe adversarial robustness to the unknown attacks is seriously overlooked.\nBesides, commonly used adaptive learning and fine-tuning technique is\nunsuitable for adversarial defense since it is essentially a zero-shot problem\nwhen deployed. Thus, to tackle this challenge, we propose an attack-agnostic\ndefense method named Meta Invariance Defense (MID). Specifically, various\ncombinations of adversarial attacks are randomly sampled from a manually\nconstructed Attacker Pool to constitute different defense tasks against unknown\nattacks, in which a student encoder is supervised by multi-consistency\ndistillation to learn the attack-invariant features via a meta principle. The\nproposed MID has two merits: 1) Full distillation from pixel-, feature- and\nprediction-level between benign and adversarial samples facilitates the\ndiscovery of attack-invariance. 2) The model simultaneously achieves robustness\nto the imperceptible adversarial perturbations in high-level image\nclassification and attack-suppression in low-level robust image regeneration.\nTheoretical and empirical studies on numerous benchmarks such as ImageNet\nverify the generalizable robustness and superiority of MID under various\nattacks.\n", "date": "2024-04-04", "categories": []}
{"id": "2404.02287", "title": "One Noise to Rule Them All: Multi-View Adversarial Attacks with\n  Universal Perturbation", "abstract": "  This paper presents a novel universal perturbation method for generating\nrobust multi-view adversarial examples in 3D object recognition. Unlike\nconventional attacks limited to single views, our approach operates on multiple\n2D images, offering a practical and scalable solution for enhancing model\nscalability and robustness. This generalizable method bridges the gap between\n2D perturbations and 3D-like attack capabilities, making it suitable for\nreal-world applications.\n  Existing adversarial attacks may become ineffective when images undergo\ntransformations like changes in lighting, camera position, or natural\ndeformations. We address this challenge by crafting a single universal noise\nperturbation applicable to various object views. Experiments on diverse\nrendered 3D objects demonstrate the effectiveness of our approach. The\nuniversal perturbation successfully identified a single adversarial noise for\neach given set of 3D object renders from multiple poses and viewpoints.\nCompared to single-view attacks, our universal attacks lower classification\nconfidence across multiple viewing angles, especially at low noise levels. A\nsample implementation is made available at\nhttps://github.com/memoatwit/UniversalPerturbation.\n", "date": "2024-04-02", "categories": []}
{"id": "2404.02585", "title": "Unsegment Anything by Simulating Deformation", "abstract": "  Foundation segmentation models, while powerful, pose a significant risk: they\nenable users to effortlessly extract any objects from any digital content with\na single click, potentially leading to copyright infringement or malicious\nmisuse. To mitigate this risk, we introduce a new task \"Anything Unsegmentable\"\nto grant any image \"the right to be unsegmented\". The ambitious pursuit of the\ntask is to achieve highly transferable adversarial attacks against all\nprompt-based segmentation models, regardless of model parameterizations and\nprompts. We highlight the non-transferable and heterogeneous nature of\nprompt-specific adversarial noises. Our approach focuses on disrupting image\nencoder features to achieve prompt-agnostic attacks. Intriguingly, targeted\nfeature attacks exhibit better transferability compared to untargeted ones,\nsuggesting the optimal update direction aligns with the image manifold. Based\non the observations, we design a novel attack named Unsegment Anything by\nSimulating Deformation (UAD). Our attack optimizes a differentiable deformation\nfunction to create a target deformed image, which alters structural information\nwhile preserving achievable feature distance by adversarial example. Extensive\nexperiments verify the effectiveness of our approach, compromising a variety of\npromptable segmentation models with different architectures and prompt\ninterfaces. We release the code at\nhttps://github.com/jiahaolu97/anything-unsegmentable.\n", "date": "2024-04-03", "categories": []}
{"id": "2404.02660", "title": "Adversarial Attacks and Dimensionality in Text Classifiers", "abstract": "  Adversarial attacks on machine learning algorithms have been a key deterrent\nto the adoption of AI in many real-world use cases. They significantly\nundermine the ability of high-performance neural networks by forcing\nmisclassifications. These attacks introduce minute and structured perturbations\nor alterations in the test samples, imperceptible to human annotators in\ngeneral, but trained neural networks and other models are sensitive to it.\nHistorically, adversarial attacks have been first identified and studied in the\ndomain of image processing. In this paper, we study adversarial examples in the\nfield of natural language processing, specifically text classification tasks.\nWe investigate the reasons for adversarial vulnerability, particularly in\nrelation to the inherent dimensionality of the model. Our key finding is that\nthere is a very strong correlation between the embedding dimensionality of the\nadversarial samples and their effectiveness on models tuned with input samples\nwith same embedding dimension. We utilize this sensitivity to design an\nadversarial defense mechanism. We use ensemble models of varying inherent\ndimensionality to thwart the attacks. This is tested on multiple datasets for\nits efficacy in providing robustness. We also study the problem of measuring\nadversarial perturbation using different distance metrics. For all of the\naforementioned studies, we have run tests on multiple models with varying\ndimensionality and used a word-vector level adversarial attack to substantiate\nthe findings.\n", "date": "2024-04-03", "categories": []}
{"id": "2310.13076", "title": "PatchCURE: Improving Certifiable Robustness, Model Utility, and\n  Computation Efficiency of Adversarial Patch Defenses", "abstract": "  State-of-the-art defenses against adversarial patch attacks can now achieve\nstrong certifiable robustness with a marginal drop in model utility. However,\nthis impressive performance typically comes at the cost of 10-100x more\ninference-time computation compared to undefended models -- the research\ncommunity has witnessed an intense three-way trade-off between certifiable\nrobustness, model utility, and computation efficiency. In this paper, we\npropose a defense framework named PatchCURE to approach this trade-off problem.\nPatchCURE provides sufficient \"knobs\" for tuning defense performance and allows\nus to build a family of defenses: the most robust PatchCURE instance can match\nthe performance of any existing state-of-the-art defense (without efficiency\nconsiderations); the most efficient PatchCURE instance has similar inference\nefficiency as undefended models. Notably, PatchCURE achieves state-of-the-art\nrobustness and utility performance across all different efficiency levels,\ne.g., 16-23% absolute clean accuracy and certified robust accuracy advantages\nover prior defenses when requiring computation efficiency to be close to\nundefended models. The family of PatchCURE defenses enables us to flexibly\nchoose appropriate defenses to satisfy given computation and/or utility\nconstraints in practice.\n", "date": "2023-10-19", "categories": []}
{"id": "2403.19080", "title": "MMCert: Provable Defense against Adversarial Attacks to Multi-modal\n  Models", "abstract": "  Different from a unimodal model whose input is from a single modality, the\ninput (called multi-modal input) of a multi-modal model is from multiple\nmodalities such as image, 3D points, audio, text, etc. Similar to unimodal\nmodels, many existing studies show that a multi-modal model is also vulnerable\nto adversarial perturbation, where an attacker could add small perturbation to\nall modalities of a multi-modal input such that the multi-modal model makes\nincorrect predictions for it. Existing certified defenses are mostly designed\nfor unimodal models, which achieve sub-optimal certified robustness guarantees\nwhen extended to multi-modal models as shown in our experimental results. In\nour work, we propose MMCert, the first certified defense against adversarial\nattacks to a multi-modal model. We derive a lower bound on the performance of\nour MMCert under arbitrary adversarial attacks with bounded perturbations to\nboth modalities (e.g., in the context of auto-driving, we bound the number of\nchanged pixels in both RGB image and depth image). We evaluate our MMCert using\ntwo benchmark datasets: one for the multi-modal road segmentation task and the\nother for the multi-modal emotion recognition task. Moreover, we compare our\nMMCert with a state-of-the-art certified defense extended from unimodal models.\nOur experimental results show that our MMCert outperforms the baseline.\n", "date": "2024-03-28", "categories": []}
{"id": "2404.01828", "title": "Defense without Forgetting: Continual Adversarial Defense with\n  Anisotropic & Isotropic Pseudo Replay", "abstract": "  Deep neural networks have demonstrated susceptibility to adversarial attacks.\nAdversarial defense techniques often focus on one-shot setting to maintain\nrobustness against attack. However, new attacks can emerge in sequences in\nreal-world deployment scenarios. As a result, it is crucial for a defense model\nto constantly adapt to new attacks, but the adaptation process can lead to\ncatastrophic forgetting of previously defended against attacks. In this paper,\nwe discuss for the first time the concept of continual adversarial defense\nunder a sequence of attacks, and propose a lifelong defense baseline called\nAnisotropic \\& Isotropic Replay (AIR), which offers three advantages: (1)\nIsotropic replay ensures model consistency in the neighborhood distribution of\nnew data, indirectly aligning the output preference between old and new tasks.\n(2) Anisotropic replay enables the model to learn a compromise data manifold\nwith fresh mixed semantics for further replay constraints and potential future\nattacks. (3) A straightforward regularizer mitigates the 'plasticity-stability'\ntrade-off by aligning model output between new and old tasks. Experiment\nresults demonstrate that AIR can approximate or even exceed the empirical\nperformance upper bounds achieved by Joint Training.\n", "date": "2024-04-02", "categories": []}
{"id": "2404.01907", "title": "Humanizing Machine-Generated Content: Evading AI-Text Detection through\n  Adversarial Attack", "abstract": "  With the development of large language models (LLMs), detecting whether text\nis generated by a machine becomes increasingly challenging in the face of\nmalicious use cases like the spread of false information, protection of\nintellectual property, and prevention of academic plagiarism. While\nwell-trained text detectors have demonstrated promising performance on unseen\ntest data, recent research suggests that these detectors have vulnerabilities\nwhen dealing with adversarial attacks such as paraphrasing. In this paper, we\npropose a framework for a broader class of adversarial attacks, designed to\nperform minor perturbations in machine-generated content to evade detection. We\nconsider two attack settings: white-box and black-box, and employ adversarial\nlearning in dynamic scenarios to assess the potential enhancement of the\ncurrent detection model's robustness against such attacks. The empirical\nresults reveal that the current detection models can be compromised in as\nlittle as 10 seconds, leading to the misclassification of machine-generated\ntext as human-written content. Furthermore, we explore the prospect of\nimproving the model's robustness over iterative adversarial learning. Although\nsome improvements in model robustness are observed, practical applications\nstill face significant challenges. These findings shed light on the future\ndevelopment of AI-text detectors, emphasizing the need for more accurate and\nrobust detection methods.\n", "date": "2024-04-02", "categories": []}
{"id": "2404.02067", "title": "Red-Teaming Segment Anything Model", "abstract": "  Foundation models have emerged as pivotal tools, tackling many complex tasks\nthrough pre-training on vast datasets and subsequent fine-tuning for specific\napplications. The Segment Anything Model is one of the first and most\nwell-known foundation models for computer vision segmentation tasks. This work\npresents a multi-faceted red-teaming analysis that tests the Segment Anything\nModel against challenging tasks: (1) We analyze the impact of style transfer on\nsegmentation masks, demonstrating that applying adverse weather conditions and\nraindrops to dashboard images of city roads significantly distorts generated\nmasks. (2) We focus on assessing whether the model can be used for attacks on\nprivacy, such as recognizing celebrities' faces, and show that the model\npossesses some undesired knowledge in this task. (3) Finally, we check how\nrobust the model is to adversarial attacks on segmentation masks under text\nprompts. We not only show the effectiveness of popular white-box attacks and\nresistance to black-box attacks but also introduce a novel approach - Focused\nIterative Gradient Attack (FIGA) that combines white-box approaches to\nconstruct an efficient attack resulting in a smaller number of modified pixels.\nAll of our testing methods and analyses indicate a need for enhanced safety\nmeasures in foundation models for image segmentation.\n", "date": "2024-04-02", "categories": []}
{"id": "2305.04067", "title": "The Best Defense is Attack: Repairing Semantics in Textual Adversarial\n  Examples", "abstract": "  Recent studies have revealed the vulnerability of pre-trained language models\nto adversarial attacks. Existing adversarial defense techniques attempt to\nreconstruct adversarial examples within feature or text spaces. However, these\nmethods struggle to effectively repair the semantics in adversarial examples,\nresulting in unsatisfactory performance and limiting their practical utility.\nTo repair the semantics in adversarial examples, we introduce a novel approach\nnamed Reactive Perturbation Defocusing (Rapid). Rapid employs an adversarial\ndetector to identify fake labels of adversarial examples and leverage\nadversarial attackers to repair the semantics in adversarial examples. Our\nextensive experimental results conducted on four public datasets, convincingly\ndemonstrate the effectiveness of Rapid in various adversarial attack scenarios.\nTo address the problem of defense performance validation in previous works, we\nprovide a demonstration of adversarial detection and repair based on our work,\nwhich can be easily evaluated at https://tinyurl.com/22ercuf8.\n", "date": "2023-05-06", "categories": []}
{"id": "2311.17516", "title": "MMA-Diffusion: MultiModal Attack on Diffusion Models", "abstract": "  In recent years, Text-to-Image (T2I) models have seen remarkable\nadvancements, gaining widespread adoption. However, this progress has\ninadvertently opened avenues for potential misuse, particularly in generating\ninappropriate or Not-Safe-For-Work (NSFW) content. Our work introduces\nMMA-Diffusion, a framework that presents a significant and realistic threat to\nthe security of T2I models by effectively circumventing current defensive\nmeasures in both open-source models and commercial online services. Unlike\nprevious approaches, MMA-Diffusion leverages both textual and visual modalities\nto bypass safeguards like prompt filters and post-hoc safety checkers, thus\nexposing and highlighting the vulnerabilities in existing defense mechanisms.\n", "date": "2023-11-29", "categories": []}
{"id": "2312.05716", "title": "Initialization Matters for Adversarial Transfer Learning", "abstract": "  With the prevalence of the Pretraining-Finetuning paradigm in transfer\nlearning, the robustness of downstream tasks has become a critical concern. In\nthis work, we delve into adversarial robustness in transfer learning and reveal\nthe critical role of initialization, including both the pretrained model and\nthe linear head. First, we discover the necessity of an adversarially robust\npretrained model. Specifically, we reveal that with a standard pretrained\nmodel, Parameter-Efficient Finetuning (PEFT) methods either fail to be\nadversarially robust or continue to exhibit significantly degraded adversarial\nrobustness on downstream tasks, even with adversarial training during\nfinetuning. Leveraging a robust pretrained model, surprisingly, we observe that\na simple linear probing can outperform full finetuning and other PEFT methods\nwith random initialization on certain datasets. We further identify that linear\nprobing excels in preserving robustness from the robust pretraining. Based on\nthis, we propose Robust Linear Initialization (RoLI) for adversarial\nfinetuning, which initializes the linear head with the weights obtained by\nadversarial linear probing to maximally inherit the robustness from\npretraining. Across five different image classification datasets, we\ndemonstrate the effectiveness of RoLI and achieve new state-of-the-art results.\nOur code is available at \\url{https://github.com/DongXzz/RoLI}.\n", "date": "2023-12-10", "categories": []}
{"id": "2312.09935", "title": "LogoStyleFool: Vitiating Video Recognition Systems via Logo Style\n  Transfer", "abstract": "  Video recognition systems are vulnerable to adversarial examples. Recent\nstudies show that style transfer-based and patch-based unrestricted\nperturbations can effectively improve attack efficiency. These attacks,\nhowever, face two main challenges: 1) Adding large stylized perturbations to\nall pixels reduces the naturalness of the video and such perturbations can be\neasily detected. 2) Patch-based video attacks are not extensible to targeted\nattacks due to the limited search space of reinforcement learning that has been\nwidely used in video attacks recently. In this paper, we focus on the video\nblack-box setting and propose a novel attack framework named LogoStyleFool by\nadding a stylized logo to the clean video. We separate the attack into three\nstages: style reference selection, reinforcement-learning-based logo style\ntransfer, and perturbation optimization. We solve the first challenge by\nscaling down the perturbation range to a regional logo, while the second\nchallenge is addressed by complementing an optimization stage after\nreinforcement learning. Experimental results substantiate the overall\nsuperiority of LogoStyleFool over three state-of-the-art patch-based attacks in\nterms of attack performance and semantic preservation. Meanwhile, LogoStyleFool\nstill maintains its performance against two existing patch-based defense\nmethods. We believe that our research is beneficial in increasing the attention\nof the security community to such subregional style transfer attacks.\n", "date": "2023-12-15", "categories": []}
{"id": "2404.00461", "title": "Shortcuts Arising from Contrast: Effective and Covert Clean-Label\n  Attacks in Prompt-Based Learning", "abstract": "  Prompt-based learning paradigm has demonstrated remarkable efficacy in\nenhancing the adaptability of pretrained language models (PLMs), particularly\nin few-shot scenarios. However, this learning paradigm has been shown to be\nvulnerable to backdoor attacks. The current clean-label attack, employing a\nspecific prompt as a trigger, can achieve success without the need for external\ntriggers and ensure correct labeling of poisoned samples, which is more\nstealthy compared to the poisoned-label attack, but on the other hand, it faces\nsignificant issues with false activations and poses greater challenges,\nnecessitating a higher rate of poisoning. Using conventional negative data\naugmentation methods, we discovered that it is challenging to trade off between\neffectiveness and stealthiness in a clean-label setting. In addressing this\nissue, we are inspired by the notion that a backdoor acts as a shortcut and\nposit that this shortcut stems from the contrast between the trigger and the\ndata utilized for poisoning. In this study, we propose a method named\nContrastive Shortcut Injection (CSI), by leveraging activation values,\nintegrates trigger design and data selection strategies to craft stronger\nshortcut features. With extensive experiments on full-shot and few-shot text\nclassification tasks, we empirically validate CSI's high effectiveness and high\nstealthiness at low poisoning rates. Notably, we found that the two approaches\nplay leading roles in full-shot and few-shot settings, respectively.\n", "date": "2024-03-30", "categories": []}
{"id": "2404.00540", "title": "Embodied Active Defense: Leveraging Recurrent Feedback to Counter\n  Adversarial Patches", "abstract": "  The vulnerability of deep neural networks to adversarial patches has\nmotivated numerous defense strategies for boosting model robustness. However,\nthe prevailing defenses depend on single observation or pre-established\nadversary information to counter adversarial patches, often failing to be\nconfronted with unseen or adaptive adversarial attacks and easily exhibiting\nunsatisfying performance in dynamic 3D environments. Inspired by active human\nperception and recurrent feedback mechanisms, we develop Embodied Active\nDefense (EAD), a proactive defensive strategy that actively contextualizes\nenvironmental information to address misaligned adversarial patches in 3D\nreal-world settings. To achieve this, EAD develops two central recurrent\nsub-modules, i.e., a perception module and a policy module, to implement two\ncritical functions of active vision. These models recurrently process a series\nof beliefs and observations, facilitating progressive refinement of their\ncomprehension of the target object and enabling the development of strategic\nactions to counter adversarial patches in 3D environments. To optimize learning\nefficiency, we incorporate a differentiable approximation of environmental\ndynamics and deploy patches that are agnostic to the adversary strategies.\nExtensive experiments demonstrate that EAD substantially enhances robustness\nagainst a variety of patches within just a few steps through its action policy\nin safety-critical tasks (e.g., face recognition and object detection), without\ncompromising standard accuracy. Furthermore, due to the attack-agnostic\ncharacteristic, EAD facilitates excellent generalization to unseen attacks,\ndiminishing the averaged attack success rate by 95 percent across a range of\nunseen adversarial attacks.\n", "date": "2024-03-31", "categories": []}
{"id": "2209.11964", "title": "Strong Transferable Adversarial Attacks via Ensembled Asymptotically\n  Normal Distribution Learning", "abstract": "  Strong adversarial examples are crucial for evaluating and enhancing the\nrobustness of deep neural networks. However, the performance of popular attacks\nis usually sensitive, for instance, to minor image transformations, stemming\nfrom limited information -- typically only one input example, a handful of\nwhite-box source models, and undefined defense strategies. Hence, the crafted\nadversarial examples are prone to overfit the source model, which hampers their\ntransferability to unknown architectures. In this paper, we propose an approach\nnamed Multiple Asymptotically Normal Distribution Attacks (MultiANDA) which\nexplicitly characterize adversarial perturbations from a learned distribution.\nSpecifically, we approximate the posterior distribution over the perturbations\nby taking advantage of the asymptotic normality property of stochastic gradient\nascent (SGA), then employ the deep ensemble strategy as an effective proxy for\nBayesian marginalization in this process, aiming to estimate a mixture of\nGaussians that facilitates a more thorough exploration of the potential\noptimization space. The approximated posterior essentially describes the\nstationary distribution of SGA iterations, which captures the geometric\ninformation around the local optimum. Thus, MultiANDA allows drawing an\nunlimited number of adversarial perturbations for each input and reliably\nmaintains the transferability. Our proposed method outperforms ten\nstate-of-the-art black-box attacks on deep learning models with or without\ndefenses through extensive experiments on seven normally trained and seven\ndefense models.\n", "date": "2022-09-24", "categories": []}
{"id": "2310.00648", "title": "PETA: Parameter-Efficient Trojan Attacks", "abstract": "  Parameter-efficient fine-tuning (PEFT) enables efficient adaptation of\npre-trained language models (PLMs) to specific tasks. By tuning only a minimal\nset of (extra) parameters, PEFT achieves performance that is comparable to\nstandard fine-tuning. However, despite its prevalent use, the security\nimplications of PEFT remain largely unexplored. In this paper, we take the\ninitial steps and present PETA, a novel trojan attack that compromises the\nweights of PLMs by accounting for downstream adaptation through bilevel\noptimization: the upper-level objective embeds the backdoor into a model while\nthe lower-level objective simulates PEFT to both retain the PLM's task-specific\nperformance and ensure that the backdoor persists after fine-tuning. With\nextensive evaluation across a variety of downstream tasks and trigger designs,\nwe demonstrate PETA's effectiveness in terms of both attack success rate and\nclean accuracy, even when the attacker does not have full knowledge of the\nvictim user's training process.\n", "date": "2023-10-01", "categories": []}
{"id": "2310.18274", "title": "LipSim: A Provably Robust Perceptual Similarity Metric", "abstract": "  Recent years have seen growing interest in developing and applying perceptual\nsimilarity metrics. Research has shown the superiority of perceptual metrics\nover pixel-wise metrics in aligning with human perception and serving as a\nproxy for the human visual system. On the other hand, as perceptual metrics\nrely on neural networks, there is a growing concern regarding their resilience,\ngiven the established vulnerability of neural networks to adversarial attacks.\nIt is indeed logical to infer that perceptual metrics may inherit both the\nstrengths and shortcomings of neural networks. In this work, we demonstrate the\nvulnerability of state-of-the-art perceptual similarity metrics based on an\nensemble of ViT-based feature extractors to adversarial attacks. We then\npropose a framework to train a robust perceptual similarity metric called\nLipSim (Lipschitz Similarity Metric) with provable guarantees. By leveraging\n1-Lipschitz neural networks as the backbone, LipSim provides guarded areas\naround each data point and certificates for all perturbations within an\n$\\ell_2$ ball. Finally, a comprehensive set of experiments shows the\nperformance of LipSim in terms of natural and certified scores and on the image\nretrieval application. The code is available at\nhttps://github.com/SaraGhazanfari/LipSim.\n", "date": "2023-10-27", "categories": []}
{"id": "2211.01579", "title": "Data-free Defense of Black Box Models Against Adversarial Attacks", "abstract": "  Several companies often safeguard their trained deep models (i.e., details of\narchitecture, learnt weights, training details etc.) from third-party users by\nexposing them only as black boxes through APIs. Moreover, they may not even\nprovide access to the training data due to proprietary reasons or sensitivity\nconcerns. In this work, we propose a novel defense mechanism for black box\nmodels against adversarial attacks in a data-free set up. We construct\nsynthetic data via generative model and train surrogate network using model\nstealing techniques. To minimize adversarial contamination on perturbed\nsamples, we propose 'wavelet noise remover' (WNR) that performs discrete\nwavelet decomposition on input images and carefully select only a few important\ncoefficients determined by our 'wavelet coefficient selection module' (WCSM).\nTo recover the high-frequency content of the image after noise removal via WNR,\nwe further train a 'regenerator' network with an objective to retrieve the\ncoefficients such that the reconstructed image yields similar to original\npredictions on the surrogate model. At test time, WNR combined with trained\nregenerator network is prepended to the black box network, resulting in a high\nboost in adversarial accuracy. Our method improves the adversarial accuracy on\nCIFAR-10 by 38.98% and 32.01% on state-of-the-art Auto Attack compared to\nbaseline, even when the attacker uses surrogate architecture (Alexnet-half and\nAlexnet) similar to the black box architecture (Alexnet) with same model\nstealing strategy as defender. The code is available at\nhttps://github.com/vcl-iisc/data-free-black-box-defense\n", "date": "2022-11-03", "categories": []}
{"id": "2402.17729", "title": "Towards Fairness-Aware Adversarial Learning", "abstract": "  Although adversarial training (AT) has proven effective in enhancing the\nmodel's robustness, the recently revealed issue of fairness in robustness has\nnot been well addressed, i.e. the robust accuracy varies significantly among\ndifferent categories. In this paper, instead of uniformly evaluating the\nmodel's average class performance, we delve into the issue of robust fairness,\nby considering the worst-case distribution across various classes. We propose a\nnovel learning paradigm, named Fairness-Aware Adversarial Learning (FAAL). As a\ngeneralization of conventional AT, we re-define the problem of adversarial\ntraining as a min-max-max framework, to ensure both robustness and fairness of\nthe trained model. Specifically, by taking advantage of distributional robust\noptimization, our method aims to find the worst distribution among different\ncategories, and the solution is guaranteed to obtain the upper bound\nperformance with high probability. In particular, FAAL can fine-tune an unfair\nrobust model to be fair within only two epochs, without compromising the\noverall clean and robust accuracies. Extensive experiments on various image\ndatasets validate the superior performance and efficiency of the proposed FAAL\ncompared to other state-of-the-art methods.\n", "date": "2024-02-27", "categories": []}
{"id": "2403.19009", "title": "Towards Sustainable SecureML: Quantifying Carbon Footprint of\n  Adversarial Machine Learning", "abstract": "  The widespread adoption of machine learning (ML) across various industries\nhas raised sustainability concerns due to its substantial energy usage and\ncarbon emissions. This issue becomes more pressing in adversarial ML, which\nfocuses on enhancing model security against different network-based attacks.\nImplementing defenses in ML systems often necessitates additional computational\nresources and network security measures, exacerbating their environmental\nimpacts. In this paper, we pioneer the first investigation into adversarial\nML's carbon footprint, providing empirical evidence connecting greater model\nrobustness to higher emissions. Addressing the critical need to quantify this\ntrade-off, we introduce the Robustness Carbon Trade-off Index (RCTI). This\nnovel metric, inspired by economic elasticity principles, captures the\nsensitivity of carbon emissions to changes in adversarial robustness. We\ndemonstrate the RCTI through an experiment involving evasion attacks, analyzing\nthe interplay between robustness against attacks, performance, and carbon\nemissions.\n", "date": "2024-03-27", "categories": []}
{"id": "2403.19150", "title": "Towards Understanding Dual BN In Hybrid Adversarial Training", "abstract": "  There is a growing concern about applying batch normalization (BN) in\nadversarial training (AT), especially when the model is trained on both\nadversarial samples and clean samples (termed Hybrid-AT). With the assumption\nthat adversarial and clean samples are from two different domains, a common\npractice in prior works is to adopt Dual BN, where BN and BN are used for\nadversarial and clean branches, respectively. A popular belief for motivating\nDual BN is that estimating normalization statistics of this mixture\ndistribution is challenging and thus disentangling it for normalization\nachieves stronger robustness. In contrast to this belief, we reveal that\ndisentangling statistics plays a less role than disentangling affine parameters\nin model training. This finding aligns with prior work (Rebuffi et al., 2023),\nand we build upon their research for further investigations. We demonstrate\nthat the domain gap between adversarial and clean samples is not very large,\nwhich is counter-intuitive considering the significant influence of adversarial\nperturbation on the model accuracy. We further propose a two-task hypothesis\nwhich serves as the empirical foundation and a unified framework for Hybrid-AT\nimprovement. We also investigate Dual BN in test-time and reveal that affine\nparameters characterize the robustness during inference. Overall, our work\nsheds new light on understanding the mechanism of Dual BN in Hybrid-AT and its\nunderlying justification.\n", "date": "2024-03-28", "categories": []}
{"id": "2403.19326", "title": "MedBN: Robust Test-Time Adaptation against Malicious Test Samples", "abstract": "  Test-time adaptation (TTA) has emerged as a promising solution to address\nperformance decay due to unforeseen distribution shifts between training and\ntest data. While recent TTA methods excel in adapting to test data variations,\nsuch adaptability exposes a model to vulnerability against malicious examples,\nan aspect that has received limited attention. Previous studies have uncovered\nsecurity vulnerabilities within TTA even when a small proportion of the test\nbatch is maliciously manipulated. In response to the emerging threat, we\npropose median batch normalization (MedBN), leveraging the robustness of the\nmedian for statistics estimation within the batch normalization layer during\ntest-time inference. Our method is algorithm-agnostic, thus allowing seamless\nintegration with existing TTA frameworks. Our experimental results on benchmark\ndatasets, including CIFAR10-C, CIFAR100-C and ImageNet-C, consistently\ndemonstrate that MedBN outperforms existing approaches in maintaining robust\nperformance across different attack scenarios, encompassing both instant and\ncumulative attacks. Through extensive experiments, we show that our approach\nsustains the performance even in the absence of attacks, achieving a practical\nbalance between robustness and performance.\n", "date": "2024-03-28", "categories": []}
{"id": "2302.06912", "title": "Regret-Based Defense in Adversarial Reinforcement Learning", "abstract": "  Deep Reinforcement Learning (DRL) policies have been shown to be vulnerable\nto small adversarial noise in observations. Such adversarial noise can have\ndisastrous consequences in safety-critical environments. For instance, a\nself-driving car receiving adversarially perturbed sensory observations about\nnearby signs (e.g., a stop sign physically altered to be perceived as a speed\nlimit sign) or objects (e.g., cars altered to be recognized as trees) can be\nfatal. Existing approaches for making RL algorithms robust to an\nobservation-perturbing adversary have focused on reactive approaches that\niteratively improve against adversarial examples generated at each iteration.\nWhile such approaches have been shown to provide improvements over regular RL\nmethods, they are reactive and can fare significantly worse if certain\ncategories of adversarial examples are not generated during training. To that\nend, we pursue a more proactive approach that relies on directly optimizing a\nwell-studied robustness measure, regret instead of expected value. We provide a\nprincipled approach that minimizes maximum regret over a \"neighborhood\" of\nobservations to the received \"observation\". Our regret criterion can be used to\nmodify existing value- and policy-based Deep RL methods. We demonstrate that\nour approaches provide a significant improvement in performance across a wide\nvariety of benchmarks against leading approaches for robust Deep RL.\n", "date": "2023-02-14", "categories": []}
{"id": "2403.11656", "title": "LocalStyleFool: Regional Video Style Transfer Attack Using Segment\n  Anything Model", "abstract": "  Previous work has shown that well-crafted adversarial perturbations can\nthreaten the security of video recognition systems. Attackers can invade such\nmodels with a low query budget when the perturbations are semantic-invariant,\nsuch as StyleFool. Despite the query efficiency, the naturalness of the minutia\nareas still requires amelioration, since StyleFool leverages style transfer to\nall pixels in each frame. To close the gap, we propose LocalStyleFool, an\nimproved black-box video adversarial attack that superimposes regional\nstyle-transfer-based perturbations on videos. Benefiting from the popularity\nand scalably usability of Segment Anything Model (SAM), we first extract\ndifferent regions according to semantic information and then track them through\nthe video stream to maintain the temporal consistency. Then, we add\nstyle-transfer-based perturbations to several regions selected based on the\nassociative criterion of transfer-based gradient information and regional area.\nPerturbation fine adjustment is followed to make stylized videos adversarial.\nWe demonstrate that LocalStyleFool can improve both intra-frame and inter-frame\nnaturalness through a human-assessed survey, while maintaining competitive\nfooling rate and query efficiency. Successful experiments on the\nhigh-resolution dataset also showcase that scrupulous segmentation of SAM helps\nto improve the scalability of adversarial attacks under high-resolution data.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.17301", "title": "Physical 3D Adversarial Attacks against Monocular Depth Estimation in\n  Autonomous Driving", "abstract": "  Deep learning-based monocular depth estimation (MDE), extensively applied in\nautonomous driving, is known to be vulnerable to adversarial attacks. Previous\nphysical attacks against MDE models rely on 2D adversarial patches, so they\nonly affect a small, localized region in the MDE map but fail under various\nviewpoints. To address these limitations, we propose 3D Depth Fool\n(3D$^2$Fool), the first 3D texture-based adversarial attack against MDE models.\n3D$^2$Fool is specifically optimized to generate 3D adversarial textures\nagnostic to model types of vehicles and to have improved robustness in bad\nweather conditions, such as rain and fog. Experimental results validate the\nsuperior performance of our 3D$^2$Fool across various scenarios, including\nvehicles, MDE models, weather conditions, and viewpoints. Real-world\nexperiments with printed 3D textures on physical vehicle models further\ndemonstrate that our 3D$^2$Fool can cause an MDE error of over 10 meters.\n", "date": "2024-03-26", "categories": []}
{"id": "2403.18423", "title": "SemRoDe: Macro Adversarial Training to Learn Representations That are\n  Robust to Word-Level Attacks", "abstract": "  Language models (LMs) are indispensable tools for natural language processing\ntasks, but their vulnerability to adversarial attacks remains a concern. While\ncurrent research has explored adversarial training techniques, their\nimprovements to defend against word-level attacks have been limited. In this\nwork, we propose a novel approach called Semantic Robust Defence (SemRoDe), a\nMacro Adversarial Training strategy to enhance the robustness of LMs. Drawing\ninspiration from recent studies in the image domain, we investigate and later\nconfirm that in a discrete data setting such as language, adversarial samples\ngenerated via word substitutions do indeed belong to an adversarial domain\nexhibiting a high Wasserstein distance from the base domain. Our method learns\na robust representation that bridges these two domains. We hypothesize that if\nsamples were not projected into an adversarial domain, but instead to a domain\nwith minimal shift, it would improve attack robustness. We align the domains by\nincorporating a new distance-based objective. With this, our model is able to\nlearn more generalized representations by aligning the model's high-level\noutput features and therefore better handling unseen adversarial samples. This\nmethod can be generalized across word embeddings, even when they share minimal\noverlap at both vocabulary and word-substitution levels. To evaluate the\neffectiveness of our approach, we conduct experiments on BERT and RoBERTa\nmodels on three datasets. The results demonstrate promising state-of-the-art\nrobustness.\n", "date": "2024-03-27", "categories": []}
{"id": "2403.18587", "title": "The Impact of Uniform Inputs on Activation Sparsity and Energy-Latency\n  Attacks in Computer Vision", "abstract": "  Resource efficiency plays an important role for machine learning nowadays.\nThe energy and decision latency are two critical aspects to ensure a\nsustainable and practical application. Unfortunately, the energy consumption\nand decision latency are not robust against adversaries. Researchers have\nrecently demonstrated that attackers can compute and submit so-called sponge\nexamples at inference time to increase the energy consumption and decision\nlatency of neural networks. In computer vision, the proposed strategy crafts\ninputs with less activation sparsity which could otherwise be used to\naccelerate the computation. In this paper, we analyze the mechanism how these\nenergy-latency attacks reduce activation sparsity. In particular, we find that\ninput uniformity is a key enabler. A uniform image, that is, an image with\nmostly flat, uniformly colored surfaces, triggers more activations due to a\nspecific interplay of convolution, batch normalization, and ReLU activation.\nBased on these insights, we propose two new simple, yet effective strategies\nfor crafting sponge examples: sampling images from a probability distribution\nand identifying dense, yet inconspicuous inputs in natural datasets. We\nempirically examine our findings in a comprehensive evaluation with multiple\nimage classification models and show that our attack achieves the same sparsity\neffect as prior sponge-example methods, but at a fraction of computation\neffort. We also show that our sponge examples transfer between different neural\nnetworks. Finally, we discuss applications of our findings for the good by\nimproving efficiency by increasing sparsity.\n", "date": "2024-03-27", "categories": []}
{"id": "2306.11990", "title": "Physics-constrained Attack against Convolution-based Human Motion\n  Prediction", "abstract": "  Human motion prediction has achieved a brilliant performance with the help of\nconvolution-based neural networks. However, currently, there is no work\nevaluating the potential risk in human motion prediction when facing\nadversarial attacks. The adversarial attack will encounter problems against\nhuman motion prediction in naturalness and data scale. To solve the problems\nabove, we propose a new adversarial attack method that generates the worst-case\nperturbation by maximizing the human motion predictor's prediction error with\nphysical constraints. Specifically, we introduce a novel adaptable scheme that\nfacilitates the attack to suit the scale of the target pose and two physical\nconstraints to enhance the naturalness of the adversarial example. The\nevaluating experiments on three datasets show that the prediction errors of all\ntarget models are enlarged significantly, which means current convolution-based\nhuman motion prediction models are vulnerable to the proposed attack. Based on\nthe experimental results, we provide insights on how to enhance the adversarial\nrobustness of the human motion predictor and how to improve the adversarial\nattack against human motion prediction.\n", "date": "2023-06-21", "categories": []}
{"id": "2403.17188", "title": "LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning", "abstract": "  Backdoor attack poses a significant security threat to Deep Learning\napplications. Existing attacks are often not evasive to established backdoor\ndetection techniques. This susceptibility primarily stems from the fact that\nthese attacks typically leverage a universal trigger pattern or transformation\nfunction, such that the trigger can cause misclassification for any input. In\nresponse to this, recent papers have introduced attacks using sample-specific\ninvisible triggers crafted through special transformation functions. While\nthese approaches manage to evade detection to some extent, they reveal\nvulnerability to existing backdoor mitigation techniques. To address and\nenhance both evasiveness and resilience, we introduce a novel backdoor attack\nLOTUS. Specifically, it leverages a secret function to separate samples in the\nvictim class into a set of partitions and applies unique triggers to different\npartitions. Furthermore, LOTUS incorporates an effective trigger focusing\nmechanism, ensuring only the trigger corresponding to the partition can induce\nthe backdoor behavior. Extensive experimental results show that LOTUS can\nachieve high attack success rate across 4 datasets and 7 model structures, and\neffectively evading 13 backdoor detection and mitigation techniques. The code\nis available at https://github.com/Megum1/LOTUS.\n", "date": "2024-03-25", "categories": []}
{"id": "2403.17520", "title": "Boosting Adversarial Training via Fisher-Rao Norm-based Regularization", "abstract": "  Adversarial training is extensively utilized to improve the adversarial\nrobustness of deep neural networks. Yet, mitigating the degradation of standard\ngeneralization performance in adversarial-trained models remains an open\nproblem. This paper attempts to resolve this issue through the lens of model\ncomplexity. First, We leverage the Fisher-Rao norm, a geometrically invariant\nmetric for model complexity, to establish the non-trivial bounds of the\nCross-Entropy Loss-based Rademacher complexity for a ReLU-activated Multi-Layer\nPerceptron. Then we generalize a complexity-related variable, which is\nsensitive to the changes in model width and the trade-off factors in\nadversarial training. Moreover, intensive empirical evidence validates that\nthis variable highly correlates with the generalization gap of Cross-Entropy\nloss between adversarial-trained and standard-trained models, especially during\nthe initial and final phases of the training process. Building upon this\nobservation, we propose a novel regularization framework, called Logit-Oriented\nAdversarial Training (LOAT), which can mitigate the trade-off between\nrobustness and accuracy while imposing only a negligible increase in\ncomputational overhead. Our extensive experiments demonstrate that the proposed\nregularization strategy can boost the performance of the prevalent adversarial\ntraining algorithms, including PGD-AT, TRADES, TRADES (LSE), MART, and DM-AT,\nacross various network architectures. Our code will be available at\nhttps://github.com/TrustAI/LOAT.\n", "date": "2024-03-26", "categories": []}
{"id": "2308.10299", "title": "Boosting Adversarial Transferability by Block Shuffle and Rotation", "abstract": "  Adversarial examples mislead deep neural networks with imperceptible\nperturbations and have brought significant threats to deep learning. An\nimportant aspect is their transferability, which refers to their ability to\ndeceive other models, thus enabling attacks in the black-box setting. Though\nvarious methods have been proposed to boost transferability, the performance\nstill falls short compared with white-box attacks. In this work, we observe\nthat existing input transformation based attacks, one of the mainstream\ntransfer-based attacks, result in different attention heatmaps on various\nmodels, which might limit the transferability. We also find that breaking the\nintrinsic relation of the image can disrupt the attention heatmap of the\noriginal image. Based on this finding, we propose a novel input transformation\nbased attack called block shuffle and rotation (BSR). Specifically, BSR splits\nthe input image into several blocks, then randomly shuffles and rotates these\nblocks to construct a set of new images for gradient calculation. Empirical\nevaluations on the ImageNet dataset demonstrate that BSR could achieve\nsignificantly better transferability than the existing input transformation\nbased methods under single-model and ensemble-model settings. Combining BSR\nwith the current input transformation method can further improve the\ntransferability, which significantly outperforms the state-of-the-art methods.\nCode is available at https://github.com/Trustworthy-AI-Group/BSR\n", "date": "2023-08-20", "categories": []}
{"id": "2403.03967", "title": "Effect of Ambient-Intrinsic Dimension Gap on Adversarial Vulnerability", "abstract": "  The existence of adversarial attacks on machine learning models imperceptible\nto a human is still quite a mystery from a theoretical perspective. In this\nwork, we introduce two notions of adversarial attacks: natural or on-manifold\nattacks, which are perceptible by a human/oracle, and unnatural or off-manifold\nattacks, which are not. We argue that the existence of the off-manifold attacks\nis a natural consequence of the dimension gap between the intrinsic and ambient\ndimensions of the data. For 2-layer ReLU networks, we prove that even though\nthe dimension gap does not affect generalization performance on samples drawn\nfrom the observed data space, it makes the clean-trained model more vulnerable\nto adversarial perturbations in the off-manifold direction of the data space.\nOur main results provide an explicit relationship between the\n$\\ell_2,\\ell_{\\infty}$ attack strength of the on/off-manifold attack and the\ndimension gap.\n", "date": "2024-03-06", "categories": []}
{"id": "2403.15467", "title": "Don't be a Fool: Pooling Strategies in Offensive Language Detection from\n  User-Intended Adversarial Attacks", "abstract": "  Offensive language detection is an important task for filtering out abusive\nexpressions and improving online user experiences. However, malicious users\noften attempt to avoid filtering systems through the involvement of textual\nnoises. In this paper, we propose these evasions as user-intended adversarial\nattacks that insert special symbols or leverage the distinctive features of the\nKorean language. Furthermore, we introduce simple yet effective pooling\nstrategies in a layer-wise manner to defend against the proposed attacks,\nfocusing on the preceding layers not just the last layer to capture both\noffensiveness and token embeddings. We demonstrate that these pooling\nstrategies are more robust to performance degradation even when the attack rate\nis increased, without directly training of such patterns. Notably, we found\nthat models pre-trained on clean texts could achieve a comparable performance\nin detecting attacked offensive language, to models pre-trained on noisy texts\nby employing these pooling strategies.\n", "date": "2024-03-20", "categories": []}
{"id": "2403.16176", "title": "Subspace Defense: Discarding Adversarial Perturbations by Learning a\n  Subspace for Clean Signals", "abstract": "  Deep neural networks (DNNs) are notoriously vulnerable to adversarial attacks\nthat place carefully crafted perturbations on normal examples to fool DNNs. To\nbetter understand such attacks, a characterization of the features carried by\nadversarial examples is needed. In this paper, we tackle this challenge by\ninspecting the subspaces of sample features through spectral analysis. We first\nempirically show that the features of either clean signals or adversarial\nperturbations are redundant and span in low-dimensional linear subspaces\nrespectively with minimal overlap, and the classical low-dimensional subspace\nprojection can suppress perturbation features out of the subspace of clean\nsignals. This makes it possible for DNNs to learn a subspace where only\nfeatures of clean signals exist while those of perturbations are discarded,\nwhich can facilitate the distinction of adversarial examples. To prevent the\nresidual perturbations that is inevitable in subspace learning, we propose an\nindependence criterion to disentangle clean signals from perturbations.\nExperimental results show that the proposed strategy enables the model to\ninherently suppress adversaries, which not only boosts model robustness but\nalso motivates new directions of effective adversarial defense.\n", "date": "2024-03-24", "categories": []}
{"id": "2403.16365", "title": "Generating Potent Poisons and Backdoors from Scratch with Guided\n  Diffusion", "abstract": "  Modern neural networks are often trained on massive datasets that are web\nscraped with minimal human inspection. As a result of this insecure curation\npipeline, an adversary can poison or backdoor the resulting model by uploading\nmalicious data to the internet and waiting for a victim to scrape and train on\nit. Existing approaches for creating poisons and backdoors start with randomly\nsampled clean data, called base samples, and then modify those samples to craft\npoisons. However, some base samples may be significantly more amenable to\npoisoning than others. As a result, we may be able to craft more potent poisons\nby carefully choosing the base samples. In this work, we use guided diffusion\nto synthesize base samples from scratch that lead to significantly more potent\npoisons and backdoors than previous state-of-the-art attacks. Our Guided\nDiffusion Poisoning (GDP) base samples can be combined with any downstream\npoisoning or backdoor attack to boost its effectiveness. Our implementation\ncode is publicly available at: https://github.com/hsouri/GDP .\n", "date": "2024-03-25", "categories": []}
{"id": "2403.16405", "title": "Ensemble Adversarial Defense via Integration of Multiple Dispersed Low\n  Curvature Models", "abstract": "  The integration of an ensemble of deep learning models has been extensively\nexplored to enhance defense against adversarial attacks. The diversity among\nsub-models increases the attack cost required to deceive the majority of the\nensemble, thereby improving the adversarial robustness. While existing\napproaches mainly center on increasing diversity in feature representations or\ndispersion of first-order gradients with respect to input, the limited\ncorrelation between these diversity metrics and adversarial robustness\nconstrains the performance of ensemble adversarial defense. In this work, we\naim to enhance ensemble diversity by reducing attack transferability. We\nidentify second-order gradients, which depict the loss curvature, as a key\nfactor in adversarial robustness. Computing the Hessian matrix involved in\nsecond-order gradients is computationally expensive. To address this, we\napproximate the Hessian-vector product using differential approximation. Given\nthat low curvature provides better robustness, our ensemble model was designed\nto consider the influence of curvature among different sub-models. We introduce\na novel regularizer to train multiple more-diverse low-curvature network\nmodels. Extensive experiments across various datasets demonstrate that our\nensemble model exhibits superior robustness against a range of attacks,\nunderscoring the effectiveness of our approach.\n", "date": "2024-03-25", "categories": []}
{"id": "2403.16782", "title": "The Anatomy of Adversarial Attacks: Concept-based XAI Dissection", "abstract": "  Adversarial attacks (AAs) pose a significant threat to the reliability and\nrobustness of deep neural networks. While the impact of these attacks on model\npredictions has been extensively studied, their effect on the learned\nrepresentations and concepts within these models remains largely unexplored. In\nthis work, we perform an in-depth analysis of the influence of AAs on the\nconcepts learned by convolutional neural networks (CNNs) using eXplainable\nartificial intelligence (XAI) techniques. Through an extensive set of\nexperiments across various network architectures and targeted AA techniques, we\nunveil several key findings. First, AAs induce substantial alterations in the\nconcept composition within the feature space, introducing new concepts or\nmodifying existing ones. Second, the adversarial perturbation itself can be\nlinearly decomposed into a set of latent vector components, with a subset of\nthese being responsible for the attack's success. Notably, we discover that\nthese components are target-specific, i.e., are similar for a given target\nclass throughout different AA techniques and starting classes. Our findings\nprovide valuable insights into the nature of AAs and their impact on learned\nrepresentations, paving the way for the development of more robust and\ninterpretable deep learning models, as well as effective defenses against\nadversarial threats.\n", "date": "2024-03-25", "categories": []}
{"id": "2401.11170", "title": "Inducing High Energy-Latency of Large Vision-Language Models with\n  Verbose Images", "abstract": "  Large vision-language models (VLMs) such as GPT-4 have achieved exceptional\nperformance across various multi-modal tasks. However, the deployment of VLMs\nnecessitates substantial energy consumption and computational resources. Once\nattackers maliciously induce high energy consumption and latency time\n(energy-latency cost) during inference of VLMs, it will exhaust computational\nresources. In this paper, we explore this attack surface about availability of\nVLMs and aim to induce high energy-latency cost during inference of VLMs. We\nfind that high energy-latency cost during inference of VLMs can be manipulated\nby maximizing the length of generated sequences. To this end, we propose\nverbose images, with the goal of crafting an imperceptible perturbation to\ninduce VLMs to generate long sentences during inference. Concretely, we design\nthree loss objectives. First, a loss is proposed to delay the occurrence of\nend-of-sequence (EOS) token, where EOS token is a signal for VLMs to stop\ngenerating further tokens. Moreover, an uncertainty loss and a token diversity\nloss are proposed to increase the uncertainty over each generated token and the\ndiversity among all tokens of the whole generated sequence, respectively, which\ncan break output dependency at token-level and sequence-level. Furthermore, a\ntemporal weight adjustment algorithm is proposed, which can effectively balance\nthese losses. Extensive experiments demonstrate that our verbose images can\nincrease the length of generated sequences by 7.87 times and 8.56 times\ncompared to original images on MS-COCO and ImageNet datasets, which presents\npotential challenges for various applications. Our code is available at\nhttps://github.com/KuofengGao/Verbose_Images.\n", "date": "2024-01-20", "categories": []}
{"id": "2403.14731", "title": "Reversible Jump Attack to Textual Classifiers with Modification\n  Reduction", "abstract": "  Recent studies on adversarial examples expose vulnerabilities of natural\nlanguage processing (NLP) models. Existing techniques for generating\nadversarial examples are typically driven by deterministic hierarchical rules\nthat are agnostic to the optimal adversarial examples, a strategy that often\nresults in adversarial samples with a suboptimal balance between magnitudes of\nchanges and attack successes. To this end, in this research we propose two\nalgorithms, Reversible Jump Attack (RJA) and Metropolis-Hasting Modification\nReduction (MMR), to generate highly effective adversarial examples and to\nimprove the imperceptibility of the examples, respectively. RJA utilizes a\nnovel randomization mechanism to enlarge the search space and efficiently\nadapts to a number of perturbed words for adversarial examples. With these\ngenerated adversarial examples, MMR applies the Metropolis-Hasting sampler to\nenhance the imperceptibility of adversarial examples. Extensive experiments\ndemonstrate that RJA-MMR outperforms current state-of-the-art methods in attack\nperformance, imperceptibility, fluency and grammar correctness.\n", "date": "2024-03-21", "categories": []}
{"id": "2403.14778", "title": "Diffusion Attack: Leveraging Stable Diffusion for Naturalistic Image\n  Attacking", "abstract": "  In Virtual Reality (VR), adversarial attack remains a significant security\nthreat. Most deep learning-based methods for physical and digital adversarial\nattacks focus on enhancing attack performance by crafting adversarial examples\nthat contain large printable distortions that are easy for human observers to\nidentify. However, attackers rarely impose limitations on the naturalness and\ncomfort of the appearance of the generated attack image, resulting in a\nnoticeable and unnatural attack. To address this challenge, we propose a\nframework to incorporate style transfer to craft adversarial inputs of natural\nstyles that exhibit minimal detectability and maximum natural appearance, while\nmaintaining superior attack capabilities.\n", "date": "2024-03-21", "categories": []}
{"id": "2403.15207", "title": "Robust optimization for adversarial learning with finite sample\n  complexity guarantees", "abstract": "  Decision making and learning in the presence of uncertainty has attracted\nsignificant attention in view of the increasing need to achieve robust and\nreliable operations. In the case where uncertainty stems from the presence of\nadversarial attacks this need is becoming more prominent. In this paper we\nfocus on linear and nonlinear classification problems and propose a novel\nadversarial training method for robust classifiers, inspired by Support Vector\nMachine (SVM) margins. We view robustness under a data driven lens, and derive\nfinite sample complexity bounds for both linear and non-linear classifiers in\nbinary and multi-class scenarios. Notably, our bounds match natural\nclassifiers' complexity. Our algorithm minimizes a worst-case surrogate loss\nusing Linear Programming (LP) and Second Order Cone Programming (SOCP) for\nlinear and non-linear models. Numerical experiments on the benchmark MNIST and\nCIFAR10 datasets show our approach's comparable performance to state-of-the-art\nmethods, without needing adversarial examples during training. Our work offers\na comprehensive framework for enhancing binary linear and non-linear classifier\nrobustness, embedding robustness in learning under the presence of adversaries.\n", "date": "2024-03-22", "categories": []}
{"id": "2403.13507", "title": "FMM-Attack: A Flow-based Multi-modal Adversarial Attack on Video-based\n  LLMs", "abstract": "  Despite the remarkable performance of video-based large language models\n(LLMs), their adversarial threat remains unexplored. To fill this gap, we\npropose the first adversarial attack tailored for video-based LLMs by crafting\nflow-based multi-modal adversarial perturbations on a small fraction of frames\nwithin a video, dubbed FMM-Attack. Extensive experiments show that our attack\ncan effectively induce video-based LLMs to generate incorrect answers when\nvideos are added with imperceptible adversarial perturbations. Intriguingly,\nour FMM-Attack can also induce garbling in the model output, prompting\nvideo-based LLMs to hallucinate. Overall, our observations inspire a further\nunderstanding of multi-modal robustness and safety-related feature alignment\nacross different modalities, which is of great importance for various large\nmulti-modal models. Our code is available at\nhttps://github.com/THU-Kingmin/FMM-Attack.\n", "date": "2024-03-20", "categories": []}
{"id": "2403.14489", "title": "Adversary-Robust Graph-Based Learning of WSIs", "abstract": "  Enhancing the robustness of deep learning models against adversarial attacks\nis crucial, especially in critical domains like healthcare where significant\nfinancial interests heighten the risk of such attacks. Whole slide images\n(WSIs) are high-resolution, digitized versions of tissue samples mounted on\nglass slides, scanned using sophisticated imaging equipment. The digital\nanalysis of WSIs presents unique challenges due to their gigapixel size and\nmulti-resolution storage format. In this work, we aim at improving the\nrobustness of cancer Gleason grading classification systems against adversarial\nattacks, addressing challenges at both the image and graph levels. As regards\nthe proposed algorithm, we develop a novel and innovative graph-based model\nwhich utilizes GNN to extract features from the graph representation of WSIs. A\ndenoising module, along with a pooling layer is incorporated to manage the\nimpact of adversarial attacks on the WSIs. The process concludes with a\ntransformer module that classifies various grades of prostate cancer based on\nthe processed data. To assess the effectiveness of the proposed method, we\nconducted a comparative analysis using two scenarios. Initially, we trained and\ntested the model without the denoiser using WSIs that had not been exposed to\nany attack. We then introduced a range of attacks at either the image or graph\nlevel and processed them through the proposed network. The performance of the\nmodel was evaluated in terms of accuracy and kappa scores. The results from\nthis comparison showed a significant improvement in cancer diagnosis accuracy,\nhighlighting the robustness and efficiency of the proposed method in handling\nadversarial challenges in the context of medical imaging.\n", "date": "2024-03-21", "categories": []}
{"id": "2403.13017", "title": "Impart: An Imperceptible and Effective Label-Specific Backdoor Attack", "abstract": "  Backdoor attacks have been shown to impose severe threats to real\nsecurity-critical scenarios. Although previous works can achieve high attack\nsuccess rates, they either require access to victim models which may\nsignificantly reduce their threats in practice, or perform visually noticeable\nin stealthiness. Besides, there is still room to improve the attack success\nrates in the scenario that different poisoned samples may have different target\nlabels (a.k.a., the all-to-all setting). In this study, we propose a novel\nimperceptible backdoor attack framework, named Impart, in the scenario where\nthe attacker has no access to the victim model. Specifically, in order to\nenhance the attack capability of the all-to-all setting, we first propose a\nlabel-specific attack. Different from previous works which try to find an\nimperceptible pattern and add it to the source image as the poisoned image, we\nthen propose to generate perturbations that align with the target label in the\nimage feature by a surrogate model. In this way, the generated poisoned images\nare attached with knowledge about the target class, which significantly\nenhances the attack capability.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.13018", "title": "Invisible Backdoor Attack Through Singular Value Decomposition", "abstract": "  With the widespread application of deep learning across various domains,\nconcerns about its security have grown significantly. Among these, backdoor\nattacks pose a serious security threat to deep neural networks (DNNs). In\nrecent years, backdoor attacks on neural networks have become increasingly\nsophisticated, aiming to compromise the security and trustworthiness of models\nby implanting hidden, unauthorized functionalities or triggers, leading to\nmisleading predictions or behaviors. To make triggers less perceptible and\nimperceptible, various invisible backdoor attacks have been proposed. However,\nmost of them only consider invisibility in the spatial domain, making it easy\nfor recent defense methods to detect the generated toxic images.To address\nthese challenges, this paper proposes an invisible backdoor attack called DEBA.\nDEBA leverages the mathematical properties of Singular Value Decomposition\n(SVD) to embed imperceptible backdoors into models during the training phase,\nthereby causing them to exhibit predefined malicious behavior under specific\ntrigger conditions. Specifically, we first perform SVD on images, and then\nreplace the minor features of trigger images with those of clean images, using\nthem as triggers to ensure the effectiveness of the attack. As minor features\nare scattered throughout the entire image, the major features of clean images\nare preserved, making poisoned images visually indistinguishable from clean\nones. Extensive experimental evaluations demonstrate that DEBA is highly\neffective, maintaining high perceptual quality and a high attack success rate\nfor poisoned images. Furthermore, we assess the performance of DEBA under\nexisting defense measures, showing that it is robust and capable of\nsignificantly evading and resisting the effects of these defense measures.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.13523", "title": "Have You Poisoned My Data? Defending Neural Networks against Data\n  Poisoning", "abstract": "  The unprecedented availability of training data fueled the rapid development\nof powerful neural networks in recent years. However, the need for such large\namounts of data leads to potential threats such as poisoning attacks:\nadversarial manipulations of the training data aimed at compromising the\nlearned model to achieve a given adversarial goal.\n  This paper investigates defenses against clean-label poisoning attacks and\nproposes a novel approach to detect and filter poisoned datapoints in the\ntransfer learning setting. We define a new characteristic vector representation\nof datapoints and show that it effectively captures the intrinsic properties of\nthe data distribution. Through experimental analysis, we demonstrate that\neffective poisons can be successfully differentiated from clean points in the\ncharacteristic vector space. We thoroughly evaluate our proposed approach and\ncompare it to existing state-of-the-art defenses using multiple architectures,\ndatasets, and poison budgets. Our evaluation shows that our proposal\noutperforms existing approaches in defense rate and final trained model\nperformance across all experimental settings.\n", "date": "2024-03-20", "categories": []}
{"id": "2306.11035", "title": "Adversarial Training Should Be Cast as a Non-Zero-Sum Game", "abstract": "  One prominent approach toward resolving the adversarial vulnerability of deep\nneural networks is the two-player zero-sum paradigm of adversarial training, in\nwhich predictors are trained against adversarially chosen perturbations of\ndata. Despite the promise of this approach, algorithms based on this paradigm\nhave not engendered sufficient levels of robustness and suffer from\npathological behavior like robust overfitting. To understand this shortcoming,\nwe first show that the commonly used surrogate-based relaxation used in\nadversarial training algorithms voids all guarantees on the robustness of\ntrained classifiers. The identification of this pitfall informs a novel\nnon-zero-sum bilevel formulation of adversarial training, wherein each player\noptimizes a different objective function. Our formulation yields a simple\nalgorithmic framework that matches and in some cases outperforms\nstate-of-the-art attacks, attains comparable levels of robustness to standard\nadversarial training algorithms, and does not suffer from robust overfitting.\n", "date": "2023-06-19", "categories": []}
{"id": "2309.01582", "title": "Improving Visual Quality and Transferability of Adversarial Attacks on\n  Face Recognition Simultaneously with Adversarial Restoration", "abstract": "  Adversarial face examples possess two critical properties: Visual Quality and\nTransferability. However, existing approaches rarely address these properties\nsimultaneously, leading to subpar results. To address this issue, we propose a\nnovel adversarial attack technique known as Adversarial Restoration\n(AdvRestore), which enhances both visual quality and transferability of\nadversarial face examples by leveraging a face restoration prior. In our\napproach, we initially train a Restoration Latent Diffusion Model (RLDM)\ndesigned for face restoration. Subsequently, we employ the inference process of\nRLDM to generate adversarial face examples. The adversarial perturbations are\napplied to the intermediate features of RLDM. Additionally, by treating RLDM\nface restoration as a sibling task, the transferability of the generated\nadversarial face examples is further improved. Our experimental results\nvalidate the effectiveness of the proposed attack method.\n", "date": "2023-09-04", "categories": []}
{"id": "2403.10801", "title": "Securely Fine-tuning Pre-trained Encoders Against Adversarial Examples", "abstract": "  With the evolution of self-supervised learning, the pre-training paradigm has\nemerged as a predominant solution within the deep learning landscape. Model\nproviders furnish pre-trained encoders designed to function as versatile\nfeature extractors, enabling downstream users to harness the benefits of\nexpansive models with minimal effort through fine-tuning. Nevertheless, recent\nworks have exposed a vulnerability in pre-trained encoders, highlighting their\nsusceptibility to downstream-agnostic adversarial examples (DAEs) meticulously\ncrafted by attackers. The lingering question pertains to the feasibility of\nfortifying the robustness of downstream models against DAEs, particularly in\nscenarios where the pre-trained encoders are publicly accessible to the\nattackers.\n  In this paper, we initially delve into existing defensive mechanisms against\nadversarial examples within the pre-training paradigm. Our findings reveal that\nthe failure of current defenses stems from the domain shift between\npre-training data and downstream tasks, as well as the sensitivity of encoder\nparameters. In response to these challenges, we propose Genetic\nEvolution-Nurtured Adversarial Fine-tuning (Gen-AF), a two-stage adversarial\nfine-tuning approach aimed at enhancing the robustness of downstream models.\nOur extensive experiments, conducted across ten self-supervised training\nmethods and six datasets, demonstrate that Gen-AF attains high testing accuracy\nand robust testing accuracy against state-of-the-art DAEs.\n", "date": "2024-03-16", "categories": []}
{"id": "2403.12077", "title": "Evaluating Robustness of Generative Search Engine on Adversarial Factual\n  Questions", "abstract": "  Generative search engines have the potential to transform how people seek\ninformation online, but generated responses from existing large language models\n(LLMs)-backed generative search engines may not always be accurate.\nNonetheless, retrieval-augmented generation exacerbates safety concerns, since\nadversaries may successfully evade the entire system by subtly manipulating the\nmost vulnerable part of a claim. To this end, we propose evaluating the\nrobustness of generative search engines in the realistic and high-risk setting,\nwhere adversaries have only black-box system access and seek to deceive the\nmodel into returning incorrect responses. Through a comprehensive human\nevaluation of various generative search engines, such as Bing Chat,\nPerplexityAI, and YouChat across diverse queries, we demonstrate the\neffectiveness of adversarial factual questions in inducing incorrect responses.\nMoreover, retrieval-augmented generation exhibits a higher susceptibility to\nfactual errors compared to LLMs without retrieval. These findings highlight the\npotential security risks of these systems and emphasize the need for rigorous\nevaluation before deployment.\n", "date": "2024-02-25", "categories": []}
{"id": "2403.12693", "title": "As Firm As Their Foundations: Can open-sourced foundation models be used\n  to create adversarial examples for downstream tasks?", "abstract": "  Foundation models pre-trained on web-scale vision-language data, such as\nCLIP, are widely used as cornerstones of powerful machine learning systems.\nWhile pre-training offers clear advantages for downstream learning, it also\nendows downstream models with shared adversarial vulnerabilities that can be\neasily identified through the open-sourced foundation model. In this work, we\nexpose such vulnerabilities in CLIP's downstream models and show that\nfoundation models can serve as a basis for attacking their downstream systems.\nIn particular, we propose a simple yet effective adversarial attack strategy\ntermed Patch Representation Misalignment (PRM). Solely based on open-sourced\nCLIP vision encoders, this method produces adversaries that simultaneously fool\nmore than 20 downstream models spanning 4 common vision-language tasks\n(semantic segmentation, object detection, image captioning and visual\nquestion-answering). Our findings highlight the concerning safety risks\nintroduced by the extensive usage of public foundational models in the\ndevelopment of downstream systems, calling for extra caution in these\nscenarios.\n", "date": "2024-03-19", "categories": []}
{"id": "2304.00436", "title": "Instance-Level Trojan Attacks on Visual Question Answering via\n  Adversarial Learning in Neuron Activation Space", "abstract": "  Trojan attacks embed perturbations in input data leading to malicious\nbehavior in neural network models. A combination of various Trojans in\ndifferent modalities enables an adversary to mount a sophisticated attack on\nmultimodal learning such as Visual Question Answering (VQA). However,\nmultimodal Trojans in conventional methods are susceptible to parameter\nadjustment during processes such as fine-tuning. To this end, we propose an\ninstance-level multimodal Trojan attack on VQA that efficiently adapts to\nfine-tuned models through a dual-modality adversarial learning method. This\nmethod compromises two specific neurons in a specific perturbation layer in the\npretrained model to produce overly large neuron activations. Then, a malicious\ncorrelation between these overactive neurons and the malicious output of a\nfine-tuned model is established through adversarial learning. Extensive\nexperiments are conducted using the VQA-v2 dataset, based on a wide range of\nmetrics including sample efficiency, stealthiness, and robustness. The proposed\nattack demonstrates enhanced performance with diverse vision and text Trojans\ntailored for each sample. We demonstrate that the proposed attack can be\nefficiently adapted to different fine-tuned models, by injecting only a few\nshots of Trojan samples. Moreover, we investigate the attack performance under\nconventional defenses, where the defenses cannot effectively mitigate the\nattack.\n", "date": "2023-04-02", "categories": []}
{"id": "2305.13991", "title": "Expressive Losses for Verified Robustness via Convex Combinations", "abstract": "  In order to train networks for verified adversarial robustness, it is common\nto over-approximate the worst-case loss over perturbation regions, resulting in\nnetworks that attain verifiability at the expense of standard performance. As\nshown in recent work, better trade-offs between accuracy and robustness can be\nobtained by carefully coupling adversarial training with over-approximations.\nWe hypothesize that the expressivity of a loss function, which we formalize as\nthe ability to span a range of trade-offs between lower and upper bounds to the\nworst-case loss through a single parameter (the over-approximation\ncoefficient), is key to attaining state-of-the-art performance. To support our\nhypothesis, we show that trivial expressive losses, obtained via convex\ncombinations between adversarial attacks and IBP bounds, yield state-of-the-art\nresults across a variety of settings in spite of their conceptual simplicity.\nWe provide a detailed analysis of the relationship between the\nover-approximation coefficient and performance profiles across different\nexpressive losses, showing that, while expressivity is essential, better\napproximations of the worst-case loss are not necessarily linked to superior\nrobustness-accuracy trade-offs.\n", "date": "2023-05-23", "categories": []}
{"id": "2306.13587", "title": "Creating Valid Adversarial Examples of Malware", "abstract": "  Machine learning is becoming increasingly popular as a go-to approach for\nmany tasks due to its world-class results. As a result, antivirus developers\nare incorporating machine learning models into their products. While these\nmodels improve malware detection capabilities, they also carry the disadvantage\nof being susceptible to adversarial attacks. Although this vulnerability has\nbeen demonstrated for many models in white-box settings, a black-box attack is\nmore applicable in practice for the domain of malware detection. We present a\ngenerator of adversarial malware examples using reinforcement learning\nalgorithms. The reinforcement learning agents utilize a set of\nfunctionality-preserving modifications, thus creating valid adversarial\nexamples. Using the proximal policy optimization (PPO) algorithm, we achieved\nan evasion rate of 53.84% against the gradient-boosted decision tree (GBDT)\nmodel. The PPO agent previously trained against the GBDT classifier scored an\nevasion rate of 11.41% against the neural network-based classifier MalConv and\nan average evasion rate of 2.31% against top antivirus programs. Furthermore,\nwe discovered that random application of our functionality-preserving portable\nexecutable modifications successfully evades leading antivirus engines, with an\naverage evasion rate of 11.65%. These findings indicate that machine\nlearning-based models used in malware detection systems are vulnerable to\nadversarial attacks and that better safeguards need to be taken to protect\nthese systems.\n", "date": "2023-06-23", "categories": []}
{"id": "2310.17645", "title": "PubDef: Defending Against Transfer Attacks From Public Models", "abstract": "  Adversarial attacks have been a looming and unaddressed threat in the\nindustry. However, through a decade-long history of the robustness evaluation\nliterature, we have learned that mounting a strong or optimal attack is\nchallenging. It requires both machine learning and domain expertise. In other\nwords, the white-box threat model, religiously assumed by a large majority of\nthe past literature, is unrealistic. In this paper, we propose a new practical\nthreat model where the adversary relies on transfer attacks through publicly\navailable surrogate models. We argue that this setting will become the most\nprevalent for security-sensitive applications in the future. We evaluate the\ntransfer attacks in this setting and propose a specialized defense method based\non a game-theoretic perspective. The defenses are evaluated under 24 public\nmodels and 11 attack algorithms across three datasets (CIFAR-10, CIFAR-100, and\nImageNet). Under this threat model, our defense, PubDef, outperforms the\nstate-of-the-art white-box adversarial training by a large margin with almost\nno loss in the normal accuracy. For instance, on ImageNet, our defense achieves\n62% accuracy under the strongest transfer attack vs only 36% of the best\nadversarially trained model. Its accuracy when not under attack is only 2%\nlower than that of an undefended model (78% vs 80%). We release our code at\nhttps://github.com/wagner-group/pubdef.\n", "date": "2023-10-26", "categories": []}
{"id": "2401.06637", "title": "Adversarial Examples are Misaligned in Diffusion Model Manifolds", "abstract": "  In recent years, diffusion models (DMs) have drawn significant attention for\ntheir success in approximating data distributions, yielding state-of-the-art\ngenerative results. Nevertheless, the versatility of these models extends\nbeyond their generative capabilities to encompass various vision applications,\nsuch as image inpainting, segmentation, adversarial robustness, among others.\nThis study is dedicated to the investigation of adversarial attacks through the\nlens of diffusion models. However, our objective does not involve enhancing the\nadversarial robustness of image classifiers. Instead, our focus lies in\nutilizing the diffusion model to detect and analyze the anomalies introduced by\nthese attacks on images. To that end, we systematically examine the alignment\nof the distributions of adversarial examples when subjected to the process of\ntransformation using diffusion models. The efficacy of this approach is\nassessed across CIFAR-10 and ImageNet datasets, including varying image sizes\nin the latter. The results demonstrate a notable capacity to discriminate\neffectively between benign and attacked images, providing compelling evidence\nthat adversarial instances do not align with the learned manifold of the DMs.\n", "date": "2024-01-12", "categories": []}
{"id": "2403.10562", "title": "Counter-Samples: A Stateless Strategy to Neutralize Black Box\n  Adversarial Attacks", "abstract": "  Our paper presents a novel defence against black box attacks, where attackers\nuse the victim model as an oracle to craft their adversarial examples. Unlike\ntraditional preprocessing defences that rely on sanitizing input samples, our\nstateless strategy counters the attack process itself. For every query we\nevaluate a counter-sample instead, where the counter-sample is the original\nsample optimized against the attacker's objective. By countering every black\nbox query with a targeted white box optimization, our strategy effectively\nintroduces an asymmetry to the game to the defender's advantage. This defence\nnot only effectively misleads the attacker's search for an adversarial example,\nit also preserves the model's accuracy on legitimate inputs and is generic to\nmultiple types of attacks.\n  We demonstrate that our approach is remarkably effective against\nstate-of-the-art black box attacks and outperforms existing defences for both\nthe CIFAR-10 and ImageNet datasets. Additionally, we also show that the\nproposed defence is robust against strong adversaries as well.\n", "date": "2024-03-14", "categories": []}
{"id": "2403.10935", "title": "Understanding Robustness of Visual State Space Models for Image\n  Classification", "abstract": "  Visual State Space Model (VMamba) has recently emerged as a promising\narchitecture, exhibiting remarkable performance in various computer vision\ntasks. However, its robustness has not yet been thoroughly studied. In this\npaper, we delve into the robustness of this architecture through comprehensive\ninvestigations from multiple perspectives. Firstly, we investigate its\nrobustness to adversarial attacks, employing both whole-image and\npatch-specific adversarial attacks. Results demonstrate superior adversarial\nrobustness compared to Transformer architectures while revealing scalability\nweaknesses. Secondly, the general robustness of VMamba is assessed against\ndiverse scenarios, including natural adversarial examples, out-of-distribution\ndata, and common corruptions. VMamba exhibits exceptional generalizability with\nout-of-distribution data but shows scalability weaknesses against natural\nadversarial examples and common corruptions. Additionally, we explore VMamba's\ngradients and back-propagation during white-box attacks, uncovering unique\nvulnerabilities and defensive capabilities of its novel components. Lastly, the\nsensitivity of VMamba to image structure variations is examined, highlighting\nvulnerabilities associated with the distribution of disturbance areas and\nspatial information, with increased susceptibility closer to the image center.\nThrough these comprehensive studies, we contribute to a deeper understanding of\nVMamba's robustness, providing valuable insights for refining and advancing the\ncapabilities of deep neural networks in computer vision applications.\n", "date": "2024-03-16", "categories": []}
{"id": "2403.11082", "title": "RobustSentEmbed: Robust Sentence Embeddings Using Adversarial\n  Self-Supervised Contrastive Learning", "abstract": "  Pre-trained language models (PLMs) have consistently demonstrated outstanding\nperformance across a diverse spectrum of natural language processing tasks.\nNevertheless, despite their success with unseen data, current PLM-based\nrepresentations often exhibit poor robustness in adversarial settings. In this\npaper, we introduce RobustSentEmbed, a self-supervised sentence embedding\nframework designed to improve both generalization and robustness in diverse\ntext representation tasks and against a diverse set of adversarial attacks.\nThrough the generation of high-risk adversarial perturbations and their\nutilization in a novel objective function, RobustSentEmbed adeptly learns\nhigh-quality and robust sentence embeddings. Our experiments confirm the\nsuperiority of RobustSentEmbed over state-of-the-art representations.\nSpecifically, Our framework achieves a significant reduction in the success\nrate of various adversarial attacks, notably reducing the BERTAttack success\nrate by almost half (from 75.51\\% to 38.81\\%). The framework also yields\nimprovements of 1.59\\% and 0.23\\% in semantic textual similarity tasks and\nvarious transfer tasks, respectively.\n", "date": "2024-03-17", "categories": []}
{"id": "2403.11448", "title": "Robust Overfitting Does Matter: Test-Time Adversarial Purification With\n  FGSM", "abstract": "  Numerous studies have demonstrated the susceptibility of deep neural networks\n(DNNs) to subtle adversarial perturbations, prompting the development of many\nadvanced adversarial defense methods aimed at mitigating adversarial attacks.\nCurrent defense strategies usually train DNNs for a specific adversarial attack\nmethod and can achieve good robustness in defense against this type of\nadversarial attack. Nevertheless, when subjected to evaluations involving\nunfamiliar attack modalities, empirical evidence reveals a pronounced\ndeterioration in the robustness of DNNs. Meanwhile, there is a trade-off\nbetween the classification accuracy of clean examples and adversarial examples.\nMost defense methods often sacrifice the accuracy of clean examples in order to\nimprove the adversarial robustness of DNNs. To alleviate these problems and\nenhance the overall robust generalization of DNNs, we propose the Test-Time\nPixel-Level Adversarial Purification (TPAP) method. This approach is based on\nthe robust overfitting characteristic of DNNs to the fast gradient sign method\n(FGSM) on training and test datasets. It utilizes FGSM for adversarial\npurification, to process images for purifying unknown adversarial perturbations\nfrom pixels at testing time in a \"counter changes with changelessness\" manner,\nthereby enhancing the defense capability of DNNs against various unknown\nadversarial attacks. Extensive experimental results show that our method can\neffectively improve both overall robust generalization of DNNs, notably over\nprevious methods.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.11833", "title": "SSCAE -- Semantic, Syntactic, and Context-aware natural language\n  Adversarial Examples generator", "abstract": "  Machine learning models are vulnerable to maliciously crafted Adversarial\nExamples (AEs). Training a machine learning model with AEs improves its\nrobustness and stability against adversarial attacks. It is essential to\ndevelop models that produce high-quality AEs. Developing such models has been\nmuch slower in natural language processing (NLP) than in areas such as computer\nvision. This paper introduces a practical and efficient adversarial attack\nmodel called SSCAE for \\textbf{S}emantic, \\textbf{S}yntactic, and\n\\textbf{C}ontext-aware natural language \\textbf{AE}s generator. SSCAE\nidentifies important words and uses a masked language model to generate an\nearly set of substitutions. Next, two well-known language models are employed\nto evaluate the initial set in terms of semantic and syntactic characteristics.\nWe introduce (1) a dynamic threshold to capture more efficient perturbations\nand (2) a local greedy search to generate high-quality AEs. As a black-box\nmethod, SSCAE generates humanly imperceptible and context-aware AEs that\npreserve semantic consistency and the source language's syntactical and\ngrammatical requirements. The effectiveness and superiority of the proposed\nSSCAE model are illustrated with fifteen comparative experiments and extensive\nsensitivity analysis for parameter optimization. SSCAE outperforms the existing\nmodels in all experiments while maintaining a higher semantic consistency with\na lower query number and a comparable perturbation rate.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.10073", "title": "Revisiting Adversarial Training under Long-Tailed Distributions", "abstract": "  Deep neural networks are vulnerable to adversarial attacks, often leading to\nerroneous outputs. Adversarial training has been recognized as one of the most\neffective methods to counter such attacks. However, existing adversarial\ntraining techniques have predominantly been tested on balanced datasets,\nwhereas real-world data often exhibit a long-tailed distribution, casting doubt\non the efficacy of these methods in practical scenarios.\n  In this paper, we delve into adversarial training under long-tailed\ndistributions. Through an analysis of the previous work \"RoBal\", we discover\nthat utilizing Balanced Softmax Loss alone can achieve performance comparable\nto the complete RoBal approach while significantly reducing training overheads.\nAdditionally, we reveal that, similar to uniform distributions, adversarial\ntraining under long-tailed distributions also suffers from robust overfitting.\nTo address this, we explore data augmentation as a solution and unexpectedly\ndiscover that, unlike results obtained with balanced data, data augmentation\nnot only effectively alleviates robust overfitting but also significantly\nimproves robustness. We further investigate the reasons behind the improvement\nof robustness through data augmentation and identify that it is attributable to\nthe increased diversity of examples. Extensive experiments further corroborate\nthat data augmentation alone can significantly improve robustness. Finally,\nbuilding on these findings, we demonstrate that compared to RoBal, the\ncombination of BSL and data augmentation leads to a +6.66% improvement in model\nrobustness under AutoAttack on CIFAR-10-LT. Our code is available at\nhttps://github.com/NISPLab/AT-BSL .\n", "date": "2024-03-15", "categories": []}
{"id": "2403.10076", "title": "Benchmarking Adversarial Robustness of Image Shadow Removal with\n  Shadow-adaptive Attacks", "abstract": "  Shadow removal is a task aimed at erasing regional shadows present in images\nand reinstating visually pleasing natural scenes with consistent illumination.\nWhile recent deep learning techniques have demonstrated impressive performance\nin image shadow removal, their robustness against adversarial attacks remains\nlargely unexplored. Furthermore, many existing attack frameworks typically\nallocate a uniform budget for perturbations across the entire input image,\nwhich may not be suitable for attacking shadow images. This is primarily due to\nthe unique characteristic of spatially varying illumination within shadow\nimages. In this paper, we propose a novel approach, called shadow-adaptive\nadversarial attack. Different from standard adversarial attacks, our attack\nbudget is adjusted based on the pixel intensity in different regions of shadow\nimages. Consequently, the optimized adversarial noise in the shadowed regions\nbecomes visually less perceptible while permitting a greater tolerance for\nperturbations in non-shadow regions. The proposed shadow-adaptive attacks\nnaturally align with the varying illumination distribution in shadow images,\nresulting in perturbations that are less conspicuous. Building on this, we\nconduct a comprehensive empirical evaluation of existing shadow removal\nmethods, subjecting them to various levels of attack on publicly available\ndatasets.\n", "date": "2024-03-15", "categories": []}
{"id": "2403.09101", "title": "Soften to Defend: Towards Adversarial Robustness via Self-Guided Label\n  Refinement", "abstract": "  Adversarial training (AT) is currently one of the most effective ways to\nobtain the robustness of deep neural networks against adversarial attacks.\nHowever, most AT methods suffer from robust overfitting, i.e., a significant\ngeneralization gap in adversarial robustness between the training and testing\ncurves. In this paper, we first identify a connection between robust\noverfitting and the excessive memorization of noisy labels in AT from a view of\ngradient norm. As such label noise is mainly caused by a distribution mismatch\nand improper label assignments, we are motivated to propose a label refinement\napproach for AT. Specifically, our Self-Guided Label Refinement first\nself-refines a more accurate and informative label distribution from\nover-confident hard labels, and then it calibrates the training by dynamically\nincorporating knowledge from self-distilled models into the current model and\nthus requiring no external teachers. Empirical results demonstrate that our\nmethod can simultaneously boost the standard accuracy and robust performance\nacross multiple benchmark datasets, attack types, and architectures. In\naddition, we also provide a set of analyses from the perspectives of\ninformation theory to dive into our method and suggest the importance of soft\nlabels for robust generalization.\n", "date": "2024-03-14", "categories": []}
{"id": "2308.08906", "title": "Towards a Practical Defense against Adversarial Attacks on Deep\n  Learning-based Malware Detectors via Randomized Smoothing", "abstract": "  Malware detectors based on deep learning (DL) have been shown to be\nsusceptible to malware examples that have been deliberately manipulated in\norder to evade detection, a.k.a. adversarial malware examples. More\nspecifically, it has been show that deep learning detectors are vulnerable to\nsmall changes on the input file. Given this vulnerability of deep learning\ndetectors, we propose a practical defense against adversarial malware examples\ninspired by randomized smoothing. In our work, instead of employing Gaussian or\nLaplace noise when randomizing inputs, we propose a randomized ablation-based\nsmoothing scheme that ablates a percentage of the bytes within an executable.\nDuring training, our randomized ablation-based smoothing scheme trains a base\nclassifier based on ablated versions of the executable files. At test time, the\nfinal classification for a given input executable is taken as the class most\ncommonly predicted by the classifier on a set of ablated versions of the\noriginal executable. To demonstrate the suitability of our approach we have\nempirically evaluated the proposed ablation-based model against various\nstate-of-the-art evasion attacks on the BODMAS dataset. Results show greater\nrobustness and generalization capabilities to adversarial malware examples in\ncomparison to a non-smoothed classifier.\n", "date": "2023-08-17", "categories": []}
{"id": "2403.07942", "title": "Attacking Transformers with Feature Diversity Adversarial Perturbation", "abstract": "  Understanding the mechanisms behind Vision Transformer (ViT), particularly\nits vulnerability to adversarial perturba tions, is crucial for addressing\nchallenges in its real-world applications. Existing ViT adversarial attackers\nrely on la bels to calculate the gradient for perturbation, and exhibit low\ntransferability to other structures and tasks. In this paper, we present a\nlabel-free white-box attack approach for ViT-based models that exhibits strong\ntransferability to various black box models, including most ViT variants, CNNs,\nand MLPs, even for models developed for other modalities. Our inspira tion\ncomes from the feature collapse phenomenon in ViTs, where the critical\nattention mechanism overly depends on the low-frequency component of features,\ncausing the features in middle-to-end layers to become increasingly similar and\neventually collapse. We propose the feature diversity attacker to naturally\naccelerate this process and achieve remarkable performance and transferability.\n", "date": "2024-03-10", "categories": []}
{"id": "2403.07943", "title": "Revisiting Edge Perturbation for Graph Neural Network in Graph Data\n  Augmentation and Attack", "abstract": "  Edge perturbation is a basic method to modify graph structures. It can be\ncategorized into two veins based on their effects on the performance of graph\nneural networks (GNNs), i.e., graph data augmentation and attack. Surprisingly,\nboth veins of edge perturbation methods employ the same operations, yet yield\nopposite effects on GNNs' accuracy. A distinct boundary between these methods\nin using edge perturbation has never been clearly defined. Consequently,\ninappropriate perturbations may lead to undesirable outcomes, necessitating\nprecise adjustments to achieve desired effects. Therefore, questions of ``why\nedge perturbation has a two-faced effect?'' and ``what makes edge perturbation\nflexible and effective?'' still remain unanswered.\n  In this paper, we will answer these questions by proposing a unified\nformulation and establishing a clear boundary between two categories of edge\nperturbation methods. Specifically, we conduct experiments to elucidate the\ndifferences and similarities between these methods and theoretically unify the\nworkflow of these methods by casting it to one optimization problem. Then, we\ndevise Edge Priority Detector (EPD) to generate a novel priority metric,\nbridging these methods up in the workflow. Experiments show that EPD can make\naugmentation or attack flexibly and achieve comparable or superior performance\nto other counterparts with less time overhead.\n", "date": "2024-03-10", "categories": []}
{"id": "2403.08170", "title": "Versatile Defense Against Adversarial Attacks on Image Recognition", "abstract": "  Adversarial attacks present a significant security risk to image recognition\ntasks. Defending against these attacks in a real-life setting can be compared\nto the way antivirus software works, with a key consideration being how well\nthe defense can adapt to new and evolving attacks. Another important factor is\nthe resources involved in terms of time and cost for training defense models\nand updating the model database. Training many models that are specific to each\ntype of attack can be time-consuming and expensive. Ideally, we should be able\nto train one single model that can handle a wide range of attacks. It appears\nthat a defense method based on image-to-image translation may be capable of\nthis. The proposed versatile defense approach in this paper only requires\ntraining one model to effectively resist various unknown adversarial attacks.\nThe trained model has successfully improved the classification accuracy from\nnearly zero to an average of 86%, performing better than other defense methods\nproposed in prior studies. When facing the PGD attack and the MI-FGSM attack,\nversatile defense model even outperforms the attack-specific models trained\nbased on these two attacks. The robustness check also shows that our versatile\ndefense model performs stably regardless with the attack strength.\n", "date": "2024-03-13", "categories": []}
{"id": "2403.08208", "title": "Advancing Security in AI Systems: A Novel Approach to Detecting\n  Backdoors in Deep Neural Networks", "abstract": "  In the rapidly evolving landscape of communication and network security, the\nincreasing reliance on deep neural networks (DNNs) and cloud services for data\nprocessing presents a significant vulnerability: the potential for backdoors\nthat can be exploited by malicious actors. Our approach leverages advanced\ntensor decomposition algorithms Independent Vector Analysis (IVA), Multiset\nCanonical Correlation Analysis (MCCA), and Parallel Factor Analysis (PARAFAC2)\nto meticulously analyze the weights of pre-trained DNNs and distinguish between\nbackdoored and clean models effectively. The key strengths of our method lie in\nits domain independence, adaptability to various network architectures, and\nability to operate without access to the training data of the scrutinized\nmodels. This not only ensures versatility across different application\nscenarios but also addresses the challenge of identifying backdoors without\nprior knowledge of the specific triggers employed to alter network behavior. We\nhave applied our detection pipeline to three distinct computer vision datasets,\nencompassing both image classification and object detection tasks. The results\ndemonstrate a marked improvement in both accuracy and efficiency over existing\nbackdoor detection methods. This advancement enhances the security of deep\nlearning and AI in networked systems, providing essential cybersecurity against\nevolving threats in emerging technologies.\n", "date": "2024-03-13", "categories": []}
{"id": "2403.07463", "title": "Backdoor Attack with Mode Mixture Latent Modification", "abstract": "  Backdoor attacks become a significant security concern for deep neural\nnetworks in recent years. An image classification model can be compromised if\nmalicious backdoors are injected into it. This corruption will cause the model\nto function normally on clean images but predict a specific target label when\ntriggers are present. Previous research can be categorized into two genres:\npoisoning a portion of the dataset with triggered images for users to train the\nmodel from scratch, or training a backdoored model alongside a triggered image\ngenerator. Both approaches require significant amount of attackable parameters\nfor optimization to establish a connection between the trigger and the target\nlabel, which may raise suspicions as more people become aware of the existence\nof backdoor attacks. In this paper, we propose a backdoor attack paradigm that\nonly requires minimal alterations (specifically, the output layer) to a clean\nmodel in order to inject the backdoor under the guise of fine-tuning. To\nachieve this, we leverage mode mixture samples, which are located between\ndifferent modes in latent space, and introduce a novel method for conducting\nbackdoor attacks. We evaluate the effectiveness of our method on four popular\nbenchmark datasets: MNIST, CIFAR-10, GTSRB, and TinyImageNet.\n", "date": "2024-03-12", "categories": []}
{"id": "2301.13096", "title": "Language-Driven Anchors for Zero-Shot Adversarial Robustness", "abstract": "  Deep Neural Networks (DNNs) are known to be susceptible to adversarial\nattacks. Previous researches mainly focus on improving adversarial robustness\nin the fully supervised setting, leaving the challenging domain of zero-shot\nadversarial robustness an open question. In this work, we investigate this\ndomain by leveraging the recent advances in large vision-language models, such\nas CLIP, to introduce zero-shot adversarial robustness to DNNs. We propose\nLAAT, a Language-driven, Anchor-based Adversarial Training strategy. LAAT\nutilizes the features of a text encoder for each category as fixed anchors\n(normalized feature embeddings) for each category, which are then employed for\nadversarial training. By leveraging the semantic consistency of the text\nencoders, LAAT aims to enhance the adversarial robustness of the image model on\nnovel categories. However, naively using text encoders leads to poor results.\nThrough analysis, we identified the issue to be the high cosine similarity\nbetween text encoders. We then design an expansion algorithm and an alignment\ncross-entropy loss to alleviate the problem. Our experimental results\ndemonstrated that LAAT significantly improves zero-shot adversarial robustness\nover state-of-the-art methods. LAAT has the potential to enhance adversarial\nrobustness by large-scale multimodal models, especially when labeled data is\nunavailable during training.\n", "date": "2023-01-30", "categories": []}
{"id": "2306.16614", "title": "Group-based Robustness: A General Framework for Customized Robustness in\n  the Real World", "abstract": "  Machine-learning models are known to be vulnerable to evasion attacks that\nperturb model inputs to induce misclassifications. In this work, we identify\nreal-world scenarios where the true threat cannot be assessed accurately by\nexisting attacks. Specifically, we find that conventional metrics measuring\ntargeted and untargeted robustness do not appropriately reflect a model's\nability to withstand attacks from one set of source classes to another set of\ntarget classes. To address the shortcomings of existing methods, we formally\ndefine a new metric, termed group-based robustness, that complements existing\nmetrics and is better-suited for evaluating model performance in certain attack\nscenarios. We show empirically that group-based robustness allows us to\ndistinguish between models' vulnerability against specific threat models in\nsituations where traditional robustness metrics do not apply. Moreover, to\nmeasure group-based robustness efficiently and accurately, we 1) propose two\nloss functions and 2) identify three new attack strategies. We show empirically\nthat with comparable success rates, finding evasive samples using our new loss\nfunctions saves computation by a factor as large as the number of targeted\nclasses, and finding evasive samples using our new attack strategies saves time\nby up to 99\\% compared to brute-force search methods. Finally, we propose a\ndefense method that increases group-based robustness by up to 3.52$\\times$.\n", "date": "2023-06-29", "categories": []}
{"id": "2310.12713", "title": "Learn from the Past: A Proxy Guided Adversarial Defense Framework with\n  Self Distillation Regularization", "abstract": "  Adversarial Training (AT), pivotal in fortifying the robustness of deep\nlearning models, is extensively adopted in practical applications. However,\nprevailing AT methods, relying on direct iterative updates for target model's\ndefense, frequently encounter obstacles such as unstable training and\ncatastrophic overfitting. In this context, our work illuminates the potential\nof leveraging the target model's historical states as a proxy to provide\neffective initialization and defense prior, which results in a general proxy\nguided defense framework, `LAST' ({\\bf L}earn from the P{\\bf ast}).\nSpecifically, LAST derives response of the proxy model as dynamically learned\nfast weights, which continuously corrects the update direction of the target\nmodel. Besides, we introduce a self-distillation regularized defense objective,\ningeniously designed to steer the proxy model's update trajectory without\nresorting to external teacher models, thereby ameliorating the impact of\ncatastrophic overfitting on performance. Extensive experiments and ablation\nstudies showcase the framework's efficacy in markedly improving model\nrobustness (e.g., up to 9.2\\% and 20.3\\% enhancement in robust accuracy on\nCIFAR10 and CIFAR100 datasets, respectively) and training stability. These\nimprovements are consistently observed across various model architectures,\nlarger datasets, perturbation sizes, and attack modalities, affirming LAST's\nability to consistently refine both single-step and multi-step AT strategies.\nThe code will be available at~\\url{https://github.com/callous-youth/LAST}.\n", "date": "2023-10-19", "categories": []}
{"id": "2401.09574", "title": "Towards Scalable and Robust Model Versioning", "abstract": "  As the deployment of deep learning models continues to expand across\nindustries, the threat of malicious incursions aimed at gaining access to these\ndeployed models is on the rise. Should an attacker gain access to a deployed\nmodel, whether through server breaches, insider attacks, or model inversion\ntechniques, they can then construct white-box adversarial attacks to manipulate\nthe model's classification outcomes, thereby posing significant risks to\norganizations that rely on these models for critical tasks. Model owners need\nmechanisms to protect themselves against such losses without the necessity of\nacquiring fresh training data - a process that typically demands substantial\ninvestments in time and capital.\n  In this paper, we explore the feasibility of generating multiple versions of\na model that possess different attack properties, without acquiring new\ntraining data or changing model architecture. The model owner can deploy one\nversion at a time and replace a leaked version immediately with a new version.\nThe newly deployed model version can resist adversarial attacks generated\nleveraging white-box access to one or all previously leaked versions. We show\ntheoretically that this can be accomplished by incorporating parameterized\nhidden distributions into the model training data, forcing the model to learn\ntask-irrelevant features uniquely defined by the chosen data. Additionally,\noptimal choices of hidden distributions can produce a sequence of model\nversions capable of resisting compound transferability attacks over time.\nLeveraging our analytical insights, we design and implement a practical model\nversioning method for DNN classifiers, which leads to significant robustness\nimprovements over existing methods. We believe our work presents a promising\ndirection for safeguarding DNN services beyond their initial deployment.\n", "date": "2024-01-17", "categories": []}
{"id": "2403.04954", "title": "Fooling Neural Networks for Motion Forecasting via Adversarial Attacks", "abstract": "  Human motion prediction is still an open problem, which is extremely\nimportant for autonomous driving and safety applications. Although there are\ngreat advances in this area, the widely studied topic of adversarial attacks\nhas not been applied to multi-regression models such as GCNs and MLP-based\narchitectures in human motion prediction. This work intends to reduce this gap\nusing extensive quantitative and qualitative experiments in state-of-the-art\narchitectures similar to the initial stages of adversarial attacks in image\nclassification. The results suggest that models are susceptible to attacks even\non low levels of perturbation. We also show experiments with 3D transformations\nthat affect the model performance, in particular, we show that most models are\nsensitive to simple rotations and translations which do not alter joint\ndistances. We conclude that similar to earlier CNN models, motion forecasting\ntasks are susceptible to small perturbations and simple 3D transformations.\n", "date": "2024-03-07", "categories": []}
{"id": "2403.06014", "title": "Hard-label based Small Query Black-box Adversarial Attack", "abstract": "  We consider the hard label based black box adversarial attack setting which\nsolely observes predicted classes from the target model. Most of the attack\nmethods in this setting suffer from impractical number of queries required to\nachieve a successful attack. One approach to tackle this drawback is utilising\nthe adversarial transferability between white box surrogate models and black\nbox target model. However, the majority of the methods adopting this approach\nare soft label based to take the full advantage of zeroth order optimisation.\nUnlike mainstream methods, we propose a new practical setting of hard label\nbased attack with an optimisation process guided by a pretrained surrogate\nmodel. Experiments show the proposed method significantly improves the query\nefficiency of the hard label based black-box attack across various target model\narchitectures. We find the proposed method achieves approximately 5 times\nhigher attack success rate compared to the benchmarks, especially at the small\nquery budgets as 100 and 250.\n", "date": "2024-03-09", "categories": []}
{"id": "2403.06661", "title": "epsilon-Mesh Attack: A Surface-based Adversarial Point Cloud Attack for\n  Facial Expression Recognition", "abstract": "  Point clouds and meshes are widely used 3D data structures for many computer\nvision applications. While the meshes represent the surfaces of an object,\npoint cloud represents sampled points from the surface which is also the output\nof modern sensors such as LiDAR and RGB-D cameras. Due to the wide application\narea of point clouds and the recent advancements in deep neural networks,\nstudies focusing on robust classification of the 3D point cloud data emerged.\nTo evaluate the robustness of deep classifier networks, a common method is to\nuse adversarial attacks where the gradient direction is followed to change the\ninput slightly. The previous studies on adversarial attacks are generally\nevaluated on point clouds of daily objects. However, considering 3D faces,\nthese adversarial attacks tend to affect the person's facial structure more\nthan the desired amount and cause malformation. Specifically for facial\nexpressions, even a small adversarial attack can have a significant effect on\nthe face structure. In this paper, we suggest an adversarial attack called\n$\\epsilon$-Mesh Attack, which operates on point cloud data via limiting\nperturbations to be on the mesh surface. We also parameterize our attack by\n$\\epsilon$ to scale the perturbation mesh. Our surface-based attack has tighter\nperturbation bounds compared to $L_2$ and $L_\\infty$ norm bounded attacks that\noperate on unit-ball. Even though our method has additional constraints, our\nexperiments on CoMA, Bosphorus and FaceWarehouse datasets show that\n$\\epsilon$-Mesh Attack (Perpendicular) successfully confuses trained DGCNN and\nPointNet models $99.72\\%$ and $97.06\\%$ of the time, with indistinguishable\nfacial deformations. The code is available at\nhttps://github.com/batuceng/e-mesh-attack.\n", "date": "2024-03-11", "categories": []}
{"id": "2403.06698", "title": "PCLD: Point Cloud Layerwise Diffusion for Adversarial Purification", "abstract": "  Point clouds are extensively employed in a variety of real-world applications\nsuch as robotics, autonomous driving and augmented reality. Despite the recent\nsuccess of point cloud neural networks, especially for safety-critical tasks,\nit is essential to also ensure the robustness of the model. A typical way to\nassess a model's robustness is through adversarial attacks, where test-time\nexamples are generated based on gradients to deceive the model. While many\ndifferent defense mechanisms are studied in 2D, studies on 3D point clouds have\nbeen relatively limited in the academic field. Inspired from PointDP, which\ndenoises the network inputs by diffusion, we propose Point Cloud Layerwise\nDiffusion (PCLD), a layerwise diffusion based 3D point cloud defense strategy.\nUnlike PointDP, we propagated the diffusion denoising after each layer to\nincrementally enhance the results. We apply our defense method to different\ntypes of commonly used point cloud models and adversarial attacks to evaluate\nits robustness. Our experiments demonstrate that the proposed defense method\nachieved results that are comparable to or surpass those of existing\nmethodologies, establishing robustness through a novel technique. Code is\navailable at https://github.com/batuceng/diffusion-layer-robustness-pc.\n", "date": "2024-03-11", "categories": []}
{"id": "2403.06798", "title": "Dynamic Perturbation-Adaptive Adversarial Training on Medical Image\n  Classification", "abstract": "  Remarkable successes were made in Medical Image Classification (MIC)\nrecently, mainly due to wide applications of convolutional neural networks\n(CNNs). However, adversarial examples (AEs) exhibited imperceptible similarity\nwith raw data, raising serious concerns on network robustness. Although\nadversarial training (AT), in responding to malevolent AEs, was recognized as\nan effective approach to improve robustness, it was challenging to overcome\ngeneralization decline of networks caused by the AT. In this paper, in order to\nreserve high generalization while improving robustness, we proposed a dynamic\nperturbation-adaptive adversarial training (DPAAT) method, which placed AT in a\ndynamic learning environment to generate adaptive data-level perturbations and\nprovided a dynamically updated criterion by loss information collections to\nhandle the disadvantage of fixed perturbation sizes in conventional AT methods\nand the dependence on external transference. Comprehensive testing on\ndermatology HAM10000 dataset showed that the DPAAT not only achieved better\nrobustness improvement and generalization preservation but also significantly\nenhanced mean average precision and interpretability on various CNNs,\nindicating its great potential as a generic adversarial training method on the\nMIC.\n", "date": "2024-03-11", "categories": []}
{"id": "2403.05247", "title": "Hide in Thicket: Generating Imperceptible and Rational Adversarial\n  Perturbations on 3D Point Clouds", "abstract": "  Adversarial attack methods based on point manipulation for 3D point cloud\nclassification have revealed the fragility of 3D models, yet the adversarial\nexamples they produce are easily perceived or defended against. The trade-off\nbetween the imperceptibility and adversarial strength leads most point attack\nmethods to inevitably introduce easily detectable outlier points upon a\nsuccessful attack. Another promising strategy, shape-based attack, can\neffectively eliminate outliers, but existing methods often suffer significant\nreductions in imperceptibility due to irrational deformations. We find that\nconcealing deformation perturbations in areas insensitive to human eyes can\nachieve a better trade-off between imperceptibility and adversarial strength,\nspecifically in parts of the object surface that are complex and exhibit\ndrastic curvature changes. Therefore, we propose a novel shape-based\nadversarial attack method, HiT-ADV, which initially conducts a two-stage search\nfor attack regions based on saliency and imperceptibility scores, and then adds\ndeformation perturbations in each attack region using Gaussian kernel\nfunctions. Additionally, HiT-ADV is extendable to physical attack. We propose\nthat by employing benign resampling and benign rigid transformations, we can\nfurther enhance physical adversarial strength with little sacrifice to\nimperceptibility. Extensive experiments have validated the superiority of our\nmethod in terms of adversarial and imperceptible properties in both digital and\nphysical spaces. Our code is avaliable at: https://github.com/TRLou/HiT-ADV.\n", "date": "2024-03-08", "categories": []}
{"id": "2403.05365", "title": "The Impact of Quantization on the Robustness of Transformer-based Text\n  Classifiers", "abstract": "  Transformer-based models have made remarkable advancements in various NLP\nareas. Nevertheless, these models often exhibit vulnerabilities when confronted\nwith adversarial attacks. In this paper, we explore the effect of quantization\non the robustness of Transformer-based models. Quantization usually involves\nmapping a high-precision real number to a lower-precision value, aiming at\nreducing the size of the model at hand. To the best of our knowledge, this work\nis the first application of quantization on the robustness of NLP models. In\nour experiments, we evaluate the impact of quantization on BERT and DistilBERT\nmodels in text classification using SST-2, Emotion, and MR datasets. We also\nevaluate the performance of these models against TextFooler, PWWS, and PSO\nadversarial attacks. Our findings show that quantization significantly improves\n(by an average of 18.68%) the adversarial accuracy of the models. Furthermore,\nwe compare the effect of quantization versus that of the adversarial training\napproach on robustness. Our experiments indicate that quantization increases\nthe robustness of the model by 18.80% on average compared to adversarial\ntraining without imposing any extra computational overhead during training.\nTherefore, our results highlight the effectiveness of quantization in improving\nthe robustness of NLP models.\n", "date": "2024-03-08", "categories": []}
{"id": "2403.04070", "title": "Improving Adversarial Training using Vulnerability-Aware Perturbation\n  Budget", "abstract": "  Adversarial Training (AT) effectively improves the robustness of Deep Neural\nNetworks (DNNs) to adversarial attacks. Generally, AT involves training DNN\nmodels with adversarial examples obtained within a pre-defined, fixed\nperturbation bound. Notably, individual natural examples from which these\nadversarial examples are crafted exhibit varying degrees of intrinsic\nvulnerabilities, and as such, crafting adversarial examples with fixed\nperturbation radius for all instances may not sufficiently unleash the potency\nof AT. Motivated by this observation, we propose two simple, computationally\ncheap vulnerability-aware reweighting functions for assigning perturbation\nbounds to adversarial examples used for AT, named Margin-Weighted Perturbation\nBudget (MWPB) and Standard-Deviation-Weighted Perturbation Budget (SDWPB). The\nproposed methods assign perturbation radii to individual adversarial samples\nbased on the vulnerability of their corresponding natural examples.\nExperimental results show that the proposed methods yield genuine improvements\nin the robustness of AT algorithms against various adversarial attacks.\n", "date": "2024-03-06", "categories": []}
{"id": "2307.00907", "title": "Enhancing the Robustness of QMIX against State-adversarial Attacks", "abstract": "  Deep reinforcement learning (DRL) performance is generally impacted by\nstate-adversarial attacks, a perturbation applied to an agent's observation.\nMost recent research has concentrated on robust single-agent reinforcement\nlearning (SARL) algorithms against state-adversarial attacks. Still, there has\nyet to be much work on robust multi-agent reinforcement learning. Using QMIX,\none of the popular cooperative multi-agent reinforcement algorithms, as an\nexample, we discuss four techniques to improve the robustness of SARL\nalgorithms and extend them to multi-agent scenarios. To increase the robustness\nof multi-agent reinforcement learning (MARL) algorithms, we train models using\na variety of attacks in this research. We then test the models taught using the\nother attacks by subjecting them to the corresponding attacks throughout the\ntraining phase. In this way, we organize and summarize techniques for enhancing\nrobustness when used with MARL.\n", "date": "2023-07-03", "categories": []}
{"id": "2402.15653", "title": "Low-Frequency Black-Box Backdoor Attack via Evolutionary Algorithm", "abstract": "  While convolutional neural networks (CNNs) have achieved success in computer\nvision tasks, it is vulnerable to backdoor attacks. Such attacks could mislead\nthe victim model to make attacker-chosen prediction with a specific trigger\npattern. Until now, the trigger injection of existing attacks is mainly limited\nto spatial domain. Recent works take advantage of perceptual properties of\nplanting specific patterns in the frequency domain, which only reflect\nindistinguishable pixel-wise perturbations in pixel domain. However, in the\nblack-box setup, the inaccessibility of training process often renders more\ncomplex trigger designs. Existing frequency attacks simply handcraft the\nmagnitude of spectrum, introducing anomaly frequency disparities between clean\nand poisoned data and taking risks of being removed by image processing\noperations (such as lossy compression and filtering). In this paper, we propose\na robust low-frequency black-box backdoor attack (LFBA), which minimally\nperturbs low-frequency components of frequency spectrum and maintains the\nperceptual similarity in spatial space simultaneously. The key insight of our\nattack restrict the search for the optimal trigger to low-frequency region that\ncan achieve high attack effectiveness, robustness against image transformation\ndefenses and stealthiness in dual space. We utilize simulated annealing (SA), a\nform of evolutionary algorithm, to optimize the properties of frequency trigger\nincluding the number of manipulated frequency bands and the perturbation of\neach frequency component, without relying on the knowledge from the victim\nclassifier. Extensive experiments on real-world datasets verify the\neffectiveness and robustness of LFBA against image processing operations and\nthe state-of-the-art backdoor defenses, as well as its inherent stealthiness in\nboth spatial and frequency space, making it resilient against frequency\ninspection.\n", "date": "2024-02-23", "categories": []}
{"id": "2402.19322", "title": "Verification of Neural Networks' Global Robustness", "abstract": "  Neural networks are successful in various applications but are also\nsusceptible to adversarial attacks. To show the safety of network classifiers,\nmany verifiers have been introduced to reason about the local robustness of a\ngiven input to a given perturbation. While successful, local robustness cannot\ngeneralize to unseen inputs. Several works analyze global robustness\nproperties, however, neither can provide a precise guarantee about the cases\nwhere a network classifier does not change its classification. In this work, we\npropose a new global robustness property for classifiers aiming at finding the\nminimal globally robust bound, which naturally extends the popular local\nrobustness property for classifiers. We introduce VHAGaR, an anytime verifier\nfor computing this bound. VHAGaR relies on three main ideas: encoding the\nproblem as a mixed-integer programming and pruning the search space by\nidentifying dependencies stemming from the perturbation or the network's\ncomputation and generalizing adversarial attacks to unknown inputs. We evaluate\nVHAGaR on several datasets and classifiers and show that, given a three hour\ntimeout, the average gap between the lower and upper bound on the minimal\nglobally robust bound computed by VHAGaR is 1.9, while the gap of an existing\nglobal robustness verifier is 154.7. Moreover, VHAGaR is 130.6x faster than\nthis verifier. Our results further indicate that leveraging dependencies and\nadversarial attacks makes VHAGaR 78.6x faster.\n", "date": "2024-02-29", "categories": []}
{"id": "2403.03674", "title": "Adversarial Infrared Geometry: Using Geometry to Perform Adversarial\n  Attack against Infrared Pedestrian Detectors", "abstract": "  Currently, infrared imaging technology enjoys widespread usage, with infrared\nobject detection technology experiencing a surge in prominence. While previous\nstudies have delved into physical attacks on infrared object detectors, the\nimplementation of these techniques remains complex. For instance, some\napproaches entail the use of bulb boards or infrared QR suits as perturbations\nto execute attacks, which entail costly optimization and cumbersome deployment\nprocesses. Other methodologies involve the utilization of irregular aerogel as\nphysical perturbations for infrared attacks, albeit at the expense of\noptimization expenses and perceptibility issues. In this study, we propose a\nnovel infrared physical attack termed Adversarial Infrared Geometry\n(\\textbf{AdvIG}), which facilitates efficient black-box query attacks by\nmodeling diverse geometric shapes (lines, triangles, ellipses) and optimizing\ntheir physical parameters using Particle Swarm Optimization (PSO). Extensive\nexperiments are conducted to evaluate the effectiveness, stealthiness, and\nrobustness of AdvIG. In digital attack experiments, line, triangle, and ellipse\npatterns achieve attack success rates of 93.1\\%, 86.8\\%, and 100.0\\%,\nrespectively, with average query times of 71.7, 113.1, and 2.57, respectively,\nthereby confirming the efficiency of AdvIG. Physical attack experiments are\nconducted to assess the attack success rate of AdvIG at different distances. On\naverage, the line, triangle, and ellipse achieve attack success rates of\n61.1\\%, 61.2\\%, and 96.2\\%, respectively. Further experiments are conducted to\ncomprehensively analyze AdvIG, including ablation experiments, transfer attack\nexperiments, and adversarial defense mechanisms. Given the superior performance\nof our method as a simple and efficient black-box adversarial attack in both\ndigital and physical environments, we advocate for widespread attention to\nAdvIG.\n", "date": "2024-03-06", "categories": []}
{"id": "2302.13056", "title": "SATBA: An Invisible Backdoor Attack Based On Spatial Attention", "abstract": "  Backdoor attack has emerged as a novel and concerning threat to AI security.\nThese attacks involve the training of Deep Neural Network (DNN) on datasets\nthat contain hidden trigger patterns. Although the poisoned model behaves\nnormally on benign samples, it exhibits abnormal behavior on samples containing\nthe trigger pattern. However, most existing backdoor attacks suffer from two\nsignificant drawbacks: their trigger patterns are visible and easy to detect by\nbackdoor defense or even human inspection, and their injection process results\nin the loss of natural sample features and trigger patterns, thereby reducing\nthe attack success rate and model accuracy. In this paper, we propose a novel\nbackdoor attack named SATBA that overcomes these limitations using spatial\nattention and an U-net based model. The attack process begins by using spatial\nattention to extract meaningful data features and generate trigger patterns\nassociated with clean images. Then, an U-shaped model is used to embed these\ntrigger patterns into the original data without causing noticeable feature\nloss. We evaluate our attack on three prominent image classification DNN across\nthree standard datasets. The results demonstrate that SATBA achieves high\nattack success rate while maintaining robustness against backdoor defenses.\nFurthermore, we conduct extensive image similarity experiments to emphasize the\nstealthiness of our attack strategy. Overall, SATBA presents a promising\napproach to backdoor attack, addressing the shortcomings of previous methods\nand showcasing its effectiveness in evading detection and maintaining high\nattack success rate.\n", "date": "2023-02-25", "categories": []}
{"id": "2309.16487", "title": "Towards Poisoning Fair Representations", "abstract": "  Fair machine learning seeks to mitigate model prediction bias against certain\ndemographic subgroups such as elder and female. Recently, fair representation\nlearning (FRL) trained by deep neural networks has demonstrated superior\nperformance, whereby representations containing no demographic information are\ninferred from the data and then used as the input to classification or other\ndownstream tasks. Despite the development of FRL methods, their vulnerability\nunder data poisoning attack, a popular protocol to benchmark model robustness\nunder adversarial scenarios, is under-explored. Data poisoning attacks have\nbeen developed for classical fair machine learning methods which incorporate\nfairness constraints into shallow-model classifiers. Nonetheless, these attacks\nfall short in FRL due to notably different fairness goals and model\narchitectures. This work proposes the first data poisoning framework attacking\nFRL. We induce the model to output unfair representations that contain as much\ndemographic information as possible by injecting carefully crafted poisoning\nsamples into the training data. This attack entails a prohibitive bilevel\noptimization, wherefore an effective approximated solution is proposed. A\ntheoretical analysis on the needed number of poisoning samples is derived and\nsheds light on defending against the attack. Experiments on benchmark fairness\ndatasets and state-of-the-art fair representation learning models demonstrate\nthe superiority of our attack.\n", "date": "2023-09-28", "categories": []}
{"id": "2402.12426", "title": "Attacks on Node Attributes in Graph Neural Networks", "abstract": "  Graphs are commonly used to model complex networks prevalent in modern social\nmedia and literacy applications. Our research investigates the vulnerability of\nthese graphs through the application of feature based adversarial attacks,\nfocusing on both decision time attacks and poisoning attacks. In contrast to\nstate of the art models like Net Attack and Meta Attack, which target node\nattributes and graph structure, our study specifically targets node attributes.\nFor our analysis, we utilized the text dataset Hellaswag and graph datasets\nCora and CiteSeer, providing a diverse basis for evaluation. Our findings\nindicate that decision time attacks using Projected Gradient Descent (PGD) are\nmore potent compared to poisoning attacks that employ Mean Node Embeddings and\nGraph Contrastive Learning strategies. This provides insights for graph data\nsecurity, pinpointing where graph-based models are most vulnerable and thereby\ninforming the development of stronger defense mechanisms against such attacks.\n", "date": "2024-02-19", "categories": []}
{"id": "2403.02723", "title": "Minimum Topology Attacks for Graph Neural Networks", "abstract": "  With the great popularity of Graph Neural Networks (GNNs), their robustness\nto adversarial topology attacks has received significant attention. Although\nmany attack methods have been proposed, they mainly focus on fixed-budget\nattacks, aiming at finding the most adversarial perturbations within a fixed\nbudget for target node. However, considering the varied robustness of each\nnode, there is an inevitable dilemma caused by the fixed budget, i.e., no\nsuccessful perturbation is found when the budget is relatively small, while if\nit is too large, the yielding redundant perturbations will hurt the\ninvisibility. To break this dilemma, we propose a new type of topology attack,\nnamed minimum-budget topology attack, aiming to adaptively find the minimum\nperturbation sufficient for a successful attack on each node. To this end, we\npropose an attack model, named MiBTack, based on a dynamic projected gradient\ndescent algorithm, which can effectively solve the involving non-convex\nconstraint optimization on discrete topology. Extensive results on three GNNs\nand four real-world datasets show that MiBTack can successfully lead all target\nnodes misclassified with the minimum perturbation edges. Moreover, the obtained\nminimum budget can be used to measure node robustness, so we can explore the\nrelationships of robustness, topology, and uncertainty for nodes, which is\nbeyond what the current fixed-budget topology attacks can offer.\n", "date": "2024-03-05", "categories": []}
{"id": "2403.02950", "title": "A general approach to enhance the survivability of backdoor attacks by\n  decision path coupling", "abstract": "  Backdoor attacks have been one of the emerging security threats to deep\nneural networks (DNNs), leading to serious consequences. One of the mainstream\nbackdoor defenses is model reconstruction-based. Such defenses adopt model\nunlearning or pruning to eliminate backdoors. However, little attention has\nbeen paid to survive from such defenses. To bridge the gap, we propose Venom,\nthe first generic backdoor attack enhancer to improve the survivability of\nexisting backdoor attacks against model reconstruction-based defenses. We\nformalize Venom as a binary-task optimization problem. The first is the\noriginal backdoor attack task to preserve the original attack capability, while\nthe second is the attack enhancement task to improve the attack survivability.\nTo realize the second task, we propose attention imitation loss to force the\ndecision path of poisoned samples in backdoored models to couple with the\ncrucial decision path of benign samples, which makes backdoors difficult to\neliminate. Our extensive evaluation on two DNNs and three datasets has\ndemonstrated that Venom significantly improves the survivability of eight\nstate-of-the-art attacks against eight state-of-the-art defenses without\nimpacting the capability of the original attacks.\n", "date": "2024-03-05", "categories": []}
{"id": "2403.02995", "title": "Mitigating Label Flipping Attacks in Malicious URL Detectors Using\n  Ensemble Trees", "abstract": "  Malicious URLs provide adversarial opportunities across various industries,\nincluding transportation, healthcare, energy, and banking which could be\ndetrimental to business operations. Consequently, the detection of these URLs\nis of crucial importance; however, current Machine Learning (ML) models are\nsusceptible to backdoor attacks. These attacks involve manipulating a small\npercentage of training data labels, such as Label Flipping (LF), which changes\nbenign labels to malicious ones and vice versa. This manipulation results in\nmisclassification and leads to incorrect model behavior. Therefore, integrating\ndefense mechanisms into the architecture of ML models becomes an imperative\nconsideration to fortify against potential attacks.\n  The focus of this study is on backdoor attacks in the context of URL\ndetection using ensemble trees. By illuminating the motivations behind such\nattacks, highlighting the roles of attackers, and emphasizing the critical\nimportance of effective defense strategies, this paper contributes to the\nongoing efforts to fortify ML models against adversarial threats within the ML\ndomain in network security. We propose an innovative alarm system that detects\nthe presence of poisoned labels and a defense mechanism designed to uncover the\noriginal class labels with the aim of mitigating backdoor attacks on ensemble\ntree classifiers. We conducted a case study using the Alexa and Phishing Site\nURL datasets and showed that LF attacks can be addressed using our proposed\ndefense mechanism. Our experimental results prove that the LF attack achieved\nan Attack Success Rate (ASR) between 50-65% within 2-5%, and the innovative\ndefense method successfully detected poisoned labels with an accuracy of up to\n100%.\n", "date": "2024-03-05", "categories": []}
{"id": "2303.09105", "title": "Rethinking Model Ensemble in Transfer-based Adversarial Attacks", "abstract": "  It is widely recognized that deep learning models lack robustness to\nadversarial examples. An intriguing property of adversarial examples is that\nthey can transfer across different models, which enables black-box attacks\nwithout any knowledge of the victim model. An effective strategy to improve the\ntransferability is attacking an ensemble of models. However, previous works\nsimply average the outputs of different models, lacking an in-depth analysis on\nhow and why model ensemble methods can strongly improve the transferability. In\nthis paper, we rethink the ensemble in adversarial attacks and define the\ncommon weakness of model ensemble with two properties: 1) the flatness of loss\nlandscape; and 2) the closeness to the local optimum of each model. We\nempirically and theoretically show that both properties are strongly correlated\nwith the transferability and propose a Common Weakness Attack (CWA) to generate\nmore transferable adversarial examples by promoting these two properties.\nExperimental results on both image classification and object detection tasks\nvalidate the effectiveness of our approach to improving the adversarial\ntransferability, especially when attacking adversarially trained models. We\nalso successfully apply our method to attack a black-box large vision-language\nmodel -- Google's Bard, showing the practical effectiveness. Code is available\nat \\url{https://github.com/huanranchen/AdversarialAttacks}.\n", "date": "2023-03-16", "categories": []}
{"id": "2304.14614", "title": "Fusion is Not Enough: Single Modal Attacks on Fusion Models for 3D\n  Object Detection", "abstract": "  Multi-sensor fusion (MSF) is widely used in autonomous vehicles (AVs) for\nperception, particularly for 3D object detection with camera and LiDAR sensors.\nThe purpose of fusion is to capitalize on the advantages of each modality while\nminimizing its weaknesses. Advanced deep neural network (DNN)-based fusion\ntechniques have demonstrated the exceptional and industry-leading performance.\nDue to the redundant information in multiple modalities, MSF is also recognized\nas a general defence strategy against adversarial attacks. In this paper, we\nattack fusion models from the camera modality that is considered to be of\nlesser importance in fusion but is more affordable for attackers. We argue that\nthe weakest link of fusion models depends on their most vulnerable modality,\nand propose an attack framework that targets advanced camera-LiDAR fusion-based\n3D object detection models through camera-only adversarial attacks. Our\napproach employs a two-stage optimization-based strategy that first thoroughly\nevaluates vulnerable image areas under adversarial attacks, and then applies\ndedicated attack strategies for different fusion models to generate deployable\npatches. The evaluations with six advanced camera-LiDAR fusion models and one\ncamera-only model indicate that our attacks successfully compromise all of\nthem. Our approach can either decrease the mean average precision (mAP) of\ndetection performance from 0.824 to 0.353, or degrade the detection score of a\ntarget object from 0.728 to 0.156, demonstrating the efficacy of our proposed\nattack framework. Code is available.\n", "date": "2023-04-28", "categories": []}
{"id": "2310.13822", "title": "Adversarial Attacks on Fairness of Graph Neural Networks", "abstract": "  Fairness-aware graph neural networks (GNNs) have gained a surge of attention\nas they can reduce the bias of predictions on any demographic group (e.g.,\nfemale) in graph-based applications. Although these methods greatly improve the\nalgorithmic fairness of GNNs, the fairness can be easily corrupted by carefully\ndesigned adversarial attacks. In this paper, we investigate the problem of\nadversarial attacks on fairness of GNNs and propose G-FairAttack, a general\nframework for attacking various types of fairness-aware GNNs in terms of\nfairness with an unnoticeable effect on prediction utility. In addition, we\npropose a fast computation technique to reduce the time complexity of\nG-FairAttack. The experimental study demonstrates that G-FairAttack\nsuccessfully corrupts the fairness of different types of GNNs while keeping the\nattack unnoticeable. Our study on fairness attacks sheds light on potential\nvulnerabilities in fairness-aware GNNs and guides further research on the\nrobustness of GNNs in terms of fairness.\n", "date": "2023-10-20", "categories": []}
{"id": "2311.00859", "title": "Optimal Cost Constrained Adversarial Attacks For Multiple Agent Systems", "abstract": "  Finding optimal adversarial attack strategies is an important topic in\nreinforcement learning and the Markov decision process. Previous studies\nusually assume one all-knowing coordinator (attacker) for whom attacking\ndifferent recipient (victim) agents incurs uniform costs. However, in reality,\ninstead of using one limitless central attacker, the attacks often need to be\nperformed by distributed attack agents. We formulate the problem of performing\noptimal adversarial agent-to-agent attacks using distributed attack agents, in\nwhich we impose distinct cost constraints on each different attacker-victim\npair. We propose an optimal method integrating within-step static constrained\nattack-resource allocation optimization and between-step dynamic programming to\nachieve the optimal adversarial attack in a multi-agent system. Our numerical\nresults show that the proposed attacks can significantly reduce the rewards\nreceived by the attacked agents.\n", "date": "2023-11-01", "categories": []}
{"id": "2312.08675", "title": "AVA: Inconspicuous Attribute Variation-based Adversarial Attack\n  bypassing DeepFake Detection", "abstract": "  While DeepFake applications are becoming popular in recent years, their\nabuses pose a serious privacy threat. Unfortunately, most related detection\nalgorithms to mitigate the abuse issues are inherently vulnerable to\nadversarial attacks because they are built atop DNN-based classification\nmodels, and the literature has demonstrated that they could be bypassed by\nintroducing pixel-level perturbations. Though corresponding mitigation has been\nproposed, we have identified a new attribute-variation-based adversarial attack\n(AVA) that perturbs the latent space via a combination of Gaussian prior and\nsemantic discriminator to bypass such mitigation. It perturbs the semantics in\nthe attribute space of DeepFake images, which are inconspicuous to human beings\n(e.g., mouth open) but can result in substantial differences in DeepFake\ndetection. We evaluate our proposed AVA attack on nine state-of-the-art\nDeepFake detection algorithms and applications. The empirical results\ndemonstrate that AVA attack defeats the state-of-the-art black box attacks\nagainst DeepFake detectors and achieves more than a 95% success rate on two\ncommercial DeepFake detectors. Moreover, our human study indicates that\nAVA-generated DeepFake images are often imperceptible to humans, which presents\nhuge security and privacy concerns.\n", "date": "2023-12-14", "categories": []}
{"id": "2402.11837", "title": "Self-Guided Robust Graph Structure Refinement", "abstract": "  Recent studies have revealed that GNNs are vulnerable to adversarial attacks.\nTo defend against such attacks, robust graph structure refinement (GSR) methods\naim at minimizing the effect of adversarial edges based on node features, graph\nstructure, or external information. However, we have discovered that existing\nGSR methods are limited by narrowassumptions, such as assuming clean node\nfeatures, moderate structural attacks, and the availability of external clean\ngraphs, resulting in the restricted applicability in real-world scenarios. In\nthis paper, we propose a self-guided GSR framework (SG-GSR), which utilizes a\nclean sub-graph found within the given attacked graph itself. Furthermore, we\npropose a novel graph augmentation and a group-training strategy to handle the\ntwo technical challenges in the clean sub-graph extraction: 1) loss of\nstructural information, and 2) imbalanced node degree distribution. Extensive\nexperiments demonstrate the effectiveness of SG-GSR under various scenarios\nincluding non-targeted attacks, targeted attacks, feature attacks, e-commerce\nfraud, and noisy node labels. Our code is available at\nhttps://github.com/yeonjun-in/torch-SG-GSR.\n", "date": "2024-02-19", "categories": []}
{"id": "2403.01118", "title": "Adversarial Testing for Visual Grounding via Image-Aware Property\n  Reduction", "abstract": "  Due to the advantages of fusing information from various modalities,\nmultimodal learning is gaining increasing attention. Being a fundamental task\nof multimodal learning, Visual Grounding (VG), aims to locate objects in images\nthrough natural language expressions. Ensuring the quality of VG models\npresents significant challenges due to the complex nature of the task. In the\nblack box scenario, existing adversarial testing techniques often fail to fully\nexploit the potential of both modalities of information. They typically apply\nperturbations based solely on either the image or text information,\ndisregarding the crucial correlation between the two modalities, which would\nlead to failures in test oracles or an inability to effectively challenge VG\nmodels. To this end, we propose PEELING, a text perturbation approach via\nimage-aware property reduction for adversarial testing of the VG model. The\ncore idea is to reduce the property-related information in the original\nexpression meanwhile ensuring the reduced expression can still uniquely\ndescribe the original object in the image. To achieve this, PEELING first\nconducts the object and properties extraction and recombination to generate\ncandidate property reduction expressions. It then selects the satisfied\nexpressions that accurately describe the original object while ensuring no\nother objects in the image fulfill the expression, through querying the image\nwith a visual understanding technique. We evaluate PEELING on the\nstate-of-the-art VG model, i.e. OFA-VG, involving three commonly used datasets.\nResults show that the adversarial tests generated by PEELING achieves 21.4% in\nMultiModal Impact score (MMI), and outperforms state-of-the-art baselines for\nimages and texts by 8.2%--15.1%.\n", "date": "2024-03-02", "categories": []}
{"id": "2403.01210", "title": "SAR-AE-SFP: SAR Imagery Adversarial Example in Real Physics domain with\n  Target Scattering Feature Parameters", "abstract": "  Deep neural network-based Synthetic Aperture Radar (SAR) target recognition\nmodels are susceptible to adversarial examples. Current adversarial example\ngeneration methods for SAR imagery primarily operate in the 2D digital domain,\nknown as image adversarial examples. Recent work, while considering SAR imaging\nscatter mechanisms, fails to account for the actual imaging process, rendering\nattacks in the three-dimensional physical domain infeasible, termed pseudo\nphysics adversarial examples. To address these challenges, this paper proposes\nSAR-AE-SFP-Attack, a method to generate real physics adversarial examples by\naltering the scattering feature parameters of target objects. Specifically, we\niteratively optimize the coherent energy accumulation of the target echo by\nperturbing the reflection coefficient and scattering coefficient in the\nscattering feature parameters of the three-dimensional target object, and\nobtain the adversarial example after echo signal processing and imaging\nprocessing in the RaySAR simulator. Experimental results show that compared to\ndigital adversarial attack methods, SAR-AE-SFP Attack significantly improves\nattack efficiency on CNN-based models (over 30\\%) and Transformer-based models\n(over 13\\%), demonstrating significant transferability of attack effects across\ndifferent models and perspectives.\n", "date": "2024-03-02", "categories": []}
{"id": "2403.02329", "title": "COMMIT: Certifying Robustness of Multi-Sensor Fusion Systems against\n  Semantic Attacks", "abstract": "  Multi-sensor fusion systems (MSFs) play a vital role as the perception module\nin modern autonomous vehicles (AVs). Therefore, ensuring their robustness\nagainst common and realistic adversarial semantic transformations, such as\nrotation and shifting in the physical world, is crucial for the safety of AVs.\nWhile empirical evidence suggests that MSFs exhibit improved robustness\ncompared to single-modal models, they are still vulnerable to adversarial\nsemantic transformations. Despite the proposal of empirical defenses, several\nworks show that these defenses can be attacked again by new adaptive attacks.\nSo far, there is no certified defense proposed for MSFs. In this work, we\npropose the first robustness certification framework COMMIT certify robustness\nof multi-sensor fusion systems against semantic attacks. In particular, we\npropose a practical anisotropic noise mechanism that leverages randomized\nsmoothing with multi-modal data and performs a grid-based splitting method to\ncharacterize complex semantic transformations. We also propose efficient\nalgorithms to compute the certification in terms of object detection accuracy\nand IoU for large-scale MSF models. Empirically, we evaluate the efficacy of\nCOMMIT in different settings and provide a comprehensive benchmark of certified\nrobustness for different MSF models using the CARLA simulation platform. We\nshow that the certification for MSF models is at most 48.39% higher than that\nof single-modal models, which validates the advantages of MSF models. We\nbelieve our certification framework and benchmark will contribute an important\nstep towards certifiably robust AVs in practice.\n", "date": "2024-03-04", "categories": []}
{"id": "2211.08229", "title": "CorruptEncoder: Data Poisoning based Backdoor Attacks to Contrastive\n  Learning", "abstract": "  Contrastive learning (CL) pre-trains general-purpose encoders using an\nunlabeled pre-training dataset, which consists of images or image-text pairs.\nCL is vulnerable to data poisoning based backdoor attacks (DPBAs), in which an\nattacker injects poisoned inputs into the pre-training dataset so the encoder\nis backdoored. However, existing DPBAs achieve limited effectiveness. In this\nwork, we take the first step to analyze the limitations of existing backdoor\nattacks and propose new DPBAs called CorruptEncoder to CL. CorruptEncoder\nintroduces a new attack strategy to create poisoned inputs and uses a\ntheory-guided method to maximize attack effectiveness. Our experiments show\nthat CorruptEncoder substantially outperforms existing DPBAs. In particular,\nCorruptEncoder is the first DPBA that achieves more than 90% attack success\nrates with only a few (3) reference images and a small poisoning ratio 0.5%.\nMoreover, we also propose a defense, called localized cropping, to defend\nagainst DPBAs. Our results show that our defense can reduce the effectiveness\nof DPBAs, but it sacrifices the utility of the encoder, highlighting the need\nfor new defenses.\n", "date": "2022-11-15", "categories": []}
{"id": "2212.06776", "title": "Unfolding Local Growth Rate Estimates for (Almost) Perfect Adversarial\n  Detection", "abstract": "  Convolutional neural networks (CNN) define the state-of-the-art solution on\nmany perceptual tasks. However, current CNN approaches largely remain\nvulnerable against adversarial perturbations of the input that have been\ncrafted specifically to fool the system while being quasi-imperceptible to the\nhuman eye. In recent years, various approaches have been proposed to defend\nCNNs against such attacks, for example by model hardening or by adding explicit\ndefence mechanisms. Thereby, a small \"detector\" is included in the network and\ntrained on the binary classification task of distinguishing genuine data from\ndata containing adversarial perturbations. In this work, we propose a simple\nand light-weight detector, which leverages recent findings on the relation\nbetween networks' local intrinsic dimensionality (LID) and adversarial attacks.\nBased on a re-interpretation of the LID measure and several simple adaptations,\nwe surpass the state-of-the-art on adversarial detection by a significant\nmargin and reach almost perfect results in terms of F1-score for several\nnetworks and datasets. Sources available at:\nhttps://github.com/adverML/multiLID\n", "date": "2022-12-13", "categories": []}
{"id": "2402.05493", "title": "Investigating White-Box Attacks for On-Device Models", "abstract": "  Numerous mobile apps have leveraged deep learning capabilities. However,\non-device models are vulnerable to attacks as they can be easily extracted from\ntheir corresponding mobile apps. Existing on-device attacking approaches only\ngenerate black-box attacks, which are far less effective and efficient than\nwhite-box strategies. This is because mobile deep learning frameworks like\nTFLite do not support gradient computing, which is necessary for white-box\nattacking algorithms. Thus, we argue that existing findings may underestimate\nthe harmfulness of on-device attacks. To this end, we conduct a study to answer\nthis research question: Can on-device models be directly attacked via white-box\nstrategies? We first systematically analyze the difficulties of transforming\nthe on-device model to its debuggable version, and propose a Reverse\nEngineering framework for On-device Models (REOM), which automatically reverses\nthe compiled on-device TFLite model to the debuggable model. Specifically, REOM\nfirst transforms compiled on-device models into Open Neural Network Exchange\nformat, then removes the non-debuggable parts, and converts them to the\ndebuggable DL models format that allows attackers to exploit in a white-box\nsetting. Our experimental results show that our approach is effective in\nachieving automated transformation among 244 TFLite models. Compared with\nprevious attacks using surrogate models, REOM enables attackers to achieve\nhigher attack success rates with a hundred times smaller attack perturbations.\nIn addition, because the ONNX platform has plenty of tools for model format\nexchanging, the proposed method based on the ONNX platform can be adapted to\nother model formats. Our findings emphasize the need for developers to\ncarefully consider their model deployment strategies, and use white-box methods\nto evaluate the vulnerability of on-device models.\n", "date": "2024-02-08", "categories": []}
{"id": "2402.18787", "title": "Enhancing the \"Immunity\" of Mixture-of-Experts Networks for Adversarial\n  Defense", "abstract": "  Recent studies have revealed the vulnerability of Deep Neural Networks (DNNs)\nto adversarial examples, which can easily fool DNNs into making incorrect\npredictions. To mitigate this deficiency, we propose a novel adversarial\ndefense method called \"Immunity\" (Innovative MoE with MUtual information \\&\npositioN stabilITY) based on a modified Mixture-of-Experts (MoE) architecture\nin this work. The key enhancements to the standard MoE are two-fold: 1)\nintegrating of Random Switch Gates (RSGs) to obtain diverse network structures\nvia random permutation of RSG parameters at evaluation time, despite of RSGs\nbeing determined after one-time training; 2) devising innovative Mutual\nInformation (MI)-based and Position Stability-based loss functions by\ncapitalizing on Grad-CAM's explanatory power to increase the diversity and the\ncausality of expert networks. Notably, our MI-based loss operates directly on\nthe heatmaps, thereby inducing subtler negative impacts on the classification\nperformance when compared to other losses of the same type, theoretically.\nExtensive evaluation validates the efficacy of the proposed approach in\nimproving adversarial robustness against a wide range of attacks.\n", "date": "2024-02-29", "categories": []}
{"id": "2402.18792", "title": "MPAT: Building Robust Deep Neural Networks against Textual Adversarial\n  Attacks", "abstract": "  Deep neural networks have been proven to be vulnerable to adversarial\nexamples and various methods have been proposed to defend against adversarial\nattacks for natural language processing tasks. However, previous defense\nmethods have limitations in maintaining effective defense while ensuring the\nperformance of the original task. In this paper, we propose a malicious\nperturbation based adversarial training method (MPAT) for building robust deep\nneural networks against textual adversarial attacks. Specifically, we construct\na multi-level malicious example generation strategy to generate adversarial\nexamples with malicious perturbations, which are used instead of original\ninputs for model training. Additionally, we employ a novel training objective\nfunction to ensure achieving the defense goal without compromising the\nperformance on the original task. We conduct comprehensive experiments to\nevaluate our defense method by attacking five victim models on three benchmark\ndatasets. The result demonstrates that our method is more effective against\nmalicious adversarial attacks compared with previous defense methods while\nmaintaining or further improving the performance on the original task.\n", "date": "2024-02-29", "categories": []}
{"id": "2402.19355", "title": "Unraveling Adversarial Examples against Speaker Identification --\n  Techniques for Attack Detection and Victim Model Classification", "abstract": "  Adversarial examples have proven to threaten speaker identification systems,\nand several countermeasures against them have been proposed. In this paper, we\npropose a method to detect the presence of adversarial examples, i.e., a binary\nclassifier distinguishing between benign and adversarial examples. We build\nupon and extend previous work on attack type classification by exploring new\narchitectures. Additionally, we introduce a method for identifying the victim\nmodel on which the adversarial attack is carried out. To achieve this, we\ngenerate a new dataset containing multiple attacks performed against various\nvictim models. We achieve an AUC of 0.982 for attack detection, with no more\nthan a 0.03 drop in performance for unknown attacks. Our attack classification\naccuracy (excluding benign) reaches 86.48% across eight attack types using our\nLightResNet34 architecture, while our victim model classification accuracy\nreaches 72.28% across four victim models.\n", "date": "2024-02-29", "categories": []}
{"id": "2207.06154", "title": "On the Robustness of Bayesian Neural Networks to Adversarial Attacks", "abstract": "  Vulnerability to adversarial attacks is one of the principal hurdles to the\nadoption of deep learning in safety-critical applications. Despite significant\nefforts, both practical and theoretical, training deep learning models robust\nto adversarial attacks is still an open problem. In this paper, we analyse the\ngeometry of adversarial attacks in the large-data, overparameterized limit for\nBayesian Neural Networks (BNNs). We show that, in the limit, vulnerability to\ngradient-based attacks arises as a result of degeneracy in the data\ndistribution, i.e., when the data lies on a lower-dimensional submanifold of\nthe ambient space. As a direct consequence, we demonstrate that in this limit\nBNN posteriors are robust to gradient-based adversarial attacks. Crucially, we\nprove that the expected gradient of the loss with respect to the BNN posterior\ndistribution is vanishing, even when each neural network sampled from the\nposterior is vulnerable to gradient-based attacks. Experimental results on the\nMNIST, Fashion MNIST, and half moons datasets, representing the finite data\nregime, with BNNs trained with Hamiltonian Monte Carlo and Variational\nInference, support this line of arguments, showing that BNNs can display both\nhigh accuracy on clean data and robustness to both gradient-based and\ngradient-free based adversarial attacks.\n", "date": "2022-07-13", "categories": []}
{"id": "2310.10010", "title": "Black-box Targeted Adversarial Attack on Segment Anything (SAM)", "abstract": "  Deep recognition models are widely vulnerable to adversarial examples, which\nchange the model output by adding quasi-imperceptible perturbation to the image\ninput. Recently, Segment Anything Model (SAM) has emerged to become a popular\nfoundation model in computer vision due to its impressive generalization to\nunseen data and tasks. Realizing flexible attacks on SAM is beneficial for\nunderstanding the robustness of SAM in the adversarial context. To this end,\nthis work aims to achieve a targeted adversarial attack (TAA) on SAM.\nSpecifically, under a certain prompt, the goal is to make the predicted mask of\nan adversarial example resemble that of a given target image. The task of TAA\non SAM has been realized in a recent arXiv work in the white-box setup by\nassuming access to prompt and model, which is thus less practical. To address\nthe issue of prompt dependence, we propose a simple yet effective approach by\nonly attacking the image encoder. Moreover, we propose a novel regularization\nloss to enhance the cross-model transferability by increasing the feature\ndominance of adversarial images over random natural images. Extensive\nexperiments verify the effectiveness of our proposed simple techniques to\nconduct a successful black-box TAA on SAM.\n", "date": "2023-10-16", "categories": []}
{"id": "2401.11618", "title": "Efficient local linearity regularization to overcome catastrophic\n  overfitting", "abstract": "  Catastrophic overfitting (CO) in single-step adversarial training (AT)\nresults in abrupt drops in the adversarial test accuracy (even down to 0%). For\nmodels trained with multi-step AT, it has been observed that the loss function\nbehaves locally linearly with respect to the input, this is however lost in\nsingle-step AT. To address CO in single-step AT, several methods have been\nproposed to enforce local linearity of the loss via regularization. However,\nthese regularization terms considerably slow down training due to Double\nBackpropagation. Instead, in this work, we introduce a regularization term,\ncalled ELLE, to mitigate CO effectively and efficiently in classical AT\nevaluations, as well as some more difficult regimes, e.g., large adversarial\nperturbations and long training schedules. Our regularization term can be\ntheoretically linked to curvature of the loss function and is computationally\ncheaper than previous methods by avoiding Double Backpropagation. Our thorough\nexperimental validation demonstrates that our work does not suffer from CO,\neven in challenging settings where previous works suffer from it. We also\nnotice that adapting our regularization parameter during training (ELLE-A)\ngreatly improves the performance, specially in large $\\epsilon$ setups. Our\nimplementation is available in https://github.com/LIONS-EPFL/ELLE .\n", "date": "2024-01-21", "categories": []}
{"id": "2402.15808", "title": "Optimal Zero-Shot Detector for Multi-Armed Attacks", "abstract": "  This paper explores a scenario in which a malicious actor employs a\nmulti-armed attack strategy to manipulate data samples, offering them various\navenues to introduce noise into the dataset. Our central objective is to\nprotect the data by detecting any alterations to the input. We approach this\ndefensive strategy with utmost caution, operating in an environment where the\ndefender possesses significantly less information compared to the attacker.\nSpecifically, the defender is unable to utilize any data samples for training a\ndefense model or verifying the integrity of the channel. Instead, the defender\nrelies exclusively on a set of pre-existing detectors readily available \"off\nthe shelf\". To tackle this challenge, we derive an innovative\ninformation-theoretic defense approach that optimally aggregates the decisions\nmade by these detectors, eliminating the need for any training data. We further\nexplore a practical use-case scenario for empirical evaluation, where the\nattacker possesses a pre-trained classifier and launches well-known adversarial\nattacks against it. Our experiments highlight the effectiveness of our proposed\nsolution, even in scenarios that deviate from the optimal setup.\n", "date": "2024-02-24", "categories": []}
{"id": "2402.17533", "title": "Black-box Adversarial Attacks Against Image Quality Assessment Models", "abstract": "  The goal of No-Reference Image Quality Assessment (NR-IQA) is to predict the\nperceptual quality of an image in line with its subjective evaluation. To put\nthe NR-IQA models into practice, it is essential to study their potential\nloopholes for model refinement. This paper makes the first attempt to explore\nthe black-box adversarial attacks on NR-IQA models. Specifically, we first\nformulate the attack problem as maximizing the deviation between the estimated\nquality scores of original and perturbed images, while restricting the\nperturbed image distortions for visual quality preservation. Under such\nformulation, we then design a Bi-directional loss function to mislead the\nestimated quality scores of adversarial examples towards an opposite direction\nwith maximum deviation. On this basis, we finally develop an efficient and\neffective black-box attack method against NR-IQA models. Extensive experiments\nreveal that all the evaluated NR-IQA models are vulnerable to the proposed\nattack method. And the generated perturbations are not transferable, enabling\nthem to serve the investigation of specialities of disparate IQA models.\n", "date": "2024-02-27", "categories": []}
{"id": "2402.18027", "title": "Breaking the Black-Box: Confidence-Guided Model Inversion Attack for\n  Distribution Shift", "abstract": "  Model inversion attacks (MIAs) seek to infer the private training data of a\ntarget classifier by generating synthetic images that reflect the\ncharacteristics of the target class through querying the model. However, prior\nstudies have relied on full access to the target model, which is not practical\nin real-world scenarios. Additionally, existing black-box MIAs assume that the\nimage prior and target model follow the same distribution. However, when\nconfronted with diverse data distribution settings, these methods may result in\nsuboptimal performance in conducting attacks. To address these limitations,\nthis paper proposes a \\textbf{C}onfidence-\\textbf{G}uided \\textbf{M}odel\n\\textbf{I}nversion attack method called CG-MI, which utilizes the latent space\nof a pre-trained publicly available generative adversarial network (GAN) as\nprior information and gradient-free optimizer, enabling high-resolution MIAs\nacross different data distributions in a black-box setting. Our experiments\ndemonstrate that our method significantly \\textbf{outperforms the SOTA\nblack-box MIA by more than 49\\% for Celeba and 58\\% for Facescrub in different\ndistribution settings}. Furthermore, our method exhibits the ability to\ngenerate high-quality images \\textbf{comparable to those produced by white-box\nattacks}. Our method provides a practical and effective solution for black-box\nmodel inversion attacks.\n", "date": "2024-02-28", "categories": []}
{"id": "2402.18211", "title": "Catastrophic Overfitting: A Potential Blessing in Disguise", "abstract": "  Fast Adversarial Training (FAT) has gained increasing attention within the\nresearch community owing to its efficacy in improving adversarial robustness.\nParticularly noteworthy is the challenge posed by catastrophic overfitting (CO)\nin this field. Although existing FAT approaches have made strides in mitigating\nCO, the ascent of adversarial robustness occurs with a non-negligible decline\nin classification accuracy on clean samples. To tackle this issue, we initially\nemploy the feature activation differences between clean and adversarial\nexamples to analyze the underlying causes of CO. Intriguingly, our findings\nreveal that CO can be attributed to the feature coverage induced by a few\nspecific pathways. By intentionally manipulating feature activation differences\nin these pathways with well-designed regularization terms, we can effectively\nmitigate and induce CO, providing further evidence for this observation.\nNotably, models trained stably with these terms exhibit superior performance\ncompared to prior FAT work. On this basis, we harness CO to achieve `attack\nobfuscation', aiming to bolster model performance. Consequently, the models\nsuffering from CO can attain optimal classification accuracy on both clean and\nadversarial data when adding random noise to inputs during evaluation. We also\nvalidate their robustness against transferred adversarial examples and the\nnecessity of inducing CO to improve robustness. Hence, CO may not be a problem\nthat has to be solved.\n", "date": "2024-02-28", "categories": []}
{"id": "2310.06958", "title": "Comparing the Robustness of Modern No-Reference Image- and Video-Quality\n  Metrics to Adversarial Attacks", "abstract": "  Nowadays, neural-network-based image- and video-quality metrics perform\nbetter than traditional methods. However, they also became more vulnerable to\nadversarial attacks that increase metrics' scores without improving visual\nquality. The existing benchmarks of quality metrics compare their performance\nin terms of correlation with subjective quality and calculation time.\nNonetheless, the adversarial robustness of image-quality metrics is also an\narea worth researching. This paper analyses modern metrics' robustness to\ndifferent adversarial attacks. We adapted adversarial attacks from computer\nvision tasks and compared attacks' efficiency against 15 no-reference image-\nand video-quality metrics. Some metrics showed high resistance to adversarial\nattacks, which makes their usage in benchmarks safer than vulnerable metrics.\nThe benchmark accepts submissions of new metrics for researchers who want to\nmake their metrics more robust to attacks or to find such metrics for their\nneeds. The latest results can be found online:\nhttps://videoprocessing.ai/benchmarks/metrics-robustness.html.\n", "date": "2023-10-10", "categories": []}
{"id": "2401.12461", "title": "Fast Adversarial Training against Textual Adversarial Attacks", "abstract": "  Many adversarial defense methods have been proposed to enhance the\nadversarial robustness of natural language processing models. However, most of\nthem introduce additional pre-set linguistic knowledge and assume that the\nsynonym candidates used by attackers are accessible, which is an ideal\nassumption. We delve into adversarial training in the embedding space and\npropose a Fast Adversarial Training (FAT) method to improve the model\nrobustness in the synonym-unaware scenario from the perspective of single-step\nperturbation generation and perturbation initialization. Based on the\nobservation that the adversarial perturbations crafted by single-step and\nmulti-step gradient ascent are similar, FAT uses single-step gradient ascent to\ncraft adversarial examples in the embedding space to expedite the training\nprocess. Based on the observation that the perturbations generated on the\nidentical training sample in successive epochs are similar, FAT fully utilizes\nhistorical information when initializing the perturbation. Extensive\nexperiments demonstrate that FAT significantly boosts the robustness of BERT\nmodels in the synonym-unaware scenario, and outperforms the defense baselines\nunder various attacks with character-level and word-level modifications.\n", "date": "2024-01-23", "categories": []}
{"id": "2402.13946", "title": "AttackGNN: Red-Teaming GNNs in Hardware Security Using Reinforcement\n  Learning", "abstract": "  Machine learning has shown great promise in addressing several critical\nhardware security problems. In particular, researchers have developed novel\ngraph neural network (GNN)-based techniques for detecting intellectual property\n(IP) piracy, detecting hardware Trojans (HTs), and reverse engineering\ncircuits, to name a few. These techniques have demonstrated outstanding\naccuracy and have received much attention in the community. However, since\nthese techniques are used for security applications, it is imperative to\nevaluate them thoroughly and ensure they are robust and do not compromise the\nsecurity of integrated circuits.\n  In this work, we propose AttackGNN, the first red-team attack on GNN-based\ntechniques in hardware security. To this end, we devise a novel reinforcement\nlearning (RL) agent that generates adversarial examples, i.e., circuits,\nagainst the GNN-based techniques. We overcome three challenges related to\neffectiveness, scalability, and generality to devise a potent RL agent. We\ntarget five GNN-based techniques for four crucial classes of problems in\nhardware security: IP piracy, detecting/localizing HTs, reverse engineering,\nand hardware obfuscation. Through our approach, we craft circuits that fool all\nGNNs considered in this work. For instance, to evade IP piracy detection, we\ngenerate adversarial pirated circuits that fool the GNN-based defense into\nclassifying our crafted circuits as not pirated. For attacking HT localization\nGNN, our attack generates HT-infested circuits that fool the defense on all\ntested circuits. We obtain a similar 100% success rate against GNNs for all\nclasses of problems.\n", "date": "2024-02-21", "categories": []}
{"id": "2402.16926", "title": "On the (In)feasibility of ML Backdoor Detection as an Hypothesis Testing\n  Problem", "abstract": "  We introduce a formal statistical definition for the problem of backdoor\ndetection in machine learning systems and use it to analyze the feasibility of\nsuch problems, providing evidence for the utility and applicability of our\ndefinition. The main contributions of this work are an impossibility result and\nan achievability result for backdoor detection. We show a no-free-lunch\ntheorem, proving that universal (adversary-unaware) backdoor detection is\nimpossible, except for very small alphabet sizes. Thus, we argue, that backdoor\ndetection methods need to be either explicitly, or implicitly adversary-aware.\nHowever, our work does not imply that backdoor detection cannot work in\nspecific scenarios, as evidenced by successful backdoor detection methods in\nthe scientific literature. Furthermore, we connect our definition to the\nprobably approximately correct (PAC) learnability of the out-of-distribution\ndetection problem.\n", "date": "2024-02-26", "categories": []}
{"id": "2402.17104", "title": "Adversarial Perturbations of Physical Signals", "abstract": "  We investigate the vulnerability of computer-vision-based signal classifiers\nto adversarial perturbations of their inputs, where the signals and\nperturbations are subject to physical constraints. We consider a scenario in\nwhich a source and interferer emit signals that propagate as waves to a\ndetector, which attempts to classify the source by analyzing the spectrogram of\nthe signal it receives using a pre-trained neural network. By solving\nPDE-constrained optimization problems, we construct interfering signals that\ncause the detector to misclassify the source even though the perturbations to\nthe spectrogram of the received signal are nearly imperceptible. Though such\nproblems can have millions of decision variables, we introduce methods to solve\nthem efficiently. Our experiments demonstrate that one can compute effective\nand physically realizable adversarial perturbations for a variety of machine\nlearning models under various physical conditions.\n", "date": "2024-02-27", "categories": []}
{"id": "2402.15570", "title": "Fast Adversarial Attacks on Language Models In One GPU Minute", "abstract": "  In this paper, we introduce a novel class of fast, beam search-based\nadversarial attack (BEAST) for Language Models (LMs). BEAST employs\ninterpretable parameters, enabling attackers to balance between attack speed,\nsuccess rate, and the readability of adversarial prompts. The computational\nefficiency of BEAST facilitates us to investigate its applications on LMs for\njailbreaking, eliciting hallucinations, and privacy attacks. Our gradient-free\ntargeted attack can jailbreak aligned LMs with high attack success rates within\none minute. For instance, BEAST can jailbreak Vicuna-7B-v1.5 under one minute\nwith a success rate of 89% when compared to a gradient-based baseline that\ntakes over an hour to achieve 70% success rate using a single Nvidia RTX A6000\n48GB GPU. Additionally, we discover a unique outcome wherein our untargeted\nattack induces hallucinations in LM chatbots. Through human evaluations, we\nfind that our untargeted attack causes Vicuna-7B-v1.5 to produce ~15% more\nincorrect outputs when compared to LM outputs in the absence of our attack. We\nalso learn that 22% of the time, BEAST causes Vicuna to generate outputs that\nare not relevant to the original prompt. Further, we use BEAST to generate\nadversarial prompts in a few seconds that can boost the performance of existing\nmembership inference attacks for LMs. We believe that our fast attack, BEAST,\nhas the potential to accelerate research in LM security and privacy. Our\ncodebase is publicly available at https://github.com/vinusankars/BEAST.\n", "date": "2024-02-23", "categories": []}
{"id": "2402.15586", "title": "Distilling Adversarial Robustness Using Heterogeneous Teachers", "abstract": "  Achieving resiliency against adversarial attacks is necessary prior to\ndeploying neural network classifiers in domains where misclassification incurs\nsubstantial costs, e.g., self-driving cars or medical imaging. Recent work has\ndemonstrated that robustness can be transferred from an adversarially trained\nteacher to a student model using knowledge distillation. However, current\nmethods perform distillation using a single adversarial and vanilla teacher and\nconsider homogeneous architectures (i.e., residual networks) that are\nsusceptible to misclassify examples from similar adversarial subspaces. In this\nwork, we develop a defense framework against adversarial attacks by distilling\nadversarial robustness using heterogeneous teachers (DARHT). In DARHT, the\nstudent model explicitly represents teacher logits in a student-teacher feature\nmap and leverages multiple teachers that exhibit low adversarial example\ntransferability (i.e., exhibit high performance on dissimilar adversarial\nexamples). Experiments on classification tasks in both white-box and black-box\nscenarios demonstrate that DARHT achieves state-of-the-art clean and robust\naccuracies when compared to competing adversarial training and distillation\nmethods in the CIFAR-10, CIFAR-100, and Tiny ImageNet datasets. Comparisons\nwith homogeneous and heterogeneous teacher sets suggest that leveraging\nteachers with low adversarial example transferability increases student model\nrobustness.\n", "date": "2024-02-23", "categories": []}
{"id": "2402.15959", "title": "Towards Robust Image Stitching: An Adaptive Resistance Learning against\n  Compatible Attacks", "abstract": "  Image stitching seamlessly integrates images captured from varying\nperspectives into a single wide field-of-view image. Such integration not only\nbroadens the captured scene but also augments holistic perception in computer\nvision applications. Given a pair of captured images, subtle perturbations and\ndistortions which go unnoticed by the human visual system tend to attack the\ncorrespondence matching, impairing the performance of image stitching\nalgorithms. In light of this challenge, this paper presents the first attempt\nto improve the robustness of image stitching against adversarial attacks.\nSpecifically, we introduce a stitching-oriented attack~(SoA), tailored to\namplify the alignment loss within overlapping regions, thereby targeting the\nfeature matching procedure. To establish an attack resistant model, we delve\ninto the robustness of stitching architecture and develop an adaptive\nadversarial training~(AAT) to balance attack resistance with stitching\nprecision. In this way, we relieve the gap between the routine adversarial\ntraining and benign models, ensuring resilience without quality compromise.\nComprehensive evaluation across real-world and synthetic datasets validate the\ndeterioration of SoA on stitching performance. Furthermore, AAT emerges as a\nmore robust solution against adversarial perturbations, delivering superior\nstitching results. Code is available at:https://github.com/Jzy2017/TRIS.\n", "date": "2024-02-25", "categories": []}
{"id": "2402.16005", "title": "Adversarial-Robust Transfer Learning for Medical Imaging via Domain\n  Assimilation", "abstract": "  In the field of Medical Imaging, extensive research has been dedicated to\nleveraging its potential in uncovering critical diagnostic features in\npatients. Artificial Intelligence (AI)-driven medical diagnosis relies on\nsophisticated machine learning and deep learning models to analyze, detect, and\nidentify diseases from medical images. Despite the remarkable performance of\nthese models, characterized by high accuracy, they grapple with trustworthiness\nissues. The introduction of a subtle perturbation to the original image\nempowers adversaries to manipulate the prediction output, redirecting it to\nother targeted or untargeted classes. Furthermore, the scarcity of publicly\navailable medical images, constituting a bottleneck for reliable training, has\nled contemporary algorithms to depend on pretrained models grounded on a large\nset of natural images -- a practice referred to as transfer learning. However,\na significant {\\em domain discrepancy} exists between natural and medical\nimages, which causes AI models resulting from transfer learning to exhibit\nheightened {\\em vulnerability} to adversarial attacks. This paper proposes a\n{\\em domain assimilation} approach that introduces texture and color adaptation\ninto transfer learning, followed by a texture preservation component to\nsuppress undesired distortion. We systematically analyze the performance of\ntransfer learning in the face of various adversarial attacks under different\ndata modalities, with the overarching goal of fortifying the model's robustness\nand security in medical imaging tasks. The results demonstrate high\neffectiveness in reducing attack efficacy, contributing toward more trustworthy\ntransfer learning in biomedical applications.\n", "date": "2024-02-25", "categories": []}
{"id": "2402.16397", "title": "Investigating Deep Watermark Security: An Adversarial Transferability\n  Perspective", "abstract": "  The rise of generative neural networks has triggered an increased demand for\nintellectual property (IP) protection in generated content. Deep watermarking\ntechniques, recognized for their flexibility in IP protection, have garnered\nsignificant attention. However, the surge in adversarial transferable attacks\nposes unprecedented challenges to the security of deep watermarking\ntechniques-an area currently lacking systematic investigation. This study fills\nthis gap by introducing two effective transferable attackers to assess the\nvulnerability of deep watermarks against erasure and tampering risks.\nSpecifically, we initially define the concept of local sample density,\nutilizing it to deduce theorems on the consistency of model outputs. Upon\ndiscovering that perturbing samples towards high sample density regions (HSDR)\nof the target class enhances targeted adversarial transferability, we propose\nthe Easy Sample Selection (ESS) mechanism and the Easy Sample Matching Attack\n(ESMA) method. Additionally, we propose the Bottleneck Enhanced Mixup (BEM)\nthat integrates information bottleneck theory to reduce the generator's\ndependence on irrelevant noise. Experiments show a significant enhancement in\nthe success rate of targeted transfer attacks for both ESMA and BEM-ESMA\nmethods. We further conduct a comprehensive evaluation using ESMA and BEM-ESMA\nas measurements, considering model architecture and watermark encoding length,\nand achieve some impressive findings.\n", "date": "2024-02-26", "categories": []}
{"id": "2402.16470", "title": "Unveiling Vulnerability of Self-Attention", "abstract": "  Pre-trained language models (PLMs) are shown to be vulnerable to minor word\nchanges, which poses a big threat to real-world systems. While previous studies\ndirectly focus on manipulating word inputs, they are limited by their means of\ngenerating adversarial samples, lacking generalization to versatile real-world\nattack. This paper studies the basic structure of transformer-based PLMs, the\nself-attention (SA) mechanism. (1) We propose a powerful perturbation technique\n\\textit{HackAttend}, which perturbs the attention scores within the SA matrices\nvia meticulously crafted attention masks. We show that state-of-the-art PLMs\nfall into heavy vulnerability that minor attention perturbations $(1\\%)$ can\nproduce a very high attack success rate $(98\\%)$. Our paper expands the\nconventional text attack of word perturbations to more general structural\nperturbations. (2) We introduce \\textit{S-Attend}, a novel smoothing technique\nthat effectively makes SA robust via structural perturbations. We empirically\ndemonstrate that this simple yet effective technique achieves robust\nperformance on par with adversarial training when facing various text\nattackers. Code is publicly available at \\url{github.com/liongkj/HackAttend}.\n", "date": "2024-02-26", "categories": []}
{"id": "2402.16586", "title": "Improving the JPEG-resistance of Adversarial Attacks on Face Recognition\n  by Interpolation Smoothing", "abstract": "  JPEG compression can significantly impair the performance of adversarial face\nexamples, which previous adversarial attacks on face recognition (FR) have not\nadequately addressed. Considering this challenge, we propose a novel\nadversarial attack on FR that aims to improve the resistance of adversarial\nexamples against JPEG compression. Specifically, during the iterative process\nof generating adversarial face examples, we interpolate the adversarial face\nexamples into a smaller size. Then we utilize these interpolated adversarial\nface examples to create the adversarial examples in the next iteration.\nSubsequently, we restore the adversarial face examples to their original size\nby interpolating. Throughout the entire process, our proposed method can smooth\nthe adversarial perturbations, effectively mitigating the presence of\nhigh-frequency signals in the crafted adversarial face examples that are\ntypically eliminated by JPEG compression. Our experimental results demonstrate\nthe effectiveness of our proposed method in improving the JPEG-resistance of\nadversarial face examples.\n", "date": "2024-02-26", "categories": []}
{"id": "2308.10779", "title": "Spear and Shield: Adversarial Attacks and Defense Methods for\n  Model-Based Link Prediction on Continuous-Time Dynamic Graphs", "abstract": "  Real-world graphs are dynamic, constantly evolving with new interactions,\nsuch as financial transactions in financial networks. Temporal Graph Neural\nNetworks (TGNNs) have been developed to effectively capture the evolving\npatterns in dynamic graphs. While these models have demonstrated their\nsuperiority, being widely adopted in various important fields, their\nvulnerabilities against adversarial attacks remain largely unexplored. In this\npaper, we propose T-SPEAR, a simple and effective adversarial attack method for\nlink prediction on continuous-time dynamic graphs, focusing on investigating\nthe vulnerabilities of TGNNs. Specifically, before the training procedure of a\nvictim model, which is a TGNN for link prediction, we inject edge perturbations\nto the data that are unnoticeable in terms of the four constraints we propose,\nand yet effective enough to cause malfunction of the victim model. Moreover, we\npropose a robust training approach T-SHIELD to mitigate the impact of\nadversarial attacks. By using edge filtering and enforcing temporal smoothness\nto node embeddings, we enhance the robustness of the victim model. Our\nexperimental study shows that T-SPEAR significantly degrades the victim model's\nperformance on link prediction tasks, and even more, our attacks are\ntransferable to other TGNNs, which differ from the victim model assumed by the\nattacker. Moreover, we demonstrate that T-SHIELD effectively filters out\nadversarial edges and exhibits robustness against adversarial attacks,\nsurpassing the link prediction performance of the naive TGNN by up to 11.2%\nunder T-SPEAR.\n", "date": "2023-08-21", "categories": []}
{"id": "2402.14937", "title": "SoK: Analyzing Adversarial Examples: A Framework to Study Adversary\n  Knowledge", "abstract": "  Adversarial examples are malicious inputs to machine learning models that\ntrigger a misclassification. This type of attack has been studied for close to\na decade, and we find that there is a lack of study and formalization of\nadversary knowledge when mounting attacks. This has yielded a complex space of\nattack research with hard-to-compare threat models and attacks. We focus on the\nimage classification domain and provide a theoretical framework to study\nadversary knowledge inspired by work in order theory. We present an adversarial\nexample game, inspired by cryptographic games, to standardize attacks. We\nsurvey recent attacks in the image classification domain and classify their\nadversary's knowledge in our framework. From this systematization, we compile\nresults that both confirm existing beliefs about adversary knowledge, such as\nthe potency of information about the attacked model as well as allow us to\nderive new conclusions on the difficulty associated with the white-box and\ntransferable threat models, for example, that transferable attacks might not be\nas difficult as previously thought.\n", "date": "2024-02-22", "categories": []}
{"id": "2402.14977", "title": "Mudjacking: Patching Backdoor Vulnerabilities in Foundation Models", "abstract": "  Foundation model has become the backbone of the AI ecosystem. In particular,\na foundation model can be used as a general-purpose feature extractor to build\nvarious downstream classifiers. However, foundation models are vulnerable to\nbackdoor attacks and a backdoored foundation model is a single-point-of-failure\nof the AI ecosystem, e.g., multiple downstream classifiers inherit the backdoor\nvulnerabilities simultaneously. In this work, we propose Mudjacking, the first\nmethod to patch foundation models to remove backdoors. Specifically, given a\nmisclassified trigger-embedded input detected after a backdoored foundation\nmodel is deployed, Mudjacking adjusts the parameters of the foundation model to\nremove the backdoor. We formulate patching a foundation model as an\noptimization problem and propose a gradient descent based method to solve it.\nWe evaluate Mudjacking on both vision and language foundation models, eleven\nbenchmark datasets, five existing backdoor attacks, and thirteen adaptive\nbackdoor attacks. Our results show that Mudjacking can remove backdoor from a\nfoundation model while maintaining its utility.\n", "date": "2024-02-22", "categories": []}
{"id": "2402.14988", "title": "Verifiable Boosted Tree Ensembles", "abstract": "  Verifiable learning advocates for training machine learning models amenable\nto efficient security verification. Prior research demonstrated that specific\nclasses of decision tree ensembles -- called large-spread ensembles -- allow\nfor robustness verification in polynomial time against any norm-based attacker.\nThis study expands prior work on verifiable learning from basic ensemble\nmethods (i.e., hard majority voting) to advanced boosted tree ensembles, such\nas those trained using XGBoost or LightGBM. Our formal results indicate that\nrobustness verification is achievable in polynomial time when considering\nattackers based on the $L_\\infty$-norm, but remains NP-hard for other\nnorm-based attackers. Nevertheless, we present a pseudo-polynomial time\nalgorithm to verify robustness against attackers based on the $L_p$-norm for\nany $p \\in \\mathbb{N} \\cup \\{0\\}$, which in practice grants excellent\nperformance. Our experimental evaluation shows that large-spread boosted\nensembles are accurate enough for practical adoption, while being amenable to\nefficient security verification.\n", "date": "2024-02-22", "categories": []}
{"id": "2402.15218", "title": "BSPA: Exploring Black-box Stealthy Prompt Attacks against Image\n  Generators", "abstract": "  Extremely large image generators offer significant transformative potential\nacross diverse sectors. It allows users to design specific prompts to generate\nrealistic images through some black-box APIs. However, some studies reveal that\nimage generators are notably susceptible to attacks and generate Not Suitable\nFor Work (NSFW) contents by manually designed toxin texts, especially\nimperceptible to human observers. We urgently need a multitude of universal and\ntransferable prompts to improve the safety of image generators, especially\nblack-box-released APIs. Nevertheless, they are constrained by labor-intensive\ndesign processes and heavily reliant on the quality of the given instructions.\nTo achieve this, we introduce a black-box stealthy prompt attack (BSPA) that\nadopts a retriever to simulate attacks from API users. It can effectively\nharness filter scores to tune the retrieval space of sensitive words for\nmatching the input prompts, thereby crafting stealthy prompts tailored for\nimage generators. Significantly, this approach is model-agnostic and requires\nno internal access to the model's features, ensuring its applicability to a\nwide range of image generators. Building on BSPA, we have constructed an\nautomated prompt tool and a comprehensive prompt attack dataset (NSFWeval).\nExtensive experiments demonstrate that BSPA effectively explores the security\nvulnerabilities in a variety of state-of-the-art available black-box models,\nincluding Stable Diffusion XL, Midjourney, and DALL-E 2/3. Furthermore, we\ndevelop a resilient text filter and offer targeted recommendations to ensure\nthe security of image generators against prompt attacks in the future.\n", "date": "2024-02-23", "categories": []}
{"id": "2401.17499", "title": "AdvGPS: Adversarial GPS for Multi-Agent Perception Attack", "abstract": "  The multi-agent perception system collects visual data from sensors located\non various agents and leverages their relative poses determined by GPS signals\nto effectively fuse information, mitigating the limitations of single-agent\nsensing, such as occlusion. However, the precision of GPS signals can be\ninfluenced by a range of factors, including wireless transmission and\nobstructions like buildings. Given the pivotal role of GPS signals in\nperception fusion and the potential for various interference, it becomes\nimperative to investigate whether specific GPS signals can easily mislead the\nmulti-agent perception system. To address this concern, we frame the task as an\nadversarial attack challenge and introduce \\textsc{AdvGPS}, a method capable of\ngenerating adversarial GPS signals which are also stealthy for individual\nagents within the system, significantly reducing object detection accuracy. To\nenhance the success rates of these attacks in a black-box scenario, we\nintroduce three types of statistically sensitive natural discrepancies:\nappearance-based discrepancy, distribution-based discrepancy, and task-aware\ndiscrepancy. Our extensive experiments on the OPV2V dataset demonstrate that\nthese attacks substantially undermine the performance of state-of-the-art\nmethods, showcasing remarkable transferability across different point cloud\nbased 3D detection systems. This alarming revelation underscores the pressing\nneed to address security implications within multi-agent perception systems,\nthereby underscoring a critical area of research.\n", "date": "2024-01-30", "categories": []}
{"id": "2402.02695", "title": "Exploiting Class Probabilities for Black-box Sentence-level Attacks", "abstract": "  Sentence-level attacks craft adversarial sentences that are synonymous with\ncorrectly-classified sentences but are misclassified by the text classifiers.\nUnder the black-box setting, classifiers are only accessible through their\nfeedback to queried inputs, which is predominately available in the form of\nclass probabilities. Even though utilizing class probabilities results in\nstronger attacks, due to the challenges of using them for sentence-level\nattacks, existing attacks use either no feedback or only the class labels.\nOvercoming the challenges, we develop a novel algorithm that uses class\nprobabilities for black-box sentence-level attacks, investigate the\neffectiveness of using class probabilities on the attack's success, and examine\nthe question if it is worthy or practical to use class probabilities by\nblack-box sentence-level attacks. We conduct extensive evaluations of our\nattack comparing with the baselines across various classifiers and benchmark\ndatasets.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.13287", "title": "Manipulating hidden-Markov-model inferences by corrupting batch data", "abstract": "  Time-series models typically assume untainted and legitimate streams of data.\nHowever, a self-interested adversary may have incentive to corrupt this data,\nthereby altering a decision maker's inference. Within the broader field of\nadversarial machine learning, this research provides a novel, probabilistic\nperspective toward the manipulation of hidden Markov model inferences via\ncorrupted data. In particular, we provision a suite of corruption problems for\nfiltering, smoothing, and decoding inferences leveraging an adversarial risk\nanalysis approach. Multiple stochastic programming models are set forth that\nincorporate realistic uncertainties and varied attacker objectives. Three\ngeneral solution methods are developed by alternatively viewing the problem\nfrom frequentist and Bayesian perspectives. The efficacy of each method is\nillustrated via extensive, empirical testing. The developed methods are\ncharacterized by their solution quality and computational effort, resulting in\na stratification of techniques across varying problem-instance architectures.\nThis research highlights the weaknesses of hidden Markov models under\nadversarial activity, thereby motivating the need for robustification\ntechniques to ensure their security.\n", "date": "2024-02-19", "categories": []}
{"id": "2402.13487", "title": "Stealthy Adversarial Attacks on Stochastic Multi-Armed Bandits", "abstract": "  Adversarial attacks against stochastic multi-armed bandit (MAB) algorithms\nhave been extensively studied in the literature. In this work, we focus on\nreward poisoning attacks and find most existing attacks can be easily detected\nby our proposed detection method based on the test of homogeneity, due to their\naggressive nature in reward manipulations. This motivates us to study the\nnotion of stealthy attack against stochastic MABs and investigate the resulting\nattackability. Our analysis shows that against two popularly employed MAB\nalgorithms, UCB1 and $\\epsilon$-greedy, the success of a stealthy attack\ndepends on the environmental conditions and the realized reward of the arm\npulled in the first round. We also analyze the situation for general MAB\nalgorithms equipped with our attack detection method and find that it is\npossible to have a stealthy attack that almost always succeeds. This brings new\ninsights into the security risks of MAB algorithms.\n", "date": "2024-02-21", "categories": []}
{"id": "2402.13851", "title": "VL-Trojan: Multimodal Instruction Backdoor Attacks against\n  Autoregressive Visual Language Models", "abstract": "  Autoregressive Visual Language Models (VLMs) showcase impressive few-shot\nlearning capabilities in a multimodal context. Recently, multimodal instruction\ntuning has been proposed to further enhance instruction-following abilities.\nHowever, we uncover the potential threat posed by backdoor attacks on\nautoregressive VLMs during instruction tuning. Adversaries can implant a\nbackdoor by injecting poisoned samples with triggers embedded in instructions\nor images, enabling malicious manipulation of the victim model's predictions\nwith predefined triggers. Nevertheless, the frozen visual encoder in\nautoregressive VLMs imposes constraints on the learning of conventional image\ntriggers. Additionally, adversaries may encounter restrictions in accessing the\nparameters and architectures of the victim model. To address these challenges,\nwe propose a multimodal instruction backdoor attack, namely VL-Trojan. Our\napproach facilitates image trigger learning through an isolating and clustering\nstrategy and enhance black-box-attack efficacy via an iterative character-level\ntext trigger generation method. Our attack successfully induces target outputs\nduring inference, significantly surpassing baselines (+62.52\\%) in ASR.\nMoreover, it demonstrates robustness across various model scales and few-shot\nin-context reasoning scenarios.\n", "date": "2024-02-21", "categories": []}
{"id": "2402.13987", "title": "A Simple and Yet Fairly Effective Defense for Graph Neural Networks", "abstract": "  Graph Neural Networks (GNNs) have emerged as the dominant approach for\nmachine learning on graph-structured data. However, concerns have arisen\nregarding the vulnerability of GNNs to small adversarial perturbations.\nExisting defense methods against such perturbations suffer from high time\ncomplexity and can negatively impact the model's performance on clean graphs.\nTo address these challenges, this paper introduces NoisyGNNs, a novel defense\nmethod that incorporates noise into the underlying model's architecture. We\nestablish a theoretical connection between noise injection and the enhancement\nof GNN robustness, highlighting the effectiveness of our approach. We further\nconduct extensive empirical evaluations on the node classification task to\nvalidate our theoretical findings, focusing on two popular GNNs: the GCN and\nGIN. The results demonstrate that NoisyGNN achieves superior or comparable\ndefense performance to existing methods while minimizing added time complexity.\nThe NoisyGNN approach is model-agnostic, allowing it to be integrated with\ndifferent GNN architectures. Successful combinations of our NoisyGNN approach\nwith existing defense techniques demonstrate even further improved adversarial\ndefense results. Our code is publicly available at:\nhttps://github.com/Sennadir/NoisyGNN.\n", "date": "2024-02-21", "categories": []}
{"id": "2304.02688", "title": "Going Further: Flatness at the Rescue of Early Stopping for Adversarial\n  Example Transferability", "abstract": "  Transferability is the property of adversarial examples to be misclassified\nby other models than the surrogate model for which they were crafted. Previous\nresearch has shown that early stopping the training of the surrogate model\nsubstantially increases transferability. A common hypothesis to explain this is\nthat deep neural networks (DNNs) first learn robust features, which are more\ngeneric, thus a better surrogate. Then, at later epochs, DNNs learn non-robust\nfeatures, which are more brittle, hence worst surrogate. First, we tend to\nrefute this hypothesis, using transferability as a proxy for representation\nsimilarity. We then establish links between transferability and the exploration\nof the loss landscape in parameter space, focusing on sharpness, which is\naffected by early stopping. This leads us to evaluate surrogate models trained\nwith seven minimizers that minimize both loss value and loss sharpness. Among\nthem, SAM consistently outperforms early stopping by up to 28.8 percentage\npoints. We discover that the strong SAM regularization from large flat\nneighborhoods tightly links to transferability. Finally, the best\nsharpness-aware minimizers prove competitive with other training methods and\ncomplement existing transferability techniques.\n", "date": "2023-04-05", "categories": []}
{"id": "2305.17342", "title": "Rethinking Adversarial Policies: A Generalized Attack Formulation and\n  Provable Defense in RL", "abstract": "  Most existing works focus on direct perturbations to the victim's\nstate/action or the underlying transition dynamics to demonstrate the\nvulnerability of reinforcement learning agents to adversarial attacks. However,\nsuch direct manipulations may not be always realizable. In this paper, we\nconsider a multi-agent setting where a well-trained victim agent $\\nu$ is\nexploited by an attacker controlling another agent $\\alpha$ with an\n\\textit{adversarial policy}. Previous models do not account for the possibility\nthat the attacker may only have partial control over $\\alpha$ or that the\nattack may produce easily detectable \"abnormal\" behaviors. Furthermore, there\nis a lack of provably efficient defenses against these adversarial policies. To\naddress these limitations, we introduce a generalized attack framework that has\nthe flexibility to model to what extent the adversary is able to control the\nagent, and allows the attacker to regulate the state distribution shift and\nproduce stealthier adversarial policies. Moreover, we offer a provably\nefficient defense with polynomial convergence to the most robust victim policy\nthrough adversarial training with timescale separation. This stands in sharp\ncontrast to supervised learning, where adversarial training typically provides\nonly \\textit{empirical} defenses. Using the Robosumo competition experiments,\nwe show that our generalized attack formulation results in much stealthier\nadversarial policies when maintaining the same winning rate as baselines.\nAdditionally, our adversarial training approach yields stable learning dynamics\nand less exploitable victim policies.\n", "date": "2023-05-27", "categories": []}
{"id": "2402.12626", "title": "Indiscriminate Data Poisoning Attacks on Pre-trained Feature Extractors", "abstract": "  Machine learning models have achieved great success in supervised learning\ntasks for end-to-end training, which requires a large amount of labeled data\nthat is not always feasible. Recently, many practitioners have shifted to\nself-supervised learning methods that utilize cheap unlabeled data to learn a\ngeneral feature extractor via pre-training, which can be further applied to\npersonalized downstream tasks by simply training an additional linear layer\nwith limited labeled data. However, such a process may also raise concerns\nregarding data poisoning attacks. For instance, indiscriminate data poisoning\nattacks, which aim to decrease model utility by injecting a small number of\npoisoned data into the training set, pose a security risk to machine learning\nmodels, but have only been studied for end-to-end supervised learning. In this\npaper, we extend the exploration of the threat of indiscriminate attacks on\ndownstream tasks that apply pre-trained feature extractors. Specifically, we\npropose two types of attacks: (1) the input space attacks, where we modify\nexisting attacks to directly craft poisoned data in the input space. However,\ndue to the difficulty of optimization under constraints, we further propose (2)\nthe feature targeted attacks, where we mitigate the challenge with three\nstages, firstly acquiring target parameters for the linear head; secondly\nfinding poisoned features by treating the learned feature representations as a\ndataset; and thirdly inverting the poisoned features back to the input space.\nOur experiments examine such attacks in popular downstream tasks of fine-tuning\non the same dataset and transfer learning that considers domain adaptation.\nEmpirical results reveal that transfer learning is more vulnerable to our\nattacks. Additionally, input space attacks are a strong threat if no\ncountermeasures are posed, but are otherwise weaker than feature targeted\nattacks.\n", "date": "2024-02-20", "categories": []}
{"id": "2402.12673", "title": "Beyond Worst-case Attacks: Robust RL with Adaptive Defense via\n  Non-dominated Policies", "abstract": "  In light of the burgeoning success of reinforcement learning (RL) in diverse\nreal-world applications, considerable focus has been directed towards ensuring\nRL policies are robust to adversarial attacks during test time. Current\napproaches largely revolve around solving a minimax problem to prepare for\npotential worst-case scenarios. While effective against strong attacks, these\nmethods often compromise performance in the absence of attacks or the presence\nof only weak attacks. To address this, we study policy robustness under the\nwell-accepted state-adversarial attack model, extending our focus beyond only\nworst-case attacks. We first formalize this task at test time as a regret\nminimization problem and establish its intrinsic hardness in achieving\nsublinear regret when the baseline policy is from a general continuous policy\nclass, $\\Pi$. This finding prompts us to \\textit{refine} the baseline policy\nclass $\\Pi$ prior to test time, aiming for efficient adaptation within a finite\npolicy class $\\Tilde{\\Pi}$, which can resort to an adversarial bandit\nsubroutine. In light of the importance of a small, finite $\\Tilde{\\Pi}$, we\npropose a novel training-time algorithm to iteratively discover\n\\textit{non-dominated policies}, forming a near-optimal and minimal\n$\\Tilde{\\Pi}$, thereby ensuring both robustness and test-time efficiency.\nEmpirical validation on the Mujoco corroborates the superiority of our approach\nin terms of natural and robust performance, as well as adaptability to various\nattack scenarios.\n", "date": "2024-02-20", "categories": []}
{"id": "2307.11729", "title": "OUTFOX: LLM-Generated Essay Detection Through In-Context Learning with\n  Adversarially Generated Examples", "abstract": "  Large Language Models (LLMs) have achieved human-level fluency in text\ngeneration, making it difficult to distinguish between human-written and\nLLM-generated texts. This poses a growing risk of misuse of LLMs and demands\nthe development of detectors to identify LLM-generated texts. However, existing\ndetectors lack robustness against attacks: they degrade detection accuracy by\nsimply paraphrasing LLM-generated texts. Furthermore, a malicious user might\nattempt to deliberately evade the detectors based on detection results, but\nthis has not been assumed in previous studies. In this paper, we propose\nOUTFOX, a framework that improves the robustness of LLM-generated-text\ndetectors by allowing both the detector and the attacker to consider each\nother's output. In this framework, the attacker uses the detector's prediction\nlabels as examples for in-context learning and adversarially generates essays\nthat are harder to detect, while the detector uses the adversarially generated\nessays as examples for in-context learning to learn to detect essays from a\nstrong attacker. Experiments in the domain of student essays show that the\nproposed detector improves the detection performance on the attacker-generated\ntexts by up to +41.3 points F1-score. Furthermore, the proposed detector shows\na state-of-the-art detection performance: up to 96.9 points F1-score, beating\nexisting detectors on non-attacked texts. Finally, the proposed attacker\ndrastically degrades the performance of detectors by up to -57.0 points\nF1-score, massively outperforming the baseline paraphrasing method for evading\ndetection.\n", "date": "2023-07-21", "categories": []}
{"id": "2308.08925", "title": "A White-Box False Positive Adversarial Attack Method on Contrastive Loss\n  Based Offline Handwritten Signature Verification Models", "abstract": "  In this paper, we tackle the challenge of white-box false positive\nadversarial attacks on contrastive loss based offline handwritten signature\nverification models. We propose a novel attack method that treats the attack as\na style transfer between closely related but distinct writing styles. To guide\nthe generation of deceptive images, we introduce two new loss functions that\nenhance the attack success rate by perturbing the Euclidean distance between\nthe embedding vectors of the original and synthesized samples, while ensuring\nminimal perturbations by reducing the difference between the generated image\nand the original image. Our method demonstrates state-of-the-art performance in\nwhite-box attacks on contrastive loss based offline handwritten signature\nverification models, as evidenced by our experiments. The key contributions of\nthis paper include a novel false positive attack method, two new loss\nfunctions, effective style transfer in handwriting styles, and superior\nperformance in white-box false positive attacks compared to other white-box\nattack methods.\n", "date": "2023-08-17", "categories": []}
{"id": "2311.11509", "title": "Token-Level Adversarial Prompt Detection Based on Perplexity Measures\n  and Contextual Information", "abstract": "  In recent years, Large Language Models (LLM) have emerged as pivotal tools in\nvarious applications. However, these models are susceptible to adversarial\nprompt attacks, where attackers can carefully curate input strings that mislead\nLLMs into generating incorrect or undesired outputs. Previous work has revealed\nthat with relatively simple yet effective attacks based on discrete\noptimization, it is possible to generate adversarial prompts that bypass\nmoderation and alignment of the models. This vulnerability to adversarial\nprompts underscores a significant concern regarding the robustness and\nreliability of LLMs. Our work aims to address this concern by introducing a\nnovel approach to detecting adversarial prompts at a token level, leveraging\nthe LLM's capability to predict the next token's probability. We measure the\ndegree of the model's perplexity, where tokens predicted with high probability\nare considered normal, and those exhibiting high perplexity are flagged as\nadversarial. Additionaly, our method also integrates context understanding by\nincorporating neighboring token information to encourage the detection of\ncontiguous adversarial prompt sequences. To this end, we design two algorithms\nfor adversarial prompt detection: one based on optimization techniques and\nanother on Probabilistic Graphical Models (PGM). Both methods are equipped with\nefficient solving methods, ensuring efficient adversarial prompt detection. Our\ntoken-level detection result can be visualized as heatmap overlays on the text\nsequence, allowing for a clearer and more intuitive representation of which\npart of the text may contain adversarial prompts.\n", "date": "2023-11-20", "categories": []}
{"id": "2402.10983", "title": "Quantum-Inspired Analysis of Neural Network Vulnerabilities: The Role of\n  Conjugate Variables in System Attacks", "abstract": "  Neural networks demonstrate inherent vulnerability to small, non-random\nperturbations, emerging as adversarial attacks. Such attacks, born from the\ngradient of the loss function relative to the input, are discerned as input\nconjugates, revealing a systemic fragility within the network structure.\nIntriguingly, a mathematical congruence manifests between this mechanism and\nthe quantum physics' uncertainty principle, casting light on a hitherto\nunanticipated interdisciplinarity. This inherent susceptibility within neural\nnetwork systems is generally intrinsic, highlighting not only the innate\nvulnerability of these networks but also suggesting potential advancements in\nthe interdisciplinary area for understanding these black-box networks.\n", "date": "2024-02-16", "categories": []}
{"id": "2402.11083", "title": "VQAttack: Transferable Adversarial Attacks on Visual Question Answering\n  via Pre-trained Models", "abstract": "  Visual Question Answering (VQA) is a fundamental task in computer vision and\nnatural language process fields. Although the ``pre-training & finetuning''\nlearning paradigm significantly improves the VQA performance, the adversarial\nrobustness of such a learning paradigm has not been explored. In this paper, we\ndelve into a new problem: using a pre-trained multimodal source model to create\nadversarial image-text pairs and then transferring them to attack the target\nVQA models. Correspondingly, we propose a novel VQAttack model, which can\niteratively generate both image and text perturbations with the designed\nmodules: the large language model (LLM)-enhanced image attack and the\ncross-modal joint attack module. At each iteration, the LLM-enhanced image\nattack module first optimizes the latent representation-based loss to generate\nfeature-level image perturbations. Then it incorporates an LLM to further\nenhance the image perturbations by optimizing the designed masked answer\nanti-recovery loss. The cross-modal joint attack module will be triggered at a\nspecific iteration, which updates the image and text perturbations\nsequentially. Notably, the text perturbation updates are based on both the\nlearned gradients in the word embedding space and word synonym-based\nsubstitution. Experimental results on two VQA datasets with five validated\nmodels demonstrate the effectiveness of the proposed VQAttack in the\ntransferable attack setting, compared with state-of-the-art baselines. This\nwork reveals a significant blind spot in the ``pre-training & fine-tuning''\nparadigm on VQA tasks. Source codes will be released.\n", "date": "2024-02-16", "categories": []}
{"id": "2402.11120", "title": "DART: A Principled Approach to Adversarially Robust Unsupervised Domain\n  Adaptation", "abstract": "  Distribution shifts and adversarial examples are two major challenges for\ndeploying machine learning models. While these challenges have been studied\nindividually, their combination is an important topic that remains relatively\nunder-explored. In this work, we study the problem of adversarial robustness\nunder a common setting of distribution shift - unsupervised domain adaptation\n(UDA). Specifically, given a labeled source domain $D_S$ and an unlabeled\ntarget domain $D_T$ with related but different distributions, the goal is to\nobtain an adversarially robust model for $D_T$. The absence of target domain\nlabels poses a unique challenge, as conventional adversarial robustness\ndefenses cannot be directly applied to $D_T$. To address this challenge, we\nfirst establish a generalization bound for the adversarial target loss, which\nconsists of (i) terms related to the loss on the data, and (ii) a measure of\nworst-case domain divergence. Motivated by this bound, we develop a novel\nunified defense framework called Divergence Aware adveRsarial Training (DART),\nwhich can be used in conjunction with a variety of standard UDA methods; e.g.,\nDANN [Ganin and Lempitsky, 2015]. DART is applicable to general threat models,\nincluding the popular $\\ell_p$-norm model, and does not require heuristic\nregularizers or architectural changes. We also release DomainRobust: a testbed\nfor evaluating robustness of UDA models to adversarial attacks. DomainRobust\nconsists of 4 multi-domain benchmark datasets (with 46 source-target pairs) and\n7 meta-algorithms with a total of 11 variants. Our large-scale experiments\ndemonstrate that on average, DART significantly enhances model robustness on\nall benchmarks compared to the state of the art, while maintaining competitive\nstandard accuracy. The relative improvement in robustness from DART reaches up\nto 29.2% on the source-target domain pairs considered.\n", "date": "2024-02-16", "categories": []}
{"id": "2402.11733", "title": "The Effectiveness of Random Forgetting for Robust Generalization", "abstract": "  Deep neural networks are susceptible to adversarial attacks, which can\ncompromise their performance and accuracy. Adversarial Training (AT) has\nemerged as a popular approach for protecting neural networks against such\nattacks. However, a key challenge of AT is robust overfitting, where the\nnetwork's robust performance on test data deteriorates with further training,\nthus hindering generalization. Motivated by the concept of active forgetting in\nthe brain, we introduce a novel learning paradigm called \"Forget to Mitigate\nOverfitting (FOMO)\". FOMO alternates between the forgetting phase, which\nrandomly forgets a subset of weights and regulates the model's information\nthrough weight reinitialization, and the relearning phase, which emphasizes\nlearning generalizable features. Our experiments on benchmark datasets and\nadversarial attacks show that FOMO alleviates robust overfitting by\nsignificantly reducing the gap between the best and last robust test accuracy\nwhile improving the state-of-the-art robustness. Furthermore, FOMO provides a\nbetter trade-off between standard and robust accuracy, outperforming baseline\nadversarial methods. Finally, our framework is robust to AutoAttacks and\nincreases generalization in many real-world scenarios.\n", "date": "2024-02-18", "categories": []}
{"id": "2402.12100", "title": "Groot: Adversarial Testing for Generative Text-to-Image Models with\n  Tree-based Semantic Transformation", "abstract": "  With the prevalence of text-to-image generative models, their safety becomes\na critical concern. adversarial testing techniques have been developed to probe\nwhether such models can be prompted to produce Not-Safe-For-Work (NSFW)\ncontent. However, existing solutions face several challenges, including low\nsuccess rate and inefficiency. We introduce Groot, the first automated\nframework leveraging tree-based semantic transformation for adversarial testing\nof text-to-image models. Groot employs semantic decomposition and sensitive\nelement drowning strategies in conjunction with LLMs to systematically refine\nadversarial prompts. Our comprehensive evaluation confirms the efficacy of\nGroot, which not only exceeds the performance of current state-of-the-art\napproaches but also achieves a remarkable success rate (93.66%) on leading\ntext-to-image models such as DALL-E 3 and Midjourney.\n", "date": "2024-02-19", "categories": []}
{"id": "2402.12187", "title": "Adversarial Feature Alignment: Balancing Robustness and Accuracy in Deep\n  Learning via Adversarial Training", "abstract": "  Deep learning models continue to advance in accuracy, yet they remain\nvulnerable to adversarial attacks, which often lead to the misclassification of\nadversarial examples. Adversarial training is used to mitigate this problem by\nincreasing robustness against these attacks. However, this approach typically\nreduces a model's standard accuracy on clean, non-adversarial samples. The\nnecessity for deep learning models to balance both robustness and accuracy for\nsecurity is obvious, but achieving this balance remains challenging, and the\nunderlying reasons are yet to be clarified. This paper proposes a novel\nadversarial training method called Adversarial Feature Alignment (AFA), to\naddress these problems. Our research unveils an intriguing insight:\nmisalignment within the feature space often leads to misclassification,\nregardless of whether the samples are benign or adversarial. AFA mitigates this\nrisk by employing a novel optimization algorithm based on contrastive learning\nto alleviate potential feature misalignment. Through our evaluations, we\ndemonstrate the superior performance of AFA. The baseline AFA delivers higher\nrobust accuracy than previous adversarial contrastive learning methods while\nminimizing the drop in clean accuracy to 1.86% and 8.91% on CIFAR10 and\nCIFAR100, respectively, in comparison to cross-entropy. We also show that joint\noptimization of AFA and TRADES, accompanied by data augmentation using a recent\ndiffusion model, achieves state-of-the-art accuracy and robustness.\n", "date": "2024-02-19", "categories": []}
{"id": "2402.00418", "title": "Benchmarking Transferable Adversarial Attacks", "abstract": "  The robustness of deep learning models against adversarial attacks remains a\npivotal concern. This study presents, for the first time, an exhaustive review\nof the transferability aspect of adversarial attacks. It systematically\ncategorizes and critically evaluates various methodologies developed to augment\nthe transferability of adversarial attacks. This study encompasses a spectrum\nof techniques, including Generative Structure, Semantic Similarity, Gradient\nEditing, Target Modification, and Ensemble Approach. Concurrently, this paper\nintroduces a benchmark framework \\textit{TAA-Bench}, integrating ten leading\nmethodologies for adversarial attack transferability, thereby providing a\nstandardized and systematic platform for comparative analysis across diverse\nmodel architectures. Through comprehensive scrutiny, we delineate the efficacy\nand constraints of each method, shedding light on their underlying operational\nprinciples and practical utility. This review endeavors to be a quintessential\nresource for both scholars and practitioners in the field, charting the complex\nterrain of adversarial transferability and setting a foundation for future\nexplorations in this vital sector. The associated codebase is accessible at:\nhttps://github.com/KxPlaug/TAA-Bench\n", "date": "2024-02-01", "categories": []}
{"id": "2402.10283", "title": "Backdoor Attack against One-Class Sequential Anomaly Detection Models", "abstract": "  Deep anomaly detection on sequential data has garnered significant attention\ndue to the wide application scenarios. However, deep learning-based models face\na critical security threat - their vulnerability to backdoor attacks. In this\npaper, we explore compromising deep sequential anomaly detection models by\nproposing a novel backdoor attack strategy. The attack approach comprises two\nprimary steps, trigger generation and backdoor injection. Trigger generation is\nto derive imperceptible triggers by crafting perturbed samples from the benign\nnormal data, of which the perturbed samples are still normal. The backdoor\ninjection is to properly inject the backdoor triggers to comprise the model\nonly for the samples with triggers. The experimental results demonstrate the\neffectiveness of our proposed attack strategy by injecting backdoors on two\nwell-established one-class anomaly detection models.\n", "date": "2024-02-15", "categories": []}
{"id": "2402.10470", "title": "Theoretical Understanding of Learning from Adversarial Perturbations", "abstract": "  It is not fully understood why adversarial examples can deceive neural\nnetworks and transfer between different networks. To elucidate this, several\nstudies have hypothesized that adversarial perturbations, while appearing as\nnoises, contain class features. This is supported by empirical evidence showing\nthat networks trained on mislabeled adversarial examples can still generalize\nwell to correctly labeled test samples. However, a theoretical understanding of\nhow perturbations include class features and contribute to generalization is\nlimited. In this study, we provide a theoretical framework for understanding\nlearning from perturbations using a one-hidden-layer network trained on\nmutually orthogonal samples. Our results highlight that various adversarial\nperturbations, even perturbations of a few pixels, contain sufficient class\nfeatures for generalization. Moreover, we reveal that the decision boundary\nwhen learning from perturbations matches that from standard samples except for\nspecific regions under mild conditions. The code is available at\nhttps://github.com/s-kumano/learning-from-adversarial-perturbations.\n", "date": "2024-02-16", "categories": []}
{"id": "2303.17255", "title": "Fooling the Image Dehazing Models by First Order Gradient", "abstract": "  The research on the single image dehazing task has been widely explored.\nHowever, as far as we know, no comprehensive study has been conducted on the\nrobustness of the well-trained dehazing models. Therefore, there is no evidence\nthat the dehazing networks can resist malicious attacks. In this paper, we\nfocus on designing a group of attack methods based on first order gradient to\nverify the robustness of the existing dehazing algorithms. By analyzing the\ngeneral purpose of image dehazing task, four attack methods are proposed, which\nare predicted dehazed image attack, hazy layer mask attack, haze-free image\nattack and haze-preserved attack. The corresponding experiments are conducted\non six datasets with different scales. Further, the defense strategy based on\nadversarial training is adopted for reducing the negative effects caused by\nmalicious attacks. In summary, this paper defines a new challenging problem for\nthe image dehazing area, which can be called as adversarial attack on dehazing\nnetworks (AADN). Code and Supplementary Material are available at\nhttps://github.com/Xiaofeng-life/AADN Dehazing.\n", "date": "2023-03-30", "categories": []}
{"id": "2310.16955", "title": "Break it, Imitate it, Fix it: Robustness by Generating Human-Like\n  Attacks", "abstract": "  Real-world natural language processing systems need to be robust to human\nadversaries. Collecting examples of human adversaries for training is an\neffective but expensive solution. On the other hand, training on synthetic\nattacks with small perturbations - such as word-substitution - does not\nactually improve robustness to human adversaries. In this paper, we propose an\nadversarial training framework that uses limited human adversarial examples to\ngenerate more useful adversarial examples at scale. We demonstrate the\nadvantages of this system on the ANLI and hate speech detection benchmark\ndatasets - both collected via an iterative, adversarial\nhuman-and-model-in-the-loop procedure. Compared to training only on observed\nhuman attacks, also training on our synthetic adversarial examples improves\nmodel robustness to future rounds. In ANLI, we see accuracy gains on the\ncurrent set of attacks (44.1%$\\,\\to\\,$50.1%) and on two future unseen rounds of\nhuman generated attacks (32.5%$\\,\\to\\,$43.4%, and 29.4%$\\,\\to\\,$40.2%). In hate\nspeech detection, we see AUC gains on current attacks (0.76 $\\to$ 0.84) and a\nfuture round (0.77 $\\to$ 0.79). Attacks from methods that do not learn the\ndistribution of existing human adversaries, meanwhile, degrade robustness.\n", "date": "2023-10-25", "categories": []}
{"id": "2402.09671", "title": "Exploiting Alpha Transparency In Language And Vision-Based AI Systems", "abstract": "  This investigation reveals a novel exploit derived from PNG image file\nformats, specifically their alpha transparency layer, and its potential to fool\nmultiple AI vision systems. Our method uses this alpha layer as a clandestine\nchannel invisible to human observers but fully actionable by AI image\nprocessors. The scope tested for the vulnerability spans representative vision\nsystems from Apple, Microsoft, Google, Salesforce, Nvidia, and Facebook,\nhighlighting the attack's potential breadth. This vulnerability challenges the\nsecurity protocols of existing and fielded vision systems, from medical imaging\nto autonomous driving technologies. Our experiments demonstrate that the\naffected systems, which rely on convolutional neural networks or the latest\nmultimodal language models, cannot quickly mitigate these vulnerabilities\nthrough simple patches or updates. Instead, they require retraining and\narchitectural changes, indicating a persistent hole in multimodal technologies\nwithout some future adversarial hardening against such vision-language\nexploits.\n", "date": "2024-02-15", "categories": []}
{"id": "2402.09674", "title": "PAL: Proxy-Guided Black-Box Attack on Large Language Models", "abstract": "  Large Language Models (LLMs) have surged in popularity in recent months, but\nthey have demonstrated concerning capabilities to generate harmful content when\nmanipulated. While techniques like safety fine-tuning aim to minimize harmful\nuse, recent works have shown that LLMs remain vulnerable to attacks that elicit\ntoxic responses. In this work, we introduce the Proxy-Guided Attack on LLMs\n(PAL), the first optimization-based attack on LLMs in a black-box query-only\nsetting. In particular, it relies on a surrogate model to guide the\noptimization and a sophisticated loss designed for real-world LLM APIs. Our\nattack achieves 84% attack success rate (ASR) on GPT-3.5-Turbo and 48% on\nLlama-2-7B, compared to 4% for the current state of the art. We also propose\nGCG++, an improvement to the GCG attack that reaches 94% ASR on white-box\nLlama-2-7B, and the Random-Search Attack on LLMs (RAL), a strong but simple\nbaseline for query-based attacks. We believe the techniques proposed in this\nwork will enable more comprehensive safety testing of LLMs and, in the long\nterm, the development of better security guardrails. The code can be found at\nhttps://github.com/chawins/pal.\n", "date": "2024-02-15", "categories": []}
{"id": "2402.09874", "title": "Camouflage is all you need: Evaluating and Enhancing Language Model\n  Robustness Against Camouflage Adversarial Attacks", "abstract": "  Adversarial attacks represent a substantial challenge in Natural Language\nProcessing (NLP). This study undertakes a systematic exploration of this\nchallenge in two distinct phases: vulnerability evaluation and resilience\nenhancement of Transformer-based models under adversarial attacks.\n  In the evaluation phase, we assess the susceptibility of three Transformer\nconfigurations, encoder-decoder, encoder-only, and decoder-only setups, to\nadversarial attacks of escalating complexity across datasets containing\noffensive language and misinformation. Encoder-only models manifest a 14% and\n21% performance drop in offensive language detection and misinformation\ndetection tasks, respectively. Decoder-only models register a 16% decrease in\nboth tasks, while encoder-decoder models exhibit a maximum performance drop of\n14% and 26% in the respective tasks.\n  The resilience-enhancement phase employs adversarial training, integrating\npre-camouflaged and dynamically altered data. This approach effectively reduces\nthe performance drop in encoder-only models to an average of 5% in offensive\nlanguage detection and 2% in misinformation detection tasks. Decoder-only\nmodels, occasionally exceeding original performance, limit the performance drop\nto 7% and 2% in the respective tasks. Although not surpassing the original\nperformance, Encoder-decoder models can reduce the drop to an average of 6% and\n2% respectively.\n  Results suggest a trade-off between performance and robustness, with some\nmodels maintaining similar performance while gaining robustness. Our study and\nadversarial training techniques have been incorporated into an open-source tool\nfor generating camouflaged datasets. However, methodology effectiveness depends\non the specific camouflage technique and data encountered, emphasizing the need\nfor continued exploration.\n", "date": "2024-02-15", "categories": []}
{"id": "2402.10196", "title": "A Trembling House of Cards? Mapping Adversarial Attacks against Language\n  Agents", "abstract": "  Language agents powered by large language models (LLMs) have seen exploding\ndevelopment. Their capability of using language as a vehicle for thought and\ncommunication lends an incredible level of flexibility and versatility. People\nhave quickly capitalized on this capability to connect LLMs to a wide range of\nexternal components and environments: databases, tools, the Internet, robotic\nembodiment, etc. Many believe an unprecedentedly powerful automation technology\nis emerging. However, new automation technologies come with new safety risks,\nespecially for intricate systems like language agents. There is a surprisingly\nlarge gap between the speed and scale of their development and deployment and\nour understanding of their safety risks. Are we building a house of cards? In\nthis position paper, we present the first systematic effort in mapping\nadversarial attacks against language agents. We first present a unified\nconceptual framework for agents with three major components: Perception, Brain,\nand Action. Under this framework, we present a comprehensive discussion and\npropose 12 potential attack scenarios against different components of an agent,\ncovering different attack strategies (e.g., input manipulation, adversarial\ndemonstrations, jailbreaking, backdoors). We also draw connections to\nsuccessful attack strategies previously applied to LLMs. We emphasize the\nurgency to gain a thorough understanding of language agent risks before their\nwidespread deployment.\n", "date": "2024-02-15", "categories": []}
{"id": "2306.02895", "title": "Evading Black-box Classifiers Without Breaking Eggs", "abstract": "  Decision-based evasion attacks repeatedly query a black-box classifier to\ngenerate adversarial examples. Prior work measures the cost of such attacks by\nthe total number of queries made to the classifier. We argue this metric is\nflawed. Most security-critical machine learning systems aim to weed out \"bad\"\ndata (e.g., malware, harmful content, etc). Queries to such systems carry a\nfundamentally asymmetric cost: queries detected as \"bad\" come at a higher cost\nbecause they trigger additional security filters, e.g., usage throttling or\naccount suspension. Yet, we find that existing decision-based attacks issue a\nlarge number of \"bad\" queries, which likely renders them ineffective against\nsecurity-critical systems. We then design new attacks that reduce the number of\nbad queries by $1.5$-$7.3\\times$, but often at a significant increase in total\n(non-bad) queries. We thus pose it as an open problem to build black-box\nattacks that are more effective under realistic cost metrics.\n", "date": "2023-06-05", "categories": []}
{"id": "2310.00076", "title": "Robustness of AI-Image Detectors: Fundamental Limits and Practical\n  Attacks", "abstract": "  In light of recent advancements in generative AI models, it has become\nessential to distinguish genuine content from AI-generated one to prevent the\nmalicious usage of fake materials as authentic ones and vice versa. Various\ntechniques have been introduced for identifying AI-generated images, with\nwatermarking emerging as a promising approach. In this paper, we analyze the\nrobustness of various AI-image detectors including watermarking and\nclassifier-based deepfake detectors. For watermarking methods that introduce\nsubtle image perturbations (i.e., low perturbation budget methods), we reveal a\nfundamental trade-off between the evasion error rate (i.e., the fraction of\nwatermarked images detected as non-watermarked ones) and the spoofing error\nrate (i.e., the fraction of non-watermarked images detected as watermarked\nones) upon an application of diffusion purification attack. To validate our\ntheoretical findings, we also provide empirical evidence demonstrating that\ndiffusion purification effectively removes low perturbation budget watermarks\nby applying minimal changes to images. The diffusion purification attack is\nineffective for high perturbation watermarking methods where notable changes\nare applied to images. In this case, we develop a model substitution\nadversarial attack that can successfully remove watermarks. Moreover, we show\nthat watermarking methods are vulnerable to spoofing attacks where the attacker\naims to have real images identified as watermarked ones, damaging the\nreputation of the developers. In particular, with black-box access to the\nwatermarking method, a watermarked noise image can be generated and added to\nreal images, causing them to be incorrectly classified as watermarked. Finally,\nwe extend our theory to characterize a fundamental trade-off between the\nrobustness and reliability of classifier-based deep fake detectors and\ndemonstrate it through experiments.\n", "date": "2023-09-29", "categories": []}
{"id": "2310.17534", "title": "SoK: Pitfalls in Evaluating Black-Box Attacks", "abstract": "  Numerous works study black-box attacks on image classifiers. However, these\nworks make different assumptions on the adversary's knowledge and current\nliterature lacks a cohesive organization centered around the threat model. To\nsystematize knowledge in this area, we propose a taxonomy over the threat space\nspanning the axes of feedback granularity, the access of interactive queries,\nand the quality and quantity of the auxiliary data available to the attacker.\nOur new taxonomy provides three key insights. 1) Despite extensive literature,\nnumerous under-explored threat spaces exist, which cannot be trivially solved\nby adapting techniques from well-explored settings. We demonstrate this by\nestablishing a new state-of-the-art in the less-studied setting of access to\ntop-k confidence scores by adapting techniques from well-explored settings of\naccessing the complete confidence vector, but show how it still falls short of\nthe more restrictive setting that only obtains the prediction label,\nhighlighting the need for more research. 2) Identification the threat model of\ndifferent attacks uncovers stronger baselines that challenge prior\nstate-of-the-art claims. We demonstrate this by enhancing an initially weaker\nbaseline (under interactive query access) via surrogate models, effectively\noverturning claims in the respective paper. 3) Our taxonomy reveals\ninteractions between attacker knowledge that connect well to related areas,\nsuch as model inversion and extraction attacks. We discuss how advances in\nother areas can enable potentially stronger black-box attacks. Finally, we\nemphasize the need for a more realistic assessment of attack success by\nfactoring in local attack runtime. This approach reveals the potential for\ncertain attacks to achieve notably higher success rates and the need to\nevaluate attacks in diverse and harder settings, highlighting the need for\nbetter selection criteria.\n", "date": "2023-10-26", "categories": []}
{"id": "2311.09790", "title": "Breaking Boundaries: Balancing Performance and Robustness in Deep\n  Wireless Traffic Forecasting", "abstract": "  Balancing the trade-off between accuracy and robustness is a long-standing\nchallenge in time series forecasting. While most of existing robust algorithms\nhave achieved certain suboptimal performance on clean data, sustaining the same\nperformance level in the presence of data perturbations remains extremely hard.\nIn this paper, we study a wide array of perturbation scenarios and propose\nnovel defense mechanisms against adversarial attacks using real-world telecom\ndata. We compare our strategy against two existing adversarial training\nalgorithms under a range of maximal allowed perturbations, defined using\n$\\ell_{\\infty}$-norm, $\\in [0.1,0.4]$. Our findings reveal that our hybrid\nstrategy, which is composed of a classifier to detect adversarial examples, a\ndenoiser to eliminate noise from the perturbed data samples, and a standard\nforecaster, achieves the best performance on both clean and perturbed data. Our\noptimal model can retain up to $92.02\\%$ the performance of the original\nforecasting model in terms of Mean Squared Error (MSE) on clean data, while\nbeing more robust than the standard adversarially trained models on perturbed\ndata. Its MSE is 2.71$\\times$ and 2.51$\\times$ lower than those of comparing\nmethods on normal and perturbed data, respectively. In addition, the components\nof our models can be trained in parallel, resulting in better computational\nefficiency. Our results indicate that we can optimally balance the trade-off\nbetween the performance and robustness of forecasting models by improving the\nclassifier and denoiser, even in the presence of sophisticated and destructive\npoisoning attacks.\n", "date": "2023-11-16", "categories": []}
{"id": "2402.08695", "title": "Game of Trojans: Adaptive Adversaries Against Output-based\n  Trojaned-Model Detectors", "abstract": "  We propose and analyze an adaptive adversary that can retrain a Trojaned DNN\nand is also aware of SOTA output-based Trojaned model detectors. We show that\nsuch an adversary can ensure (1) high accuracy on both trigger-embedded and\nclean samples and (2) bypass detection. Our approach is based on an observation\nthat the high dimensionality of the DNN parameters provides sufficient degrees\nof freedom to simultaneously achieve these objectives. We also enable SOTA\ndetectors to be adaptive by allowing retraining to recalibrate their\nparameters, thus modeling a co-evolution of parameters of a Trojaned model and\ndetectors. We then show that this co-evolution can be modeled as an iterative\ngame, and prove that the resulting (optimal) solution of this interactive game\nleads to the adversary successfully achieving the above objectives. In\naddition, we provide a greedy algorithm for the adversary to select a minimum\nnumber of input samples for embedding triggers. We show that for cross-entropy\nor log-likelihood loss functions used by the DNNs, the greedy algorithm\nprovides provable guarantees on the needed number of trigger-embedded input\nsamples. Extensive experiments on four diverse datasets -- MNIST, CIFAR-10,\nCIFAR-100, and SpeechCommand -- reveal that the adversary effectively evades\nfour SOTA output-based Trojaned model detectors: MNTD, NeuralCleanse, STRIP,\nand TABOR.\n", "date": "2024-02-12", "categories": []}
{"id": "2402.08763", "title": "Enhancing Robustness of Indoor Robotic Navigation with Free-Space\n  Segmentation Models Against Adversarial Attacks", "abstract": "  Endeavors in indoor robotic navigation rely on the accuracy of segmentation\nmodels to identify free space in RGB images. However, deep learning models are\nvulnerable to adversarial attacks, posing a significant challenge to their\nreal-world deployment. In this study, we identify vulnerabilities within the\nhidden layers of neural networks and introduce a practical approach to\nreinforce traditional adversarial training. Our method incorporates a novel\ndistance loss function, minimizing the gap between hidden layers in clean and\nadversarial images. Experiments demonstrate satisfactory performance in\nimproving the model's robustness against adversarial perturbations.\n", "date": "2024-02-13", "categories": []}
{"id": "2402.09316", "title": "Only My Model On My Data: A Privacy Preserving Approach Protecting one\n  Model and Deceiving Unauthorized Black-Box Models", "abstract": "  Deep neural networks are extensively applied to real-world tasks, such as\nface recognition and medical image classification, where privacy and data\nprotection are critical. Image data, if not protected, can be exploited to\ninfer personal or contextual information. Existing privacy preservation\nmethods, like encryption, generate perturbed images that are unrecognizable to\neven humans. Adversarial attack approaches prohibit automated inference even\nfor authorized stakeholders, limiting practical incentives for commercial and\nwidespread adaptation. This pioneering study tackles an unexplored practical\nprivacy preservation use case by generating human-perceivable images that\nmaintain accurate inference by an authorized model while evading other\nunauthorized black-box models of similar or dissimilar objectives, and\naddresses the previous research gaps. The datasets employed are ImageNet, for\nimage classification, Celeba-HQ dataset, for identity classification, and\nAffectNet, for emotion classification. Our results show that the generated\nimages can successfully maintain the accuracy of a protected model and degrade\nthe average accuracy of the unauthorized black-box models to 11.97%, 6.63%, and\n55.51% on ImageNet, Celeba-HQ, and AffectNet datasets, respectively.\n", "date": "2024-02-14", "categories": []}
{"id": "2303.11793", "title": "Bridging Optimal Transport and Jacobian Regularization by Optimal\n  Trajectory for Enhanced Adversarial Defense", "abstract": "  Deep neural networks, particularly in vision tasks, are notably susceptible\nto adversarial perturbations. To overcome this challenge, developing a robust\nclassifier is crucial. In light of the recent advancements in the robustness of\nclassifiers, we delve deep into the intricacies of adversarial training and\nJacobian regularization, two pivotal defenses. Our work is the first carefully\nanalyzes and characterizes these two schools of approaches, both theoretically\nand empirically, to demonstrate how each approach impacts the robust learning\nof a classifier. Next, we propose our novel Optimal Transport with Jacobian\nregularization method, dubbed OTJR, bridging the input Jacobian regularization\nwith the a output representation alignment by leveraging the optimal transport\ntheory. In particular, we employ the Sliced Wasserstein distance that can\nefficiently push the adversarial samples' representations closer to those of\nclean samples, regardless of the number of classes within the dataset. The SW\ndistance provides the adversarial samples' movement directions, which are much\nmore informative and powerful for the Jacobian regularization. Our empirical\nevaluations set a new standard in the domain, with our method achieving\ncommendable accuracies of 52.57% on CIFAR-10 and 28.3% on CIFAR-100 datasets\nunder the AutoAttack. Further validating our model's practicality, we conducted\nreal-world tests by subjecting internet-sourced images to online adversarial\nattacks. These demonstrations highlight our model's capability to counteract\nsophisticated adversarial perturbations, affirming its significance and\napplicability in real-world scenarios.\n", "date": "2023-03-21", "categories": []}
{"id": "2312.09533", "title": "Adversarial Robustness on Image Classification with $k$-means", "abstract": "  In this paper we explore the challenges and strategies for enhancing the\nrobustness of $k$-means clustering algorithms against adversarial\nmanipulations. We evaluate the vulnerability of clustering algorithms to\nadversarial attacks, emphasising the associated security risks. Our study\ninvestigates the impact of incremental attack strength on training, introduces\nthe concept of transferability between supervised and unsupervised models, and\nhighlights the sensitivity of unsupervised models to sample distributions. We\nadditionally introduce and evaluate an adversarial training method that\nimproves testing performance in adversarial scenarios, and we highlight the\nimportance of various parameters in the proposed training method, such as\ncontinuous learning, centroid initialisation, and adversarial step-count.\n", "date": "2023-12-15", "categories": []}
{"id": "2402.07480", "title": "Topological safeguard for evasion attack interpreting the neural\n  networks' behavior", "abstract": "  In the last years, Deep Learning technology has been proposed in different\nfields, bringing many advances in each of them, but identifying new threats in\nthese solutions regarding cybersecurity. Those implemented models have brought\nseveral vulnerabilities associated with Deep Learning technology. Moreover,\nthose allow taking advantage of the implemented model, obtaining private\ninformation, and even modifying the model's decision-making. Therefore,\ninterest in studying those vulnerabilities/attacks and designing defenses to\navoid or fight them is gaining prominence among researchers. In particular, the\nwidely known evasion attack is being analyzed by researchers; thus, several\ndefenses to avoid such a threat can be found in the literature. Since the\npresentation of the L-BFG algorithm, this threat concerns the research\ncommunity. However, it continues developing new and ingenious countermeasures\nsince there is no perfect defense for all the known evasion algorithms. In this\nwork, a novel detector of evasion attacks is developed. It focuses on the\ninformation of the activations of the neurons given by the model when an input\nsample is injected. Moreover, it puts attention to the topology of the targeted\ndeep learning model to analyze the activations according to which neurons are\nconnecting. This approach has been decided because the literature shows that\nthe targeted model's topology contains essential information about if the\nevasion attack occurs. For this purpose, a huge data preprocessing is required\nto introduce all this information in the detector, which uses the Graph\nConvolutional Neural Network (GCN) technology. Thus, it understands the\ntopology of the target model, obtaining promising results and improving the\noutcomes presented in the literature related to similar defenses.\n", "date": "2024-02-12", "categories": []}
{"id": "2402.08577", "title": "Test-Time Backdoor Attacks on Multimodal Large Language Models", "abstract": "  Backdoor attacks are commonly executed by contaminating training data, such\nthat a trigger can activate predetermined harmful effects during the test\nphase. In this work, we present AnyDoor, a test-time backdoor attack against\nmultimodal large language models (MLLMs), which involves injecting the backdoor\ninto the textual modality using adversarial test images (sharing the same\nuniversal perturbation), without requiring access to or modification of the\ntraining data. AnyDoor employs similar techniques used in universal adversarial\nattacks, but distinguishes itself by its ability to decouple the timing of\nsetup and activation of harmful effects. In our experiments, we validate the\neffectiveness of AnyDoor against popular MLLMs such as LLaVA-1.5, MiniGPT-4,\nInstructBLIP, and BLIP-2, as well as provide comprehensive ablation studies.\nNotably, because the backdoor is injected by a universal perturbation, AnyDoor\ncan dynamically change its backdoor trigger prompts/harmful effects, exposing a\nnew challenge for defending against backdoor attacks. Our project page is\navailable at https://sail-sg.github.io/AnyDoor/.\n", "date": "2024-02-13", "categories": []}
{"id": "2402.08586", "title": "Faster Repeated Evasion Attacks in Tree Ensembles", "abstract": "  Tree ensembles are one of the most widely used model classes. However, these\nmodels are susceptible to adversarial examples, i.e., slightly perturbed\nexamples that elicit a misprediction. There has been significant research on\ndesigning approaches to construct such examples for tree ensembles. But this is\na computationally challenging problem that often must be solved a large number\nof times (e.g., for all examples in a training set). This is compounded by the\nfact that current approaches attempt to find such examples from scratch. In\ncontrast, we exploit the fact that multiple similar problems are being solved.\nSpecifically, our approach exploits the insight that adversarial examples for\ntree ensembles tend to perturb a consistent but relatively small set of\nfeatures. We show that we can quickly identify this set of features and use\nthis knowledge to speedup constructing adversarial examples.\n", "date": "2024-02-13", "categories": []}
{"id": "2306.14043", "title": "Machine Learning needs Better Randomness Standards: Randomised Smoothing\n  and PRNG-based attacks", "abstract": "  Randomness supports many critical functions in the field of machine learning\n(ML) including optimisation, data selection, privacy, and security. ML systems\noutsource the task of generating or harvesting randomness to the compiler, the\ncloud service provider or elsewhere in the toolchain. Yet there is a long\nhistory of attackers exploiting poor randomness, or even creating it -- as when\nthe NSA put backdoors in random number generators to break cryptography. In\nthis paper we consider whether attackers can compromise an ML system using only\nthe randomness on which they commonly rely. We focus our effort on Randomised\nSmoothing, a popular approach to train certifiably robust models, and to\ncertify specific input datapoints of an arbitrary model. We choose Randomised\nSmoothing since it is used for both security and safety -- to counteract\nadversarial examples and quantify uncertainty respectively. Under the hood, it\nrelies on sampling Gaussian noise to explore the volume around a data point to\ncertify that a model is not vulnerable to adversarial examples. We demonstrate\nan entirely novel attack, where an attacker backdoors the supplied randomness\nto falsely certify either an overestimate or an underestimate of robustness for\nup to 81 times. We demonstrate that such attacks are possible, that they\nrequire very small changes to randomness to succeed, and that they are hard to\ndetect. As an example, we hide an attack in the random number generator and\nshow that the randomness tests suggested by NIST fail to detect it. We advocate\nupdating the NIST guidelines on random number testing to make them more\nappropriate for safety-critical and security-critical machine-learning\napplications.\n", "date": "2023-06-24", "categories": []}
{"id": "2308.10743", "title": "Enhancing Adversarial Attacks: The Similar Target Method", "abstract": "  Deep neural networks are vulnerable to adversarial examples, posing a threat\nto the models' applications and raising security concerns. An intriguing\nproperty of adversarial examples is their strong transferability. Several\nmethods have been proposed to enhance transferability, including ensemble\nattacks which have demonstrated their efficacy. However, prior approaches\nsimply average logits, probabilities, or losses for model ensembling, lacking a\ncomprehensive analysis of how and why model ensembling significantly improves\ntransferability. In this paper, we propose a similar targeted attack method\nnamed Similar Target~(ST). By promoting cosine similarity between the gradients\nof each model, our method regularizes the optimization direction to\nsimultaneously attack all surrogate models. This strategy has been proven to\nenhance generalization ability. Experimental results on ImageNet validate the\neffectiveness of our approach in improving adversarial transferability. Our\nmethod outperforms state-of-the-art attackers on 18 discriminative classifiers\nand adversarially trained models.\n", "date": "2023-08-21", "categories": []}
{"id": "2311.16526", "title": "On robust overfitting: adversarial training induced distribution matters", "abstract": "  Adversarial training may be regarded as standard training with a modified\nloss function. But its generalization error appears much larger than standard\ntraining under standard loss. This phenomenon, known as robust overfitting, has\nattracted significant research attention and remains largely as a mystery. In\nthis paper, we first show empirically that robust overfitting correlates with\nthe increasing generalization difficulty of the perturbation-induced\ndistributions along the trajectory of adversarial training (specifically\nPGD-based adversarial training). We then provide a novel upper bound for\ngeneralization error with respect to the perturbation-induced distributions, in\nwhich a notion of the perturbation operator, referred to \"local dispersion\",\nplays an important role. Experimental results are presented to validate the\nusefulness of the bound and various additional insights are provided.\n", "date": "2023-11-28", "categories": []}
{"id": "2402.06655", "title": "Adversarial Text Purification: A Large Language Model Approach for\n  Defense", "abstract": "  Adversarial purification is a defense mechanism for safeguarding classifiers\nagainst adversarial attacks without knowing the type of attacks or training of\nthe classifier. These techniques characterize and eliminate adversarial\nperturbations from the attacked inputs, aiming to restore purified samples that\nretain similarity to the initially attacked ones and are correctly classified\nby the classifier. Due to the inherent challenges associated with\ncharacterizing noise perturbations for discrete inputs, adversarial text\npurification has been relatively unexplored. In this paper, we investigate the\neffectiveness of adversarial purification methods in defending text\nclassifiers. We propose a novel adversarial text purification that harnesses\nthe generative capabilities of Large Language Models (LLMs) to purify\nadversarial text without the need to explicitly characterize the discrete noise\nperturbations. We utilize prompt engineering to exploit LLMs for recovering the\npurified examples for given adversarial examples such that they are\nsemantically similar and correctly classified. Our proposed method demonstrates\nremarkable performance over various classifiers, improving their accuracy under\nthe attack by over 65% on average.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.06957", "title": "Architectural Neural Backdoors from First Principles", "abstract": "  While previous research backdoored neural networks by changing their\nparameters, recent work uncovered a more insidious threat: backdoors embedded\nwithin the definition of the network's architecture. This involves injecting\ncommon architectural components, such as activation functions and pooling\nlayers, to subtly introduce a backdoor behavior that persists even after (full\nre-)training. However, the full scope and implications of architectural\nbackdoors have remained largely unexplored. Bober-Irizar et al. [2023]\nintroduced the first architectural backdoor; they showed how to create a\nbackdoor for a checkerboard pattern, but never explained how to target an\narbitrary trigger pattern of choice. In this work we construct an arbitrary\ntrigger detector which can be used to backdoor an architecture with no human\nsupervision. This leads us to revisit the concept of architecture backdoors and\ntaxonomise them, describing 12 distinct types. To gauge the difficulty of\ndetecting such backdoors, we conducted a user study, revealing that ML\ndevelopers can only identify suspicious components in common model definitions\nas backdoors in 37% of cases, while they surprisingly preferred backdoored\nmodels in 33% of cases. To contextualize these results, we find that language\nmodels outperform humans at the detection of backdoors. Finally, we discuss\ndefenses against architectural backdoors, emphasizing the need for robust and\ncomprehensive strategies to safeguard the integrity of ML systems.\n", "date": "2024-02-10", "categories": []}
{"id": "2402.07183", "title": "A Random Ensemble of Encrypted Vision Transformers for Adversarially\n  Robust Defense", "abstract": "  Deep neural networks (DNNs) are well known to be vulnerable to adversarial\nexamples (AEs). In previous studies, the use of models encrypted with a secret\nkey was demonstrated to be robust against white-box attacks, but not against\nblack-box ones. In this paper, we propose a novel method using the vision\ntransformer (ViT) that is a random ensemble of encrypted models for enhancing\nrobustness against both white-box and black-box attacks. In addition, a\nbenchmark attack method, called AutoAttack, is applied to models to test\nadversarial robustness objectively. In experiments, the method was demonstrated\nto be robust against not only white-box attacks but also black-box ones in an\nimage classification task on the CIFAR-10 and ImageNet datasets. The method was\nalso compared with the state-of-the-art in a standardized benchmark for\nadversarial robustness, RobustBench, and it was verified to outperform\nconventional defenses in terms of clean accuracy and robust accuracy.\n", "date": "2024-02-11", "categories": []}
{"id": "2402.07347", "title": "Accuracy of TextFooler black box adversarial attacks on 01 loss sign\n  activation neural network ensemble", "abstract": "  Recent work has shown the defense of 01 loss sign activation neural networks\nagainst image classification adversarial attacks. A public challenge to attack\nthe models on CIFAR10 dataset remains undefeated. We ask the following question\nin this study: are 01 loss sign activation neural networks hard to deceive with\na popular black box text adversarial attack program called TextFooler? We study\nthis question on four popular text classification datasets: IMDB reviews, Yelp\nreviews, MR sentiment classification, and AG news classification. We find that\nour 01 loss sign activation network is much harder to attack with TextFooler\ncompared to sigmoid activation cross entropy and binary neural networks. We\nalso study a 01 loss sign activation convolutional neural network with a novel\nglobal pooling step specific to sign activation networks. With this new\nvariation we see a significant gain in adversarial accuracy rendering\nTextFooler practically useless against it. We make our code freely available at\n\\url{https://github.com/zero-one-loss/wordcnn01} and\n\\url{https://github.com/xyzacademic/mlp01example}. Our work here suggests that\n01 loss sign activation networks could be further developed to create fool\nproof models against text adversarial attacks.\n", "date": "2024-02-12", "categories": []}
{"id": "2402.07496", "title": "Understanding Deep Learning defenses Against Adversarial Examples\n  Through Visualizations for Dynamic Risk Assessment", "abstract": "  In recent years, Deep Neural Network models have been developed in different\nfields, where they have brought many advances. However, they have also started\nto be used in tasks where risk is critical. A misdiagnosis of these models can\nlead to serious accidents or even death. This concern has led to an interest\namong researchers to study possible attacks on these models, discovering a long\nlist of vulnerabilities, from which every model should be defended. The\nadversarial example attack is a widely known attack among researchers, who have\ndeveloped several defenses to avoid such a threat. However, these defenses are\nas opaque as a deep neural network model, how they work is still unknown. This\nis why visualizing how they change the behavior of the target model is\ninteresting in order to understand more precisely how the performance of the\ndefended model is being modified. For this work, some defenses, against\nadversarial example attack, have been selected in order to visualize the\nbehavior modification of each of them in the defended model. Adversarial\ntraining, dimensionality reduction and prediction similarity were the selected\ndefenses, which have been developed using a model composed by convolution\nneural network layers and dense neural network layers. In each defense, the\nbehavior of the original model has been compared with the behavior of the\ndefended model, representing the target model by a graph in a visualization.\n", "date": "2024-02-12", "categories": []}
{"id": "2303.01734", "title": "AdvART: Adversarial Art for Camouflaged Object Detection Attacks", "abstract": "  Physical adversarial attacks pose a significant practical threat as it\ndeceives deep learning systems operating in the real world by producing\nprominent and maliciously designed physical perturbations. Emphasizing the\nevaluation of naturalness is crucial in such attacks, as humans can readily\ndetect and eliminate unnatural manipulations. To overcome this limitation,\nrecent work has proposed leveraging generative adversarial networks (GANs) to\ngenerate naturalistic patches, which may not catch human's attention. However,\nthese approaches suffer from a limited latent space which leads to an\ninevitable trade-off between naturalness and attack efficiency. In this paper,\nwe propose a novel approach to generate naturalistic and inconspicuous\nadversarial patches. Specifically, we redefine the optimization problem by\nintroducing an additional loss term to the cost function. This term works as a\nsemantic constraint to ensure that the generated camouflage pattern holds\nsemantic meaning rather than arbitrary patterns. The additional term leverages\nsimilarity metrics to construct a similarity loss that we optimize within the\nglobal objective function. Our technique is based on directly manipulating the\npixel values in the patch, which gives higher flexibility and larger space\ncompared to the GAN-based techniques that are based on indirectly optimizing\nthe patch by modifying the latent vector. Our attack achieves superior success\nrate of up to 91.19\\% and 72\\%, respectively, in the digital world and when\ndeployed in smart cameras at the edge compared to the GAN-based technique.\n", "date": "2023-03-03", "categories": []}
{"id": "2402.06249", "title": "Anomaly Unveiled: Securing Image Classification against Adversarial\n  Patch Attacks", "abstract": "  Adversarial patch attacks pose a significant threat to the practical\ndeployment of deep learning systems. However, existing research primarily\nfocuses on image pre-processing defenses, which often result in reduced\nclassification accuracy for clean images and fail to effectively counter\nphysically feasible attacks. In this paper, we investigate the behavior of\nadversarial patches as anomalies within the distribution of image information\nand leverage this insight to develop a robust defense strategy. Our proposed\ndefense mechanism utilizes a clustering-based technique called DBSCAN to\nisolate anomalous image segments, which is carried out by a three-stage\npipeline consisting of Segmenting, Isolating, and Blocking phases to identify\nand mitigate adversarial noise. Upon identifying adversarial components, we\nneutralize them by replacing them with the mean pixel value, surpassing\nalternative replacement options. Our model-agnostic defense mechanism is\nevaluated across multiple models and datasets, demonstrating its effectiveness\nin countering various adversarial patch attacks in image classification tasks.\nOur proposed approach significantly improves accuracy, increasing from 38.8\\%\nwithout the defense to 67.1\\% with the defense against LaVAN and GoogleAp\nattacks, surpassing prominent state-of-the-art methods such as LGS (53.86\\%)\nand Jujutsu (60\\%)\n", "date": "2024-02-09", "categories": []}
{"id": "2402.05526", "title": "Buffer Overflow in Mixture of Experts", "abstract": "  Mixture of Experts (MoE) has become a key ingredient for scaling large\nfoundation models while keeping inference costs steady. We show that expert\nrouting strategies that have cross-batch dependencies are vulnerable to\nattacks. Malicious queries can be sent to a model and can affect a model's\noutput on other benign queries if they are grouped in the same batch. We\ndemonstrate this via a proof-of-concept attack in a toy experimental setting.\n", "date": "2024-02-08", "categories": []}
{"id": "2402.04325", "title": "Enhance DNN Adversarial Robustness and Efficiency via Injecting Noise to\n  Non-Essential Neurons", "abstract": "  Deep Neural Networks (DNNs) have revolutionized a wide range of industries,\nfrom healthcare and finance to automotive, by offering unparalleled\ncapabilities in data analysis and decision-making. Despite their transforming\nimpact, DNNs face two critical challenges: the vulnerability to adversarial\nattacks and the increasing computational costs associated with more complex and\nlarger models. In this paper, we introduce an effective method designed to\nsimultaneously enhance adversarial robustness and execution efficiency. Unlike\nprior studies that enhance robustness via uniformly injecting noise, we\nintroduce a non-uniform noise injection algorithm, strategically applied at\neach DNN layer to disrupt adversarial perturbations introduced in attacks. By\nemploying approximation techniques, our approach identifies and protects\nessential neurons while strategically introducing noise into non-essential\nneurons. Our experimental results demonstrate that our method successfully\nenhances both robustness and efficiency across several attack scenarios, model\narchitectures, and datasets.\n", "date": "2024-02-06", "categories": []}
{"id": "2208.10373", "title": "Reversing Skin Cancer Adversarial Examples by Multiscale Diffusive and\n  Denoising Aggregation Mechanism", "abstract": "  Reliable skin cancer diagnosis models play an essential role in early\nscreening and medical intervention. Prevailing computer-aided skin cancer\nclassification systems employ deep learning approaches. However, recent studies\nreveal their extreme vulnerability to adversarial attacks -- often\nimperceptible perturbations to significantly reduce the performances of skin\ncancer diagnosis models. To mitigate these threats, this work presents a\nsimple, effective, and resource-efficient defense framework by reverse\nengineering adversarial perturbations in skin cancer images. Specifically, a\nmultiscale image pyramid is first established to better preserve discriminative\nstructures in the medical imaging domain. To neutralize adversarial effects,\nskin images at different scales are then progressively diffused by injecting\nisotropic Gaussian noises to move the adversarial examples to the clean image\nmanifold. Crucially, to further reverse adversarial noises and suppress\nredundant injected noises, a novel multiscale denoising mechanism is carefully\ndesigned that aggregates image information from neighboring scales. We\nevaluated the defensive effectiveness of our method on ISIC 2019, a largest\nskin cancer multiclass classification dataset. Experimental results demonstrate\nthat the proposed method can successfully reverse adversarial perturbations\nfrom different attacks and significantly outperform some state-of-the-art\nmethods in defending skin cancer diagnosis models.\n", "date": "2022-08-22", "categories": []}
{"id": "2310.04655", "title": "VLATTACK: Multimodal Adversarial Attacks on Vision-Language Tasks via\n  Pre-trained Models", "abstract": "  Vision-Language (VL) pre-trained models have shown their superiority on many\nmultimodal tasks. However, the adversarial robustness of such models has not\nbeen fully explored. Existing approaches mainly focus on exploring the\nadversarial robustness under the white-box setting, which is unrealistic. In\nthis paper, we aim to investigate a new yet practical task to craft image and\ntext perturbations using pre-trained VL models to attack black-box fine-tuned\nmodels on different downstream tasks. Towards this end, we propose VLATTACK to\ngenerate adversarial samples by fusing perturbations of images and texts from\nboth single-modal and multimodal levels. At the single-modal level, we propose\na new block-wise similarity attack (BSA) strategy to learn image perturbations\nfor disrupting universal representations. Besides, we adopt an existing text\nattack strategy to generate text perturbations independent of the image-modal\nattack. At the multimodal level, we design a novel iterative cross-search\nattack (ICSA) method to update adversarial image-text pairs periodically,\nstarting with the outputs from the single-modal level. We conduct extensive\nexperiments to attack five widely-used VL pre-trained models for six tasks.\nExperimental results show that VLATTACK achieves the highest attack success\nrates on all tasks compared with state-of-the-art baselines, which reveals a\nblind spot in the deployment of pre-trained VL models. Source codes can be\nfound at https://github.com/ericyinyzy/VLAttack.\n", "date": "2023-10-07", "categories": []}
{"id": "2402.02309", "title": "Jailbreaking Attack against Multimodal Large Language Model", "abstract": "  This paper focuses on jailbreaking attacks against multi-modal large language\nmodels (MLLMs), seeking to elicit MLLMs to generate objectionable responses to\nharmful user queries. A maximum likelihood-based algorithm is proposed to find\nan \\emph{image Jailbreaking Prompt} (imgJP), enabling jailbreaks against MLLMs\nacross multiple unseen prompts and images (i.e., data-universal property). Our\napproach exhibits strong model-transferability, as the generated imgJP can be\ntransferred to jailbreak various models, including MiniGPT-v2, LLaVA,\nInstructBLIP, and mPLUG-Owl2, in a black-box manner. Moreover, we reveal a\nconnection between MLLM-jailbreaks and LLM-jailbreaks. As a result, we\nintroduce a construction-based method to harness our approach for\nLLM-jailbreaks, demonstrating greater efficiency than current state-of-the-art\nmethods. The code is available here. \\textbf{Warning: some content generated by\nlanguage models may be offensive to some readers.}\n", "date": "2024-02-04", "categories": []}
{"id": "2402.02732", "title": "A Generative Approach to Surrogate-based Black-box Attacks", "abstract": "  Surrogate-based black-box attacks have exposed the heightened vulnerability\nof DNNs. These attacks are designed to craft adversarial examples for any\nsamples with black-box target feedback for only a given set of samples.\nState-of-the-art surrogate-based attacks involve training a discriminative\nsurrogate that mimics the target's outputs. The goal is to learn the decision\nboundaries of the target. The surrogate is then attacked by white-box attacks\nto craft adversarial examples similar to the original samples but belong to\nother classes. With limited samples, the discriminative surrogate fails to\naccurately learn the target's decision boundaries, and these surrogate-based\nattacks suffer from low success rates. Different from the discriminative\napproach, we propose a generative surrogate that learns the distribution of\nsamples residing on or close to the target's decision boundaries. The\ndistribution learned by the generative surrogate can be used to craft\nadversarial examples that have imperceptible differences from the original\nsamples but belong to other classes. The proposed generative approach results\nin attacks with remarkably high attack success rates on various targets and\ndatasets.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.03477", "title": "Arabic Synonym BERT-based Adversarial Examples for Text Classification", "abstract": "  Text classification systems have been proven vulnerable to adversarial text\nexamples, modified versions of the original text examples that are often\nunnoticed by human eyes, yet can force text classification models to alter\ntheir classification. Often, research works quantifying the impact of\nadversarial text attacks have been applied only to models trained in English.\nIn this paper, we introduce the first word-level study of adversarial attacks\nin Arabic. Specifically, we use a synonym (word-level) attack using a Masked\nLanguage Modeling (MLM) task with a BERT model in a black-box setting to assess\nthe robustness of the state-of-the-art text classification models to\nadversarial attacks in Arabic. To evaluate the grammatical and semantic\nsimilarities of the newly produced adversarial examples using our synonym\nBERT-based attack, we invite four human evaluators to assess and compare the\nproduced adversarial examples with their original examples. We also study the\ntransferability of these newly produced Arabic adversarial examples to various\nmodels and investigate the effectiveness of defense mechanisms against these\nadversarial examples on the BERT models. We find that fine-tuned BERT models\nwere more susceptible to our synonym attacks than the other Deep Neural\nNetworks (DNN) models like WordCNN and WordLSTM we trained. We also find that\nfine-tuned BERT models were more susceptible to transferred attacks. We,\nlastly, find that fine-tuned BERT models successfully regain at least 2% in\naccuracy after applying adversarial training as an initial defense mechanism.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.03576", "title": "Generalization Properties of Adversarial Training for $\\ell_0$-Bounded\n  Adversarial Attacks", "abstract": "  We have widely observed that neural networks are vulnerable to small additive\nperturbations to the input causing misclassification. In this paper, we focus\non the $\\ell_0$-bounded adversarial attacks, and aim to theoretically\ncharacterize the performance of adversarial training for an important class of\ntruncated classifiers. Such classifiers are shown to have strong performance\nempirically, as well as theoretically in the Gaussian mixture model, in the\n$\\ell_0$-adversarial setting. The main contribution of this paper is to prove a\nnovel generalization bound for the binary classification setting with\n$\\ell_0$-bounded adversarial perturbation that is distribution-independent.\nDeriving a generalization bound in this setting has two main challenges: (i)\nthe truncated inner product which is highly non-linear; and (ii) maximization\nover the $\\ell_0$ ball due to adversarial training is non-convex and highly\nnon-smooth. To tackle these challenges, we develop new coding techniques for\nbounding the combinatorial dimension of the truncated hypothesis class.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.03705", "title": "FoolSDEdit: Deceptively Steering Your Edits Towards Targeted\n  Attribute-aware Distribution", "abstract": "  Guided image synthesis methods, like SDEdit based on the diffusion model,\nexcel at creating realistic images from user inputs such as stroke paintings.\nHowever, existing efforts mainly focus on image quality, often overlooking a\nkey point: the diffusion model represents a data distribution, not individual\nimages. This introduces a low but critical chance of generating images that\ncontradict user intentions, raising ethical concerns. For example, a user\ninputting a stroke painting with female characteristics might, with some\nprobability, get male faces from SDEdit. To expose this potential\nvulnerability, we aim to build an adversarial attack forcing SDEdit to generate\na specific data distribution aligned with a specified attribute (e.g., female),\nwithout changing the input's attribute characteristics. We propose the Targeted\nAttribute Generative Attack (TAGA), using an attribute-aware objective function\nand optimizing the adversarial noise added to the input stroke painting.\nEmpirical studies reveal that traditional adversarial noise struggles with\nTAGA, while natural perturbations like exposure and motion blur easily alter\ngenerated images' attributes. To execute effective attacks, we introduce\nFoolSDEdit: We design a joint adversarial exposure and blur attack, adding\nexposure and motion blur to the stroke painting and optimizing them together.\nWe optimize the execution strategy of various perturbations, framing it as a\nnetwork architecture search problem. We create the SuperPert, a graph\nrepresenting diverse execution strategies for different perturbations. After\ntraining, we obtain the optimized execution strategy for effective TAGA against\nSDEdit. Comprehensive experiments on two datasets show our method compelling\nSDEdit to generate a targeted attribute-aware data distribution, significantly\noutperforming baselines.\n", "date": "2024-02-06", "categories": []}
{"id": "2402.03951", "title": "Boosting Adversarial Transferability across Model Genus by\n  Deformation-Constrained Warping", "abstract": "  Adversarial examples generated by a surrogate model typically exhibit limited\ntransferability to unknown target systems. To address this problem, many\ntransferability enhancement approaches (e.g., input transformation and model\naugmentation) have been proposed. However, they show poor performances in\nattacking systems having different model genera from the surrogate model. In\nthis paper, we propose a novel and generic attacking strategy, called\nDeformation-Constrained Warping Attack (DeCoWA), that can be effectively\napplied to cross model genus attack. Specifically, DeCoWA firstly augments\ninput examples via an elastic deformation, namely Deformation-Constrained\nWarping (DeCoW), to obtain rich local details of the augmented input. To avoid\nsevere distortion of global semantics led by random deformation, DeCoW further\nconstrains the strength and direction of the warping transformation by a novel\nadaptive control strategy. Extensive experiments demonstrate that the\ntransferable examples crafted by our DeCoWA on CNN surrogates can significantly\nhinder the performance of Transformers (and vice versa) on various tasks,\nincluding image classification, video action recognition, and audio\nrecognition. Code is made available at https://github.com/LinQinLiang/DeCoWA.\n", "date": "2024-02-06", "categories": []}
{"id": "2402.04010", "title": "Efficient Availability Attacks against Supervised and Contrastive\n  Learning Simultaneously", "abstract": "  Availability attacks can prevent the unauthorized use of private data and\ncommercial datasets by generating imperceptible noise and making unlearnable\nexamples before release. Ideally, the obtained unlearnability prevents\nalgorithms from training usable models. When supervised learning (SL)\nalgorithms have failed, a malicious data collector possibly resorts to\ncontrastive learning (CL) algorithms to bypass the protection. Through\nevaluation, we have found that most of the existing methods are unable to\nachieve both supervised and contrastive unlearnability, which poses risks to\ndata protection. Different from recent methods based on contrastive error\nminimization, we employ contrastive-like data augmentations in supervised error\nminimization or maximization frameworks to obtain attacks effective for both SL\nand CL. Our proposed AUE and AAP attacks achieve state-of-the-art worst-case\nunlearnability across SL and CL algorithms with less computation consumption,\nshowcasing prospects in real-world applications.\n", "date": "2024-02-06", "categories": []}
{"id": "2306.14275", "title": "Enhancing Adversarial Training via Reweighting Optimization Trajectory", "abstract": "  Despite the fact that adversarial training has become the de facto method for\nimproving the robustness of deep neural networks, it is well-known that vanilla\nadversarial training suffers from daunting robust overfitting, resulting in\nunsatisfactory robust generalization. A number of approaches have been proposed\nto address these drawbacks such as extra regularization, adversarial weights\nperturbation, and training with more data over the last few years. However, the\nrobust generalization improvement is yet far from satisfactory. In this paper,\nwe approach this challenge with a brand new perspective -- refining historical\noptimization trajectories. We propose a new method named \\textbf{Weighted\nOptimization Trajectories (WOT)} that leverages the optimization trajectories\nof adversarial training in time. We have conducted extensive experiments to\ndemonstrate the effectiveness of WOT under various state-of-the-art adversarial\nattacks. Our results show that WOT integrates seamlessly with the existing\nadversarial training methods and consistently overcomes the robust overfitting\nissue, resulting in better adversarial robustness. For example, WOT boosts the\nrobust accuracy of AT-PGD under AA-$L_{\\infty}$ attack by 1.53\\% $\\sim$ 6.11\\%\nand meanwhile increases the clean accuracy by 0.55\\%$\\sim$5.47\\% across SVHN,\nCIFAR-10, CIFAR-100, and Tiny-ImageNet datasets.\n", "date": "2023-06-25", "categories": []}
{"id": "2309.08999", "title": "Context-aware Adversarial Attack on Named Entity Recognition", "abstract": "  In recent years, large pre-trained language models (PLMs) have achieved\nremarkable performance on many natural language processing benchmarks. Despite\ntheir success, prior studies have shown that PLMs are vulnerable to attacks\nfrom adversarial examples. In this work, we focus on the named entity\nrecognition task and study context-aware adversarial attack methods to examine\nthe model's robustness. Specifically, we propose perturbing the most\ninformative words for recognizing entities to create adversarial examples and\ninvestigate different candidate replacement methods to generate natural and\nplausible adversarial examples. Experiments and analyses show that our methods\nare more effective in deceiving the model into making wrong predictions than\nstrong baselines.\n", "date": "2023-09-16", "categories": []}
{"id": "2310.06112", "title": "Theoretical Analysis of Robust Overfitting for Wide DNNs: An NTK\n  Approach", "abstract": "  Adversarial training (AT) is a canonical method for enhancing the robustness\nof deep neural networks (DNNs). However, recent studies empirically\ndemonstrated that it suffers from robust overfitting, i.e., a long time AT can\nbe detrimental to the robustness of DNNs. This paper presents a theoretical\nexplanation of robust overfitting for DNNs. Specifically, we non-trivially\nextend the neural tangent kernel (NTK) theory to AT and prove that an\nadversarially trained wide DNN can be well approximated by a linearized DNN.\nMoreover, for squared loss, closed-form AT dynamics for the linearized DNN can\nbe derived, which reveals a new AT degeneration phenomenon: a long-term AT will\nresult in a wide DNN degenerates to that obtained without AT and thus cause\nrobust overfitting. Based on our theoretical results, we further design a\nmethod namely Adv-NTK, the first AT algorithm for infinite-width DNNs.\nExperiments on real-world datasets show that Adv-NTK can help infinite-width\nDNNs enhance comparable robustness to that of their finite-width counterparts,\nwhich in turn justifies our theoretical findings. The code is available at\nhttps://github.com/fshp971/adv-ntk.\n", "date": "2023-10-09", "categories": []}
{"id": "2312.00050", "title": "Elijah: Eliminating Backdoors Injected in Diffusion Models via\n  Distribution Shift", "abstract": "  Diffusion models (DM) have become state-of-the-art generative models because\nof their capability to generate high-quality images from noises without\nadversarial training. However, they are vulnerable to backdoor attacks as\nreported by recent studies. When a data input (e.g., some Gaussian noise) is\nstamped with a trigger (e.g., a white patch), the backdoored model always\ngenerates the target image (e.g., an improper photo). However, effective\ndefense strategies to mitigate backdoors from DMs are underexplored. To bridge\nthis gap, we propose the first backdoor detection and removal framework for\nDMs. We evaluate our framework Elijah on hundreds of DMs of 3 types including\nDDPM, NCSN and LDM, with 13 samplers against 3 existing backdoor attacks.\nExtensive experiments show that our approach can have close to 100% detection\naccuracy and reduce the backdoor effects to close to zero without significantly\nsacrificing the model utility.\n", "date": "2023-11-27", "categories": []}
{"id": "2401.11373", "title": "Finding a Needle in the Adversarial Haystack: A Targeted Paraphrasing\n  Approach For Uncovering Edge Cases with Minimal Distribution Distortion", "abstract": "  Adversarial attacks against language models(LMs) are a significant concern.\nIn particular, adversarial samples exploit the model's sensitivity to small\ninput changes. While these changes appear insignificant on the semantics of the\ninput sample, they result in significant decay in model performance. In this\npaper, we propose Targeted Paraphrasing via RL (TPRL), an approach to\nautomatically learn a policy to generate challenging samples that most likely\nimprove the model's performance. TPRL leverages FLAN T5, a language model, as a\ngenerator and employs a self learned policy using a proximal policy gradient to\ngenerate the adversarial examples automatically. TPRL's reward is based on the\nconfusion induced in the classifier, preserving the original text meaning\nthrough a Mutual Implication score. We demonstrate and evaluate TPRL's\neffectiveness in discovering natural adversarial attacks and improving model\nperformance through extensive experiments on four diverse NLP classification\ntasks via Automatic and Human evaluation. TPRL outperforms strong baselines,\nexhibits generalizability across classifiers and datasets, and combines the\nstrengths of language modeling and reinforcement learning to generate diverse\nand influential adversarial examples.\n", "date": "2024-01-21", "categories": []}
{"id": "2402.01806", "title": "HQA-Attack: Toward High Quality Black-Box Hard-Label Adversarial Attack\n  on Text", "abstract": "  Black-box hard-label adversarial attack on text is a practical and\nchallenging task, as the text data space is inherently discrete and\nnon-differentiable, and only the predicted label is accessible. Research on\nthis problem is still in the embryonic stage and only a few methods are\navailable. Nevertheless, existing methods rely on the complex heuristic\nalgorithm or unreliable gradient estimation strategy, which probably fall into\nthe local optimum and inevitably consume numerous queries, thus are difficult\nto craft satisfactory adversarial examples with high semantic similarity and\nlow perturbation rate in a limited query budget. To alleviate above issues, we\npropose a simple yet effective framework to generate high quality textual\nadversarial examples under the black-box hard-label attack scenarios, named\nHQA-Attack. Specifically, after initializing an adversarial example randomly,\nHQA-attack first constantly substitutes original words back as many as\npossible, thus shrinking the perturbation rate. Then it leverages the synonym\nset of the remaining changed words to further optimize the adversarial example\nwith the direction which can improve the semantic similarity and satisfy the\nadversarial condition simultaneously. In addition, during the optimizing\nprocedure, it searches a transition synonym word for each changed word, thus\navoiding traversing the whole synonym set and reducing the query number to some\nextent. Extensive experimental results on five text classification datasets,\nthree natural language inference datasets and two real-world APIs have shown\nthat the proposed HQA-Attack method outperforms other strong baselines\nsignificantly.\n", "date": "2024-02-02", "categories": []}
{"id": "2402.02154", "title": "Evaluating the Robustness of Off-Road Autonomous Driving Segmentation\n  against Adversarial Attacks: A Dataset-Centric analysis", "abstract": "  This study investigates the vulnerability of semantic segmentation models to\nadversarial input perturbations, in the domain of off-road autonomous driving.\nDespite good performance in generic conditions, the state-of-the-art\nclassifiers are often susceptible to (even) small perturbations, ultimately\nresulting in inaccurate predictions with high confidence. Prior research has\ndirected their focus on making models more robust by modifying the architecture\nand training with noisy input images, but has not explored the influence of\ndatasets in adversarial attacks. Our study aims to address this gap by\nexamining the impact of non-robust features in off-road datasets and comparing\nthe effects of adversarial attacks on different segmentation network\narchitectures. To enable this, a robust dataset is created consisting of only\nrobust features and training the networks on this robustified dataset. We\npresent both qualitative and quantitative analysis of our findings, which have\nimportant implications on improving the robustness of machine learning models\nin off-road autonomous driving applications. Additionally, this work\ncontributes to the safe navigation of autonomous robot Unimog U5023 in rough\noff-road unstructured environments by evaluating the robustness of segmentation\noutputs. The code is publicly available at\nhttps://github.com/rohtkumar/adversarial_attacks_ on_segmentation\n", "date": "2024-02-03", "categories": []}
{"id": "2402.02600", "title": "Evading Deep Learning-Based Malware Detectors via Obfuscation: A Deep\n  Reinforcement Learning Approach", "abstract": "  Adversarial Malware Generation (AMG), the generation of adversarial malware\nvariants to strengthen Deep Learning (DL)-based malware detectors has emerged\nas a crucial tool in the development of proactive cyberdefense. However, the\nmajority of extant works offer subtle perturbations or additions to executable\nfiles and do not explore full-file obfuscation. In this study, we show that an\nopen-source encryption tool coupled with a Reinforcement Learning (RL)\nframework can successfully obfuscate malware to evade state-of-the-art malware\ndetection engines and outperform techniques that use advanced modification\nmethods. Our results show that the proposed method improves the evasion rate\nfrom 27%-49% compared to widely-used state-of-the-art reinforcement\nlearning-based methods.\n", "date": "2024-02-04", "categories": []}
{"id": "2402.02739", "title": "DisDet: Exploring Detectability of Backdoor Attack on Diffusion Models", "abstract": "  In the exciting generative AI era, the diffusion model has emerged as a very\npowerful and widely adopted content generation and editing tool for various\ndata modalities, making the study of their potential security risks very\nnecessary and critical. Very recently, some pioneering works have shown the\nvulnerability of the diffusion model against backdoor attacks, calling for\nin-depth analysis and investigation of the security challenges of this popular\nand fundamental AI technique.\n  In this paper, for the first time, we systematically explore the\ndetectability of the poisoned noise input for the backdoored diffusion models,\nan important performance metric yet little explored in the existing works.\nStarting from the perspective of a defender, we first analyze the properties of\nthe trigger pattern in the existing diffusion backdoor attacks, discovering the\nimportant role of distribution discrepancy in Trojan detection. Based on this\nfinding, we propose a low-cost trigger detection mechanism that can effectively\nidentify the poisoned input noise. We then take a further step to study the\nsame problem from the attack side, proposing a backdoor attack strategy that\ncan learn the unnoticeable trigger to evade our proposed detection scheme.\n  Empirical evaluations across various diffusion models and datasets\ndemonstrate the effectiveness of the proposed trigger detection and\ndetection-evading attack strategy. For trigger detection, our distribution\ndiscrepancy-based solution can achieve a 100\\% detection rate for the Trojan\ntriggers used in the existing works. For evading trigger detection, our\nproposed stealthy trigger design approach performs end-to-end learning to make\nthe distribution of poisoned noise input approach that of benign noise,\nenabling nearly 100\\% detection pass rate with very high attack and benign\nperformance for the backdoored diffusion models.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.03095", "title": "Transcending Adversarial Perturbations: Manifold-Aided Adversarial\n  Examples with Legitimate Semantics", "abstract": "  Deep neural networks were significantly vulnerable to adversarial examples\nmanipulated by malicious tiny perturbations. Although most conventional\nadversarial attacks ensured the visual imperceptibility between adversarial\nexamples and corresponding raw images by minimizing their geometric distance,\nthese constraints on geometric distance led to limited attack transferability,\ninferior visual quality, and human-imperceptible interpretability. In this\npaper, we proposed a supervised semantic-transformation generative model to\ngenerate adversarial examples with real and legitimate semantics, wherein an\nunrestricted adversarial manifold containing continuous semantic variations was\nconstructed for the first time to realize a legitimate transition from\nnon-adversarial examples to adversarial ones. Comprehensive experiments on\nMNIST and industrial defect datasets showed that our adversarial examples not\nonly exhibited better visual quality but also achieved superior attack\ntransferability and more effective explanations for model vulnerabilities,\nindicating their great potential as generic adversarial examples. The code and\npre-trained models were available at https://github.com/shuaili1027/MAELS.git.\n", "date": "2024-02-05", "categories": []}
{"id": "2305.01219", "title": "Prompt as Triggers for Backdoor Attack: Examining the Vulnerability in\n  Language Models", "abstract": "  The prompt-based learning paradigm, which bridges the gap between\npre-training and fine-tuning, achieves state-of-the-art performance on several\nNLP tasks, particularly in few-shot settings. Despite being widely applied,\nprompt-based learning is vulnerable to backdoor attacks. Textual backdoor\nattacks are designed to introduce targeted vulnerabilities into models by\npoisoning a subset of training samples through trigger injection and label\nmodification. However, they suffer from flaws such as abnormal natural language\nexpressions resulting from the trigger and incorrect labeling of poisoned\nsamples. In this study, we propose ProAttack, a novel and efficient method for\nperforming clean-label backdoor attacks based on the prompt, which uses the\nprompt itself as a trigger. Our method does not require external triggers and\nensures correct labeling of poisoned samples, improving the stealthy nature of\nthe backdoor attack. With extensive experiments on rich-resource and few-shot\ntext classification tasks, we empirically validate ProAttack's competitive\nperformance in textual backdoor attacks. Notably, in the rich-resource setting,\nProAttack achieves state-of-the-art attack success rates in the clean-label\nbackdoor attack benchmark without external triggers.\n", "date": "2023-05-02", "categories": []}
{"id": "2310.15085", "title": "On the Detection of Image-Scaling Attacks in Machine Learning", "abstract": "  Image scaling is an integral part of machine learning and computer vision\nsystems. Unfortunately, this preprocessing step is vulnerable to so-called\nimage-scaling attacks where an attacker makes unnoticeable changes to an image\nso that it becomes a new image after scaling. This opens up new ways for\nattackers to control the prediction or to improve poisoning and backdoor\nattacks. While effective techniques exist to prevent scaling attacks, their\ndetection has not been rigorously studied yet. Consequently, it is currently\nnot possible to reliably spot these attacks in practice.\n  This paper presents the first in-depth systematization and analysis of\ndetection methods for image-scaling attacks. We identify two general detection\nparadigms and derive novel methods from them that are simple in design yet\nsignificantly outperform previous work. We demonstrate the efficacy of these\nmethods in a comprehensive evaluation with all major learning platforms and\nscaling algorithms. First, we show that image-scaling attacks modifying the\nentire scaled image can be reliably detected even under an adaptive adversary.\nSecond, we find that our methods provide strong detection performance even if\nonly minor parts of the image are manipulated. As a result, we can introduce a\nnovel protection layer against image-scaling attacks.\n", "date": "2023-10-23", "categories": []}
{"id": "2402.01220", "title": "Delving into Decision-based Black-box Attacks on Semantic Segmentation", "abstract": "  Semantic segmentation is a fundamental visual task that finds extensive\ndeployment in applications with security-sensitive considerations. Nonetheless,\nrecent work illustrates the adversarial vulnerability of semantic segmentation\nmodels to white-box attacks. However, its adversarial robustness against\nblack-box attacks has not been fully explored. In this paper, we present the\nfirst exploration of black-box decision-based attacks on semantic segmentation.\nFirst, we analyze the challenges that semantic segmentation brings to\ndecision-based attacks through the case study. Then, to address these\nchallenges, we first propose a decision-based attack on semantic segmentation,\ncalled Discrete Linear Attack (DLA). Based on random search and proxy index, we\nutilize the discrete linear noises for perturbation exploration and calibration\nto achieve efficient attack efficiency. We conduct adversarial robustness\nevaluation on 5 models from Cityscapes and ADE20K under 8 attacks. DLA shows\nits formidable power on Cityscapes by dramatically reducing PSPNet's mIoU from\nan impressive 77.83% to a mere 2.14% with just 50 queries.\n", "date": "2024-02-02", "categories": []}
{"id": "2402.01227", "title": "STAA-Net: A Sparse and Transferable Adversarial Attack for Speech\n  Emotion Recognition", "abstract": "  Speech contains rich information on the emotions of humans, and Speech\nEmotion Recognition (SER) has been an important topic in the area of\nhuman-computer interaction. The robustness of SER models is crucial,\nparticularly in privacy-sensitive and reliability-demanding domains like\nprivate healthcare. Recently, the vulnerability of deep neural networks in the\naudio domain to adversarial attacks has become a popular area of research.\nHowever, prior works on adversarial attacks in the audio domain primarily rely\non iterative gradient-based techniques, which are time-consuming and prone to\noverfitting the specific threat model. Furthermore, the exploration of sparse\nperturbations, which have the potential for better stealthiness, remains\nlimited in the audio domain. To address these challenges, we propose a\ngenerator-based attack method to generate sparse and transferable adversarial\nexamples to deceive SER models in an end-to-end and efficient manner. We\nevaluate our method on two widely-used SER datasets, Database of Elicited Mood\nin Speech (DEMoS) and Interactive Emotional dyadic MOtion CAPture (IEMOCAP),\nand demonstrate its ability to generate successful sparse adversarial examples\nin an efficient manner. Moreover, our generated adversarial examples exhibit\nmodel-agnostic transferability, enabling effective adversarial attacks on\nadvanced victim models.\n", "date": "2024-02-02", "categories": []}
{"id": "2402.00304", "title": "Invariance-powered Trustworthy Defense via Remove Then Restore", "abstract": "  Adversarial attacks pose a challenge to the deployment of deep neural\nnetworks (DNNs), while previous defense models overlook the generalization to\nvarious attacks. Inspired by targeted therapies for cancer, we view adversarial\nsamples as local lesions of natural benign samples, because a key finding is\nthat salient attack in an adversarial sample dominates the attacking process,\nwhile trivial attack unexpectedly provides trustworthy evidence for obtaining\ngeneralizable robustness. Based on this finding, a Pixel Surgery and Semantic\nRegeneration (PSSR) model following the targeted therapy mechanism is\ndeveloped, which has three merits: 1) To remove the salient attack, a\nscore-based Pixel Surgery module is proposed, which retains the trivial attack\nas a kind of invariance information. 2) To restore the discriminative content,\na Semantic Regeneration module based on a conditional alignment extrapolator is\nproposed, which achieves pixel and semantic consistency. 3) To further\nharmonize robustness and accuracy, an intractable problem, a self-augmentation\nregularizer with adversarial R-drop is designed. Experiments on numerous\nbenchmarks show the superiority of PSSR.\n", "date": "2024-02-01", "categories": []}
{"id": "2402.00576", "title": "Tropical Decision Boundaries for Neural Networks Are Robust Against\n  Adversarial Attacks", "abstract": "  We introduce a simple, easy to implement, and computationally efficient\ntropical convolutional neural network architecture that is robust against\nadversarial attacks. We exploit the tropical nature of piece-wise linear neural\nnetworks by embedding the data in the tropical projective torus in a single\nhidden layer which can be added to any model. We study the geometry of its\ndecision boundary theoretically and show its robustness against adversarial\nattacks on image datasets using computational experiments.\n", "date": "2024-02-01", "categories": []}
{"id": "2401.17523", "title": "Game-Theoretic Unlearnable Example Generator", "abstract": "  Unlearnable example attacks are data poisoning attacks aiming to degrade the\nclean test accuracy of deep learning by adding imperceptible perturbations to\nthe training samples, which can be formulated as a bi-level optimization\nproblem. However, directly solving this optimization problem is intractable for\ndeep neural networks. In this paper, we investigate unlearnable example attacks\nfrom a game-theoretic perspective, by formulating the attack as a nonzero sum\nStackelberg game. First, the existence of game equilibria is proved under the\nnormal setting and the adversarial training setting. It is shown that the game\nequilibrium gives the most powerful poison attack in that the victim has the\nlowest test accuracy among all networks within the same hypothesis space, when\ncertain loss functions are used. Second, we propose a novel attack method,\ncalled the Game Unlearnable Example (GUE), which has three main gradients. (1)\nThe poisons are obtained by directly solving the equilibrium of the Stackelberg\ngame with a first-order algorithm. (2) We employ an autoencoder-like generative\nnetwork model as the poison attacker. (3) A novel payoff function is introduced\nto evaluate the performance of the poison. Comprehensive experiments\ndemonstrate that GUE can effectively poison the model in various scenarios.\nFurthermore, the GUE still works by using a relatively small percentage of the\ntraining data to train the generator, and the poison generator can generalize\nto unseen data well. Our implementation code can be found at\nhttps://github.com/hong-xian/gue.\n", "date": "2024-01-31", "categories": []}
{"id": "2303.18136", "title": "Machine-learned Adversarial Attacks against Fault Prediction Systems in\n  Smart Electrical Grids", "abstract": "  In smart electrical grids, fault detection tasks may have a high impact on\nsociety due to their economic and critical implications. In the recent years,\nnumerous smart grid applications, such as defect detection and load\nforecasting, have embraced data-driven methodologies. The purpose of this study\nis to investigate the challenges associated with the security of machine\nlearning (ML) applications in the smart grid scenario. Indeed, the robustness\nand security of these data-driven algorithms have not been extensively studied\nin relation to all power grid applications. We demonstrate first that the deep\nneural network method used in the smart grid is susceptible to adversarial\nperturbation. Then, we highlight how studies on fault localization and type\nclassification illustrate the weaknesses of present ML algorithms in smart\ngrids to various adversarial attacks\n", "date": "2023-03-28", "categories": []}
{"id": "2401.06031", "title": "GE-AdvGAN: Improving the transferability of adversarial samples by\n  gradient editing-based adversarial generative model", "abstract": "  Adversarial generative models, such as Generative Adversarial Networks\n(GANs), are widely applied for generating various types of data, i.e., images,\ntext, and audio. Accordingly, its promising performance has led to the\nGAN-based adversarial attack methods in the white-box and black-box attack\nscenarios. The importance of transferable black-box attacks lies in their\nability to be effective across different models and settings, more closely\naligning with real-world applications. However, it remains challenging to\nretain the performance in terms of transferable adversarial examples for such\nmethods. Meanwhile, we observe that some enhanced gradient-based transferable\nadversarial attack algorithms require prolonged time for adversarial sample\ngeneration. Thus, in this work, we propose a novel algorithm named GE-AdvGAN to\nenhance the transferability of adversarial samples whilst improving the\nalgorithm's efficiency. The main approach is via optimising the training\nprocess of the generator parameters. With the functional and characteristic\nsimilarity analysis, we introduce a novel gradient editing (GE) mechanism and\nverify its feasibility in generating transferable samples on various models.\nMoreover, by exploring the frequency domain information to determine the\ngradient editing direction, GE-AdvGAN can generate highly transferable\nadversarial samples while minimizing the execution time in comparison to the\nstate-of-the-art transferable adversarial attack algorithms. The performance of\nGE-AdvGAN is comprehensively evaluated by large-scale experiments on different\ndatasets, which results demonstrate the superiority of our algorithm. The code\nfor our algorithm is available at: https://github.com/LMBTough/GE-advGAN\n", "date": "2024-01-11", "categories": []}
{"id": "2401.16766", "title": "Detection and Recovery Against Deep Neural Network Fault Injection\n  Attacks Based on Contrastive Learning", "abstract": "  Deep Neural Network (DNN) models when implemented on executing devices as the\ninference engines are susceptible to Fault Injection Attacks (FIAs) that\nmanipulate model parameters to disrupt inference execution with disastrous\nperformance. This work introduces Contrastive Learning (CL) of visual\nrepresentations i.e., a self-supervised learning approach into the deep\nlearning training and inference pipeline to implement DNN inference engines\nwith self-resilience under FIAs. Our proposed CL based FIA Detection and\nRecovery (CFDR) framework features (i) real-time detection with only a single\nbatch of testing data and (ii) fast recovery effective even with only a small\namount of unlabeled testing data. Evaluated with the CIFAR-10 dataset on\nmultiple types of FIAs, our CFDR shows promising detection and recovery\neffectiveness.\n", "date": "2024-01-30", "categories": []}
{"id": "2304.08767", "title": "Masked Language Model Based Textual Adversarial Example Detection", "abstract": "  Adversarial attacks are a serious threat to the reliable deployment of\nmachine learning models in safety-critical applications. They can misguide\ncurrent models to predict incorrectly by slightly modifying the inputs.\nRecently, substantial work has shown that adversarial examples tend to deviate\nfrom the underlying data manifold of normal examples, whereas pre-trained\nmasked language models can fit the manifold of normal NLP data. To explore how\nto use the masked language model in adversarial detection, we propose a novel\ntextual adversarial example detection method, namely Masked Language\nModel-based Detection (MLMD), which can produce clearly distinguishable signals\nbetween normal examples and adversarial examples by exploring the changes in\nmanifolds induced by the masked language model. MLMD features a plug and play\nusage (i.e., no need to retrain the victim model) for adversarial defense and\nit is agnostic to classification tasks, victim model's architectures, and\nto-be-defended attack methods. We evaluate MLMD on various benchmark textual\ndatasets, widely studied machine learning models, and state-of-the-art (SOTA)\nadversarial attacks (in total $3*4*4 = 48$ settings). Experimental results show\nthat MLMD can achieve strong performance, with detection accuracy up to 0.984,\n0.967, and 0.901 on AG-NEWS, IMDB, and SST-2 datasets, respectively.\nAdditionally, MLMD is superior, or at least comparable to, the SOTA detection\ndefenses in detection accuracy and F1 score. Among many defenses based on the\noff-manifold assumption of adversarial examples, this work offers a new angle\nfor capturing the manifold change. The code for this work is openly accessible\nat \\url{https://github.com/mlmddetection/MLMDdetection}.\n", "date": "2023-04-18", "categories": []}
{"id": "2401.15248", "title": "Better Representations via Adversarial Training in Pre-Training: A\n  Theoretical Perspective", "abstract": "  Pre-training is known to generate universal representations for downstream\ntasks in large-scale deep learning such as large language models. Existing\nliterature, e.g., \\cite{kim2020adversarial}, empirically observe that the\ndownstream tasks can inherit the adversarial robustness of the pre-trained\nmodel. We provide theoretical justifications for this robustness inheritance\nphenomenon. Our theoretical results reveal that feature purification plays an\nimportant role in connecting the adversarial robustness of the pre-trained\nmodel and the downstream tasks in two-layer neural networks. Specifically, we\nshow that (i) with adversarial training, each hidden node tends to pick only\none (or a few) feature; (ii) without adversarial training, the hidden nodes can\nbe vulnerable to attacks. This observation is valid for both supervised\npre-training and contrastive learning. With purified nodes, it turns out that\nclean training is enough to achieve adversarial robustness in downstream tasks.\n", "date": "2024-01-26", "categories": []}
{"id": "2401.15817", "title": "Transparency Attacks: How Imperceptible Image Layers Can Fool AI\n  Perception", "abstract": "  This paper investigates a novel algorithmic vulnerability when imperceptible\nimage layers confound multiple vision models into arbitrary label assignments\nand captions. We explore image preprocessing methods to introduce stealth\ntransparency, which triggers AI misinterpretation of what the human eye\nperceives. The research compiles a broad attack surface to investigate the\nconsequences ranging from traditional watermarking, steganography, and\nbackground-foreground miscues. We demonstrate dataset poisoning using the\nattack to mislabel a collection of grayscale landscapes and logos using either\na single attack layer or randomly selected poisoning classes. For example, a\nmilitary tank to the human eye is a mislabeled bridge to object classifiers\nbased on convolutional networks (YOLO, etc.) and vision transformers (ViT,\nGPT-Vision, etc.). A notable attack limitation stems from its dependency on the\nbackground (hidden) layer in grayscale as a rough match to the transparent\nforeground image that the human eye perceives. This dependency limits the\npractical success rate without manual tuning and exposes the hidden layers when\nplaced on the opposite display theme (e.g., light background, light transparent\nforeground visible, works best against a light theme image viewer or browser).\nThe stealth transparency confounds established vision systems, including\nevading facial recognition and surveillance, digital watermarking, content\nfiltering, dataset curating, automotive and drone autonomy, forensic evidence\ntampering, and retail product misclassifying. This method stands in contrast to\ntraditional adversarial attacks that typically focus on modifying pixel values\nin ways that are either slightly perceptible or entirely imperceptible for both\nhumans and machines.\n", "date": "2024-01-29", "categories": []}
{"id": "2312.13628", "title": "Where and How to Attack? A Causality-Inspired Recipe for Generating\n  Counterfactual Adversarial Examples", "abstract": "  Deep neural networks (DNNs) have been demonstrated to be vulnerable to\nwell-crafted \\emph{adversarial examples}, which are generated through either\nwell-conceived $\\mathcal{L}_p$-norm restricted or unrestricted attacks.\nNevertheless, the majority of those approaches assume that adversaries can\nmodify any features as they wish, and neglect the causal generating process of\nthe data, which is unreasonable and unpractical. For instance, a modification\nin income would inevitably impact features like the debt-to-income ratio within\na banking system. By considering the underappreciated causal generating\nprocess, first, we pinpoint the source of the vulnerability of DNNs via the\nlens of causality, then give theoretical results to answer \\emph{where to\nattack}. Second, considering the consequences of the attack interventions on\nthe current state of the examples to generate more realistic adversarial\nexamples, we propose CADE, a framework that can generate\n\\textbf{C}ounterfactual \\textbf{AD}versarial \\textbf{E}xamples to answer\n\\emph{how to attack}. The empirical results demonstrate CADE's effectiveness,\nas evidenced by its competitive performance across diverse attack scenarios,\nincluding white-box, transfer-based, and random intervention attacks.\n", "date": "2023-12-21", "categories": []}
{"id": "2401.14948", "title": "Conserve-Update-Revise to Cure Generalization and Robustness Trade-off\n  in Adversarial Training", "abstract": "  Adversarial training improves the robustness of neural networks against\nadversarial attacks, albeit at the expense of the trade-off between standard\nand robust generalization. To unveil the underlying factors driving this\nphenomenon, we examine the layer-wise learning capabilities of neural networks\nduring the transition from a standard to an adversarial setting. Our empirical\nfindings demonstrate that selectively updating specific layers while preserving\nothers can substantially enhance the network's learning capacity. We therefore\npropose CURE, a novel training framework that leverages a gradient prominence\ncriterion to perform selective conservation, updating, and revision of weights.\nImportantly, CURE is designed to be dataset- and architecture-agnostic,\nensuring its applicability across various scenarios. It effectively tackles\nboth memorization and overfitting issues, thus enhancing the trade-off between\nrobustness and generalization and additionally, this training approach also\naids in mitigating \"robust overfitting\". Furthermore, our study provides\nvaluable insights into the mechanisms of selective adversarial training and\noffers a promising avenue for future research.\n", "date": "2024-01-26", "categories": []}
{"id": "2302.01757", "title": "RS-Del: Edit Distance Robustness Certificates for Sequence Classifiers\n  via Randomized Deletion", "abstract": "  Randomized smoothing is a leading approach for constructing classifiers that\nare certifiably robust against adversarial examples. Existing work on\nrandomized smoothing has focused on classifiers with continuous inputs, such as\nimages, where $\\ell_p$-norm bounded adversaries are commonly studied. However,\nthere has been limited work for classifiers with discrete or variable-size\ninputs, such as for source code, which require different threat models and\nsmoothing mechanisms. In this work, we adapt randomized smoothing for discrete\nsequence classifiers to provide certified robustness against edit\ndistance-bounded adversaries. Our proposed smoothing mechanism randomized\ndeletion (RS-Del) applies random deletion edits, which are (perhaps\nsurprisingly) sufficient to confer robustness against adversarial deletion,\ninsertion and substitution edits. Our proof of certification deviates from the\nestablished Neyman-Pearson approach, which is intractable in our setting, and\nis instead organized around longest common subsequences. We present a case\nstudy on malware detection--a binary classification problem on byte sequences\nwhere classifier evasion is a well-established threat model. When applied to\nthe popular MalConv malware detection model, our smoothing mechanism RS-Del\nachieves a certified accuracy of 91% at an edit distance radius of 128 bytes.\n", "date": "2023-01-31", "categories": []}
{"id": "2306.13119", "title": "Adversarial Resilience in Sequential Prediction via Abstention", "abstract": "  We study the problem of sequential prediction in the stochastic setting with\nan adversary that is allowed to inject clean-label adversarial (or\nout-of-distribution) examples. Algorithms designed to handle purely stochastic\ndata tend to fail in the presence of such adversarial examples, often leading\nto erroneous predictions. This is undesirable in many high-stakes applications\nsuch as medical recommendations, where abstaining from predictions on\nadversarial examples is preferable to misclassification. On the other hand,\nassuming fully adversarial data leads to very pessimistic bounds that are often\nvacuous in practice.\n  To capture this motivation, we propose a new model of sequential prediction\nthat sits between the purely stochastic and fully adversarial settings by\nallowing the learner to abstain from making a prediction at no cost on\nadversarial examples. Assuming access to the marginal distribution on the\nnon-adversarial examples, we design a learner whose error scales with the VC\ndimension (mirroring the stochastic setting) of the hypothesis class, as\nopposed to the Littlestone dimension which characterizes the fully adversarial\nsetting. Furthermore, we design a learner for VC dimension~1 classes, which\nworks even in the absence of access to the marginal distribution. Our key\ntechnical contribution is a novel measure for quantifying uncertainty for\nlearning VC classes, which may be of independent interest.\n", "date": "2023-06-22", "categories": []}
{"id": "2401.12236", "title": "The Surprising Harmfulness of Benign Overfitting for Adversarial\n  Robustness", "abstract": "  Recent empirical and theoretical studies have established the generalization\ncapabilities of large machine learning models that are trained to\n(approximately or exactly) fit noisy data. In this work, we prove a surprising\nresult that even if the ground truth itself is robust to adversarial examples,\nand the benignly overfitted model is benign in terms of the ``standard''\nout-of-sample risk objective, this benign overfitting process can be harmful\nwhen out-of-sample data are subject to adversarial manipulation. More\nspecifically, our main results contain two parts: (i) the min-norm estimator in\noverparameterized linear model always leads to adversarial vulnerability in the\n``benign overfitting'' setting; (ii) we verify an asymptotic trade-off result\nbetween the standard risk and the ``adversarial'' risk of every ridge\nregression estimator, implying that under suitable conditions these two items\ncannot both be small at the same time by any single choice of the ridge\nregularization parameter. Furthermore, under the lazy training regime, we\ndemonstrate parallel results on two-layer neural tangent kernel (NTK) model,\nwhich align with empirical observations in deep neural networks. Our finding\nprovides theoretical insights into the puzzling phenomenon observed in\npractice, where the true target function (e.g., human) is robust against\nadverasrial attack, while beginly overfitted neural networks lead to models\nthat are not robust.\n", "date": "2024-01-19", "categories": []}
{"id": "2401.14031", "title": "Sparse and Transferable Universal Singular Vectors Attack", "abstract": "  The research in the field of adversarial attacks and models' vulnerability is\none of the fundamental directions in modern machine learning. Recent studies\nreveal the vulnerability phenomenon, and understanding the mechanisms behind\nthis is essential for improving neural network characteristics and\ninterpretability. In this paper, we propose a novel sparse universal white-box\nadversarial attack. Our approach is based on truncated power iteration\nproviding sparsity to $(p,q)$-singular vectors of the hidden layers of Jacobian\nmatrices. Using the ImageNet benchmark validation subset, we analyze the\nproposed method in various settings, achieving results comparable to dense\nbaselines with more than a 50% fooling rate while damaging only 5% of pixels\nand utilizing 256 samples for perturbation fitting. We also show that our\nalgorithm admits higher attack magnitude without affecting the human ability to\nsolve the task. Furthermore, we investigate that the constructed perturbations\nare highly transferable among different models without significantly decreasing\nthe fooling rate. Our findings demonstrate the vulnerability of\nstate-of-the-art models to sparse attacks and highlight the importance of\ndeveloping robust machine learning systems.\n", "date": "2024-01-25", "categories": []}
{"id": "2401.13205", "title": "Boosting the Transferability of Adversarial Examples via Local Mixup and\n  Adaptive Step Size", "abstract": "  Adversarial examples are one critical security threat to various visual\napplications, where injected human-imperceptible perturbations can confuse the\noutput.Generating transferable adversarial examples in the black-box setting is\ncrucial but challenging in practice. Existing input-diversity-based methods\nadopt different image transformations, but may be inefficient due to\ninsufficient input diversity and an identical perturbation step size. Motivated\nby the fact that different image regions have distinctive weights in\nclassification, this paper proposes a black-box adversarial generative\nframework by jointly designing enhanced input diversity and adaptive step\nsizes. We design local mixup to randomly mix a group of transformed adversarial\nimages, strengthening the input diversity. For precise adversarial generation,\nwe project the perturbation into the $tanh$ space to relax the boundary\nconstraint. Moreover, the step sizes of different regions can be dynamically\nadjusted by integrating a second-order momentum.Extensive experiments on\nImageNet validate that our framework can achieve superior transferability\ncompared to state-of-the-art baselines.\n", "date": "2024-01-24", "categories": []}
{"id": "2401.13624", "title": "Can overfitted deep neural networks in adversarial training generalize?\n  -- An approximation viewpoint", "abstract": "  Adversarial training is a widely used method to improve the robustness of\ndeep neural networks (DNNs) over adversarial perturbations. However, it is\nempirically observed that adversarial training on over-parameterized networks\noften suffers from the \\textit{robust overfitting}: it can achieve almost zero\nadversarial training error while the robust generalization performance is not\npromising. In this paper, we provide a theoretical understanding of the\nquestion of whether overfitted DNNs in adversarial training can generalize from\nan approximation viewpoint. Specifically, our main results are summarized into\nthree folds: i) For classification, we prove by construction the existence of\ninfinitely many adversarial training classifiers on over-parameterized DNNs\nthat obtain arbitrarily small adversarial training error (overfitting), whereas\nachieving good robust generalization error under certain conditions concerning\nthe data quality, well separated, and perturbation level. ii) Linear\nover-parameterization (meaning that the number of parameters is only slightly\nlarger than the sample size) is enough to ensure such existence if the target\nfunction is smooth enough. iii) For regression, our results demonstrate that\nthere also exist infinitely many overfitted DNNs with linear\nover-parameterization in adversarial training that can achieve almost optimal\nrates of convergence for the standard generalization error. Overall, our\nanalysis points out that robust overfitting can be avoided but the required\nmodel capacity will depend on the smoothness of the target function, while a\nrobust generalization gap is inevitable. We hope our analysis will give a\nbetter understanding of the mathematical foundations of robustness in DNNs from\nan approximation view.\n", "date": "2024-01-24", "categories": []}
{"id": "2305.13208", "title": "Iterative Adversarial Attack on Image-guided Story Ending Generation", "abstract": "  Multimodal learning involves developing models that can integrate information\nfrom various sources like images and texts. In this field, multimodal text\ngeneration is a crucial aspect that involves processing data from multiple\nmodalities and outputting text. The image-guided story ending generation\n(IgSEG) is a particularly significant task, targeting on an understanding of\ncomplex relationships between text and image data with a complete story text\nending. Unfortunately, deep neural networks, which are the backbone of recent\nIgSEG models, are vulnerable to adversarial samples. Current adversarial attack\nmethods mainly focus on single-modality data and do not analyze adversarial\nattacks for multimodal text generation tasks that use cross-modal information.\nTo this end, we propose an iterative adversarial attack method\n(Iterative-attack) that fuses image and text modality attacks, allowing for an\nattack search for adversarial text and image in an more effective iterative\nway. Experimental results demonstrate that the proposed method outperforms\nexisting single-modal and non-iterative multimodal attack methods, indicating\nthe potential for improving the adversarial robustness of multimodal text\ngeneration models, such as multimodal machine translation, multimodal question\nanswering, etc.\n", "date": "2023-05-16", "categories": []}
{"id": "2401.12532", "title": "DAFA: Distance-Aware Fair Adversarial Training", "abstract": "  The disparity in accuracy between classes in standard training is amplified\nduring adversarial training, a phenomenon termed the robust fairness problem.\nExisting methodologies aimed to enhance robust fairness by sacrificing the\nmodel's performance on easier classes in order to improve its performance on\nharder ones. However, we observe that under adversarial attacks, the majority\nof the model's predictions for samples from the worst class are biased towards\nclasses similar to the worst class, rather than towards the easy classes.\nThrough theoretical and empirical analysis, we demonstrate that robust fairness\ndeteriorates as the distance between classes decreases. Motivated by these\ninsights, we introduce the Distance-Aware Fair Adversarial training (DAFA)\nmethodology, which addresses robust fairness by taking into account the\nsimilarities between classes. Specifically, our method assigns distinct loss\nweights and adversarial margins to each class and adjusts them to encourage a\ntrade-off in robustness among similar classes. Experimental results across\nvarious datasets demonstrate that our method not only maintains average robust\naccuracy but also significantly improves the worst robust accuracy, indicating\na marked improvement in robust fairness compared to existing methods.\n", "date": "2024-01-23", "categories": []}
{"id": "2306.06071", "title": "Adversarial Attack On Yolov5 For Traffic And Road Sign Detection", "abstract": "  This paper implements and investigates popular adversarial attacks on the\nYOLOv5 Object Detection algorithm. The paper explores the vulnerability of the\nYOLOv5 to adversarial attacks in the context of traffic and road sign\ndetection. The paper investigates the impact of different types of attacks,\nincluding the Limited memory Broyden Fletcher Goldfarb Shanno (L-BFGS), the\nFast Gradient Sign Method (FGSM) attack, the Carlini and Wagner (C&W) attack,\nthe Basic Iterative Method (BIM) attack, the Projected Gradient Descent (PGD)\nattack, One Pixel Attack, and the Universal Adversarial Perturbations attack on\nthe accuracy of YOLOv5 in detecting traffic and road signs. The results show\nthat YOLOv5 is susceptible to these attacks, with misclassification rates\nincreasing as the magnitude of the perturbations increases. We also explain the\nresults using saliency maps. The findings of this paper have important\nimplications for the safety and reliability of object detection algorithms used\nin traffic and transportation systems, highlighting the need for more robust\nand secure models to ensure their effectiveness in real-world applications.\n", "date": "2023-05-27", "categories": []}
{"id": "2312.00157", "title": "Universal Backdoor Attacks", "abstract": "  Web-scraped datasets are vulnerable to data poisoning, which can be used for\nbackdooring deep image classifiers during training. Since training on large\ndatasets is expensive, a model is trained once and re-used many times. Unlike\nadversarial examples, backdoor attacks often target specific classes rather\nthan any class learned by the model. One might expect that targeting many\nclasses through a naive composition of attacks vastly increases the number of\npoison samples. We show this is not necessarily true and more efficient,\nuniversal data poisoning attacks exist that allow controlling\nmisclassifications from any source class into any target class with a small\nincrease in poison samples. Our idea is to generate triggers with salient\ncharacteristics that the model can learn. The triggers we craft exploit a\nphenomenon we call inter-class poison transferability, where learning a trigger\nfrom one class makes the model more vulnerable to learning triggers for other\nclasses. We demonstrate the effectiveness and robustness of our universal\nbackdoor attacks by controlling models with up to 6,000 classes while poisoning\nonly 0.15% of the training dataset. Our source code is available at\nhttps://github.com/Ben-Schneider-code/Universal-Backdoor-Attacks.\n", "date": "2023-11-30", "categories": []}
{"id": "2312.13118", "title": "LRS: Enhancing Adversarial Transferability through Lipschitz Regularized\n  Surrogate", "abstract": "  The transferability of adversarial examples is of central importance to\ntransfer-based black-box adversarial attacks. Previous works for generating\ntransferable adversarial examples focus on attacking \\emph{given} pretrained\nsurrogate models while the connections between surrogate models and adversarial\ntrasferability have been overlooked. In this paper, we propose {\\em Lipschitz\nRegularized Surrogate} (LRS) for transfer-based black-box attacks, a novel\napproach that transforms surrogate models towards favorable adversarial\ntransferability. Using such transformed surrogate models, any existing\ntransfer-based black-box attack can run without any change, yet achieving much\nbetter performance. Specifically, we impose Lipschitz regularization on the\nloss landscape of surrogate models to enable a smoother and more controlled\noptimization process for generating more transferable adversarial examples. In\naddition, this paper also sheds light on the connection between the inner\nproperties of surrogate models and adversarial transferability, where three\nfactors are identified: smaller local Lipschitz constant, smoother loss\nlandscape, and stronger adversarial robustness. We evaluate our proposed LRS\napproach by attacking state-of-the-art standard deep neural networks and\ndefense models. The results demonstrate significant improvement on the attack\nsuccess rates and transferability. Our code is available at\nhttps://github.com/TrustAIoT/LRS.\n", "date": "2023-12-20", "categories": []}
{"id": "2401.07991", "title": "Robustness Against Adversarial Attacks via Learning Confined Adversarial\n  Polytopes", "abstract": "  Deep neural networks (DNNs) could be deceived by generating\nhuman-imperceptible perturbations of clean samples. Therefore, enhancing the\nrobustness of DNNs against adversarial attacks is a crucial task. In this\npaper, we aim to train robust DNNs by limiting the set of outputs reachable via\na norm-bounded perturbation added to a clean sample. We refer to this set as\nadversarial polytope, and each clean sample has a respective adversarial\npolytope. Indeed, if the respective polytopes for all the samples are compact\nsuch that they do not intersect the decision boundaries of the DNN, then the\nDNN is robust against adversarial samples. Hence, the inner-working of our\nalgorithm is based on learning \\textbf{c}onfined \\textbf{a}dversarial\n\\textbf{p}olytopes (CAP). By conducting a thorough set of experiments, we\ndemonstrate the effectiveness of CAP over existing adversarial robustness\nmethods in improving the robustness of models against state-of-the-art attacks\nincluding AutoAttack.\n", "date": "2024-01-15", "categories": []}
{"id": "2401.11126", "title": "CARE: Ensemble Adversarial Robustness Evaluation Against Adaptive\n  Attackers for Security Applications", "abstract": "  Ensemble defenses, are widely employed in various security-related\napplications to enhance model performance and robustness. The widespread\nadoption of these techniques also raises many questions: Are general ensembles\ndefenses guaranteed to be more robust than individuals? Will stronger adaptive\nattacks defeat existing ensemble defense strategies as the cybersecurity arms\nrace progresses? Can ensemble defenses achieve adversarial robustness to\ndifferent types of attacks simultaneously and resist the continually adjusted\nadaptive attacks? Unfortunately, these critical questions remain unresolved as\nthere are no platforms for comprehensive evaluation of ensemble adversarial\nattacks and defenses in the cybersecurity domain. In this paper, we propose a\ngeneral Cybersecurity Adversarial Robustness Evaluation (CARE) platform aiming\nto bridge this gap.\n", "date": "2024-01-20", "categories": []}
{"id": "2401.11224", "title": "Susceptibility of Adversarial Attack on Medical Image Segmentation\n  Models", "abstract": "  The nature of deep neural networks has given rise to a variety of attacks,\nbut little work has been done to address the effect of adversarial attacks on\nsegmentation models trained on MRI datasets. In light of the grave consequences\nthat such attacks could cause, we explore four models from the U-Net family and\nexamine their responses to the Fast Gradient Sign Method (FGSM) attack. We\nconduct FGSM attacks on each of them and experiment with various schemes to\nconduct the attacks. In this paper, we find that medical imaging segmentation\nmodels are indeed vulnerable to adversarial attacks and that there is a\nnegligible correlation between parameter size and adversarial attack success.\nFurthermore, we show that using a different loss function than the one used for\ntraining yields higher adversarial attack success, contrary to what the FGSM\nauthors suggested. In future efforts, we will conduct the experiments detailed\nin this paper with more segmentation models and different attacks. We will also\nattempt to find ways to counteract the attacks by using model ensembles or\nspecial data augmentations. Our code is available at\nhttps://github.com/ZhongxuanWang/adv_attk\n", "date": "2024-01-20", "categories": []}
{"id": "2401.11543", "title": "How Robust Are Energy-Based Models Trained With Equilibrium Propagation?", "abstract": "  Deep neural networks (DNNs) are easily fooled by adversarial perturbations\nthat are imperceptible to humans. Adversarial training, a process where\nadversarial examples are added to the training set, is the current\nstate-of-the-art defense against adversarial attacks, but it lowers the model's\naccuracy on clean inputs, is computationally expensive, and offers less\nrobustness to natural noise. In contrast, energy-based models (EBMs), which\nwere designed for efficient implementation in neuromorphic hardware and\nphysical systems, incorporate feedback connections from each layer to the\nprevious layer, yielding a recurrent, deep-attractor architecture which we\nhypothesize should make them naturally robust. Our work is the first to explore\nthe robustness of EBMs to both natural corruptions and adversarial attacks,\nwhich we do using the CIFAR-10 and CIFAR-100 datasets. We demonstrate that EBMs\nare more robust than transformers and display comparable robustness to\nadversarially-trained DNNs on gradient-based (white-box) attacks, query-based\n(black-box) attacks, and natural perturbations without sacrificing clean\naccuracy, and without the need for adversarial training or additional training\ntechniques.\n", "date": "2024-01-21", "categories": []}
{"id": "2401.11902", "title": "A Training-Free Defense Framework for Robust Learned Image Compression", "abstract": "  We study the robustness of learned image compression models against\nadversarial attacks and present a training-free defense technique based on\nsimple image transform functions. Recent learned image compression models are\nvulnerable to adversarial attacks that result in poor compression rate, low\nreconstruction quality, or weird artifacts. To address the limitations, we\npropose a simple but effective two-way compression algorithm with random input\ntransforms, which is conveniently applicable to existing image compression\nmodels. Unlike the na\\\"ive approaches, our approach preserves the original\nrate-distortion performance of the models on clean images. Moreover, the\nproposed algorithm requires no additional training or modification of existing\nmodels, making it more practical. We demonstrate the effectiveness of the\nproposed techniques through extensive experiments under multiple compression\nmodels, evaluation metrics, and attack scenarios.\n", "date": "2024-01-22", "categories": []}
{"id": "2301.10766", "title": "On the Adversarial Robustness of Camera-based 3D Object Detection", "abstract": "  In recent years, camera-based 3D object detection has gained widespread\nattention for its ability to achieve high performance with low computational\ncost. However, the robustness of these methods to adversarial attacks has not\nbeen thoroughly examined, especially when considering their deployment in\nsafety-critical domains like autonomous driving. In this study, we conduct the\nfirst comprehensive investigation of the robustness of leading camera-based 3D\nobject detection approaches under various adversarial conditions. We\nsystematically analyze the resilience of these models under two attack\nsettings: white-box and black-box; focusing on two primary objectives:\nclassification and localization. Additionally, we delve into two types of\nadversarial attack techniques: pixel-based and patch-based. Our experiments\nyield four interesting findings: (a) bird's-eye-view-based representations\nexhibit stronger robustness against localization attacks; (b)\ndepth-estimation-free approaches have the potential to show stronger\nrobustness; (c) accurate depth estimation effectively improves robustness for\ndepth-estimation-based methods; (d) incorporating multi-frame benign inputs can\neffectively mitigate adversarial attacks. We hope our findings can steer the\ndevelopment of future camera-based object detection models with enhanced\nadversarial robustness.\n", "date": "2023-01-25", "categories": []}
{"id": "2401.10405", "title": "Differentially Private and Adversarially Robust Machine Learning: An\n  Empirical Evaluation", "abstract": "  Malicious adversaries can attack machine learning models to infer sensitive\ninformation or damage the system by launching a series of evasion attacks.\nAlthough various work addresses privacy and security concerns, they focus on\nindividual defenses, but in practice, models may undergo simultaneous attacks.\nThis study explores the combination of adversarial training and differentially\nprivate training to defend against simultaneous attacks. While\ndifferentially-private adversarial training, as presented in DP-Adv,\noutperforms the other state-of-the-art methods in performance, it lacks formal\nprivacy guarantees and empirical validation. Thus, in this work, we benchmark\nthe performance of this technique using a membership inference attack and\nempirically show that the resulting approach is as private as non-robust\nprivate models. This work also highlights the need to explore privacy\nguarantees in dynamic training paradigms.\n", "date": "2024-01-18", "categories": []}
{"id": "2401.10657", "title": "FIMBA: Evaluating the Robustness of AI in Genomics via Feature\n  Importance Adversarial Attacks", "abstract": "  With the steady rise of the use of AI in bio-technical applications and the\nwidespread adoption of genomics sequencing, an increasing amount of AI-based\nalgorithms and tools is entering the research and production stage affecting\ncritical decision-making streams like drug discovery and clinical outcomes.\nThis paper demonstrates the vulnerability of AI models often utilized\ndownstream tasks on recognized public genomics datasets. We undermine model\nrobustness by deploying an attack that focuses on input transformation while\nmimicking the real data and confusing the model decision-making, ultimately\nyielding a pronounced deterioration in model performance. Further, we enhance\nour approach by generating poisoned data using a variational autoencoder-based\nmodel. Our empirical findings unequivocally demonstrate a decline in model\nperformance, underscored by diminished accuracy and an upswing in false\npositives and false negatives. Furthermore, we analyze the resulting\nadversarial samples via spectral analysis yielding conclusions for\ncountermeasures against such attacks.\n", "date": "2024-01-19", "categories": []}
{"id": "2401.09945", "title": "HGAttack: Transferable Heterogeneous Graph Adversarial Attack", "abstract": "  Heterogeneous Graph Neural Networks (HGNNs) are increasingly recognized for\ntheir performance in areas like the web and e-commerce, where resilience\nagainst adversarial attacks is crucial. However, existing adversarial attack\nmethods, which are primarily designed for homogeneous graphs, fall short when\napplied to HGNNs due to their limited ability to address the structural and\nsemantic complexity of HGNNs. This paper introduces HGAttack, the first\ndedicated gray box evasion attack method for heterogeneous graphs. We design a\nnovel surrogate model to closely resemble the behaviors of the target HGNN and\nutilize gradient-based methods for perturbation generation. Specifically, the\nproposed surrogate model effectively leverages heterogeneous information by\nextracting meta-path induced subgraphs and applying GNNs to learn node\nembeddings with distinct semantics from each subgraph. This approach improves\nthe transferability of generated attacks on the target HGNN and significantly\nreduces memory costs. For perturbation generation, we introduce a\nsemantics-aware mechanism that leverages subgraph gradient information to\nautonomously identify vulnerable edges across a wide range of relations within\na constrained perturbation budget. We validate HGAttack's efficacy with\ncomprehensive experiments on three datasets, providing empirical analyses of\nits generated perturbations. Outperforming baseline methods, HGAttack\ndemonstrated significant efficacy in diminishing the performance of target HGNN\nmodels, affirming the effectiveness of our approach in evaluating the\nrobustness of HGNNs against adversarial attacks.\n", "date": "2024-01-18", "categories": []}
{"id": "2401.10091", "title": "Power in Numbers: Robust reading comprehension by finetuning with four\n  adversarial sentences per example", "abstract": "  Recent models have achieved human level performance on the Stanford Question\nAnswering Dataset when using F1 scores to evaluate the reading comprehension\ntask. Yet, teaching machines to comprehend text has not been solved in the\ngeneral case. By appending one adversarial sentence to the context paragraph,\npast research has shown that the F1 scores from reading comprehension models\ndrop almost in half. In this paper, I replicate past adversarial research with\na new model, ELECTRA-Small, and demonstrate that the new model's F1 score drops\nfrom 83.9% to 29.2%. To improve ELECTRA-Small's resistance to this attack, I\nfinetune the model on SQuAD v1.1 training examples with one to five adversarial\nsentences appended to the context paragraph. Like past research, I find that\nthe finetuned model on one adversarial sentence does not generalize well across\nevaluation datasets. However, when finetuned on four or five adversarial\nsentences the model attains an F1 score of more than 70% on most evaluation\ndatasets with multiple appended and prepended adversarial sentences. The\nresults suggest that with enough examples we can make models robust to\nadversarial attacks.\n", "date": "2024-01-18", "categories": []}
{"id": "2305.16494", "title": "Diffusion-Based Adversarial Sample Generation for Improved Stealthiness\n  and Controllability", "abstract": "  Neural networks are known to be susceptible to adversarial samples: small\nvariations of natural examples crafted to deliberately mislead the models.\nWhile they can be easily generated using gradient-based techniques in digital\nand physical scenarios, they often differ greatly from the actual data\ndistribution of natural images, resulting in a trade-off between strength and\nstealthiness. In this paper, we propose a novel framework dubbed\nDiffusion-Based Projected Gradient Descent (Diff-PGD) for generating realistic\nadversarial samples. By exploiting a gradient guided by a diffusion model,\nDiff-PGD ensures that adversarial samples remain close to the original data\ndistribution while maintaining their effectiveness. Moreover, our framework can\nbe easily customized for specific tasks such as digital attacks, physical-world\nattacks, and style-based attacks. Compared with existing methods for generating\nnatural-style adversarial samples, our framework enables the separation of\noptimizing adversarial loss from other surrogate losses (e.g.,\ncontent/smoothness/style loss), making it more stable and controllable.\nFinally, we demonstrate that the samples generated using Diff-PGD have better\ntransferability and anti-purification power than traditional gradient-based\nmethods. Code will be released in https://github.com/xavihart/Diff-PGD\n", "date": "2023-05-25", "categories": []}
{"id": "2401.08725", "title": "Revealing Vulnerabilities in Stable Diffusion via Targeted Attacks", "abstract": "  Recent developments in text-to-image models, particularly Stable Diffusion,\nhave marked significant achievements in various applications. With these\nadvancements, there are growing safety concerns about the vulnerability of the\nmodel that malicious entities exploit to generate targeted harmful images.\nHowever, the existing methods in the vulnerability of the model mainly evaluate\nthe alignment between the prompt and generated images, but fall short in\nrevealing the vulnerability associated with targeted image generation. In this\nstudy, we formulate the problem of targeted adversarial attack on Stable\nDiffusion and propose a framework to generate adversarial prompts.\nSpecifically, we design a gradient-based embedding optimization method to craft\nreliable adversarial prompts that guide stable diffusion to generate specific\nimages. Furthermore, after obtaining successful adversarial prompts, we reveal\nthe mechanisms that cause the vulnerability of the model. Extensive experiments\non two targeted attack tasks demonstrate the effectiveness of our method in\ntargeted attacks. The code can be obtained in\nhttps://github.com/datar001/Revealing-Vulnerabilities-in-Stable-Diffusion-via-Targeted-Attacks.\n", "date": "2024-01-16", "categories": []}
{"id": "2401.09191", "title": "An Optimal Transport Approach for Computing Adversarial Training Lower\n  Bounds in Multiclass Classification", "abstract": "  Despite the success of deep learning-based algorithms, it is widely known\nthat neural networks may fail to be robust. A popular paradigm to enforce\nrobustness is adversarial training (AT), however, this introduces many\ncomputational and theoretical difficulties. Recent works have developed a\nconnection between AT in the multiclass classification setting and\nmultimarginal optimal transport (MOT), unlocking a new set of tools to study\nthis problem. In this paper, we leverage the MOT connection to propose\ncomputationally tractable numerical algorithms for computing universal lower\nbounds on the optimal adversarial risk and identifying optimal classifiers. We\npropose two main algorithms based on linear programming (LP) and entropic\nregularization (Sinkhorn). Our key insight is that one can harmlessly truncate\nthe higher order interactions between classes, preventing the combinatorial run\ntimes typically encountered in MOT problems. We validate these results with\nexperiments on MNIST and CIFAR-$10$, which demonstrate the tractability of our\napproach.\n", "date": "2024-01-17", "categories": []}
{"id": "2308.16562", "title": "The Power of MEME: Adversarial Malware Creation with Model-Based\n  Reinforcement Learning", "abstract": "  Due to the proliferation of malware, defenders are increasingly turning to\nautomation and machine learning as part of the malware detection tool-chain.\nHowever, machine learning models are susceptible to adversarial attacks,\nrequiring the testing of model and product robustness. Meanwhile, attackers\nalso seek to automate malware generation and evasion of antivirus systems, and\ndefenders try to gain insight into their methods. This work proposes a new\nalgorithm that combines Malware Evasion and Model Extraction (MEME) attacks.\nMEME uses model-based reinforcement learning to adversarially modify Windows\nexecutable binary samples while simultaneously training a surrogate model with\na high agreement with the target model to evade. To evaluate this method, we\ncompare it with two state-of-the-art attacks in adversarial malware creation,\nusing three well-known published models and one antivirus product as targets.\nResults show that MEME outperforms the state-of-the-art methods in terms of\nevasion capabilities in almost all cases, producing evasive malware with an\nevasion rate in the range of 32-73%. It also produces surrogate models with a\nprediction label agreement with the respective target models between 97-99%.\nThe surrogate could be used to fine-tune and improve the evasion rate in the\nfuture.\n", "date": "2023-08-31", "categories": []}
{"id": "2312.02708", "title": "Provable Adversarial Robustness for Group Equivariant Tasks: Graphs,\n  Point Clouds, Molecules, and More", "abstract": "  A machine learning model is traditionally considered robust if its prediction\nremains (almost) constant under input perturbations with small norm. However,\nreal-world tasks like molecular property prediction or point cloud segmentation\nhave inherent equivariances, such as rotation or permutation equivariance. In\nsuch tasks, even perturbations with large norm do not necessarily change an\ninput's semantic content. Furthermore, there are perturbations for which a\nmodel's prediction explicitly needs to change. For the first time, we propose a\nsound notion of adversarial robustness that accounts for task equivariance. We\nthen demonstrate that provable robustness can be achieved by (1) choosing a\nmodel that matches the task's equivariances (2) certifying traditional\nadversarial robustness. Certification methods are, however, unavailable for\nmany models, such as those with continuous equivariances. We close this gap by\ndeveloping the framework of equivariance-preserving randomized smoothing, which\nenables architecture-agnostic certification. We additionally derive the first\narchitecture-specific graph edit distance certificates, i.e. sound robustness\nguarantees for isomorphism equivariant tasks like node classification. Overall,\na sound notion of robustness is an important prerequisite for future work at\nthe intersection of robust and geometric machine learning.\n", "date": "2023-12-05", "categories": []}
{"id": "2401.02727", "title": "Enhancing targeted transferability via feature space fine-tuning", "abstract": "  Adversarial examples (AEs) have been extensively studied due to their\npotential for privacy protection and inspiring robust neural networks. Yet,\nmaking a targeted AE transferable across unknown models remains challenging. In\nthis paper, to alleviate the overfitting dilemma common in an AE crafted by\nexisting simple iterative attacks, we propose fine-tuning it in the feature\nspace. Specifically, starting with an AE generated by a baseline attack, we\nencourage the features conducive to the target class and discourage the\nfeatures to the original class in a middle layer of the source model. Extensive\nexperiments demonstrate that only a few iterations of fine-tuning can boost\nexisting attacks' targeted transferability nontrivially and universally. Our\nresults also verify that the simple iterative attacks can yield comparable or\neven better transferability than the resource-intensive methods, which rest on\ntraining target-specific classifiers or generators with additional data. The\ncode is available at: github.com/zengh5/TA_feature_FT.\n", "date": "2024-01-05", "categories": []}
{"id": "2401.07087", "title": "Exploring Adversarial Attacks against Latent Diffusion Model from the\n  Perspective of Adversarial Transferability", "abstract": "  Recently, many studies utilized adversarial examples (AEs) to raise the cost\nof malicious image editing and copyright violation powered by latent diffusion\nmodels (LDMs). Despite their successes, a few have studied the surrogate model\nthey used to generate AEs. In this paper, from the perspective of adversarial\ntransferability, we investigate how the surrogate model's property influences\nthe performance of AEs for LDMs. Specifically, we view the time-step sampling\nin the Monte-Carlo-based (MC-based) adversarial attack as selecting surrogate\nmodels. We find that the smoothness of surrogate models at different time steps\ndiffers, and we substantially improve the performance of the MC-based AEs by\nselecting smoother surrogate models. In the light of the theoretical framework\non adversarial transferability in image classification, we also conduct a\ntheoretical analysis to explain why smooth surrogate models can also boost AEs\nfor LDMs.\n", "date": "2024-01-13", "categories": []}
{"id": "2401.07188", "title": "Left-right Discrepancy for Adversarial Attack on Stereo Networks", "abstract": "  Stereo matching neural networks often involve a Siamese structure to extract\nintermediate features from left and right images. The similarity between these\nintermediate left-right features significantly impacts the accuracy of\ndisparity estimation. In this paper, we introduce a novel adversarial attack\napproach that generates perturbation noise specifically designed to maximize\nthe discrepancy between left and right image features. Extensive experiments\ndemonstrate the superior capability of our method to induce larger prediction\nerrors in stereo neural networks, e.g. outperforming existing state-of-the-art\nattack methods by 219% MAE on the KITTI dataset and 85% MAE on the Scene Flow\ndataset. Additionally, we extend our approach to include a proxy network\nblack-box attack method, eliminating the need for access to stereo neural\nnetwork. This method leverages an arbitrary network from a different vision\ntask as a proxy to generate adversarial noise, effectively causing the stereo\nnetwork to produce erroneous predictions. Our findings highlight a notable\nsensitivity of stereo networks to discrepancies in shallow layer features,\noffering valuable insights that could guide future research in enhancing the\nrobustness of stereo vision systems.\n", "date": "2024-01-14", "categories": []}
{"id": "2401.08255", "title": "A Generative Adversarial Attack for Multilingual Text Classifiers", "abstract": "  Current adversarial attack algorithms, where an adversary changes a text to\nfool a victim model, have been repeatedly shown to be effective against text\nclassifiers. These attacks, however, generally assume that the victim model is\nmonolingual and cannot be used to target multilingual victim models, a\nsignificant limitation given the increased use of these models. For this\nreason, in this work we propose an approach to fine-tune a multilingual\nparaphrase model with an adversarial objective so that it becomes able to\ngenerate effective adversarial examples against multilingual classifiers. The\ntraining objective incorporates a set of pre-trained models to ensure text\nquality and language consistency of the generated text. In addition, all the\nmodels are suitably connected to the generator by vocabulary-mapping matrices,\nallowing for full end-to-end differentiability of the overall training\npipeline. The experimental validation over two multilingual datasets and five\nlanguages has shown the effectiveness of the proposed approach compared to\nexisting baselines, particularly in terms of query efficiency. We also provide\na detailed analysis of the generated attacks and discuss limitations and\nopportunities for future research.\n", "date": "2024-01-16", "categories": []}
{"id": "2303.02669", "title": "Consistent Valid Physically-Realizable Adversarial Attack against\n  Crowd-flow Prediction Models", "abstract": "  Recent works have shown that deep learning (DL) models can effectively learn\ncity-wide crowd-flow patterns, which can be used for more effective urban\nplanning and smart city management. However, DL models have been known to\nperform poorly on inconspicuous adversarial perturbations. Although many works\nhave studied these adversarial perturbations in general, the adversarial\nvulnerabilities of deep crowd-flow prediction models in particular have\nremained largely unexplored. In this paper, we perform a rigorous analysis of\nthe adversarial vulnerabilities of DL-based crowd-flow prediction models under\nmultiple threat settings, making three-fold contributions. (1) We propose\nCaV-detect by formally identifying two novel properties - Consistency and\nValidity - of the crowd-flow prediction inputs that enable the detection of\nstandard adversarial inputs with 0% false acceptance rate (FAR). (2) We\nleverage universal adversarial perturbations and an adaptive adversarial loss\nto present adaptive adversarial attacks to evade CaV-detect defense. (3) We\npropose CVPR, a Consistent, Valid and Physically-Realizable adversarial attack,\nthat explicitly inducts the consistency and validity priors in the perturbation\ngeneration mechanism. We find out that although the crowd-flow models are\nvulnerable to adversarial perturbations, it is extremely challenging to\nsimulate these perturbations in physical settings, notably when CaV-detect is\nin place. We also show that CVPR attack considerably outperforms the adaptively\nmodified standard attacks in FAR and adversarial loss metrics. We conclude with\nuseful insights emerging from our work and highlight promising future research\ndirections.\n", "date": "2023-03-05", "categories": []}
{"id": "2207.03689", "title": "Guiding the retraining of convolutional neural networks against\n  adversarial inputs", "abstract": "  Background: When using deep learning models, there are many possible\nvulnerabilities and some of the most worrying are the adversarial inputs, which\ncan cause wrong decisions with minor perturbations. Therefore, it becomes\nnecessary to retrain these models against adversarial inputs, as part of the\nsoftware testing process addressing the vulnerability to these inputs.\nFurthermore, for an energy efficient testing and retraining, data scientists\nneed support on which are the best guidance metrics and optimal dataset\nconfigurations.\n  Aims: We examined four guidance metrics for retraining convolutional neural\nnetworks and three retraining configurations. Our goal is to improve the models\nagainst adversarial inputs regarding accuracy, resource utilization and time\nfrom the point of view of a data scientist in the context of image\nclassification.\n  Method: We conducted an empirical study in two datasets for image\nclassification. We explore: (a) the accuracy, resource utilization and time of\nretraining convolutional neural networks by ordering new training set by four\ndifferent guidance metrics (neuron coverage, likelihood-based surprise\nadequacy, distance-based surprise adequacy and random), (b) the accuracy and\nresource utilization of retraining convolutional neural networks with three\ndifferent configurations (from scratch and augmented dataset, using weights and\naugmented dataset, and using weights and only adversarial inputs).\n  Results: We reveal that retraining with adversarial inputs from original\nweights and by ordering with surprise adequacy metrics gives the best model\nw.r.t. the used metrics.\n  Conclusions: Although more studies are necessary, we recommend data\nscientists to use the above configuration and metrics to deal with the\nvulnerability to adversarial inputs of deep learning models, as they can\nimprove their models against adversarial inputs without using many inputs.\n", "date": "2022-07-08", "categories": []}
{"id": "2210.08159", "title": "Dynamics-aware Adversarial Attack of Adaptive Neural Networks", "abstract": "  In this paper, we investigate the dynamics-aware adversarial attack problem\nof adaptive neural networks. Most existing adversarial attack algorithms are\ndesigned under a basic assumption -- the network architecture is fixed\nthroughout the attack process. However, this assumption does not hold for many\nrecently proposed adaptive neural networks, which adaptively deactivate\nunnecessary execution units based on inputs to improve computational\nefficiency. It results in a serious issue of lagged gradient, making the\nlearned attack at the current step ineffective due to the architecture change\nafterward. To address this issue, we propose a Leaded Gradient Method (LGM) and\nshow the significant effects of the lagged gradient. More specifically, we\nreformulate the gradients to be aware of the potential dynamic changes of\nnetwork architectures, so that the learned attack better \"leads\" the next step\nthan the dynamics-unaware methods when network architecture changes\ndynamically. Extensive experiments on representative types of adaptive neural\nnetworks for both 2D images and 3D point clouds show that our LGM achieves\nimpressive adversarial attack performance compared with the dynamic-unaware\nattack methods. Code is available at https://github.com/antao97/LGM.\n", "date": "2022-10-15", "categories": []}
{"id": "2401.05998", "title": "Combating Adversarial Attacks with Multi-Agent Debate", "abstract": "  While state-of-the-art language models have achieved impressive results, they\nremain susceptible to inference-time adversarial attacks, such as adversarial\nprompts generated by red teams arXiv:2209.07858. One approach proposed to\nimprove the general quality of language model generations is multi-agent\ndebate, where language models self-evaluate through discussion and feedback\narXiv:2305.14325. We implement multi-agent debate between current\nstate-of-the-art language models and evaluate models' susceptibility to red\nteam attacks in both single- and multi-agent settings. We find that multi-agent\ndebate can reduce model toxicity when jailbroken or less capable models are\nforced to debate with non-jailbroken or more capable models. We also find\nmarginal improvements through the general usage of multi-agent interactions. We\nfurther perform adversarial prompt content classification via embedding\nclustering, and analyze the susceptibility of different models to different\ntypes of attack topics.\n", "date": "2024-01-11", "categories": []}
{"id": "2304.12707", "title": "Lyapunov-Stable Deep Equilibrium Models", "abstract": "  Deep equilibrium (DEQ) models have emerged as a promising class of implicit\nlayer models, which abandon traditional depth by solving for the fixed points\nof a single nonlinear layer. Despite their success, the stability of the fixed\npoints for these models remains poorly understood. By considering DEQ models as\nnonlinear dynamic systems, we propose a robust DEQ model named LyaDEQ with\nguaranteed provable stability via Lyapunov theory. The crux of our method is\nensuring the Lyapunov stability of the DEQ model's fixed points, which enables\nthe proposed model to resist minor initial perturbations. To avoid poor\nadversarial defense due to Lyapunov-stable fixed points being located near each\nother, we orthogonalize the layers after the Lyapunov stability module to\nseparate different fixed points. We evaluate LyaDEQ models under well-known\nadversarial attacks, and experimental results demonstrate significant\nimprovement in robustness. Furthermore, we show that the LyaDEQ model can be\ncombined with other defense methods, such as adversarial training, to achieve\neven better adversarial robustness.\n", "date": "2023-04-25", "categories": []}
{"id": "2308.00319", "title": "LimeAttack: Local Explainable Method for Textual Hard-Label Adversarial\n  Attack", "abstract": "  Natural language processing models are vulnerable to adversarial examples.\nPrevious textual adversarial attacks adopt gradients or confidence scores to\ncalculate word importance ranking and generate adversarial examples. However,\nthis information is unavailable in the real world. Therefore, we focus on a\nmore realistic and challenging setting, named hard-label attack, in which the\nattacker can only query the model and obtain a discrete prediction label.\nExisting hard-label attack algorithms tend to initialize adversarial examples\nby random substitution and then utilize complex heuristic algorithms to\noptimize the adversarial perturbation. These methods require a lot of model\nqueries and the attack success rate is restricted by adversary initialization.\nIn this paper, we propose a novel hard-label attack algorithm named LimeAttack,\nwhich leverages a local explainable method to approximate word importance\nranking, and then adopts beam search to find the optimal solution. Extensive\nexperiments show that LimeAttack achieves the better attacking performance\ncompared with existing hard-label attack under the same query budget. In\naddition, we evaluate the effectiveness of LimeAttack on large language models,\nand results indicate that adversarial examples remain a significant threat to\nlarge language models. The adversarial examples crafted by LimeAttack are\nhighly transferable and effectively improve model robustness in adversarial\ntraining.\n", "date": "2023-08-01", "categories": []}
{"id": "2312.01468", "title": "Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in\n  Autonomous Driving", "abstract": "  Our study assesses the adversarial robustness of LiDAR-camera fusion models\nin 3D object detection. We introduce an attack technique that, by simply adding\na limited number of physically constrained adversarial points above a car, can\nmake the car undetectable by the fusion model. Experimental results reveal that\neven without changes to the image data channel, the fusion model can be\ndeceived solely by manipulating the LiDAR data channel. This finding raises\nsafety concerns in the field of autonomous driving. Further, we explore how the\nquantity of adversarial points, the distance between the front-near car and the\nLiDAR-equipped car, and various angular factors affect the attack success rate.\nWe believe our research can contribute to the understanding of multi-sensor\nrobustness, offering insights and guidance to enhance the safety of autonomous\ndriving.\n", "date": "2023-12-03", "categories": []}
{"id": "2401.03514", "title": "ROIC-DM: Robust Text Inference and Classification via Diffusion Model", "abstract": "  While language models have made many milestones in text inference and\nclassification tasks, they remain susceptible to adversarial attacks that can\nlead to unforeseen outcomes. Existing works alleviate this problem by equipping\nlanguage models with defense patches. However, these defense strategies often\nrely on impractical assumptions or entail substantial sacrifices in model\nperformance. Consequently, enhancing the resilience of the target model using\nsuch defense mechanisms is a formidable challenge. This paper introduces an\ninnovative model for robust text inference and classification, built upon\ndiffusion models (ROIC-DM). Benefiting from its training involving denoising\nstages, ROIC-DM inherently exhibits greater robustness compared to conventional\nlanguage models. Moreover, ROIC-DM can attain comparable, and in some cases,\nsuperior performance to language models, by effectively incorporating them as\nadvisory components. Extensive experiments conducted with several strong\ntextual adversarial attacks on three datasets demonstrate that (1) ROIC-DM\noutperforms traditional language models in robustness, even when the latter are\nfortified with advanced defense mechanisms; (2) ROIC-DM can achieve comparable\nand even better performance than traditional language models by using them as\nadvisors.\n", "date": "2024-01-07", "categories": []}
{"id": "2211.09782", "title": "Assessing Neural Network Robustness via Adversarial Pivotal Tuning", "abstract": "  The robustness of image classifiers is essential to their deployment in the\nreal world. The ability to assess this resilience to manipulations or\ndeviations from the training data is thus crucial. These modifications have\ntraditionally consisted of minimal changes that still manage to fool\nclassifiers, and modern approaches are increasingly robust to them. Semantic\nmanipulations that modify elements of an image in meaningful ways have thus\ngained traction for this purpose. However, they have primarily been limited to\nstyle, color, or attribute changes. While expressive, these manipulations do\nnot make use of the full capabilities of a pretrained generative model. In this\nwork, we aim to bridge this gap. We show how a pretrained image generator can\nbe used to semantically manipulate images in a detailed, diverse, and\nphotorealistic way while still preserving the class of the original image.\nInspired by recent GAN-based image inversion methods, we propose a method\ncalled Adversarial Pivotal Tuning (APT). Given an image, APT first finds a\npivot latent space input that reconstructs the image using a pretrained\ngenerator. It then adjusts the generator's weights to create small yet semantic\nmanipulations in order to fool a pretrained classifier. APT preserves the full\nexpressive editing capabilities of the generative model. We demonstrate that\nAPT is capable of a wide range of class-preserving semantic image manipulations\nthat fool a variety of pretrained classifiers. Finally, we show that\nclassifiers that are robust to other benchmarks are not robust to APT\nmanipulations and suggest a method to improve them. Code available at:\nhttps://captaine.github.io/apt/\n", "date": "2022-11-17", "categories": []}
{"id": "2312.14260", "title": "Elevating Defenses: Bridging Adversarial Training and Watermarking for\n  Model Resilience", "abstract": "  Machine learning models are being used in an increasing number of critical\napplications; thus, securing their integrity and ownership is critical. Recent\nstudies observed that adversarial training and watermarking have a conflicting\ninteraction. This work introduces a novel framework to integrate adversarial\ntraining with watermarking techniques to fortify against evasion attacks and\nprovide confident model verification in case of intellectual property theft. We\nuse adversarial training together with adversarial watermarks to train a robust\nwatermarked model. The key intuition is to use a higher perturbation budget to\ngenerate adversarial watermarks compared to the budget used for adversarial\ntraining, thus avoiding conflict. We use the MNIST and Fashion-MNIST datasets\nto evaluate our proposed technique on various model stealing attacks. The\nresults obtained consistently outperform the existing baseline in terms of\nrobustness performance and further prove the resilience of this defense against\npruning and fine-tuning removal attacks.\n", "date": "2023-12-21", "categories": []}
{"id": "2401.03156", "title": "Data-Dependent Stability Analysis of Adversarial Training", "abstract": "  Stability analysis is an essential aspect of studying the generalization\nability of deep learning, as it involves deriving generalization bounds for\nstochastic gradient descent-based training algorithms. Adversarial training is\nthe most widely used defense against adversarial example attacks. However,\nprevious generalization bounds for adversarial training have not included\ninformation regarding the data distribution. In this paper, we fill this gap by\nproviding generalization bounds for stochastic gradient descent-based\nadversarial training that incorporate data distribution information. We utilize\nthe concepts of on-average stability and high-order approximate Lipschitz\nconditions to examine how changes in data distribution and adversarial budget\ncan affect robust generalization gaps. Our derived generalization bounds for\nboth convex and non-convex losses are at least as good as the uniform\nstability-based counterparts which do not include data distribution\ninformation. Furthermore, our findings demonstrate how distribution shifts from\ndata poisoning attacks can impact robust generalization.\n", "date": "2024-01-06", "categories": []}
{"id": "2401.03488", "title": "Data-Driven Subsampling in the Presence of an Adversarial Actor", "abstract": "  Deep learning based automatic modulation classification (AMC) has received\nsignificant attention owing to its potential applications in both military and\ncivilian use cases. Recently, data-driven subsampling techniques have been\nutilized to overcome the challenges associated with computational complexity\nand training time for AMC. Beyond these direct advantages of data-driven\nsubsampling, these methods also have regularizing properties that may improve\nthe adversarial robustness of the modulation classifier. In this paper, we\ninvestigate the effects of an adversarial attack on an AMC system that employs\ndeep learning models both for AMC and for subsampling. Our analysis shows that\nsubsampling itself is an effective deterrent to adversarial attacks. We also\nuncover the most efficient subsampling strategy when an adversarial attack on\nboth the classifier and the subsampler is anticipated.\n", "date": "2024-01-07", "categories": []}
{"id": "2401.03582", "title": "Invisible Reflections: Leveraging Infrared Laser Reflections to Target\n  Traffic Sign Perception", "abstract": "  All vehicles must follow the rules that govern traffic behavior, regardless\nof whether the vehicles are human-driven or Connected Autonomous Vehicles\n(CAVs). Road signs indicate locally active rules, such as speed limits and\nrequirements to yield or stop. Recent research has demonstrated attacks, such\nas adding stickers or projected colored patches to signs, that cause CAV\nmisinterpretation, resulting in potential safety issues. Humans can see and\npotentially defend against these attacks. But humans can not detect what they\ncan not observe. We have developed an effective physical-world attack that\nleverages the sensitivity of filterless image sensors and the properties of\nInfrared Laser Reflections (ILRs), which are invisible to humans. The attack is\ndesigned to affect CAV cameras and perception, undermining traffic sign\nrecognition by inducing misclassification. In this work, we formulate the\nthreat model and requirements for an ILR-based traffic sign perception attack\nto succeed. We evaluate the effectiveness of the ILR attack with real-world\nexperiments against two major traffic sign recognition architectures on four\nIR-sensitive cameras. Our black-box optimization methodology allows the attack\nto achieve up to a 100% attack success rate in indoor, static scenarios and a\n>80.5% attack success rate in our outdoor, moving vehicle scenarios. We find\nthe latest state-of-the-art certifiable defense is ineffective against ILR\nattacks as it mis-certifies >33.5% of cases. To address this, we propose a\ndetection strategy based on the physical properties of IR laser reflections\nwhich can detect 96% of ILR attacks.\n", "date": "2024-01-07", "categories": []}
{"id": "2208.12511", "title": "Lower Difficulty and Better Robustness: A Bregman Divergence Perspective\n  for Adversarial Training", "abstract": "  In this paper, we investigate on improving the adversarial robustness\nobtained in adversarial training (AT) via reducing the difficulty of\noptimization. To better study this problem, we build a novel Bregman divergence\nperspective for AT, in which AT can be viewed as the sliding process of the\ntraining data points on the negative entropy curve. Based on this perspective,\nwe analyze the learning objectives of two typical AT methods, i.e., PGD-AT and\nTRADES, and we find that the optimization process of TRADES is easier than\nPGD-AT for that TRADES separates PGD-AT. In addition, we discuss the function\nof entropy in TRADES, and we find that models with high entropy can be better\nrobustness learners. Inspired by the above findings, we propose two methods,\ni.e., FAIT and MER, which can both not only reduce the difficulty of\noptimization under the 10-step PGD adversaries, but also provide better\nrobustness. Our work suggests that reducing the difficulty of optimization\nunder the 10-step PGD adversaries is a promising approach for enhancing the\nadversarial robustness in AT.\n", "date": "2022-08-26", "categories": []}
{"id": "2401.02633", "title": "A Random Ensemble of Encrypted models for Enhancing Robustness against\n  Adversarial Examples", "abstract": "  Deep neural networks (DNNs) are well known to be vulnerable to adversarial\nexamples (AEs). In addition, AEs have adversarial transferability, which means\nAEs generated for a source model can fool another black-box model (target\nmodel) with a non-trivial probability. In previous studies, it was confirmed\nthat the vision transformer (ViT) is more robust against the property of\nadversarial transferability than convolutional neural network (CNN) models such\nas ConvMixer, and moreover encrypted ViT is more robust than ViT without any\nencryption. In this article, we propose a random ensemble of encrypted ViT\nmodels to achieve much more robust models. In experiments, the proposed scheme\nis verified to be more robust against not only black-box attacks but also\nwhite-box ones than convention methods.\n", "date": "2024-01-05", "categories": []}
{"id": "2401.02652", "title": "Adaptive Discounting of Training Time Attacks", "abstract": "  Among the most insidious attacks on Reinforcement Learning (RL) solutions are\ntraining-time attacks (TTAs) that create loopholes and backdoors in the learned\nbehaviour. Not limited to a simple disruption, constructive TTAs (C-TTAs) are\nnow available, where the attacker forces a specific, target behaviour upon a\ntraining RL agent (victim). However, even state-of-the-art C-TTAs focus on\ntarget behaviours that could be naturally adopted by the victim if not for a\nparticular feature of the environment dynamics, which C-TTAs exploit. In this\nwork, we show that a C-TTA is possible even when the target behaviour is\nun-adoptable due to both environment dynamics as well as non-optimality with\nrespect to the victim objective(s). To find efficient attacks in this context,\nwe develop a specialised flavour of the DDPG algorithm, which we term\ngammaDDPG, that learns this stronger version of C-TTA. gammaDDPG dynamically\nalters the attack policy planning horizon based on the victim's current\nbehaviour. This improves effort distribution throughout the attack timeline and\nreduces the effect of uncertainty the attacker has about the victim. To\ndemonstrate the features of our method and better relate the results to prior\nresearch, we borrow a 3D grid domain from a state-of-the-art C-TTA for our\nexperiments. Code is available at \"bit.ly/github-rb-gDDPG\".\n", "date": "2024-01-05", "categories": []}
{"id": "2302.09457", "title": "Attacks in Adversarial Machine Learning: A Systematic Survey from the\n  Life-cycle Perspective", "abstract": "  Adversarial machine learning (AML) studies the adversarial phenomenon of\nmachine learning, which may make inconsistent or unexpected predictions with\nhumans. Some paradigms have been recently developed to explore this adversarial\nphenomenon occurring at different stages of a machine learning system, such as\nbackdoor attack occurring at the pre-training, in-training and inference stage;\nweight attack occurring at the post-training, deployment and inference stage;\nadversarial attack occurring at the inference stage. However, although these\nadversarial paradigms share a common goal, their developments are almost\nindependent, and there is still no big picture of AML. In this work, we aim to\nprovide a unified perspective to the AML community to systematically review the\noverall progress of this field. We firstly provide a general definition about\nAML, and then propose a unified mathematical framework to covering existing\nattack paradigms. According to the proposed unified framework, we build a full\ntaxonomy to systematically categorize and review existing representative\nmethods for each paradigm. Besides, using this unified framework, it is easy to\nfigure out the connections and differences among different attack paradigms,\nwhich may inspire future researchers to develop more advanced attack paradigms.\nFinally, to facilitate the viewing of the built taxonomy and the related\nliterature in adversarial machine learning, we further provide a website, \\ie,\n\\url{http://adversarial-ml.com}, where the taxonomies and literature will be\ncontinuously updated.\n", "date": "2023-02-19", "categories": []}
{"id": "2311.09024", "title": "Fast Certification of Vision-Language Models Using Incremental\n  Randomized Smoothing", "abstract": "  A key benefit of deep vision-language models such as CLIP is that they enable\nzero-shot open vocabulary classification; the user has the ability to define\nnovel class labels via natural language prompts at inference time. However,\nwhile CLIP-based zero-shot classifiers have demonstrated competitive\nperformance across a range of domain shifts, they remain highly vulnerable to\nadversarial attacks. Therefore, ensuring the robustness of such models is\ncrucial for their reliable deployment in the wild.\n  In this work, we introduce Open Vocabulary Certification (OVC), a fast\ncertification method designed for open-vocabulary models like CLIP via\nrandomized smoothing techniques. Given a base \"training\" set of prompts and\ntheir corresponding certified CLIP classifiers, OVC relies on the observation\nthat a classifier with a novel prompt can be viewed as a perturbed version of\nnearby classifiers in the base training set. Therefore, OVC can rapidly certify\nthe novel classifier using a variation of incremental randomized smoothing. By\nusing a caching trick, we achieve approximately two orders of magnitude\nacceleration in the certification process for novel prompts. To achieve further\n(heuristic) speedups, OVC approximates the embedding space at a given input\nusing a multivariate normal distribution bypassing the need for sampling via\nforward passes through the vision backbone. We demonstrate the effectiveness of\nOVC on through experimental evaluation using multiple vision-language backbones\non the CIFAR-10 and ImageNet test datasets.\n", "date": "2023-11-15", "categories": []}
{"id": "2311.16124", "title": "DiffAttack: Evasion Attacks Against Diffusion-Based Adversarial\n  Purification", "abstract": "  Diffusion-based purification defenses leverage diffusion models to remove\ncrafted perturbations of adversarial examples and achieve state-of-the-art\nrobustness. Recent studies show that even advanced attacks cannot break such\ndefenses effectively, since the purification process induces an extremely deep\ncomputational graph which poses the potential problem of gradient obfuscation,\nhigh memory cost, and unbounded randomness. In this paper, we propose a unified\nframework DiffAttack to perform effective and efficient attacks against\ndiffusion-based purification defenses, including both DDPM and score-based\napproaches. In particular, we propose a deviated-reconstruction loss at\nintermediate diffusion steps to induce inaccurate density gradient estimation\nto tackle the problem of vanishing/exploding gradients. We also provide a\nsegment-wise forwarding-backwarding algorithm, which leads to memory-efficient\ngradient backpropagation. We validate the attack effectiveness of DiffAttack\ncompared with existing adaptive attacks on CIFAR-10 and ImageNet. We show that\nDiffAttack decreases the robust accuracy of models compared with SOTA attacks\nby over 20% on CIFAR-10 under $\\ell_\\infty$ attack $(\\epsilon=8/255)$, and over\n10% on ImageNet under $\\ell_\\infty$ attack $(\\epsilon=4/255)$. We conduct a\nseries of ablations studies, and we find 1) DiffAttack with the\ndeviated-reconstruction loss added over uniformly sampled time steps is more\neffective than that added over only initial/final steps, and 2) diffusion-based\npurification with a moderate diffusion length is more robust under DiffAttack.\n", "date": "2023-10-27", "categories": []}
{"id": "2312.15228", "title": "Adversarial Data Poisoning for Fake News Detection: How to Make a Model\n  Misclassify a Target News without Modifying It", "abstract": "  Fake news detection models are critical to countering disinformation but can\nbe manipulated through adversarial attacks. In this position paper, we analyze\nhow an attacker can compromise the performance of an online learning detector\non specific news content without being able to manipulate the original target\nnews. In some contexts, such as social networks, where the attacker cannot\nexert complete control over all the information, this scenario can indeed be\nquite plausible. Therefore, we show how an attacker could potentially introduce\npoisoning data into the training data to manipulate the behavior of an online\nlearning method. Our initial findings reveal varying susceptibility of logistic\nregression models based on complexity and attack type.\n", "date": "2023-12-23", "categories": []}
{"id": "2401.02031", "title": "Spy-Watermark: Robust Invisible Watermarking for Backdoor Attack", "abstract": "  Backdoor attack aims to deceive a victim model when facing backdoor instances\nwhile maintaining its performance on benign data. Current methods use manual\npatterns or special perturbations as triggers, while they often overlook the\nrobustness against data corruption, making backdoor attacks easy to defend in\npractice. To address this issue, we propose a novel backdoor attack method\nnamed Spy-Watermark, which remains effective when facing data collapse and\nbackdoor defense. Therein, we introduce a learnable watermark embedded in the\nlatent domain of images, serving as the trigger. Then, we search for a\nwatermark that can withstand collapse during image decoding, cooperating with\nseveral anti-collapse operations to further enhance the resilience of our\ntrigger against data corruption. Extensive experiments are conducted on\nCIFAR10, GTSRB, and ImageNet datasets, demonstrating that Spy-Watermark\novertakes ten state-of-the-art methods in terms of robustness and stealthiness.\n", "date": "2024-01-04", "categories": []}
{"id": "2401.02342", "title": "Evasive Hardware Trojan through Adversarial Power Trace", "abstract": "  The globalization of the Integrated Circuit (IC) supply chain, driven by\ntime-to-market and cost considerations, has made ICs vulnerable to hardware\nTrojans (HTs). Against this threat, a promising approach is to use Machine\nLearning (ML)-based side-channel analysis, which has the advantage of being a\nnon-intrusive method, along with efficiently detecting HTs under golden\nchip-free settings. In this paper, we question the trustworthiness of ML-based\nHT detection via side-channel analysis. We introduce a HT obfuscation (HTO)\napproach to allow HTs to bypass this detection method. Rather than\ntheoretically misleading the model by simulated adversarial traces, a key\naspect of our approach is the design and implementation of adversarial noise as\npart of the circuitry, alongside the HT. We detail HTO methodologies for ASICs\nand FPGAs, and evaluate our approach using TrustHub benchmark. Interestingly,\nwe found that HTO can be implemented with only a single transistor for ASIC\ndesigns to generate adversarial power traces that can fool the defense with\n100% efficiency. We also efficiently implemented our approach on a Spartan 6\nXilinx FPGA using 2 different variants: (i) DSP slices-based, and (ii)\nring-oscillator-based design. Additionally, we assess the efficiency of\ncountermeasures like spectral domain analysis, and we show that an adaptive\nattacker can still design evasive HTOs by constraining the design with a\nspectral noise budget. In addition, while adversarial training (AT) offers\nhigher protection against evasive HTs, AT models suffer from a considerable\nutility loss, potentially rendering them unsuitable for such security\napplication. We believe this research represents a significant step in\nunderstanding and exploiting ML vulnerabilities in a hardware security context,\nand we make all resources and designs openly available online:\nhttps://dev.d18uu4lqwhbmka.amplifyapp.com\n", "date": "2024-01-04", "categories": []}
{"id": "2312.13863", "title": "Manipulating Trajectory Prediction with Backdoors", "abstract": "  Autonomous vehicles ought to predict the surrounding agents' trajectories to\nallow safe maneuvers in uncertain and complex traffic situations. As companies\nincreasingly apply trajectory prediction in the real world, security becomes a\nrelevant concern. In this paper, we focus on backdoors - a security threat\nacknowledged in other fields but so far overlooked for trajectory prediction.\nTo this end, we describe and investigate four triggers that could affect\ntrajectory prediction. We then show that these triggers (for example, a braking\nvehicle), when correlated with a desired output (for example, a curve) during\ntraining, cause the desired output of a state-of-the-art trajectory prediction\nmodel. In other words, the model has good benign performance but is vulnerable\nto backdoors. This is the case even if the trigger maneuver is performed by a\nnon-casual agent behind the target vehicle. As a side-effect, our analysis\nreveals interesting limitations within trajectory prediction models. Finally,\nwe evaluate a range of defenses against backdoors. While some, like simple\noffroad checks, do not enable detection for all triggers, clustering is a\npromising candidate to support manual inspection to find backdoors.\n", "date": "2023-12-21", "categories": []}
{"id": "2401.01575", "title": "Enhancing Generalization of Invisible Facial Privacy Cloak via Gradient\n  Accumulation", "abstract": "  The blooming of social media and face recognition (FR) systems has increased\npeople's concern about privacy and security. A new type of adversarial privacy\ncloak (class-universal) can be applied to all the images of regular users, to\nprevent malicious FR systems from acquiring their identity information. In this\nwork, we discover the optimization dilemma in the existing methods -- the local\noptima problem in large-batch optimization and the gradient information\nelimination problem in small-batch optimization. To solve these problems, we\npropose Gradient Accumulation (GA) to aggregate multiple small-batch gradients\ninto a one-step iterative gradient to enhance the gradient stability and reduce\nthe usage of quantization operations. Experiments show that our proposed method\nachieves high performance on the Privacy-Commons dataset against black-box face\nrecognition models.\n", "date": "2024-01-03", "categories": []}
{"id": "2211.16808", "title": "Efficient Adversarial Input Generation via Neural Net Patching", "abstract": "  The generation of adversarial inputs has become a crucial issue in\nestablishing the robustness and trustworthiness of deep neural nets, especially\nwhen they are used in safety-critical application domains such as autonomous\nvehicles and precision medicine. However, the problem poses multiple practical\nchallenges, including scalability issues owing to large-sized networks, and the\ngeneration of adversarial inputs that lack important qualities such as\nnaturalness and output-impartiality. This problem shares its end goal with the\ntask of patching neural nets where small changes in some of the network's\nweights need to be discovered so that upon applying these changes, the modified\nnet produces the desirable output for a given set of inputs. We exploit this\nconnection by proposing to obtain an adversarial input from a patch, with the\nunderlying observation that the effect of changing the weights can also be\nbrought about by changing the inputs instead. Thus, this paper presents a novel\nway to generate input perturbations that are adversarial for a given network by\nusing an efficient network patching technique. We note that the proposed method\nis significantly more effective than the prior state-of-the-art techniques.\n", "date": "2022-11-30", "categories": []}
{"id": "2309.11766", "title": "Dictionary Attack on IMU-based Gait Authentication", "abstract": "  We present a novel adversarial model for authentication systems that use gait\npatterns recorded by the inertial measurement unit (IMU) built into\nsmartphones. The attack idea is inspired by and named after the concept of a\ndictionary attack on knowledge (PIN or password) based authentication systems.\nIn particular, this work investigates whether it is possible to build a\ndictionary of IMUGait patterns and use it to launch an attack or find an\nimitator who can actively reproduce IMUGait patterns that match the target's\nIMUGait pattern. Nine physically and demographically diverse individuals walked\nat various levels of four predefined controllable and adaptable gait factors\n(speed, step length, step width, and thigh-lift), producing 178 unique IMUGait\npatterns. Each pattern attacked a wide variety of user authentication models.\nThe deeper analysis of error rates (before and after the attack) challenges the\nbelief that authentication systems based on IMUGait patterns are the most\ndifficult to spoof; further research is needed on adversarial models and\nassociated countermeasures.\n", "date": "2023-09-21", "categories": []}
{"id": "2306.06874", "title": "VillanDiffusion: A Unified Backdoor Attack Framework for Diffusion\n  Models", "abstract": "  Diffusion Models (DMs) are state-of-the-art generative models that learn a\nreversible corruption process from iterative noise addition and denoising. They\nare the backbone of many generative AI applications, such as text-to-image\nconditional generation. However, recent studies have shown that basic\nunconditional DMs (e.g., DDPM and DDIM) are vulnerable to backdoor injection, a\ntype of output manipulation attack triggered by a maliciously embedded pattern\nat model input. This paper presents a unified backdoor attack framework\n(VillanDiffusion) to expand the current scope of backdoor analysis for DMs. Our\nframework covers mainstream unconditional and conditional DMs (denoising-based\nand score-based) and various training-free samplers for holistic evaluations.\nExperiments show that our unified framework facilitates the backdoor analysis\nof different DM configurations and provides new insights into caption-based\nbackdoor attacks on DMs. Our code is available on GitHub:\n\\url{https://github.com/IBM/villandiffusion}\n", "date": "2023-06-12", "categories": []}
{"id": "2312.17301", "title": "Explainability-Based Adversarial Attack on Graphs Through Edge\n  Perturbation", "abstract": "  Despite the success of graph neural networks (GNNs) in various domains, they\nexhibit susceptibility to adversarial attacks. Understanding these\nvulnerabilities is crucial for developing robust and secure applications. In\nthis paper, we investigate the impact of test time adversarial attacks through\nedge perturbations which involve both edge insertions and deletions. A novel\nexplainability-based method is proposed to identify important nodes in the\ngraph and perform edge perturbation between these nodes. The proposed method is\ntested for node classification with three different architectures and datasets.\nThe results suggest that introducing edges between nodes of different classes\nhas higher impact as compared to removing edges among nodes within the same\nclass.\n", "date": "2023-12-28", "categories": []}
{"id": "2301.02288", "title": "gRoMA: a Tool for Measuring the Global Robustness of Deep Neural\n  Networks", "abstract": "  Deep neural networks (DNNs) are at the forefront of cutting-edge technology,\nand have been achieving remarkable performance in a variety of complex tasks.\nNevertheless, their integration into safety-critical systems, such as in the\naerospace or automotive domains, poses a significant challenge due to the\nthreat of adversarial inputs: perturbations in inputs that might cause the DNN\nto make grievous mistakes. Multiple studies have demonstrated that even modern\nDNNs are susceptible to adversarial inputs, and this risk must thus be measured\nand mitigated to allow the deployment of DNNs in critical settings. Here, we\npresent gRoMA (global Robustness Measurement and Assessment), an innovative and\nscalable tool that implements a probabilistic approach to measure the global\ncategorial robustness of a DNN. Specifically, gRoMA measures the probability of\nencountering adversarial inputs for a specific output category. Our tool\noperates on pre-trained, black-box classification DNNs, and generates input\nsamples belonging to an output category of interest. It measures the DNN's\nsusceptibility to adversarial inputs around these inputs, and aggregates the\nresults to infer the overall global categorial robustness of the DNN up to some\nsmall bounded statistical error.\n  We evaluate our tool on the popular Densenet DNN model over the CIFAR10\ndataset. Our results reveal significant gaps in the robustness of the different\noutput categories. This experiment demonstrates the usefulness and scalability\nof our approach and its potential for allowing DNNs to be deployed within\ncritical systems of interest.\n", "date": "2023-01-05", "categories": []}
{"id": "2303.01234", "title": "Frauds Bargain Attack: Generating Adversarial Text Samples via Word\n  Manipulation Process", "abstract": "  Recent research has revealed that natural language processing (NLP) models\nare vulnerable to adversarial examples. However, the current techniques for\ngenerating such examples rely on deterministic heuristic rules, which fail to\nproduce optimal adversarial examples. In response, this study proposes a new\nmethod called the Fraud's Bargain Attack (FBA), which uses a randomization\nmechanism to expand the search space and produce high-quality adversarial\nexamples with a higher probability of success. FBA uses the Metropolis-Hasting\nsampler, a type of Markov Chain Monte Carlo sampler, to improve the selection\nof adversarial examples from all candidates generated by a customized\nstochastic process called the Word Manipulation Process (WMP). The WMP method\nmodifies individual words in a contextually-aware manner through insertion,\nremoval, or substitution. Through extensive experiments, this study\ndemonstrates that FBA outperforms other methods in terms of attack success\nrate, imperceptibility and sentence quality.\n", "date": "2023-03-01", "categories": []}
{"id": "2311.12914", "title": "Attention Deficit is Ordered! Fooling Deformable Vision Transformers\n  with Collaborative Adversarial Patches", "abstract": "  The latest generation of transformer-based vision models has proven to be\nsuperior to Convolutional Neural Network (CNN)-based models across several\nvision tasks, largely attributed to their remarkable prowess in relation\nmodeling. Deformable vision transformers significantly reduce the quadratic\ncomplexity of attention modeling by using sparse attention structures, enabling\nthem to incorporate features across different scales and be used in large-scale\napplications, such as multi-view vision systems. Recent work has demonstrated\nadversarial attacks against conventional vision transformers; we show that\nthese attacks do not transfer to deformable transformers due to their sparse\nattention structure. Specifically, attention in deformable transformers is\nmodeled using pointers to the most relevant other tokens. In this work, we\ncontribute for the first time adversarial attacks that manipulate the attention\nof deformable transformers, redirecting it to focus on irrelevant parts of the\nimage. We also develop new collaborative attacks where a source patch\nmanipulates attention to point to a target patch, which contains the\nadversarial noise to fool the model. In our experiments, we observe that\naltering less than 1% of the patched area in the input field results in a\ncomplete drop to 0% AP in single-view object detection using MS COCO and a 0%\nMODA in multi-view object detection using Wildtrack.\n", "date": "2023-11-21", "categories": []}
{"id": "2312.11285", "title": "Adv-Diffusion: Imperceptible Adversarial Face Identity Attack via Latent\n  Diffusion Model", "abstract": "  Adversarial attacks involve adding perturbations to the source image to cause\nmisclassification by the target model, which demonstrates the potential of\nattacking face recognition models. Existing adversarial face image generation\nmethods still can't achieve satisfactory performance because of low\ntransferability and high detectability. In this paper, we propose a unified\nframework Adv-Diffusion that can generate imperceptible adversarial identity\nperturbations in the latent space but not the raw pixel space, which utilizes\nstrong inpainting capabilities of the latent diffusion model to generate\nrealistic adversarial images. Specifically, we propose the identity-sensitive\nconditioned diffusion generative model to generate semantic perturbations in\nthe surroundings. The designed adaptive strength-based adversarial perturbation\nalgorithm can ensure both attack transferability and stealthiness. Extensive\nqualitative and quantitative experiments on the public FFHQ and CelebA-HQ\ndatasets prove the proposed method achieves superior performance compared with\nthe state-of-the-art methods without an extra generative model training\nprocess. The source code is available at\nhttps://github.com/kopper-xdu/Adv-Diffusion.\n", "date": "2023-12-18", "categories": []}
{"id": "2312.16339", "title": "Universal Pyramid Adversarial Training for Improved ViT Performance", "abstract": "  Recently, Pyramid Adversarial training (Herrmann et al., 2022) has been shown\nto be very effective for improving clean accuracy and distribution-shift\nrobustness of vision transformers. However, due to the iterative nature of\nadversarial training, the technique is up to 7 times more expensive than\nstandard training. To make the method more efficient, we propose Universal\nPyramid Adversarial training, where we learn a single pyramid adversarial\npattern shared across the whole dataset instead of the sample-wise patterns.\nWith our proposed technique, we decrease the computational cost of Pyramid\nAdversarial training by up to 70% while retaining the majority of its benefit\non clean performance and distribution-shift robustness. In addition, to the\nbest of our knowledge, we are also the first to find that universal adversarial\ntraining can be leveraged to improve clean model performance.\n", "date": "2023-12-26", "categories": []}
{"id": "2312.16401", "title": "Natural Adversarial Patch Generation Method Based on Latent Diffusion\n  Model", "abstract": "  Recently, some research show that deep neural networks are vulnerable to the\nadversarial attacks, the well-trainned samples or patches could be used to\ntrick the neural network detector or human visual perception. However, these\nadversarial patches, with their conspicuous and unusual patterns, lack\ncamouflage and can easily raise suspicion in the real world. To solve this\nproblem, this paper proposed a novel adversarial patch method called the Latent\nDiffusion Patch (LDP), in which, a pretrained encoder is first designed to\ncompress the natural images into a feature space with key characteristics. Then\ntrains the diffusion model using the above feature space. Finally, explore the\nlatent space of the pretrained diffusion model using the image denoising\ntechnology. It polishes the patches and images through the powerful natural\nabilities of diffusion models, making them more acceptable to the human visual\nsystem. Experimental results, both digital and physical worlds, show that LDPs\nachieve a visual subjectivity score of 87.3%, while still maintaining effective\nattack capabilities.\n", "date": "2023-12-27", "categories": []}
{"id": "2312.16880", "title": "Adversarial Attacks on Image Classification Models: Analysis and Defense", "abstract": "  The notion of adversarial attacks on image classification models based on\nconvolutional neural networks (CNN) is introduced in this work. To classify\nimages, deep learning models called CNNs are frequently used. However, when the\nnetworks are subject to adversarial attacks, extremely potent and previously\ntrained CNN models that perform quite effectively on image datasets for image\nclassification tasks may perform poorly. In this work, one well-known\nadversarial attack known as the fast gradient sign method (FGSM) is explored\nand its adverse effects on the performances of image classification models are\nexamined. The FGSM attack is simulated on three pre-trained image classifier\nCNN architectures, ResNet-101, AlexNet, and RegNetY 400MF using randomly chosen\nimages from the ImageNet dataset. The classification accuracies of the models\nare computed in the absence and presence of the attack to demonstrate the\ndetrimental effect of the attack on the performances of the classifiers.\nFinally, a mechanism is proposed to defend against the FGSM attack based on a\nmodified defensive distillation-based approach. Extensive results are presented\nfor the validation of the proposed scheme.\n", "date": "2023-12-28", "categories": []}
{"id": "2312.16957", "title": "Attack Tree Analysis for Adversarial Evasion Attacks", "abstract": "  Recently, the evolution of deep learning has promoted the application of\nmachine learning (ML) to various systems. However, there are ML systems, such\nas autonomous vehicles, that cause critical damage when they misclassify.\nConversely, there are ML-specific attacks called adversarial attacks based on\nthe characteristics of ML systems. For example, one type of adversarial attack\nis an evasion attack, which uses minute perturbations called \"adversarial\nexamples\" to intentionally misclassify classifiers. Therefore, it is necessary\nto analyze the risk of ML-specific attacks in introducing ML base systems. In\nthis study, we propose a quantitative evaluation method for analyzing the risk\nof evasion attacks using attack trees. The proposed method consists of the\nextension of the conventional attack tree to analyze evasion attacks and the\nsystematic construction method of the extension. In the extension of the\nconventional attack tree, we introduce ML and conventional attack nodes to\nrepresent various characteristics of evasion attacks. In the systematic\nconstruction process, we propose a procedure to construct the attack tree. The\nprocedure consists of three steps: (1) organizing information about attack\nmethods in the literature to a matrix, (2) identifying evasion attack scenarios\nfrom methods in the matrix, and (3) constructing the attack tree from the\nidentified scenarios using a pattern. Finally, we conducted experiments on\nthree ML image recognition systems to demonstrate the versatility and\neffectiveness of our proposed method.\n", "date": "2023-12-28", "categories": []}
{"id": "2312.06199", "title": "Towards Transferable Adversarial Attacks with Centralized Perturbation", "abstract": "  Adversarial transferability enables black-box attacks on unknown victim deep\nneural networks (DNNs), rendering attacks viable in real-world scenarios.\nCurrent transferable attacks create adversarial perturbation over the entire\nimage, resulting in excessive noise that overfit the source model.\nConcentrating perturbation to dominant image regions that are model-agnostic is\ncrucial to improving adversarial efficacy. However, limiting perturbation to\nlocal regions in the spatial domain proves inadequate in augmenting\ntransferability. To this end, we propose a transferable adversarial attack with\nfine-grained perturbation optimization in the frequency domain, creating\ncentralized perturbation. We devise a systematic pipeline to dynamically\nconstrain perturbation optimization to dominant frequency coefficients. The\nconstraint is optimized in parallel at each iteration, ensuring the directional\nalignment of perturbation optimization with model prediction. Our approach\nallows us to centralize perturbation towards sample-specific important\nfrequency features, which are shared by DNNs, effectively mitigating source\nmodel overfitting. Experiments demonstrate that by dynamically centralizing\nperturbation on dominating frequency coefficients, crafted adversarial examples\nexhibit stronger transferability, and allowing them to bypass various defenses.\n", "date": "2023-12-11", "categories": []}
{"id": "2312.15172", "title": "Pre-trained Trojan Attacks for Visual Recognition", "abstract": "  Pre-trained vision models (PVMs) have become a dominant component due to\ntheir exceptional performance when fine-tuned for downstream tasks. However,\nthe presence of backdoors within PVMs poses significant threats. Unfortunately,\nexisting studies primarily focus on backdooring PVMs for the classification\ntask, neglecting potential inherited backdoors in downstream tasks such as\ndetection and segmentation. In this paper, we propose the Pre-trained Trojan\nattack, which embeds backdoors into a PVM, enabling attacks across various\ndownstream vision tasks. We highlight the challenges posed by cross-task\nactivation and shortcut connections in successful backdoor attacks. To achieve\neffective trigger activation in diverse tasks, we stylize the backdoor trigger\npatterns with class-specific textures, enhancing the recognition of\ntask-irrelevant low-level features associated with the target class in the\ntrigger pattern. Moreover, we address the issue of shortcut connections by\nintroducing a context-free learning pipeline for poison training. In this\napproach, triggers without contextual backgrounds are directly utilized as\ntraining data, diverging from the conventional use of clean images.\nConsequently, we establish a direct shortcut from the trigger to the target\nclass, mitigating the shortcut connection issue. We conducted extensive\nexperiments to thoroughly validate the effectiveness of our attacks on\ndownstream detection and segmentation tasks. Additionally, we showcase the\npotential of our approach in more practical scenarios, including large vision\nmodels and 3D object detection in autonomous driving. This paper aims to raise\nawareness of the potential threats associated with applying PVMs in practical\nscenarios. Our codes will be available upon paper publication.\n", "date": "2023-12-23", "categories": []}
{"id": "2312.14217", "title": "Adversarial Infrared Curves: An Attack on Infrared Pedestrian Detectors\n  in the Physical World", "abstract": "  Deep neural network security is a persistent concern, with considerable\nresearch on visible light physical attacks but limited exploration in the\ninfrared domain. Existing approaches, like white-box infrared attacks using\nbulb boards and QR suits, lack realism and stealthiness. Meanwhile, black-box\nmethods with cold and hot patches often struggle to ensure robustness. To\nbridge these gaps, we propose Adversarial Infrared Curves (AdvIC). Using\nParticle Swarm Optimization, we optimize two Bezier curves and employ cold\npatches in the physical realm to introduce perturbations, creating infrared\ncurve patterns for physical sample generation. Our extensive experiments\nconfirm AdvIC's effectiveness, achieving 94.8\\% and 67.2\\% attack success rates\nfor digital and physical attacks, respectively. Stealthiness is demonstrated\nthrough a comparative analysis, and robustness assessments reveal AdvIC's\nsuperiority over baseline methods. When deployed against diverse advanced\ndetectors, AdvIC achieves an average attack success rate of 76.8\\%, emphasizing\nits robust nature. we explore adversarial defense strategies against AdvIC and\nexamine its impact under various defense mechanisms. Given AdvIC's substantial\nsecurity implications for real-world vision-based applications, urgent\nattention and mitigation efforts are warranted.\n", "date": "2023-12-21", "categories": []}
{"id": "2312.14218", "title": "AutoAugment Input Transformation for Highly Transferable Targeted\n  Attacks", "abstract": "  Deep Neural Networks (DNNs) are widely acknowledged to be susceptible to\nadversarial examples, wherein imperceptible perturbations are added to clean\nexamples through diverse input transformation attacks. However, these methods\noriginally designed for non-targeted attacks exhibit low success rates in\ntargeted attacks. Recent targeted adversarial attacks mainly pay attention to\ngradient optimization, attempting to find the suitable perturbation direction.\nHowever, few of them are dedicated to input transformation.In this work, we\nobserve a positive correlation between the logit/probability of the target\nclass and diverse input transformation methods in targeted attacks. To this\nend, we propose a novel targeted adversarial attack called AutoAugment Input\nTransformation (AAIT). Instead of relying on hand-made strategies, AAIT\nsearches for the optimal transformation policy from a transformation space\ncomprising various operations. Then, AAIT crafts adversarial examples using the\nfound optimal transformation policy to boost the adversarial transferability in\ntargeted attacks. Extensive experiments conducted on CIFAR-10 and\nImageNet-Compatible datasets demonstrate that the proposed AAIT surpasses other\ntransfer-based targeted attacks significantly.\n", "date": "2023-12-21", "categories": []}
{"id": "2312.14407", "title": "AdvCloak: Customized Adversarial Cloak for Privacy Protection", "abstract": "  With extensive face images being shared on social media, there has been a\nnotable escalation in privacy concerns. In this paper, we propose AdvCloak, an\ninnovative framework for privacy protection using generative models. AdvCloak\nis designed to automatically customize class-wise adversarial masks that can\nmaintain superior image-level naturalness while providing enhanced\nfeature-level generalization ability. Specifically, AdvCloak sequentially\noptimizes the generative adversarial networks by employing a two-stage training\nstrategy. This strategy initially focuses on adapting the masks to the unique\nindividual faces via image-specific training and then enhances their\nfeature-level generalization ability to diverse facial variations of\nindividuals via person-specific training. To fully utilize the limited training\ndata, we combine AdvCloak with several general geometric modeling methods, to\nbetter describe the feature subspace of source identities. Extensive\nquantitative and qualitative evaluations on both common and celebrity datasets\ndemonstrate that AdvCloak outperforms existing state-of-the-art methods in\nterms of efficiency and effectiveness.\n", "date": "2023-12-22", "categories": []}
{"id": "2312.14677", "title": "MEAOD: Model Extraction Attack against Object Detectors", "abstract": "  The widespread use of deep learning technology across various industries has\nmade deep neural network models highly valuable and, as a result, attractive\ntargets for potential attackers. Model extraction attacks, particularly\nquery-based model extraction attacks, allow attackers to replicate a substitute\nmodel with comparable functionality to the victim model and present a\nsignificant threat to the confidentiality and security of MLaaS platforms.\nWhile many studies have explored threats of model extraction attacks against\nclassification models in recent years, object detection models, which are more\nfrequently used in real-world scenarios, have received less attention. In this\npaper, we investigate the challenges and feasibility of query-based model\nextraction attacks against object detection models and propose an effective\nattack method called MEAOD. It selects samples from the attacker-possessed\ndataset to construct an efficient query dataset using active learning and\nenhances the categories with insufficient objects. We additionally improve the\nextraction effectiveness by updating the annotations of the query dataset.\nAccording to our gray-box and black-box scenarios experiments, we achieve an\nextraction performance of over 70% under the given condition of a 10k query\nbudget.\n", "date": "2023-12-22", "categories": []}
{"id": "2210.14404", "title": "Adversarial Purification with the Manifold Hypothesis", "abstract": "  In this work, we formulate a novel framework for adversarial robustness using\nthe manifold hypothesis. This framework provides sufficient conditions for\ndefending against adversarial examples. We develop an adversarial purification\nmethod with this framework. Our method combines manifold learning with\nvariational inference to provide adversarial robustness without the need for\nexpensive adversarial training. Experimentally, our approach can provide\nadversarial robustness even if attackers are aware of the existence of the\ndefense. In addition, our method can also serve as a test-time defense\nmechanism for variational autoencoders.\n", "date": "2022-10-26", "categories": []}
{"id": "2307.15043", "title": "Universal and Transferable Adversarial Attacks on Aligned Language\n  Models", "abstract": "  Because \"out-of-the-box\" large language models are capable of generating a\ngreat deal of objectionable content, recent work has focused on aligning these\nmodels in an attempt to prevent undesirable generation. While there has been\nsome success at circumventing these measures -- so-called \"jailbreaks\" against\nLLMs -- these attacks have required significant human ingenuity and are brittle\nin practice. In this paper, we propose a simple and effective attack method\nthat causes aligned language models to generate objectionable behaviors.\nSpecifically, our approach finds a suffix that, when attached to a wide range\nof queries for an LLM to produce objectionable content, aims to maximize the\nprobability that the model produces an affirmative response (rather than\nrefusing to answer). However, instead of relying on manual engineering, our\napproach automatically produces these adversarial suffixes by a combination of\ngreedy and gradient-based search techniques, and also improves over past\nautomatic prompt generation methods.\n  Surprisingly, we find that the adversarial prompts generated by our approach\nare quite transferable, including to black-box, publicly released LLMs.\nSpecifically, we train an adversarial attack suffix on multiple prompts (i.e.,\nqueries asking for many different types of objectionable content), as well as\nmultiple models (in our case, Vicuna-7B and 13B). When doing so, the resulting\nattack suffix is able to induce objectionable content in the public interfaces\nto ChatGPT, Bard, and Claude, as well as open source LLMs such as LLaMA-2-Chat,\nPythia, Falcon, and others. In total, this work significantly advances the\nstate-of-the-art in adversarial attacks against aligned language models,\nraising important questions about how such systems can be prevented from\nproducing objectionable information. Code is available at\ngithub.com/llm-attacks/llm-attacks.\n", "date": "2023-07-27", "categories": []}
{"id": "2311.17458", "title": "Quantum Neural Networks under Depolarization Noise: Exploring White-Box\n  Attacks and Defenses", "abstract": "  Leveraging the unique properties of quantum mechanics, Quantum Machine\nLearning (QML) promises computational breakthroughs and enriched perspectives\nwhere traditional systems reach their boundaries. However, similarly to\nclassical machine learning, QML is not immune to adversarial attacks. Quantum\nadversarial machine learning has become instrumental in highlighting the weak\npoints of QML models when faced with adversarial crafted feature vectors.\nDiving deep into this domain, our exploration shines light on the interplay\nbetween depolarization noise and adversarial robustness. While previous results\nenhanced robustness from adversarial threats through depolarization noise, our\nfindings paint a different picture. Interestingly, adding depolarization noise\ndiscontinued the effect of providing further robustness for a multi-class\nclassification scenario. Consolidating our findings, we conducted experiments\nwith a multi-class classifier adversarially trained on gate-based quantum\nsimulators, further elucidating this unexpected behavior.\n", "date": "2023-11-29", "categories": []}
{"id": "2312.13575", "title": "ARBiBench: Benchmarking Adversarial Robustness of Binarized Neural\n  Networks", "abstract": "  Network binarization exhibits great potential for deployment on\nresource-constrained devices due to its low computational cost. Despite the\ncritical importance, the security of binarized neural networks (BNNs) is rarely\ninvestigated. In this paper, we present ARBiBench, a comprehensive benchmark to\nevaluate the robustness of BNNs against adversarial perturbations on CIFAR-10\nand ImageNet. We first evaluate the robustness of seven influential BNNs on\nvarious white-box and black-box attacks. The results reveal that 1) The\nadversarial robustness of BNNs exhibits a completely opposite performance on\nthe two datasets under white-box attacks. 2) BNNs consistently exhibit better\nadversarial robustness under black-box attacks. 3) Different BNNs exhibit\ncertain similarities in their robustness performance. Then, we conduct\nexperiments to analyze the adversarial robustness of BNNs based on these\ninsights. Our research contributes to inspiring future research on enhancing\nthe robustness of BNNs and advancing their application in real-world scenarios.\n", "date": "2023-12-21", "categories": []}
{"id": "2303.06854", "title": "Robust Contrastive Language-Image Pre-training against Data Poisoning\n  and Backdoor Attacks", "abstract": "  Contrastive vision-language representation learning has achieved\nstate-of-the-art performance for zero-shot classification, by learning from\nmillions of image-caption pairs crawled from the internet. However, the massive\ndata that powers large multimodal models such as CLIP, makes them extremely\nvulnerable to various types of targeted data poisoning and backdoor attacks.\nDespite this vulnerability, robust contrastive vision-language pre-training\nagainst such attacks has remained unaddressed. In this work, we propose ROCLIP,\nthe first effective method for robust pre-training multimodal vision-language\nmodels against targeted data poisoning and backdoor attacks. ROCLIP effectively\nbreaks the association between poisoned image-caption pairs by considering a\nrelatively large and varying pool of random captions, and matching every image\nwith the text that is most similar to it in the pool instead of its own\ncaption, every few epochs.It also leverages image and text augmentations to\nfurther strengthen the defense and improve the performance of the model. Our\nextensive experiments show that ROCLIP renders state-of-the-art targeted data\npoisoning and backdoor attacks ineffective during pre-training CLIP models. In\nparticular, ROCLIP decreases the success rate for targeted data poisoning\nattacks from 93.75% to 12.5% and that of backdoor attacks down to 0%, while\nimproving the model's linear probe performance by 10% and maintains a similar\nzero shot performance compared to CLIP. By increasing the frequency of\nmatching, ROCLIP is able to defend strong attacks, which add up to 1% poisoned\nexamples to the data, and successfully maintain a low attack success rate of\n12.5%, while trading off the performance on some tasks.\n", "date": "2023-03-13", "categories": []}
{"id": "2305.10701", "title": "Personalization as a Shortcut for Few-Shot Backdoor Attack against\n  Text-to-Image Diffusion Models", "abstract": "  Although recent personalization methods have democratized high-resolution\nimage synthesis by enabling swift concept acquisition with minimal examples and\nlightweight computation, they also present an exploitable avenue for high\naccessible backdoor attacks. This paper investigates a critical and unexplored\naspect of text-to-image (T2I) diffusion models - their potential vulnerability\nto backdoor attacks via personalization. Our study focuses on a zero-day\nbackdoor vulnerability prevalent in two families of personalization methods,\nepitomized by Textual Inversion and DreamBooth.Compared to traditional backdoor\nattacks, our proposed method can facilitate more precise, efficient, and easily\naccessible attacks with a lower barrier to entry. We provide a comprehensive\nreview of personalization in T2I diffusion models, highlighting the operation\nand exploitation potential of this backdoor vulnerability. To be specific, by\nstudying the prompt processing of Textual Inversion and DreamBooth, we have\ndevised dedicated backdoor attacks according to the different ways of dealing\nwith unseen tokens and analyzed the influence of triggers and concept images on\nthe attack effect. Through comprehensive empirical study, we endorse the\nutilization of the nouveau-token backdoor attack due to its impressive\neffectiveness, stealthiness, and integrity, markedly outperforming the\nlegacy-token backdoor attack.\n", "date": "2023-05-18", "categories": []}
{"id": "2308.03108", "title": "SAAM: Stealthy Adversarial Attack on Monocular Depth Estimation", "abstract": "  In this paper, we investigate the vulnerability of MDE to adversarial\npatches. We propose a novel \\underline{S}tealthy \\underline{A}dversarial\n\\underline{A}ttacks on \\underline{M}DE (SAAM) that compromises MDE by either\ncorrupting the estimated distance or causing an object to seamlessly blend into\nits surroundings. Our experiments, demonstrate that the designed stealthy patch\nsuccessfully causes a DNN-based MDE to misestimate the depth of objects. In\nfact, our proposed adversarial patch achieves a significant 60\\% depth error\nwith 99\\% ratio of the affected region. Importantly, despite its adversarial\nnature, the patch maintains a naturalistic appearance, making it inconspicuous\nto human observers. We believe that this work sheds light on the threat of\nadversarial attacks in the context of MDE on edge devices. We hope it raises\nawareness within the community about the potential real-life harm of such\nattacks and encourages further research into developing more robust and\nadaptive defense mechanisms.\n", "date": "2023-08-06", "categories": []}
{"id": "2312.07392", "title": "ReRoGCRL: Representation-based Robustness in Goal-Conditioned\n  Reinforcement Learning", "abstract": "  While Goal-Conditioned Reinforcement Learning (GCRL) has gained attention,\nits algorithmic robustness against adversarial perturbations remains\nunexplored. The attacks and robust representation training methods that are\ndesigned for traditional RL become less effective when applied to GCRL. To\naddress this challenge, we first propose the Semi-Contrastive Representation\nattack, a novel approach inspired by the adversarial contrastive attack. Unlike\nexisting attacks in RL, it only necessitates information from the policy\nfunction and can be seamlessly implemented during deployment. Then, to mitigate\nthe vulnerability of existing GCRL algorithms, we introduce Adversarial\nRepresentation Tactics, which combines Semi-Contrastive Adversarial\nAugmentation with Sensitivity-Aware Regularizer to improve the adversarial\nrobustness of the underlying RL agent against various types of perturbations.\nExtensive experiments validate the superior performance of our attack and\ndefence methods across multiple state-of-the-art GCRL algorithms. Our tool\nReRoGCRL is available at https://github.com/TrustAI/ReRoGCRL.\n", "date": "2023-12-12", "categories": []}
{"id": "2312.12768", "title": "Mutual-modality Adversarial Attack with Semantic Perturbation", "abstract": "  Adversarial attacks constitute a notable threat to machine learning systems,\ngiven their potential to induce erroneous predictions and classifications.\nHowever, within real-world contexts, the essential specifics of the deployed\nmodel are frequently treated as a black box, consequently mitigating the\nvulnerability to such attacks. Thus, enhancing the transferability of the\nadversarial samples has become a crucial area of research, which heavily relies\non selecting appropriate surrogate models. To address this challenge, we\npropose a novel approach that generates adversarial attacks in a\nmutual-modality optimization scheme. Our approach is accomplished by leveraging\nthe pre-trained CLIP model. Firstly, we conduct a visual attack on the clean\nimage that causes semantic perturbations on the aligned embedding space with\nthe other textual modality. Then, we apply the corresponding defense on the\ntextual modality by updating the prompts, which forces the re-matching on the\nperturbed embedding space. Finally, to enhance the attack transferability, we\nutilize the iterative training strategy on the visual attack and the textual\ndefense, where the two processes optimize from each other. We evaluate our\napproach on several benchmark datasets and demonstrate that our mutual-modal\nattack strategy can effectively produce high-transferable attacks, which are\nstable regardless of the target networks. Our approach outperforms\nstate-of-the-art attack methods and can be readily deployed as a plug-and-play\nsolution.\n", "date": "2023-12-20", "categories": []}
{"id": "2312.12904", "title": "PGN: A perturbation generation network against deep reinforcement\n  learning", "abstract": "  Deep reinforcement learning has advanced greatly and applied in many areas.\nIn this paper, we explore the vulnerability of deep reinforcement learning by\nproposing a novel generative model for creating effective adversarial examples\nto attack the agent. Our proposed model can achieve both targeted attacks and\nuntargeted attacks. Considering the specificity of deep reinforcement learning,\nwe propose the action consistency ratio as a measure of stealthiness, and a new\nmeasurement index of effectiveness and stealthiness. Experiment results show\nthat our method can ensure the effectiveness and stealthiness of attack\ncompared with other algorithms. Moreover, our methods are considerably faster\nand thus can achieve rapid and efficient verification of the vulnerability of\ndeep reinforcement learning.\n", "date": "2023-12-20", "categories": []}
{"id": "2312.13131", "title": "Scaling Compute Is Not All You Need for Adversarial Robustness", "abstract": "  The last six years have witnessed significant progress in adversarially\nrobust deep learning. As evidenced by the CIFAR-10 dataset category in\nRobustBench benchmark, the accuracy under $\\ell_\\infty$ adversarial\nperturbations improved from 44\\% in \\citet{Madry2018Towards} to 71\\% in\n\\citet{peng2023robust}. Although impressive, existing state-of-the-art is still\nfar from satisfactory. It is further observed that best-performing models are\noften very large models adversarially trained by industrial labs with\nsignificant computational budgets. In this paper, we aim to understand: ``how\nmuch longer can computing power drive adversarial robustness advances?\" To\nanswer this question, we derive \\emph{scaling laws for adversarial robustness}\nwhich can be extrapolated in the future to provide an estimate of how much cost\nwe would need to pay to reach a desired level of robustness. We show that\nincreasing the FLOPs needed for adversarial training does not bring as much\nadvantage as it does for standard training in terms of performance\nimprovements. Moreover, we find that some of the top-performing techniques are\ndifficult to exactly reproduce, suggesting that they are not robust enough for\nminor changes in the training setup. Our analysis also uncovers potentially\nworthwhile directions to pursue in future research. Finally, we make our\nbenchmarking framework (built on top of \\texttt{timm}~\\citep{rw2019timm})\npublicly available to facilitate future analysis in efficient robust deep\nlearning.\n", "date": "2023-12-20", "categories": []}
{"id": "2211.09273", "title": "Privacy against Real-Time Speech Emotion Detection via Acoustic\n  Adversarial Evasion of Machine Learning", "abstract": "  Smart speaker voice assistants (VAs) such as Amazon Echo and Google Home have\nbeen widely adopted due to their seamless integration with smart home devices\nand the Internet of Things (IoT) technologies. These VA services raise privacy\nconcerns, especially due to their access to our speech. This work considers one\nsuch use case: the unaccountable and unauthorized surveillance of a user's\nemotion via speech emotion recognition (SER). This paper presents DARE-GP, a\nsolution that creates additive noise to mask users' emotional information while\npreserving the transcription-relevant portions of their speech. DARE-GP does\nthis by using a constrained genetic programming approach to learn the spectral\nfrequency traits that depict target users' emotional content, and then\ngenerating a universal adversarial audio perturbation that provides this\nprivacy protection. Unlike existing works, DARE-GP provides: a) real-time\nprotection of previously unheard utterances, b) against previously unseen\nblack-box SER classifiers, c) while protecting speech transcription, and d)\ndoes so in a realistic, acoustic environment. Further, this evasion is robust\nagainst defenses employed by a knowledgeable adversary. The evaluations in this\nwork culminate with acoustic evaluations against two off-the-shelf commercial\nsmart speakers using a small-form-factor (raspberry pi) integrated with a\nwake-word system to evaluate the efficacy of its real-world, real-time\ndeployment.\n", "date": "2022-11-17", "categories": []}
{"id": "2312.11500", "title": "A Red Teaming Framework for Securing AI in Maritime Autonomous Systems", "abstract": "  Artificial intelligence (AI) is being ubiquitously adopted to automate\nprocesses in science and industry. However, due to its often intricate and\nopaque nature, AI has been shown to possess inherent vulnerabilities which can\nbe maliciously exploited with adversarial AI, potentially putting AI users and\ndevelopers at both cyber and physical risk. In addition, there is insufficient\ncomprehension of the real-world effects of adversarial AI and an inadequacy of\nAI security examinations; therefore, the growing threat landscape is unknown\nfor many AI solutions. To mitigate this issue, we propose one of the first red\nteam frameworks for evaluating the AI security of maritime autonomous systems.\nThe framework provides operators with a proactive (secure by design) and\nreactive (post-deployment evaluation) response to securing AI technology today\nand in the future. This framework is a multi-part checklist, which can be\ntailored to different systems and requirements. We demonstrate this framework\nto be highly effective for a red team to use to uncover numerous\nvulnerabilities within a real-world maritime autonomous systems AI, ranging\nfrom poisoning to adversarial patch attacks. The lessons learned from\nsystematic AI red teaming can help prevent MAS-related catastrophic events in a\nworld with increasing uptake and reliance on mission-critical AI.\n", "date": "2023-12-08", "categories": []}
{"id": "2312.11510", "title": "QuadAttack: A Quadratic Programming Approach to Ordered Top-K Attacks", "abstract": "  The adversarial vulnerability of Deep Neural Networks (DNNs) has been\nwell-known and widely concerned, often under the context of learning top-$1$\nattacks (e.g., fooling a DNN to classify a cat image as dog). This paper shows\nthat the concern is much more serious by learning significantly more aggressive\nordered top-$K$ clear-box~\\footnote{ This is often referred to as\nwhite/black-box attacks in the literature. We choose to adopt neutral\nterminology, clear/opaque-box attacks in this paper, and omit the prefix\nclear-box for simplicity.} targeted attacks proposed in Adversarial\nDistillation. We propose a novel and rigorous quadratic programming (QP) method\nof learning ordered top-$K$ attacks with low computing cost, dubbed as\n\\textbf{QuadAttac$K$}. Our QuadAttac$K$ directly solves the QP to satisfy the\nattack constraint in the feature embedding space (i.e., the input space to the\nfinal linear classifier), which thus exploits the semantics of the feature\nembedding space (i.e., the principle of class coherence). With the optimized\nfeature embedding vector perturbation, it then computes the adversarial\nperturbation in the data space via the vanilla one-step back-propagation. In\nexperiments, the proposed QuadAttac$K$ is tested in the ImageNet-1k\nclassification using ResNet-50, DenseNet-121, and Vision Transformers (ViT-B\nand DEiT-S). It successfully pushes the boundary of successful ordered top-$K$\nattacks from $K=10$ up to $K=20$ at a cheap budget ($1\\times 60$) and further\nimproves attack success rates for $K=5$ for all tested models, while retaining\nthe performance for $K=1$.\n", "date": "2023-12-12", "categories": []}
{"id": "2303.00215", "title": "Single Image Backdoor Inversion via Robust Smoothed Classifiers", "abstract": "  Backdoor inversion, a central step in many backdoor defenses, is a\nreverse-engineering process to recover the hidden backdoor trigger inserted\ninto a machine learning model. Existing approaches tackle this problem by\nsearching for a backdoor pattern that is able to flip a set of clean images\ninto the target class, while the exact size needed of this support set is\nrarely investigated. In this work, we present a new approach for backdoor\ninversion, which is able to recover the hidden backdoor with as few as a single\nimage. Insipired by recent advances in adversarial robustness, our method\nSmoothInv starts from a single clean image, and then performs projected\ngradient descent towards the target class on a robust smoothed version of the\noriginal backdoored classifier. We find that backdoor patterns emerge naturally\nfrom such optimization process. Compared to existing backdoor inversion\nmethods, SmoothInv introduces minimum optimization variables and does not\nrequire complex regularization schemes. We perform a comprehensive quantitative\nand qualitative study on backdoored classifiers obtained from existing backdoor\nattacks. We demonstrate that SmoothInv consistently recovers successful\nbackdoors from single images: for backdoored ImageNet classifiers, our\nreconstructed backdoors have close to 100% attack success rates. We also show\nthat they maintain high fidelity to the underlying true backdoors. Last, we\npropose and analyze two countermeasures to our approach and show that SmoothInv\nremains robust in the face of an adaptive attacker. Our code is available at\nhttps://github.com/locuslab/smoothinv.\n", "date": "2023-03-01", "categories": []}
{"id": "2303.11611", "title": "Out of Thin Air: Exploring Data-Free Adversarial Robustness Distillation", "abstract": "  Adversarial Robustness Distillation (ARD) is a promising task to solve the\nissue of limited adversarial robustness of small capacity models while\noptimizing the expensive computational costs of Adversarial Training (AT).\nDespite the good robust performance, the existing ARD methods are still\nimpractical to deploy in natural high-security scenes due to these methods rely\nentirely on original or publicly available data with a similar distribution. In\nfact, these data are almost always private, specific, and distinctive for\nscenes that require high robustness. To tackle these issues, we propose a\nchallenging but significant task called Data-Free Adversarial Robustness\nDistillation (DFARD), which aims to train small, easily deployable, robust\nmodels without relying on data. We demonstrate that the challenge lies in the\nlower upper bound of knowledge transfer information, making it crucial to\nmining and transferring knowledge more efficiently. Inspired by human\neducation, we design a plug-and-play Interactive Temperature Adjustment (ITA)\nstrategy to improve the efficiency of knowledge transfer and propose an\nAdaptive Generator Balance (AGB) module to retain more data information. Our\nmethod uses adaptive hyperparameters to avoid a large number of parameter\ntuning, which significantly outperforms the combination of existing techniques.\nMeanwhile, our method achieves stable and reliable performance on multiple\nbenchmarks.\n", "date": "2023-03-21", "categories": []}
{"id": "2312.09494", "title": "No-Skim: Towards Efficiency Robustness Evaluation on Skimming-based\n  Language Models", "abstract": "  To reduce the computation cost and the energy consumption in large language\nmodels (LLM), skimming-based acceleration dynamically drops unimportant tokens\nof the input sequence progressively along layers of the LLM while preserving\nthe tokens of semantic importance. However, our work for the first time reveals\nthe acceleration may be vulnerable to Denial-of-Service (DoS) attacks. In this\npaper, we propose No-Skim, a general framework to help the owners of\nskimming-based LLM to understand and measure the robustness of their\nacceleration scheme. Specifically, our framework searches minimal and\nunnoticeable perturbations at character-level and token-level to generate\nadversarial inputs that sufficiently increase the remaining token ratio, thus\nincreasing the computation cost and energy consumption. We systematically\nevaluate the vulnerability of the skimming acceleration in various LLM\narchitectures including BERT and RoBERTa on the GLUE benchmark. In the worst\ncase, the perturbation found by No-Skim substantially increases the running\ncost of LLM by over 145% on average. Moreover, No-Skim extends the evaluation\nframework to various scenarios, making the evaluation conductible with\ndifferent level of knowledge.\n", "date": "2023-12-15", "categories": []}
{"id": "2312.10329", "title": "Perturbation-Invariant Adversarial Training for Neural Ranking Models:\n  Improving the Effectiveness-Robustness Trade-Off", "abstract": "  Neural ranking models (NRMs) have shown great success in information\nretrieval (IR). But their predictions can easily be manipulated using\nadversarial examples, which are crafted by adding imperceptible perturbations\nto legitimate documents. This vulnerability raises significant concerns about\ntheir reliability and hinders the widespread deployment of NRMs. By\nincorporating adversarial examples into training data, adversarial training has\nbecome the de facto defense approach to adversarial attacks against NRMs.\nHowever, this defense mechanism is subject to a trade-off between effectiveness\nand adversarial robustness. In this study, we establish theoretical guarantees\nregarding the effectiveness-robustness trade-off in NRMs. We decompose the\nrobust ranking error into two components, i.e., a natural ranking error for\neffectiveness evaluation and a boundary ranking error for assessing adversarial\nrobustness. Then, we define the perturbation invariance of a ranking model and\nprove it to be a differentiable upper bound on the boundary ranking error for\nattainable computation. Informed by our theoretical analysis, we design a novel\n\\emph{perturbation-invariant adversarial training} (PIAT) method for ranking\nmodels to achieve a better effectiveness-robustness trade-off. We design a\nregularized surrogate loss, in which one term encourages the effectiveness to\nbe maximized while the regularization term encourages the output to be smooth,\nso as to improve adversarial robustness. Experimental results on several\nranking models demonstrate the superiority of PITA compared to existing\nadversarial defenses.\n", "date": "2023-12-16", "categories": []}
{"id": "2312.10508", "title": "TrojFair: Trojan Fairness Attacks", "abstract": "  Deep learning models have been incorporated into high-stakes sectors,\nincluding healthcare diagnosis, loan approvals, and candidate recruitment,\namong others. Consequently, any bias or unfairness in these models can harm\nthose who depend on such models. In response, many algorithms have emerged to\nensure fairness in deep learning. However, while the potential for harm is\nsubstantial, the resilience of these fair deep learning models against\nmalicious attacks has never been thoroughly explored, especially in the context\nof emerging Trojan attacks. Moving beyond prior research, we aim to fill this\nvoid by introducing \\textit{TrojFair}, a Trojan fairness attack. Unlike\nexisting attacks, TrojFair is model-agnostic and crafts a Trojaned model that\nfunctions accurately and equitably for clean inputs. However, it displays\ndiscriminatory behaviors \\text{-} producing both incorrect and unfair results\n\\text{-} for specific groups with tainted inputs containing a trigger. TrojFair\nis a stealthy Fairness attack that is resilient to existing model fairness\naudition detectors since the model for clean inputs is fair. TrojFair achieves\na target group attack success rate exceeding $88.77\\%$, with an average\naccuracy loss less than $0.44\\%$. It also maintains a high discriminative score\nbetween the target and non-target groups across various datasets and models.\n", "date": "2023-12-16", "categories": []}
{"id": "2312.10713", "title": "Synthesizing Black-box Anti-forensics DeepFakes with High Visual Quality", "abstract": "  DeepFake, an AI technology for creating facial forgeries, has garnered global\nattention. Amid such circumstances, forensics researchers focus on developing\ndefensive algorithms to counter these threats. In contrast, there are\ntechniques developed for enhancing the aggressiveness of DeepFake, e.g.,\nthrough anti-forensics attacks, to disrupt forensic detectors. However, such\nattacks often sacrifice image visual quality for improved undetectability. To\naddress this issue, we propose a method to generate novel adversarial\nsharpening masks for launching black-box anti-forensics attacks. Unlike many\nexisting arts, with such perturbations injected, DeepFakes could achieve high\nanti-forensics performance while exhibiting pleasant sharpening visual effects.\nAfter experimental evaluations, we prove that the proposed method could\nsuccessfully disrupt the state-of-the-art DeepFake detectors. Besides, compared\nwith the images processed by existing DeepFake anti-forensics methods, the\nvisual qualities of anti-forensics DeepFakes rendered by the proposed method\nare significantly refined.\n", "date": "2023-12-17", "categories": []}
{"id": "2312.10903", "title": "Robust Node Representation Learning via Graph Variational Diffusion\n  Networks", "abstract": "  Node representation learning by using Graph Neural Networks (GNNs) has been\nwidely explored. However, in recent years, compelling evidence has revealed\nthat GNN-based node representation learning can be substantially deteriorated\nby delicately-crafted perturbations in a graph structure. To learn robust node\nrepresentation in the presence of perturbations, various works have been\nproposed to safeguard GNNs. Within these existing works, Bayesian label\ntransition has been proven to be more effective, but this method is extensively\nreliant on a well-built prior distribution. The variational inference could\naddress this limitation by sampling the latent node embedding from a Gaussian\nprior distribution. Besides, leveraging the Gaussian distribution (noise) in\nhidden layers is an appealing strategy to strengthen the robustness of GNNs.\nHowever, our experiments indicate that such a strategy can cause over-smoothing\nissues during node aggregation. In this work, we propose the Graph Variational\nDiffusion Network (GVDN), a new node encoder that effectively manipulates\nGaussian noise to safeguard robustness on perturbed graphs while alleviating\nover-smoothing issues through two mechanisms: Gaussian diffusion and node\nembedding propagation. Thanks to these two mechanisms, our model can generate\nrobust node embeddings for recovery. Specifically, we design a retraining\nmechanism using the generated node embedding to recover the performance of node\nclassifications in the presence of perturbations. The experiments verify the\neffectiveness of our proposed model across six public datasets.\n", "date": "2023-12-18", "categories": []}
{"id": "2312.10911", "title": "The Pros and Cons of Adversarial Robustness", "abstract": "  Robustness is widely regarded as a fundamental problem in the analysis of\nmachine learning (ML) models. Most often robustness equates with deciding the\nnon-existence of adversarial examples, where adversarial examples denote\nsituations where small changes on some inputs cause a change in the prediction.\nThe perceived importance of ML model robustness explains the continued progress\nobserved for most of the last decade. Whereas robustness is often assessed\nlocally, i.e. given some target point in feature space, robustness can also be\ndefined globally, i.e. where any point in feature space can be considered. The\nimportance of ML model robustness is illustrated for example by the existence\nof competitions evaluating the progress of robustness tools, namely in the case\nof neural networks (NNs) but also by efforts towards robustness certification.\nMore recently, robustness tools have also been used for computing rigorous\nexplanations of ML models. In contrast with the observed successes of\nrobustness, this paper uncovers some limitations with existing definitions of\nrobustness, both global and local, but also with efforts towards robustness\ncertification. The paper also investigates uses of adversarial examples besides\nthose related with robustness.\n", "date": "2023-12-18", "categories": []}
{"id": "2208.09602", "title": "Exploring Adversarial Robustness of Vision Transformers in the Spectral\n  Perspective", "abstract": "  The Vision Transformer has emerged as a powerful tool for image\nclassification tasks, surpassing the performance of convolutional neural\nnetworks (CNNs). Recently, many researchers have attempted to understand the\nrobustness of Transformers against adversarial attacks. However, previous\nresearches have focused solely on perturbations in the spatial domain. This\npaper proposes an additional perspective that explores the adversarial\nrobustness of Transformers against frequency-selective perturbations in the\nspectral domain. To facilitate comparison between these two domains, an attack\nframework is formulated as a flexible tool for implementing attacks on images\nin the spatial and spectral domains. The experiments reveal that Transformers\nrely more on phase and low frequency information, which can render them more\nvulnerable to frequency-selective attacks than CNNs. This work offers new\ninsights into the properties and adversarial robustness of Transformers.\n", "date": "2022-08-20", "categories": []}
{"id": "2312.09636", "title": "A Malware Classification Survey on Adversarial Attacks and Defences", "abstract": "  As the number and complexity of malware attacks continue to increase, there\nis an urgent need for effective malware detection systems. While deep learning\nmodels are effective at detecting malware, they are vulnerable to adversarial\nattacks. Attacks like this can create malicious files that are resistant to\ndetection, creating a significant cybersecurity risk. Recent research has seen\nthe development of several adversarial attack and response approaches aiming at\nstrengthening deep learning models' resilience to such attacks. This survey\nstudy offers an in-depth look at current research in adversarial attack and\ndefensive strategies for malware classification in cybersecurity. The methods\nare classified into four categories: generative models, feature-based\napproaches, ensemble methods, and hybrid tactics. The article outlines\ncutting-edge procedures within each area, assessing their benefits and\ndrawbacks. Each topic presents cutting-edge approaches and explores their\nadvantages and disadvantages. In addition, the study discusses the datasets and\nassessment criteria that are often utilized on this subject. Finally, it\nidentifies open research difficulties and suggests future study options. This\ndocument is a significant resource for malware categorization and cyber\nsecurity researchers and practitioners.\n", "date": "2023-12-15", "categories": []}
{"id": "2310.06372", "title": "Leveraging Diffusion-Based Image Variations for Robust Training on\n  Poisoned Data", "abstract": "  Backdoor attacks pose a serious security threat for training neural networks\nas they surreptitiously introduce hidden functionalities into a model. Such\nbackdoors remain silent during inference on clean inputs, evading detection due\nto inconspicuous behavior. However, once a specific trigger pattern appears in\nthe input data, the backdoor activates, causing the model to execute its\nconcealed function. Detecting such poisoned samples within vast datasets is\nvirtually impossible through manual inspection. To address this challenge, we\npropose a novel approach that enables model training on potentially poisoned\ndatasets by utilizing the power of recent diffusion models. Specifically, we\ncreate synthetic variations of all training samples, leveraging the inherent\nresilience of diffusion models to potential trigger patterns in the data. By\ncombining this generative approach with knowledge distillation, we produce\nstudent models that maintain their general performance on the task while\nexhibiting robust resistance to backdoor triggers.\n", "date": "2023-10-10", "categories": []}
{"id": "2310.15140", "title": "AutoDAN: Interpretable Gradient-Based Adversarial Attacks on Large\n  Language Models", "abstract": "  Safety alignment of Large Language Models (LLMs) can be compromised with\nmanual jailbreak attacks and (automatic) adversarial attacks. Recent studies\nsuggest that defending against these attacks is possible: adversarial attacks\ngenerate unlimited but unreadable gibberish prompts, detectable by\nperplexity-based filters; manual jailbreak attacks craft readable prompts, but\ntheir limited number due to the necessity of human creativity allows for easy\nblocking. In this paper, we show that these solutions may be too optimistic. We\nintroduce AutoDAN, an interpretable, gradient-based adversarial attack that\nmerges the strengths of both attack types. Guided by the dual goals of\njailbreak and readability, AutoDAN optimizes and generates tokens one by one\nfrom left to right, resulting in readable prompts that bypass perplexity\nfilters while maintaining high attack success rates. Notably, these prompts,\ngenerated from scratch using gradients, are interpretable and diverse, with\nemerging strategies commonly seen in manual jailbreak attacks. They also\ngeneralize to unforeseen harmful behaviors and transfer to black-box LLMs\nbetter than their unreadable counterparts when using limited training data or a\nsingle proxy model. Furthermore, we show the versatility of AutoDAN by\nautomatically leaking system prompts using a customized objective. Our work\noffers a new way to red-team LLMs and understand jailbreak mechanisms via\ninterpretability.\n", "date": "2023-10-23", "categories": []}
{"id": "2312.08622", "title": "Scalable Ensemble-based Detection Method against Adversarial Attacks for\n  speaker verification", "abstract": "  Automatic speaker verification (ASV) is highly susceptible to adversarial\nattacks. Purification modules are usually adopted as a pre-processing to\nmitigate adversarial noise. However, they are commonly implemented across\ndiverse experimental settings, rendering direct comparisons challenging. This\npaper comprehensively compares mainstream purification techniques in a unified\nframework. We find these methods often face a trade-off between user experience\nand security, as they struggle to simultaneously maintain genuine sample\nperformance and reduce adversarial perturbations. To address this challenge,\nsome efforts have extended purification modules to encompass detection\ncapabilities, aiming to alleviate the trade-off. However, advanced purification\nmodules will always come into the stage to surpass previous detection method.\nAs a result, we further propose an easy-to-follow ensemble approach that\nintegrates advanced purification modules for detection, achieving\nstate-of-the-art (SOTA) performance in countering adversarial noise. Our\nensemble method has great potential due to its compatibility with future\nadvanced purification techniques.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.08651", "title": "Towards Inductive Robustness: Distilling and Fostering Wave-induced\n  Resonance in Transductive GCNs Against Graph Adversarial Attacks", "abstract": "  Graph neural networks (GNNs) have recently been shown to be vulnerable to\nadversarial attacks, where slight perturbations in the graph structure can lead\nto erroneous predictions. However, current robust models for defending against\nsuch attacks inherit the transductive limitations of graph convolutional\nnetworks (GCNs). As a result, they are constrained by fixed structures and do\nnot naturally generalize to unseen nodes. Here, we discover that transductive\nGCNs inherently possess a distillable robustness, achieved through a\nwave-induced resonance process. Based on this, we foster this resonance to\nfacilitate inductive and robust learning. Specifically, we first prove that the\nsignal formed by GCN-driven message passing (MP) is equivalent to the\nedge-based Laplacian wave, where, within a wave system, resonance can naturally\nemerge between the signal and its transmitting medium. This resonance provides\ninherent resistance to malicious perturbations inflicted on the signal system.\nWe then prove that merely three MP iterations within GCNs can induce signal\nresonance between nodes and edges, manifesting as a coupling between nodes and\ntheir distillable surrounding local subgraph. Consequently, we present Graph\nResonance-fostering Network (GRN) to foster this resonance via learning node\nrepresentations from their distilled resonating subgraphs. By capturing the\nedge-transmitted signals within this subgraph and integrating them with the\nnode signal, GRN embeds these combined signals into the central node's\nrepresentation. This node-wise embedding approach allows for generalization to\nunseen nodes. We validate our theoretical findings with experiments, and\ndemonstrate that GRN generalizes robustness to unseen nodes, whilst maintaining\nstate-of-the-art classification accuracy on perturbed graphs.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.08877", "title": "May the Noise be with you: Adversarial Training without Adversarial\n  Examples", "abstract": "  In this paper, we investigate the following question: Can we obtain\nadversarially-trained models without training on adversarial examples? Our\nintuition is that training a model with inherent stochasticity, i.e.,\noptimizing the parameters by minimizing a stochastic loss function, yields a\nrobust expectation function that is non-stochastic. In contrast to related\nmethods that introduce noise at the input level, our proposed approach\nincorporates inherent stochasticity by embedding Gaussian noise within the\nlayers of the NN model at training time. We model the propagation of noise\nthrough the layers, introducing a closed-form stochastic loss function that\nencapsulates a noise variance parameter. Additionally, we contribute a\nformalized noise-aware gradient, enabling the optimization of model parameters\nwhile accounting for stochasticity. Our experimental results confirm that the\nexpectation model of a stochastic architecture trained on benign distribution\nis adversarially robust. Interestingly, we find that the impact of the applied\nGaussian noise's standard deviation on both robustness and baseline accuracy\nclosely mirrors the impact of the noise magnitude employed in adversarial\ntraining. Our work contributes adversarially trained networks using a\ncompletely different approach, with empirically similar robustness to\nadversarial training.\n", "date": "2023-12-12", "categories": []}
{"id": "2312.08890", "title": "Defenses in Adversarial Machine Learning: A Survey", "abstract": "  Adversarial phenomenon has been widely observed in machine learning (ML)\nsystems, especially in those using deep neural networks, describing that ML\nsystems may produce inconsistent and incomprehensible predictions with humans\nat some particular cases. This phenomenon poses a serious security threat to\nthe practical application of ML systems, and several advanced attack paradigms\nhave been developed to explore it, mainly including backdoor attacks, weight\nattacks, and adversarial examples. For each individual attack paradigm, various\ndefense paradigms have been developed to improve the model robustness against\nthe corresponding attack paradigm. However, due to the independence and\ndiversity of these defense paradigms, it is difficult to examine the overall\nrobustness of an ML system against different kinds of attacks.This survey aims\nto build a systematic review of all existing defense paradigms from a unified\nperspective. Specifically, from the life-cycle perspective, we factorize a\ncomplete machine learning system into five stages, including pre-training,\ntraining, post-training, deployment, and inference stages, respectively. Then,\nwe present a clear taxonomy to categorize and review representative defense\nmethods at each individual stage. The unified perspective and presented\ntaxonomies not only facilitate the analysis of the mechanism of each defense\nparadigm but also help us to understand connections and differences among\ndifferent defense paradigms, which may inspire future research to develop more\nadvanced, comprehensive defenses.\n", "date": "2023-12-13", "categories": []}
{"id": "2312.08898", "title": "Detection and Defense of Unlearnable Examples", "abstract": "  Privacy preserving has become increasingly critical with the emergence of\nsocial media. Unlearnable examples have been proposed to avoid leaking personal\ninformation on the Internet by degrading generalization abilities of deep\nlearning models. However, our study reveals that unlearnable examples are\neasily detectable. We provide theoretical results on linear separability of\ncertain unlearnable poisoned dataset and simple network based detection methods\nthat can identify all existing unlearnable examples, as demonstrated by\nextensive experiments. Detectability of unlearnable examples with simple\nnetworks motivates us to design a novel defense method. We propose using\nstronger data augmentations coupled with adversarial noises generated by simple\nnetworks, to degrade the detectability and thus provide effective defense\nagainst unlearnable examples with a lower cost. Adversarial training with large\nbudgets is a widely-used defense method on unlearnable examples. We establish\nquantitative criteria between the poison and adversarial budgets which\ndetermine the existence of robust unlearnable examples or the failure of the\nadversarial defense.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.09057", "title": "On the Difficulty of Defending Contrastive Learning against Backdoor\n  Attacks", "abstract": "  Recent studies have shown that contrastive learning, like supervised\nlearning, is highly vulnerable to backdoor attacks wherein malicious functions\nare injected into target models, only to be activated by specific triggers.\nHowever, thus far it remains under-explored how contrastive backdoor attacks\nfundamentally differ from their supervised counterparts, which impedes the\ndevelopment of effective defenses against the emerging threat.\n  This work represents a solid step toward answering this critical question.\nSpecifically, we define TRL, a unified framework that encompasses both\nsupervised and contrastive backdoor attacks. Through the lens of TRL, we\nuncover that the two types of attacks operate through distinctive mechanisms:\nin supervised attacks, the learning of benign and backdoor tasks tends to occur\nindependently, while in contrastive attacks, the two tasks are deeply\nintertwined both in their representations and throughout their learning\nprocesses. This distinction leads to the disparate learning dynamics and\nfeature distributions of supervised and contrastive attacks. More importantly,\nwe reveal that the specificities of contrastive backdoor attacks entail\nimportant implications from a defense perspective: existing defenses for\nsupervised attacks are often inadequate and not easily retrofitted to\ncontrastive attacks. We also explore several alternative defenses and discuss\ntheir potential challenges. Our findings highlight the need for defenses\ntailored to the specificities of contrastive backdoor attacks, pointing to\npromising directions for future research.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.09078", "title": "Coevolutionary Algorithm for Building Robust Decision Trees under\n  Minimax Regret", "abstract": "  In recent years, there has been growing interest in developing robust machine\nlearning (ML) models that can withstand adversarial attacks, including one of\nthe most widely adopted, efficient, and interpretable ML algorithms-decision\ntrees (DTs). This paper proposes a novel coevolutionary algorithm (CoEvoRDT)\ndesigned to create robust DTs capable of handling noisy high-dimensional data\nin adversarial contexts. Motivated by the limitations of traditional DT\nalgorithms, we leverage adaptive coevolution to allow DTs to evolve and learn\nfrom interactions with perturbed input data. CoEvoRDT alternately evolves\ncompeting populations of DTs and perturbed features, enabling construction of\nDTs with desired properties. CoEvoRDT is easily adaptable to various target\nmetrics, allowing the use of tailored robustness criteria such as minimax\nregret. Furthermore, CoEvoRDT has potential to improve the results of other\nstate-of-the-art methods by incorporating their outcomes (DTs they produce)\ninto the initial population and optimize them in the process of coevolution.\nInspired by the game theory, CoEvoRDT utilizes mixed Nash equilibrium to\nenhance convergence. The method is tested on 20 popular datasets and shows\nsuperior performance compared to 4 state-of-the-art algorithms. It outperformed\nall competing methods on 13 datasets with adversarial accuracy metrics, and on\nall 20 considered datasets with minimax regret. Strong experimental results and\nflexibility in choosing the error measure make CoEvoRDT a promising approach\nfor constructing robust DTs in real-world applications.\n", "date": "2023-12-14", "categories": []}
{"id": "2306.04064", "title": "Transferable Adversarial Robustness for Categorical Data via Universal\n  Robust Embeddings", "abstract": "  Research on adversarial robustness is primarily focused on image and text\ndata. Yet, many scenarios in which lack of robustness can result in serious\nrisks, such as fraud detection, medical diagnosis, or recommender systems often\ndo not rely on images or text but instead on tabular data. Adversarial\nrobustness in tabular data poses two serious challenges. First, tabular\ndatasets often contain categorical features, and therefore cannot be tackled\ndirectly with existing optimization procedures. Second, in the tabular domain,\nalgorithms that are not based on deep networks are widely used and offer great\nperformance, but algorithms to enhance robustness are tailored to neural\nnetworks (e.g. adversarial training).\n  In this paper, we tackle both challenges. We present a method that allows us\nto train adversarially robust deep networks for tabular data and to transfer\nthis robustness to other classifiers via universal robust embeddings tailored\nto categorical data. These embeddings, created using a bilevel alternating\nminimization framework, can be transferred to boosted trees or random forests\nmaking them robust without the need for adversarial training while preserving\ntheir high accuracy on tabular data. We show that our methods outperform\nexisting techniques within a practical threat model suitable for tabular data.\n", "date": "2023-06-06", "categories": []}
{"id": "2309.13857", "title": "Adversarial Attacks on Video Object Segmentation with Hard Region\n  Discovery", "abstract": "  Video object segmentation has been applied to various computer vision tasks,\nsuch as video editing, autonomous driving, and human-robot interaction.\nHowever, the methods based on deep neural networks are vulnerable to\nadversarial examples, which are the inputs attacked by almost\nhuman-imperceptible perturbations, and the adversary (i.e., attacker) will fool\nthe segmentation model to make incorrect pixel-level predictions. This will\nrise the security issues in highly-demanding tasks because small perturbations\nto the input video will result in potential attack risks. Though adversarial\nexamples have been extensively used for classification, it is rarely studied in\nvideo object segmentation. Existing related methods in computer vision either\nrequire prior knowledge of categories or cannot be directly applied due to the\nspecial design for certain tasks, failing to consider the pixel-wise region\nattack. Hence, this work develops an object-agnostic adversary that has\nadversarial impacts on VOS by first-frame attacking via hard region discovery.\nParticularly, the gradients from the segmentation model are exploited to\ndiscover the easily confused region, in which it is difficult to identify the\npixel-wise objects from the background in a frame. This provides a hardness map\nthat helps to generate perturbations with a stronger adversarial power for\nattacking the first frame. Empirical studies on three benchmarks indicate that\nour attacker significantly degrades the performance of several state-of-the-art\nvideo object segmentation models.\n", "date": "2023-09-25", "categories": []}
{"id": "2309.14585", "title": "DifAttack: Query-Efficient Black-Box Attack via Disentangled Feature\n  Space", "abstract": "  This work investigates efficient score-based black-box adversarial attacks\nwith a high Attack Success Rate (ASR) and good generalizability. We design a\nnovel attack method based on a Disentangled Feature space, called DifAttack,\nwhich differs significantly from the existing ones operating over the entire\nfeature space. Specifically, DifAttack firstly disentangles an image's latent\nfeature into an adversarial feature and a visual feature, where the former\ndominates the adversarial capability of an image, while the latter largely\ndetermines its visual appearance. We train an autoencoder for the\ndisentanglement by using pairs of clean images and their Adversarial Examples\n(AEs) generated from available surrogate models via white-box attack methods.\nEventually, DifAttack iteratively optimizes the adversarial feature according\nto the query feedback from the victim model until a successful AE is generated,\nwhile keeping the visual feature unaltered. In addition, due to the avoidance\nof using surrogate models' gradient information when optimizing AEs for\nblack-box models, our proposed DifAttack inherently possesses better attack\ncapability in the open-set scenario, where the training dataset of the victim\nmodel is unknown. Extensive experimental results demonstrate that our method\nachieves significant improvements in ASR and query efficiency simultaneously,\nespecially in the targeted attack and open-set scenarios. The code is available\nat https://github.com/csjunjun/DifAttack.git.\n", "date": "2023-09-26", "categories": []}
{"id": "2312.04730", "title": "DeceptPrompt: Exploiting LLM-driven Code Generation via Adversarial\n  Natural Language Instructions", "abstract": "  With the advancement of Large Language Models (LLMs), significant progress\nhas been made in code generation, enabling LLMs to transform natural language\ninto programming code. These Code LLMs have been widely accepted by massive\nusers and organizations. However, a dangerous nature is hidden in the code,\nwhich is the existence of fatal vulnerabilities. While some LLM providers have\nattempted to address these issues by aligning with human guidance, these\nefforts fall short of making Code LLMs practical and robust. Without a deep\nunderstanding of the performance of the LLMs under the practical worst cases,\nit would be concerning to apply them to various real-world applications. In\nthis paper, we answer the critical issue: Are existing Code LLMs immune to\ngenerating vulnerable code? If not, what is the possible maximum severity of\nthis issue in practical deployment scenarios? In this paper, we introduce\nDeceptPrompt, a novel algorithm that can generate adversarial natural language\ninstructions that drive the Code LLMs to generate functionality correct code\nwith vulnerabilities. DeceptPrompt is achieved through a systematic\nevolution-based algorithm with a fine grain loss design. The unique advantage\nof DeceptPrompt enables us to find natural prefix/suffix with totally benign\nand non-directional semantic meaning, meanwhile, having great power in inducing\nthe Code LLMs to generate vulnerable code. This feature can enable us to\nconduct the almost-worstcase red-teaming on these LLMs in a real scenario,\nwhere users are using natural language. Our extensive experiments and analyses\non DeceptPrompt not only validate the effectiveness of our approach but also\nshed light on the huge weakness of LLMs in the code generation task. When\napplying the optimized prefix/suffix, the attack success rate (ASR) will\nimprove by average 50% compared with no prefix/suffix applying.\n", "date": "2023-12-07", "categories": []}
{"id": "2312.08193", "title": "Universal Adversarial Framework to Improve Adversarial Robustness for\n  Diabetic Retinopathy Detection", "abstract": "  Diabetic Retinopathy (DR) is a prevalent illness associated with Diabetes\nwhich, if left untreated, can result in irreversible blindness. Deep Learning\nbased systems are gradually being introduced as automated support for clinical\ndiagnosis. Since healthcare has always been an extremely important domain\ndemanding error-free performance, any adversaries could pose a big threat to\nthe applicability of such systems. In this work, we use Universal Adversarial\nPerturbations (UAPs) to quantify the vulnerability of Medical Deep Neural\nNetworks (DNNs) for detecting DR. To the best of our knowledge, this is the\nvery first attempt that works on attacking complete fine-grained classification\nof DR images using various UAPs. Also, as a part of this work, we use UAPs to\nfine-tune the trained models to defend against adversarial samples. We\nexperiment on several models and observe that the performance of such models\ntowards unseen adversarial attacks gets boosted on average by $3.41$\nCohen-kappa value and maximum by $31.92$ Cohen-kappa value. The performance\ndegradation on normal data upon ensembling the fine-tuned models was found to\nbe statistically insignificant using t-test, highlighting the benefits of\nUAP-based adversarial fine-tuning.\n", "date": "2023-12-13", "categories": []}
{"id": "2209.01962", "title": "Adversarial Detection: Attacking Object Detection in Real Time", "abstract": "  Intelligent robots rely on object detection models to perceive the\nenvironment. Following advances in deep learning security it has been revealed\nthat object detection models are vulnerable to adversarial attacks. However,\nprior research primarily focuses on attacking static images or offline videos.\nTherefore, it is still unclear if such attacks could jeopardize real-world\nrobotic applications in dynamic environments. This paper bridges this gap by\npresenting the first real-time online attack against object detection models.\nWe devise three attacks that fabricate bounding boxes for nonexistent objects\nat desired locations. The attacks achieve a success rate of about 90% within\nabout 20 iterations. The demo video is available at\nhttps://youtu.be/zJZ1aNlXsMU.\n", "date": "2022-09-05", "categories": []}
{"id": "2312.06991", "title": "Attacking the Loop: Adversarial Attacks on Graph-based Loop Closure\n  Detection", "abstract": "  With the advancement in robotics, it is becoming increasingly common for\nlarge factories and warehouses to incorporate visual SLAM (vSLAM) enabled\nautomated robots that operate closely next to humans. This makes any\nadversarial attacks on vSLAM components potentially detrimental to humans\nworking alongside them. Loop Closure Detection (LCD) is a crucial component in\nvSLAM that minimizes the accumulation of drift in mapping, since even a small\ndrift can accumulate into a significant drift over time. A prior work by Kim et\nal., SymbioLCD2, unified visual features and semantic objects into a single\ngraph structure for finding loop closure candidates. While this provided a\nperformance improvement over visual feature-based LCD, it also created a single\npoint of vulnerability for potential graph-based adversarial attacks. Unlike\npreviously reported visual-patch based attacks, small graph perturbations are\nfar more challenging to detect, making them a more significant threat. In this\npaper, we present Adversarial-LCD, a novel black-box evasion attack framework\nthat employs an eigencentrality-based perturbation method and an SVM-RBF\nsurrogate model with a Weisfeiler-Lehman feature extractor for attacking\ngraph-based LCD. Our evaluation shows that the attack performance of\nAdversarial-LCD with the SVM-RBF surrogate model was superior to that of other\nmachine learning surrogate algorithms, including SVM-linear, SVM-polynomial,\nand Bayesian classifier, demonstrating the effectiveness of our attack\nframework. Furthermore, we show that our eigencentrality-based perturbation\nmethod outperforms other algorithms, such as Random-walk and Shortest-path,\nhighlighting the efficiency of Adversarial-LCD's perturbation selection method.\n", "date": "2023-12-12", "categories": []}
{"id": "2312.07022", "title": "EdgePruner: Poisoned Edge Pruning in Graph Contrastive Learning", "abstract": "  Graph Contrastive Learning (GCL) is unsupervised graph representation\nlearning that can obtain useful representation of unknown nodes. The node\nrepresentation can be utilized as features of downstream tasks. However, GCL is\nvulnerable to poisoning attacks as with existing learning models. A\nstate-of-the-art defense cannot sufficiently negate adverse effects by poisoned\ngraphs although such a defense introduces adversarial training in the GCL. To\nachieve further improvement, pruning adversarial edges is important. To the\nbest of our knowledge, the feasibility remains unexplored in the GCL domain. In\nthis paper, we propose a simple defense for GCL, EdgePruner. We focus on the\nfact that the state-of-the-art poisoning attack on GCL tends to mainly add\nadversarial edges to create poisoned graphs, which means that pruning edges is\nimportant to sanitize the graphs. Thus, EdgePruner prunes edges that contribute\nto minimizing the contrastive loss based on the node representation obtained\nafter training on poisoned graphs by GCL. Furthermore, we focus on the fact\nthat nodes with distinct features are connected by adversarial edges in\npoisoned graphs. Thus, we introduce feature similarity between neighboring\nnodes to help more appropriately determine adversarial edges. This similarity\nis helpful in further eliminating adverse effects from poisoned graphs on\nvarious datasets. Finally, EdgePruner outputs a graph that yields the minimum\ncontrastive loss as the sanitized graph. Our results demonstrate that pruning\nadversarial edges is feasible on six datasets. EdgePruner can improve the\naccuracy of node classification under the attack by up to 5.55% compared with\nthat of the state-of-the-art defense. Moreover, we show that EdgePruner is\nimmune to an adaptive attack.\n", "date": "2023-12-12", "categories": []}
{"id": "2312.07067", "title": "Focus on Hiders: Exploring Hidden Threats for Enhancing Adversarial\n  Training", "abstract": "  Adversarial training is often formulated as a min-max problem, however,\nconcentrating only on the worst adversarial examples causes alternating\nrepetitive confusion of the model, i.e., previously defended or correctly\nclassified samples are not defensible or accurately classifiable in subsequent\nadversarial training. We characterize such non-ignorable samples as \"hiders\",\nwhich reveal the hidden high-risk regions within the secure area obtained\nthrough adversarial training and prevent the model from finding the real worst\ncases. We demand the model to prevent hiders when defending against adversarial\nexamples for improving accuracy and robustness simultaneously. By rethinking\nand redefining the min-max optimization problem for adversarial training, we\npropose a generalized adversarial training algorithm called Hider-Focused\nAdversarial Training (HFAT). HFAT introduces the iterative evolution\noptimization strategy to simplify the optimization problem and employs an\nauxiliary model to reveal hiders, effectively combining the optimization\ndirections of standard adversarial training and prevention hiders. Furthermore,\nwe introduce an adaptive weighting mechanism that facilitates the model in\nadaptively adjusting its focus between adversarial examples and hiders during\ndifferent training periods. We demonstrate the effectiveness of our method\nbased on extensive experiments, and ensure that HFAT can provide higher\nrobustness and accuracy.\n", "date": "2023-12-12", "categories": []}
{"id": "2312.07158", "title": "Cost Aware Untargeted Poisoning Attack against Graph Neural Networks,", "abstract": "  Graph Neural Networks (GNNs) have become widely used in the field of graph\nmining. However, these networks are vulnerable to structural perturbations.\nWhile many research efforts have focused on analyzing vulnerability through\npoisoning attacks, we have identified an inefficiency in current attack losses.\nThese losses steer the attack strategy towards modifying edges targeting\nmisclassified nodes or resilient nodes, resulting in a waste of structural\nadversarial perturbation. To address this issue, we propose a novel attack loss\nframework called the Cost Aware Poisoning Attack (CA-attack) to improve the\nallocation of the attack budget by dynamically considering the classification\nmargins of nodes. Specifically, it prioritizes nodes with smaller positive\nmargins while postponing nodes with negative margins. Our experiments\ndemonstrate that the proposed CA-attack significantly enhances existing attack\nstrategies\n", "date": "2023-12-12", "categories": []}
{"id": "2312.07245", "title": "DTA: Distribution Transform-based Attack for Query-Limited Scenario", "abstract": "  In generating adversarial examples, the conventional black-box attack methods\nrely on sufficient feedback from the to-be-attacked models by repeatedly\nquerying until the attack is successful, which usually results in thousands of\ntrials during an attack. This may be unacceptable in real applications since\nMachine Learning as a Service Platform (MLaaS) usually only returns the final\nresult (i.e., hard-label) to the client and a system equipped with certain\ndefense mechanisms could easily detect malicious queries. By contrast, a\nfeasible way is a hard-label attack that simulates an attacked action being\npermitted to conduct a limited number of queries. To implement this idea, in\nthis paper, we bypass the dependency on the to-be-attacked model and benefit\nfrom the characteristics of the distributions of adversarial examples to\nreformulate the attack problem in a distribution transform manner and propose a\ndistribution transform-based attack (DTA). DTA builds a statistical mapping\nfrom the benign example to its adversarial counterparts by tackling the\nconditional likelihood under the hard-label black-box settings. In this way, it\nis no longer necessary to query the target model frequently. A well-trained DTA\nmodel can directly and efficiently generate a batch of adversarial examples for\na certain input, which can be used to attack un-seen models based on the\nassumed transferability. Furthermore, we surprisingly find that the\nwell-trained DTA model is not sensitive to the semantic spaces of the training\ndataset, meaning that the model yields acceptable attack performance on other\ndatasets. Extensive experiments validate the effectiveness of the proposed idea\nand the superiority of DTA over the state-of-the-art.\n", "date": "2023-12-12", "categories": []}
{"id": "2312.07258", "title": "SSTA: Salient Spatially Transformed Attack", "abstract": "  Extensive studies have demonstrated that deep neural networks (DNNs) are\nvulnerable to adversarial attacks, which brings a huge security risk to the\nfurther application of DNNs, especially for the AI models developed in the real\nworld. Despite the significant progress that has been made recently, existing\nattack methods still suffer from the unsatisfactory performance of escaping\nfrom being detected by naked human eyes due to the formulation of adversarial\nexample (AE) heavily relying on a noise-adding manner. Such mentioned\nchallenges will significantly increase the risk of exposure and result in an\nattack to be failed. Therefore, in this paper, we propose the Salient Spatially\nTransformed Attack (SSTA), a novel framework to craft imperceptible AEs, which\nenhance the stealthiness of AEs by estimating a smooth spatial transform metric\non a most critical area to generate AEs instead of adding external noise to the\nwhole image. Compared to state-of-the-art baselines, extensive experiments\nindicated that SSTA could effectively improve the imperceptibility of the AEs\nwhile maintaining a 100\\% attack success rate.\n", "date": "2023-12-12", "categories": []}
{"id": "2312.07389", "title": "Eroding Trust In Aerial Imagery: Comprehensive Analysis and Evaluation\n  Of Adversarial Attacks In Geospatial Systems", "abstract": "  In critical operations where aerial imagery plays an essential role, the\nintegrity and trustworthiness of data are paramount. The emergence of\nadversarial attacks, particularly those that exploit control over labels or\nemploy physically feasible trojans, threatens to erode that trust, making the\nanalysis and mitigation of these attacks a matter of urgency. We demonstrate\nhow adversarial attacks can degrade confidence in geospatial systems,\nspecifically focusing on scenarios where the attacker's control over labels is\nrestricted and the use of realistic threat vectors. Proposing and evaluating\nseveral innovative attack methodologies, including those tailored to overhead\nimages, we empirically show their threat to remote sensing systems using\nhigh-quality SpaceNet datasets. Our experimentation reflects the unique\nchallenges posed by aerial imagery, and these preliminary results not only\nreveal the potential risks but also highlight the non-trivial nature of the\nproblem compared to recent works.\n", "date": "2023-12-12", "categories": []}
{"id": "2211.05523", "title": "Impact of Adversarial Training on Robustness and Generalizability of\n  Language Models", "abstract": "  Adversarial training is widely acknowledged as the most effective defense\nagainst adversarial attacks. However, it is also well established that\nachieving both robustness and generalization in adversarially trained models\ninvolves a trade-off. The goal of this work is to provide an in depth\ncomparison of different approaches for adversarial training in language models.\nSpecifically, we study the effect of pre-training data augmentation as well as\ntraining time input perturbations vs. embedding space perturbations on the\nrobustness and generalization of transformer-based language models. Our\nfindings suggest that better robustness can be achieved by pre-training data\naugmentation or by training with input space perturbation. However, training\nwith embedding space perturbation significantly improves generalization. A\nlinguistic correlation analysis of neurons of the learned models reveals that\nthe improved generalization is due to 'more specialized' neurons. To the best\nof our knowledge, this is the first work to carry out a deep qualitative\nanalysis of different methods of generating adversarial examples in adversarial\ntraining of language models.\n", "date": "2022-11-10", "categories": []}
{"id": "2305.16771", "title": "Robust Nonparametric Regression under Poisoning Attack", "abstract": "  This paper studies robust nonparametric regression, in which an adversarial\nattacker can modify the values of up to $q$ samples from a training dataset of\nsize $N$. Our initial solution is an M-estimator based on Huber loss\nminimization. Compared with simple kernel regression, i.e. the Nadaraya-Watson\nestimator, this method can significantly weaken the impact of malicious samples\non the regression performance. We provide the convergence rate as well as the\ncorresponding minimax lower bound. The result shows that, with proper bandwidth\nselection, $\\ell_\\infty$ error is minimax optimal. The $\\ell_2$ error is\noptimal with relatively small $q$, but is suboptimal with larger $q$. The\nreason is that this estimator is vulnerable if there are many attacked samples\nconcentrating in a small region. To address this issue, we propose a correction\nmethod by projecting the initial estimate to the space of Lipschitz functions.\nThe final estimate is nearly minimax optimal for arbitrary $q$, up to a $\\ln N$\nfactor.\n", "date": "2023-05-26", "categories": []}
{"id": "2308.11070", "title": "Temporal-Distributed Backdoor Attack Against Video Based Action\n  Recognition", "abstract": "  Deep neural networks (DNNs) have achieved tremendous success in various\napplications including video action recognition, yet remain vulnerable to\nbackdoor attacks (Trojans). The backdoor-compromised model will mis-classify to\nthe target class chosen by the attacker when a test instance (from a non-target\nclass) is embedded with a specific trigger, while maintaining high accuracy on\nattack-free instances. Although there are extensive studies on backdoor attacks\nagainst image data, the susceptibility of video-based systems under backdoor\nattacks remains largely unexplored. Current studies are direct extensions of\napproaches proposed for image data, e.g., the triggers are independently\nembedded within the frames, which tend to be detectable by existing defenses.\nIn this paper, we introduce a simple yet effective backdoor attack against\nvideo data. Our proposed attack, adding perturbations in a transformed domain,\nplants an imperceptible, temporally distributed trigger across the video\nframes, and is shown to be resilient to existing defensive strategies. The\neffectiveness of the proposed attack is demonstrated by extensive experiments\nwith various well-known models on two video recognition benchmarks, UCF101 and\nHMDB51, and a sign language recognition benchmark, Greek Sign Language (GSL)\ndataset. We delve into the impact of several influential factors on our\nproposed attack and identify an intriguing effect termed \"collateral damage\"\nthrough extensive studies.\n", "date": "2023-08-21", "categories": []}
{"id": "2312.05502", "title": "Poisoning $\\times$ Evasion: Symbiotic Adversarial Robustness for Graph\n  Neural Networks", "abstract": "  It is well-known that deep learning models are vulnerable to small input\nperturbations. Such perturbed instances are called adversarial examples.\nAdversarial examples are commonly crafted to fool a model either at training\ntime (poisoning) or test time (evasion). In this work, we study the symbiosis\nof poisoning and evasion. We show that combining both threat models can\nsubstantially improve the devastating efficacy of adversarial attacks.\nSpecifically, we study the robustness of Graph Neural Networks (GNNs) under\nstructure perturbations and devise a memory-efficient adaptive end-to-end\nattack for the novel threat model using first-order optimization.\n", "date": "2023-12-09", "categories": []}
{"id": "2312.06163", "title": "Adversarial Camera Patch: An Effective and Robust Physical-World Attack\n  on Object Detectors", "abstract": "  Nowadays, the susceptibility of deep neural networks (DNNs) has garnered\nsignificant attention. Researchers are exploring patch-based physical attacks,\nyet traditional approaches, while effective, often result in conspicuous\npatches covering target objects. This leads to easy detection by human\nobservers. Recently, novel camera-based physical attacks have emerged,\nleveraging camera patches to execute stealthy attacks. These methods circumvent\ntarget object modifications by introducing perturbations directly to the camera\nlens, achieving a notable breakthrough in stealthiness. However, prevailing\ncamera-based strategies necessitate the deployment of multiple patches on the\ncamera lens, which introduces complexity. To address this issue, we propose an\nAdversarial Camera Patch (ADCP).\n", "date": "2023-12-11", "categories": []}
{"id": "2312.06568", "title": "Sparse but Strong: Crafting Adversarially Robust Graph Lottery Tickets", "abstract": "  Graph Lottery Tickets (GLTs), comprising a sparse adjacency matrix and a\nsparse graph neural network (GNN), can significantly reduce the inference\nlatency and compute footprint compared to their dense counterparts. Despite\nthese benefits, their performance against adversarial structure perturbations\nremains to be fully explored. In this work, we first investigate the resilience\nof GLTs against different structure perturbation attacks and observe that they\nare highly vulnerable and show a large drop in classification accuracy. Based\non this observation, we then present an adversarially robust graph\nsparsification (ARGS) framework that prunes the adjacency matrix and the GNN\nweights by optimizing a novel loss function capturing the graph homophily\nproperty and information associated with both the true labels of the train\nnodes and the pseudo labels of the test nodes. By iteratively applying ARGS to\nprune both the perturbed graph adjacency matrix and the GNN model weights, we\ncan find adversarially robust graph lottery tickets that are highly sparse yet\nachieve competitive performance under different untargeted training-time\nstructure attacks. Evaluations conducted on various benchmarks, considering\ndifferent poisoning structure attacks, namely, PGD, MetaAttack, Meta-PGD, and\nPR-BCD demonstrate that the GLTs generated by ARGS can significantly improve\nthe robustness, even when subjected to high levels of sparsity.\n", "date": "2023-12-11", "categories": []}
{"id": "2312.03777", "title": "On the Robustness of Large Multimodal Models Against Image Adversarial\n  Attacks", "abstract": "  Recent advances in instruction tuning have led to the development of\nState-of-the-Art Large Multimodal Models (LMMs). Given the novelty of these\nmodels, the impact of visual adversarial attacks on LMMs has not been\nthoroughly examined. We conduct a comprehensive study of the robustness of\nvarious LMMs against different adversarial attacks, evaluated across tasks\nincluding image classification, image captioning, and Visual Question Answer\n(VQA). We find that in general LMMs are not robust to visual adversarial\ninputs. However, our findings suggest that context provided to the model via\nprompts, such as questions in a QA pair helps to mitigate the effects of visual\nadversarial inputs. Notably, the LMMs evaluated demonstrated remarkable\nresilience to such attacks on the ScienceQA task with only an 8.10% drop in\nperformance compared to their visual counterparts which dropped 99.73%. We also\npropose a new approach to real-world image classification which we term query\ndecomposition. By incorporating existence queries into our input prompt we\nobserve diminished attack effectiveness and improvements in image\nclassification accuracy. This research highlights a previously under-explored\nfacet of LMM robustness and sets the stage for future work aimed at\nstrengthening the resilience of multimodal systems in adversarial environments.\n", "date": "2023-12-06", "categories": []}
{"id": "2312.04748", "title": "Forcing Generative Models to Degenerate Ones: The Power of Data\n  Poisoning Attacks", "abstract": "  Growing applications of large language models (LLMs) trained by a third party\nraise serious concerns on the security vulnerability of LLMs.It has been\ndemonstrated that malicious actors can covertly exploit these vulnerabilities\nin LLMs through poisoning attacks aimed at generating undesirable outputs.\nWhile poisoning attacks have received significant attention in the image domain\n(e.g., object detection), and classification tasks, their implications for\ngenerative models, particularly in the realm of natural language generation\n(NLG) tasks, remain poorly understood. To bridge this gap, we perform a\ncomprehensive exploration of various poisoning techniques to assess their\neffectiveness across a range of generative tasks. Furthermore, we introduce a\nrange of metrics designed to quantify the success and stealthiness of poisoning\nattacks specifically tailored to NLG tasks. Through extensive experiments on\nmultiple NLG tasks, LLMs and datasets, we show that it is possible to\nsuccessfully poison an LLM during the fine-tuning stage using as little as 1\\%\nof the total tuning data samples. Our paper presents the first systematic\napproach to comprehend poisoning attacks targeting NLG tasks considering a wide\nrange of triggers and attack settings. We hope our findings will assist the AI\nsecurity community in devising appropriate defenses against such threats.\n", "date": "2023-12-07", "categories": []}
{"id": "2312.04802", "title": "MimicDiffusion: Purifying Adversarial Perturbation via Mimicking Clean\n  Diffusion Model", "abstract": "  Deep neural networks (DNNs) are vulnerable to adversarial perturbation, where\nan imperceptible perturbation is added to the image that can fool the DNNs.\nDiffusion-based adversarial purification focuses on using the diffusion model\nto generate a clean image against such adversarial attacks. Unfortunately, the\ngenerative process of the diffusion model is also inevitably affected by\nadversarial perturbation since the diffusion model is also a deep network where\nits input has adversarial perturbation. In this work, we propose\nMimicDiffusion, a new diffusion-based adversarial purification technique, that\ndirectly approximates the generative process of the diffusion model with the\nclean image as input. Concretely, we analyze the differences between the guided\nterms using the clean image and the adversarial sample. After that, we first\nimplement MimicDiffusion based on Manhattan distance. Then, we propose two\nguidance to purify the adversarial perturbation and approximate the clean\ndiffusion model. Extensive experiments on three image datasets including\nCIFAR-10, CIFAR-100, and ImageNet with three classifier backbones including\nWideResNet-70-16, WideResNet-28-10, and ResNet50 demonstrate that\nMimicDiffusion significantly performs better than the state-of-the-art\nbaselines. On CIFAR-10, CIFAR-100, and ImageNet, it achieves 92.67\\%, 61.35\\%,\nand 61.53\\% average robust accuracy, which are 18.49\\%, 13.23\\%, and 17.64\\%\nhigher, respectively. The code is available in the supplementary material.\n", "date": "2023-12-08", "categories": []}
{"id": "2312.04879", "title": "HC-Ref: Hierarchical Constrained Refinement for Robust Adversarial\n  Training of GNNs", "abstract": "  Recent studies have shown that attackers can catastrophically reduce the\nperformance of GNNs by maliciously modifying the graph structure or node\nfeatures on the graph. Adversarial training, which has been shown to be one of\nthe most effective defense mechanisms against adversarial attacks in computer\nvision, holds great promise for enhancing the robustness of GNNs. There is\nlimited research on defending against attacks by performing adversarial\ntraining on graphs, and it is crucial to delve deeper into this approach to\noptimize its effectiveness. Therefore, based on robust adversarial training on\ngraphs, we propose a hierarchical constraint refinement framework (HC-Ref) that\nenhances the anti-perturbation capabilities of GNNs and downstream classifiers\nseparately, ultimately leading to improved robustness. We propose corresponding\nadversarial regularization terms that are conducive to adaptively narrowing the\ndomain gap between the normal part and the perturbation part according to the\ncharacteristics of different layers, promoting the smoothness of the predicted\ndistribution of both parts. Moreover, existing research on graph robust\nadversarial training primarily concentrates on training from the standpoint of\nnode feature perturbations and seldom takes into account alterations in the\ngraph structure. This limitation makes it challenging to prevent attacks based\non topological changes in the graph. This paper generates adversarial examples\nby utilizing graph structure perturbations, offering an effective approach to\ndefend against attack methods that are based on topological changes. Extensive\nexperiments on two real-world graph benchmarks show that HC-Ref successfully\nresists various attacks and has better node classification performance compared\nto several baseline methods.\n", "date": "2023-12-08", "categories": []}
{"id": "2312.04913", "title": "SA-Attack: Improving Adversarial Transferability of Vision-Language\n  Pre-training Models via Self-Augmentation", "abstract": "  Current Visual-Language Pre-training (VLP) models are vulnerable to\nadversarial examples. These adversarial examples present substantial security\nrisks to VLP models, as they can leverage inherent weaknesses in the models,\nresulting in incorrect predictions. In contrast to white-box adversarial\nattacks, transfer attacks (where the adversary crafts adversarial examples on a\nwhite-box model to fool another black-box model) are more reflective of\nreal-world scenarios, thus making them more meaningful for research. By\nsummarizing and analyzing existing research, we identified two factors that can\ninfluence the efficacy of transfer attacks on VLP models: inter-modal\ninteraction and data diversity. Based on these insights, we propose a\nself-augment-based transfer attack method, termed SA-Attack. Specifically,\nduring the generation of adversarial images and adversarial texts, we apply\ndifferent data augmentation methods to the image modality and text modality,\nrespectively, with the aim of improving the adversarial transferability of the\ngenerated adversarial images and texts. Experiments conducted on the FLickr30K\nand COCO datasets have validated the effectiveness of our method. Our code will\nbe available after this paper is accepted.\n", "date": "2023-12-08", "categories": []}
{"id": "2207.13326", "title": "Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets\n  Graph Signal Processing", "abstract": "  With the increasing attention in various 3D safety-critical applications,\npoint cloud learning models have been shown to be vulnerable to adversarial\nattacks. Although existing 3D attack methods achieve high success rates, they\ndelve into the data space with point-wise perturbation, which may neglect the\ngeometric characteristics. Instead, we propose point cloud attacks from a new\nperspective -- the graph spectral domain attack, aiming to perturb graph\ntransform coefficients in the spectral domain that corresponds to varying\ncertain geometric structure. Specifically, leveraging on graph signal\nprocessing, we first adaptively transform the coordinates of points onto the\nspectral domain via graph Fourier transform (GFT) for compact representation.\nThen, we analyze the influence of different spectral bands on the geometric\nstructure, based on which we propose to perturb the GFT coefficients via a\nlearnable graph spectral filter. Considering the low-frequency components\nmainly contribute to the rough shape of the 3D object, we further introduce a\nlow-frequency constraint to limit perturbations within imperceptible\nhigh-frequency components. Finally, the adversarial point cloud is generated by\ntransforming the perturbed spectral representation back to the data domain via\nthe inverse GFT. Experimental results demonstrate the effectiveness of the\nproposed attack in terms of both the imperceptibility and attack success rates.\n", "date": "2022-07-27", "categories": []}
{"id": "2302.10722", "title": "Characterizing the Optimal 0-1 Loss for Multi-class Classification with\n  a Test-time Attacker", "abstract": "  Finding classifiers robust to adversarial examples is critical for their safe\ndeployment. Determining the robustness of the best possible classifier under a\ngiven threat model for a given data distribution and comparing it to that\nachieved by state-of-the-art training methods is thus an important diagnostic\ntool. In this paper, we find achievable information-theoretic lower bounds on\nloss in the presence of a test-time attacker for multi-class classifiers on any\ndiscrete dataset. We provide a general framework for finding the optimal 0-1\nloss that revolves around the construction of a conflict hypergraph from the\ndata and adversarial constraints. We further define other variants of the\nattacker-classifier game that determine the range of the optimal loss more\nefficiently than the full-fledged hypergraph construction. Our evaluation\nshows, for the first time, an analysis of the gap to optimal robustness for\nclassifiers in the multi-class setting on benchmark datasets.\n", "date": "2023-02-21", "categories": []}
{"id": "2312.03979", "title": "Node-aware Bi-smoothing: Certified Robustness against Graph Injection\n  Attacks", "abstract": "  Deep Graph Learning (DGL) has emerged as a crucial technique across various\ndomains. However, recent studies have exposed vulnerabilities in DGL models,\nsuch as susceptibility to evasion and poisoning attacks. While empirical and\nprovable robustness techniques have been developed to defend against graph\nmodification attacks (GMAs), the problem of certified robustness against graph\ninjection attacks (GIAs) remains largely unexplored. To bridge this gap, we\nintroduce the node-aware bi-smoothing framework, which is the first certifiably\nrobust approach for general node classification tasks against GIAs. Notably,\nthe proposed node-aware bi-smoothing scheme is model-agnostic and is applicable\nfor both evasion and poisoning attacks. Through rigorous theoretical analysis,\nwe establish the certifiable conditions of our smoothing scheme. We also\nexplore the practical implications of our node-aware bi-smoothing schemes in\ntwo contexts: as an empirical defense approach against real-world GIAs and in\nthe context of recommendation systems. Furthermore, we extend two\nstate-of-the-art certified robustness frameworks to address node injection\nattacks and compare our approach against them. Extensive evaluations\ndemonstrate the effectiveness of our proposed certificates.\n", "date": "2023-12-07", "categories": []}
{"id": "2312.04032", "title": "RoAST: Robustifying Language Models via Adversarial Perturbation with\n  Selective Training", "abstract": "  Fine-tuning pre-trained language models (LMs) has become the de facto\nstandard in many NLP tasks. Nevertheless, fine-tuned LMs are still prone to\nrobustness issues, such as adversarial robustness and model calibration.\nSeveral perspectives of robustness for LMs have been studied independently, but\nlacking a unified consideration in multiple perspectives. In this paper, we\npropose Robustifying LMs via Adversarial perturbation with Selective Training\n(RoAST), a simple yet effective fine-tuning technique to enhance the\nmulti-perspective robustness of LMs in a unified way. RoAST effectively\nincorporates two important sources for the model robustness, robustness on the\nperturbed inputs and generalizable knowledge in pre-trained LMs. To be\nspecific, RoAST introduces adversarial perturbation during fine-tuning while\nthe model parameters are selectively updated upon their relative importance to\nminimize unnecessary deviation. Under a unified evaluation of fine-tuned LMs by\nincorporating four representative perspectives of model robustness, we\ndemonstrate the effectiveness of RoAST compared to state-of-the-art fine-tuning\nmethods on six different types of LMs, which indicates its usefulness in\npractice.\n", "date": "2023-12-07", "categories": []}
{"id": "2312.04403", "title": "OT-Attack: Enhancing Adversarial Transferability of Vision-Language\n  Models via Optimal Transport Optimization", "abstract": "  Vision-language pre-training (VLP) models demonstrate impressive abilities in\nprocessing both images and text. However, they are vulnerable to multi-modal\nadversarial examples (AEs). Investigating the generation of\nhigh-transferability adversarial examples is crucial for uncovering VLP models'\nvulnerabilities in practical scenarios. Recent works have indicated that\nleveraging data augmentation and image-text modal interactions can enhance the\ntransferability of adversarial examples for VLP models significantly. However,\nthey do not consider the optimal alignment problem between dataaugmented\nimage-text pairs. This oversight leads to adversarial examples that are overly\ntailored to the source model, thus limiting improvements in transferability. In\nour research, we first explore the interplay between image sets produced\nthrough data augmentation and their corresponding text sets. We find that\naugmented image samples can align optimally with certain texts while exhibiting\nless relevance to others. Motivated by this, we propose an Optimal\nTransport-based Adversarial Attack, dubbed OT-Attack. The proposed method\nformulates the features of image and text sets as two distinct distributions\nand employs optimal transport theory to determine the most efficient mapping\nbetween them. This optimal mapping informs our generation of adversarial\nexamples to effectively counteract the overfitting issues. Extensive\nexperiments across various network architectures and datasets in image-text\nmatching tasks reveal that our OT-Attack outperforms existing state-of-the-art\nmethods in terms of adversarial transferability.\n", "date": "2023-12-07", "categories": []}
{"id": "2303.07320", "title": "Model-tuning Via Prompts Makes NLP Models Adversarially Robust", "abstract": "  In recent years, NLP practitioners have converged on the following practice:\n(i) import an off-the-shelf pretrained (masked) language model; (ii) append a\nmultilayer perceptron atop the CLS token's hidden representation (with randomly\ninitialized weights); and (iii) fine-tune the entire model on a downstream task\n(MLP-FT). This procedure has produced massive gains on standard NLP benchmarks,\nbut these models remain brittle, even to mild adversarial perturbations. In\nthis work, we demonstrate surprising gains in adversarial robustness enjoyed by\nModel-tuning Via Prompts (MVP), an alternative method of adapting to downstream\ntasks. Rather than appending an MLP head to make output prediction, MVP appends\na prompt template to the input, and makes prediction via text\ninfilling/completion. Across 5 NLP datasets, 4 adversarial attacks, and 3\ndifferent models, MVP improves performance against adversarial substitutions by\nan average of 8% over standard methods and even outperforms adversarial\ntraining-based state-of-art defenses by 3.5%. By combining MVP with adversarial\ntraining, we achieve further improvements in adversarial robustness while\nmaintaining performance on unperturbed examples. Finally, we conduct ablations\nto investigate the mechanism underlying these gains. Notably, we find that the\nmain causes of vulnerability of MLP-FT can be attributed to the misalignment\nbetween pre-training and fine-tuning tasks, and the randomly initialized MLP\nparameters.\n", "date": "2023-03-13", "categories": []}
{"id": "2311.09428", "title": "Beyond Detection: Unveiling Fairness Vulnerabilities in Abusive Language\n  Models", "abstract": "  This work investigates the potential of undermining both fairness and\ndetection performance in abusive language detection. In a dynamic and complex\ndigital world, it is crucial to investigate the vulnerabilities of these\ndetection models to adversarial fairness attacks to improve their fairness\nrobustness. We propose a simple yet effective framework FABLE that leverages\nbackdoor attacks as they allow targeted control over the fairness and detection\nperformance. FABLE explores three types of trigger designs (i.e., rare,\nartificial, and natural triggers) and novel sampling strategies. Specifically,\nthe adversary can inject triggers into samples in the minority group with the\nfavored outcome (i.e., \"non-abusive\") and flip their labels to the unfavored\noutcome, i.e., \"abusive\". Experiments on benchmark datasets demonstrate the\neffectiveness of FABLE attacking fairness and utility in abusive language\ndetection.\n", "date": "2023-11-15", "categories": []}
{"id": "2312.01679", "title": "Adversarial Medical Image with Hierarchical Feature Hiding", "abstract": "  Deep learning based methods for medical images can be easily compromised by\nadversarial examples (AEs), posing a great security flaw in clinical\ndecision-making. It has been discovered that conventional adversarial attacks\nlike PGD which optimize the classification logits, are easy to distinguish in\nthe feature space, resulting in accurate reactive defenses. To better\nunderstand this phenomenon and reassess the reliability of the reactive\ndefenses for medical AEs, we thoroughly investigate the characteristic of\nconventional medical AEs. Specifically, we first theoretically prove that\nconventional adversarial attacks change the outputs by continuously optimizing\nvulnerable features in a fixed direction, thereby leading to outlier\nrepresentations in the feature space. Then, a stress test is conducted to\nreveal the vulnerability of medical images, by comparing with natural images.\nInterestingly, this vulnerability is a double-edged sword, which can be\nexploited to hide AEs. We then propose a simple-yet-effective hierarchical\nfeature constraint (HFC), a novel add-on to conventional white-box attacks,\nwhich assists to hide the adversarial feature in the target feature\ndistribution. The proposed method is evaluated on three medical datasets, both\n2D and 3D, with different modalities. The experimental results demonstrate the\nsuperiority of HFC, \\emph{i.e.,} it bypasses an array of state-of-the-art\nadversarial medical AE detectors more efficiently than competing adaptive\nattacks, which reveals the deficiencies of medical reactive defense and allows\nto develop more robust defenses in future.\n", "date": "2023-12-04", "categories": []}
{"id": "2312.03030", "title": "Generating Visually Realistic Adversarial Patch", "abstract": "  Deep neural networks (DNNs) are vulnerable to various types of adversarial\nexamples, bringing huge threats to security-critical applications. Among these,\nadversarial patches have drawn increasing attention due to their good\napplicability to fool DNNs in the physical world. However, existing works often\ngenerate patches with meaningless noise or patterns, making it conspicuous to\nhumans. To address this issue, we explore how to generate visually realistic\nadversarial patches to fool DNNs. Firstly, we analyze that a high-quality\nadversarial patch should be realistic, position irrelevant, and printable to be\ndeployed in the physical world. Based on this analysis, we propose an effective\nattack called VRAP, to generate visually realistic adversarial patches.\nSpecifically, VRAP constrains the patch in the neighborhood of a real image to\nensure the visual reality, optimizes the patch at the poorest position for\nposition irrelevance, and adopts Total Variance loss as well as gamma\ntransformation to make the generated patch printable without losing\ninformation. Empirical evaluations on the ImageNet dataset demonstrate that the\nproposed VRAP exhibits outstanding attack performance in the digital world.\nMoreover, the generated adversarial patches can be disguised as the scrawl or\nlogo in the physical world to fool the deep models without being detected,\nbringing significant threats to DNNs-enabled applications.\n", "date": "2023-12-05", "categories": []}
{"id": "2312.03520", "title": "Defense Against Adversarial Attacks using Convolutional Auto-Encoders", "abstract": "  Deep learning models, while achieving state-of-the-art performance on many\ntasks, are susceptible to adversarial attacks that exploit inherent\nvulnerabilities in their architectures. Adversarial attacks manipulate the\ninput data with imperceptible perturbations, causing the model to misclassify\nthe data or produce erroneous outputs. This work is based on enhancing the\nrobustness of targeted classifier models against adversarial attacks. To\nachieve this, an convolutional autoencoder-based approach is employed that\neffectively counters adversarial perturbations introduced to the input images.\nBy generating images closely resembling the input images, the proposed\nmethodology aims to restore the model's accuracy.\n", "date": "2023-12-06", "categories": []}
{"id": "2312.02207", "title": "TranSegPGD: Improving Transferability of Adversarial Examples on\n  Semantic Segmentation", "abstract": "  Transferability of adversarial examples on image classification has been\nsystematically explored, which generates adversarial examples in black-box\nmode. However, the transferability of adversarial examples on semantic\nsegmentation has been largely overlooked. In this paper, we propose an\neffective two-stage adversarial attack strategy to improve the transferability\nof adversarial examples on semantic segmentation, dubbed TranSegPGD.\nSpecifically, at the first stage, every pixel in an input image is divided into\ndifferent branches based on its adversarial property. Different branches are\nassigned different weights for optimization to improve the adversarial\nperformance of all pixels.We assign high weights to the loss of the\nhard-to-attack pixels to misclassify all pixels. At the second stage, the\npixels are divided into different branches based on their transferable property\nwhich is dependent on Kullback-Leibler divergence. Different branches are\nassigned different weights for optimization to improve the transferability of\nthe adversarial examples. We assign high weights to the loss of the\nhigh-transferability pixels to improve the transferability of adversarial\nexamples. Extensive experiments with various segmentation models are conducted\non PASCAL VOC 2012 and Cityscapes datasets to demonstrate the effectiveness of\nthe proposed method. The proposed adversarial attack method can achieve\nstate-of-the-art performance.\n", "date": "2023-12-03", "categories": []}
{"id": "2312.02236", "title": "Rethinking Adversarial Training with Neural Tangent Kernel", "abstract": "  Adversarial training (AT) is an important and attractive topic in deep\nlearning security, exhibiting mysteries and odd properties. Recent studies of\nneural network training dynamics based on Neural Tangent Kernel (NTK) make it\npossible to reacquaint AT and deeply analyze its properties. In this paper, we\nperform an in-depth investigation of AT process and properties with NTK, such\nas NTK evolution. We uncover three new findings that are missed in previous\nworks. First, we disclose the impact of data normalization on AT and the\nimportance of unbiased estimators in batch normalization layers. Second, we\nexperimentally explore the kernel dynamics and propose more time-saving AT\nmethods. Third, we study the spectrum feature inside the kernel to address the\ncatastrophic overfitting problem. To the best of our knowledge, it is the first\nwork leveraging the observations of kernel dynamics to improve existing AT\nmethods.\n", "date": "2023-12-04", "categories": []}
{"id": "2312.02237", "title": "Singular Regularization with Information Bottleneck Improves Model's\n  Adversarial Robustness", "abstract": "  Adversarial examples are one of the most severe threats to deep learning\nmodels. Numerous works have been proposed to study and defend adversarial\nexamples. However, these works lack analysis of adversarial information or\nperturbation, which cannot reveal the mystery of adversarial examples and lose\nproper interpretation. In this paper, we aim to fill this gap by studying\nadversarial information as unstructured noise, which does not have a clear\npattern. Specifically, we provide some empirical studies with singular value\ndecomposition, by decomposing images into several matrices, to analyze\nadversarial information for different attacks. Based on the analysis, we\npropose a new module to regularize adversarial information and combine\ninformation bottleneck theory, which is proposed to theoretically restrict\nintermediate representations. Therefore, our method is interpretable. Moreover,\nthe fashion of our design is a novel principle that is general and unified.\nEquipped with our new module, we evaluate two popular model structures on two\nmainstream datasets with various adversarial attacks. The results indicate that\nthe improvement in robust accuracy is significant. On the other hand, we prove\nthat our method is efficient with only a few additional parameters and able to\nbe explained under regional faithfulness analysis.\n", "date": "2023-12-04", "categories": []}
{"id": "2312.02780", "title": "Scaling Laws for Adversarial Attacks on Language Model Activations", "abstract": "  We explore a class of adversarial attacks targeting the activations of\nlanguage models. By manipulating a relatively small subset of model\nactivations, $a$, we demonstrate the ability to control the exact prediction of\na significant number (in some cases up to 1000) of subsequent tokens $t$. We\nempirically verify a scaling law where the maximum number of target tokens\n$t_\\mathrm{max}$ predicted depends linearly on the number of tokens $a$ whose\nactivations the attacker controls as $t_\\mathrm{max} = \\kappa a$. We find that\nthe number of bits of control in the input space needed to control a single bit\nin the output space (what we call attack resistance $\\chi$) is remarkably\nconstant between $\\approx 16$ and $\\approx 25$ over 2 orders of magnitude of\nmodel sizes for different language models. Compared to attacks on tokens,\nattacks on activations are predictably much stronger, however, we identify a\nsurprising regularity where one bit of input steered either via activations or\nvia tokens is able to exert control over a similar amount of output bits. This\ngives support for the hypothesis that adversarial attacks are a consequence of\ndimensionality mismatch between the input and output spaces. A practical\nimplication of the ease of attacking language model activations instead of\ntokens is for multi-modal and selected retrieval models, where additional data\nsources are added as activations directly, sidestepping the tokenized input.\nThis opens up a new, broad attack surface. By using language models as a\ncontrollable test-bed to study adversarial attacks, we were able to experiment\nwith input-output dimensions that are inaccessible in computer vision,\nespecially where the output dimension dominates.\n", "date": "2023-12-05", "categories": []}
{"id": "2312.02912", "title": "Realistic Scatterer Based Adversarial Attacks on SAR Image Classifiers", "abstract": "  Adversarial attacks have highlighted the vulnerability of classifiers based\non machine learning for Synthetic Aperture Radar (SAR) Automatic Target\nRecognition (ATR) tasks. An adversarial attack perturbs SAR images of on-ground\ntargets such that the classifiers are misled into making incorrect predictions.\nHowever, many existing attacking techniques rely on arbitrary manipulation of\nSAR images while overlooking the feasibility of executing the attacks on\nreal-world SAR imagery. Instead, adversarial attacks should be able to be\nimplemented by physical actions, for example, placing additional false objects\nas scatterers around the on-ground target to perturb the SAR image and fool the\nSAR ATR.\n  In this paper, we propose the On-Target Scatterer Attack (OTSA), a\nscatterer-based physical adversarial attack. To ensure the feasibility of its\nphysical execution, we enforce a constraint on the positioning of the\nscatterers. Specifically, we restrict the scatterers to be placed only on the\ntarget instead of in the shadow regions or the background. To achieve this, we\nintroduce a positioning score based on Gaussian kernels and formulate an\noptimization problem for our OTSA attack. Using a gradient ascent method to\nsolve the optimization problem, the OTSA can generate a vector of parameters\ndescribing the positions, shapes, sizes and amplitudes of the scatterers to\nguide the physical execution of the attack that will mislead SAR image\nclassifiers. The experimental results show that our attack obtains\nsignificantly higher success rates under the positioning constraint compared\nwith the existing method.\n", "date": "2023-12-05", "categories": []}
{"id": "2207.13381", "title": "Look Closer to Your Enemy: Learning to Attack via Teacher-Student\n  Mimicking", "abstract": "  Deep neural networks have significantly advanced person re-identification\n(ReID) applications in the realm of the industrial internet, yet they remain\nvulnerable. Thus, it is crucial to study the robustness of ReID systems, as\nthere are risks of adversaries using these vulnerabilities to compromise\nindustrial surveillance systems. Current adversarial methods focus on\ngenerating attack samples using misclassification feedback from victim models\n(VMs), neglecting VM's cognitive processes. We seek to address this by\nproducing authentic ReID attack instances through VM cognition decryption. This\napproach boasts advantages like better transferability to open-set ReID tests,\neasier VM misdirection, and enhanced creation of realistic and undetectable\nassault images. However, the task of deciphering the cognitive mechanism in VM\nis widely considered to be a formidable challenge. In this paper, we propose a\nnovel inconspicuous and controllable ReID attack baseline, LCYE (Look Closer to\nYour Enemy), to generate adversarial query images. Specifically, LCYE first\ndistills VM's knowledge via teacher-student memory mimicking the proxy task.\nThis knowledge prior serves as an unambiguous cryptographic token,\nencapsulating elements deemed indispensable and plausible by the VM, with the\nintent of facilitating precise adversarial misdirection. Further, benefiting\nfrom the multiple opposing task framework of LCYE, we investigate the\ninterpretability and generalization of ReID models from the view of the\nadversarial attack, including cross-domain adaption, cross-model consensus, and\nonline learning process. Extensive experiments on four ReID benchmarks show\nthat our method outperforms other state-of-the-art attackers with a large\nmargin in white-box, black-box, and target attacks. The source code can be\nfound at https://github.com/MingjieWang0606/LCYE-attack_reid.\n", "date": "2022-07-27", "categories": []}
{"id": "2210.00094", "title": "Improving Robustness with Adaptive Weight Decay", "abstract": "  We propose adaptive weight decay, which automatically tunes the\nhyper-parameter for weight decay during each training iteration. For\nclassification problems, we propose changing the value of the weight decay\nhyper-parameter on the fly based on the strength of updates from the\nclassification loss (i.e., gradient of cross-entropy), and the regularization\nloss (i.e., $\\ell_2$-norm of the weights). We show that this simple\nmodification can result in large improvements in adversarial robustness -- an\narea which suffers from robust overfitting -- without requiring extra data\nacross various datasets and architecture choices. For example, our\nreformulation results in $20\\%$ relative robustness improvement for CIFAR-100,\nand $10\\%$ relative robustness improvement on CIFAR-10 comparing to the best\ntuned hyper-parameters of traditional weight decay resulting in models that\nhave comparable performance to SOTA robustness methods. In addition, this\nmethod has other desirable properties, such as less sensitivity to learning\nrate, and smaller weight norms, which the latter contributes to robustness to\noverfitting to label noise, and pruning.\n", "date": "2022-09-30", "categories": []}
{"id": "2303.09051", "title": "Robust Evaluation of Diffusion-Based Adversarial Purification", "abstract": "  We question the current evaluation practice on diffusion-based purification\nmethods. Diffusion-based purification methods aim to remove adversarial effects\nfrom an input data point at test time. The approach gains increasing attention\nas an alternative to adversarial training due to the disentangling between\ntraining and testing. Well-known white-box attacks are often employed to\nmeasure the robustness of the purification. However, it is unknown whether\nthese attacks are the most effective for the diffusion-based purification since\nthe attacks are often tailored for adversarial training. We analyze the current\npractices and provide a new guideline for measuring the robustness of\npurification methods against adversarial attacks. Based on our analysis, we\nfurther propose a new purification strategy improving robustness compared to\nthe current diffusion-based purification methods.\n", "date": "2023-03-16", "categories": []}
{"id": "2306.15427", "title": "Adversarial Training for Graph Neural Networks: Pitfalls, Solutions, and\n  New Directions", "abstract": "  Despite its success in the image domain, adversarial training did not (yet)\nstand out as an effective defense for Graph Neural Networks (GNNs) against\ngraph structure perturbations. In the pursuit of fixing adversarial training\n(1) we show and overcome fundamental theoretical as well as practical\nlimitations of the adopted graph learning setting in prior work; (2) we reveal\nthat more flexible GNNs based on learnable graph diffusion are able to adjust\nto adversarial perturbations, while the learned message passing scheme is\nnaturally interpretable; (3) we introduce the first attack for structure\nperturbations that, while targeting multiple nodes at once, is capable of\nhandling global (graph-level) as well as local (node-level) constraints.\nIncluding these contributions, we demonstrate that adversarial training is a\nstate-of-the-art defense against adversarial structure perturbations.\n", "date": "2023-06-27", "categories": []}
{"id": "2307.08596", "title": "Omnipotent Adversarial Training in the Wild", "abstract": "  Adversarial training is an important topic in robust deep learning, but the\ncommunity lacks attention to its practical usage. In this paper, we aim to\nresolve a real-world challenge, i.e., training a model on an imbalanced and\nnoisy dataset to achieve high clean accuracy and adversarial robustness, with\nour proposed Omnipotent Adversarial Training (OAT) strategy. OAT consists of\ntwo innovative methodologies to address the imperfection in the training set.\nWe first introduce an oracle into the adversarial training process to help the\nmodel learn a correct data-label conditional distribution. This\ncarefully-designed oracle can provide correct label annotations for adversarial\ntraining. We further propose logits adjustment adversarial training to overcome\nthe data imbalance issue, which can help the model learn a Bayes-optimal\ndistribution. Our comprehensive evaluation results show that OAT outperforms\nother baselines by more than 20% clean accuracy improvement and 10% robust\naccuracy improvement under complex combinations of data imbalance and label\nnoise scenarios. The code can be found in https://github.com/GuanlinLee/OAT.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.10205", "title": "Alleviating the Effect of Data Imbalance on Adversarial Training", "abstract": "  In this paper, we study adversarial training on datasets that obey the\nlong-tailed distribution, which is practical but rarely explored in previous\nworks. Compared with conventional adversarial training on balanced datasets,\nthis process falls into the dilemma of generating uneven adversarial examples\n(AEs) and an unbalanced feature embedding space, causing the resulting model to\nexhibit low robustness and accuracy on tail data. To combat that, we\ntheoretically analyze the lower bound of the robust risk to train a model on a\nlong-tailed dataset to obtain the key challenges in addressing the\naforementioned dilemmas. Based on it, we propose a new adversarial training\nframework -- Re-balancing Adversarial Training (REAT). This framework consists\nof two components: (1) a new training strategy inspired by the effective number\nto guide the model to generate more balanced and informative AEs; (2) a\ncarefully constructed penalty function to force a satisfactory feature space.\nEvaluation results on different datasets and model structures prove that REAT\ncan effectively enhance the model's robustness and preserve the model's clean\naccuracy. The code can be found in https://github.com/GuanlinLee/REAT.\n", "date": "2023-07-14", "categories": []}
{"id": "2312.01789", "title": "Two-stage optimized unified adversarial patch for attacking\n  visible-infrared cross-modal detectors in the physical world", "abstract": "  Currently, many studies have addressed security concerns related to visible\nand infrared detectors independently. In practical scenarios, utilizing\ncross-modal detectors for tasks proves more reliable than relying on\nsingle-modal detectors. Despite this, there is a lack of comprehensive security\nevaluations for cross-modal detectors. While existing research has explored the\nfeasibility of attacks against cross-modal detectors, the implementation of a\nrobust attack remains unaddressed. This work introduces the Two-stage Optimized\nUnified Adversarial Patch (TOUAP) designed for performing attacks against\nvisible-infrared cross-modal detectors in real-world, black-box settings. The\nTOUAP employs a two-stage optimization process: firstly, PSO optimizes an\nirregular polygonal infrared patch to attack the infrared detector; secondly,\nthe color QR code is optimized, and the shape information of the infrared patch\nfrom the first stage is used as a mask. The resulting irregular polygon visible\nmodal patch executes an attack on the visible detector. Through extensive\nexperiments conducted in both digital and physical environments, we validate\nthe effectiveness and robustness of the proposed method. As the TOUAP surpasses\nbaseline performance, we advocate for its widespread attention.\n", "date": "2023-12-04", "categories": []}
{"id": "2212.02003", "title": "Bayesian Learning with Information Gain Provably Bounds Risk for a\n  Robust Adversarial Defense", "abstract": "  We present a new algorithm to learn a deep neural network model robust\nagainst adversarial attacks. Previous algorithms demonstrate an adversarially\ntrained Bayesian Neural Network (BNN) provides improved robustness. We\nrecognize the adversarial learning approach for approximating the multi-modal\nposterior distribution of a Bayesian model can lead to mode collapse;\nconsequently, the model's achievements in robustness and performance are\nsub-optimal. Instead, we first propose preventing mode collapse to better\napproximate the multi-modal posterior distribution. Second, based on the\nintuition that a robust model should ignore perturbations and only consider the\ninformative content of the input, we conceptualize and formulate an information\ngain objective to measure and force the information learned from both benign\nand adversarial training instances to be similar. Importantly. we prove and\ndemonstrate that minimizing the information gain objective allows the\nadversarial risk to approach the conventional empirical risk. We believe our\nefforts provide a step toward a basis for a principled method of adversarially\ntraining BNNs. Our model demonstrate significantly improved robustness--up to\n20%--compared with adversarial training and Adv-BNN under PGD attacks with\n0.035 distortion on both CIFAR-10 and STL-10 datasets.\n", "date": "2022-12-05", "categories": []}
{"id": "2307.00309", "title": "Adversarial Attacks and Defenses on 3D Point Cloud Classification: A\n  Survey", "abstract": "  Deep learning has successfully solved a wide range of tasks in 2D vision as a\ndominant AI technique. Recently, deep learning on 3D point clouds is becoming\nincreasingly popular for addressing various tasks in this field. Despite\nremarkable achievements, deep learning algorithms are vulnerable to adversarial\nattacks. These attacks are imperceptible to the human eye but can easily fool\ndeep neural networks in the testing and deployment stage. To encourage future\nresearch, this survey summarizes the current progress on adversarial attack and\ndefense techniques on point cloud classification.This paper first introduces\nthe principles and characteristics of adversarial attacks and summarizes and\nanalyzes adversarial example generation methods in recent years. Additionally,\nit provides an overview of defense strategies, organized into data-focused and\nmodel-focused methods. Finally, it presents several current challenges and\npotential future research directions in this domain.\n", "date": "2023-07-01", "categories": []}
{"id": "2312.00105", "title": "Improving the Robustness of Quantized Deep Neural Networks to White-Box\n  Attacks using Stochastic Quantization and Information-Theoretic Ensemble\n  Training", "abstract": "  Most real-world applications that employ deep neural networks (DNNs) quantize\nthem to low precision to reduce the compute needs. We present a method to\nimprove the robustness of quantized DNNs to white-box adversarial attacks. We\nfirst tackle the limitation of deterministic quantization to fixed ``bins'' by\nintroducing a differentiable Stochastic Quantizer (SQ). We explore the\nhypothesis that different quantizations may collectively be more robust than\neach quantized DNN. We formulate a training objective to encourage different\nquantized DNNs to learn different representations of the input image. The\ntraining objective captures diversity and accuracy via mutual information\nbetween ensemble members. Through experimentation, we demonstrate substantial\nimprovement in robustness against $L_\\infty$ attacks even if the attacker is\nallowed to backpropagate through SQ (e.g., > 50\\% accuracy to PGD(5/255) on\nCIFAR10 without adversarial training), compared to vanilla DNNs as well as\nexisting ensembles of quantized DNNs. We extend the method to detect attacks\nand generate robustness profiles in the adversarial information plane (AIP),\ntowards a unified analysis of different threat models by correlating the MI and\naccuracy.\n", "date": "2023-11-30", "categories": []}
{"id": "2312.00173", "title": "Fool the Hydra: Adversarial Attacks against Multi-view Object Detection\n  Systems", "abstract": "  Adversarial patches exemplify the tangible manifestation of the threat posed\nby adversarial attacks on Machine Learning (ML) models in real-world scenarios.\nRobustness against these attacks is of the utmost importance when designing\ncomputer vision applications, especially for safety-critical domains such as\nCCTV systems. In most practical situations, monitoring open spaces requires\nmulti-view systems to overcome acquisition challenges such as occlusion\nhandling. Multiview object systems are able to combine data from multiple\nviews, and reach reliable detection results even in difficult environments.\nDespite its importance in real-world vision applications, the vulnerability of\nmultiview systems to adversarial patches is not sufficiently investigated. In\nthis paper, we raise the following question: Does the increased performance and\ninformation sharing across views offer as a by-product robustness to\nadversarial patches? We first conduct a preliminary analysis showing promising\nrobustness against off-the-shelf adversarial patches, even in an extreme\nsetting where we consider patches applied to all views by all persons in\nWildtrack benchmark. However, we challenged this observation by proposing two\nnew attacks: (i) In the first attack, targeting a multiview CNN, we maximize\nthe global loss by proposing gradient projection to the different views and\naggregating the obtained local gradients. (ii) In the second attack, we focus\non a Transformer-based multiview framework. In addition to the focal loss, we\nalso maximize the transformer-specific loss by dissipating its attention\nblocks. Our results show a large degradation in the detection performance of\nvictim multiview systems with our first patch attack reaching an attack success\nrate of 73% , while our second proposed attack reduced the performance of its\ntarget detector by 62%\n", "date": "2023-11-30", "categories": []}
{"id": "2301.12993", "title": "Benchmarking Robustness to Adversarial Image Obfuscations", "abstract": "  Automated content filtering and moderation is an important tool that allows\nonline platforms to build striving user communities that facilitate cooperation\nand prevent abuse. Unfortunately, resourceful actors try to bypass automated\nfilters in a bid to post content that violate platform policies and codes of\nconduct. To reach this goal, these malicious actors may obfuscate policy\nviolating images (e.g. overlay harmful images by carefully selected benign\nimages or visual patterns) to prevent machine learning models from reaching the\ncorrect decision. In this paper, we invite researchers to tackle this specific\nissue and present a new image benchmark. This benchmark, based on ImageNet,\nsimulates the type of obfuscations created by malicious actors. It goes beyond\nImageNet-$\\textrm{C}$ and ImageNet-$\\bar{\\textrm{C}}$ by proposing general,\ndrastic, adversarial modifications that preserve the original content intent.\nIt aims to tackle a more common adversarial threat than the one considered by\n$\\ell_p$-norm bounded adversaries. We evaluate 33 pretrained models on the\nbenchmark and train models with different augmentations, architectures and\ntraining methods on subsets of the obfuscations to measure generalization. We\nhope this benchmark will encourage researchers to test their models and methods\nand try to find new approaches that are more robust to these obfuscations.\n", "date": "2023-01-30", "categories": []}
{"id": "2305.08192", "title": "Diffusion Models for Imperceptible and Transferable Adversarial Attack", "abstract": "  Many existing adversarial attacks generate $L_p$-norm perturbations on image\nRGB space. Despite some achievements in transferability and attack success\nrate, the crafted adversarial examples are easily perceived by human eyes.\nTowards visual imperceptibility, some recent works explore unrestricted attacks\nwithout $L_p$-norm constraints, yet lacking transferability of attacking\nblack-box models. In this work, we propose a novel imperceptible and\ntransferable attack by leveraging both the generative and discriminative power\nof diffusion models. Specifically, instead of direct manipulation in pixel\nspace, we craft perturbations in the latent space of diffusion models. Combined\nwith well-designed content-preserving structures, we can generate\nhuman-insensitive perturbations embedded with semantic clues. For better\ntransferability, we further \"deceive\" the diffusion model which can be viewed\nas an implicit recognition surrogate, by distracting its attention away from\nthe target regions. To our knowledge, our proposed method, DiffAttack, is the\nfirst that introduces diffusion models into the adversarial attack field.\nExtensive experiments on various model structures, datasets, and defense\nmethods have demonstrated the superiority of our attack over the existing\nattack methods.\n", "date": "2023-05-14", "categories": []}
{"id": "2306.00974", "title": "Discovering Failure Modes of Text-guided Diffusion Models via\n  Adversarial Search", "abstract": "  Text-guided diffusion models (TDMs) are widely applied but can fail\nunexpectedly. Common failures include: (i) natural-looking text prompts\ngenerating images with the wrong content, or (ii) different random samples of\nthe latent variables that generate vastly different, and even unrelated,\noutputs despite being conditioned on the same text prompt. In this work, we aim\nto study and understand the failure modes of TDMs in more detail. To achieve\nthis, we propose SAGE, the first adversarial search method on TDMs that\nsystematically explores the discrete prompt space and the high-dimensional\nlatent space, to automatically discover undesirable behaviors and failure cases\nin image generation. We use image classifiers as surrogate loss functions\nduring searching, and employ human inspections to validate the identified\nfailures. For the first time, our method enables efficient exploration of both\nthe discrete and intricate human language space and the challenging latent\nspace, overcoming the gradient vanishing problem. Then, we demonstrate the\neffectiveness of SAGE on five widely used generative models and reveal four\ntypical failure modes: (1) We find a variety of natural text prompts that\ngenerate images failing to capture the semantics of input texts. We further\ndiscuss the underlying causes and potential solutions based on the results. (2)\nWe find regions in the latent space that lead to distorted images independent\nof the text prompt, suggesting that parts of the latent space are not\nwell-structured. (3) We also find latent samples that result in natural-looking\nimages unrelated to the text prompt, implying a possible misalignment between\nthe latent and prompt spaces. (4) By appending a single adversarial token\nembedding to any input prompts, we can generate a variety of specified target\nobjects. Project page: https://sage-diffusion.github.io/\n", "date": "2023-06-01", "categories": []}
{"id": "2311.08592", "title": "AART: AI-Assisted Red-Teaming with Diverse Data Generation for New\n  LLM-powered Applications", "abstract": "  Adversarial testing of large language models (LLMs) is crucial for their safe\nand responsible deployment. We introduce a novel approach for automated\ngeneration of adversarial evaluation datasets to test the safety of LLM\ngenerations on new downstream applications. We call it AI-assisted Red-Teaming\n(AART) - an automated alternative to current manual red-teaming efforts. AART\noffers a data generation and augmentation pipeline of reusable and customizable\nrecipes that reduce human effort significantly and enable integration of\nadversarial testing earlier in new product development. AART generates\nevaluation datasets with high diversity of content characteristics critical for\neffective adversarial testing (e.g. sensitive and harmful concepts, specific to\na wide range of cultural and geographic regions and application scenarios). The\ndata generation is steered by AI-assisted recipes to define, scope and\nprioritize diversity within the application context. This feeds into a\nstructured LLM-generation process that scales up evaluation priorities.\nCompared to some state-of-the-art tools, AART shows promising results in terms\nof concept coverage and data quality.\n", "date": "2023-11-14", "categories": []}
{"id": "2311.17400", "title": "Improving the Robustness of Transformer-based Large Language Models with\n  Dynamic Attention", "abstract": "  Transformer-based models, such as BERT and GPT, have been widely adopted in\nnatural language processing (NLP) due to their exceptional performance.\nHowever, recent studies show their vulnerability to textual adversarial attacks\nwhere the model's output can be misled by intentionally manipulating the text\ninputs. Despite various methods that have been proposed to enhance the model's\nrobustness and mitigate this vulnerability, many require heavy consumption\nresources (e.g., adversarial training) or only provide limited protection\n(e.g., defensive dropout). In this paper, we propose a novel method called\ndynamic attention, tailored for the transformer architecture, to enhance the\ninherent robustness of the model itself against various adversarial attacks.\nOur method requires no downstream task knowledge and does not incur additional\ncosts. The proposed dynamic attention consists of two modules: (I) attention\nrectification, which masks or weakens the attention value of the chosen tokens,\nand (ii) dynamic modeling, which dynamically builds the set of candidate\ntokens. Extensive experiments demonstrate that dynamic attention significantly\nmitigates the impact of adversarial attacks, improving up to 33\\% better\nperformance than previous methods against widely-used adversarial attacks. The\nmodel-level design of dynamic attention enables it to be easily combined with\nother defense methods (e.g., adversarial training) to further enhance the\nmodel's robustness. Furthermore, we demonstrate that dynamic attention\npreserves the state-of-the-art robustness space of the original model compared\nto other dynamic modeling methods.\n", "date": "2023-11-29", "categories": []}
{"id": "2305.10665", "title": "Content-based Unrestricted Adversarial Attack", "abstract": "  Unrestricted adversarial attacks typically manipulate the semantic content of\nan image (e.g., color or texture) to create adversarial examples that are both\neffective and photorealistic, demonstrating their ability to deceive human\nperception and deep neural networks with stealth and success. However, current\nworks usually sacrifice unrestricted degrees and subjectively select some image\ncontent to guarantee the photorealism of unrestricted adversarial examples,\nwhich limits its attack performance. To ensure the photorealism of adversarial\nexamples and boost attack performance, we propose a novel unrestricted attack\nframework called Content-based Unrestricted Adversarial Attack. By leveraging a\nlow-dimensional manifold that represents natural images, we map the images onto\nthe manifold and optimize them along its adversarial direction. Therefore,\nwithin this framework, we implement Adversarial Content Attack based on Stable\nDiffusion and can generate high transferable unrestricted adversarial examples\nwith various adversarial contents. Extensive experimentation and visualization\ndemonstrate the efficacy of ACA, particularly in surpassing state-of-the-art\nattacks by an average of 13.3-50.4% and 16.8-48.0% in normally trained models\nand defense methods, respectively.\n", "date": "2023-05-18", "categories": []}
{"id": "2308.06703", "title": "Understanding the robustness difference between stochastic gradient\n  descent and adaptive gradient methods", "abstract": "  Stochastic gradient descent (SGD) and adaptive gradient methods, such as Adam\nand RMSProp, have been widely used in training deep neural networks. We\nempirically show that while the difference between the standard generalization\nperformance of models trained using these methods is small, those trained using\nSGD exhibit far greater robustness under input perturbations. Notably, our\ninvestigation demonstrates the presence of irrelevant frequencies in natural\ndatasets, where alterations do not affect models' generalization performance.\nHowever, models trained with adaptive methods show sensitivity to these\nchanges, suggesting that their use of irrelevant frequencies can lead to\nsolutions sensitive to perturbations. To better understand this difference, we\nstudy the learning dynamics of gradient descent (GD) and sign gradient descent\n(signGD) on a synthetic dataset that mirrors natural signals. With a\nthree-dimensional input space, the models optimized with GD and signGD have\nstandard risks close to zero but vary in their adversarial risks. Our result\nshows that linear models' robustness to $\\ell_2$-norm bounded changes is\ninversely proportional to the model parameters' weight norm: a smaller weight\nnorm implies better robustness. In the context of deep learning, our\nexperiments show that SGD-trained neural networks have smaller Lipschitz\nconstants, explaining the better robustness to input perturbations than those\ntrained with adaptive gradient methods.\n", "date": "2023-08-13", "categories": []}
{"id": "2311.17087", "title": "Rethinking Mixup for Improving the Adversarial Transferability", "abstract": "  Mixup augmentation has been widely integrated to generate adversarial\nexamples with superior adversarial transferability when immigrating from a\nsurrogate model to other models. However, the underlying mechanism influencing\nthe mixup's effect on transferability remains unexplored. In this work, we\nposit that the adversarial examples located at the convergence of decision\nboundaries across various categories exhibit better transferability and\nidentify that Admix tends to steer the adversarial examples towards such\nregions. However, we find the constraint on the added image in Admix decays its\ncapability, resulting in limited transferability. To address such an issue, we\npropose a new input transformation-based attack called Mixing the Image but\nSeparating the gradienT (MIST). Specifically, MIST randomly mixes the input\nimage with a randomly shifted image and separates the gradient of each loss\nitem for each mixed image. To counteract the imprecise gradient, MIST\ncalculates the gradient on several mixed images for each input sample.\nExtensive experimental results on the ImageNet dataset demonstrate that MIST\noutperforms existing SOTA input transformation-based attacks with a clear\nmargin on both Convolutional Neural Networks (CNNs) and Vision Transformers\n(ViTs) w/wo defense mechanisms, supporting MIST's high effectiveness and\ngenerality.\n", "date": "2023-11-28", "categories": []}
{"id": "2311.17128", "title": "Vulnerability Analysis of Transformer-based Optical Character\n  Recognition to Adversarial Attacks", "abstract": "  Recent advancements in Optical Character Recognition (OCR) have been driven\nby transformer-based models. OCR systems are critical in numerous high-stakes\ndomains, yet their vulnerability to adversarial attack remains largely\nuncharted territory, raising concerns about security and compliance with\nemerging AI regulations. In this work we present a novel framework to assess\nthe resilience of Transformer-based OCR (TrOCR) models. We develop and assess\nalgorithms for both targeted and untargeted attacks. For the untargeted case,\nwe measure the Character Error Rate (CER), while for the targeted case we use\nthe success ratio. We find that TrOCR is highly vulnerable to untargeted\nattacks and somewhat less vulnerable to targeted attacks. On a benchmark\nhandwriting data set, untargeted attacks can cause a CER of more than 1 without\nbeing noticeable to the eye. With a similar perturbation size, targeted attacks\ncan lead to success rates of around $25\\%$ -- here we attacked single tokens,\nrequiring TrOCR to output the tenth most likely token from a large vocabulary.\n", "date": "2023-11-28", "categories": []}
{"id": "2311.17332", "title": "NeRFTAP: Enhancing Transferability of Adversarial Patches on Face\n  Recognition using Neural Radiance Fields", "abstract": "  Face recognition (FR) technology plays a crucial role in various\napplications, but its vulnerability to adversarial attacks poses significant\nsecurity concerns. Existing research primarily focuses on transferability to\ndifferent FR models, overlooking the direct transferability to victim's face\nimages, which is a practical threat in real-world scenarios. In this study, we\npropose a novel adversarial attack method that considers both the\ntransferability to the FR model and the victim's face image, called NeRFTAP.\nLeveraging NeRF-based 3D-GAN, we generate new view face images for the source\nand target subjects to enhance transferability of adversarial patches. We\nintroduce a style consistency loss to ensure the visual similarity between the\nadversarial UV map and the target UV map under a 0-1 mask, enhancing the\neffectiveness and naturalness of the generated adversarial face images.\nExtensive experiments and evaluations on various FR models demonstrate the\nsuperiority of our approach over existing attack techniques. Our work provides\nvaluable insights for enhancing the robustness of FR systems in practical\nadversarial settings.\n", "date": "2023-11-29", "categories": []}
{"id": "2311.17608", "title": "Adversarial Robust Memory-Based Continual Learner", "abstract": "  Despite the remarkable advances that have been made in continual learning,\nthe adversarial vulnerability of such methods has not been fully discussed. We\ndelve into the adversarial robustness of memory-based continual learning\nalgorithms and observe limited robustness improvement by directly applying\nadversarial training techniques. Preliminary studies reveal the twin challenges\nfor building adversarial robust continual learners: accelerated forgetting in\ncontinual learning and gradient obfuscation in adversarial robustness. In this\nstudy, we put forward a novel adversarial robust memory-based continual learner\nthat adjusts data logits to mitigate the forgetting of pasts caused by\nadversarial samples. Furthermore, we devise a gradient-based data selection\nmechanism to overcome the gradient obfuscation caused by limited stored data.\nThe proposed approach can widely integrate with existing memory-based continual\nlearning as well as adversarial training algorithms in a plug-and-play way.\nExtensive experiments on Split-CIFAR10/100 and Split-Tiny-ImageNet demonstrate\nthe effectiveness of our approach, achieving up to 8.13% higher accuracy for\nadversarial data.\n", "date": "2023-11-29", "categories": []}
{"id": "2302.07221", "title": "On the Role of Randomization in Adversarially Robust Classification", "abstract": "  Deep neural networks are known to be vulnerable to small adversarial\nperturbations in test data. To defend against adversarial attacks,\nprobabilistic classifiers have been proposed as an alternative to deterministic\nones. However, literature has conflicting findings on the effectiveness of\nprobabilistic classifiers in comparison to deterministic ones. In this paper,\nwe clarify the role of randomization in building adversarially robust\nclassifiers. Given a base hypothesis set of deterministic classifiers, we show\nthe conditions under which a randomized ensemble outperforms the hypothesis set\nin adversarial risk, extending previous results. Additionally, we show that for\nany probabilistic binary classifier (including randomized ensembles), there\nexists a deterministic classifier that outperforms it. Finally, we give an\nexplicit description of the deterministic hypothesis set that contains such a\ndeterministic classifier for many types of commonly used probabilistic\nclassifiers, i.e. randomized ensembles and parametric/input noise injection.\n", "date": "2023-02-14", "categories": []}
{"id": "2311.16118", "title": "Imperceptible CMOS camera dazzle for adversarial attacks on deep neural\n  networks", "abstract": "  Despite the outstanding performance of deep neural networks, they are\nvulnerable to adversarial attacks. While there are many invisible attacks in\nthe digital domain, most physical world adversarial attacks are visible. Here\nwe present an invisible optical adversarial attack that uses a light source to\ndazzle a CMOS camera with a rolling shutter. We present the photopic conditions\nrequired to keep the attacking light source completely invisible while\nsufficiently jamming the captured image so that a deep neural network applied\nto it is deceived.\n", "date": "2023-10-22", "categories": []}
{"id": "2311.16478", "title": "RetouchUAA: Unconstrained Adversarial Attack via Image Retouching", "abstract": "  Deep Neural Networks (DNNs) are susceptible to adversarial examples.\nConventional attacks generate controlled noise-like perturbations that fail to\nreflect real-world scenarios and hard to interpretable. In contrast, recent\nunconstrained attacks mimic natural image transformations occurring in the real\nworld for perceptible but inconspicuous attacks, yet compromise realism due to\nneglect of image post-processing and uncontrolled attack direction. In this\npaper, we propose RetouchUAA, an unconstrained attack that exploits a real-life\nperturbation: image retouching styles, highlighting its potential threat to\nDNNs. Compared to existing attacks, RetouchUAA offers several notable\nadvantages. Firstly, RetouchUAA excels in generating interpretable and\nrealistic perturbations through two key designs: the image retouching attack\nframework and the retouching style guidance module. The former custom-designed\nhuman-interpretability retouching framework for adversarial attack by\nlinearizing images while modelling the local processing and retouching\ndecision-making in human retouching behaviour, provides an explicit and\nreasonable pipeline for understanding the robustness of DNNs against\nretouching. The latter guides the adversarial image towards standard retouching\nstyles, thereby ensuring its realism. Secondly, attributed to the design of the\nretouching decision regularization and the persistent attack strategy,\nRetouchUAA also exhibits outstanding attack capability and defense robustness,\nposing a heavy threat to DNNs. Experiments on ImageNet and Place365 reveal that\nRetouchUAA achieves nearly 100\\% white-box attack success against three DNNs,\nwhile achieving a better trade-off between image naturalness, transferability\nand defense robustness than baseline attacks.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.16577", "title": "Efficient Key-Based Adversarial Defense for ImageNet by Using\n  Pre-trained Model", "abstract": "  In this paper, we propose key-based defense model proliferation by leveraging\npre-trained models and utilizing recent efficient fine-tuning techniques on\nImageNet-1k classification. First, we stress that deploying key-based models on\nedge devices is feasible with the latest model deployment advancements, such as\nApple CoreML, although the mainstream enterprise edge artificial intelligence\n(Edge AI) has been focused on the Cloud. Then, we point out that the previous\nkey-based defense on on-device image classification is impractical for two\nreasons: (1) training many classifiers from scratch is not feasible, and (2)\nkey-based defenses still need to be thoroughly tested on large datasets like\nImageNet. To this end, we propose to leverage pre-trained models and utilize\nefficient fine-tuning techniques to proliferate key-based models even on\nlimited computing resources. Experiments were carried out on the ImageNet-1k\ndataset using adaptive and non-adaptive attacks. The results show that our\nproposed fine-tuned key-based models achieve a superior classification accuracy\n(more than 10% increase) compared to the previous key-based models on\nclassifying clean and adversarial examples.\n", "date": "2023-11-28", "categories": []}
{"id": "2303.10594", "title": "AdaptGuard: Defending Against Universal Attacks for Model Adaptation", "abstract": "  Model adaptation aims at solving the domain transfer problem under the\nconstraint of only accessing the pretrained source models. With the increasing\nconsiderations of data privacy and transmission efficiency, this paradigm has\nbeen gaining recent popularity. This paper studies the vulnerability to\nuniversal attacks transferred from the source domain during model adaptation\nalgorithms due to the existence of malicious providers. We explore both\nuniversal adversarial perturbations and backdoor attacks as loopholes on the\nsource side and discover that they still survive in the target models after\nadaptation. To address this issue, we propose a model preprocessing framework,\nnamed AdaptGuard, to improve the security of model adaptation algorithms.\nAdaptGuard avoids direct use of the risky source parameters through knowledge\ndistillation and utilizes the pseudo adversarial samples under adjusted radius\nto enhance the robustness. AdaptGuard is a plug-and-play module that requires\nneither robust pretrained models nor any changes for the following model\nadaptation algorithms. Extensive results on three commonly used datasets and\ntwo popular adaptation methods validate that AdaptGuard can effectively defend\nagainst universal attacks and maintain clean accuracy in the target domain\nsimultaneously. We hope this research will shed light on the safety and\nrobustness of transfer learning. Code is available at\nhttps://github.com/TomSheng21/AdaptGuard.\n", "date": "2023-03-19", "categories": []}
{"id": "2311.15339", "title": "Adversarial Purification of Information Masking", "abstract": "  Adversarial attacks meticulously generate minuscule, imperceptible\nperturbations to images to deceive neural networks. Counteracting these,\nadversarial purification methods seek to transform adversarial input samples\ninto clean output images to defend against adversarial attacks. Nonetheless,\nextent generative models fail to effectively eliminate adversarial\nperturbations, yielding less-than-ideal purification results. We emphasize the\npotential threat of residual adversarial perturbations to target models,\nquantitatively establishing a relationship between perturbation scale and\nattack capability. Notably, the residual perturbations on the purified image\nprimarily stem from the same-position patch and similar patches of the\nadversarial sample. We propose a novel adversarial purification approach named\nInformation Mask Purification (IMPure), aims to extensively eliminate\nadversarial perturbations. To obtain an adversarial sample, we first mask part\nof the patches information, then reconstruct the patches to resist adversarial\nperturbations from the patches. We reconstruct all patches in parallel to\nobtain a cohesive image. Then, in order to protect the purified samples against\npotential similar regional perturbations, we simulate this risk by randomly\nmixing the purified samples with the input samples before inputting them into\nthe feature extraction network. Finally, we establish a combined constraint of\npixel loss and perceptual loss to augment the model's reconstruction\nadaptability. Extensive experiments on the ImageNet dataset with three\nclassifier models demonstrate that our approach achieves state-of-the-art\nresults against nine adversarial attack methods. Implementation code and\npre-trained weights can be accessed at\n\\textcolor{blue}{https://github.com/NoWindButRain/IMPure}.\n", "date": "2023-11-26", "categories": []}
{"id": "2311.15551", "title": "Instruct2Attack: Language-Guided Semantic Adversarial Attacks", "abstract": "  We propose Instruct2Attack (I2A), a language-guided semantic attack that\ngenerates semantically meaningful perturbations according to free-form language\ninstructions. We make use of state-of-the-art latent diffusion models, where we\nadversarially guide the reverse diffusion process to search for an adversarial\nlatent code conditioned on the input image and text instruction. Compared to\nexisting noise-based and semantic attacks, I2A generates more natural and\ndiverse adversarial examples while providing better controllability and\ninterpretability. We further automate the attack process with GPT-4 to generate\ndiverse image-specific text instructions. We show that I2A can successfully\nbreak state-of-the-art deep neural networks even under strong adversarial\ndefenses, and demonstrate great transferability among a variety of network\narchitectures.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.15782", "title": "Relationship between Model Compression and Adversarial Robustness: A\n  Review of Current Evidence", "abstract": "  Increasing the model capacity is a known approach to enhance the adversarial\nrobustness of deep learning networks. On the other hand, various model\ncompression techniques, including pruning and quantization, can reduce the size\nof the network while preserving its accuracy. Several recent studies have\naddressed the relationship between model compression and adversarial\nrobustness, while some experiments have reported contradictory results. This\nwork summarizes available evidence and discusses possible explanations for the\nobserved effects.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.07553", "title": "An Extensive Study on Adversarial Attack against Pre-trained Models of\n  Code", "abstract": "  Transformer-based pre-trained models of code (PTMC) have been widely utilized\nand have achieved state-of-the-art performance in many mission-critical\napplications. However, they can be vulnerable to adversarial attacks through\nidentifier substitution or coding style transformation, which can significantly\ndegrade accuracy and may further incur security concerns. Although several\napproaches have been proposed to generate adversarial examples for PTMC, the\neffectiveness and efficiency of such approaches, especially on different code\nintelligence tasks, has not been well understood. To bridge this gap, this\nstudy systematically analyzes five state-of-the-art adversarial attack\napproaches from three perspectives: effectiveness, efficiency, and the quality\nof generated examples. The results show that none of the five approaches\nbalances all these perspectives. Particularly, approaches with a high attack\nsuccess rate tend to be time-consuming; the adversarial code they generate\noften lack naturalness, and vice versa. To address this limitation, we explore\nthe impact of perturbing identifiers under different contexts and find that\nidentifier substitution within for and if statements is the most effective.\nBased on these findings, we propose a new approach that prioritizes different\ntypes of statements for various tasks and further utilizes beam search to\ngenerate adversarial examples. Evaluation results show that it outperforms the\nstate-of-the-art ALERT in terms of both effectiveness and efficiency while\npreserving the naturalness of the generated adversarial examples.\n", "date": "2023-11-13", "categories": []}
{"id": "2311.11995", "title": "BrainWash: A Poisoning Attack to Forget in Continual Learning", "abstract": "  Continual learning has gained substantial attention within the deep learning\ncommunity, offering promising solutions to the challenging problem of\nsequential learning. Yet, a largely unexplored facet of this paradigm is its\nsusceptibility to adversarial attacks, especially with the aim of inducing\nforgetting. In this paper, we introduce \"BrainWash,\" a novel data poisoning\nmethod tailored to impose forgetting on a continual learner. By adding the\nBrainWash noise to a variety of baselines, we demonstrate how a trained\ncontinual learner can be induced to forget its previously learned tasks\ncatastrophically, even when using these continual learning baselines. An\nimportant feature of our approach is that the attacker requires no access to\nprevious tasks' data and is armed merely with the model's current parameters\nand the data belonging to the most recent task. Our extensive experiments\nhighlight the efficacy of BrainWash, showcasing degradation in performance\nacross various regularization-based continual learning methods.\n", "date": "2023-11-20", "categories": []}
{"id": "2311.14450", "title": "Segment (Almost) Nothing: Prompt-Agnostic Adversarial Attacks on\n  Segmentation Models", "abstract": "  General purpose segmentation models are able to generate (semantic)\nsegmentation masks from a variety of prompts, including visual (points, boxed,\netc.) and textual (object names) ones. In particular, input images are\npre-processed by an image encoder to obtain embedding vectors which are later\nused for mask predictions. Existing adversarial attacks target the end-to-end\ntasks, i.e. aim at altering the segmentation mask predicted for a specific\nimage-prompt pair. However, this requires running an individual attack for each\nnew prompt for the same image. We propose instead to generate prompt-agnostic\nadversarial attacks by maximizing the $\\ell_2$-distance, in the latent space,\nbetween the embedding of the original and perturbed images. Since the encoding\nprocess only depends on the image, distorted image representations will cause\nperturbations in the segmentation masks for a variety of prompts. We show that\neven imperceptible $\\ell_\\infty$-bounded perturbations of radius\n$\\epsilon=1/255$ are often sufficient to drastically modify the masks predicted\nwith point, box and text prompts by recently proposed foundation models for\nsegmentation. Moreover, we explore the possibility of creating universal, i.e.\nnon image-specific, attacks which can be readily applied to any input without\nfurther computational cost.\n", "date": "2023-11-24", "categories": []}
{"id": "2306.15755", "title": "Adversarial Backdoor Attack by Naturalistic Data Poisoning on Trajectory\n  Prediction in Autonomous Driving", "abstract": "  In autonomous driving, behavior prediction is fundamental for safe motion\nplanning, hence the security and robustness of prediction models against\nadversarial attacks are of paramount importance. We propose a novel adversarial\nbackdoor attack against trajectory prediction models as a means of studying\ntheir potential vulnerabilities. Our attack affects the victim at training time\nvia naturalistic, hence stealthy, poisoned samples crafted using a novel\ntwo-step approach. First, the triggers are crafted by perturbing the trajectory\nof attacking vehicle and then disguised by transforming the scene using a\nbi-level optimization technique. The proposed attack does not depend on a\nparticular model architecture and operates in a black-box manner, thus can be\neffective without any knowledge of the victim model. We conduct extensive\nempirical studies using state-of-the-art prediction models on two benchmark\ndatasets using metrics customized for trajectory prediction. We show that the\nproposed attack is highly effective, as it can significantly hinder the\nperformance of prediction models, unnoticeable by the victims, and efficient as\nit forces the victim to generate malicious behavior even under constrained\nconditions. Via ablative studies, we analyze the impact of different attack\ndesign choices followed by an evaluation of existing defence mechanisms against\nthe proposed attack.\n", "date": "2023-06-27", "categories": []}
{"id": "2311.12857", "title": "Adversarial sample generation and training using geometric masks for\n  accurate and resilient license plate character recognition", "abstract": "  Reading dirty license plates accurately in moving vehicles is challenging for\nautomatic license plate recognition systems. Moreover, license plates are often\nintentionally tampered with a malicious intent to avoid police apprehension.\nUsually, such groups and individuals know how to fool the existing recognition\nsystems by making minor unnoticeable plate changes. Designing and developing\ndeep learning methods resilient to such real-world 'attack' practices remains\nan active research problem. As a solution, this work develops a resilient\nmethod to recognize license plate characters. Extracting 1057 character images\nfrom 160 Nepalese vehicles, as the first step, we trained several standard deep\nconvolutional neural networks to obtain 99.5% character classification\naccuracy. On adversarial images generated to simulate malicious tampering,\nhowever, our model's accuracy dropped to 25%. Next, we enriched our dataset by\ngenerating and adding geometrically masked images, retrained our models, and\ninvestigated the models' predictions. The proposed approach of training with\ngenerated adversarial images helped our adversarial attack-aware license plate\ncharacter recognition (AA-LPCR) model achieves an accuracy of 99.7%. This\nnear-perfect accuracy demonstrates that the proposed idea of random geometric\nmasking is highly effective for improving the accuracy of license plate\nrecognition models. Furthermore, by performing interpretability studies to\nunderstand why our models work, we identify and highlight attack-prone regions\nin the input character images. In sum, although Nepal's embossed license plate\ndetection systems are vulnerable to malicious attacks, our findings suggest\nthat these systems can be upgraded to close to 100% resilience.\n", "date": "2023-10-25", "categories": []}
{"id": "2311.12858", "title": "RAEDiff: Denoising Diffusion Probabilistic Models Based Reversible\n  Adversarial Examples Self-Generation and Self-Recovery", "abstract": "  Collected and annotated datasets, which are obtained through extensive\nefforts, are effective for training Deep Neural Network (DNN) models. However,\nthese datasets are susceptible to be misused by unauthorized users, resulting\nin infringement of Intellectual Property (IP) rights owned by the dataset\ncreators. Reversible Adversarial Exsamples (RAE) can help to solve the issues\nof IP protection for datasets. RAEs are adversarial perturbed images that can\nbe restored to the original. As a cutting-edge approach, RAE scheme can serve\nthe purposes of preventing unauthorized users from engaging in malicious model\ntraining, as well as ensuring the legitimate usage of authorized users.\nNevertheless, in the existing work, RAEs still rely on the embedded auxiliary\ninformation for restoration, which may compromise their adversarial abilities.\nIn this paper, a novel self-generation and self-recovery method, named as\nRAEDiff, is introduced for generating RAEs based on a Denoising Diffusion\nProbabilistic Models (DDPM). It diffuses datasets into a Biased Gaussian\nDistribution (BGD) and utilizes the prior knowledge of the DDPM for generating\nand recovering RAEs. The experimental results demonstrate that RAEDiff\neffectively self-generates adversarial perturbations for DNN models, including\nArtificial Intelligence Generated Content (AIGC) models, while also exhibiting\nsignificant self-recovery capabilities.\n", "date": "2023-10-25", "categories": []}
{"id": "2311.13233", "title": "A Survey of Adversarial CAPTCHAs on its History, Classification and\n  Generation", "abstract": "  Completely Automated Public Turing test to tell Computers and Humans Apart,\nshort for CAPTCHA, is an essential and relatively easy way to defend against\nmalicious attacks implemented by bots. The security and usability trade-off\nlimits the use of massive geometric transformations to interfere deep model\nrecognition and deep models even outperformed humans in complex CAPTCHAs. The\ndiscovery of adversarial examples provides an ideal solution to the security\nand usability trade-off by integrating adversarial examples and CAPTCHAs to\ngenerate adversarial CAPTCHAs that can fool the deep models. In this paper, we\nextend the definition of adversarial CAPTCHAs and propose a classification\nmethod for adversarial CAPTCHAs. Then we systematically review some commonly\nused methods to generate adversarial examples and methods that are successfully\nused to generate adversarial CAPTCHAs. Also, we analyze some defense methods\nthat can be used to defend adversarial CAPTCHAs, indicating potential threats\nto adversarial CAPTCHAs. Finally, we discuss some possible future research\ndirections for adversarial CAPTCHAs at the end of this paper.\n", "date": "2023-11-22", "categories": []}
{"id": "2311.13244", "title": "Hard Label Black Box Node Injection Attack on Graph Neural Networks", "abstract": "  While graph neural networks have achieved state-of-the-art performances in\nmany real-world tasks including graph classification and node classification,\nrecent works have demonstrated they are also extremely vulnerable to\nadversarial attacks. Most previous works have focused on attacking node\nclassification networks under impractical white-box scenarios. In this work, we\nwill propose a non-targeted Hard Label Black Box Node Injection Attack on Graph\nNeural Networks, which to the best of our knowledge, is the first of its kind.\nUnder this setting, more real world tasks can be studied because our attack\nassumes no prior knowledge about (1): the model architecture of the GNN we are\nattacking; (2): the model's gradients; (3): the output logits of the target GNN\nmodel. Our attack is based on an existing edge perturbation attack, from which\nwe restrict the optimization process to formulate a node injection attack. In\nthe work, we will evaluate the performance of the attack using three datasets,\nCOIL-DEL, IMDB-BINARY, and NCI1.\n", "date": "2023-11-22", "categories": []}
{"id": "2311.13445", "title": "Transfer Attacks and Defenses for Large Language Models on Coding Tasks", "abstract": "  Modern large language models (LLMs), such as ChatGPT, have demonstrated\nimpressive capabilities for coding tasks including writing and reasoning about\ncode. They improve upon previous neural network models of code, such as\ncode2seq or seq2seq, that already demonstrated competitive results when\nperforming tasks such as code summarization and identifying code\nvulnerabilities. However, these previous code models were shown vulnerable to\nadversarial examples, i.e. small syntactic perturbations that do not change the\nprogram's semantics, such as the inclusion of \"dead code\" through false\nconditions or the addition of inconsequential print statements, designed to\n\"fool\" the models. LLMs can also be vulnerable to the same adversarial\nperturbations but a detailed study on this concern has been lacking so far. In\nthis paper we aim to investigate the effect of adversarial perturbations on\ncoding tasks with LLMs. In particular, we study the transferability of\nadversarial examples, generated through white-box attacks on smaller code\nmodels, to LLMs. Furthermore, to make the LLMs more robust against such\nadversaries without incurring the cost of retraining, we propose prompt-based\ndefenses that involve modifying the prompt to include additional information\nsuch as examples of adversarially perturbed code and explicit instructions for\nreversing adversarial perturbations. Our experiments show that adversarial\nexamples obtained with a smaller code model are indeed transferable, weakening\nthe LLMs' performance. The proposed defenses show promise in improving the\nmodel's resilience, paving the way to more robust defensive solutions for LLMs\nin code-related applications.\n", "date": "2023-11-22", "categories": []}
{"id": "2305.11618", "title": "DAP: A Dynamic Adversarial Patch for Evading Person Detectors", "abstract": "  Patch-based adversarial attacks were proven to compromise the robustness and\nreliability of computer vision systems. However, their conspicuous and easily\ndetectable nature challenge their practicality in real-world setting. To\naddress this, recent work has proposed using Generative Adversarial Networks\n(GANs) to generate naturalistic patches that may not attract human attention.\nHowever, such approaches suffer from a limited latent space making it\nchallenging to produce a patch that is efficient, stealthy, and robust to\nmultiple real-world transformations. This paper introduces a novel approach\nthat produces a Dynamic Adversarial Patch (DAP) designed to overcome these\nlimitations. DAP maintains a naturalistic appearance while optimizing attack\nefficiency and robustness to real-world transformations. The approach involves\nredefining the optimization problem and introducing a novel objective function\nthat incorporates a similarity metric to guide the patch's creation. Unlike\nGAN-based techniques, the DAP directly modifies pixel values within the patch,\nproviding increased flexibility and adaptability to multiple transformations.\nFurthermore, most clothing-based physical attacks assume static objects and\nignore the possible transformations caused by non-rigid deformation due to\nchanges in a person's pose. To address this limitation, a 'Creases\nTransformation' (CT) block is introduced, enhancing the patch's resilience to a\nvariety of real-world distortions. Experimental results demonstrate that the\nproposed approach outperforms state-of-the-art attacks, achieving a success\nrate of up to 82.28% in the digital world when targeting the YOLOv7 detector\nand 65% in the physical world when targeting YOLOv3tiny detector deployed in\nedge-based smart cameras.\n", "date": "2023-05-19", "categories": []}
{"id": "2306.12685", "title": "Rethinking the Backward Propagation for Adversarial Transferability", "abstract": "  Transfer-based attacks generate adversarial examples on the surrogate model,\nwhich can mislead other black-box models without access, making it promising to\nattack real-world applications. Recently, several works have been proposed to\nboost adversarial transferability, in which the surrogate model is usually\noverlooked. In this work, we identify that non-linear layers (e.g., ReLU,\nmax-pooling, etc.) truncate the gradient during backward propagation, making\nthe gradient w.r.t. input image imprecise to the loss function. We hypothesize\nand empirically validate that such truncation undermines the transferability of\nadversarial examples. Based on these findings, we propose a novel method called\nBackward Propagation Attack (BPA) to increase the relevance between the\ngradient w.r.t. input image and loss function so as to generate adversarial\nexamples with higher transferability. Specifically, BPA adopts a non-monotonic\nfunction as the derivative of ReLU and incorporates softmax with temperature to\nsmooth the derivative of max-pooling, thereby mitigating the information loss\nduring the backward propagation of gradients. Empirical results on the ImageNet\ndataset demonstrate that not only does our method substantially boost the\nadversarial transferability, but it is also general to existing transfer-based\nattacks. Code is available at https://github.com/Trustworthy-AI-Group/RPA.\n", "date": "2023-06-22", "categories": []}
{"id": "2311.12051", "title": "Boost Adversarial Transferability by Uniform Scale and Mix Mask Method", "abstract": "  Adversarial examples generated from surrogate models often possess the\nability to deceive other black-box models, a property known as transferability.\nRecent research has focused on enhancing adversarial transferability, with\ninput transformation being one of the most effective approaches. However,\nexisting input transformation methods suffer from two issues. Firstly, certain\nmethods, such as the Scale-Invariant Method, employ exponentially decreasing\nscale invariant parameters that decrease the adaptability in generating\neffective adversarial examples across multiple scales. Secondly, most mixup\nmethods only linearly combine candidate images with the source image, leading\nto reduced features blending effectiveness. To address these challenges, we\npropose a framework called Uniform Scale and Mix Mask Method (US-MM) for\nadversarial example generation. The Uniform Scale approach explores the upper\nand lower boundaries of perturbation with a linear factor, minimizing the\nnegative impact of scale copies. The Mix Mask method introduces masks into the\nmixing process in a nonlinear manner, significantly improving the effectiveness\nof mixing strategies. Ablation experiments are conducted to validate the\neffectiveness of each component in US-MM and explore the effect of\nhyper-parameters. Empirical evaluations on standard ImageNet datasets\ndemonstrate that US-MM achieves an average of 7% better transfer attack success\nrate compared to state-of-the-art methods.\n", "date": "2023-11-18", "categories": []}
{"id": "2311.12773", "title": "Iris Presentation Attack: Assessing the Impact of Combining Vanadium\n  Dioxide Films with Artificial Eyes", "abstract": "  Iris recognition systems, operating in the near infrared spectrum (NIR), have\ndemonstrated vulnerability to presentation attacks, where an adversary uses\nartifacts such as cosmetic contact lenses, artificial eyes or printed iris\nimages in order to circumvent the system. At the same time, a number of\neffective presentation attack detection (PAD) methods have been developed.\nThese methods have demonstrated success in detecting artificial eyes (e.g.,\nfake Van Dyke eyes) as presentation attacks. In this work, we seek to alter the\noptical characteristics of artificial eyes by affixing Vanadium Dioxide (VO2)\nfilms on their surface in various spatial configurations. VO2 films can be used\nto selectively transmit NIR light and can, therefore, be used to regulate the\namount of NIR light from the object that is captured by the iris sensor. We\nstudy the impact of such images produced by the sensor on two state-of-the-art\niris PA detection methods. We observe that the addition of VO2 films on the\nsurface of artificial eyes can cause the PA detection methods to misclassify\nthem as bonafide eyes in some cases. This represents a vulnerability that must\nbe systematically analyzed and effectively addressed.\n", "date": "2023-11-21", "categories": []}
{"id": "2306.07613", "title": "Revisiting and Advancing Adversarial Training Through A Simple Baseline", "abstract": "  In this paper, we delve into the essential components of adversarial training\nwhich is a pioneering defense technique against adversarial attacks. We\nindicate that some factors such as the loss function, learning rate scheduler,\nand data augmentation, which are independent of the model architecture, will\ninfluence adversarial robustness and generalization. When these factors are\ncontrolled for, we introduce a simple baseline approach, termed SimpleAT, that\nperforms competitively with recent methods and mitigates robust overfitting. We\nconduct extensive experiments on CIFAR-10/100 and Tiny-ImageNet, which validate\nthe robustness of SimpleAT against state-of-the-art adversarial attackers such\nas AutoAttack. Our results also demonstrate that SimpleAT exhibits good\nperformance in the presence of various image corruptions, such as those found\nin the CIFAR-10-C. In addition, we empirically show that SimpleAT is capable of\nreducing the variance in model predictions, which is considered the primary\ncontributor to robust overfitting. Our results also reveal the connections\nbetween SimpleAT and many advanced state-of-the-art adversarial defense\nmethods.\n", "date": "2023-06-13", "categories": []}
{"id": "2307.02500", "title": "Interpretable Computer Vision Models through Adversarial Training:\n  Unveiling the Robustness-Interpretability Connection", "abstract": "  With the perpetual increase of complexity of the state-of-the-art deep neural\nnetworks, it becomes a more and more challenging task to maintain their\ninterpretability. Our work aims to evaluate the effects of adversarial training\nutilized to produce robust models - less vulnerable to adversarial attacks. It\nhas been shown to make computer vision models more interpretable.\nInterpretability is as essential as robustness when we deploy the models to the\nreal world. To prove the correlation between these two problems, we extensively\nexamine the models using local feature-importance methods (SHAP, Integrated\nGradients) and feature visualization techniques (Representation Inversion,\nClass Specific Image Generation). Standard models, compared to robust are more\nsusceptible to adversarial attacks, and their learned representations are less\nmeaningful to humans. Conversely, these models focus on distinctive regions of\nthe images which support their predictions. Moreover, the features learned by\nthe robust model are closer to the real ones.\n", "date": "2023-07-04", "categories": []}
{"id": "2310.03358", "title": "Enhancing Robust Representation in Adversarial Training: Alignment and\n  Exclusion Criteria", "abstract": "  Deep neural networks are vulnerable to adversarial noise. Adversarial\nTraining (AT) has been demonstrated to be the most effective defense strategy\nto protect neural networks from being fooled. However, we find AT omits to\nlearning robust features, resulting in poor performance of adversarial\nrobustness. To address this issue, we highlight two criteria of robust\nrepresentation: (1) Exclusion: \\emph{the feature of examples keeps away from\nthat of other classes}; (2) Alignment: \\emph{the feature of natural and\ncorresponding adversarial examples is close to each other}. These motivate us\nto propose a generic framework of AT to gain robust representation, by the\nasymmetric negative contrast and reverse attention. Specifically, we design an\nasymmetric negative contrast based on predicted probabilities, to push away\nexamples of different classes in the feature space. Moreover, we propose to\nweight feature by parameters of the linear classifier as the reverse attention,\nto obtain class-aware feature and pull close the feature of the same class.\nEmpirical evaluations on three benchmark datasets show our methods greatly\nadvance the robustness of AT and achieve state-of-the-art performance.\n", "date": "2023-10-05", "categories": []}
{"id": "2311.07780", "title": "Parrot-Trained Adversarial Examples: Pushing the Practicality of\n  Black-Box Audio Attacks against Speaker Recognition Models", "abstract": "  Audio adversarial examples (AEs) have posed significant security challenges\nto real-world speaker recognition systems. Most black-box attacks still require\ncertain information from the speaker recognition model to be effective (e.g.,\nkeeping probing and requiring the knowledge of similarity scores). This work\naims to push the practicality of the black-box attacks by minimizing the\nattacker's knowledge about a target speaker recognition model. Although it is\nnot feasible for an attacker to succeed with completely zero knowledge, we\nassume that the attacker only knows a short (or a few seconds) speech sample of\na target speaker. Without any probing to gain further knowledge about the\ntarget model, we propose a new mechanism, called parrot training, to generate\nAEs against the target model. Motivated by recent advancements in voice\nconversion (VC), we propose to use the one short sentence knowledge to generate\nmore synthetic speech samples that sound like the target speaker, called parrot\nspeech. Then, we use these parrot speech samples to train a parrot-trained(PT)\nsurrogate model for the attacker. Under a joint transferability and perception\nframework, we investigate different ways to generate AEs on the PT model\n(called PT-AEs) to ensure the PT-AEs can be generated with high transferability\nto a black-box target model with good human perceptual quality. Real-world\nexperiments show that the resultant PT-AEs achieve the attack success rates of\n45.8% - 80.8% against the open-source models in the digital-line scenario and\n47.9% - 58.3% against smart devices, including Apple HomePod (Siri), Amazon\nEcho, and Google Home, in the over-the-air scenario.\n", "date": "2023-11-13", "categories": []}
{"id": "2311.10919", "title": "PACOL: Poisoning Attacks Against Continual Learners", "abstract": "  Continual learning algorithms are typically exposed to untrusted sources that\ncontain training data inserted by adversaries and bad actors. An adversary can\ninsert a small number of poisoned samples, such as mislabeled samples from\npreviously learned tasks, or intentional adversarial perturbed samples, into\nthe training datasets, which can drastically reduce the model's performance. In\nthis work, we demonstrate that continual learning systems can be manipulated by\nmalicious misinformation and present a new category of data poisoning attacks\nspecific for continual learners, which we refer to as {\\em Poisoning Attacks\nAgainst Continual Learners} (PACOL). The effectiveness of labeling flipping\nattacks inspires PACOL; however, PACOL produces attack samples that do not\nchange the sample's label and produce an attack that causes catastrophic\nforgetting. A comprehensive set of experiments shows the vulnerability of\ncommonly used generative replay and regularization-based continual learning\napproaches against attack methods. We evaluate the ability of label-flipping\nand a new adversarial poison attack, namely PACOL proposed in this work, to\nforce the continual learning system to forget the knowledge of a learned\ntask(s). More specifically, we compared the performance degradation of\ncontinual learning systems trained on benchmark data streams with and without\npoisoning attacks. Moreover, we discuss the stealthiness of the attacks in\nwhich we test the success rate of data sanitization defense and other outlier\ndetection-based defenses for filtering out adversarial samples.\n", "date": "2023-11-18", "categories": []}
{"id": "2311.11017", "title": "Improving Adversarial Transferability by Stable Diffusion", "abstract": "  Deep neural networks (DNNs) are susceptible to adversarial examples, which\nintroduce imperceptible perturbations to benign samples, deceiving DNN\npredictions. While some attack methods excel in the white-box setting, they\noften struggle in the black-box scenario, particularly against models fortified\nwith defense mechanisms. Various techniques have emerged to enhance the\ntransferability of adversarial attacks for the black-box scenario. Among these,\ninput transformation-based attacks have demonstrated their effectiveness. In\nthis paper, we explore the potential of leveraging data generated by Stable\nDiffusion to boost adversarial transferability. This approach draws inspiration\nfrom recent research that harnessed synthetic data generated by Stable\nDiffusion to enhance model generalization. In particular, previous work has\nhighlighted the correlation between the presence of both real and synthetic\ndata and improved model generalization. Building upon this insight, we\nintroduce a novel attack method called Stable Diffusion Attack Method (SDAM),\nwhich incorporates samples generated by Stable Diffusion to augment input\nimages. Furthermore, we propose a fast variant of SDAM to reduce computational\noverhead while preserving high adversarial transferability. Our extensive\nexperimental results demonstrate that our method outperforms state-of-the-art\nbaselines by a substantial margin. Moreover, our approach is compatible with\nexisting transfer-based attacks to further enhance adversarial transferability.\n", "date": "2023-11-18", "categories": []}
{"id": "2311.11544", "title": "Understanding Variation in Subpopulation Susceptibility to Poisoning\n  Attacks", "abstract": "  Machine learning is susceptible to poisoning attacks, in which an attacker\ncontrols a small fraction of the training data and chooses that data with the\ngoal of inducing some behavior unintended by the model developer in the trained\nmodel. We consider a realistic setting in which the adversary with the ability\nto insert a limited number of data points attempts to control the model's\nbehavior on a specific subpopulation. Inspired by previous observations on\ndisparate effectiveness of random label-flipping attacks on different\nsubpopulations, we investigate the properties that can impact the effectiveness\nof state-of-the-art poisoning attacks against different subpopulations. For a\nfamily of 2-dimensional synthetic datasets, we empirically find that dataset\nseparability plays a dominant role in subpopulation vulnerability for less\nseparable datasets. However, well-separated datasets exhibit more dependence on\nindividual subpopulation properties. We further discover that a crucial\nsubpopulation property is captured by the difference in loss on the clean\ndataset between the clean model and a target model that misclassifies the\nsubpopulation, and a subpopulation is much easier to attack if the loss\ndifference is small. This property also generalizes to high-dimensional\nbenchmark datasets. For the Adult benchmark dataset, we show that we can find\nsemantically-meaningful subpopulation properties that are related to the\nsusceptibilities of a selected group of subpopulations. The results in this\npaper are accompanied by a fully interactive web-based visualization of\nsubpopulation poisoning attacks found at\nhttps://uvasrg.github.io/visualizing-poisoning\n", "date": "2023-11-20", "categories": []}
{"id": "2311.11753", "title": "AdvGen: Physical Adversarial Attack on Face Presentation Attack\n  Detection Systems", "abstract": "  Evaluating the risk level of adversarial images is essential for safely\ndeploying face authentication models in the real world. Popular approaches for\nphysical-world attacks, such as print or replay attacks, suffer from some\nlimitations, like including physical and geometrical artifacts. Recently,\nadversarial attacks have gained attraction, which try to digitally deceive the\nlearning strategy of a recognition system using slight modifications to the\ncaptured image. While most previous research assumes that the adversarial image\ncould be digitally fed into the authentication systems, this is not always the\ncase for systems deployed in the real world. This paper demonstrates the\nvulnerability of face authentication systems to adversarial images in physical\nworld scenarios. We propose AdvGen, an automated Generative Adversarial\nNetwork, to simulate print and replay attacks and generate adversarial images\nthat can fool state-of-the-art PADs in a physical domain attack setting. Using\nthis attack strategy, the attack success rate goes up to 82.01%. We test AdvGen\nextensively on four datasets and ten state-of-the-art PADs. We also demonstrate\nthe effectiveness of our attack by conducting experiments in a realistic,\nphysical environment.\n", "date": "2023-11-20", "categories": []}
{"id": "2311.11861", "title": "Generating Valid and Natural Adversarial Examples with Large Language\n  Models", "abstract": "  Deep learning-based natural language processing (NLP) models, particularly\npre-trained language models (PLMs), have been revealed to be vulnerable to\nadversarial attacks. However, the adversarial examples generated by many\nmainstream word-level adversarial attack models are neither valid nor natural,\nleading to the loss of semantic maintenance, grammaticality, and human\nimperceptibility. Based on the exceptional capacity of language understanding\nand generation of large language models (LLMs), we propose LLM-Attack, which\naims at generating both valid and natural adversarial examples with LLMs. The\nmethod consists of two stages: word importance ranking (which searches for the\nmost vulnerable words) and word synonym replacement (which substitutes them\nwith their synonyms obtained from LLMs). Experimental results on the Movie\nReview (MR), IMDB, and Yelp Review Polarity datasets against the baseline\nadversarial attack models illustrate the effectiveness of LLM-Attack, and it\noutperforms the baselines in human and GPT-4 evaluation by a significant\nmargin. The model can generate adversarial examples that are typically valid\nand natural, with the preservation of semantic meaning, grammaticality, and\nhuman imperceptibility.\n", "date": "2023-11-20", "categories": []}
{"id": "2311.10366", "title": "Breaking Temporal Consistency: Generating Video Universal Adversarial\n  Perturbations Using Image Models", "abstract": "  As video analysis using deep learning models becomes more widespread, the\nvulnerability of such models to adversarial attacks is becoming a pressing\nconcern. In particular, Universal Adversarial Perturbation (UAP) poses a\nsignificant threat, as a single perturbation can mislead deep learning models\non entire datasets. We propose a novel video UAP using image data and image\nmodel. This enables us to take advantage of the rich image data and image\nmodel-based studies available for video applications. However, there is a\nchallenge that image models are limited in their ability to analyze the\ntemporal aspects of videos, which is crucial for a successful video attack. To\naddress this challenge, we introduce the Breaking Temporal Consistency (BTC)\nmethod, which is the first attempt to incorporate temporal information into\nvideo attacks using image models. We aim to generate adversarial videos that\nhave opposite patterns to the original. Specifically, BTC-UAP minimizes the\nfeature similarity between neighboring frames in videos. Our approach is simple\nbut effective at attacking unseen video models. Additionally, it is applicable\nto videos of varying lengths and invariant to temporal shifts. Our approach\nsurpasses existing methods in terms of effectiveness on various datasets,\nincluding ImageNet, UCF-101, and Kinetics-400.\n", "date": "2023-11-17", "categories": []}
{"id": "2311.09473", "title": "JAB: Joint Adversarial Prompting and Belief Augmentation", "abstract": "  With the recent surge of language models in different applications, attention\nto safety and robustness of these models has gained significant importance.\nHere we introduce a joint framework in which we simultaneously probe and\nimprove the robustness of a black-box target model via adversarial prompting\nand belief augmentation using iterative feedback loops. This framework utilizes\nan automated red teaming approach to probe the target model, along with a\nbelief augmenter to generate instructions for the target model to improve its\nrobustness to those adversarial probes. Importantly, the adversarial model and\nthe belief generator leverage the feedback from past interactions to improve\nthe effectiveness of the adversarial prompts and beliefs, respectively. In our\nexperiments, we demonstrate that such a framework can reduce toxic content\ngeneration both in dynamic cases where an adversary directly interacts with a\ntarget model and static cases where we use a static benchmark dataset to\nevaluate our model.\n", "date": "2023-11-16", "categories": []}
{"id": "2310.14421", "title": "On existence, uniqueness and scalability of adversarial robustness\n  measures for AI classifiers", "abstract": "  Simply-verifiable mathematical conditions for existence, uniqueness and\nexplicit analytical computation of minimal adversarial paths (MAP) and minimal\nadversarial distances (MAD) for (locally) uniquely-invertible classifiers, for\ngeneralized linear models (GLM), and for entropic AI (EAI) are formulated and\nproven. Practical computation of MAP and MAD, their comparison and\ninterpretations for various classes of AI tools (for neuronal networks, boosted\nrandom forests, GLM and EAI) are demonstrated on the common synthetic\nbenchmarks: on a double Swiss roll spiral and its extensions, as well as on the\ntwo biomedical data problems (for the health insurance claim predictions, and\nfor the heart attack lethality classification). On biomedical applications it\nis demonstrated how MAP provides unique minimal patient-specific\nrisk-mitigating interventions in the predefined subsets of accessible control\nvariables.\n", "date": "2023-10-19", "categories": []}
{"id": "2311.08539", "title": "Physical Adversarial Examples for Multi-Camera Systems", "abstract": "  Neural networks build the foundation of several intelligent systems, which,\nhowever, are known to be easily fooled by adversarial examples. Recent advances\nmade these attacks possible even in air-gapped scenarios, where the autonomous\nsystem observes its surroundings by, e.g., a camera. We extend these ideas in\nour research and evaluate the robustness of multi-camera setups against such\nphysical adversarial examples. This scenario becomes ever more important with\nthe rise in popularity of autonomous vehicles, which fuse the information of\nseveral cameras for their driving decision. While we find that multi-camera\nsetups provide some robustness towards past attack methods, we see that this\nadvantage reduces when optimizing on multiple perspectives at once. We propose\na novel attack method that we call Transcender-MC, where we incorporate online\n3D renderings and perspective projections in the training process. Moreover, we\nmotivate that certain data augmentation techniques can facilitate the\ngeneration of successful adversarial examples even further. Transcender-MC is\n11% more effective in successfully attacking multi-camera setups than\nstate-of-the-art methods. Our findings offer valuable insights regarding the\nresilience of object detection in a setup with multiple cameras and motivate\nthe need of developing adequate defense mechanisms against them.\n", "date": "2023-11-14", "categories": []}
{"id": "2311.09014", "title": "Adversarial Attacks to Reward Machine-based Reinforcement Learning", "abstract": "  In recent years, Reward Machines (RMs) have stood out as a simple yet\neffective automata-based formalism for exposing and exploiting task structure\nin reinforcement learning settings. Despite their relevance, little to no\nattention has been directed to the study of their security implications and\nrobustness to adversarial scenarios, likely due to their recent appearance in\nthe literature. With my thesis, I aim to provide the first analysis of the\nsecurity of RM-based reinforcement learning techniques, with the hope of\nmotivating further research in the field, and I propose and evaluate a novel\nclass of attacks on RM-based techniques: blinding attacks.\n", "date": "2023-11-15", "categories": []}
{"id": "2311.07928", "title": "Towards Improving Robustness Against Common Corruptions in Object\n  Detectors Using Adversarial Contrastive Learning", "abstract": "  Neural networks have revolutionized various domains, exhibiting remarkable\naccuracy in tasks like natural language processing and computer vision.\nHowever, their vulnerability to slight alterations in input samples poses\nchallenges, particularly in safety-critical applications like autonomous\ndriving. Current approaches, such as introducing distortions during training,\nfall short in addressing unforeseen corruptions. This paper proposes an\ninnovative adversarial contrastive learning framework to enhance neural network\nrobustness simultaneously against adversarial attacks and common corruptions.\nBy generating instance-wise adversarial examples and optimizing contrastive\nloss, our method fosters representations that resist adversarial perturbations\nand remain robust in real-world scenarios. Subsequent contrastive learning then\nstrengthens the similarity between clean samples and their adversarial\ncounterparts, fostering representations resistant to both adversarial attacks\nand common distortions. By focusing on improving performance under adversarial\nand real-world conditions, our approach aims to bolster the robustness of\nneural networks in safety-critical applications, such as autonomous vehicles\nnavigating unpredictable weather conditions. We anticipate that this framework\nwill contribute to advancing the reliability of neural networks in challenging\nenvironments, facilitating their widespread adoption in mission-critical\nscenarios.\n", "date": "2023-11-14", "categories": []}
{"id": "2311.08265", "title": "On The Relationship Between Universal Adversarial Attacks And Sparse\n  Representations", "abstract": "  The prominent success of neural networks, mainly in computer vision tasks, is\nincreasingly shadowed by their sensitivity to small, barely perceivable\nadversarial perturbations in image input.\n  In this work, we aim at explaining this vulnerability through the framework\nof sparsity.\n  We show the connection between adversarial attacks and sparse\nrepresentations, with a focus on explaining the universality and\ntransferability of adversarial examples in neural networks.\n  To this end, we show that sparse coding algorithms, and the neural\nnetwork-based learned iterative shrinkage thresholding algorithm (LISTA) among\nthem, suffer from this sensitivity, and that common attacks on neural networks\ncan be expressed as attacks on the sparse representation of the input image.\nThe phenomenon that we observe holds true also when the network is agnostic to\nthe sparse representation and dictionary, and thus can provide a possible\nexplanation for the universality and transferability of adversarial attacks.\n  The code is available at\nhttps://github.com/danawr/adversarial_attacks_and_sparse_representations.\n", "date": "2023-11-14", "categories": []}
{"id": "2305.03626", "title": "Verifiable Learning for Robust Tree Ensembles", "abstract": "  Verifying the robustness of machine learning models against evasion attacks\nat test time is an important research problem. Unfortunately, prior work\nestablished that this problem is NP-hard for decision tree ensembles, hence\nbound to be intractable for specific inputs. In this paper, we identify a\nrestricted class of decision tree ensembles, called large-spread ensembles,\nwhich admit a security verification algorithm running in polynomial time. We\nthen propose a new approach called verifiable learning, which advocates the\ntraining of such restricted model classes which are amenable for efficient\nverification. We show the benefits of this idea by designing a new training\nalgorithm that automatically learns a large-spread decision tree ensemble from\nlabelled data, thus enabling its security verification in polynomial time.\nExperimental results on public datasets confirm that large-spread ensembles\ntrained using our algorithm can be verified in a matter of seconds, using\nstandard commercial hardware. Moreover, large-spread ensembles are more robust\nthan traditional ensembles against evasion attacks, at the cost of an\nacceptable loss of accuracy in the non-adversarial setting.\n", "date": "2023-05-05", "categories": []}
{"id": "2305.12082", "title": "SneakyPrompt: Jailbreaking Text-to-image Generative Models", "abstract": "  Text-to-image generative models such as Stable Diffusion and DALL$\\cdot$E\nraise many ethical concerns due to the generation of harmful images such as\nNot-Safe-for-Work (NSFW) ones. To address these ethical concerns, safety\nfilters are often adopted to prevent the generation of NSFW images. In this\nwork, we propose SneakyPrompt, the first automated attack framework, to\njailbreak text-to-image generative models such that they generate NSFW images\neven if safety filters are adopted. Given a prompt that is blocked by a safety\nfilter, SneakyPrompt repeatedly queries the text-to-image generative model and\nstrategically perturbs tokens in the prompt based on the query results to\nbypass the safety filter. Specifically, SneakyPrompt utilizes reinforcement\nlearning to guide the perturbation of tokens. Our evaluation shows that\nSneakyPrompt successfully jailbreaks DALL$\\cdot$E 2 with closed-box safety\nfilters to generate NSFW images. Moreover, we also deploy several\nstate-of-the-art, open-source safety filters on a Stable Diffusion model. Our\nevaluation shows that SneakyPrompt not only successfully generates NSFW images,\nbut also outperforms existing text adversarial attacks when extended to\njailbreak text-to-image generative models, in terms of both the number of\nqueries and qualities of the generated NSFW images. SneakyPrompt is open-source\nand available at this repository:\n\\url{https://github.com/Yuchen413/text2image_safety}.\n", "date": "2023-05-20", "categories": []}
{"id": "2311.06361", "title": "CALLOC: Curriculum Adversarial Learning for Secure and Robust Indoor\n  Localization", "abstract": "  Indoor localization has become increasingly vital for many applications from\ntracking assets to delivering personalized services. Yet, achieving pinpoint\naccuracy remains a challenge due to variations across indoor environments and\ndevices used to assist with localization. Another emerging challenge is\nadversarial attacks on indoor localization systems that not only threaten\nservice integrity but also reduce localization accuracy. To combat these\nchallenges, we introduce CALLOC, a novel framework designed to resist\nadversarial attacks and variations across indoor environments and devices that\nreduce system accuracy and reliability. CALLOC employs a novel adaptive\ncurriculum learning approach with a domain specific lightweight scaled-dot\nproduct attention neural network, tailored for adversarial and variation\nresilience in practical use cases with resource constrained mobile devices.\nExperimental evaluations demonstrate that CALLOC can achieve improvements of up\nto 6.03x in mean error and 4.6x in worst-case error against state-of-the-art\nindoor localization frameworks, across diverse building floorplans, mobile\ndevices, and adversarial attacks scenarios.\n", "date": "2023-11-10", "categories": []}
{"id": "2311.07417", "title": "Mitigating Backdoors within Deep Neural Networks in Data-limited\n  Configuration", "abstract": "  As the capacity of deep neural networks (DNNs) increases, their need for huge\namounts of data significantly grows. A common practice is to outsource the\ntraining process or collect more data over the Internet, which introduces the\nrisks of a backdoored DNN. A backdoored DNN shows normal behavior on clean data\nwhile behaving maliciously once a trigger is injected into a sample at the test\ntime. In such cases, the defender faces multiple difficulties. First, the\navailable clean dataset may not be sufficient for fine-tuning and recovering\nthe backdoored DNN. Second, it is impossible to recover the trigger in many\nreal-world applications without information about it. In this paper, we\nformulate some characteristics of poisoned neurons. This backdoor\nsuspiciousness score can rank network neurons according to their activation\nvalues, weights, and their relationship with other neurons in the same layer.\nOur experiments indicate the proposed method decreases the chance of attacks\nbeing successful by more than 50% with a tiny clean dataset, i.e., ten clean\nsamples for the CIFAR-10 dataset, without significantly deteriorating the\nmodel's performance. Moreover, the proposed method runs three times as fast as\nbaselines.\n", "date": "2023-11-13", "categories": []}
{"id": "2307.01073", "title": "What Distributions are Robust to Indiscriminate Poisoning Attacks for\n  Linear Learners?", "abstract": "  We study indiscriminate poisoning for linear learners where an adversary\ninjects a few crafted examples into the training data with the goal of forcing\nthe induced model to incur higher test error. Inspired by the observation that\nlinear learners on some datasets are able to resist the best known attacks even\nwithout any defenses, we further investigate whether datasets can be inherently\nrobust to indiscriminate poisoning attacks for linear learners. For theoretical\nGaussian distributions, we rigorously characterize the behavior of an optimal\npoisoning attack, defined as the poisoning strategy that attains the maximum\nrisk of the induced model at a given poisoning budget. Our results prove that\nlinear learners can indeed be robust to indiscriminate poisoning if the\nclass-wise data distributions are well-separated with low variance and the size\nof the constraint set containing all permissible poisoning points is also\nsmall. These findings largely explain the drastic variation in empirical attack\nperformance of the state-of-the-art poisoning attacks on linear learners across\nbenchmark datasets, making an important initial step towards understanding the\nunderlying reasons some learning tasks are vulnerable to data poisoning\nattacks.\n", "date": "2023-07-03", "categories": []}
{"id": "2311.05992", "title": "Robust Adversarial Attacks Detection for Deep Learning based Relative\n  Pose Estimation for Space Rendezvous", "abstract": "  Research on developing deep learning techniques for autonomous spacecraft\nrelative navigation challenges is continuously growing in recent years.\nAdopting those techniques offers enhanced performance. However, such approaches\nalso introduce heightened apprehensions regarding the trustability and security\nof such deep learning methods through their susceptibility to adversarial\nattacks. In this work, we propose a novel approach for adversarial attack\ndetection for deep neural network-based relative pose estimation schemes based\non the explainability concept. We develop for an orbital rendezvous scenario an\ninnovative relative pose estimation technique adopting our proposed\nConvolutional Neural Network (CNN), which takes an image from the chaser's\nonboard camera and outputs accurately the target's relative position and\nrotation. We perturb seamlessly the input images using adversarial attacks that\nare generated by the Fast Gradient Sign Method (FGSM). The adversarial attack\ndetector is then built based on a Long Short Term Memory (LSTM) network which\ntakes the explainability measure namely SHapley Value from the CNN-based pose\nestimator and flags the detection of adversarial attacks when acting.\nSimulation results show that the proposed adversarial attack detector achieves\na detection accuracy of 99.21%. Both the deep relative pose estimator and\nadversarial attack detector are then tested on real data captured from our\nlaboratory-designed setup. The experimental results from our\nlaboratory-designed setup demonstrate that the proposed adversarial attack\ndetector achieves an average detection accuracy of 96.29%.\n", "date": "2023-11-10", "categories": []}
{"id": "2302.01056", "title": "Beyond Pretrained Features: Noisy Image Modeling Provides Adversarial\n  Defense", "abstract": "  Recent advancements in masked image modeling (MIM) have made it a prevailing\nframework for self-supervised visual representation learning. The MIM\npretrained models, like most deep neural network methods, remain vulnerable to\nadversarial attacks, limiting their practical application, and this issue has\nreceived little research attention. In this paper, we investigate how this\npowerful self-supervised learning paradigm can provide adversarial robustness\nto downstream classifiers. During the exploration, we find that noisy image\nmodeling (NIM), a simple variant of MIM that adopts denoising as the pre-text\ntask, reconstructs noisy images surprisingly well despite severe corruption.\nMotivated by this observation, we propose an adversarial defense method,\nreferred to as De^3, by exploiting the pretrained decoder for denoising.\nThrough De^3, NIM is able to enhance adversarial robustness beyond providing\npretrained features. Furthermore, we incorporate a simple modification,\nsampling the noise scale hyperparameter from random distributions, and enable\nthe defense to achieve a better and tunable trade-off between accuracy and\nrobustness. Experimental results demonstrate that, in terms of adversarial\nrobustness, NIM is superior to MIM thanks to its effective denoising\ncapability. Moreover, the defense provided by NIM achieves performance on par\nwith adversarial training while offering the extra tunability advantage. Source\ncode and models are available at https://github.com/youzunzhi/NIM-AdvDef.\n", "date": "2023-02-02", "categories": []}
{"id": "2305.03807", "title": "Evading Watermark based Detection of AI-Generated Content", "abstract": "  A generative AI model can generate extremely realistic-looking content,\nposing growing challenges to the authenticity of information. To address the\nchallenges, watermark has been leveraged to detect AI-generated content.\nSpecifically, a watermark is embedded into an AI-generated content before it is\nreleased. A content is detected as AI-generated if a similar watermark can be\ndecoded from it. In this work, we perform a systematic study on the robustness\nof such watermark-based AI-generated content detection. We focus on\nAI-generated images. Our work shows that an attacker can post-process a\nwatermarked image via adding a small, human-imperceptible perturbation to it,\nsuch that the post-processed image evades detection while maintaining its\nvisual quality. We show the effectiveness of our attack both theoretically and\nempirically. Moreover, to evade detection, our adversarial post-processing\nmethod adds much smaller perturbations to AI-generated images and thus better\nmaintain their visual quality than existing popular post-processing methods\nsuch as JPEG compression, Gaussian blur, and Brightness/Contrast. Our work\nshows the insufficiency of existing watermark-based detection of AI-generated\ncontent, highlighting the urgent needs of new methods. Our code is publicly\navailable: https://github.com/zhengyuan-jiang/WEvade.\n", "date": "2023-05-05", "categories": []}
{"id": "2306.06136", "title": "Robustness Testing for Multi-Agent Reinforcement Learning: State\n  Perturbations on Critical Agents", "abstract": "  Multi-Agent Reinforcement Learning (MARL) has been widely applied in many\nfields such as smart traffic and unmanned aerial vehicles. However, most MARL\nalgorithms are vulnerable to adversarial perturbations on agent states.\nRobustness testing for a trained model is an essential step for confirming the\ntrustworthiness of the model against unexpected perturbations. This work\nproposes a novel Robustness Testing framework for MARL that attacks states of\nCritical Agents (RTCA). The RTCA has two innovations: 1) a Differential\nEvolution (DE) based method to select critical agents as victims and to advise\nthe worst-case joint actions on them; and 2) a team cooperation policy\nevaluation method employed as the objective function for the optimization of\nDE. Then, adversarial state perturbations of the critical agents are generated\nbased on the worst-case joint actions. This is the first robustness testing\nframework with varying victim agents. RTCA demonstrates outstanding performance\nin terms of the number of victim agents and destroying cooperation policies.\n", "date": "2023-06-09", "categories": []}
{"id": "2310.18626", "title": "Benchmark Generation Framework with Customizable Distortions for Image\n  Classifier Robustness", "abstract": "  We present a novel framework for generating adversarial benchmarks to\nevaluate the robustness of image classification models. Our framework allows\nusers to customize the types of distortions to be optimally applied to images,\nwhich helps address the specific distortions relevant to their deployment. The\nbenchmark can generate datasets at various distortion levels to assess the\nrobustness of different image classifiers. Our results show that the\nadversarial samples generated by our framework with any of the image\nclassification models, like ResNet-50, Inception-V3, and VGG-16, are effective\nand transferable to other models causing them to fail. These failures happen\neven when these models are adversarially retrained using state-of-the-art\ntechniques, demonstrating the generalizability of our adversarial samples. We\nachieve competitive performance in terms of net $L_2$ distortion compared to\nstate-of-the-art benchmark techniques on CIFAR-10 and ImageNet; however, we\ndemonstrate our framework achieves such results with simple distortions like\nGaussian noise without introducing unnatural artifacts or color bleeds. This is\nmade possible by a model-based reinforcement learning (RL) agent and a\ntechnique that reduces a deep tree search of the image for model sensitivity to\nperturbations, to a one-level analysis and action. The flexibility of choosing\ndistortions and setting classification probability thresholds for multiple\nclasses makes our framework suitable for algorithmic audits.\n", "date": "2023-10-28", "categories": []}
{"id": "2311.04503", "title": "Constrained Adaptive Attacks: Realistic Evaluation of Adversarial\n  Examples and Robust Training of Deep Neural Networks for Tabular Data", "abstract": "  State-of-the-art deep learning models for tabular data have recently achieved\nacceptable performance to be deployed in industrial settings. However, the\nrobustness of these models remains scarcely explored. Contrary to computer\nvision, there is to date no realistic protocol to properly evaluate the\nadversarial robustness of deep tabular models due to intrinsic properties of\ntabular data such as categorical features, immutability, and feature\nrelationship constraints. To fill this gap, we propose CAA, the first efficient\nevasion attack for constrained tabular deep learning models. CAA is an\niterative parameter-free attack that combines gradient and search attacks to\ngenerate adversarial examples under constraints. We leverage CAA to build a\nbenchmark of deep tabular models across three popular use cases: credit\nscoring, phishing and botnet attacks detection. Our benchmark supports ten\nthreat models with increasing capabilities of the attacker, and reflects\nreal-world attack scenarios for each use case. Overall, our results demonstrate\nhow domain knowledge, adversarial training, and attack budgets impact the\nrobustness assessment of deep tabular models and provide security practitioners\nwith a set of recommendations to improve the robustness of deep tabular models\nagainst various evasion attack scenarios.\n", "date": "2023-11-08", "categories": []}
{"id": "2209.15042", "title": "Generalizability of Adversarial Robustness Under Distribution Shifts", "abstract": "  Recent progress in empirical and certified robustness promises to deliver\nreliable and deployable Deep Neural Networks (DNNs). Despite that success, most\nexisting evaluations of DNN robustness have been done on images sampled from\nthe same distribution on which the model was trained. However, in the real\nworld, DNNs may be deployed in dynamic environments that exhibit significant\ndistribution shifts. In this work, we take a first step towards thoroughly\ninvestigating the interplay between empirical and certified adversarial\nrobustness on one hand and domain generalization on another. To do so, we train\nrobust models on multiple domains and evaluate their accuracy and robustness on\nan unseen domain. We observe that: (1) both empirical and certified robustness\ngeneralize to unseen domains, and (2) the level of generalizability does not\ncorrelate well with input visual similarity, measured by the FID between source\nand target domains. We also extend our study to cover a real-world medical\napplication, in which adversarial augmentation significantly boosts the\ngeneralization of robustness with minimal effect on clean data accuracy.\n", "date": "2022-09-29", "categories": []}
{"id": "2210.09364", "title": "Probabilistic Categorical Adversarial Attack & Adversarial Training", "abstract": "  The existence of adversarial examples brings huge concern for people to apply\nDeep Neural Networks (DNNs) in safety-critical tasks. However, how to generate\nadversarial examples with categorical data is an important problem but lack of\nextensive exploration. Previously established methods leverage greedy search\nmethod, which can be very time-consuming to conduct successful attack. This\nalso limits the development of adversarial training and potential defenses for\ncategorical data. To tackle this problem, we propose Probabilistic Categorical\nAdversarial Attack (PCAA), which transfers the discrete optimization problem to\na continuous problem that can be solved efficiently by Projected Gradient\nDescent. In our paper, we theoretically analyze its optimality and time\ncomplexity to demonstrate its significant advantage over current greedy based\nattacks. Moreover, based on our attack, we propose an efficient adversarial\ntraining framework. Through a comprehensive empirical study, we justify the\neffectiveness of our proposed attack and defense algorithms.\n", "date": "2022-10-17", "categories": []}
{"id": "2309.00543", "title": "Curating Naturally Adversarial Datasets for Learning-Enabled Medical\n  Cyber-Physical Systems", "abstract": "  Deep learning models have shown promising predictive accuracy for time-series\nhealthcare applications. However, ensuring the robustness of these models is\nvital for building trustworthy AI systems. Existing research predominantly\nfocuses on robustness to synthetic adversarial examples, crafted by adding\nimperceptible perturbations to clean input data. However, these synthetic\nadversarial examples do not accurately reflect the most challenging real-world\nscenarios, especially in the context of healthcare data. Consequently,\nrobustness to synthetic adversarial examples may not necessarily translate to\nrobustness against naturally occurring adversarial examples, which is highly\ndesirable for trustworthy AI. We propose a method to curate datasets comprised\nof natural adversarial examples to evaluate model robustness. The method relies\non probabilistic labels obtained from automated weakly-supervised labeling that\ncombines noisy and cheap-to-obtain labeling heuristics. Based on these labels,\nour method adversarially orders the input data and uses this ordering to\nconstruct a sequence of increasingly adversarial datasets. Our evaluation on\nsix medical case studies and three non-medical case studies demonstrates the\nefficacy and statistical validity of our approach to generating naturally\nadversarial datasets\n", "date": "2023-09-01", "categories": []}
{"id": "2311.03373", "title": "Unscrambling the Rectification of Adversarial Attacks Transferability\n  across Computer Networks", "abstract": "  Convolutional neural networks (CNNs) models play a vital role in achieving\nstate-of-the-art performances in various technological fields. CNNs are not\nlimited to Natural Language Processing (NLP) or Computer Vision (CV) but also\nhave substantial applications in other technological domains, particularly in\ncybersecurity. The reliability of CNN's models can be compromised because of\ntheir susceptibility to adversarial attacks, which can be generated\neffortlessly, easily applied, and transferred in real-world scenarios.\n  In this paper, we present a novel and comprehensive method to improve the\nstrength of attacks and assess the transferability of adversarial examples in\nCNNs when such strength changes, as well as whether the transferability\nproperty issue exists in computer network applications. In the context of our\nstudy, we initially examined six distinct modes of attack: the Carlini and\nWagner (C&W), Fast Gradient Sign Method (FGSM), Iterative Fast Gradient Sign\nMethod (I-FGSM), Jacobian-based Saliency Map (JSMA), Limited-memory Broyden\nfletcher Goldfarb Shanno (L-BFGS), and Projected Gradient Descent (PGD) attack.\nWe applied these attack techniques on two popular datasets: the CIC and UNSW\ndatasets. The outcomes of our experiment demonstrate that an improvement in\ntransferability occurs in the targeted scenarios for FGSM, JSMA, LBFGS, and\nother attacks. Our findings further indicate that the threats to security posed\nby adversarial examples, even in computer network applications, necessitate the\ndevelopment of novel defense mechanisms to enhance the security of DL-based\ntechniques.\n", "date": "2023-10-26", "categories": []}
{"id": "2311.03566", "title": "Measuring Adversarial Datasets", "abstract": "  In the era of widespread public use of AI systems across various domains,\nensuring adversarial robustness has become increasingly vital to maintain\nsafety and prevent undesirable errors. Researchers have curated various\nadversarial datasets (through perturbations) for capturing model deficiencies\nthat cannot be revealed in standard benchmark datasets. However, little is\nknown about how these adversarial examples differ from the original data\npoints, and there is still no methodology to measure the intended and\nunintended consequences of those adversarial transformations. In this research,\nwe conducted a systematic survey of existing quantifiable metrics that describe\ntext instances in NLP tasks, among dimensions of difficulty, diversity, and\ndisagreement. We selected several current adversarial effect datasets and\ncompared the distributions between the original and their adversarial\ncounterparts. The results provide valuable insights into what makes these\ndatasets more challenging from a metrics perspective and whether they align\nwith underlying assumptions.\n", "date": "2023-11-06", "categories": []}
{"id": "2310.14265", "title": "CT-GAT: Cross-Task Generative Adversarial Attack based on\n  Transferability", "abstract": "  Neural network models are vulnerable to adversarial examples, and adversarial\ntransferability further increases the risk of adversarial attacks. Current\nmethods based on transferability often rely on substitute models, which can be\nimpractical and costly in real-world scenarios due to the unavailability of\ntraining data and the victim model's structural details. In this paper, we\npropose a novel approach that directly constructs adversarial examples by\nextracting transferable features across various tasks. Our key insight is that\nadversarial transferability can extend across different tasks. Specifically, we\ntrain a sequence-to-sequence generative model named CT-GAT using adversarial\nsample data collected from multiple tasks to acquire universal adversarial\nfeatures and generate adversarial examples for different tasks. We conduct\nexperiments on ten distinct datasets, and the results demonstrate that our\nmethod achieves superior attack performance with small cost.\n", "date": "2023-10-22", "categories": []}
{"id": "2311.01478", "title": "Adversary ML Resilience in Autonomous Driving Through Human Centered\n  Perception Mechanisms", "abstract": "  Physical adversarial attacks on road signs are continuously exploiting\nvulnerabilities in modern day autonomous vehicles (AVs) and impeding their\nability to correctly classify what type of road sign they encounter. Current\nmodels cannot generalize input data well, resulting in overfitting or\nunderfitting. In overfitting, the model memorizes the input data but cannot\ngeneralize to new scenarios. In underfitting, the model does not learn enough\nof the input data to accurately classify these road signs. This paper explores\nthe resilience of autonomous driving systems against three main physical\nadversarial attacks (tape, graffiti, illumination), specifically targeting\nobject classifiers. Several machine learning models were developed and\nevaluated on two distinct datasets: road signs (stop signs, speed limit signs,\ntraffic lights, and pedestrian crosswalk signs) and geometric shapes (octagons,\ncircles, squares, and triangles). The study compared algorithm performance\nunder different conditions, including clean and adversarial training and\ntesting on these datasets. To build robustness against attacks, defense\ntechniques like adversarial training and transfer learning were implemented.\nResults demonstrated transfer learning models played a crucial role in\nperformance by allowing knowledge gained from shape training to improve\ngeneralizability of road sign classification, despite the datasets being\ncompletely different. The paper suggests future research directions, including\nhuman-in-the-loop validation, security analysis, real-world testing, and\nexplainable AI for transparency. This study aims to contribute to improving\nsecurity and robustness of object classifiers in autonomous vehicles and\nmitigating adversarial example impacts on driving systems.\n", "date": "2023-11-02", "categories": []}
{"id": "2311.01563", "title": "Assist Is Just as Important as the Goal: Image Resurfacing to Aid\n  Model's Robust Prediction", "abstract": "  Adversarial patches threaten visual AI models in the real world. The number\nof patches in a patch attack is variable and determines the attack's potency in\na specific environment. Most existing defenses assume a single patch in the\nscene, and the multiple patch scenarios are shown to overcome them. This paper\npresents a model-agnostic defense against patch attacks based on total\nvariation for image resurfacing (TVR). The TVR is an image-cleansing method\nthat processes images to remove probable adversarial regions. TVR can be\nutilized solely or augmented with a defended model, providing multi-level\nsecurity for robust prediction. TVR nullifies the influence of patches in a\nsingle image scan with no prior assumption on the number of patches in the\nscene. We validate TVR on the ImageNet-Patch benchmark dataset and with\nreal-world physical objects, demonstrating its ability to mitigate patch\nattack.\n", "date": "2023-11-02", "categories": []}
{"id": "2311.01642", "title": "Robust Adversarial Reinforcement Learning via Bounded Rationality\n  Curricula", "abstract": "  Robustness against adversarial attacks and distribution shifts is a\nlong-standing goal of Reinforcement Learning (RL). To this end, Robust\nAdversarial Reinforcement Learning (RARL) trains a protagonist against\ndestabilizing forces exercised by an adversary in a competitive zero-sum Markov\ngame, whose optimal solution, i.e., rational strategy, corresponds to a Nash\nequilibrium. However, finding Nash equilibria requires facing complex saddle\npoint optimization problems, which can be prohibitive to solve, especially for\nhigh-dimensional control. In this paper, we propose a novel approach for\nadversarial RL based on entropy regularization to ease the complexity of the\nsaddle point optimization problem. We show that the solution of this\nentropy-regularized problem corresponds to a Quantal Response Equilibrium\n(QRE), a generalization of Nash equilibria that accounts for bounded\nrationality, i.e., agents sometimes play random actions instead of optimal\nones. Crucially, the connection between the entropy-regularized objective and\nQRE enables free modulation of the rationality of the agents by simply tuning\nthe temperature coefficient. We leverage this insight to propose our novel\nalgorithm, Quantal Adversarial RL (QARL), which gradually increases the\nrationality of the adversary in a curriculum fashion until it is fully\nrational, easing the complexity of the optimization problem while retaining\nrobustness. We provide extensive evidence of QARL outperforming RARL and recent\nbaselines across several MuJoCo locomotion and navigation problems in overall\nperformance and robustness.\n", "date": "2023-11-03", "categories": []}
{"id": "2311.01698", "title": "Adversarial Attacks on Cooperative Multi-agent Bandits", "abstract": "  Cooperative multi-agent multi-armed bandits (CMA2B) consider the\ncollaborative efforts of multiple agents in a shared multi-armed bandit game.\nWe study latent vulnerabilities exposed by this collaboration and consider\nadversarial attacks on a few agents with the goal of influencing the decisions\nof the rest. More specifically, we study adversarial attacks on CMA2B in both\nhomogeneous settings, where agents operate with the same arm set, and\nheterogeneous settings, where agents have distinct arm sets. In the homogeneous\nsetting, we propose attack strategies that, by targeting just one agent,\nconvince all agents to select a particular target arm $T-o(T)$ times while\nincurring $o(T)$ attack costs in $T$ rounds. In the heterogeneous setting, we\nprove that a target arm attack requires linear attack costs and propose attack\nstrategies that can force a maximum number of agents to suffer linear regrets\nwhile incurring sublinear costs and only manipulating the observations of a few\ntarget agents. Numerical experiments validate the effectiveness of our proposed\nattack strategies.\n", "date": "2023-11-03", "categories": []}
{"id": "2311.01873", "title": "Efficient Black-Box Adversarial Attacks on Neural Text Detectors", "abstract": "  Neural text detectors are models trained to detect whether a given text was\ngenerated by a language model or written by a human. In this paper, we\ninvestigate three simple and resource-efficient strategies (parameter tweaking,\nprompt engineering, and character-level mutations) to alter texts generated by\nGPT-3.5 that are unsuspicious or unnoticeable for humans but cause\nmisclassification by neural text detectors. The results show that especially\nparameter tweaking and character-level mutations are effective strategies.\n", "date": "2023-11-03", "categories": []}
{"id": "2304.13410", "title": "Improving Adversarial Transferability via Intermediate-level\n  Perturbation Decay", "abstract": "  Intermediate-level attacks that attempt to perturb feature representations\nfollowing an adversarial direction drastically have shown favorable performance\nin crafting transferable adversarial examples. Existing methods in this\ncategory are normally formulated with two separate stages, where a directional\nguide is required to be determined at first and the scalar projection of the\nintermediate-level perturbation onto the directional guide is enlarged\nthereafter. The obtained perturbation deviates from the guide inevitably in the\nfeature space, and it is revealed in this paper that such a deviation may lead\nto sub-optimal attack. To address this issue, we develop a novel\nintermediate-level method that crafts adversarial examples within a single\nstage of optimization. In particular, the proposed method, named\nintermediate-level perturbation decay (ILPD), encourages the intermediate-level\nperturbation to be in an effective adversarial direction and to possess a great\nmagnitude simultaneously. In-depth discussion verifies the effectiveness of our\nmethod. Experimental results show that it outperforms state-of-the-arts by\nlarge margins in attacking various victim models on ImageNet (+10.07% on\naverage) and CIFAR-10 (+3.88% on average). Our code is at\nhttps://github.com/qizhangli/ILPD-attack.\n", "date": "2023-04-26", "categories": []}
{"id": "2306.05225", "title": "Boosting Adversarial Transferability by Achieving Flat Local Maxima", "abstract": "  Transfer-based attack adopts the adversarial examples generated on the\nsurrogate model to attack various models, making it applicable in the physical\nworld and attracting increasing interest. Recently, various adversarial attacks\nhave emerged to boost adversarial transferability from different perspectives.\nIn this work, inspired by the observation that flat local minima are correlated\nwith good generalization, we assume and empirically validate that adversarial\nexamples at a flat local region tend to have good transferability by\nintroducing a penalized gradient norm to the original loss function. Since\ndirectly optimizing the gradient regularization norm is computationally\nexpensive and intractable for generating adversarial examples, we propose an\napproximation optimization method to simplify the gradient update of the\nobjective function. Specifically, we randomly sample an example and adopt a\nfirst-order procedure to approximate the curvature of Hessian/vector product,\nwhich makes computing more efficient by interpolating two neighboring\ngradients. Meanwhile, in order to obtain a more stable gradient direction, we\nrandomly sample multiple examples and average the gradients of these examples\nto reduce the variance due to random sampling during the iterative process.\nExtensive experimental results on the ImageNet-compatible dataset show that the\nproposed method can generate adversarial examples at flat local regions, and\nsignificantly improve the adversarial transferability on either normally\ntrained models or adversarially trained models than the state-of-the-art\nattacks. Our codes are available at:\nhttps://github.com/Trustworthy-AI-Group/PGN.\n", "date": "2023-06-08", "categories": []}
{"id": "2310.17403", "title": "Detection Defenses: An Empty Promise against Adversarial Patch Attacks\n  on Optical Flow", "abstract": "  Adversarial patches undermine the reliability of optical flow predictions\nwhen placed in arbitrary scene locations. Therefore, they pose a realistic\nthreat to real-world motion detection and its downstream applications.\nPotential remedies are defense strategies that detect and remove adversarial\npatches, but their influence on the underlying motion prediction has not been\ninvestigated. In this paper, we thoroughly examine the currently available\ndetect-and-remove defenses ILP and LGS for a wide selection of state-of-the-art\noptical flow methods, and illuminate their side effects on the quality and\nrobustness of the final flow predictions. In particular, we implement\ndefense-aware attacks to investigate whether current defenses are able to\nwithstand attacks that take the defense mechanism into account. Our experiments\nyield two surprising results: Detect-and-remove defenses do not only lower the\noptical flow quality on benign scenes, in doing so, they also harm the\nrobustness under patch attacks for all tested optical flow methods except\nFlowNetC. As currently employed detect-and-remove defenses fail to deliver the\npromised adversarial robustness for optical flow, they evoke a false sense of\nsecurity. The code is available at\nhttps://github.com/cv-stuttgart/DetectionDefenses.\n", "date": "2023-10-26", "categories": []}
{"id": "2310.18477", "title": "Understanding and Improving Ensemble Adversarial Defense", "abstract": "  The strategy of ensemble has become popular in adversarial defense, which\ntrains multiple base classifiers to defend against adversarial attacks in a\ncooperative manner. Despite the empirical success, theoretical explanations on\nwhy an ensemble of adversarially trained classifiers is more robust than single\nones remain unclear. To fill in this gap, we develop a new error theory\ndedicated to understanding ensemble adversarial defense, demonstrating a\nprovable 0-1 loss reduction on challenging sample sets in an adversarial\ndefense scenario. Guided by this theory, we propose an effective approach to\nimprove ensemble adversarial defense, named interactive global adversarial\ntraining (iGAT). The proposal includes (1) a probabilistic distributing rule\nthat selectively allocates to different base classifiers adversarial examples\nthat are globally challenging to the ensemble, and (2) a regularization term to\nrescue the severest weaknesses of the base classifiers. Being tested over\nvarious existing ensemble adversarial defense techniques, iGAT is capable of\nboosting their performance by increases up to 17% evaluated using CIFAR10 and\nCIFAR100 datasets under both white-box and black-box attacks.\n", "date": "2023-10-27", "categories": []}
{"id": "2311.01323", "title": "Towards Evaluating Transfer-based Attacks Systematically, Practically,\n  and Fairly", "abstract": "  The adversarial vulnerability of deep neural networks (DNNs) has drawn great\nattention due to the security risk of applying these models in real-world\napplications. Based on transferability of adversarial examples, an increasing\nnumber of transfer-based methods have been developed to fool black-box DNN\nmodels whose architecture and parameters are inaccessible. Although tremendous\neffort has been exerted, there still lacks a standardized benchmark that could\nbe taken advantage of to compare these methods systematically, fairly, and\npractically. Our investigation shows that the evaluation of some methods needs\nto be more reasonable and more thorough to verify their effectiveness, to\navoid, for example, unfair comparison and insufficient consideration of\npossible substitute/victim models. Therefore, we establish a transfer-based\nattack benchmark (TA-Bench) which implements 30+ methods. In this paper, we\nevaluate and compare them comprehensively on 25 popular substitute/victim\nmodels on ImageNet. New insights about the effectiveness of these methods are\ngained and guidelines for future evaluations are provided. Code at:\nhttps://github.com/qizhangli/TA-Bench.\n", "date": "2023-11-02", "categories": []}
{"id": "2310.20175", "title": "LFAA: Crafting Transferable Targeted Adversarial Examples with\n  Low-Frequency Perturbations", "abstract": "  Deep neural networks are susceptible to adversarial attacks, which pose a\nsignificant threat to their security and reliability in real-world\napplications. The most notable adversarial attacks are transfer-based attacks,\nwhere an adversary crafts an adversarial example to fool one model, which can\nalso fool other models. While previous research has made progress in improving\nthe transferability of untargeted adversarial examples, the generation of\ntargeted adversarial examples that can transfer between models remains a\nchallenging task. In this work, we present a novel approach to generate\ntransferable targeted adversarial examples by exploiting the vulnerability of\ndeep neural networks to perturbations on high-frequency components of images.\nWe observe that replacing the high-frequency component of an image with that of\nanother image can mislead deep models, motivating us to craft perturbations\ncontaining high-frequency information to achieve targeted attacks. To this end,\nwe propose a method called Low-Frequency Adversarial Attack (\\name), which\ntrains a conditional generator to generate targeted adversarial perturbations\nthat are then added to the low-frequency component of the image. Extensive\nexperiments on ImageNet demonstrate that our proposed approach significantly\noutperforms state-of-the-art methods, improving targeted attack success rates\nby a margin from 3.2\\% to 15.5\\%.\n", "date": "2023-10-31", "categories": []}
{"id": "2311.00428", "title": "NEO-KD: Knowledge-Distillation-Based Adversarial Training for Robust\n  Multi-Exit Neural Networks", "abstract": "  While multi-exit neural networks are regarded as a promising solution for\nmaking efficient inference via early exits, combating adversarial attacks\nremains a challenging problem. In multi-exit networks, due to the high\ndependency among different submodels, an adversarial example targeting a\nspecific exit not only degrades the performance of the target exit but also\nreduces the performance of all other exits concurrently. This makes multi-exit\nnetworks highly vulnerable to simple adversarial attacks. In this paper, we\npropose NEO-KD, a knowledge-distillation-based adversarial training strategy\nthat tackles this fundamental challenge based on two key contributions. NEO-KD\nfirst resorts to neighbor knowledge distillation to guide the output of the\nadversarial examples to tend to the ensemble outputs of neighbor exits of clean\ndata. NEO-KD also employs exit-wise orthogonal knowledge distillation for\nreducing adversarial transferability across different submodels. The result is\na significantly improved robustness against adversarial attacks. Experimental\nresults on various datasets/models show that our method achieves the best\nadversarial accuracy with reduced computation budgets, compared to the\nbaselines relying on existing adversarial training or knowledge distillation\ntechniques for multi-exit networks.\n", "date": "2023-11-01", "categories": []}
{"id": "2311.00441", "title": "Improving Robustness for Vision Transformer with a Simple Dynamic\n  Scanning Augmentation", "abstract": "  Vision Transformer (ViT) has demonstrated promising performance in computer\nvision tasks, comparable to state-of-the-art neural networks. Yet, this new\ntype of deep neural network architecture is vulnerable to adversarial attacks\nlimiting its capabilities in terms of robustness. This article presents a novel\ncontribution aimed at further improving the accuracy and robustness of ViT,\nparticularly in the face of adversarial attacks. We propose an augmentation\ntechnique called `Dynamic Scanning Augmentation' that leverages dynamic input\nsequences to adaptively focus on different patches, thereby maintaining\nperformance and robustness. Our detailed investigations reveal that this\nadaptability to the input sequence induces significant changes in the attention\nmechanism of ViT, even for the same image. We introduce four variations of\nDynamic Scanning Augmentation, outperforming ViT in terms of both robustness to\nadversarial attacks and accuracy against natural images, with one variant\nshowing comparable results. By integrating our augmentation technique, we\nobserve a substantial increase in ViT's robustness, improving it from $17\\%$ to\n$92\\%$ measured across different types of adversarial attacks. These findings,\ntogether with other comprehensive tests, indicate that Dynamic Scanning\nAugmentation enhances accuracy and robustness by promoting a more adaptive type\nof attention. In conclusion, this work contributes to the ongoing research on\nVision Transformers by introducing Dynamic Scanning Augmentation as a technique\nfor improving the accuracy and robustness of ViT. The observed results\nhighlight the potential of this approach in advancing computer vision tasks and\nmerit further exploration in future studies.\n", "date": "2023-11-01", "categories": []}
{"id": "2305.11132", "title": "Attacks on Online Learners: a Teacher-Student Analysis", "abstract": "  Machine learning models are famously vulnerable to adversarial attacks: small\nad-hoc perturbations of the data that can catastrophically alter the model\npredictions. While a large literature has studied the case of test-time attacks\non pre-trained models, the important case of attacks in an online learning\nsetting has received little attention so far. In this work, we use a\ncontrol-theoretical perspective to study the scenario where an attacker may\nperturb data labels to manipulate the learning dynamics of an online learner.\nWe perform a theoretical analysis of the problem in a teacher-student setup,\nconsidering different attack strategies, and obtaining analytical results for\nthe steady state of simple linear learners. These results enable us to prove\nthat a discontinuous transition in the learner's accuracy occurs when the\nattack strength exceeds a critical threshold. We then study empirically attacks\non learners with complex architectures using real data, confirming the insights\nof our theoretical analysis. Our findings show that greedy attacks can be\nextremely efficient, especially when data stream in small batches.\n", "date": "2023-05-18", "categories": []}
{"id": "2306.06815", "title": "TrojLLM: A Black-box Trojan Prompt Attack on Large Language Models", "abstract": "  Large Language Models (LLMs) are progressively being utilized as machine\nlearning services and interface tools for various applications. However, the\nsecurity implications of LLMs, particularly in relation to adversarial and\nTrojan attacks, remain insufficiently examined. In this paper, we propose\nTrojLLM, an automatic and black-box framework to effectively generate universal\nand stealthy triggers. When these triggers are incorporated into the input\ndata, the LLMs' outputs can be maliciously manipulated. Moreover, the framework\nalso supports embedding Trojans within discrete prompts, enhancing the overall\neffectiveness and precision of the triggers' attacks. Specifically, we propose\na trigger discovery algorithm for generating universal triggers for various\ninputs by querying victim LLM-based APIs using few-shot data samples.\nFurthermore, we introduce a novel progressive Trojan poisoning algorithm\ndesigned to generate poisoned prompts that retain efficacy and transferability\nacross a diverse range of models. Our experiments and results demonstrate\nTrojLLM's capacity to effectively insert Trojans into text prompts in\nreal-world black-box LLM APIs including GPT-3.5 and GPT-4, while maintaining\nexceptional performance on clean test sets. Our work sheds light on the\npotential security risks in current models and offers a potential defensive\napproach. The source code of TrojLLM is available at\nhttps://github.com/UCF-ML-Research/TrojLLM.\n", "date": "2023-06-12", "categories": []}
{"id": "2310.19152", "title": "BERT Lost Patience Won't Be Robust to Adversarial Slowdown", "abstract": "  In this paper, we systematically evaluate the robustness of multi-exit\nlanguage models against adversarial slowdown. To audit their robustness, we\ndesign a slowdown attack that generates natural adversarial text bypassing\nearly-exit points. We use the resulting WAFFLE attack as a vehicle to conduct a\ncomprehensive evaluation of three multi-exit mechanisms with the GLUE benchmark\nagainst adversarial slowdown. We then show our attack significantly reduces the\ncomputational savings provided by the three methods in both white-box and\nblack-box settings. The more complex a mechanism is, the more vulnerable it is\nto adversarial slowdown. We also perform a linguistic analysis of the perturbed\ntext inputs, identifying common perturbation patterns that our attack\ngenerates, and comparing them with standard adversarial text attacks. Moreover,\nwe show that adversarial training is ineffective in defeating our slowdown\nattack, but input sanitization with a conversational model, e.g., ChatGPT, can\nremove perturbations effectively. This result suggests that future work is\nneeded for developing efficient yet robust multi-exit models. Our code is\navailable at: https://github.com/ztcoalson/WAFFLE\n", "date": "2023-10-29", "categories": []}
{"id": "2211.09565", "title": "Towards Good Practices in Evaluating Transfer Adversarial Attacks", "abstract": "  Transfer adversarial attacks raise critical security concerns in real-world,\nblack-box scenarios. However, the actual progress of this field is difficult to\nassess due to two common limitations in existing evaluations. First, different\nmethods are often not systematically and fairly evaluated in a one-to-one\ncomparison. Second, only transferability is evaluated but another key attack\nproperty, stealthiness, is largely overlooked. In this work, we design good\npractices to address these limitations, and we present the first comprehensive\nevaluation of transfer attacks, covering 23 representative attacks against 9\ndefenses on ImageNet. In particular, we propose to categorize existing attacks\ninto five categories, which enables our systematic category-wise analyses.\nThese analyses lead to new findings that even challenge existing knowledge and\nalso help determine the optimal attack hyperparameters for our attack-wise\ncomprehensive evaluation. We also pay particular attention to stealthiness, by\nadopting diverse imperceptibility metrics and looking into new, finer-grained\ncharacteristics. Overall, our new insights into transferability and\nstealthiness lead to actionable good practices for future evaluations.\n", "date": "2022-11-17", "categories": []}
{"id": "2301.12549", "title": "Unlocking Deterministic Robustness Certification on ImageNet", "abstract": "  Despite the promise of Lipschitz-based methods for provably-robust deep\nlearning with deterministic guarantees, current state-of-the-art results are\nlimited to feed-forward Convolutional Networks (ConvNets) on low-dimensional\ndata, such as CIFAR-10. This paper investigates strategies for expanding\ncertifiably robust training to larger, deeper models. A key challenge in\ncertifying deep networks is efficient calculation of the Lipschitz bound for\nresidual blocks found in ResNet and ViT architectures. We show that fast ways\nof bounding the Lipschitz constant for conventional ResNets are loose, and show\nhow to address this by designing a new residual block, leading to the\n\\emph{Linear ResNet} (LiResNet) architecture. We then introduce \\emph{Efficient\nMargin MAximization} (EMMA), a loss function that stabilizes robust training by\nsimultaneously penalizing worst-case adversarial examples from \\emph{all}\nclasses. Together, these contributions yield new \\emph{state-of-the-art} robust\naccuracy on CIFAR-10/100 and Tiny-ImageNet under $\\ell_2$ perturbations.\nMoreover, for the first time, we are able to scale up fast deterministic\nrobustness guarantees to ImageNet, demonstrating that this approach to robust\nlearning can be applied to real-world applications.\n  We release our code on Github: \\url{https://github.com/klasleino/gloro}.\n", "date": "2023-01-29", "categories": []}
{"id": "2303.01870", "title": "Revisiting Adversarial Training for ImageNet: Architectures, Training\n  and Generalization across Threat Models", "abstract": "  While adversarial training has been extensively studied for ResNet\narchitectures and low resolution datasets like CIFAR, much less is known for\nImageNet. Given the recent debate about whether transformers are more robust\nthan convnets, we revisit adversarial training on ImageNet comparing ViTs and\nConvNeXts. Extensive experiments show that minor changes in architecture, most\nnotably replacing PatchStem with ConvStem, and training scheme have a\nsignificant impact on the achieved robustness. These changes not only increase\nrobustness in the seen $\\ell_\\infty$-threat model, but even more so improve\ngeneralization to unseen $\\ell_1/\\ell_2$-attacks. Our modified ConvNeXt,\nConvNeXt + ConvStem, yields the most robust $\\ell_\\infty$-models across\ndifferent ranges of model parameters and FLOPs, while our ViT + ConvStem yields\nthe best generalization to unseen threat models.\n", "date": "2023-03-03", "categories": []}
{"id": "2307.04333", "title": "Enhancing Adversarial Robustness via Score-Based Optimization", "abstract": "  Adversarial attacks have the potential to mislead deep neural network\nclassifiers by introducing slight perturbations. Developing algorithms that can\nmitigate the effects of these attacks is crucial for ensuring the safe use of\nartificial intelligence. Recent studies have suggested that score-based\ndiffusion models are effective in adversarial defenses. However, existing\ndiffusion-based defenses rely on the sequential simulation of the reversed\nstochastic differential equations of diffusion models, which are\ncomputationally inefficient and yield suboptimal results. In this paper, we\nintroduce a novel adversarial defense scheme named ScoreOpt, which optimizes\nadversarial samples at test-time, towards original clean data in the direction\nguided by score-based priors. We conduct comprehensive experiments on multiple\ndatasets, including CIFAR10, CIFAR100 and ImageNet. Our experimental results\ndemonstrate that our approach outperforms existing adversarial defenses in\nterms of both robustness performance and inference speed.\n", "date": "2023-07-10", "categories": []}
{"id": "2310.06182", "title": "PAC-Bayesian Spectrally-Normalized Bounds for Adversarially Robust\n  Generalization", "abstract": "  Deep neural networks (DNNs) are vulnerable to adversarial attacks. It is\nfound empirically that adversarially robust generalization is crucial in\nestablishing defense algorithms against adversarial attacks. Therefore, it is\ninteresting to study the theoretical guarantee of robust generalization. This\npaper focuses on norm-based complexity, based on a PAC-Bayes approach\n(Neyshabur et al., 2017). The main challenge lies in extending the key\ningredient, which is a weight perturbation bound in standard settings, to the\nrobust settings. Existing attempts heavily rely on additional strong\nassumptions, leading to loose bounds. In this paper, we address this issue and\nprovide a spectrally-normalized robust generalization bound for DNNs. Compared\nto existing bounds, our bound offers two significant advantages: Firstly, it\ndoes not depend on additional assumptions. Secondly, it is considerably\ntighter, aligning with the bounds of standard generalization. Therefore, our\nresult provides a different perspective on understanding robust generalization:\nThe mismatch terms between standard and robust generalization bounds shown in\nprevious studies do not contribute to the poor robust generalization. Instead,\nthese disparities solely due to mathematical issues. Finally, we extend the\nmain result to adversarial robustness against general non-$\\ell_p$ attacks and\nother neural network architectures.\n", "date": "2023-10-09", "categories": []}
{"id": "2310.18603", "title": "Large Language Models Are Better Adversaries: Exploring Generative\n  Clean-Label Backdoor Attacks Against Text Classifiers", "abstract": "  Backdoor attacks manipulate model predictions by inserting innocuous triggers\ninto training and test data. We focus on more realistic and more challenging\nclean-label attacks where the adversarial training examples are correctly\nlabeled. Our attack, LLMBkd, leverages language models to automatically insert\ndiverse style-based triggers into texts. We also propose a poison selection\ntechnique to improve the effectiveness of both LLMBkd as well as existing\ntextual backdoor attacks. Lastly, we describe REACT, a baseline defense to\nmitigate backdoor attacks via antidote training examples. Our evaluations\ndemonstrate LLMBkd's effectiveness and efficiency, where we consistently\nachieve high attack success rates across a wide range of styles with little\neffort and no model training.\n", "date": "2023-10-28", "categories": []}
{"id": "2310.18762", "title": "Purify++: Improving Diffusion-Purification with Advanced Diffusion\n  Models and Control of Randomness", "abstract": "  Adversarial attacks can mislead neural network classifiers. The defense\nagainst adversarial attacks is important for AI safety. Adversarial\npurification is a family of approaches that defend adversarial attacks with\nsuitable pre-processing. Diffusion models have been shown to be effective for\nadversarial purification. Despite their success, many aspects of diffusion\npurification still remain unexplored. In this paper, we investigate and improve\nupon three limiting designs of diffusion purification: the use of an improved\ndiffusion model, advanced numerical simulation techniques, and optimal control\nof randomness. Based on our findings, we propose Purify++, a new diffusion\npurification algorithm that is now the state-of-the-art purification method\nagainst several adversarial attacks. Our work presents a systematic exploration\nof the limits of diffusion purification methods.\n", "date": "2023-10-28", "categories": []}
{"id": "2310.18933", "title": "Label Poisoning is All You Need", "abstract": "  In a backdoor attack, an adversary injects corrupted data into a model's\ntraining dataset in order to gain control over its predictions on images with a\nspecific attacker-defined trigger. A typical corrupted training example\nrequires altering both the image, by applying the trigger, and the label.\nModels trained on clean images, therefore, were considered safe from backdoor\nattacks. However, in some common machine learning scenarios, the training\nlabels are provided by potentially malicious third-parties. This includes\ncrowd-sourced annotation and knowledge distillation. We, hence, investigate a\nfundamental question: can we launch a successful backdoor attack by only\ncorrupting labels? We introduce a novel approach to design label-only backdoor\nattacks, which we call FLIP, and demonstrate its strengths on three datasets\n(CIFAR-10, CIFAR-100, and Tiny-ImageNet) and four architectures (ResNet-32,\nResNet-18, VGG-19, and Vision Transformer). With only 2% of CIFAR-10 labels\ncorrupted, FLIP achieves a near-perfect attack success rate of 99.4% while\nsuffering only a 1.8% drop in the clean test accuracy. Our approach builds upon\nthe recent advances in trajectory matching, originally introduced for dataset\ndistillation.\n", "date": "2023-10-29", "categories": []}
{"id": "2310.18975", "title": "Blacksmith: Fast Adversarial Training of Vision Transformers via a\n  Mixture of Single-step and Multi-step Methods", "abstract": "  Despite the remarkable success achieved by deep learning algorithms in\nvarious domains, such as computer vision, they remain vulnerable to adversarial\nperturbations. Adversarial Training (AT) stands out as one of the most\neffective solutions to address this issue; however, single-step AT can lead to\nCatastrophic Overfitting (CO). This scenario occurs when the adversarially\ntrained network suddenly loses robustness against multi-step attacks like\nProjected Gradient Descent (PGD). Although several approaches have been\nproposed to address this problem in Convolutional Neural Networks (CNNs), we\nfound out that they do not perform well when applied to Vision Transformers\n(ViTs). In this paper, we propose Blacksmith, a novel training strategy to\novercome the CO problem, specifically in ViTs. Our approach utilizes either of\nPGD-2 or Fast Gradient Sign Method (FGSM) randomly in a mini-batch during the\nadversarial training of the neural network. This will increase the diversity of\nour training attacks, which could potentially mitigate the CO issue. To manage\nthe increased training time resulting from this combination, we craft the PGD-2\nattack based on only the first half of the layers, while FGSM is applied\nend-to-end. Through our experiments, we demonstrate that our novel method\neffectively prevents CO, achieves PGD-2 level performance, and outperforms\nother existing techniques including N-FGSM, which is the state-of-the-art\nmethod in fast training for CNNs.\n", "date": "2023-10-29", "categories": []}
{"id": "2310.19038", "title": "Boosting Decision-Based Black-Box Adversarial Attack with Gradient\n  Priors", "abstract": "  Decision-based methods have shown to be effective in black-box adversarial\nattacks, as they can obtain satisfactory performance and only require to access\nthe final model prediction. Gradient estimation is a critical step in black-box\nadversarial attacks, as it will directly affect the query efficiency. Recent\nworks have attempted to utilize gradient priors to facilitate score-based\nmethods to obtain better results. However, these gradient priors still suffer\nfrom the edge gradient discrepancy issue and the successive iteration gradient\ndirection issue, thus are difficult to simply extend to decision-based methods.\nIn this paper, we propose a novel Decision-based Black-box Attack framework\nwith Gradient Priors (DBA-GP), which seamlessly integrates the data-dependent\ngradient prior and time-dependent prior into the gradient estimation procedure.\nFirst, by leveraging the joint bilateral filter to deal with each random\nperturbation, DBA-GP can guarantee that the generated perturbations in edge\nlocations are hardly smoothed, i.e., alleviating the edge gradient discrepancy,\nthus remaining the characteristics of the original image as much as possible.\nSecond, by utilizing a new gradient updating strategy to automatically adjust\nthe successive iteration gradient direction, DBA-GP can accelerate the\nconvergence speed, thus improving the query efficiency. Extensive experiments\nhave demonstrated that the proposed method outperforms other strong baselines\nsignificantly.\n", "date": "2023-10-29", "categories": []}
{"id": "2310.19156", "title": "Poisoning Retrieval Corpora by Injecting Adversarial Passages", "abstract": "  Dense retrievers have achieved state-of-the-art performance in various\ninformation retrieval tasks, but to what extent can they be safely deployed in\nreal-world applications? In this work, we propose a novel attack for dense\nretrieval systems in which a malicious user generates a small number of\nadversarial passages by perturbing discrete tokens to maximize similarity with\na provided set of training queries. When these adversarial passages are\ninserted into a large retrieval corpus, we show that this attack is highly\neffective in fooling these systems to retrieve them for queries that were not\nseen by the attacker. More surprisingly, these adversarial passages can\ndirectly generalize to out-of-domain queries and corpora with a high success\nattack rate -- for instance, we find that 50 generated passages optimized on\nNatural Questions can mislead >94% of questions posed in financial documents or\nonline forums. We also benchmark and compare a range of state-of-the-art dense\nretrievers, both unsupervised and supervised. Although different systems\nexhibit varying levels of vulnerability, we show they can all be successfully\nattacked by injecting up to 500 passages, a small fraction compared to a\nretrieval corpus of millions of passages.\n", "date": "2023-10-29", "categories": []}
{"id": "2310.19177", "title": "Robustifying Language Models with Test-Time Adaptation", "abstract": "  Large-scale language models achieved state-of-the-art performance over a\nnumber of language tasks. However, they fail on adversarial language examples,\nwhich are sentences optimized to fool the language models but with similar\nsemantic meanings for humans. While prior work focuses on making the language\nmodel robust at training time, retraining for robustness is often unrealistic\nfor large-scale foundation models. Instead, we propose to make the language\nmodels robust at test time. By dynamically adapting the input sentence with\npredictions from masked words, we show that we can reverse many language\nadversarial attacks. Since our approach does not require any training, it works\nfor novel tasks at test time and can adapt to novel adversarial corruptions.\nVisualizations and empirical results on two popular sentence classification\ndatasets demonstrate that our method can repair adversarial language attacks\nover 65% o\n", "date": "2023-10-29", "categories": []}
{"id": "2310.19360", "title": "Balance, Imbalance, and Rebalance: Understanding Robust Overfitting from\n  a Minimax Game Perspective", "abstract": "  Adversarial Training (AT) has become arguably the state-of-the-art algorithm\nfor extracting robust features. However, researchers recently notice that AT\nsuffers from severe robust overfitting problems, particularly after learning\nrate (LR) decay. In this paper, we explain this phenomenon by viewing\nadversarial training as a dynamic minimax game between the model trainer and\nthe attacker. Specifically, we analyze how LR decay breaks the balance between\nthe minimax game by empowering the trainer with a stronger memorization\nability, and show such imbalance induces robust overfitting as a result of\nmemorizing non-robust features. We validate this understanding with extensive\nexperiments, and provide a holistic view of robust overfitting from the\ndynamics of both the two game players. This understanding further inspires us\nto alleviate robust overfitting by rebalancing the two players by either\nregularizing the trainer's capacity or improving the attack strength.\nExperiments show that the proposed ReBalanced Adversarial Training (ReBAT) can\nattain good robustness and does not suffer from robust overfitting even after\nvery long training. Code is available at https://github.com/PKU-ML/ReBAT.\n", "date": "2023-10-30", "categories": []}
{"id": "2310.19737", "title": "Adversarial Attacks and Defenses in Large Language Models: Old and New\n  Threats", "abstract": "  Over the past decade, there has been extensive research aimed at enhancing\nthe robustness of neural networks, yet this problem remains vastly unsolved.\nHere, one major impediment has been the overestimation of the robustness of new\ndefense approaches due to faulty defense evaluations. Flawed robustness\nevaluations necessitate rectifications in subsequent works, dangerously slowing\ndown the research and providing a false sense of security. In this context, we\nwill face substantial challenges associated with an impending adversarial arms\nrace in natural language processing, specifically with closed-source Large\nLanguage Models (LLMs), such as ChatGPT, Google Bard, or Anthropic's Claude. We\nprovide a first set of prerequisites to improve the robustness assessment of\nnew approaches and reduce the amount of faulty evaluations. Additionally, we\nidentify embedding space attacks on LLMs as another viable threat model for the\npurposes of generating malicious content in open-sourced models. Finally, we\ndemonstrate on a recently proposed defense that, without LLM-specific best\npractices in place, it is easy to overestimate the robustness of a new\napproach.\n", "date": "2023-10-30", "categories": []}
{"id": "2303.03169", "title": "A Unified Algebraic Perspective on Lipschitz Neural Networks", "abstract": "  Important research efforts have focused on the design and training of neural\nnetworks with a controlled Lipschitz constant. The goal is to increase and\nsometimes guarantee the robustness against adversarial attacks. Recent\npromising techniques draw inspirations from different backgrounds to design\n1-Lipschitz neural networks, just to name a few: convex potential layers derive\nfrom the discretization of continuous dynamical systems,\nAlmost-Orthogonal-Layer proposes a tailored method for matrix rescaling.\nHowever, it is today important to consider the recent and promising\ncontributions in the field under a common theoretical lens to better design new\nand improved layers. This paper introduces a novel algebraic perspective\nunifying various types of 1-Lipschitz neural networks, including the ones\npreviously mentioned, along with methods based on orthogonality and spectral\nmethods. Interestingly, we show that many existing techniques can be derived\nand generalized via finding analytical solutions of a common semidefinite\nprogramming (SDP) condition. We also prove that AOL biases the scaled weight to\nthe ones which are close to the set of orthogonal matrices in a certain\nmathematical manner. Moreover, our algebraic condition, combined with the\nGershgorin circle theorem, readily leads to new and diverse parameterizations\nfor 1-Lipschitz network layers. Our approach, called SDP-based Lipschitz Layers\n(SLL), allows us to design non-trivial yet efficient generalization of convex\npotential layers. Finally, the comprehensive set of experiments on image\nclassification shows that SLLs outperform previous approaches on certified\nrobust accuracy. Code is available at\nhttps://github.com/araujoalexandre/Lipschitz-SLL-Networks.\n", "date": "2023-03-06", "categories": []}
{"id": "2310.16779", "title": "Multi-scale Diffusion Denoised Smoothing", "abstract": "  Along with recent diffusion models, randomized smoothing has become one of a\nfew tangible approaches that offers adversarial robustness to models at scale,\ne.g., those of large pre-trained models. Specifically, one can perform\nrandomized smoothing on any classifier via a simple \"denoise-and-classify\"\npipeline, so-called denoised smoothing, given that an accurate denoiser is\navailable - such as diffusion model. In this paper, we present scalable methods\nto address the current trade-off between certified robustness and accuracy in\ndenoised smoothing. Our key idea is to \"selectively\" apply smoothing among\nmultiple noise scales, coined multi-scale smoothing, which can be efficiently\nimplemented with a single diffusion model. This approach also suggests a new\nobjective to compare the collective robustness of multi-scale smoothed\nclassifiers, and questions which representation of diffusion model would\nmaximize the objective. To address this, we propose to further fine-tune\ndiffusion model (a) to perform consistent denoising whenever the original image\nis recoverable, but (b) to generate rather diverse outputs otherwise. Our\nexperiments show that the proposed multi-scale smoothing scheme combined with\ndiffusion fine-tuning enables strong certified robustness available with high\nnoise level while maintaining its accuracy close to non-smoothed classifiers.\n", "date": "2023-10-25", "categories": []}
{"id": "2210.10482", "title": "Effective Targeted Attacks for Adversarial Self-Supervised Learning", "abstract": "  Recently, unsupervised adversarial training (AT) has been highlighted as a\nmeans of achieving robustness in models without any label information. Previous\nstudies in unsupervised AT have mostly focused on implementing self-supervised\nlearning (SSL) frameworks, which maximize the instance-wise classification loss\nto generate adversarial examples. However, we observe that simply maximizing\nthe self-supervised training loss with an untargeted adversarial attack often\nresults in generating ineffective adversaries that may not help improve the\nrobustness of the trained model, especially for non-contrastive SSL frameworks\nwithout negative examples. To tackle this problem, we propose a novel positive\nmining for targeted adversarial attack to generate effective adversaries for\nadversarial SSL frameworks. Specifically, we introduce an algorithm that\nselects the most confusing yet similar target example for a given instance\nbased on entropy and similarity, and subsequently perturbs the given instance\ntowards the selected target. Our method demonstrates significant enhancements\nin robustness when applied to non-contrastive SSL frameworks, and less but\nconsistent robustness improvements with contrastive SSL frameworks, on the\nbenchmark datasets.\n", "date": "2022-10-19", "categories": []}
{"id": "2210.10485", "title": "Learning Transferable Adversarial Robust Representations via Multi-view\n  Consistency", "abstract": "  Despite the success on few-shot learning problems, most meta-learned models\nonly focus on achieving good performance on clean examples and thus easily\nbreak down when given adversarially perturbed samples. While some recent works\nhave shown that a combination of adversarial learning and meta-learning could\nenhance the robustness of a meta-learner against adversarial attacks, they fail\nto achieve generalizable adversarial robustness to unseen domains and tasks,\nwhich is the ultimate goal of meta-learning. To address this challenge, we\npropose a novel meta-adversarial multi-view representation learning framework\nwith dual encoders. Specifically, we introduce the discrepancy across the two\ndifferently augmented samples of the same data instance by first updating the\nencoder parameters with them and further imposing a novel label-free\nadversarial attack to maximize their discrepancy. Then, we maximize the\nconsistency across the views to learn transferable robust representations\nacross domains and tasks. Through experimental validation on multiple\nbenchmarks, we demonstrate the effectiveness of our framework on few-shot\nlearning tasks from unseen domains, achieving over 10\\% robust accuracy\nimprovements against previous adversarial meta-learning baselines.\n", "date": "2022-10-19", "categories": []}
{"id": "2302.03857", "title": "Efficient Adversarial Contrastive Learning via Robustness-Aware Coreset\n  Selection", "abstract": "  Adversarial contrastive learning (ACL) does not require expensive data\nannotations but outputs a robust representation that withstands adversarial\nattacks and also generalizes to a wide range of downstream tasks. However, ACL\nneeds tremendous running time to generate the adversarial variants of all\ntraining data, which limits its scalability to large datasets. To speed up ACL,\nthis paper proposes a robustness-aware coreset selection (RCS) method. RCS does\nnot require label information and searches for an informative subset that\nminimizes a representational divergence, which is the distance of the\nrepresentation between natural data and their virtual adversarial variants. The\nvanilla solution of RCS via traversing all possible subsets is computationally\nprohibitive. Therefore, we theoretically transform RCS into a surrogate problem\nof submodular maximization, of which the greedy search is an efficient solution\nwith an optimality guarantee for the original problem. Empirically, our\ncomprehensive results corroborate that RCS can speed up ACL by a large margin\nwithout significantly hurting the robustness transferability. Notably, to the\nbest of our knowledge, we are the first to conduct ACL efficiently on the\nlarge-scale ImageNet-1K dataset to obtain an effective robust representation\nvia RCS. Our source code is at\nhttps://github.com/GodXuxilie/Efficient_ACL_via_RCS.\n", "date": "2023-02-08", "categories": []}
{"id": "2310.17559", "title": "Instability of computer vision models is a necessary result of the task\n  itself", "abstract": "  Adversarial examples resulting from instability of current computer vision\nmodels are an extremely important topic due to their potential to compromise\nany application. In this paper we demonstrate that instability is inevitable\ndue to a) symmetries (translational invariance) of the data, b) the categorical\nnature of the classification task, and c) the fundamental discrepancy of\nclassifying images as objects themselves. The issue is further exacerbated by\nnon-exhaustive labelling of the training data. Therefore we conclude that\ninstability is a necessary result of how the problem of computer vision is\ncurrently formulated. While the problem cannot be eliminated, through the\nanalysis of the causes, we have arrived at ways how it can be partially\nalleviated. These include i) increasing the resolution of images, ii) providing\ncontextual information for the image, iii) exhaustive labelling of training\ndata, and iv) preventing attackers from frequent access to the computer vision\nsystem.\n", "date": "2023-10-26", "categories": []}
{"id": "2305.04574", "title": "TAPS: Connecting Certified and Adversarial Training", "abstract": "  Training certifiably robust neural networks remains a notoriously hard\nproblem. On one side, adversarial training optimizes under-approximations of\nthe worst-case loss, which leads to insufficient regularization for\ncertification, while on the other, sound certified training methods optimize\nloose over-approximations, leading to over-regularization and poor (standard)\naccuracy. In this work we propose TAPS, an (unsound) certified training method\nthat combines IBP and PGD training to yield precise, although not necessarily\nsound, worst-case loss approximations, reducing over-regularization and\nincreasing certified and standard accuracies. Empirically, TAPS achieves a new\nstate-of-the-art in many settings, e.g., reaching a certified accuracy of\n$22\\%$ on TinyImageNet for $\\ell_\\infty$-perturbations with radius\n$\\epsilon=1/255$. We make our implementation and networks public at\nhttps://github.com/eth-sri/taps.\n", "date": "2023-05-08", "categories": []}
{"id": "2306.11974", "title": "Universal adversarial perturbations for multiple classification tasks\n  with quantum classifiers", "abstract": "  Quantum adversarial machine learning is an emerging field that studies the\nvulnerability of quantum learning systems against adversarial perturbations and\ndevelops possible defense strategies. Quantum universal adversarial\nperturbations are small perturbations, which can make different input samples\ninto adversarial examples that may deceive a given quantum classifier. This is\na field that was rarely looked into but worthwhile investigating because\nuniversal perturbations might simplify malicious attacks to a large extent,\ncausing unexpected devastation to quantum machine learning models. In this\npaper, we take a step forward and explore the quantum universal perturbations\nin the context of heterogeneous classification tasks. In particular, we find\nthat quantum classifiers that achieve almost state-of-the-art accuracy on two\ndifferent classification tasks can be both conclusively deceived by one\ncarefully-crafted universal perturbation. This result is explicitly\ndemonstrated with well-designed quantum continual learning models with elastic\nweight consolidation method to avoid catastrophic forgetting, as well as\nreal-life heterogeneous datasets from hand-written digits and medical MRI\nimages. Our results provide a simple and efficient way to generate universal\nperturbations on heterogeneous classification tasks and thus would provide\nvaluable guidance for future quantum learning technologies.\n", "date": "2023-06-21", "categories": []}
{"id": "2310.14480", "title": "Attention-Enhancing Backdoor Attacks Against BERT-based Models", "abstract": "  Recent studies have revealed that \\textit{Backdoor Attacks} can threaten the\nsafety of natural language processing (NLP) models. Investigating the\nstrategies of backdoor attacks will help to understand the model's\nvulnerability. Most existing textual backdoor attacks focus on generating\nstealthy triggers or modifying model weights. In this paper, we directly target\nthe interior structure of neural networks and the backdoor mechanism. We\npropose a novel Trojan Attention Loss (TAL), which enhances the Trojan behavior\nby directly manipulating the attention patterns. Our loss can be applied to\ndifferent attacking methods to boost their attack efficacy in terms of attack\nsuccessful rates and poisoning rates. It applies to not only traditional\ndirty-label attacks, but also the more challenging clean-label attacks. We\nvalidate our method on different backbone models (BERT, RoBERTa, and\nDistilBERT) and various tasks (Sentiment Analysis, Toxic Detection, and Topic\nClassification).\n", "date": "2023-10-23", "categories": []}
{"id": "2310.16335", "title": "Defense Against Model Extraction Attacks on Recommender Systems", "abstract": "  The robustness of recommender systems has become a prominent topic within the\nresearch community. Numerous adversarial attacks have been proposed, but most\nof them rely on extensive prior knowledge, such as all the white-box attacks or\nmost of the black-box attacks which assume that certain external knowledge is\navailable. Among these attacks, the model extraction attack stands out as a\npromising and practical method, involving training a surrogate model by\nrepeatedly querying the target model. However, there is a significant gap in\nthe existing literature when it comes to defending against model extraction\nattacks on recommender systems. In this paper, we introduce Gradient-based\nRanking Optimization (GRO), which is the first defense strategy designed to\ncounter such attacks. We formalize the defense as an optimization problem,\naiming to minimize the loss of the protected target model while maximizing the\nloss of the attacker's surrogate model. Since top-k ranking lists are\nnon-differentiable, we transform them into swap matrices which are instead\ndifferentiable. These swap matrices serve as input to a student model that\nemulates the surrogate model's behavior. By back-propagating the loss of the\nstudent model, we obtain gradients for the swap matrices. These gradients are\nused to compute a swap loss, which maximizes the loss of the student model. We\nconducted experiments on three benchmark datasets to evaluate the performance\nof GRO, and the results demonstrate its superior effectiveness in defending\nagainst model extraction attacks.\n", "date": "2023-10-25", "categories": []}
{"id": "2310.16540", "title": "Dual Defense: Adversarial, Traceable, and Invisible Robust Watermarking\n  against Face Swapping", "abstract": "  The malicious applications of deep forgery, represented by face swapping,\nhave introduced security threats such as misinformation dissemination and\nidentity fraud. While some research has proposed the use of robust watermarking\nmethods to trace the copyright of facial images for post-event traceability,\nthese methods cannot effectively prevent the generation of forgeries at the\nsource and curb their dissemination. To address this problem, we propose a\nnovel comprehensive active defense mechanism that combines traceability and\nadversariality, called Dual Defense. Dual Defense invisibly embeds a single\nrobust watermark within the target face to actively respond to sudden cases of\nmalicious face swapping. It disrupts the output of the face swapping model\nwhile maintaining the integrity of watermark information throughout the entire\ndissemination process. This allows for watermark extraction at any stage of\nimage tracking for traceability. Specifically, we introduce a watermark\nembedding network based on original-domain feature impersonation attack. This\nnetwork learns robust adversarial features of target facial images and embeds\nwatermarks, seeking a well-balanced trade-off between watermark invisibility,\nadversariality, and traceability through perceptual adversarial encoding\nstrategies. Extensive experiments demonstrate that Dual Defense achieves\noptimal overall defense success rates and exhibits promising universality in\nanti-face swapping tasks and dataset generalization ability. It maintains\nimpressive adversariality and traceability in both original and robust\nsettings, surpassing current forgery defense methods that possess only one of\nthese capabilities, including CMUA-Watermark, Anti-Forgery, FakeTagger, or PGD\nmethods.\n", "date": "2023-10-25", "categories": []}
{"id": "2307.14387", "title": "Coupled-Space Attacks against Random-Walk-based Anomaly Detection", "abstract": "  Random Walks-based Anomaly Detection (RWAD) is commonly used to identify\nanomalous patterns in various applications. An intriguing characteristic of\nRWAD is that the input graph can either be pre-existing or constructed from raw\nfeatures. Consequently, there are two potential attack surfaces against RWAD:\ngraph-space attacks and feature-space attacks. In this paper, we explore this\nvulnerability by designing practical coupled-space attacks, investigating the\ninterplay between graph-space and feature-space attacks. To this end, we\nconduct a thorough complexity analysis, proving that attacking RWAD is NP-hard.\nThen, we proceed to formulate the graph-space attack as a bi-level optimization\nproblem and propose two strategies to solve it: alternative iteration\n(alterI-attack) or utilizing the closed-form solution of the random walk model\n(cf-attack). Finally, we utilize the results from the graph-space attacks as\nguidance to design more powerful feature-space attacks (i.e., graph-guided\nattacks). Comprehensive experiments demonstrate that our proposed attacks are\neffective in enabling the target nodes from RWAD with a limited attack budget.\nIn addition, we conduct transfer attack experiments in a black-box setting,\nwhich show that our feature attack significantly decreases the anomaly scores\nof target nodes. Our study opens the door to studying the coupled-space attack\nagainst graph anomaly detection in which the graph space relies on the feature\nspace.\n", "date": "2023-07-26", "categories": []}
{"id": "2310.15444", "title": "Fast Propagation is Better: Accelerating Single-Step Adversarial\n  Training via Sampling Subnetworks", "abstract": "  Adversarial training has shown promise in building robust models against\nadversarial examples. A major drawback of adversarial training is the\ncomputational overhead introduced by the generation of adversarial examples. To\novercome this limitation, adversarial training based on single-step attacks has\nbeen explored. Previous work improves the single-step adversarial training from\ndifferent perspectives, e.g., sample initialization, loss regularization, and\ntraining strategy. Almost all of them treat the underlying model as a black\nbox. In this work, we propose to exploit the interior building blocks of the\nmodel to improve efficiency. Specifically, we propose to dynamically sample\nlightweight subnetworks as a surrogate model during training. By doing this,\nboth the forward and backward passes can be accelerated for efficient\nadversarial training. Besides, we provide theoretical analysis to show the\nmodel robustness can be improved by the single-step adversarial training with\nsampled subnetworks. Furthermore, we propose a novel sampling strategy where\nthe sampling varies from layer to layer and from iteration to iteration.\nCompared with previous methods, our method not only reduces the training cost\nbut also achieves better model robustness. Evaluations on a series of popular\ndatasets demonstrate the effectiveness of the proposed FB-Better. Our code has\nbeen released at https://github.com/jiaxiaojunQAQ/FP-Better.\n", "date": "2023-10-24", "categories": []}
{"id": "2310.15653", "title": "Deceptive Fairness Attacks on Graphs via Meta Learning", "abstract": "  We study deceptive fairness attacks on graphs to answer the following\nquestion: How can we achieve poisoning attacks on a graph learning model to\nexacerbate the bias deceptively? We answer this question via a bi-level\noptimization problem and propose a meta learning-based framework named FATE.\nFATE is broadly applicable with respect to various fairness definitions and\ngraph learning models, as well as arbitrary choices of manipulation operations.\nWe further instantiate FATE to attack statistical parity and individual\nfairness on graph neural networks. We conduct extensive experimental\nevaluations on real-world datasets in the task of semi-supervised node\nclassification. The experimental results demonstrate that FATE could amplify\nthe bias of graph neural networks with or without fairness consideration while\nmaintaining the utility on the downstream task. We hope this paper provides\ninsights into the adversarial robustness of fair graph learning and can shed\nlight on designing robust and fair graph learning in future studies.\n", "date": "2023-10-24", "categories": []}
{"id": "2310.15656", "title": "Momentum Gradient-based Untargeted Attack on Hypergraph Neural Networks", "abstract": "  Hypergraph Neural Networks (HGNNs) have been successfully applied in various\nhypergraph-related tasks due to their excellent higher-order representation\ncapabilities. Recent works have shown that deep learning models are vulnerable\nto adversarial attacks. Most studies on graph adversarial attacks have focused\non Graph Neural Networks (GNNs), and the study of adversarial attacks on HGNNs\nremains largely unexplored. In this paper, we try to reduce this gap. We design\na new HGNNs attack model for the untargeted attack, namely MGHGA, which focuses\non modifying node features. We consider the process of HGNNs training and use a\nsurrogate model to implement the attack before hypergraph modeling.\nSpecifically, MGHGA consists of two parts: feature selection and feature\nmodification. We use a momentum gradient mechanism to choose the attack node\nfeatures in the feature selection module. In the feature modification module,\nwe use two feature generation approaches (direct modification and sign\ngradient) to enable MGHGA to be employed on discrete and continuous datasets.\nWe conduct extensive experiments on five benchmark datasets to validate the\nattack performance of MGHGA in the node and the visual object classification\ntasks. The results show that MGHGA improves performance by an average of 2%\ncompared to the than the baselines.\n", "date": "2023-10-24", "categories": []}
{"id": "2209.08412", "title": "Characterizing Internal Evasion Attacks in Federated Learning", "abstract": "  Federated learning allows for clients in a distributed system to jointly\ntrain a machine learning model. However, clients' models are vulnerable to\nattacks during the training and testing phases. In this paper, we address the\nissue of adversarial clients performing \"internal evasion attacks\": crafting\nevasion attacks at test time to deceive other clients. For example, adversaries\nmay aim to deceive spam filters and recommendation systems trained with\nfederated learning for monetary gain. The adversarial clients have extensive\ninformation about the victim model in a federated learning setting, as weight\ninformation is shared amongst clients. We are the first to characterize the\ntransferability of such internal evasion attacks for different learning methods\nand analyze the trade-off between model accuracy and robustness depending on\nthe degree of similarities in client data. We show that adversarial training\ndefenses in the federated learning setting only display limited improvements\nagainst internal attacks. However, combining adversarial training with\npersonalized federated learning frameworks increases relative internal attack\nrobustness by 60% compared to federated adversarial training and performs well\nunder limited system resources.\n", "date": "2022-09-17", "categories": []}
{"id": "2305.00374", "title": "Enhancing Adversarial Contrastive Learning via Adversarial Invariant\n  Regularization", "abstract": "  Adversarial contrastive learning (ACL) is a technique that enhances standard\ncontrastive learning (SCL) by incorporating adversarial data to learn a robust\nrepresentation that can withstand adversarial attacks and common corruptions\nwithout requiring costly annotations. To improve transferability, the existing\nwork introduced the standard invariant regularization (SIR) to impose\nstyle-independence property to SCL, which can exempt the impact of nuisance\nstyle factors in the standard representation. However, it is unclear how the\nstyle-independence property benefits ACL-learned robust representations. In\nthis paper, we leverage the technique of causal reasoning to interpret the ACL\nand propose adversarial invariant regularization (AIR) to enforce independence\nfrom style factors. We regulate the ACL using both SIR and AIR to output the\nrobust representation. Theoretically, we show that AIR implicitly encourages\nthe representational distance between different views of natural data and their\nadversarial variants to be independent of style factors. Empirically, our\nexperimental results show that invariant regularization significantly improves\nthe performance of state-of-the-art ACL methods in terms of both standard\ngeneralization and robustness on downstream tasks. To the best of our\nknowledge, we are the first to apply causal reasoning to interpret ACL and\ndevelop AIR for enhancing ACL-learned robust representations. Our source code\nis at https://github.com/GodXuxilie/Enhancing_ACL_via_AIR.\n", "date": "2023-04-30", "categories": []}
{"id": "2305.04175", "title": "Text-to-Image Diffusion Models can be Easily Backdoored through\n  Multimodal Data Poisoning", "abstract": "  With the help of conditioning mechanisms, the state-of-the-art diffusion\nmodels have achieved tremendous success in guided image generation,\nparticularly in text-to-image synthesis. To gain a better understanding of the\ntraining process and potential risks of text-to-image synthesis, we perform a\nsystematic investigation of backdoor attack on text-to-image diffusion models\nand propose BadT2I, a general multimodal backdoor attack framework that tampers\nwith image synthesis in diverse semantic levels. Specifically, we perform\nbackdoor attacks on three levels of the vision semantics: Pixel-Backdoor,\nObject-Backdoor and Style-Backdoor. By utilizing a regularization loss, our\nmethods efficiently inject backdoors into a large-scale text-to-image diffusion\nmodel while preserving its utility with benign inputs. We conduct empirical\nexperiments on Stable Diffusion, the widely-used text-to-image diffusion model,\ndemonstrating that the large-scale diffusion model can be easily backdoored\nwithin a few fine-tuning steps. We conduct additional experiments to explore\nthe impact of different types of textual triggers, as well as the backdoor\npersistence during further training, providing insights for the development of\nbackdoor defense methods. Besides, our investigation may contribute to the\ncopyright protection of text-to-image models in the future.\n", "date": "2023-05-07", "categories": []}
{"id": "2308.00344", "title": "Kidnapping Deep Learning-based Multirotors using Optimized Flying\n  Adversarial Patches", "abstract": "  Autonomous flying robots, such as multirotors, often rely on deep learning\nmodels that make predictions based on a camera image, e.g. for pose estimation.\nThese models can predict surprising results if applied to input images outside\nthe training domain. This fault can be exploited by adversarial attacks, for\nexample, by computing small images, so-called adversarial patches, that can be\nplaced in the environment to manipulate the neural network's prediction. We\nintroduce flying adversarial patches, where multiple images are mounted on at\nleast one other flying robot and therefore can be placed anywhere in the field\nof view of a victim multirotor. By introducing the attacker robots, the system\nis extended to an adversarial multi-robot system. For an effective attack, we\ncompare three methods that simultaneously optimize multiple adversarial patches\nand their position in the input image. We show that our methods scale well with\nthe number of adversarial patches. Moreover, we demonstrate physical flights\nwith two robots, where we employ a novel attack policy that uses the computed\nadversarial patches to kidnap a robot that was supposed to follow a human.\n", "date": "2023-08-01", "categories": []}
{"id": "2310.04445", "title": "LoFT: Local Proxy Fine-tuning For Improving Transferability Of\n  Adversarial Attacks Against Large Language Model", "abstract": "  It has been shown that Large Language Model (LLM) alignments can be\ncircumvented by appending specially crafted attack suffixes with harmful\nqueries to elicit harmful responses. To conduct attacks against private target\nmodels whose characterization is unknown, public models can be used as proxies\nto fashion the attack, with successful attacks being transferred from public\nproxies to private target models. The success rate of attack depends on how\nclosely the proxy model approximates the private model. We hypothesize that for\nattacks to be transferrable, it is sufficient if the proxy can approximate the\ntarget model in the neighborhood of the harmful query. Therefore, in this\npaper, we propose \\emph{Local Fine-Tuning (LoFT)}, \\textit{i.e.}, fine-tuning\nproxy models on similar queries that lie in the lexico-semantic neighborhood of\nharmful queries to decrease the divergence between the proxy and target models.\nFirst, we demonstrate three approaches to prompt private target models to\nobtain similar queries given harmful queries. Next, we obtain data for local\nfine-tuning by eliciting responses from target models for the generated similar\nqueries. Then, we optimize attack suffixes to generate attack prompts and\nevaluate the impact of our local fine-tuning on the attack's success rate.\nExperiments show that local fine-tuning of proxy models improves attack\ntransferability and increases attack success rate by $39\\%$, $7\\%$, and $0.5\\%$\n(absolute) on target models ChatGPT, GPT-4, and Claude respectively.\n", "date": "2023-10-02", "categories": []}
{"id": "2310.10427", "title": "DANAA: Towards transferable attacks with double adversarial neuron\n  attribution", "abstract": "  While deep neural networks have excellent results in many fields, they are\nsusceptible to interference from attacking samples resulting in erroneous\njudgments. Feature-level attacks are one of the effective attack types, which\ntargets the learnt features in the hidden layers to improve its transferability\nacross different models. Yet it is observed that the transferability has been\nlargely impacted by the neuron importance estimation results. In this paper, a\ndouble adversarial neuron attribution attack method, termed `DANAA', is\nproposed to obtain more accurate feature importance estimation. In our method,\nthe model outputs are attributed to the middle layer based on an adversarial\nnon-linear path. The goal is to measure the weight of individual neurons and\nretain the features that are more important towards transferability. We have\nconducted extensive experiments on the benchmark datasets to demonstrate the\nstate-of-the-art performance of our method. Our code is available at:\nhttps://github.com/Davidjinzb/DANAA\n", "date": "2023-10-16", "categories": []}
{"id": "2310.13893", "title": "The Hidden Adversarial Vulnerabilities of Medical Federated Learning", "abstract": "  In this paper, we delve into the susceptibility of federated medical image\nanalysis systems to adversarial attacks. Our analysis uncovers a novel\nexploitation avenue: using gradient information from prior global model\nupdates, adversaries can enhance the efficiency and transferability of their\nattacks. Specifically, we demonstrate that single-step attacks (e.g. FGSM),\nwhen aptly initialized, can outperform the efficiency of their iterative\ncounterparts but with reduced computational demand. Our findings underscore the\nneed to revisit our understanding of AI security in federated healthcare\nsettings.\n", "date": "2023-10-21", "categories": []}
{"id": "2310.13950", "title": "Adversarial Image Generation by Spatial Transformation in Perceptual\n  Colorspaces", "abstract": "  Deep neural networks are known to be vulnerable to adversarial perturbations.\nThe amount of these perturbations are generally quantified using $L_p$ metrics,\nsuch as $L_0$, $L_2$ and $L_\\infty$. However, even when the measured\nperturbations are small, they tend to be noticeable by human observers since\n$L_p$ distance metrics are not representative of human perception. On the other\nhand, humans are less sensitive to changes in colorspace. In addition, pixel\nshifts in a constrained neighborhood are hard to notice. Motivated by these\nobservations, we propose a method that creates adversarial examples by applying\nspatial transformations, which creates adversarial examples by changing the\npixel locations independently to chrominance channels of perceptual colorspaces\nsuch as $YC_{b}C_{r}$ and $CIELAB$, instead of making an additive perturbation\nor manipulating pixel values directly. In a targeted white-box attack setting,\nthe proposed method is able to obtain competitive fooling rates with very high\nconfidence. The experimental evaluations show that the proposed method has\nfavorable results in terms of approximate perceptual distance between benign\nand adversarially generated images. The source code is publicly available at\nhttps://github.com/ayberkydn/stadv-torch\n", "date": "2023-10-21", "categories": []}
{"id": "2310.14001", "title": "Toward Stronger Textual Attack Detectors", "abstract": "  The landscape of available textual adversarial attacks keeps growing, posing\nsevere threats and raising concerns regarding the deep NLP system's integrity.\nHowever, the crucial problem of defending against malicious attacks has only\ndrawn the attention of the NLP community. The latter is nonetheless\ninstrumental in developing robust and trustworthy systems. This paper makes two\nimportant contributions in this line of search: (i) we introduce LAROUSSE, a\nnew framework to detect textual adversarial attacks and (ii) we introduce\nSTAKEOUT, a new benchmark composed of nine popular attack methods, three\ndatasets, and two pre-trained models. LAROUSSE is ready-to-use in production as\nit is unsupervised, hyperparameter-free, and non-differentiable, protecting it\nagainst gradient-based methods. Our new benchmark STAKEOUT allows for a robust\nevaluation framework: we conduct extensive numerical experiments which\ndemonstrate that LAROUSSE outperforms previous methods, and which allows to\nidentify interesting factors of detection rate variations.\n", "date": "2023-10-21", "categories": []}
{"id": "2310.14047", "title": "MeaeQ: Mount Model Extraction Attacks with Efficient Queries", "abstract": "  We study model extraction attacks in natural language processing (NLP) where\nattackers aim to steal victim models by repeatedly querying the open\nApplication Programming Interfaces (APIs). Recent works focus on limited-query\nbudget settings and adopt random sampling or active learning-based sampling\nstrategies on publicly available, unannotated data sources. However, these\nmethods often result in selected queries that lack task relevance and data\ndiversity, leading to limited success in achieving satisfactory results with\nlow query costs. In this paper, we propose MeaeQ (Model extraction attack with\nefficient Queries), a straightforward yet effective method to address these\nissues. Specifically, we initially utilize a zero-shot sequence inference\nclassifier, combined with API service information, to filter task-relevant data\nfrom a public text corpus instead of a problem domain-specific dataset.\nFurthermore, we employ a clustering-based data reduction technique to obtain\nrepresentative data as queries for the attack. Extensive experiments conducted\non four benchmark datasets demonstrate that MeaeQ achieves higher functional\nsimilarity to the victim model than baselines while requiring fewer queries.\nOur code is available at https://github.com/C-W-D/MeaeQ.\n", "date": "2023-10-21", "categories": []}
{"id": "2310.14504", "title": "ADoPT: LiDAR Spoofing Attack Detection Based on Point-Level Temporal\n  Consistency", "abstract": "  Deep neural networks (DNNs) are increasingly integrated into LiDAR (Light\nDetection and Ranging)-based perception systems for autonomous vehicles (AVs),\nrequiring robust performance under adversarial conditions. We aim to address\nthe challenge of LiDAR spoofing attacks, where attackers inject fake objects\ninto LiDAR data and fool AVs to misinterpret their environment and make\nerroneous decisions. However, current defense algorithms predominantly depend\non perception outputs (i.e., bounding boxes) thus face limitations in detecting\nattackers given the bounding boxes are generated by imperfect perception models\nprocessing limited points, acquired based on the ego vehicle's viewpoint. To\novercome these limitations, we propose a novel framework, named ADoPT (Anomaly\nDetection based on Point-level Temporal consistency), which quantitatively\nmeasures temporal consistency across consecutive frames and identifies abnormal\nobjects based on the coherency of point clusters. In our evaluation using the\nnuScenes dataset, our algorithm effectively counters various LiDAR spoofing\nattacks, achieving a low (< 10%) false positive ratio (FPR) and high (> 85%)\ntrue positive ratio (TPR), outperforming existing state-of-the-art defense\nmethods, CARLO and 3D-TC2. Furthermore, our evaluation demonstrates the\npromising potential for accurate attack detection across various road\nenvironments.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.14561", "title": "F$^2$AT: Feature-Focusing Adversarial Training via Disentanglement of\n  Natural and Perturbed Patterns", "abstract": "  Deep neural networks (DNNs) are vulnerable to adversarial examples crafted by\nwell-designed perturbations. This could lead to disastrous results on critical\napplications such as self-driving cars, surveillance security, and medical\ndiagnosis. At present, adversarial training is one of the most effective\ndefenses against adversarial examples. However, traditional adversarial\ntraining makes it difficult to achieve a good trade-off between clean accuracy\nand robustness since spurious features are still learned by DNNs. The intrinsic\nreason is that traditional adversarial training makes it difficult to fully\nlearn core features from adversarial examples when adversarial noise and clean\nexamples cannot be disentangled. In this paper, we disentangle the adversarial\nexamples into natural and perturbed patterns by bit-plane slicing. We assume\nthe higher bit-planes represent natural patterns and the lower bit-planes\nrepresent perturbed patterns, respectively. We propose a Feature-Focusing\nAdversarial Training (F$^2$AT), which differs from previous work in that it\nenforces the model to focus on the core features from natural patterns and\nreduce the impact of spurious features from perturbed patterns. The\nexperimental results demonstrated that F$^2$AT outperforms state-of-the-art\nmethods in clean accuracy and adversarial robustness.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.13099", "title": "No offence, Bert -- I insult only humans! Multiple addressees\n  sentence-level attack on toxicity detection neural network", "abstract": "  We introduce a simple yet efficient sentence-level attack on black-box\ntoxicity detector models. By adding several positive words or sentences to the\nend of a hateful message, we are able to change the prediction of a neural\nnetwork and pass the toxicity detection system check. This approach is shown to\nbe working on seven languages from three different language families. We also\ndescribe the defence mechanism against the aforementioned attack and discuss\nits limitations.\n", "date": "2023-10-19", "categories": []}
{"id": "2305.05896", "title": "A Black-Box Attack on Code Models via Representation Nearest Neighbor\n  Search", "abstract": "  Existing methods for generating adversarial code examples face several\nchallenges: limted availability of substitute variables, high verification\ncosts for these substitutes, and the creation of adversarial samples with\nnoticeable perturbations. To address these concerns, our proposed approach,\nRNNS, uses a search seed based on historical attacks to find potential\nadversarial substitutes. Rather than directly using the discrete substitutes,\nthey are mapped to a continuous vector space using a pre-trained variable name\nencoder. Based on the vector representation, RNNS predicts and selects better\nsubstitutes for attacks. We evaluated the performance of RNNS across six coding\ntasks encompassing three programming languages: Java, Python, and C. We\nemployed three pre-trained code models (CodeBERT, GraphCodeBERT, and CodeT5)\nthat resulted in a cumulative of 18 victim models. The results demonstrate that\nRNNS outperforms baselines in terms of ASR and QT. Furthermore, the\nperturbation of adversarial examples introduced by RNNS is smaller compared to\nthe baselines in terms of the number of replaced variables and the change in\nvariable length. Lastly, our experiments indicate that RNNS is efficient in\nattacking defended models and can be employed for adversarial training.\n", "date": "2023-05-10", "categories": []}
{"id": "2310.12243", "title": "REVAMP: Automated Simulations of Adversarial Attacks on Arbitrary\n  Objects in Realistic Scenes", "abstract": "  Deep Learning models, such as those used in an autonomous vehicle are\nvulnerable to adversarial attacks where an attacker could place an adversarial\nobject in the environment, leading to mis-classification. Generating these\nadversarial objects in the digital space has been extensively studied, however\nsuccessfully transferring these attacks from the digital realm to the physical\nrealm has proven challenging when controlling for real-world environmental\nfactors. In response to these limitations, we introduce REVAMP, an easy-to-use\nPython library that is the first-of-its-kind tool for creating attack scenarios\nwith arbitrary objects and simulating realistic environmental factors,\nlighting, reflection, and refraction. REVAMP enables researchers and\npractitioners to swiftly explore various scenarios within the digital realm by\noffering a wide range of configurable options for designing experiments and\nusing differentiable rendering to reproduce physically plausible adversarial\nobjects. We will demonstrate and invite the audience to try REVAMP to produce\nan adversarial texture on a chosen object while having control over various\nscene parameters. The audience will choose a scene, an object to attack, the\ndesired attack class, and the number of camera positions to use. Then, in real\ntime, we show how this altered texture causes the chosen object to be\nmis-classified, showcasing the potential of REVAMP in real-world scenarios.\nREVAMP is open-source and available at https://github.com/poloclub/revamp.\n", "date": "2023-10-18", "categories": []}
{"id": "2310.12432", "title": "CAT: Closed-loop Adversarial Training for Safe End-to-End Driving", "abstract": "  Driving safety is a top priority for autonomous vehicles. Orthogonal to prior\nwork handling accident-prone traffic events by algorithm designs at the policy\nlevel, we investigate a Closed-loop Adversarial Training (CAT) framework for\nsafe end-to-end driving in this paper through the lens of environment\naugmentation. CAT aims to continuously improve the safety of driving agents by\ntraining the agent on safety-critical scenarios that are dynamically generated\nover time. A novel resampling technique is developed to turn log-replay\nreal-world driving scenarios into safety-critical ones via probabilistic\nfactorization, where the adversarial traffic generation is modeled as the\nmultiplication of standard motion prediction sub-problems. Consequently, CAT\ncan launch more efficient physical attacks compared to existing safety-critical\nscenario generation methods and yields a significantly less computational cost\nin the iterative learning pipeline. We incorporate CAT into the MetaDrive\nsimulator and validate our approach on hundreds of driving scenarios imported\nfrom real-world driving datasets. Experimental results demonstrate that CAT can\neffectively generate adversarial scenarios countering the agent being trained.\nAfter training, the agent can achieve superior driving safety in both\nlog-replay and safety-critical traffic scenarios on the held-out test set. Code\nand data are available at https://metadriverse.github.io/cat.\n", "date": "2023-10-19", "categories": []}
{"id": "2310.12708", "title": "Generating Robust Adversarial Examples against Online Social Networks\n  (OSNs)", "abstract": "  Online Social Networks (OSNs) have blossomed into prevailing transmission\nchannels for images in the modern era. Adversarial examples (AEs) deliberately\ndesigned to mislead deep neural networks (DNNs) are found to be fragile against\nthe inevitable lossy operations conducted by OSNs. As a result, the AEs would\nlose their attack capabilities after being transmitted over OSNs. In this work,\nwe aim to design a new framework for generating robust AEs that can survive the\nOSN transmission; namely, the AEs before and after the OSN transmission both\npossess strong attack capabilities. To this end, we first propose a\ndifferentiable network termed SImulated OSN (SIO) to simulate the various\noperations conducted by an OSN. Specifically, the SIO network consists of two\nmodules: 1) a differentiable JPEG layer for approximating the ubiquitous JPEG\ncompression and 2) an encoder-decoder subnetwork for mimicking the remaining\noperations. Based upon the SIO network, we then formulate an optimization\nframework to generate robust AEs by enforcing model outputs with and without\npassing through the SIO to be both misled. Extensive experiments conducted over\nFacebook, WeChat and QQ demonstrate that our attack methods produce more robust\nAEs than existing approaches, especially under small distortion constraints;\nthe performance gain in terms of Attack Success Rate (ASR) could be more than\n60%. Furthermore, we build a public dataset containing more than 10,000 pairs\nof AEs processed by Facebook, WeChat or QQ, facilitating future research in the\nrobust AEs generation. The dataset and code are available at\nhttps://github.com/csjunjun/RobustOSNAttack.git.\n", "date": "2023-10-19", "categories": []}
{"id": "2302.08466", "title": "Marich: A Query-efficient Distributionally Equivalent Model Extraction\n  Attack using Public Data", "abstract": "  We study design of black-box model extraction attacks that can send minimal\nnumber of queries from a publicly available dataset to a target ML model\nthrough a predictive API with an aim to create an informative and\ndistributionally equivalent replica of the target. First, we define\ndistributionally equivalent and Max-Information model extraction attacks, and\nreduce them into a variational optimisation problem. The attacker sequentially\nsolves this optimisation problem to select the most informative queries that\nsimultaneously maximise the entropy and reduce the mismatch between the target\nand the stolen models. This leads to an active sampling-based query selection\nalgorithm, Marich, which is model-oblivious. Then, we evaluate Marich on\ndifferent text and image data sets, and different models, including CNNs and\nBERT. Marich extracts models that achieve $\\sim 60-95\\%$ of true model's\naccuracy and uses $\\sim 1,000 - 8,500$ queries from the publicly available\ndatasets, which are different from the private training datasets. Models\nextracted by Marich yield prediction distributions, which are $\\sim 2-4\\times$\ncloser to the target's distribution in comparison to the existing active\nsampling-based attacks. The extracted models also lead to $84-96\\%$ accuracy\nunder membership inference attacks. Experimental results validate that Marich\nis query-efficient, and capable of performing task-accurate, high-fidelity, and\ninformative model extraction.\n", "date": "2023-02-16", "categories": []}
{"id": "2310.10780", "title": "Demystifying Poisoning Backdoor Attacks from a Statistical Perspective", "abstract": "  The growing dependence on machine learning in real-world applications\nemphasizes the importance of understanding and ensuring its safety. Backdoor\nattacks pose a significant security risk due to their stealthy nature and\npotentially serious consequences. Such attacks involve embedding triggers\nwithin a learning model with the intention of causing malicious behavior when\nan active trigger is present while maintaining regular functionality without\nit. This paper evaluates the effectiveness of any backdoor attack incorporating\na constant trigger, by establishing tight lower and upper boundaries for the\nperformance of the compromised model on both clean and backdoor test data. The\ndeveloped theory answers a series of fundamental but previously underexplored\nproblems, including (1) what are the determining factors for a backdoor\nattack's success, (2) what is the direction of the most effective backdoor\nattack, and (3) when will a human-imperceptible trigger succeed. Our derived\nunderstanding applies to both discriminative and generative models. We also\ndemonstrate the theory by conducting experiments using benchmark datasets and\nstate-of-the-art backdoor attack scenarios.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.11597", "title": "The Efficacy of Transformer-based Adversarial Attacks in Security\n  Domains", "abstract": "  Today, the security of many domains rely on the use of Machine Learning to\ndetect threats, identify vulnerabilities, and safeguard systems from attacks.\nRecently, transformer architectures have improved the state-of-the-art\nperformance on a wide range of tasks such as malware detection and network\nintrusion detection. But, before abandoning current approaches to transformers,\nit is crucial to understand their properties and implications on cybersecurity\napplications. In this paper, we evaluate the robustness of transformers to\nadversarial samples for system defenders (i.e., resiliency to adversarial\nperturbations generated on different types of architectures) and their\nadversarial strength for system attackers (i.e., transferability of adversarial\nsamples generated by transformers to other target models). To that effect, we\nfirst fine-tune a set of pre-trained transformer, Convolutional Neural Network\n(CNN), and hybrid (an ensemble of transformer and CNN) models to solve\ndifferent downstream image-based tasks. Then, we use an attack algorithm to\ncraft 19,367 adversarial examples on each model for each task. The\ntransferability of these adversarial examples is measured by evaluating each\nset on other models to determine which models offer more adversarial strength,\nand consequently, more robustness against these attacks. We find that the\nadversarial examples crafted on transformers offer the highest transferability\nrate (i.e., 25.7% higher than the average) onto other models. Similarly,\nadversarial examples crafted on other models have the lowest rate of\ntransferability (i.e., 56.7% lower than the average) onto transformers. Our\nwork emphasizes the importance of studying transformer architectures for\nattacking and defending models in security domains, and suggests using them as\nthe primary architecture in transfer attack settings.\n", "date": "2023-10-17", "categories": []}
{"id": "2303.09767", "title": "It Is All About Data: A Survey on the Effects of Data on Adversarial\n  Robustness", "abstract": "  Adversarial examples are inputs to machine learning models that an attacker\nhas intentionally designed to confuse the model into making a mistake. Such\nexamples pose a serious threat to the applicability of machine-learning-based\nsystems, especially in life- and safety-critical domains. To address this\nproblem, the area of adversarial robustness investigates mechanisms behind\nadversarial attacks and defenses against these attacks. This survey reviews a\nparticular subset of this literature that focuses on investigating properties\nof training data in the context of model robustness under evasion attacks. It\nfirst summarizes the main properties of data leading to adversarial\nvulnerability. It then discusses guidelines and techniques for improving\nadversarial robustness by enhancing the data representation and learning\nprocedures, as well as techniques for estimating robustness guarantees given\nparticular data. Finally, it discusses gaps of knowledge and promising future\nresearch directions in this area.\n", "date": "2023-03-17", "categories": []}
{"id": "2310.10744", "title": "Fast Adversarial Label-Flipping Attack on Tabular Data", "abstract": "  Machine learning models are increasingly used in fields that require high\nreliability such as cybersecurity. However, these models remain vulnerable to\nvarious attacks, among which the adversarial label-flipping attack poses\nsignificant threats. In label-flipping attacks, the adversary maliciously flips\na portion of training labels to compromise the machine learning model. This\npaper raises significant concerns as these attacks can camouflage a highly\nskewed dataset as an easily solvable classification problem, often misleading\nmachine learning practitioners into lower defenses and miscalculations of\npotential risks. This concern amplifies in tabular data settings, where\nidentifying true labels requires expertise, allowing malicious label-flipping\nattacks to easily slip under the radar. To demonstrate this risk is inherited\nin the adversary's objective, we propose FALFA (Fast Adversarial Label-Flipping\nAttack), a novel efficient attack for crafting adversarial labels. FALFA is\nbased on transforming the adversary's objective and employs linear programming\nto reduce computational complexity. Using ten real-world tabular datasets, we\ndemonstrate FALFA's superior attack potential, highlighting the need for robust\ndefenses against such threats.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.10807", "title": "Regularization properties of adversarially-trained linear regression", "abstract": "  State-of-the-art machine learning models can be vulnerable to very small\ninput perturbations that are adversarially constructed. Adversarial training is\nan effective approach to defend against it. Formulated as a min-max problem, it\nsearches for the best solution when the training data were corrupted by the\nworst-case attacks. Linear models are among the simple models where\nvulnerabilities can be observed and are the focus of our study. In this case,\nadversarial training leads to a convex optimization problem which can be\nformulated as the minimization of a finite sum. We provide a comparative\nanalysis between the solution of adversarial training in linear regression and\nother regularization methods. Our main findings are that: (A) Adversarial\ntraining yields the minimum-norm interpolating solution in the\noverparameterized regime (more parameters than data), as long as the maximum\ndisturbance radius is smaller than a threshold. And, conversely, the\nminimum-norm interpolator is the solution to adversarial training with a given\nradius. (B) Adversarial training can be equivalent to parameter shrinking\nmethods (ridge regression and Lasso). This happens in the underparametrized\nregion, for an appropriate choice of adversarial radius and zero-mean\nsymmetrically distributed covariates. (C) For $\\ell_\\infty$-adversarial\ntraining -- as in square-root Lasso -- the choice of adversarial radius for\noptimal bounds does not depend on the additive noise variance. We confirm our\ntheoretical findings with numerical examples.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.10844", "title": "Survey of Vulnerabilities in Large Language Models Revealed by\n  Adversarial Attacks", "abstract": "  Large Language Models (LLMs) are swiftly advancing in architecture and\ncapability, and as they integrate more deeply into complex systems, the urgency\nto scrutinize their security properties grows. This paper surveys research in\nthe emerging interdisciplinary field of adversarial attacks on LLMs, a subfield\nof trustworthy ML, combining the perspectives of Natural Language Processing\nand Security. Prior work has shown that even safety-aligned LLMs (via\ninstruction tuning and reinforcement learning through human feedback) can be\nsusceptible to adversarial attacks, which exploit weaknesses and mislead AI\nsystems, as evidenced by the prevalence of `jailbreak' attacks on models like\nChatGPT and Bard. In this survey, we first provide an overview of large\nlanguage models, describe their safety alignment, and categorize existing\nresearch based on various learning structures: textual-only attacks,\nmulti-modal attacks, and additional attack methods specifically targeting\ncomplex systems, such as federated learning or multi-agent systems. We also\noffer comprehensive remarks on works that focus on the fundamental sources of\nvulnerabilities and potential defenses. To make this field more accessible to\nnewcomers, we present a systematic review of existing works, a structured\ntypology of adversarial attack concepts, and additional resources, including\nslides for presentations on related topics at the 62nd Annual Meeting of the\nAssociation for Computational Linguistics (ACL'24).\n", "date": "2023-10-16", "categories": []}
{"id": "2305.01074", "title": "Physical Adversarial Attacks for Surveillance: A Survey", "abstract": "  Modern automated surveillance techniques are heavily reliant on deep learning\nmethods. Despite the superior performance, these learning systems are\ninherently vulnerable to adversarial attacks - maliciously crafted inputs that\nare designed to mislead, or trick, models into making incorrect predictions. An\nadversary can physically change their appearance by wearing adversarial\nt-shirts, glasses, or hats or by specific behavior, to potentially avoid\nvarious forms of detection, tracking and recognition of surveillance systems;\nand obtain unauthorized access to secure properties and assets. This poses a\nsevere threat to the security and safety of modern surveillance systems. This\npaper reviews recent attempts and findings in learning and designing physical\nadversarial attacks for surveillance applications. In particular, we propose a\nframework to analyze physical adversarial attacks and provide a comprehensive\nsurvey of physical adversarial attacks on four key surveillance tasks:\ndetection, identification, tracking, and action recognition under this\nframework. Furthermore, we review and analyze strategies to defend against the\nphysical adversarial attacks and the methods for evaluating the strengths of\nthe defense. The insights in this paper present an important step in building\nresilience within surveillance systems to physical adversarial attacks.\n", "date": "2023-05-01", "categories": []}
{"id": "2309.11751", "title": "How Robust is Google's Bard to Adversarial Image Attacks?", "abstract": "  Multimodal Large Language Models (MLLMs) that integrate text and other\nmodalities (especially vision) have achieved unprecedented performance in\nvarious multimodal tasks. However, due to the unsolved adversarial robustness\nproblem of vision models, MLLMs can have more severe safety and security risks\nby introducing the vision inputs. In this work, we study the adversarial\nrobustness of Google's Bard, a competitive chatbot to ChatGPT that released its\nmultimodal capability recently, to better understand the vulnerabilities of\ncommercial MLLMs. By attacking white-box surrogate vision encoders or MLLMs,\nthe generated adversarial examples can mislead Bard to output wrong image\ndescriptions with a 22% success rate based solely on the transferability. We\nshow that the adversarial examples can also attack other MLLMs, e.g., a 26%\nattack success rate against Bing Chat and a 86% attack success rate against\nERNIE bot. Moreover, we identify two defense mechanisms of Bard, including face\ndetection and toxicity detection of images. We design corresponding attacks to\nevade these defenses, demonstrating that the current defenses of Bard are also\nvulnerable. We hope this work can deepen our understanding on the robustness of\nMLLMs and facilitate future research on defenses. Our code is available at\nhttps://github.com/thu-ml/Attack-Bard.\n  Update: GPT-4V is available at October 2023. We further evaluate its\nrobustness under the same set of adversarial examples, achieving a 45% attack\nsuccess rate.\n", "date": "2023-09-21", "categories": []}
{"id": "2310.03166", "title": "Raze to the Ground: Query-Efficient Adversarial HTML Attacks on\n  Machine-Learning Phishing Webpage Detectors", "abstract": "  Machine-learning phishing webpage detectors (ML-PWD) have been shown to\nsuffer from adversarial manipulations of the HTML code of the input webpage.\nNevertheless, the attacks recently proposed have demonstrated limited\neffectiveness due to their lack of optimizing the usage of the adopted\nmanipulations, and they focus solely on specific elements of the HTML code. In\nthis work, we overcome these limitations by first designing a novel set of\nfine-grained manipulations which allow to modify the HTML code of the input\nphishing webpage without compromising its maliciousness and visual appearance,\ni.e., the manipulations are functionality- and rendering-preserving by design.\nWe then select which manipulations should be applied to bypass the target\ndetector by a query-efficient black-box optimization algorithm. Our experiments\nshow that our attacks are able to raze to the ground the performance of current\nstate-of-the-art ML-PWD using just 30 queries, thus overcoming the weaker\nattacks developed in previous work, and enabling a much fairer robustness\nevaluation of ML-PWD.\n", "date": "2023-10-04", "categories": []}
{"id": "2310.09361", "title": "Is Certifying $\\ell_p$ Robustness Still Worthwhile?", "abstract": "  Over the years, researchers have developed myriad attacks that exploit the\nubiquity of adversarial examples, as well as defenses that aim to guard against\nthe security vulnerabilities posed by such attacks. Of particular interest to\nthis paper are defenses that provide provable guarantees against the class of\n$\\ell_p$-bounded attacks. Certified defenses have made significant progress,\ntaking robustness certification from toy models and datasets to large-scale\nproblems like ImageNet classification. While this is undoubtedly an interesting\nacademic problem, as the field has matured, its impact in practice remains\nunclear, thus we find it useful to revisit the motivation for continuing this\nline of research. There are three layers to this inquiry, which we address in\nthis paper: (1) why do we care about robustness research? (2) why do we care\nabout the $\\ell_p$-bounded threat model? And (3) why do we care about\ncertification as opposed to empirical defenses? In brief, we take the position\nthat local robustness certification indeed confers practical value to the field\nof machine learning. We focus especially on the latter two questions from\nabove. With respect to the first of the two, we argue that the $\\ell_p$-bounded\nthreat model acts as a minimal requirement for safe application of models in\nsecurity-critical domains, while at the same time, evidence has mounted\nsuggesting that local robustness may lead to downstream external benefits not\nimmediately related to robustness. As for the second, we argue that (i)\ncertification provides a resolution to the cat-and-mouse game of adversarial\nattacks; and furthermore, that (ii) perhaps contrary to popular belief, there\nmay not exist a fundamental trade-off between accuracy, robustness, and\ncertifiability, while moreover, certified training techniques constitute a\nparticularly promising way for learning robust models.\n", "date": "2023-10-13", "categories": []}
{"id": "2310.09744", "title": "Explore the Effect of Data Selection on Poison Efficiency in Backdoor\n  Attacks", "abstract": "  As the number of parameters in Deep Neural Networks (DNNs) scales, the thirst\nfor training data also increases. To save costs, it has become common for users\nand enterprises to delegate time-consuming data collection to third parties.\nUnfortunately, recent research has shown that this practice raises the risk of\nDNNs being exposed to backdoor attacks. Specifically, an attacker can\nmaliciously control the behavior of a trained model by poisoning a small\nportion of the training data. In this study, we focus on improving the\npoisoning efficiency of backdoor attacks from the sample selection perspective.\nThe existing attack methods construct such poisoned samples by randomly\nselecting some clean data from the benign set and then embedding a trigger into\nthem. However, this random selection strategy ignores that each sample may\ncontribute differently to the backdoor injection, thereby reducing the\npoisoning efficiency. To address the above problem, a new selection strategy\nnamed Improved Filtering and Updating Strategy (FUS++) is proposed.\nSpecifically, we adopt the forgetting events of the samples to indicate the\ncontribution of different poisoned samples and use the curvature of the loss\nsurface to analyses the effectiveness of this phenomenon. Accordingly, we\ncombine forgetting events and curvature of different samples to conduct a\nsimple yet efficient sample selection strategy. The experimental results on\nimage classification (CIFAR-10, CIFAR-100, ImageNet-10), text classification\n(AG News), audio classification (ESC-50), and age regression (Facial Age)\nconsistently demonstrate the effectiveness of the proposed strategy: the attack\nperformance using FUS++ is significantly higher than that using random\nselection for the same poisoning ratio.\n", "date": "2023-10-15", "categories": []}
{"id": "2310.09795", "title": "AFLOW: Developing Adversarial Examples under Extremely Noise-limited\n  Settings", "abstract": "  Extensive studies have demonstrated that deep neural networks (DNNs) are\nvulnerable to adversarial attacks. Despite the significant progress in the\nattack success rate that has been made recently, the adversarial noise\ngenerated by most of the existing attack methods is still too conspicuous to\nthe human eyes and proved to be easily detected by defense mechanisms.\nResulting that these malicious examples cannot contribute to exploring the\nvulnerabilities of existing DNNs sufficiently. Thus, to better reveal the\ndefects of DNNs and further help enhance their robustness under noise-limited\nsituations, a new inconspicuous adversarial examples generation method is\nexactly needed to be proposed. To bridge this gap, we propose a novel Normalize\nFlow-based end-to-end attack framework, called AFLOW, to synthesize\nimperceptible adversarial examples under strict constraints. Specifically,\nrather than the noise-adding manner, AFLOW directly perturbs the hidden\nrepresentation of the corresponding image to craft the desired adversarial\nexamples. Compared with existing methods, extensive experiments on three\nbenchmark datasets show that the adversarial examples built by AFLOW exhibit\nsuperiority in imperceptibility, image quality and attack capability. Even on\nrobust models, AFLOW can still achieve higher attack results than previous\nmethods.\n", "date": "2023-10-15", "categories": []}
{"id": "2310.09891", "title": "Towards Deep Learning Models Resistant to Transfer-based Adversarial\n  Attacks via Data-centric Robust Learning", "abstract": "  Transfer-based adversarial attacks raise a severe threat to real-world deep\nlearning systems since they do not require access to target models. Adversarial\ntraining (AT), which is recognized as the strongest defense against white-box\nattacks, has also guaranteed high robustness to (black-box) transfer-based\nattacks. However, AT suffers from heavy computational overhead since it\noptimizes the adversarial examples during the whole training process. In this\npaper, we demonstrate that such heavy optimization is not necessary for AT\nagainst transfer-based attacks. Instead, a one-shot adversarial augmentation\nprior to training is sufficient, and we name this new defense paradigm\nData-centric Robust Learning (DRL). Our experimental results show that DRL\noutperforms widely-used AT techniques (e.g., PGD-AT, TRADES, EAT, and FAT) in\nterms of black-box robustness and even surpasses the top-1 defense on\nRobustBench when combined with diverse data augmentations and loss\nregularizations. We also identify other benefits of DRL, for instance, the\nmodel generalization capability and robust fairness.\n", "date": "2023-10-15", "categories": []}
{"id": "2310.10036", "title": "Evading Detection Actively: Toward Anti-Forensics against Forgery\n  Localization", "abstract": "  Anti-forensics seeks to eliminate or conceal traces of tampering artifacts.\nTypically, anti-forensic methods are designed to deceive binary detectors and\npersuade them to misjudge the authenticity of an image. However, to the best of\nour knowledge, no attempts have been made to deceive forgery detectors at the\npixel level and mis-locate forged regions. Traditional adversarial attack\nmethods cannot be directly used against forgery localization due to the\nfollowing defects: 1) they tend to just naively induce the target forensic\nmodels to flip their pixel-level pristine or forged decisions; 2) their\nanti-forensics performance tends to be severely degraded when faced with the\nunseen forensic models; 3) they lose validity once the target forensic models\nare retrained with the anti-forensics images generated by them. To tackle the\nthree defects, we propose SEAR (Self-supErvised Anti-foRensics), a novel\nself-supervised and adversarial training algorithm that effectively trains\ndeep-learning anti-forensic models against forgery localization. SEAR sets a\npretext task to reconstruct perturbation for self-supervised learning. In\nadversarial training, SEAR employs a forgery localization model as a supervisor\nto explore tampering features and constructs a deep-learning concealer to erase\ncorresponding traces. We have conducted largescale experiments across diverse\ndatasets. The experimental results demonstrate that, through the combination of\nself-supervised learning and adversarial learning, SEAR successfully deceives\nthe state-of-the-art forgery localization methods, as well as tackle the three\ndefects regarding traditional adversarial attack methods mentioned above.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.10077", "title": "Prompt Packer: Deceiving LLMs through Compositional Instruction with\n  Hidden Attacks", "abstract": "  Recently, Large language models (LLMs) with powerful general capabilities\nhave been increasingly integrated into various Web applications, while\nundergoing alignment training to ensure that the generated content aligns with\nuser intent and ethics. Unfortunately, they remain the risk of generating\nharmful content like hate speech and criminal activities in practical\napplications. Current approaches primarily rely on detecting, collecting, and\ntraining against harmful prompts to prevent such risks. However, they typically\nfocused on the \"superficial\" harmful prompts with a solitary intent, ignoring\ncomposite attack instructions with multiple intentions that can easily elicit\nharmful content in real-world scenarios. In this paper, we introduce an\ninnovative technique for obfuscating harmful instructions: Compositional\nInstruction Attacks (CIA), which refers to attacking by combination and\nencapsulation of multiple instructions. CIA hides harmful prompts within\ninstructions of harmless intentions, making it impossible for the model to\nidentify underlying malicious intentions. Furthermore, we implement two\ntransformation methods, known as T-CIA and W-CIA, to automatically disguise\nharmful instructions as talking or writing tasks, making them appear harmless\nto LLMs. We evaluated CIA on GPT-4, ChatGPT, and ChatGLM2 with two safety\nassessment datasets and two harmful prompt datasets. It achieves an attack\nsuccess rate of 95%+ on safety assessment datasets, and 83%+ for GPT-4, 91%+\nfor ChatGPT (gpt-3.5-turbo backed) and ChatGLM2-6B on harmful prompt datasets.\nOur approach reveals the vulnerability of LLMs to such compositional\ninstruction attacks that harbor underlying harmful intentions, contributing\nsignificantly to LLM security development. Warning: this paper may contain\noffensive or upsetting content!\n", "date": "2023-10-16", "categories": []}
{"id": "2307.02599", "title": "Evade ChatGPT Detectors via A Single Space", "abstract": "  ChatGPT brings revolutionary social value but also raises concerns about the\nmisuse of AI-generated text. Consequently, an important question is how to\ndetect whether texts are generated by ChatGPT or by human. Existing detectors\nare built upon the assumption that there are distributional gaps between\nhuman-generated and AI-generated text. These gaps are typically identified\nusing statistical information or classifiers. Our research challenges the\ndistributional gap assumption in detectors. We find that detectors do not\neffectively discriminate the semantic and stylistic gaps between\nhuman-generated and AI-generated text. Instead, the \"subtle differences\", such\nas an extra space, become crucial for detection. Based on this discovery, we\npropose the SpaceInfi strategy to evade detection. Experiments demonstrate the\neffectiveness of this strategy across multiple benchmarks and detectors. We\nalso provide a theoretical explanation for why SpaceInfi is successful in\nevading perplexity-based detection. And we empirically show that a phenomenon\ncalled token mutation causes the evasion for language model-based detectors.\nOur findings offer new insights and challenges for understanding and\nconstructing more applicable ChatGPT detectors.\n", "date": "2023-07-05", "categories": []}
{"id": "2210.13660", "title": "Multi-SpacePhish: Extending the Evasion-space of Adversarial Attacks\n  against Phishing Website Detectors using Machine Learning", "abstract": "  Existing literature on adversarial Machine Learning (ML) focuses either on\nshowing attacks that break every ML model, or defenses that withstand most\nattacks. Unfortunately, little consideration is given to the actual feasibility\nof the attack or the defense. Moreover, adversarial samples are often crafted\nin the \"feature-space\", making the corresponding evaluations of questionable\nvalue. Simply put, the current situation does not allow to estimate the actual\nthreat posed by adversarial attacks, leading to a lack of secure ML systems.\n  We aim to clarify such confusion in this paper. By considering the\napplication of ML for Phishing Website Detection (PWD), we formalize the\n\"evasion-space\" in which an adversarial perturbation can be introduced to fool\na ML-PWD -- demonstrating that even perturbations in the \"feature-space\" are\nuseful. Then, we propose a realistic threat model describing evasion attacks\nagainst ML-PWD that are cheap to stage, and hence intrinsically more attractive\nfor real phishers. After that, we perform the first statistically validated\nassessment of state-of-the-art ML-PWD against 12 evasion attacks. Our\nevaluation shows (i) the true efficacy of evasion attempts that are more likely\nto occur; and (ii) the impact of perturbations crafted in different\nevasion-spaces. Our realistic evasion attempts induce a statistically\nsignificant degradation (3-10% at p<0.05), and their cheap cost makes them a\nsubtle threat. Notably, however, some ML-PWD are immune to our most realistic\nattacks (p=0.22).\n  Finally, as an additional contribution of this journal publication, we are\nthe first to consider the intriguing case wherein an attacker introduces\nperturbations in multiple evasion-spaces at the same time. These new results\nshow that simultaneously applying perturbations in the problem- and\nfeature-space can cause a drop in the detection rate from 0.95 to 0.\n", "date": "2022-10-24", "categories": []}
{"id": "2310.07725", "title": "Extreme Image Transformations Facilitate Robust Latent Object\n  Representations", "abstract": "  Adversarial attacks can affect the object recognition capabilities of\nmachines in wild. These can often result from spurious correlations between\ninput and class labels, and are prone to memorization in large networks. While\nnetworks are expected to do automated feature selection, it is not effective at\nthe scale of the object. Humans, however, are able to select the minimum set of\nfeatures required to form a robust representation of an object. In this work,\nwe show that finetuning any pretrained off-the-shelf network with Extreme Image\nTransformations (EIT) not only helps in learning a robust latent\nrepresentation, it also improves the performance of these networks against\ncommon adversarial attacks of various intensities. Our EIT trained networks\nshow strong activations in the object regions even when tested with more\nintense noise, showing promising generalizations across different kinds of\nadversarial attacks.\n", "date": "2023-09-19", "categories": []}
{"id": "2310.07780", "title": "Promoting Robustness of Randomized Smoothing: Two Cost-Effective\n  Approaches", "abstract": "  Randomized smoothing has recently attracted attentions in the field of\nadversarial robustness to provide provable robustness guarantees on smoothed\nneural network classifiers. However, existing works show that vanilla\nrandomized smoothing usually does not provide good robustness performance and\noften requires (re)training techniques on the base classifier in order to boost\nthe robustness of the resulting smoothed classifier. In this work, we propose\ntwo cost-effective approaches to boost the robustness of randomized smoothing\nwhile preserving its clean performance. The first approach introduces a new\nrobust training method AdvMacerwhich combines adversarial training and\nrobustness certification maximization for randomized smoothing. We show that\nAdvMacer can improve the robustness performance of randomized smoothing\nclassifiers compared to SOTA baselines, while being 3x faster to train than\nMACER baseline. The second approach introduces a post-processing method EsbRS\nwhich greatly improves the robustness certificate based on building model\nensembles. We explore different aspects of model ensembles that has not been\nstudied by prior works and propose a novel design methodology to further\nimprove robustness of the ensemble based on our theoretical analysis.\n", "date": "2023-10-11", "categories": []}
{"id": "2310.08073", "title": "Samples on Thin Ice: Re-Evaluating Adversarial Pruning of Neural\n  Networks", "abstract": "  Neural network pruning has shown to be an effective technique for reducing\nthe network size, trading desirable properties like generalization and\nrobustness to adversarial attacks for higher sparsity. Recent work has claimed\nthat adversarial pruning methods can produce sparse networks while also\npreserving robustness to adversarial examples. In this work, we first\nre-evaluate three state-of-the-art adversarial pruning methods, showing that\ntheir robustness was indeed overestimated. We then compare pruned and dense\nversions of the same models, discovering that samples on thin ice, i.e., closer\nto the unpruned model's decision boundary, are typically misclassified after\npruning. We conclude by discussing how this intuition may lead to designing\nmore effective adversarial pruning methods in future work.\n", "date": "2023-10-12", "categories": []}
{"id": "2310.08177", "title": "Improving Fast Minimum-Norm Attacks with Hyperparameter Optimization", "abstract": "  Evaluating the adversarial robustness of machine learning models using\ngradient-based attacks is challenging. In this work, we show that\nhyperparameter optimization can improve fast minimum-norm attacks by automating\nthe selection of the loss function, the optimizer and the step-size scheduler,\nalong with the corresponding hyperparameters. Our extensive evaluation\ninvolving several robust models demonstrates the improved efficacy of fast\nminimum-norm attacks when hyper-up with hyperparameter optimization. We release\nour open-source code at https://github.com/pralab/HO-FMN.\n", "date": "2023-10-12", "categories": []}
{"id": "2310.08292", "title": "Concealed Electronic Countermeasures of Radar Signal with Adversarial\n  Examples", "abstract": "  Electronic countermeasures involving radar signals are an important aspect of\nmodern warfare. Traditional electronic countermeasures techniques typically add\nlarge-scale interference signals to ensure interference effects, which can lead\nto attacks being too obvious. In recent years, AI-based attack methods have\nemerged that can effectively solve this problem, but the attack scenarios are\ncurrently limited to time domain radar signal classification. In this paper, we\nfocus on the time-frequency images classification scenario of radar signals. We\nfirst propose an attack pipeline under the time-frequency images scenario and\nDITIMI-FGSM attack algorithm with high transferability. Then, we propose\nSTFT-based time domain signal attack(STDS) algorithm to solve the problem of\nnon-invertibility in time-frequency analysis, thus obtaining the time-domain\nrepresentation of the interference signal. A large number of experiments show\nthat our attack pipeline is feasible and the proposed attack method has a high\nsuccess rate.\n", "date": "2023-10-12", "categories": []}
{"id": "2307.14539", "title": "Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal\n  Language Models", "abstract": "  We introduce new jailbreak attacks on vision language models (VLMs), which\nuse aligned LLMs and are resilient to text-only jailbreak attacks.\nSpecifically, we develop cross-modality attacks on alignment where we pair\nadversarial images going through the vision encoder with textual prompts to\nbreak the alignment of the language model. Our attacks employ a novel\ncompositional strategy that combines an image, adversarially targeted towards\ntoxic embeddings, with generic prompts to accomplish the jailbreak. Thus, the\nLLM draws the context to answer the generic prompt from the adversarial image.\nThe generation of benign-appearing adversarial images leverages a novel\nembedding-space-based methodology, operating with no access to the LLM model.\nInstead, the attacks require access only to the vision encoder and utilize one\nof our four embedding space targeting strategies. By not requiring access to\nthe LLM, the attacks lower the entry barrier for attackers, particularly when\nvision encoders such as CLIP are embedded in closed-source LLMs. The attacks\nachieve a high success rate across different VLMs, highlighting the risk of\ncross-modality alignment vulnerabilities, and the need for new alignment\napproaches for multi-modal models.\n", "date": "2023-07-26", "categories": []}
{"id": "2310.06845", "title": "RobustEdge: Low Power Adversarial Detection for Cloud-Edge Systems", "abstract": "  In practical cloud-edge scenarios, where a resource constrained edge performs\ndata acquisition and a cloud system (having sufficient resources) performs\ninference tasks with a deep neural network (DNN), adversarial robustness is\ncritical for reliability and ubiquitous deployment. Adversarial detection is a\nprime adversarial defence technique used in prior literature. However, in prior\ndetection works, the detector is attached to the classifier model and both\ndetector and classifier work in tandem to perform adversarial detection that\nrequires a high computational overhead which is not available at the low-power\nedge. Therefore, prior works can only perform adversarial detection at the\ncloud and not at the edge. This means that in case of adversarial attacks, the\nunfavourable adversarial samples must be communicated to the cloud which leads\nto energy wastage at the edge device. Therefore, a low-power edge-friendly\nadversarial detection method is required to improve the energy efficiency of\nthe edge and robustness of the cloud-based classifier. To this end, RobustEdge\nproposes Quantization-enabled Energy Separation (QES) training with \"early\ndetection and exit\" to perform edge-based low cost adversarial detection. The\nQES-trained detector implemented at the edge blocks adversarial data\ntransmission to the classifier model, thereby improving adversarial robustness\nand energy-efficiency of the Cloud-Edge system.\n", "date": "2023-09-05", "categories": []}
{"id": "2310.07084", "title": "Investigating the Adversarial Robustness of Density Estimation Using the\n  Probability Flow ODE", "abstract": "  Beyond their impressive sampling capabilities, score-based diffusion models\noffer a powerful analysis tool in the form of unbiased density estimation of a\nquery sample under the training data distribution. In this work, we investigate\nthe robustness of density estimation using the probability flow (PF) neural\nordinary differential equation (ODE) model against gradient-based likelihood\nmaximization attacks and the relation to sample complexity, where the\ncompressed size of a sample is used as a measure of its complexity. We\nintroduce and evaluate six gradient-based log-likelihood maximization attacks,\nincluding a novel reverse integration attack. Our experimental evaluations on\nCIFAR-10 show that density estimation using the PF ODE is robust against\nhigh-complexity, high-likelihood attacks, and that in some cases adversarial\nsamples are semantically meaningful, as expected from a robust estimator.\n", "date": "2023-10-10", "categories": []}
{"id": "2310.07492", "title": "Boosting Black-box Attack to Deep Neural Networks with Conditional\n  Diffusion Models", "abstract": "  Existing black-box attacks have demonstrated promising potential in creating\nadversarial examples (AE) to deceive deep learning models. Most of these\nattacks need to handle a vast optimization space and require a large number of\nqueries, hence exhibiting limited practical impacts in real-world scenarios. In\nthis paper, we propose a novel black-box attack strategy, Conditional Diffusion\nModel Attack (CDMA), to improve the query efficiency of generating AEs under\nquery-limited situations. The key insight of CDMA is to formulate the task of\nAE synthesis as a distribution transformation problem, i.e., benign examples\nand their corresponding AEs can be regarded as coming from two distinctive\ndistributions and can transform from each other with a particular converter.\nUnlike the conventional \\textit{query-and-optimization} approach, we generate\neligible AEs with direct conditional transform using the aforementioned data\nconverter, which can significantly reduce the number of queries needed. CDMA\nadopts the conditional Denoising Diffusion Probabilistic Model as the\nconverter, which can learn the transformation from clean samples to AEs, and\nensure the smooth development of perturbed noise resistant to various defense\nstrategies. We demonstrate the effectiveness and efficiency of CDMA by\ncomparing it with nine state-of-the-art black-box attacks across three\nbenchmark datasets. On average, CDMA can reduce the query count to a handful of\ntimes; in most cases, the query count is only ONE. We also show that CDMA can\nobtain $>99\\%$ attack success rate for untarget attacks over all datasets and\ntargeted attack over CIFAR-10 with the noise budget of $\\epsilon=16$.\n", "date": "2023-10-11", "categories": []}
{"id": "2310.07632", "title": "Prompt Backdoors in Visual Prompt Learning", "abstract": "  Fine-tuning large pre-trained computer vision models is infeasible for\nresource-limited users. Visual prompt learning (VPL) has thus emerged to\nprovide an efficient and flexible alternative to model fine-tuning through\nVisual Prompt as a Service (VPPTaaS). Specifically, the VPPTaaS provider\noptimizes a visual prompt given downstream data, and downstream users can use\nthis prompt together with the large pre-trained model for prediction. However,\nthis new learning paradigm may also pose security risks when the VPPTaaS\nprovider instead provides a malicious visual prompt. In this paper, we take the\nfirst step to explore such risks through the lens of backdoor attacks.\nSpecifically, we propose BadVisualPrompt, a simple yet effective backdoor\nattack against VPL. For example, poisoning $5\\%$ CIFAR10 training data leads to\nabove $99\\%$ attack success rates with only negligible model accuracy drop by\n$1.5\\%$. In particular, we identify and then address a new technical challenge\nrelated to interactions between the backdoor trigger and visual prompt, which\ndoes not exist in conventional, model-level backdoors. Moreover, we provide\nin-depth analyses of seven backdoor defenses from model, prompt, and input\nlevels. Overall, all these defenses are either ineffective or impractical to\nmitigate our BadVisualPrompt, implying the critical vulnerability of VPL.\n", "date": "2023-10-11", "categories": []}
{"id": "2302.01961", "title": "Asymmetric Certified Robustness via Feature-Convex Neural Networks", "abstract": "  Recent works have introduced input-convex neural networks (ICNNs) as learning\nmodels with advantageous training, inference, and generalization properties\nlinked to their convex structure. In this paper, we propose a novel\nfeature-convex neural network architecture as the composition of an ICNN with a\nLipschitz feature map in order to achieve adversarial robustness. We consider\nthe asymmetric binary classification setting with one \"sensitive\" class, and\nfor this class we prove deterministic, closed-form, and easily-computable\ncertified robust radii for arbitrary $\\ell_p$-norms. We theoretically justify\nthe use of these models by characterizing their decision region geometry,\nextending the universal approximation theorem for ICNN regression to the\nclassification setting, and proving a lower bound on the probability that such\nmodels perfectly fit even unstructured uniformly distributed data in\nsufficiently high dimensions. Experiments on Malimg malware classification and\nsubsets of MNIST, Fashion-MNIST, and CIFAR-10 datasets show that feature-convex\nclassifiers attain state-of-the-art certified $\\ell_1$-radii as well as\nsubstantial $\\ell_2$- and $\\ell_{\\infty}$-radii while being far more\ncomputationally efficient than any competitive baseline.\n", "date": "2023-02-03", "categories": []}
{"id": "2306.00042", "title": "Graph-based methods coupled with specific distributional distances for\n  adversarial attack detection", "abstract": "  Artificial neural networks are prone to being fooled by carefully perturbed\ninputs which cause an egregious misclassification. These \\textit{adversarial}\nattacks have been the focus of extensive research. Likewise, there has been an\nabundance of research in ways to detect and defend against them. We introduce a\nnovel approach of detection and interpretation of adversarial attacks from a\ngraph perspective. For an input image, we compute an associated sparse graph\nusing the layer-wise relevance propagation algorithm \\cite{bach15}.\nSpecifically, we only keep edges of the neural network with the highest\nrelevance values. Three quantities are then computed from the graph which are\nthen compared against those computed from the training set. The result of the\ncomparison is a classification of the image as benign or adversarial. To make\nthe comparison, two classification methods are introduced: 1) an explicit\nformula based on Wasserstein distance applied to the degree of node and 2) a\nlogistic regression. Both classification methods produce strong results which\nlead us to believe that a graph-based interpretation of adversarial attacks is\nvaluable.\n", "date": "2023-05-31", "categories": []}
{"id": "2309.09464", "title": "Reducing Adversarial Training Cost with Gradient Approximation", "abstract": "  Deep learning models have achieved state-of-the-art performances in various\ndomains, while they are vulnerable to the inputs with well-crafted but small\nperturbations, which are named after adversarial examples (AEs). Among many\nstrategies to improve the model robustness against AEs, Projected Gradient\nDescent (PGD) based adversarial training is one of the most effective methods.\nUnfortunately, the prohibitive computational overhead of generating strong\nenough AEs, due to the maximization of the loss function, sometimes makes the\nregular PGD adversarial training impractical when using larger and more\ncomplicated models. In this paper, we propose that the adversarial loss can be\napproximated by the partial sum of Taylor series. Furthermore, we approximate\nthe gradient of adversarial loss and propose a new and efficient adversarial\ntraining method, adversarial training with gradient approximation (GAAT), to\nreduce the cost of building up robust models. Additionally, extensive\nexperiments demonstrate that this efficiency improvement can be achieved\nwithout any or with very little loss in accuracy on natural and adversarial\nexamples, which show that our proposed method saves up to 60\\% of the training\ntime with comparable model test accuracy on MNIST, CIFAR-10 and CIFAR-100\ndatasets.\n", "date": "2023-09-18", "categories": []}
{"id": "2309.10916", "title": "What Learned Representations and Influence Functions Can Tell Us About\n  Adversarial Examples", "abstract": "  Adversarial examples, deliberately crafted using small perturbations to fool\ndeep neural networks, were first studied in image processing and more recently\nin NLP. While approaches to detecting adversarial examples in NLP have largely\nrelied on search over input perturbations, image processing has seen a range of\ntechniques that aim to characterise adversarial subspaces over the learned\nrepresentations.\n  In this paper, we adapt two such approaches to NLP, one based on nearest\nneighbors and influence functions and one on Mahalanobis distances. The former\nin particular produces a state-of-the-art detector when compared against\nseveral strong baselines; moreover, the novel use of influence functions\nprovides insight into how the nature of adversarial example subspaces in NLP\nrelate to those in image processing, and also how they differ depending on the\nkind of NLP task.\n", "date": "2023-09-19", "categories": []}
{"id": "2310.05947", "title": "Robust and Efficient Interference Neural Networks for Defending Against\n  Adversarial Attacks in ImageNet", "abstract": "  The existence of adversarial images has seriously affected the task of image\nrecognition and practical application of deep learning, it is also a key\nscientific problem that deep learning urgently needs to solve. By far the most\neffective approach is to train the neural network with a large number of\nadversarial examples. However, this adversarial training method requires a huge\namount of computing resources when applied to ImageNet, and has not yet\nachieved satisfactory results for high-intensity adversarial attacks. In this\npaper, we construct an interference neural network by applying additional\nbackground images and corresponding labels, and use pre-trained ResNet-152 to\nefficiently complete the training. Compared with the state-of-the-art results\nunder the PGD attack, it has a better defense effect with much smaller\ncomputing resources. This work provides new ideas for academic research and\npractical applications of effective defense against adversarial attacks.\n", "date": "2023-09-03", "categories": []}
{"id": "2310.06227", "title": "Exploring adversarial attacks in federated learning for medical imaging", "abstract": "  Federated learning offers a privacy-preserving framework for medical image\nanalysis but exposes the system to adversarial attacks. This paper aims to\nevaluate the vulnerabilities of federated learning networks in medical image\nanalysis against such attacks. Employing domain-specific MRI tumor and\npathology imaging datasets, we assess the effectiveness of known threat\nscenarios in a federated learning environment. Our tests reveal that\ndomain-specific configurations can increase the attacker's success rate\nsignificantly. The findings emphasize the urgent need for effective defense\nmechanisms and suggest a critical re-evaluation of current security protocols\nin federated medical image analysis systems.\n", "date": "2023-10-10", "categories": []}
{"id": "2310.06396", "title": "Adversarial Robustness in Graph Neural Networks: A Hamiltonian Approach", "abstract": "  Graph neural networks (GNNs) are vulnerable to adversarial perturbations,\nincluding those that affect both node features and graph topology. This paper\ninvestigates GNNs derived from diverse neural flows, concentrating on their\nconnection to various stability notions such as BIBO stability, Lyapunov\nstability, structural stability, and conservative stability. We argue that\nLyapunov stability, despite its common use, does not necessarily ensure\nadversarial robustness. Inspired by physics principles, we advocate for the use\nof conservative Hamiltonian neural flows to construct GNNs that are robust to\nadversarial attacks. The adversarial robustness of different neural flow GNNs\nis empirically compared on several benchmark datasets under a variety of\nadversarial attacks. Extensive numerical experiments demonstrate that GNNs\nleveraging conservative Hamiltonian flows with Lyapunov stability substantially\nimprove robustness against adversarial perturbations. The implementation code\nof experiments is available at\nhttps://github.com/zknus/NeurIPS-2023-HANG-Robustness.\n", "date": "2023-10-10", "categories": []}
{"id": "2310.06468", "title": "A Geometrical Approach to Evaluate the Adversarial Robustness of Deep\n  Neural Networks", "abstract": "  Deep Neural Networks (DNNs) are widely used for computer vision tasks.\nHowever, it has been shown that deep models are vulnerable to adversarial\nattacks, i.e., their performances drop when imperceptible perturbations are\nmade to the original inputs, which may further degrade the following visual\ntasks or introduce new problems such as data and privacy security. Hence,\nmetrics for evaluating the robustness of deep models against adversarial\nattacks are desired. However, previous metrics are mainly proposed for\nevaluating the adversarial robustness of shallow networks on the small-scale\ndatasets. Although the Cross Lipschitz Extreme Value for nEtwork Robustness\n(CLEVER) metric has been proposed for large-scale datasets (e.g., the ImageNet\ndataset), it is computationally expensive and its performance relies on a\ntractable number of samples. In this paper, we propose the Adversarial\nConverging Time Score (ACTS), an attack-dependent metric that quantifies the\nadversarial robustness of a DNN on a specific input. Our key observation is\nthat local neighborhoods on a DNN's output surface would have different shapes\ngiven different inputs. Hence, given different inputs, it requires different\ntime for converging to an adversarial sample. Based on this geometry meaning,\nACTS measures the converging time as an adversarial robustness metric. We\nvalidate the effectiveness and generalization of the proposed ACTS metric\nagainst different adversarial attacks on the large-scale ImageNet dataset using\nstate-of-the-art deep networks. Extensive experiments show that our ACTS metric\nis an efficient and effective adversarial metric over the previous CLEVER\nmetric.\n", "date": "2023-10-10", "categories": []}
{"id": "2207.05225", "title": "Susceptibility of Continual Learning Against Adversarial Attacks", "abstract": "  Recent continual learning approaches have primarily focused on mitigating\ncatastrophic forgetting. Nevertheless, two critical areas have remained\nrelatively unexplored: 1) evaluating the robustness of proposed methods and 2)\nensuring the security of learned tasks. This paper investigates the\nsusceptibility of continually learned tasks, including current and previously\nacquired tasks, to adversarial attacks. Specifically, we have observed that any\nclass belonging to any task can be easily targeted and misclassified as the\ndesired target class of any other task. Such susceptibility or vulnerability of\nlearned tasks to adversarial attacks raises profound concerns regarding data\nintegrity and privacy. To assess the robustness of continual learning\napproaches, we consider continual learning approaches in all three scenarios,\ni.e., task-incremental learning, domain-incremental learning, and\nclass-incremental learning. In this regard, we explore the robustness of three\nregularization-based methods, three replay-based approaches, and one hybrid\ntechnique that combines replay and exemplar approaches. We empirically\ndemonstrated that in any setting of continual learning, any class, whether\nbelonging to the current or previously learned tasks, is susceptible to\nmisclassification. Our observations identify potential limitations of continual\nlearning approaches against adversarial attacks and highlight that current\ncontinual learning algorithms could not be suitable for deployment in\nreal-world settings.\n", "date": "2022-07-11", "categories": []}
{"id": "2208.03835", "title": "On Transfer of Adversarial Robustness from Pretraining to Downstream\n  Tasks", "abstract": "  As large-scale training regimes have gained popularity, the use of pretrained\nmodels for downstream tasks has become common practice in machine learning.\nWhile pretraining has been shown to enhance the performance of models in\npractice, the transfer of robustness properties from pretraining to downstream\ntasks remains poorly understood. In this study, we demonstrate that the\nrobustness of a linear predictor on downstream tasks can be constrained by the\nrobustness of its underlying representation, regardless of the protocol used\nfor pretraining. We prove (i) a bound on the loss that holds independent of any\ndownstream task, as well as (ii) a criterion for robust classification in\nparticular. We validate our theoretical results in practical applications, show\nhow our results can be used for calibrating expectations of downstream\nrobustness, and when our results are useful for optimal transfer learning.\nTaken together, our results offer an initial step towards characterizing the\nrequirements of the representation function for reliable post-adaptation\nperformance.\n", "date": "2022-08-07", "categories": []}
{"id": "2211.16044", "title": "Model Extraction Attack against Self-supervised Speech Models", "abstract": "  Self-supervised learning (SSL) speech models generate meaningful\nrepresentations of given clips and achieve incredible performance across\nvarious downstream tasks. Model extraction attack (MEA) often refers to an\nadversary stealing the functionality of the victim model with only query\naccess. In this work, we study the MEA problem against SSL speech model with a\nsmall number of queries. We propose a two-stage framework to extract the model.\nIn the first stage, SSL is conducted on the large-scale unlabeled corpus to\npre-train a small speech model. Secondly, we actively sample a small portion of\nclips from the unlabeled corpus and query the target model with these clips to\nacquire their representations as labels for the small model's second-stage\ntraining. Experiment results show that our sampling methods can effectively\nextract the target model without knowing any information about its model\narchitecture.\n", "date": "2022-11-29", "categories": []}
{"id": "2310.05755", "title": "Deep Concept Removal", "abstract": "  We address the problem of concept removal in deep neural networks, aiming to\nlearn representations that do not encode certain specified concepts (e.g.,\ngender etc.) We propose a novel method based on adversarial linear classifiers\ntrained on a concept dataset, which helps to remove the targeted attribute\nwhile maintaining model performance. Our approach Deep Concept Removal\nincorporates adversarial probing classifiers at various layers of the network,\neffectively addressing concept entanglement and improving out-of-distribution\ngeneralization. We also introduce an implicit gradient-based technique to\ntackle the challenges associated with adversarial training using linear\nclassifiers. We evaluate the ability to remove a concept on a set of popular\ndistributionally robust optimization (DRO) benchmarks with spurious\ncorrelations, as well as out-of-distribution (OOD) generalization tasks.\n", "date": "2023-10-09", "categories": []}
{"id": "2211.04205", "title": "Preserving Semantics in Textual Adversarial Attacks", "abstract": "  The growth of hateful online content, or hate speech, has been associated\nwith a global increase in violent crimes against minorities [23]. Harmful\nonline content can be produced easily, automatically and anonymously. Even\nthough, some form of auto-detection is already achieved through text\nclassifiers in NLP, they can be fooled by adversarial attacks. To strengthen\nexisting systems and stay ahead of attackers, we need better adversarial\nattacks. In this paper, we show that up to 70% of adversarial examples\ngenerated by adversarial attacks should be discarded because they do not\npreserve semantics. We address this core weakness and propose a new, fully\nsupervised sentence embedding technique called Semantics-Preserving-Encoder\n(SPE). Our method outperforms existing sentence encoders used in adversarial\nattacks by achieving 1.2x - 5.1x better real attack success rate. We release\nour code as a plugin that can be used in any existing adversarial attack to\nimprove its quality and speed up its execution.\n", "date": "2022-11-08", "categories": []}
{"id": "2303.01338", "title": "AdvRain: Adversarial Raindrops to Attack Camera-based Smart Vision\n  Systems", "abstract": "  Vision-based perception modules are increasingly deployed in many\napplications, especially autonomous vehicles and intelligent robots. These\nmodules are being used to acquire information about the surroundings and\nidentify obstacles. Hence, accurate detection and classification are essential\nto reach appropriate decisions and take appropriate and safe actions at all\ntimes. Current studies have demonstrated that \"printed adversarial attacks\",\nknown as physical adversarial attacks, can successfully mislead perception\nmodels such as object detectors and image classifiers. However, most of these\nphysical attacks are based on noticeable and eye-catching patterns for\ngenerated perturbations making them identifiable/detectable by human eye or in\ntest drives. In this paper, we propose a camera-based inconspicuous adversarial\nattack (\\textbf{AdvRain}) capable of fooling camera-based perception systems\nover all objects of the same class. Unlike mask based fake-weather attacks that\nrequire access to the underlying computing hardware or image memory, our attack\nis based on emulating the effects of a natural weather condition (i.e.,\nRaindrops) that can be printed on a translucent sticker, which is externally\nplaced over the lens of a camera. To accomplish this, we provide an iterative\nprocess based on performing a random search aiming to identify critical\npositions to make sure that the performed transformation is adversarial for a\ntarget classifier. Our transformation is based on blurring predefined parts of\nthe captured image corresponding to the areas covered by the raindrop. We\nachieve a drop in average model accuracy of more than $45\\%$ and $40\\%$ on\nVGG19 for ImageNet and Resnet34 for Caltech-101, respectively, using only $20$\nraindrops.\n", "date": "2023-03-02", "categories": []}
{"id": "2310.03185", "title": "Misusing Tools in Large Language Models With Visual Adversarial Examples", "abstract": "  Large Language Models (LLMs) are being enhanced with the ability to use tools\nand to process multiple modalities. These new capabilities bring new benefits\nand also new security risks. In this work, we show that an attacker can use\nvisual adversarial examples to cause attacker-desired tool usage. For example,\nthe attacker could cause a victim LLM to delete calendar events, leak private\nconversations and book hotels. Different from prior work, our attacks can\naffect the confidentiality and integrity of user resources connected to the LLM\nwhile being stealthy and generalizable to multiple input prompts. We construct\nthese attacks using gradient-based adversarial training and characterize\nperformance along multiple dimensions. We find that our adversarial images can\nmanipulate the LLM to invoke tools following real-world syntax almost always\n(~98%) while maintaining high similarity to clean images (~0.9 SSIM).\nFurthermore, using human scoring and automated metrics, we find that the\nattacks do not noticeably affect the conversation (and its semantics) between\nthe user and the LLM.\n", "date": "2023-10-04", "categories": []}
{"id": "2310.03349", "title": "An Integrated Algorithm for Robust and Imperceptible Audio Adversarial\n  Examples", "abstract": "  Audio adversarial examples are audio files that have been manipulated to fool\nan automatic speech recognition (ASR) system, while still sounding benign to a\nhuman listener. Most methods to generate such samples are based on a two-step\nalgorithm: first, a viable adversarial audio file is produced, then, this is\nfine-tuned with respect to perceptibility and robustness. In this work, we\npresent an integrated algorithm that uses psychoacoustic models and room\nimpulse responses (RIR) in the generation step. The RIRs are dynamically\ncreated by a neural network during the generation process to simulate a\nphysical environment to harden our examples against transformations experienced\nin over-the-air attacks. We compare the different approaches in three\nexperiments: in a simulated environment and in a realistic over-the-air\nscenario to evaluate the robustness, and in a human study to evaluate the\nperceptibility. Our algorithms considering psychoacoustics only or in addition\nto the robustness show an improvement in the signal-to-noise ratio (SNR) as\nwell as in the human perception study, at the cost of an increased word error\nrate (WER).\n", "date": "2023-10-05", "categories": []}
{"id": "2310.03578", "title": "Targeted Adversarial Attacks on Generalizable Neural Radiance Fields", "abstract": "  Neural Radiance Fields (NeRFs) have recently emerged as a powerful tool for\n3D scene representation and rendering. These data-driven models can learn to\nsynthesize high-quality images from sparse 2D observations, enabling realistic\nand interactive scene reconstructions. However, the growing usage of NeRFs in\ncritical applications such as augmented reality, robotics, and virtual\nenvironments could be threatened by adversarial attacks.\n  In this paper we present how generalizable NeRFs can be attacked by both\nlow-intensity adversarial attacks and adversarial patches, where the later\ncould be robust enough to be used in real world applications. We also\ndemonstrate targeted attacks, where a specific, predefined output scene is\ngenerated by these attack with success.\n", "date": "2023-10-05", "categories": []}
{"id": "2310.03614", "title": "Adversarial Machine Learning for Social Good: Reframing the Adversary as\n  an Ally", "abstract": "  Deep Neural Networks (DNNs) have been the driving force behind many of the\nrecent advances in machine learning. However, research has shown that DNNs are\nvulnerable to adversarial examples -- input samples that have been perturbed to\nforce DNN-based models to make errors. As a result, Adversarial Machine\nLearning (AdvML) has gained a lot of attention, and researchers have\ninvestigated these vulnerabilities in various settings and modalities. In\naddition, DNNs have also been found to incorporate embedded bias and often\nproduce unexplainable predictions, which can result in anti-social AI\napplications. The emergence of new AI technologies that leverage Large Language\nModels (LLMs), such as ChatGPT and GPT-4, increases the risk of producing\nanti-social applications at scale. AdvML for Social Good (AdvML4G) is an\nemerging field that repurposes the AdvML bug to invent pro-social applications.\nRegulators, practitioners, and researchers should collaborate to encourage the\ndevelopment of pro-social applications and hinder the development of\nanti-social ones. In this work, we provide the first comprehensive review of\nthe emerging field of AdvML4G. This paper encompasses a taxonomy that\nhighlights the emergence of AdvML4G, a discussion of the differences and\nsimilarities between AdvML4G and AdvML, a taxonomy covering social good-related\nconcepts and aspects, an exploration of the motivations behind the emergence of\nAdvML4G at the intersection of ML4G and AdvML, and an extensive summary of the\nworks that utilize AdvML4G as an auxiliary tool for innovating pro-social\napplications. Finally, we elaborate upon various challenges and open research\nissues that require significant attention from the research community.\n", "date": "2023-10-05", "categories": []}
{"id": "2310.03664", "title": "Certification of Deep Learning Models for Medical Image Segmentation", "abstract": "  In medical imaging, segmentation models have known a significant improvement\nin the past decade and are now used daily in clinical practice. However,\nsimilar to classification models, segmentation models are affected by\nadversarial attacks. In a safety-critical field like healthcare, certifying\nmodel predictions is of the utmost importance. Randomized smoothing has been\nintroduced lately and provides a framework to certify models and obtain\ntheoretical guarantees. In this paper, we present for the first time a\ncertified segmentation baseline for medical imaging based on randomized\nsmoothing and diffusion models. Our results show that leveraging the power of\ndenoising diffusion probabilistic models helps us overcome the limits of\nrandomized smoothing. We conduct extensive experiments on five public datasets\nof chest X-rays, skin lesions, and colonoscopies, and empirically show that we\nare able to maintain high certified Dice scores even for highly perturbed\nimages. Our work represents the first attempt to certify medical image\nsegmentation models, and we aspire for it to set a foundation for future\nbenchmarks in this crucial and largely uncharted area.\n", "date": "2023-10-05", "categories": []}
{"id": "2310.03707", "title": "OMG-ATTACK: Self-Supervised On-Manifold Generation of Transferable\n  Evasion Attacks", "abstract": "  Evasion Attacks (EA) are used to test the robustness of trained neural\nnetworks by distorting input data to misguide the model into incorrect\nclassifications. Creating these attacks is a challenging task, especially with\nthe ever-increasing complexity of models and datasets. In this work, we\nintroduce a self-supervised, computationally economical method for generating\nadversarial examples, designed for the unseen black-box setting. Adapting\ntechniques from representation learning, our method generates on-manifold EAs\nthat are encouraged to resemble the data distribution. These attacks are\ncomparable in effectiveness compared to the state-of-the-art when attacking the\nmodel trained on, but are significantly more effective when attacking unseen\nmodels, as the attacks are more related to the data rather than the model\nitself. Our experiments consistently demonstrate the method is effective across\nvarious models, unseen data categories, and even defended models, suggesting a\nsignificant role for on-manifold EAs when targeting unseen models.\n", "date": "2023-10-05", "categories": []}
{"id": "2308.06887", "title": "Robustified ANNs Reveal Wormholes Between Human Category Percepts", "abstract": "  The visual object category reports of artificial neural networks (ANNs) are\nnotoriously sensitive to tiny, adversarial image perturbations. Because human\ncategory reports (aka human percepts) are thought to be insensitive to those\nsame small-norm perturbations -- and locally stable in general -- this argues\nthat ANNs are incomplete scientific models of human visual perception.\nConsistent with this, we show that when small-norm image perturbations are\ngenerated by standard ANN models, human object category percepts are indeed\nhighly stable. However, in this very same \"human-presumed-stable\" regime, we\nfind that robustified ANNs reliably discover low-norm image perturbations that\nstrongly disrupt human percepts. These previously undetectable human perceptual\ndisruptions are massive in amplitude, approaching the same level of sensitivity\nseen in robustified ANNs. Further, we show that robustified ANNs support\nprecise perceptual state interventions: they guide the construction of low-norm\nimage perturbations that strongly alter human category percepts toward specific\nprescribed percepts. These observations suggest that for arbitrary starting\npoints in image space, there exists a set of nearby \"wormholes\", each leading\nthe subject from their current category perceptual state into a semantically\nvery different state. Moreover, contemporary ANN models of biological visual\nprocessing are now accurate enough to consistently guide us to those portals.\n", "date": "2023-08-14", "categories": []}
{"id": "2310.02480", "title": "Splitting the Difference on Adversarial Training", "abstract": "  The existence of adversarial examples points to a basic weakness of deep\nneural networks. One of the most effective defenses against such examples,\nadversarial training, entails training models with some degree of robustness,\nusually at the expense of a degraded natural accuracy. Most adversarial\ntraining methods aim to learn a model that finds, for each class, a common\ndecision boundary encompassing both the clean and perturbed examples. In this\nwork, we take a fundamentally different approach by treating the perturbed\nexamples of each class as a separate class to be learned, effectively splitting\neach class into two classes: \"clean\" and \"adversarial.\" This split doubles the\nnumber of classes to be learned, but at the same time considerably simplifies\nthe decision boundaries. We provide a theoretical plausibility argument that\nsheds some light on the conditions under which our approach can be expected to\nbe beneficial. Likewise, we empirically demonstrate that our method learns\nrobust models while attaining optimal or near-optimal natural accuracy, e.g.,\non CIFAR-10 we obtain near-optimal natural accuracy of $95.01\\%$ alongside\nsignificant robustness across multiple tasks. The ability to achieve such\nnear-optimal natural accuracy, while maintaining a significant level of\nrobustness, makes our method applicable to real-world applications where\nnatural accuracy is at a premium. As a whole, our main contribution is a\ngeneral method that confers a significant level of robustness upon classifiers\nwith only minor or negligible degradation of their natural accuracy.\n", "date": "2023-10-03", "categories": []}
{"id": "2310.02544", "title": "SlowFormer: Universal Adversarial Patch for Attack on Compute and Energy\n  Efficiency of Inference Efficient Vision Transformers", "abstract": "  Recently, there has been a lot of progress in reducing the computation of\ndeep models at inference time. These methods can reduce both the computational\nneeds and power usage of deep models. Some of these approaches adaptively scale\nthe compute based on the input instance. We show that such models can be\nvulnerable to a universal adversarial patch attack, where the attacker\noptimizes for a patch that when pasted on any image, can increase the compute\nand power consumption of the model. We run experiments with three different\nefficient vision transformer methods showing that in some cases, the attacker\ncan increase the computation to the maximum possible level by simply pasting a\npatch that occupies only 8\\% of the image area. We also show that a standard\nadversarial training defense method can reduce some of the attack's success. We\nbelieve adaptive efficient methods will be necessary for the future to lower\nthe power usage of deep models, so we hope our paper encourages the community\nto study the robustness of these methods and develop better defense methods for\nthe proposed attack.\n", "date": "2023-10-04", "categories": []}
{"id": "2303.13372", "title": "DRSM: De-Randomized Smoothing on Malware Classifier Providing Certified\n  Robustness", "abstract": "  Machine Learning (ML) models have been utilized for malware detection for\nover two decades. Consequently, this ignited an ongoing arms race between\nmalware authors and antivirus systems, compelling researchers to propose\ndefenses for malware-detection models against evasion attacks. However, most if\nnot all existing defenses against evasion attacks suffer from sizable\nperformance degradation and/or can defend against only specific attacks, which\nmakes them less practical in real-world settings. In this work, we develop a\ncertified defense, DRSM (De-Randomized Smoothed MalConv), by redesigning the\nde-randomized smoothing technique for the domain of malware detection.\nSpecifically, we propose a window ablation scheme to provably limit the impact\nof adversarial bytes while maximally preserving local structures of the\nexecutables. After showing how DRSM is theoretically robust against attacks\nwith contiguous adversarial bytes, we verify its performance and certified\nrobustness experimentally, where we observe only marginal accuracy drops as the\ncost of robustness. To our knowledge, we are the first to offer certified\nrobustness in the realm of static detection of malware executables. More\nsurprisingly, through evaluating DRSM against 9 empirical attacks of different\ntypes, we observe that the proposed defense is empirically robust to some\nextent against a diverse set of attacks, some of which even fall out of the\nscope of its original threat model. In addition, we collected 15.5K recent\nbenign raw executables from diverse sources, which will be made public as a\ndataset called PACE (Publicly Accessible Collection(s) of Executables) to\nalleviate the scarcity of publicly available benign datasets for studying\nmalware detection and provide future research with more representative data of\nthe time.\n", "date": "2023-03-20", "categories": []}
{"id": "2305.09241", "title": "Unlearnable Examples Give a False Sense of Security: Piercing through\n  Unexploitable Data with Learnable Examples", "abstract": "  Safeguarding data from unauthorized exploitation is vital for privacy and\nsecurity, especially in recent rampant research in security breach such as\nadversarial/membership attacks. To this end, \\textit{unlearnable examples}\n(UEs) have been recently proposed as a compelling protection, by adding\nimperceptible perturbation to data so that models trained on them cannot\nclassify them accurately on original clean distribution. Unfortunately, we find\nUEs provide a false sense of security, because they cannot stop unauthorized\nusers from utilizing other unprotected data to remove the protection, by\nturning unlearnable data into learnable again. Motivated by this observation,\nwe formally define a new threat by introducing \\textit{learnable unauthorized\nexamples} (LEs) which are UEs with their protection removed. The core of this\napproach is a novel purification process that projects UEs onto the manifold of\nLEs. This is realized by a new joint-conditional diffusion model which denoises\nUEs conditioned on the pixel and perceptual similarity between UEs and LEs.\nExtensive experiments demonstrate that LE delivers state-of-the-art countering\nperformance against both supervised UEs and unsupervised UEs in various\nscenarios, which is the first generalizable countermeasure to UEs across\nsupervised learning and unsupervised learning. Our code is available at\n\\url{https://github.com/jiangw-0/LE_JCDP}.\n", "date": "2023-05-16", "categories": []}
{"id": "2307.10490", "title": "Abusing Images and Sounds for Indirect Instruction Injection in\n  Multi-Modal LLMs", "abstract": "  We demonstrate how images and sounds can be used for indirect prompt and\ninstruction injection in multi-modal LLMs. An attacker generates an adversarial\nperturbation corresponding to the prompt and blends it into an image or audio\nrecording. When the user asks the (unmodified, benign) model about the\nperturbed image or audio, the perturbation steers the model to output the\nattacker-chosen text and/or make the subsequent dialog follow the attacker's\ninstruction. We illustrate this attack with several proof-of-concept examples\ntargeting LLaVa and PandaGPT.\n", "date": "2023-07-19", "categories": []}
{"id": "2310.01537", "title": "Adversarial Client Detection via Non-parametric Subspace Monitoring in\n  the Internet of Federated Things", "abstract": "  The Internet of Federated Things (IoFT) represents a network of\ninterconnected systems with federated learning as the backbone, facilitating\ncollaborative knowledge acquisition while ensuring data privacy for individual\nsystems. The wide adoption of IoFT, however, is hindered by security concerns,\nparticularly the susceptibility of federated learning networks to adversarial\nattacks. In this paper, we propose an effective non-parametric approach FedRR,\nwhich leverages the low-rank features of the transmitted parameter updates\ngenerated by federated learning to address the adversarial attack problem.\nBesides, our proposed method is capable of accurately detecting adversarial\nclients and controlling the false alarm rate under the scenario with no attack\noccurring. Experiments based on digit recognition using the MNIST datasets\nvalidated the advantages of our approach.\n", "date": "2023-10-02", "categories": []}
{"id": "2310.01818", "title": "AutoLoRa: A Parameter-Free Automated Robust Fine-Tuning Framework", "abstract": "  Robust Fine-Tuning (RFT) is a low-cost strategy to obtain adversarial\nrobustness in downstream applications, without requiring a lot of computational\nresources and collecting significant amounts of data. This paper uncovers an\nissue with the existing RFT, where optimizing both adversarial and natural\nobjectives through the feature extractor (FE) yields significantly divergent\ngradient directions. This divergence introduces instability in the optimization\nprocess, thereby hindering the attainment of adversarial robustness and\nrendering RFT highly sensitive to hyperparameters. To mitigate this issue, we\npropose a low-rank (LoRa) branch that disentangles RFT into two distinct\ncomponents: optimizing natural objectives via the LoRa branch and adversarial\nobjectives via the FE. Besides, we introduce heuristic strategies for\nautomating the scheduling of the learning rate and the scalars of loss terms.\nExtensive empirical evaluations demonstrate that our proposed automated RFT\ndisentangled via the LoRa branch (AutoLoRa) achieves new state-of-the-art\nresults across a range of downstream tasks. AutoLoRa holds significant\npractical utility, as it automatically converts a pre-trained FE into an\nadversarially robust model for downstream tasks without the need for searching\nhyperparameters.\n", "date": "2023-10-03", "categories": []}
{"id": "2301.02615", "title": "Silent Killer: A Stealthy, Clean-Label, Black-Box Backdoor Attack", "abstract": "  Backdoor poisoning attacks pose a well-known risk to neural networks.\nHowever, most studies have focused on lenient threat models. We introduce\nSilent Killer, a novel attack that operates in clean-label, black-box settings,\nuses a stealthy poison and trigger and outperforms existing methods. We\ninvestigate the use of universal adversarial perturbations as triggers in\nclean-label attacks, following the success of such approaches under\npoison-label settings. We analyze the success of a naive adaptation and find\nthat gradient alignment for crafting the poison is required to ensure high\nsuccess rates. We conduct thorough experiments on MNIST, CIFAR10, and a reduced\nversion of ImageNet and achieve state-of-the-art results.\n", "date": "2023-01-05", "categories": []}
{"id": "2308.05320", "title": "Generating Transferable and Stealthy Adversarial Patch via\n  Attention-guided Adversarial Inpainting", "abstract": "  Adversarial patch attacks can fool the face recognition (FR) models via small\npatches. However, previous adversarial patch attacks often result in unnatural\npatterns that are easily noticeable. Generating transferable and stealthy\nadversarial patches that can efficiently deceive the black-box FR models while\nhaving good camouflage is challenging because of the huge stylistic difference\nbetween the source and target images. To generate transferable,\nnatural-looking, and stealthy adversarial patches, we propose an innovative\ntwo-stage attack called Adv-Inpainting, which extracts style features and\nidentity features from the attacker and target faces, respectively and then\nfills the patches with misleading and inconspicuous content guided by attention\nmaps. In the first stage, we extract multi-scale style embeddings by a\npyramid-like network and identity embeddings by a pretrained FR model and\npropose a novel Attention-guided Adaptive Instance Normalization layer (AAIN)\nto merge them via background-patch cross-attention maps. The proposed layer can\nadaptively fuse identity and style embeddings by fully exploiting priority\ncontextual information. In the second stage, we design an Adversarial Patch\nRefinement Network (APR-Net) with a novel boundary variance loss, a spatial\ndiscounted reconstruction loss, and a perceptual loss to boost the stealthiness\nfurther. Experiments demonstrate that our attack can generate adversarial\npatches with improved visual quality, better stealthiness, and stronger\ntransferability than state-of-the-art adversarial patch attacks and semantic\nattacks.\n", "date": "2023-08-10", "categories": []}
{"id": "2310.00438", "title": "Human-Producible Adversarial Examples", "abstract": "  Visual adversarial examples have so far been restricted to pixel-level image\nmanipulations in the digital world, or have required sophisticated equipment\nsuch as 2D or 3D printers to be produced in the physical real world. We present\nthe first ever method of generating human-producible adversarial examples for\nthe real world that requires nothing more complicated than a marker pen. We\ncall them $\\textbf{adversarial tags}$. First, building on top of differential\nrendering, we demonstrate that it is possible to build potent adversarial\nexamples with just lines. We find that by drawing just $4$ lines we can disrupt\na YOLO-based model in $54.8\\%$ of cases; increasing this to $9$ lines disrupts\n$81.8\\%$ of the cases tested. Next, we devise an improved method for line\nplacement to be invariant to human drawing error. We evaluate our system\nthoroughly in both digital and analogue worlds and demonstrate that our tags\ncan be applied by untrained humans. We demonstrate the effectiveness of our\nmethod for producing real-world adversarial examples by conducting a user study\nwhere participants were asked to draw over printed images using digital\nequivalents as guides. We further evaluate the effectiveness of both targeted\nand untargeted attacks, and discuss various trade-offs and method limitations,\nas well as the practical and ethical implications of our work. The source code\nwill be released publicly.\n", "date": "2023-09-30", "categories": []}
{"id": "2310.00567", "title": "Understanding the Robustness of Randomized Feature Defense Against\n  Query-Based Adversarial Attacks", "abstract": "  Recent works have shown that deep neural networks are vulnerable to\nadversarial examples that find samples close to the original image but can make\nthe model misclassify. Even with access only to the model's output, an attacker\ncan employ black-box attacks to generate such adversarial examples. In this\nwork, we propose a simple and lightweight defense against black-box attacks by\nadding random noise to hidden features at intermediate layers of the model at\ninference time. Our theoretical analysis confirms that this method effectively\nenhances the model's resilience against both score-based and decision-based\nblack-box attacks. Importantly, our defense does not necessitate adversarial\ntraining and has minimal impact on accuracy, rendering it applicable to any\npre-trained model. Our analysis also reveals the significance of selectively\nadding noise to different parts of the model based on the gradient of the\nadversarial objective function, which can be varied during the attack. We\ndemonstrate the robustness of our defense against multiple black-box attacks\nthrough extensive empirical experiments involving diverse models with various\narchitectures.\n", "date": "2023-10-01", "categories": []}
{"id": "2310.00626", "title": "GhostEncoder: Stealthy Backdoor Attacks with Dynamic Triggers to\n  Pre-trained Encoders in Self-supervised Learning", "abstract": "  Within the realm of computer vision, self-supervised learning (SSL) pertains\nto training pre-trained image encoders utilizing a substantial quantity of\nunlabeled images. Pre-trained image encoders can serve as feature extractors,\nfacilitating the construction of downstream classifiers for various tasks.\nHowever, the use of SSL has led to an increase in security research related to\nvarious backdoor attacks. Currently, the trigger patterns used in backdoor\nattacks on SSL are mostly visible or static (sample-agnostic), making backdoors\nless covert and significantly affecting the attack performance. In this work,\nwe propose GhostEncoder, the first dynamic invisible backdoor attack on SSL.\nUnlike existing backdoor attacks on SSL, which use visible or static trigger\npatterns, GhostEncoder utilizes image steganography techniques to encode hidden\ninformation into benign images and generate backdoor samples. We then fine-tune\nthe pre-trained image encoder on a manipulation dataset to inject the backdoor,\nenabling downstream classifiers built upon the backdoored encoder to inherit\nthe backdoor behavior for target downstream tasks. We evaluate GhostEncoder on\nthree downstream tasks and results demonstrate that GhostEncoder provides\npractical stealthiness on images and deceives the victim model with a high\nattack success rate without compromising its utility. Furthermore, GhostEncoder\nwithstands state-of-the-art defenses, including STRIP, STRIP-Cl, and\nSSL-Cleanse.\n", "date": "2023-10-01", "categories": []}
{"id": "2310.00633", "title": "A Survey of Robustness and Safety of 2D and 3D Deep Learning Models\n  Against Adversarial Attacks", "abstract": "  Benefiting from the rapid development of deep learning, 2D and 3D computer\nvision applications are deployed in many safe-critical systems, such as\nautopilot and identity authentication. However, deep learning models are not\ntrustworthy enough because of their limited robustness against adversarial\nattacks. The physically realizable adversarial attacks further pose fatal\nthreats to the application and human safety. Lots of papers have emerged to\ninvestigate the robustness and safety of deep learning models against\nadversarial attacks. To lead to trustworthy AI, we first construct a general\nthreat model from different perspectives and then comprehensively review the\nlatest progress of both 2D and 3D adversarial attacks. We extend the concept of\nadversarial examples beyond imperceptive perturbations and collate over 170\npapers to give an overview of deep learning model robustness against various\nadversarial attacks. To the best of our knowledge, we are the first to\nsystematically investigate adversarial attacks for 3D models, a flourishing\nfield applied to many real-world applications. In addition, we examine physical\nadversarial attacks that lead to safety violations. Last but not least, we\nsummarize present popular topics, give insights on challenges, and shed light\non future research on trustworthy AI.\n", "date": "2023-10-01", "categories": []}
{"id": "2310.00761", "title": "Counterfactual Image Generation for adversarially robust and\n  interpretable Classifiers", "abstract": "  Neural Image Classifiers are effective but inherently hard to interpret and\nsusceptible to adversarial attacks. Solutions to both problems exist, among\nothers, in the form of counterfactual examples generation to enhance\nexplainability or adversarially augment training datasets for improved\nrobustness. However, existing methods exclusively address only one of the\nissues. We propose a unified framework leveraging image-to-image translation\nGenerative Adversarial Networks (GANs) to produce counterfactual samples that\nhighlight salient regions for interpretability and act as adversarial samples\nto augment the dataset for more robustness. This is achieved by combining the\nclassifier and discriminator into a single model that attributes real images to\ntheir respective classes and flags generated images as \"fake\". We assess the\nmethod's effectiveness by evaluating (i) the produced explainability masks on a\nsemantic segmentation task for concrete cracks and (ii) the model's resilience\nagainst the Projected Gradient Descent (PGD) attack on a fruit defects\ndetection problem. Our produced saliency maps are highly descriptive, achieving\ncompetitive IoU values compared to classical segmentation models despite being\ntrained exclusively on classification labels. Furthermore, the model exhibits\nimproved robustness to adversarial attacks, and we show how the discriminator's\n\"fakeness\" value serves as an uncertainty measure of the predictions.\n", "date": "2023-10-01", "categories": []}
{"id": "2309.16706", "title": "AIR: Threats of Adversarial Attacks on Deep Learning-Based Information\n  Recovery", "abstract": "  A wireless communications system usually consists of a transmitter which\ntransmits the information and a receiver which recovers the original\ninformation from the received distorted signal. Deep learning (DL) has been\nused to improve the performance of the receiver in complicated channel\nenvironments and state-of-the-art (SOTA) performance has been achieved.\nHowever, its robustness has not been investigated. In order to evaluate the\nrobustness of DL-based information recovery models under adversarial\ncircumstances, we investigate adversarial attacks on the SOTA DL-based\ninformation recovery model, i.e., DeepReceiver. We formulate the problem as an\noptimization problem with power and peak-to-average power ratio (PAPR)\nconstraints. We design different adversarial attack methods according to the\nadversary's knowledge of DeepReceiver's model and/or testing samples. Extensive\nexperiments show that the DeepReceiver is vulnerable to the designed attack\nmethods in all of the considered scenarios. Even in the scenario of both model\nand test sample restricted, the adversary can attack the DeepReceiver and\nincrease its bit error rate (BER) above 10%. It can also be found that the\nDeepReceiver is vulnerable to adversarial perturbations even with very low\npower and limited PAPR. These results suggest that defense measures should be\ntaken to enhance the robustness of DeepReceiver.\n", "date": "2023-08-17", "categories": []}
{"id": "2309.16878", "title": "Investigating Human-Identifiable Features Hidden in Adversarial\n  Perturbations", "abstract": "  Neural networks perform exceedingly well across various machine learning\ntasks but are not immune to adversarial perturbations. This vulnerability has\nimplications for real-world applications. While much research has been\nconducted, the underlying reasons why neural networks fall prey to adversarial\nattacks are not yet fully understood. Central to our study, which explores up\nto five attack algorithms across three datasets, is the identification of\nhuman-identifiable features in adversarial perturbations. Additionally, we\nuncover two distinct effects manifesting within human-identifiable features.\nSpecifically, the masking effect is prominent in untargeted attacks, while the\ngeneration effect is more common in targeted attacks. Using pixel-level\nannotations, we extract such features and demonstrate their ability to\ncompromise target models. In addition, our findings indicate a notable extent\nof similarity in perturbations across different attack algorithms when averaged\nover multiple models. This work also provides insights into phenomena\nassociated with adversarial perturbations, such as transferability and model\ninterpretability. Our study contributes to a deeper understanding of the\nunderlying mechanisms behind adversarial attacks and offers insights for the\ndevelopment of more resilient defense strategies for neural networks.\n", "date": "2023-09-28", "categories": []}
{"id": "2309.16207", "title": "Parameter-Saving Adversarial Training: Reinforcing Multi-Perturbation\n  Robustness via Hypernetworks", "abstract": "  Adversarial training serves as one of the most popular and effective methods\nto defend against adversarial perturbations. However, most defense mechanisms\nonly consider a single type of perturbation while various attack methods might\nbe adopted to perform stronger adversarial attacks against the deployed model\nin real-world scenarios, e.g., $\\ell_2$ or $\\ell_\\infty$. Defending against\nvarious attacks can be a challenging problem since multi-perturbation\nadversarial training and its variants only achieve suboptimal robustness\ntrade-offs, due to the theoretical limit to multi-perturbation robustness for a\nsingle model. Besides, it is impractical to deploy large models in some\nstorage-efficient scenarios. To settle down these drawbacks, in this paper we\npropose a novel multi-perturbation adversarial training framework,\nparameter-saving adversarial training (PSAT), to reinforce multi-perturbation\nrobustness with an advantageous side effect of saving parameters, which\nleverages hypernetworks to train specialized models against a single\nperturbation and aggregate these specialized models to defend against multiple\nperturbations. Eventually, we extensively evaluate and compare our proposed\nmethod with state-of-the-art single/multi-perturbation robust methods against\nvarious latest attack methods on different datasets, showing the robustness\nsuperiority and parameter efficiency of our proposed method, e.g., for the\nCIFAR-10 dataset with ResNet-50 as the backbone, PSAT saves approximately 80\\%\nof parameters with achieving the state-of-the-art robustness trade-off\naccuracy.\n", "date": "2023-09-28", "categories": []}
{"id": "2309.16577", "title": "Compilation as a Defense: Enhancing DL Model Attack Robustness via\n  Tensor Optimization", "abstract": "  Adversarial Machine Learning (AML) is a rapidly growing field of security\nresearch, with an often overlooked area being model attacks through\nside-channels. Previous works show such attacks to be serious threats, though\nlittle progress has been made on efficient remediation strategies that avoid\ncostly model re-engineering. This work demonstrates a new defense against AML\nside-channel attacks using model compilation techniques, namely tensor\noptimization. We show relative model attack effectiveness decreases of up to\n43% using tensor optimization, discuss the implications, and direction of\nfuture work.\n", "date": "2023-09-20", "categories": []}
{"id": "2303.06280", "title": "Stateful Defenses for Machine Learning Models Are Not Yet Secure Against\n  Black-box Attacks", "abstract": "  Recent work has proposed stateful defense models (SDMs) as a compelling\nstrategy to defend against a black-box attacker who only has query access to\nthe model, as is common for online machine learning platforms. Such stateful\ndefenses aim to defend against black-box attacks by tracking the query history\nand detecting and rejecting queries that are \"similar\" and thus preventing\nblack-box attacks from finding useful gradients and making progress towards\nfinding adversarial attacks within a reasonable query budget. Recent SDMs\n(e.g., Blacklight and PIHA) have shown remarkable success in defending against\nstate-of-the-art black-box attacks. In this paper, we show that SDMs are highly\nvulnerable to a new class of adaptive black-box attacks. We propose a novel\nadaptive black-box attack strategy called Oracle-guided Adaptive Rejection\nSampling (OARS) that involves two stages: (1) use initial query patterns to\ninfer key properties about an SDM's defense; and, (2) leverage those extracted\nproperties to design subsequent query patterns to evade the SDM's defense while\nmaking progress towards finding adversarial inputs. OARS is broadly applicable\nas an enhancement to existing black-box attacks - we show how to apply the\nstrategy to enhance six common black-box attacks to be more effective against\ncurrent class of SDMs. For example, OARS-enhanced versions of black-box attacks\nimproved attack success rate against recent stateful defenses from almost 0% to\nto almost 100% for multiple datasets within reasonable query budgets.\n", "date": "2023-03-11", "categories": []}
{"id": "2309.14615", "title": "Gray-box Adversarial Attack of Deep Reinforcement Learning-based Trading\n  Agents", "abstract": "  In recent years, deep reinforcement learning (Deep RL) has been successfully\nimplemented as a smart agent in many systems such as complex games,\nself-driving cars, and chat-bots. One of the interesting use cases of Deep RL\nis its application as an automated stock trading agent. In general, any\nautomated trading agent is prone to manipulations by adversaries in the trading\nenvironment. Thus studying their robustness is vital for their success in\npractice. However, typical mechanism to study RL robustness, which is based on\nwhite-box gradient-based adversarial sample generation techniques (like FGSM),\nis obsolete for this use case, since the models are protected behind secure\ninternational exchange APIs, such as NASDAQ. In this research, we demonstrate\nthat a \"gray-box\" approach for attacking a Deep RL-based trading agent is\npossible by trading in the same stock market, with no extra access to the\ntrading agent. In our proposed approach, an adversary agent uses a hybrid Deep\nNeural Network as its policy consisting of Convolutional layers and\nfully-connected layers. On average, over three simulated trading market\nconfigurations, the adversary policy proposed in this research is able to\nreduce the reward values by 214.17%, which results in reducing the potential\nprofits of the baseline by 139.4%, ensemble method by 93.7%, and an automated\ntrading software developed by our industrial partner by 85.5%, while consuming\nsignificantly less budget than the victims (427.77%, 187.16%, and 66.97%,\nrespectively).\n", "date": "2023-09-26", "categories": []}
{"id": "2309.14700", "title": "Structure Invariant Transformation for better Adversarial\n  Transferability", "abstract": "  Given the severe vulnerability of Deep Neural Networks (DNNs) against\nadversarial examples, there is an urgent need for an effective adversarial\nattack to identify the deficiencies of DNNs in security-sensitive applications.\nAs one of the prevalent black-box adversarial attacks, the existing\ntransfer-based attacks still cannot achieve comparable performance with the\nwhite-box attacks. Among these, input transformation based attacks have shown\nremarkable effectiveness in boosting transferability. In this work, we find\nthat the existing input transformation based attacks transform the input image\nglobally, resulting in limited diversity of the transformed images. We\npostulate that the more diverse transformed images result in better\ntransferability. Thus, we investigate how to locally apply various\ntransformations onto the input image to improve such diversity while preserving\nthe structure of image. To this end, we propose a novel input transformation\nbased attack, called Structure Invariant Attack (SIA), which applies a random\nimage transformation onto each image block to craft a set of diverse images for\ngradient calculation. Extensive experiments on the standard ImageNet dataset\ndemonstrate that SIA exhibits much better transferability than the existing\nSOTA input transformation based attacks on CNN-based and transformer-based\nmodels, showing its generality and superiority in boosting transferability.\nCode is available at https://github.com/xiaosen-wang/SIT.\n", "date": "2023-09-26", "categories": []}
{"id": "2208.01819", "title": "Adversarial Camouflage for Node Injection Attack on Graphs", "abstract": "  Node injection attacks on Graph Neural Networks (GNNs) have received\nincreasing attention recently, due to their ability to degrade GNN performance\nwith high attack success rates. However, our study indicates that these attacks\noften fail in practical scenarios, since defense/detection methods can easily\nidentify and remove the injected nodes. To address this, we devote to\ncamouflage node injection attack, making injected nodes appear normal and\nimperceptible to defense/detection methods. Unfortunately, the non-Euclidean\nstructure of graph data and the lack of intuitive prior present great\nchallenges to the formalization, implementation, and evaluation of camouflage.\nIn this paper, we first propose and define camouflage as distribution\nsimilarity between ego networks of injected nodes and normal nodes. Then for\nimplementation, we propose an adversarial CAmouflage framework for Node\ninjection Attack, namely CANA, to improve attack performance under\ndefense/detection methods in practical scenarios. A novel camouflage metric is\nfurther designed under the guide of distribution similarity. Extensive\nexperiments demonstrate that CANA can significantly improve the attack\nperformance under defense/detection methods with higher camouflage or\nimperceptibility. This work urges us to raise awareness of the security\nvulnerabilities of GNNs in practical applications.\n", "date": "2022-08-03", "categories": []}
{"id": "2211.13775", "title": "SAGA: Spectral Adversarial Geometric Attack on 3D Meshes", "abstract": "  A triangular mesh is one of the most popular 3D data representations. As\nsuch, the deployment of deep neural networks for mesh processing is widely\nspread and is increasingly attracting more attention. However, neural networks\nare prone to adversarial attacks, where carefully crafted inputs impair the\nmodel's functionality. The need to explore these vulnerabilities is a\nfundamental factor in the future development of 3D-based applications.\nRecently, mesh attacks were studied on the semantic level, where classifiers\nare misled to produce wrong predictions. Nevertheless, mesh surfaces possess\ncomplex geometric attributes beyond their semantic meaning, and their analysis\noften includes the need to encode and reconstruct the geometry of the shape.\n  We propose a novel framework for a geometric adversarial attack on a 3D mesh\nautoencoder. In this setting, an adversarial input mesh deceives the\nautoencoder by forcing it to reconstruct a different geometric shape at its\noutput. The malicious input is produced by perturbing a clean shape in the\nspectral domain. Our method leverages the spectral decomposition of the mesh\nalong with additional mesh-related properties to obtain visually credible\nresults that consider the delicacy of surface distortions. Our code is publicly\navailable at https://github.com/StolikTomer/SAGA.\n", "date": "2022-11-24", "categories": []}
{"id": "2302.08051", "title": "Graph Adversarial Immunization for Certifiable Robustness", "abstract": "  Despite achieving great success, graph neural networks (GNNs) are vulnerable\nto adversarial attacks. Existing defenses focus on developing adversarial\ntraining or model modification. In this paper, we propose and formulate graph\nadversarial immunization, i.e., vaccinating part of graph structure to improve\ncertifiable robustness of graph against any admissible adversarial attack. We\nfirst propose edge-level immunization to vaccinate node pairs. Unfortunately,\nsuch edge-level immunization cannot defend against emerging node injection\nattacks, since it only immunizes existing node pairs. To this end, we further\npropose node-level immunization. To avoid computationally intensive\ncombinatorial optimization associated with adversarial immunization, we develop\nAdvImmune-Edge and AdvImmune-Node algorithms to effectively obtain the immune\nnode pairs or nodes. Extensive experiments demonstrate the superiority of\nAdvImmune methods. In particular, AdvImmune-Node remarkably improves the ratio\nof robust nodes by 79%, 294%, and 100%, after immunizing only 5% of nodes.\nFurthermore, AdvImmune methods show excellent defensive performance against\nvarious attacks, outperforming state-of-the-art defenses. To the best of our\nknowledge, this is the first attempt to improve certifiable robustness from\ngraph data perspective without losing performance on clean graphs, providing\nnew insights into graph adversarial learning.\n", "date": "2023-02-16", "categories": []}
{"id": "2305.14097", "title": "QFA2SR: Query-Free Adversarial Transfer Attacks to Speaker Recognition\n  Systems", "abstract": "  Current adversarial attacks against speaker recognition systems (SRSs)\nrequire either white-box access or heavy black-box queries to the target SRS,\nthus still falling behind practical attacks against proprietary commercial APIs\nand voice-controlled devices. To fill this gap, we propose QFA2SR, an effective\nand imperceptible query-free black-box attack, by leveraging the\ntransferability of adversarial voices. To improve transferability, we present\nthree novel methods, tailored loss functions, SRS ensemble, and time-freq\ncorrosion. The first one tailors loss functions to different attack scenarios.\nThe latter two augment surrogate SRSs in two different ways. SRS ensemble\ncombines diverse surrogate SRSs with new strategies, amenable to the unique\nscoring characteristics of SRSs. Time-freq corrosion augments surrogate SRSs by\nincorporating well-designed time-/frequency-domain modification functions,\nwhich simulate and approximate the decision boundary of the target SRS and\ndistortions introduced during over-the-air attacks. QFA2SR boosts the targeted\ntransferability by 20.9%-70.7% on four popular commercial APIs (Microsoft\nAzure, iFlytek, Jingdong, and TalentedSoft), significantly outperforming\nexisting attacks in query-free setting, with negligible effect on the\nimperceptibility. QFA2SR is also highly effective when launched over the air\nagainst three wide-spread voice assistants (Google Assistant, Apple Siri, and\nTMall Genie) with 60%, 46%, and 70% targeted transferability, respectively.\n", "date": "2023-05-23", "categories": []}
{"id": "2309.13245", "title": "RBFormer: Improve Adversarial Robustness of Transformer by Robust Bias", "abstract": "  Recently, there has been a surge of interest and attention in\nTransformer-based structures, such as Vision Transformer (ViT) and Vision\nMultilayer Perceptron (VMLP). Compared with the previous convolution-based\nstructures, the Transformer-based structure under investigation showcases a\ncomparable or superior performance under its distinctive attention-based input\ntoken mixer strategy. Introducing adversarial examples as a robustness\nconsideration has had a profound and detrimental impact on the performance of\nwell-established convolution-based structures. This inherent vulnerability to\nadversarial attacks has also been demonstrated in Transformer-based structures.\nIn this paper, our emphasis lies on investigating the intrinsic robustness of\nthe structure rather than introducing novel defense measures against\nadversarial attacks. To address the susceptibility to robustness issues, we\nemploy a rational structure design approach to mitigate such vulnerabilities.\nSpecifically, we enhance the adversarial robustness of the structure by\nincreasing the proportion of high-frequency structural robust biases. As a\nresult, we introduce a novel structure called Robust Bias Transformer-based\nStructure (RBFormer) that shows robust superiority compared to several existing\nbaseline structures. Through a series of extensive experiments, RBFormer\noutperforms the original structures by a significant margin, achieving an\nimpressive improvement of +16.12% and +5.04% across different evaluation\ncriteria on CIFAR-10 and ImageNet-1k, respectively.\n", "date": "2023-09-23", "categories": []}
{"id": "2309.13763", "title": "Combining Two Adversarial Attacks Against Person Re-Identification\n  Systems", "abstract": "  The field of Person Re-Identification (Re-ID) has received much attention\nrecently, driven by the progress of deep neural networks, especially for image\nclassification. The problem of Re-ID consists in identifying individuals\nthrough images captured by surveillance cameras in different scenarios.\nGovernments and companies are investing a lot of time and money in Re-ID\nsystems for use in public safety and identifying missing persons. However,\nseveral challenges remain for successfully implementing Re-ID, such as\nocclusions and light reflections in people's images. In this work, we focus on\nadversarial attacks on Re-ID systems, which can be a critical threat to the\nperformance of these systems. In particular, we explore the combination of\nadversarial attacks against Re-ID models, trying to strengthen the decrease in\nthe classification results. We conduct our experiments on three datasets:\nDukeMTMC-ReID, Market-1501, and CUHK03. We combine the use of two types of\nadversarial attacks, P-FGSM and Deep Mis-Ranking, applied to two popular Re-ID\nmodels: IDE (ResNet-50) and AlignedReID. The best result demonstrates a\ndecrease of 3.36% in the Rank-10 metric for AlignedReID applied to CUHK03. We\nalso try to use Dropout during the inference as a defense method.\n", "date": "2023-09-24", "categories": []}
{"id": "2309.13794", "title": "Projected Randomized Smoothing for Certified Adversarial Robustness", "abstract": "  Randomized smoothing is the current state-of-the-art method for producing\nprovably robust classifiers. While randomized smoothing typically yields robust\n$\\ell_2$-ball certificates, recent research has generalized provable robustness\nto different norm balls as well as anisotropic regions. This work considers a\nclassifier architecture that first projects onto a low-dimensional\napproximation of the data manifold and then applies a standard classifier. By\nperforming randomized smoothing in the low-dimensional projected space, we\ncharacterize the certified region of our smoothed composite classifier back in\nthe high-dimensional input space and prove a tractable lower bound on its\nvolume. We show experimentally on CIFAR-10 and SVHN that classifiers without\nthe initial projection are vulnerable to perturbations that are normal to the\ndata manifold and yet are captured by the certified regions of our method. We\ncompare the volume of our certified regions against various baselines and show\nthat our method improves on the state-of-the-art by many orders of magnitude.\n", "date": "2023-09-25", "categories": []}
{"id": "2309.13841", "title": "On the Effectiveness of Adversarial Samples against Ensemble\n  Learning-based Windows PE Malware Detectors", "abstract": "  Recently, there has been a growing focus and interest in applying machine\nlearning (ML) to the field of cybersecurity, particularly in malware detection\nand prevention. Several research works on malware analysis have been proposed,\noffering promising results for both academic and practical applications. In\nthese works, the use of Generative Adversarial Networks (GANs) or Reinforcement\nLearning (RL) can aid malware creators in crafting metamorphic malware that\nevades antivirus software. In this study, we propose a mutation system to\ncounteract ensemble learning-based detectors by combining GANs and an RL model,\novercoming the limitations of the MalGAN model. Our proposed FeaGAN model is\nbuilt based on MalGAN by incorporating an RL model called the Deep Q-network\nanti-malware Engines Attacking Framework (DQEAF). The RL model addresses three\nkey challenges in performing adversarial attacks on Windows Portable Executable\nmalware, including format preservation, executability preservation, and\nmaliciousness preservation. In the FeaGAN model, ensemble learning is utilized\nto enhance the malware detector's evasion ability, with the generated\nadversarial patterns. The experimental results demonstrate that 100\\% of the\nselected mutant samples preserve the format of executable files, while certain\nsuccesses in both executability preservation and maliciousness preservation are\nachieved, reaching a stable success rate.\n", "date": "2023-09-25", "categories": []}
{"id": "2309.12593", "title": "Improving Machine Learning Robustness via Adversarial Training", "abstract": "  As Machine Learning (ML) is increasingly used in solving various tasks in\nreal-world applications, it is crucial to ensure that ML algorithms are robust\nto any potential worst-case noises, adversarial attacks, and highly unusual\nsituations when they are designed. Studying ML robustness will significantly\nhelp in the design of ML algorithms. In this paper, we investigate ML\nrobustness using adversarial training in centralized and decentralized\nenvironments, where ML training and testing are conducted in one or multiple\ncomputers. In the centralized environment, we achieve a test accuracy of 65.41%\nand 83.0% when classifying adversarial examples generated by Fast Gradient Sign\nMethod and DeepFool, respectively. Comparing to existing studies, these results\ndemonstrate an improvement of 18.41% for FGSM and 47% for DeepFool. In the\ndecentralized environment, we study Federated learning (FL) robustness by using\nadversarial training with independent and identically distributed (IID) and\nnon-IID data, respectively, where CIFAR-10 is used in this research. In the IID\ndata case, our experimental results demonstrate that we can achieve such a\nrobust accuracy that it is comparable to the one obtained in the centralized\nenvironment. Moreover, in the non-IID data case, the natural accuracy drops\nfrom 66.23% to 57.82%, and the robust accuracy decreases by 25% and 23.4% in\nC&W and Projected Gradient Descent (PGD) attacks, compared to the IID data\ncase, respectively. We further propose an IID data-sharing approach, which\nallows for increasing the natural accuracy to 85.04% and the robust accuracy\nfrom 57% to 72% in C&W attacks and from 59% to 67% in PGD attacks.\n", "date": "2023-09-22", "categories": []}
{"id": "2309.11575", "title": "Distilling Adversarial Prompts from Safety Benchmarks: Report for the\n  Adversarial Nibbler Challenge", "abstract": "  Text-conditioned image generation models have recently achieved astonishing\nimage quality and alignment results. Consequently, they are employed in a\nfast-growing number of applications. Since they are highly data-driven, relying\non billion-sized datasets randomly scraped from the web, they also produce\nunsafe content. As a contribution to the Adversarial Nibbler challenge, we\ndistill a large set of over 1,000 potential adversarial inputs from existing\nsafety benchmarks. Our analysis of the gathered prompts and corresponding\nimages demonstrates the fragility of input filters and provides further\ninsights into systematic safety issues in current generative image models.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.11005", "title": "It's Simplex! Disaggregating Measures to Improve Certified Robustness", "abstract": "  Certified robustness circumvents the fragility of defences against\nadversarial attacks, by endowing model predictions with guarantees of class\ninvariance for attacks up to a calculated size. While there is value in these\ncertifications, the techniques through which we assess their performance do not\npresent a proper accounting of their strengths and weaknesses, as their\nanalysis has eschewed consideration of performance over individual samples in\nfavour of aggregated measures. By considering the potential output space of\ncertified models, this work presents two distinct approaches to improve the\nanalysis of certification mechanisms, that allow for both dataset-independent\nand dataset-dependent measures of certification performance. Embracing such a\nperspective uncovers new certification approaches, which have the potential to\nmore than double the achievable radius of certification, relative to current\nstate-of-the-art. Empirical evaluation verifies that our new approach can\ncertify $9\\%$ more samples at noise scale $\\sigma = 1$, with greater relative\nimprovements observed as the difficulty of the predictive task increases.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.11111", "title": "PRAT: PRofiling Adversarial aTtacks", "abstract": "  Intrinsic susceptibility of deep learning to adversarial examples has led to\na plethora of attack techniques with a broad common objective of fooling deep\nmodels. However, we find slight compositional differences between the\nalgorithms achieving this objective. These differences leave traces that\nprovide important clues for attacker profiling in real-life scenarios. Inspired\nby this, we introduce a novel problem of PRofiling Adversarial aTtacks (PRAT).\nGiven an adversarial example, the objective of PRAT is to identify the attack\nused to generate it. Under this perspective, we can systematically group\nexisting attacks into different families, leading to the sub-problem of attack\nfamily identification, which we also study. To enable PRAT analysis, we\nintroduce a large Adversarial Identification Dataset (AID), comprising over\n180k adversarial samples generated with 13 popular attacks for image\nspecific/agnostic white/black box setups. We use AID to devise a novel\nframework for the PRAT objective. Our framework utilizes a Transformer based\nGlobal-LOcal Feature (GLOF) module to extract an approximate signature of the\nadversarial attack, which in turn is used for the identification of the attack.\nUsing AID and our framework, we provide multiple interesting benchmark results\nfor the PRAT problem.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.11462", "title": "AudioFool: Fast, Universal and synchronization-free Cross-Domain Attack\n  on Speech Recognition", "abstract": "  Automatic Speech Recognition systems have been shown to be vulnerable to\nadversarial attacks that manipulate the command executed on the device. Recent\nresearch has focused on exploring methods to create such attacks, however, some\nissues relating to Over-The-Air (OTA) attacks have not been properly addressed.\nIn our work, we examine the needed properties of robust attacks compatible with\nthe OTA model, and we design a method of generating attacks with arbitrary such\ndesired properties, namely the invariance to synchronization, and the\nrobustness to filtering: this allows a Denial-of-Service (DoS) attack against\nASR systems. We achieve these characteristics by constructing attacks in a\nmodified frequency domain through an inverse Fourier transform. We evaluate our\nmethod on standard keyword classification tasks and analyze it in OTA, and we\nanalyze the properties of the cross-domain attacks to explain the efficiency of\nthe approach.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.10136", "title": "Efficient Low-Rank GNN Defense Against Structural Attacks", "abstract": "  Graph Neural Networks (GNNs) have been shown to possess strong representation\nabilities over graph data. However, GNNs are vulnerable to adversarial attacks,\nand even minor perturbations to the graph structure can significantly degrade\ntheir performance. Existing methods either are ineffective against\nsophisticated attacks or require the optimization of dense adjacency matrices,\nwhich is time-consuming and prone to local minima. To remedy this problem, we\npropose an Efficient Low-Rank Graph Neural Network (ELR-GNN) defense method,\nwhich aims to learn low-rank and sparse graph structures for defending against\nadversarial attacks, ensuring effective defense with greater efficiency.\nSpecifically, ELR-GNN consists of two modules: a Coarse Low-Rank Estimation\nModule and a Fine-Grained Estimation Module. The first module adopts the\ntruncated Singular Value Decomposition (SVD) to initialize the low-rank\nadjacency matrix estimation, which serves as a starting point for optimizing\nthe low-rank matrix. In the second module, the initial estimate is refined by\njointly learning a low-rank sparse graph structure with the GNN model. Sparsity\nis incorporated into the learned low-rank adjacency matrix by pruning weak\nconnections, which can reduce redundant data while maintaining valuable\ninformation. As a result, instead of using the dense adjacency matrix directly,\nELR-GNN can learn a low-rank and sparse estimate of it in a simple, efficient\nand easy to optimize manner. The experimental results demonstrate that ELR-GNN\noutperforms the state-of-the-art GNN defense methods in the literature, in\naddition to being very efficient and easy to train.\n", "date": "2023-09-18", "categories": []}
{"id": "2309.10243", "title": "Transferable Adversarial Attack on Image Tampering Localization", "abstract": "  It is significant to evaluate the security of existing digital image\ntampering localization algorithms in real-world applications. In this paper, we\npropose an adversarial attack scheme to reveal the reliability of such\ntampering localizers, which would be fooled and fail to predict altered regions\ncorrectly. Specifically, the adversarial examples based on optimization and\ngradient are implemented for white/black-box attacks. Correspondingly, the\nadversarial example is optimized via reverse gradient propagation, and the\nperturbation is added adaptively in the direction of gradient rising. The\nblack-box attack is achieved by relying on the transferability of such\nadversarial examples to different localizers. Extensive evaluations verify that\nthe proposed attack sharply reduces the localization accuracy while preserving\nhigh visual quality of the attacked images.\n", "date": "2023-09-19", "categories": []}
{"id": "2309.10586", "title": "Adversarial Attacks Against Uncertainty Quantification", "abstract": "  Machine-learning models can be fooled by adversarial examples, i.e.,\ncarefully-crafted input perturbations that force models to output wrong\npredictions. While uncertainty quantification has been recently proposed to\ndetect adversarial inputs, under the assumption that such attacks exhibit a\nhigher prediction uncertainty than pristine data, it has been shown that\nadaptive attacks specifically aimed at reducing also the uncertainty estimate\ncan easily bypass this defense mechanism. In this work, we focus on a different\nadversarial scenario in which the attacker is still interested in manipulating\nthe uncertainty estimate, but regardless of the correctness of the prediction;\nin particular, the goal is to undermine the use of machine-learning models when\ntheir outputs are consumed by a downstream module or by a human operator.\nFollowing such direction, we: \\textit{(i)} design a threat model for attacks\ntargeting uncertainty quantification; \\textit{(ii)} devise different attack\nstrategies on conceptually different UQ techniques spanning for both\nclassification and semantic segmentation problems; \\textit{(iii)} conduct a\nfirst complete and extensive analysis to compare the differences between some\nof the most employed UQ approaches under attack. Our extensive experimental\nanalysis shows that our attacks are more effective in manipulating uncertainty\nquantification measures than attacks aimed to also induce misclassifications.\n", "date": "2023-09-19", "categories": []}
{"id": "2209.14262", "title": "A Survey on Physical Adversarial Attack in Computer Vision", "abstract": "  Over the past decade, deep learning has revolutionized conventional tasks\nthat rely on hand-craft feature extraction with its strong feature learning\ncapability, leading to substantial enhancements in traditional tasks. However,\ndeep neural networks (DNNs) have been demonstrated to be vulnerable to\nadversarial examples crafted by malicious tiny noise, which is imperceptible to\nhuman observers but can make DNNs output the wrong result. Existing adversarial\nattacks can be categorized into digital and physical adversarial attacks. The\nformer is designed to pursue strong attack performance in lab environments\nwhile hardly remaining effective when applied to the physical world. In\ncontrast, the latter focus on developing physical deployable attacks, thus\nexhibiting more robustness in complex physical environmental conditions.\nRecently, with the increasing deployment of the DNN-based system in the real\nworld, strengthening the robustness of these systems is an emergency, while\nexploring physical adversarial attacks exhaustively is the precondition. To\nthis end, this paper reviews the evolution of physical adversarial attacks\nagainst DNN-based computer vision tasks, expecting to provide beneficial\ninformation for developing stronger physical adversarial attacks. Specifically,\nwe first proposed a taxonomy to categorize the current physical adversarial\nattacks and grouped them. Then, we discuss the existing physical attacks and\nfocus on the technique for improving the robustness of physical attacks under\ncomplex physical environmental conditions. Finally, we discuss the issues of\nthe current physical adversarial attacks to be solved and give promising\ndirections.\n", "date": "2022-09-28", "categories": []}
{"id": "2302.08257", "title": "On the Effect of Adversarial Training Against Invariance-based\n  Adversarial Examples", "abstract": "  Adversarial examples are carefully crafted attack points that are supposed to\nfool machine learning classifiers. In the last years, the field of adversarial\nmachine learning, especially the study of perturbation-based adversarial\nexamples, in which a perturbation that is not perceptible for humans is added\nto the images, has been studied extensively. Adversarial training can be used\nto achieve robustness against such inputs. Another type of adversarial examples\nare invariance-based adversarial examples, where the images are semantically\nmodified such that the predicted class of the model does not change, but the\nclass that is determined by humans does. How to ensure robustness against this\ntype of adversarial examples has not been explored yet. This work addresses the\nimpact of adversarial training with invariance-based adversarial examples on a\nconvolutional neural network (CNN).\n  We show that when adversarial training with invariance-based and\nperturbation-based adversarial examples is applied, it should be conducted\nsimultaneously and not consecutively. This procedure can achieve relatively\nhigh robustness against both types of adversarial examples. Additionally, we\nfind that the algorithm used for generating invariance-based adversarial\nexamples in prior work does not correctly determine the labels and therefore we\nuse human-determined labels.\n", "date": "2023-02-16", "categories": []}
{"id": "2307.10487", "title": "Attacking by Aligning: Clean-Label Backdoor Attacks on Object Detection", "abstract": "  Deep neural networks (DNNs) have shown unprecedented success in object\ndetection tasks. However, it was also discovered that DNNs are vulnerable to\nmultiple kinds of attacks, including Backdoor Attacks. Through the attack, the\nattacker manages to embed a hidden backdoor into the DNN such that the model\nbehaves normally on benign data samples, but makes attacker-specified judgments\ngiven the occurrence of a predefined trigger. Although numerous backdoor\nattacks have been experimented on image classification, backdoor attacks on\nobject detection tasks have not been properly investigated and explored. As\nobject detection has been adopted as an important module in multiple\nsecurity-sensitive applications such as autonomous driving, backdoor attacks on\nobject detection could pose even more severe threats. Inspired by the inherent\nproperty of deep learning-based object detectors, we propose a simple yet\neffective backdoor attack method against object detection without modifying the\nground truth annotations, specifically focusing on the object disappearance\nattack and object generation attack. Extensive experiments and ablation studies\nprove the effectiveness of our attack on the benchmark object detection dataset\nMSCOCO2017, on which we achieve an attack success rate of more than 92% with a\npoison rate of only 5%.\n", "date": "2023-07-19", "categories": []}
{"id": "2309.08953", "title": "Robust Backdoor Attacks on Object Detection in Real World", "abstract": "  Deep learning models are widely deployed in many applications, such as object\ndetection in various security fields. However, these models are vulnerable to\nbackdoor attacks. Most backdoor attacks were intensively studied on classified\nmodels, but little on object detection. Previous works mainly focused on the\nbackdoor attack in the digital world, but neglect the real world. Especially,\nthe backdoor attack's effect in the real world will be easily influenced by\nphysical factors like distance and illumination. In this paper, we proposed a\nvariable-size backdoor trigger to adapt to the different sizes of attacked\nobjects, overcoming the disturbance caused by the distance between the viewing\npoint and attacked object. In addition, we proposed a backdoor training named\nmalicious adversarial training, enabling the backdoor object detector to learn\nthe feature of the trigger with physical noise. The experiment results show\nthis robust backdoor attack (RBA) could enhance the attack success rate in the\nreal world.\n", "date": "2023-09-16", "categories": []}
{"id": "2309.09928", "title": "Evaluating Adversarial Robustness with Expected Viable Performance", "abstract": "  We introduce a metric for evaluating the robustness of a classifier, with\nparticular attention to adversarial perturbations, in terms of expected\nfunctionality with respect to possible adversarial perturbations. A classifier\nis assumed to be non-functional (that is, has a functionality of zero) with\nrespect to a perturbation bound if a conventional measure of performance, such\nas classification accuracy, is less than a minimally viable threshold when the\nclassifier is tested on examples from that perturbation bound. Defining\nrobustness in terms of an expected value is motivated by a domain general\napproach to robustness quantification.\n", "date": "2023-09-18", "categories": []}
{"id": "2304.10136", "title": "Diversifying the High-level Features for better Adversarial\n  Transferability", "abstract": "  Given the great threat of adversarial attacks against Deep Neural Networks\n(DNNs), numerous works have been proposed to boost transferability to attack\nreal-world applications. However, existing attacks often utilize advanced\ngradient calculation or input transformation but ignore the white-box model.\nInspired by the fact that DNNs are over-parameterized for superior performance,\nwe propose diversifying the high-level features (DHF) for more transferable\nadversarial examples. In particular, DHF perturbs the high-level features by\nrandomly transforming the high-level features and mixing them with the feature\nof benign samples when calculating the gradient at each iteration. Due to the\nredundancy of parameters, such transformation does not affect the\nclassification performance but helps identify the invariant features across\ndifferent models, leading to much better transferability. Empirical evaluations\non ImageNet dataset show that DHF could effectively improve the transferability\nof existing momentum-based attacks. Incorporated into the input\ntransformation-based attacks, DHF generates more transferable adversarial\nexamples and outperforms the baselines with a clear margin when attacking\nseveral defense models, showing its generalization to various attacks and high\neffectiveness for boosting transferability. Code is available at\nhttps://github.com/Trustworthy-AI-Group/DHF.\n", "date": "2023-04-20", "categories": []}
{"id": "2208.13049", "title": "TrojViT: Trojan Insertion in Vision Transformers", "abstract": "  Vision Transformers (ViTs) have demonstrated the state-of-the-art performance\nin various vision-related tasks. The success of ViTs motivates adversaries to\nperform backdoor attacks on ViTs. Although the vulnerability of traditional\nCNNs to backdoor attacks is well-known, backdoor attacks on ViTs are\nseldom-studied. Compared to CNNs capturing pixel-wise local features by\nconvolutions, ViTs extract global context information through patches and\nattentions. Na\\\"ively transplanting CNN-specific backdoor attacks to ViTs\nyields only a low clean data accuracy and a low attack success rate. In this\npaper, we propose a stealth and practical ViT-specific backdoor attack\n$TrojViT$. Rather than an area-wise trigger used by CNN-specific backdoor\nattacks, TrojViT generates a patch-wise trigger designed to build a Trojan\ncomposed of some vulnerable bits on the parameters of a ViT stored in DRAM\nmemory through patch salience ranking and attention-target loss. TrojViT\nfurther uses minimum-tuned parameter update to reduce the bit number of the\nTrojan. Once the attacker inserts the Trojan into the ViT model by flipping the\nvulnerable bits, the ViT model still produces normal inference accuracy with\nbenign inputs. But when the attacker embeds a trigger into an input, the ViT\nmodel is forced to classify the input to a predefined target class. We show\nthat flipping only few vulnerable bits identified by TrojViT on a ViT model\nusing the well-known RowHammer can transform the model into a backdoored one.\nWe perform extensive experiments of multiple datasets on various ViT models.\nTrojViT can classify $99.64\\%$ of test images to a target class by flipping\n$345$ bits on a ViT for ImageNet.Our codes are available at\nhttps://github.com/mxzheng/TrojViT\n", "date": "2022-08-27", "categories": []}
{"id": "2303.09858", "title": "Preventing Unauthorized AI Over-Analysis by Medical Image Adversarial\n  Watermarking", "abstract": "  The advancement of deep learning has facilitated the integration of\nArtificial Intelligence (AI) into clinical practices, particularly in\ncomputer-aided diagnosis. Given the pivotal role of medical images in various\ndiagnostic procedures, it becomes imperative to ensure the responsible and\nsecure utilization of AI techniques. However, the unauthorized utilization of\nAI for image analysis raises significant concerns regarding patient privacy and\npotential infringement on the proprietary rights of data custodians.\nConsequently, the development of pragmatic and cost-effective strategies that\nsafeguard patient privacy and uphold medical image copyrights emerges as a\ncritical necessity. In direct response to this pressing demand, we present a\npioneering solution named Medical Image Adversarial watermarking (MIAD-MARK).\nOur approach introduces watermarks that strategically mislead unauthorized AI\ndiagnostic models, inducing erroneous predictions without compromising the\nintegrity of the visual content. Importantly, our method integrates an\nauthorization protocol tailored for legitimate users, enabling the removal of\nthe MIAD-MARK through encryption-generated keys. Through extensive experiments,\nwe validate the efficacy of MIAD-MARK across three prominent medical image\ndatasets. The empirical outcomes demonstrate the substantial impact of our\napproach, notably reducing the accuracy of standard AI diagnostic models to a\nmere 8.57% under white box conditions and 45.83% in the more challenging black\nbox scenario. Additionally, our solution effectively mitigates unauthorized\nexploitation of medical images even in the presence of sophisticated watermark\nremoval networks. Notably, those AI diagnosis networks exhibit a meager average\naccuracy of 38.59% when applied to images protected by MIAD-MARK, underscoring\nthe robustness of our safeguarding mechanism.\n", "date": "2023-03-17", "categories": []}
{"id": "2306.05659", "title": "COVER: A Heuristic Greedy Adversarial Attack on Prompt-based Learning in\n  Language Models", "abstract": "  Prompt-based learning has been proved to be an effective way in pre-trained\nlanguage models (PLMs), especially in low-resource scenarios like few-shot\nsettings. However, the trustworthiness of PLMs is of paramount significance and\npotential vulnerabilities have been shown in prompt-based templates that could\nmislead the predictions of language models, causing serious security concerns.\nIn this paper, we will shed light on some vulnerabilities of PLMs, by proposing\na prompt-based adversarial attack on manual templates in black box scenarios.\nFirst of all, we design character-level and word-level heuristic approaches to\nbreak manual templates separately. Then we present a greedy algorithm for the\nattack based on the above heuristic destructive approaches. Finally, we\nevaluate our approach with the classification tasks on three variants of BERT\nseries models and eight datasets. And comprehensive experimental results\njustify the effectiveness of our approach in terms of attack success rate and\nattack speed.\n", "date": "2023-06-09", "categories": []}
{"id": "2309.07398", "title": "Semantic Adversarial Attacks via Diffusion Models", "abstract": "  Traditional adversarial attacks concentrate on manipulating clean examples in\nthe pixel space by adding adversarial perturbations. By contrast, semantic\nadversarial attacks focus on changing semantic attributes of clean examples,\nsuch as color, context, and features, which are more feasible in the real\nworld. In this paper, we propose a framework to quickly generate a semantic\nadversarial attack by leveraging recent diffusion models since semantic\ninformation is included in the latent space of well-trained diffusion models.\nThen there are two variants of this framework: 1) the Semantic Transformation\n(ST) approach fine-tunes the latent space of the generated image and/or the\ndiffusion model itself; 2) the Latent Masking (LM) approach masks the latent\nspace with another target image and local backpropagation-based interpretation\nmethods. Additionally, the ST approach can be applied in either white-box or\nblack-box settings. Extensive experiments are conducted on CelebA-HQ and AFHQ\ndatasets, and our framework demonstrates great fidelity, generalizability, and\ntransferability compared to other baselines. Our approaches achieve\napproximately 100% attack success rate in multiple settings with the best FID\nas 36.61. Code is available at\nhttps://github.com/steven202/semantic_adv_via_dm.\n", "date": "2023-09-14", "categories": []}
{"id": "2309.07428", "title": "Physical Invisible Backdoor Based on Camera Imaging", "abstract": "  Backdoor attack aims to compromise a model, which returns an adversary-wanted\noutput when a specific trigger pattern appears yet behaves normally for clean\ninputs. Current backdoor attacks require changing pixels of clean images, which\nresults in poor stealthiness of attacks and increases the difficulty of the\nphysical implementation. This paper proposes a novel physical invisible\nbackdoor based on camera imaging without changing nature image pixels.\nSpecifically, a compromised model returns a target label for images taken by a\nparticular camera, while it returns correct results for other images. To\nimplement and evaluate the proposed backdoor, we take shots of different\nobjects from multi-angles using multiple smartphones to build a new dataset of\n21,500 images. Conventional backdoor attacks work ineffectively with some\nclassical models, such as ResNet18, over the above-mentioned dataset.\nTherefore, we propose a three-step training strategy to mount the backdoor\nattack. First, we design and train a camera identification model with the phone\nIDs to extract the camera fingerprint feature. Subsequently, we elaborate a\nspecial network architecture, which is easily compromised by our backdoor\nattack, by leveraging the attributes of the CFA interpolation algorithm and\ncombining it with the feature extraction block in the camera identification\nmodel. Finally, we transfer the backdoor from the elaborated special network\narchitecture to the classical architecture model via teacher-student\ndistillation learning. Since the trigger of our method is related to the\nspecific phone, our attack works effectively in the physical world. Experiment\nresults demonstrate the feasibility of our proposed approach and robustness\nagainst various backdoor defenses.\n", "date": "2023-09-14", "categories": []}
{"id": "2301.10908", "title": "Distilling Cognitive Backdoor Patterns within an Image", "abstract": "  This paper proposes a simple method to distill and detect backdoor patterns\nwithin an image: \\emph{Cognitive Distillation} (CD). The idea is to extract the\n\"minimal essence\" from an input image responsible for the model's prediction.\nCD optimizes an input mask to extract a small pattern from the input image that\ncan lead to the same model output (i.e., logits or deep features). The\nextracted pattern can help understand the cognitive mechanism of a model on\nclean vs. backdoor images and is thus called a \\emph{Cognitive Pattern} (CP).\nUsing CD and the distilled CPs, we uncover an interesting phenomenon of\nbackdoor attacks: despite the various forms and sizes of trigger patterns used\nby different attacks, the CPs of backdoor samples are all surprisingly and\nsuspiciously small. One thus can leverage the learned mask to detect and remove\nbackdoor examples from poisoned training datasets. We conduct extensive\nexperiments to show that CD can robustly detect a wide range of advanced\nbackdoor attacks. We also show that CD can potentially be applied to help\ndetect potential biases from face datasets. Code is available at\n\\url{https://github.com/HanxunH/CognitiveDistillation}.\n", "date": "2023-01-26", "categories": []}
{"id": "2309.06527", "title": "Machine Translation Models Stand Strong in the Face of Adversarial\n  Attacks", "abstract": "  Adversarial attacks expose vulnerabilities of deep learning models by\nintroducing minor perturbations to the input, which lead to substantial\nalterations in the output. Our research focuses on the impact of such\nadversarial attacks on sequence-to-sequence (seq2seq) models, specifically\nmachine translation models. We introduce algorithms that incorporate basic text\nperturbation heuristics and more advanced strategies, such as the\ngradient-based attack, which utilizes a differentiable approximation of the\ninherently non-differentiable translation metric. Through our investigation, we\nprovide evidence that machine translation models display robustness displayed\nrobustness against best performed known adversarial attacks, as the degree of\nperturbation in the output is directly proportional to the perturbation in the\ninput. However, among underdogs, our attacks outperform alternatives, providing\nthe best relative performance. Another strong candidate is an attack based on\nmixing of individual characters.\n", "date": "2023-09-10", "categories": []}
{"id": "2309.06960", "title": "PhantomSound: Black-Box, Query-Efficient Audio Adversarial Attack via\n  Split-Second Phoneme Injection", "abstract": "  In this paper, we propose PhantomSound, a query-efficient black-box attack\ntoward voice assistants. Existing black-box adversarial attacks on voice\nassistants either apply substitution models or leverage the intermediate model\noutput to estimate the gradients for crafting adversarial audio samples.\nHowever, these attack approaches require a significant amount of queries with a\nlengthy training stage. PhantomSound leverages the decision-based attack to\nproduce effective adversarial audios, and reduces the number of queries by\noptimizing the gradient estimation. In the experiments, we perform our attack\nagainst 4 different speech-to-text APIs under 3 real-world scenarios to\ndemonstrate the real-time attack impact. The results show that PhantomSound is\npractical and robust in attacking 5 popular commercial voice controllable\ndevices over the air, and is able to bypass 3 liveness detection mechanisms\nwith >95% success rate. The benchmark result shows that PhantomSound can\ngenerate adversarial examples and launch the attack in a few minutes. We\nsignificantly enhance the query efficiency and reduce the cost of a successful\nuntargeted and targeted adversarial attack by 93.1% and 65.5% compared with the\nstate-of-the-art black-box attacks, using merely ~300 queries (~5 minutes) and\n~1,500 queries (~25 minutes), respectively.\n", "date": "2023-09-13", "categories": []}
{"id": "2309.06981", "title": "MASTERKEY: Practical Backdoor Attack Against Speaker Verification\n  Systems", "abstract": "  Speaker Verification (SV) is widely deployed in mobile systems to\nauthenticate legitimate users by using their voice traits. In this work, we\npropose a backdoor attack MASTERKEY, to compromise the SV models. Different\nfrom previous attacks, we focus on a real-world practical setting where the\nattacker possesses no knowledge of the intended victim. To design MASTERKEY, we\ninvestigate the limitation of existing poisoning attacks against unseen\ntargets. Then, we optimize a universal backdoor that is capable of attacking\narbitrary targets. Next, we embed the speaker's characteristics and semantics\ninformation into the backdoor, making it imperceptible. Finally, we estimate\nthe channel distortion and integrate it into the backdoor. We validate our\nattack on 6 popular SV models. Specifically, we poison a total of 53 models and\nuse our trigger to attack 16,430 enrolled speakers, composed of 310 target\nspeakers enrolled in 53 poisoned models. Our attack achieves 100% attack\nsuccess rate with a 15% poison rate. By decreasing the poison rate to 3%, the\nattack success rate remains around 50%. We validate our attack in 3 real-world\nscenarios and successfully demonstrate the attack through both over-the-air and\nover-the-telephony-line scenarios.\n", "date": "2023-09-13", "categories": []}
{"id": "2309.07106", "title": "Hardening RGB-D Object Recognition Systems against Adversarial Patch\n  Attacks", "abstract": "  RGB-D object recognition systems improve their predictive performances by\nfusing color and depth information, outperforming neural network architectures\nthat rely solely on colors. While RGB-D systems are expected to be more robust\nto adversarial examples than RGB-only systems, they have also been proven to be\nhighly vulnerable. Their robustness is similar even when the adversarial\nexamples are generated by altering only the original images' colors. Different\nworks highlighted the vulnerability of RGB-D systems; however, there is a\nlacking of technical explanations for this weakness. Hence, in our work, we\nbridge this gap by investigating the learned deep representation of RGB-D\nsystems, discovering that color features make the function learned by the\nnetwork more complex and, thus, more sensitive to small perturbations. To\nmitigate this problem, we propose a defense based on a detection mechanism that\nmakes RGB-D systems more robust against adversarial examples. We empirically\nshow that this defense improves the performances of RGB-D systems against\nadversarial examples even when they are computed ad-hoc to circumvent this\ndetection mechanism, and that is also more effective than adversarial training.\n", "date": "2023-09-13", "categories": []}
{"id": "2309.05879", "title": "Generalized Attacks on Face Verification Systems", "abstract": "  Face verification (FV) using deep neural network models has made tremendous\nprogress in recent years, surpassing human accuracy and seeing deployment in\nvarious applications such as border control and smartphone unlocking. However,\nFV systems are vulnerable to Adversarial Attacks, which manipulate input images\nto deceive these systems in ways usually unnoticeable to humans. This paper\nprovides an in-depth study of attacks on FV systems. We introduce the\nDodgePersonation Attack that formulates the creation of face images that\nimpersonate a set of given identities while avoiding being identified as any of\nthe identities in a separate, disjoint set. A taxonomy is proposed to provide a\nunified view of different types of Adversarial Attacks against FV systems,\nincluding Dodging Attacks, Impersonation Attacks, and Master Face Attacks.\nFinally, we propose the ''One Face to Rule Them All'' Attack which implements\nthe DodgePersonation Attack with state-of-the-art performance on a well-known\nscenario (Master Face Attack) and which can also be used for the new scenarios\nintroduced in this paper. While the state-of-the-art Master Face Attack can\nproduce a set of 9 images to cover 43.82% of the identities in their test\ndatabase, with 9 images our attack can cover 57.27% to 58.5% of these\nidentifies while giving the attacker the choice of the identity to use to\ncreate the impersonation. Moreover, the 9 generated attack images appear\nidentical to a casual observer.\n", "date": "2023-09-12", "categories": []}
{"id": "2309.05900", "title": "Adversarial Attacks Assessment of Salient Object Detection via Symbolic\n  Learning", "abstract": "  Machine learning is at the center of mainstream technology and outperforms\nclassical approaches to handcrafted feature design. Aside from its learning\nprocess for artificial feature extraction, it has an end-to-end paradigm from\ninput to output, reaching outstandingly accurate results. However, security\nconcerns about its robustness to malicious and imperceptible perturbations have\ndrawn attention since its prediction can be changed entirely. Salient object\ndetection is a research area where deep convolutional neural networks have\nproven effective but whose trustworthiness represents a significant issue\nrequiring analysis and solutions to hackers' attacks. Brain programming is a\nkind of symbolic learning in the vein of good old-fashioned artificial\nintelligence. This work provides evidence that symbolic learning robustness is\ncrucial in designing reliable visual attention systems since it can withstand\neven the most intense perturbations. We test this evolutionary computation\nmethodology against several adversarial attacks and noise perturbations using\nstandard databases and a real-world problem of a shorebird called the Snowy\nPlover portraying a visual attention task. We compare our methodology with five\ndifferent deep learning approaches, proving that they do not match the symbolic\nparadigm regarding robustness. All neural networks suffer significant\nperformance losses, while brain programming stands its ground and remains\nunaffected. Also, by studying the Snowy Plover, we remark on the importance of\nsecurity in surveillance activities regarding wildlife protection and\nconservation.\n", "date": "2023-09-12", "categories": []}
{"id": "2309.06166", "title": "Certified Robust Models with Slack Control and Large Lipschitz Constants", "abstract": "  Despite recent success, state-of-the-art learning-based models remain highly\nvulnerable to input changes such as adversarial examples. In order to obtain\ncertifiable robustness against such perturbations, recent work considers\nLipschitz-based regularizers or constraints while at the same time increasing\nprediction margin. Unfortunately, this comes at the cost of significantly\ndecreased accuracy. In this paper, we propose a Calibrated Lipschitz-Margin\nLoss (CLL) that addresses this issue and improves certified robustness by\ntackling two problems: Firstly, commonly used margin losses do not adjust the\npenalties to the shrinking output distribution; caused by minimizing the\nLipschitz constant $K$. Secondly, and most importantly, we observe that\nminimization of $K$ can lead to overly smooth decision functions. This limits\nthe model's complexity and thus reduces accuracy. Our CLL addresses these\nissues by explicitly calibrating the loss w.r.t. margin and Lipschitz constant,\nthereby establishing full control over slack and improving robustness\ncertificates even with larger Lipschitz constants. On CIFAR-10, CIFAR-100 and\nTiny-ImageNet, our models consistently outperform losses that leave the\nconstant unattended. On CIFAR-100 and Tiny-ImageNet, CLL improves upon\nstate-of-the-art deterministic $L_2$ robust accuracies. In contrast to current\ntrends, we unlock potential of much smaller models without $K=1$ constraints.\n", "date": "2023-09-12", "categories": []}
{"id": "2309.06359", "title": "Using Reed-Muller Codes for Classification with Rejection and Recovery", "abstract": "  When deploying classifiers in the real world, users expect them to respond to\ninputs appropriately. However, traditional classifiers are not equipped to\nhandle inputs which lie far from the distribution they were trained on.\nMalicious actors can exploit this defect by making adversarial perturbations\ndesigned to cause the classifier to give an incorrect output.\nClassification-with-rejection methods attempt to solve this problem by allowing\nnetworks to refuse to classify an input in which they have low confidence. This\nworks well for strongly adversarial examples, but also leads to the rejection\nof weakly perturbed images, which intuitively could be correctly classified. To\naddress these issues, we propose Reed-Muller Aggregation Networks (RMAggNet), a\nclassifier inspired by Reed-Muller error-correction codes which can correct and\nreject inputs. This paper shows that RMAggNet can minimise incorrectness while\nmaintaining good correctness over multiple adversarial attacks at different\nperturbation budgets by leveraging the ability to correct errors in the\nclassification process. This provides an alternative\nclassification-with-rejection method which can reduce the amount of additional\nprocessing in situations where a small number of incorrect classifications are\npermissible.\n", "date": "2023-09-12", "categories": []}
{"id": "2309.06377", "title": "Adversarial attacks on hybrid classical-quantum Deep Learning models for\n  Histopathological Cancer Detection", "abstract": "  We present an effective application of quantum machine learning in\nhistopathological cancer detection. The study here emphasizes two primary\napplications of hybrid classical-quantum Deep Learning models. The first\napplication is to build a classification model for histopathological cancer\ndetection using the quantum transfer learning strategy. The second application\nis to test the performance of this model for various adversarial attacks.\nRather than using a single transfer learning model, the hybrid\nclassical-quantum models are tested using multiple transfer learning models,\nespecially ResNet18, VGG-16, Inception-v3, and AlexNet as feature extractors\nand integrate it with several quantum circuit-based variational quantum\ncircuits (VQC) with high expressibility. As a result, we provide a comparative\nanalysis of classical models and hybrid classical-quantum transfer learning\nmodels for histopathological cancer detection under several adversarial\nattacks. We compared the performance accuracy of the classical model with the\nhybrid classical-quantum model using pennylane default quantum simulator. We\nalso observed that for histopathological cancer detection under several\nadversarial attacks, Hybrid Classical-Quantum (HCQ) models provided better\naccuracy than classical image classification models.\n", "date": "2023-09-08", "categories": []}
{"id": "2309.06438", "title": "Exploring Non-additive Randomness on ViT against Query-Based Black-Box\n  Attacks", "abstract": "  Deep Neural Networks can be easily fooled by small and imperceptible\nperturbations. The query-based black-box attack (QBBA) is able to create the\nperturbations using model output probabilities of image queries requiring no\naccess to the underlying models. QBBA poses realistic threats to real-world\napplications. Recently, various types of robustness have been explored to\ndefend against QBBA. In this work, we first taxonomize the stochastic defense\nstrategies against QBBA. Following our taxonomy, we propose to explore\nnon-additive randomness in models to defend against QBBA. Specifically, we\nfocus on underexplored Vision Transformers based on their flexible\narchitectures. Extensive experiments show that the proposed defense approach\nachieves effective defense, without much sacrifice in performance.\n", "date": "2023-09-12", "categories": []}
{"id": "2302.11704", "title": "Mitigating Adversarial Attacks in Deepfake Detection: An Exploration of\n  Perturbation and AI Techniques", "abstract": "  Deep learning constitutes a pivotal component within the realm of machine\nlearning, offering remarkable capabilities in tasks ranging from image\nrecognition to natural language processing. However, this very strength also\nrenders deep learning models susceptible to adversarial examples, a phenomenon\npervasive across a diverse array of applications. These adversarial examples\nare characterized by subtle perturbations artfully injected into clean images\nor videos, thereby causing deep learning algorithms to misclassify or produce\nerroneous outputs. This susceptibility extends beyond the confines of digital\ndomains, as adversarial examples can also be strategically designed to target\nhuman cognition, leading to the creation of deceptive media, such as deepfakes.\nDeepfakes, in particular, have emerged as a potent tool to manipulate public\nopinion and tarnish the reputations of public figures, underscoring the urgent\nneed to address the security and ethical implications associated with\nadversarial examples. This article delves into the multifaceted world of\nadversarial examples, elucidating the underlying principles behind their\ncapacity to deceive deep learning algorithms. We explore the various\nmanifestations of this phenomenon, from their insidious role in compromising\nmodel reliability to their impact in shaping the contemporary landscape of\ndisinformation and misinformation. To illustrate progress in combating\nadversarial examples, we showcase the development of a tailored Convolutional\nNeural Network (CNN) designed explicitly to detect deepfakes, a pivotal step\ntowards enhancing model robustness in the face of adversarial threats.\nImpressively, this custom CNN has achieved a precision rate of 76.2% on the\nDFDC dataset.\n", "date": "2023-02-22", "categories": []}
{"id": "2306.15931", "title": "Boosting Adversarial Transferability with Learnable Patch-wise Masks", "abstract": "  Adversarial examples have attracted widespread attention in security-critical\napplications because of their transferability across different models. Although\nmany methods have been proposed to boost adversarial transferability, a gap\nstill exists between capabilities and practical demand. In this paper, we argue\nthat the model-specific discriminative regions are a key factor causing\noverfitting to the source model, and thus reducing the transferability to the\ntarget model. For that, a patch-wise mask is utilized to prune the\nmodel-specific regions when calculating adversarial perturbations. To\naccurately localize these regions, we present a learnable approach to\nautomatically optimize the mask. Specifically, we simulate the target models in\nour framework, and adjust the patch-wise mask according to the feedback of the\nsimulated models. To improve the efficiency, the differential evolutionary (DE)\nalgorithm is utilized to search for patch-wise masks for a specific image.\nDuring iterative attacks, the learned masks are applied to the image to drop\nout the patches related to model-specific regions, thus making the gradients\nmore generic and improving the adversarial transferability. The proposed\napproach is a preprocessing method and can be integrated with existing methods\nto further boost the transferability. Extensive experiments on the ImageNet\ndataset demonstrate the effectiveness of our method. We incorporate the\nproposed approach with existing methods to perform ensemble attacks and achieve\nan average success rate of 93.01% against seven advanced defense methods, which\ncan effectively enhance the state-of-the-art transfer-based attack performance.\n", "date": "2023-06-28", "categories": []}
{"id": "2309.04650", "title": "Exploring Robust Features for Improving Adversarial Robustness", "abstract": "  While deep neural networks (DNNs) have revolutionized many fields, their\nfragility to carefully designed adversarial attacks impedes the usage of DNNs\nin safety-critical applications. In this paper, we strive to explore the robust\nfeatures which are not affected by the adversarial perturbations, i.e.,\ninvariant to the clean image and its adversarial examples, to improve the\nmodel's adversarial robustness. Specifically, we propose a feature\ndisentanglement model to segregate the robust features from non-robust features\nand domain specific features. The extensive experiments on four widely used\ndatasets with different attacks demonstrate that robust features obtained from\nour model improve the model's adversarial robustness compared to the\nstate-of-the-art approaches. Moreover, the trained domain discriminator is able\nto identify the domain specific features from the clean images and adversarial\nexamples almost perfectly. This enables adversarial example detection without\nincurring additional computational costs. With that, we can also specify\ndifferent classifiers for clean images and adversarial examples, thereby\navoiding any drop in clean image accuracy.\n", "date": "2023-09-09", "categories": []}
{"id": "2309.05132", "title": "DAD++: Improved Data-free Test Time Adversarial Defense", "abstract": "  With the increasing deployment of deep neural networks in safety-critical\napplications such as self-driving cars, medical imaging, anomaly detection,\netc., adversarial robustness has become a crucial concern in the reliability of\nthese networks in real-world scenarios. A plethora of works based on\nadversarial training and regularization-based techniques have been proposed to\nmake these deep networks robust against adversarial attacks. However, these\nmethods require either retraining models or training them from scratch, making\nthem infeasible to defend pre-trained models when access to training data is\nrestricted. To address this problem, we propose a test time Data-free\nAdversarial Defense (DAD) containing detection and correction frameworks.\nMoreover, to further improve the efficacy of the correction framework in cases\nwhen the detector is under-confident, we propose a soft-detection scheme\n(dubbed as \"DAD++\"). We conduct a wide range of experiments and ablations on\nseveral datasets and network architectures to show the efficacy of our proposed\napproach. Furthermore, we demonstrate the applicability of our approach in\nimparting adversarial defense at test time under data-free (or data-efficient)\napplications/setups, such as Data-free Knowledge Distillation and Source-free\nUnsupervised Domain Adaptation, as well as Semi-supervised classification\nframeworks. We observe that in all the experiments and applications, our DAD++\ngives an impressive performance against various adversarial attacks with a\nminimal drop in clean accuracy. The source code is available at:\nhttps://github.com/vcl-iisc/Improved-Data-free-Test-Time-Adversarial-Defense\n", "date": "2023-09-10", "categories": []}
{"id": "2309.05145", "title": "Outlier Robust Adversarial Training", "abstract": "  Supervised learning models are challenged by the intrinsic complexities of\ntraining data such as outliers and minority subpopulations and intentional\nattacks at inference time with adversarial samples. While traditional robust\nlearning methods and the recent adversarial training approaches are designed to\nhandle each of the two challenges, to date, no work has been done to develop\nmodels that are robust with regard to the low-quality training data and the\npotential adversarial attack at inference time simultaneously. It is for this\nreason that we introduce Outlier Robust Adversarial Training (ORAT) in this\nwork. ORAT is based on a bi-level optimization formulation of adversarial\ntraining with a robust rank-based loss function. Theoretically, we show that\nthe learning objective of ORAT satisfies the $\\mathcal{H}$-consistency in\nbinary classification, which establishes it as a proper surrogate to\nadversarial 0/1 loss. Furthermore, we analyze its generalization ability and\nprovide uniform convergence rates in high probability. ORAT can be optimized\nwith a simple algorithm. Experimental evaluations on three benchmark datasets\ndemonstrate the effectiveness and robustness of ORAT in handling outliers and\nadversarial attacks. Our code is available at\nhttps://github.com/discovershu/ORAT.\n", "date": "2023-09-10", "categories": []}
{"id": "2301.12277", "title": "Node Injection for Class-specific Network Poisoning", "abstract": "  Graph Neural Networks (GNNs) are powerful in learning rich network\nrepresentations that aid the performance of downstream tasks. However, recent\nstudies showed that GNNs are vulnerable to adversarial attacks involving node\ninjection and network perturbation. Among these, node injection attacks are\nmore practical as they don't require manipulation in the existing network and\ncan be performed more realistically. In this paper, we propose a novel problem\nstatement - a class-specific poison attack on graphs in which the attacker aims\nto misclassify specific nodes in the target class into a different class using\nnode injection. Additionally, nodes are injected in such a way that they\ncamouflage as benign nodes. We propose NICKI, a novel attacking strategy that\nutilizes an optimization-based approach to sabotage the performance of\nGNN-based node classifiers. NICKI works in two phases - it first learns the\nnode representation and then generates the features and edges of the injected\nnodes. Extensive experiments and ablation studies on four benchmark networks\nshow that NICKI is consistently better than four baseline attacking strategies\nfor misclassifying nodes in the target class. We also show that the injected\nnodes are properly camouflaged as benign, thus making the poisoned graph\nindistinguishable from its clean version w.r.t various topological properties.\n", "date": "2023-01-28", "categories": []}
{"id": "2209.04521", "title": "The Space of Adversarial Strategies", "abstract": "  Adversarial examples, inputs designed to induce worst-case behavior in\nmachine learning models, have been extensively studied over the past decade.\nYet, our understanding of this phenomenon stems from a rather fragmented pool\nof knowledge; at present, there are a handful of attacks, each with disparate\nassumptions in threat models and incomparable definitions of optimality. In\nthis paper, we propose a systematic approach to characterize worst-case (i.e.,\noptimal) adversaries. We first introduce an extensible decomposition of attacks\nin adversarial machine learning by atomizing attack components into surfaces\nand travelers. With our decomposition, we enumerate over components to create\n576 attacks (568 of which were previously unexplored). Next, we propose the\nPareto Ensemble Attack (PEA): a theoretical attack that upper-bounds attack\nperformance. With our new attacks, we measure performance relative to the PEA\non: both robust and non-robust models, seven datasets, and three extended\nlp-based threat models incorporating compute costs, formalizing the Space of\nAdversarial Strategies. From our evaluation we find that attack performance to\nbe highly contextual: the domain, model robustness, and threat model can have a\nprofound influence on attack efficacy. Our investigation suggests that future\nstudies measuring the security of machine learning should: (1) be\ncontextualized to the domain & threat models, and (2) go beyond the handful of\nknown attacks used today.\n", "date": "2022-09-09", "categories": []}
{"id": "2308.00452", "title": "A Majority Invariant Approach to Patch Robustness Certification for Deep\n  Learning Models", "abstract": "  Patch robustness certification ensures no patch within a given bound on a\nsample can manipulate a deep learning model to predict a different label.\nHowever, existing techniques cannot certify samples that cannot meet their\nstrict bars at the classifier or patch region levels. This paper proposes\nMajorCert. MajorCert firstly finds all possible label sets manipulatable by the\nsame patch region on the same sample across the underlying classifiers, then\nenumerates their combinations element-wise, and finally checks whether the\nmajority invariant of all these combinations is intact to certify samples.\n", "date": "2023-08-01", "categories": []}
{"id": "2309.03702", "title": "DiffDefense: Defending against Adversarial Attacks via Diffusion Models", "abstract": "  This paper presents a novel reconstruction method that leverages Diffusion\nModels to protect machine learning classifiers against adversarial attacks, all\nwithout requiring any modifications to the classifiers themselves. The\nsusceptibility of machine learning models to minor input perturbations renders\nthem vulnerable to adversarial attacks. While diffusion-based methods are\ntypically disregarded for adversarial defense due to their slow reverse\nprocess, this paper demonstrates that our proposed method offers robustness\nagainst adversarial threats while preserving clean accuracy, speed, and\nplug-and-play compatibility. Code at:\nhttps://github.com/HondamunigePrasannaSilva/DiffDefence.\n", "date": "2023-09-07", "categories": []}
{"id": "2208.06651", "title": "Revisiting Adversarial Attacks on Graph Neural Networks for Graph\n  Classification", "abstract": "  Graph neural networks (GNNs) have achieved tremendous success in the task of\ngraph classification and its diverse downstream real-world applications.\nDespite the huge success in learning graph representations, current GNN models\nhave demonstrated their vulnerability to potentially existent adversarial\nexamples on graph-structured data. Existing approaches are either limited to\nstructure attacks or restricted to local information, urging for the design of\na more general attack framework on graph classification, which faces\nsignificant challenges due to the complexity of generating local-node-level\nadversarial examples using the global-graph-level information. To address this\n\"global-to-local\" attack challenge, we present a novel and general framework to\ngenerate adversarial examples via manipulating graph structure and node\nfeatures. Specifically, we make use of Graph Class Activation Mapping and its\nvariant to produce node-level importance corresponding to the graph\nclassification task. Then through a heuristic design of algorithms, we can\nperform both feature and structure attacks under unnoticeable perturbation\nbudgets with the help of both node-level and subgraph-level importance.\nExperiments towards attacking four state-of-the-art graph classification models\non six real-world benchmarks verify the flexibility and effectiveness of our\nframework.\n", "date": "2022-08-13", "categories": []}
{"id": "2304.04512", "title": "Defense-Prefix for Preventing Typographic Attacks on CLIP", "abstract": "  Vision-language pre-training models (VLPs) have exhibited revolutionary\nimprovements in various vision-language tasks. In VLP, some adversarial attacks\nfool a model into false or absurd classifications. Previous studies addressed\nthese attacks by fine-tuning the model or changing its architecture. However,\nthese methods risk losing the original model's performance and are difficult to\napply to downstream tasks. In particular, their applicability to other tasks\nhas not been considered. In this study, we addressed the reduction of the\nimpact of typographic attacks on CLIP without changing the model parameters. To\nachieve this, we expand the idea of \"prefix learning\" and introduce our simple\nyet effective method: Defense-Prefix (DP), which inserts the DP token before a\nclass name to make words \"robust\" against typographic attacks. Our method can\nbe easily applied to downstream tasks, such as object detection, because the\nproposed method is independent of the model parameters. Our method\nsignificantly improves the accuracy of classification tasks for typographic\nattack datasets, while maintaining the zero-shot capabilities of the model. In\naddition, we leverage our proposed method for object detection, demonstrating\nits high applicability and effectiveness. The codes and datasets are available\nat https://github.com/azuma164/Defense-Prefix.\n", "date": "2023-04-10", "categories": []}
{"id": "2309.01866", "title": "Efficient Query-Based Attack against ML-Based Android Malware Detection\n  under Zero Knowledge Setting", "abstract": "  The widespread adoption of the Android operating system has made malicious\nAndroid applications an appealing target for attackers. Machine learning-based\n(ML-based) Android malware detection (AMD) methods are crucial in addressing\nthis problem; however, their vulnerability to adversarial examples raises\nconcerns. Current attacks against ML-based AMD methods demonstrate remarkable\nperformance but rely on strong assumptions that may not be realistic in\nreal-world scenarios, e.g., the knowledge requirements about feature space,\nmodel parameters, and training dataset. To address this limitation, we\nintroduce AdvDroidZero, an efficient query-based attack framework against\nML-based AMD methods that operates under the zero knowledge setting. Our\nextensive evaluation shows that AdvDroidZero is effective against various\nmainstream ML-based AMD methods, in particular, state-of-the-art such methods\nand real-world antivirus solutions.\n", "date": "2023-09-05", "categories": []}
{"id": "2309.02528", "title": "Adaptive Adversarial Training Does Not Increase Recourse Costs", "abstract": "  Recent work has connected adversarial attack methods and algorithmic recourse\nmethods: both seek minimal changes to an input instance which alter a model's\nclassification decision. It has been shown that traditional adversarial\ntraining, which seeks to minimize a classifier's susceptibility to malicious\nperturbations, increases the cost of generated recourse; with larger\nadversarial training radii correlating with higher recourse costs. From the\nperspective of algorithmic recourse, however, the appropriate adversarial\ntraining radius has always been unknown. Another recent line of work has\nmotivated adversarial training with adaptive training radii to address the\nissue of instance-wise variable adversarial vulnerability, showing success in\ndomains with unknown attack radii. This work studies the effects of adaptive\nadversarial training on algorithmic recourse costs. We establish that the\nimprovements in model robustness induced by adaptive adversarial training show\nlittle effect on algorithmic recourse costs, providing a potential avenue for\naffordable robustness in domains where recoursability is critical.\n", "date": "2023-09-05", "categories": []}
{"id": "2309.02752", "title": "SWAP: Exploiting Second-Ranked Logits for Adversarial Attacks on Time\n  Series", "abstract": "  Time series classification (TSC) has emerged as a critical task in various\ndomains, and deep neural models have shown superior performance in TSC tasks.\nHowever, these models are vulnerable to adversarial attacks, where subtle\nperturbations can significantly impact the prediction results. Existing\nadversarial methods often suffer from over-parameterization or random logit\nperturbation, hindering their effectiveness. Additionally, increasing the\nattack success rate (ASR) typically involves generating more noise, making the\nattack more easily detectable. To address these limitations, we propose SWAP, a\nnovel attacking method for TSC models. SWAP focuses on enhancing the confidence\nof the second-ranked logits while minimizing the manipulation of other logits.\nThis is achieved by minimizing the Kullback-Leibler divergence between the\ntarget logit distribution and the predictive logit distribution. Experimental\nresults demonstrate that SWAP achieves state-of-the-art performance, with an\nASR exceeding 50% and an 18% increase compared to existing methods.\n", "date": "2023-09-06", "categories": []}
{"id": "2309.03164", "title": "J-Guard: Journalism Guided Adversarially Robust Detection of\n  AI-generated News", "abstract": "  The rapid proliferation of AI-generated text online is profoundly reshaping\nthe information landscape. Among various types of AI-generated text,\nAI-generated news presents a significant threat as it can be a prominent source\nof misinformation online. While several recent efforts have focused on\ndetecting AI-generated text in general, these methods require enhanced\nreliability, given concerns about their vulnerability to simple adversarial\nattacks. Furthermore, due to the eccentricities of news writing, applying these\ndetection methods for AI-generated news can produce false positives,\npotentially damaging the reputation of news organizations. To address these\nchallenges, we leverage the expertise of an interdisciplinary team to develop a\nframework, J-Guard, capable of steering existing supervised AI text detectors\nfor detecting AI-generated news while boosting adversarial robustness. By\nincorporating stylistic cues inspired by the unique journalistic attributes,\nJ-Guard effectively distinguishes between real-world journalism and\nAI-generated news articles. Our experiments on news articles generated by a\nvast array of AI models, including ChatGPT (GPT3.5), demonstrate the\neffectiveness of J-Guard in enhancing detection capabilities while maintaining\nan average performance decrease of as low as 7% when faced with adversarial\nattacks.\n", "date": "2023-09-06", "categories": []}
{"id": "2309.03198", "title": "My Art My Choice: Adversarial Protection Against Unruly AI", "abstract": "  Generative AI is on the rise, enabling everyone to produce realistic content\nvia publicly available interfaces. Especially for guided image generation,\ndiffusion models are changing the creator economy by producing high quality low\ncost content. In parallel, artists are rising against unruly AI, since their\nartwork are leveraged, distributed, and dissimulated by large generative\nmodels. Our approach, My Art My Choice (MAMC), aims to empower content owners\nby protecting their copyrighted materials from being utilized by diffusion\nmodels in an adversarial fashion. MAMC learns to generate adversarially\nperturbed \"protected\" versions of images which can in turn \"break\" diffusion\nmodels. The perturbation amount is decided by the artist to balance distortion\nvs. protection of the content. MAMC is designed with a simple UNet-based\ngenerator, attacking black box diffusion models, combining several losses to\ncreate adversarial twins of the original artwork. We experiment on three\ndatasets for various image-to-image tasks, with different user control values.\nBoth protected image and diffusion output results are evaluated in visual,\nnoise, structure, pixel, and generative spaces to validate our claims. We\nbelieve that MAMC is a crucial step for preserving ownership information for AI\ngenerated content in a flawless, based-on-need, and human-centric way.\n", "date": "2023-09-06", "categories": []}
{"id": "2206.08316", "title": "Boosting the Adversarial Transferability of Surrogate Models with Dark\n  Knowledge", "abstract": "  Deep neural networks (DNNs) are vulnerable to adversarial examples. And, the\nadversarial examples have transferability, which means that an adversarial\nexample for a DNN model can fool another model with a non-trivial probability.\nThis gave birth to the transfer-based attack where the adversarial examples\ngenerated by a surrogate model are used to conduct black-box attacks. There are\nsome work on generating the adversarial examples from a given surrogate model\nwith better transferability. However, training a special surrogate model to\ngenerate adversarial examples with better transferability is relatively\nunder-explored. This paper proposes a method for training a surrogate model\nwith dark knowledge to boost the transferability of the adversarial examples\ngenerated by the surrogate model. This trained surrogate model is named dark\nsurrogate model (DSM). The proposed method for training a DSM consists of two\nkey components: a teacher model extracting dark knowledge, and the mixing\naugmentation skill enhancing dark knowledge of training data. We conducted\nextensive experiments to show that the proposed method can substantially\nimprove the adversarial transferability of surrogate models across different\narchitectures of surrogate models and optimizers for generating adversarial\nexamples, and it can be applied to other scenarios of transfer-based attack\nthat contain dark knowledge, like face verification. Our code is publicly\navailable at \\url{https://github.com/ydc123/Dark_Surrogate_Model}.\n", "date": "2022-06-16", "categories": []}
{"id": "2209.02339", "title": "TransCAB: Transferable Clean-Annotation Backdoor to Object Detection\n  with Natural Trigger in Real-World", "abstract": "  Object detection is the foundation of various critical computer-vision tasks\nsuch as segmentation, object tracking, and event detection. To train an object\ndetector with satisfactory accuracy, a large amount of data is required.\nHowever, due to the intensive workforce involved with annotating large\ndatasets, such a data curation task is often outsourced to a third party or\nrelied on volunteers. This work reveals severe vulnerabilities of such data\ncuration pipeline. We propose MACAB that crafts clean-annotated images to\nstealthily implant the backdoor into the object detectors trained on them even\nwhen the data curator can manually audit the images. We observe that the\nbackdoor effect of both misclassification and the cloaking are robustly\nachieved in the wild when the backdoor is activated with inconspicuously\nnatural physical triggers. Backdooring non-classification object detection with\nclean-annotation is challenging compared to backdooring existing image\nclassification tasks with clean-label, owing to the complexity of having\nmultiple objects within each frame, including victim and non-victim objects.\nThe efficacy of the MACAB is ensured by constructively i abusing the\nimage-scaling function used by the deep learning framework, ii incorporating\nthe proposed adversarial clean image replica technique, and iii combining\npoison data selection criteria given constrained attacking budget. Extensive\nexperiments demonstrate that MACAB exhibits more than 90% attack success rate\nunder various real-world scenes. This includes both cloaking and\nmisclassification backdoor effect even restricted with a small attack budget.\nThe poisoned samples cannot be effectively identified by state-of-the-art\ndetection techniques.The comprehensive video demo is at\nhttps://youtu.be/MA7L_LpXkp4, which is based on a poison rate of 0.14% for\nYOLOv4 cloaking backdoor and Faster R-CNN misclassification backdoor.\n", "date": "2022-09-06", "categories": []}
{"id": "2307.16489", "title": "BAGM: A Backdoor Attack for Manipulating Text-to-Image Generative Models", "abstract": "  The rise in popularity of text-to-image generative artificial intelligence\n(AI) has attracted widespread public interest. We demonstrate that this\ntechnology can be attacked to generate content that subtly manipulates its\nusers. We propose a Backdoor Attack on text-to-image Generative Models (BAGM),\nwhich upon triggering, infuses the generated images with manipulative details\nthat are naturally blended in the content. Our attack is the first to target\nthree popular text-to-image generative models across three stages of the\ngenerative process by modifying the behaviour of the embedded tokenizer, the\nlanguage model or the image generative model. Based on the penetration level,\nBAGM takes the form of a suite of attacks that are referred to as surface,\nshallow and deep attacks in this article. Given the existing gap within this\ndomain, we also contribute a comprehensive set of quantitative metrics designed\nspecifically for assessing the effectiveness of backdoor attacks on\ntext-to-image models. The efficacy of BAGM is established by attacking\nstate-of-the-art generative models, using a marketing scenario as the target\ndomain. To that end, we contribute a dataset of branded product images. Our\nembedded backdoors increase the bias towards the target outputs by more than\nfive times the usual, without compromising the model robustness or the\ngenerated content utility. By exposing generative AI's vulnerabilities, we\nencourage researchers to tackle these challenges and practitioners to exercise\ncaution when using pre-trained models. Relevant code, input prompts and\nsupplementary material can be found at https://github.com/JJ-Vice/BAGM, and the\ndataset is available at:\nhttps://ieee-dataport.org/documents/marketable-foods-mf-dataset.\n  Keywords: Generative Artificial Intelligence, Generative Models,\nText-to-Image generation, Backdoor Attacks, Trojan, Stable Diffusion.\n", "date": "2023-07-31", "categories": []}
{"id": "2308.15673", "title": "MDTD: A Multi Domain Trojan Detector for Deep Neural Networks", "abstract": "  Machine learning models that use deep neural networks (DNNs) are vulnerable\nto backdoor attacks. An adversary carrying out a backdoor attack embeds a\npredefined perturbation called a trigger into a small subset of input samples\nand trains the DNN such that the presence of the trigger in the input results\nin an adversary-desired output class. Such adversarial retraining however needs\nto ensure that outputs for inputs without the trigger remain unaffected and\nprovide high classification accuracy on clean samples. In this paper, we\npropose MDTD, a Multi-Domain Trojan Detector for DNNs, which detects inputs\ncontaining a Trojan trigger at testing time. MDTD does not require knowledge of\ntrigger-embedding strategy of the attacker and can be applied to a pre-trained\nDNN model with image, audio, or graph-based inputs. MDTD leverages an insight\nthat input samples containing a Trojan trigger are located relatively farther\naway from a decision boundary than clean samples. MDTD estimates the distance\nto a decision boundary using adversarial learning methods and uses this\ndistance to infer whether a test-time input sample is Trojaned or not. We\nevaluate MDTD against state-of-the-art Trojan detection methods across five\nwidely used image-based datasets: CIFAR100, CIFAR10, GTSRB, SVHN, and\nFlowers102; four graph-based datasets: AIDS, WinMal, Toxicant, and COLLAB; and\nthe SpeechCommand audio dataset. MDTD effectively identifies samples that\ncontain different types of Trojan triggers. We evaluate MDTD against adaptive\nattacks where an adversary trains a robust DNN to increase (decrease) distance\nof benign (Trojan) inputs from a decision boundary.\n", "date": "2023-08-30", "categories": []}
{"id": "2309.00007", "title": "When Measures are Unreliable: Imperceptible Adversarial Perturbations\n  toward Top-$k$ Multi-Label Learning", "abstract": "  With the great success of deep neural networks, adversarial learning has\nreceived widespread attention in various studies, ranging from multi-class\nlearning to multi-label learning. However, existing adversarial attacks toward\nmulti-label learning only pursue the traditional visual imperceptibility but\nignore the new perceptible problem coming from measures such as Precision@$k$\nand mAP@$k$. Specifically, when a well-trained multi-label classifier performs\nfar below the expectation on some samples, the victim can easily realize that\nthis performance degeneration stems from attack, rather than the model itself.\nTherefore, an ideal multi-labeling adversarial attack should manage to not only\ndeceive visual perception but also evade monitoring of measures. To this end,\nthis paper first proposes the concept of measure imperceptibility. Then, a\nnovel loss function is devised to generate such adversarial perturbations that\ncould achieve both visual and measure imperceptibility. Furthermore, an\nefficient algorithm, which enjoys a convex objective, is established to\noptimize this objective. Finally, extensive experiments on large-scale\nbenchmark datasets, such as PASCAL VOC 2012, MS COCO, and NUS WIDE, demonstrate\nthe superiority of our proposed method in attacking the top-$k$ multi-label\nsystems.\n", "date": "2023-07-27", "categories": []}
{"id": "2309.00614", "title": "Baseline Defenses for Adversarial Attacks Against Aligned Language\n  Models", "abstract": "  As Large Language Models quickly become ubiquitous, it becomes critical to\nunderstand their security vulnerabilities. Recent work shows that text\noptimizers can produce jailbreaking prompts that bypass moderation and\nalignment. Drawing from the rich body of work on adversarial machine learning,\nwe approach these attacks with three questions: What threat models are\npractically useful in this domain? How do baseline defense techniques perform\nin this new domain? How does LLM security differ from computer vision?\n  We evaluate several baseline defense strategies against leading adversarial\nattacks on LLMs, discussing the various settings in which each is feasible and\neffective. Particularly, we look at three types of defenses: detection\n(perplexity based), input preprocessing (paraphrase and retokenization), and\nadversarial training. We discuss white-box and gray-box settings and discuss\nthe robustness-performance trade-off for each of the defenses considered. We\nfind that the weakness of existing discrete optimizers for text, combined with\nthe relatively high costs of optimization, makes standard adaptive attacks more\nchallenging for LLMs. Future research will be needed to uncover whether more\npowerful optimizers can be developed, or whether the strength of filtering and\npreprocessing defenses is greater in the LLMs domain than it has been in\ncomputer vision.\n", "date": "2023-09-01", "categories": []}
{"id": "2309.00771", "title": "Non-Asymptotic Bounds for Adversarial Excess Risk under Misspecified\n  Models", "abstract": "  We propose a general approach to evaluating the performance of robust\nestimators based on adversarial losses under misspecified models. We first show\nthat adversarial risk is equivalent to the risk induced by a distributional\nadversarial attack under certain smoothness conditions. This ensures that the\nadversarial training procedure is well-defined. To evaluate the generalization\nperformance of the adversarial estimator, we study the adversarial excess risk.\nOur proposed analysis method includes investigations on both generalization\nerror and approximation error. We then establish non-asymptotic upper bounds\nfor the adversarial excess risk associated with Lipschitz loss functions. In\naddition, we apply our general results to adversarial training for\nclassification and regression problems. For the quadratic loss in nonparametric\nregression, we show that the adversarial excess risk bound can be improved over\nthose for a general loss.\n", "date": "2023-09-02", "categories": []}
{"id": "2309.00879", "title": "Towards Certified Probabilistic Robustness with High Accuracy", "abstract": "  Adversarial examples pose a security threat to many critical systems built on\nneural networks (such as face recognition systems, and self-driving cars).\nWhile many methods have been proposed to build robust models, how to build\ncertifiably robust yet accurate neural network models remains an open problem.\nFor example, adversarial training improves empirical robustness, but they do\nnot provide certification of the model's robustness. On the other hand,\ncertified training provides certified robustness but at the cost of a\nsignificant accuracy drop. In this work, we propose a novel approach that aims\nto achieve both high accuracy and certified probabilistic robustness. Our\nmethod has two parts, i.e., a probabilistic robust training method with an\nadditional goal of minimizing variance in terms of divergence and a runtime\ninference method for certified probabilistic robustness of the prediction. The\nlatter enables efficient certification of the model's probabilistic robustness\nat runtime with statistical guarantees. This is supported by our training\nobjective, which minimizes the variance of the model's predictions in a given\nvicinity, derived from a general definition of model robustness. Our approach\nworks for a variety of perturbations and is reasonably efficient. Our\nexperiments on multiple models trained on different datasets demonstrate that\nour approach significantly outperforms existing approaches in terms of both\ncertification rate and accuracy.\n", "date": "2023-09-02", "categories": []}
{"id": "2309.01077", "title": "Robust Adversarial Defense by Tensor Factorization", "abstract": "  As machine learning techniques become increasingly prevalent in data\nanalysis, the threat of adversarial attacks has surged, necessitating robust\ndefense mechanisms. Among these defenses, methods exploiting low-rank\napproximations for input data preprocessing and neural network (NN) parameter\nfactorization have shown potential. Our work advances this field further by\nintegrating the tensorization of input data with low-rank decomposition and\ntensorization of NN parameters to enhance adversarial defense. The proposed\napproach demonstrates significant defense capabilities, maintaining robust\naccuracy even when subjected to the strongest known auto-attacks. Evaluations\nagainst leading-edge robust performance benchmarks reveal that our results not\nonly hold their ground against the best defensive methods available but also\nexceed all current defense strategies that rely on tensor factorizations. This\nstudy underscores the potential of integrating tensorization and low-rank\ndecomposition as a robust defense against adversarial attacks in machine\nlearning.\n", "date": "2023-09-03", "categories": []}
{"id": "2309.01099", "title": "Enhancing Infrared Small Target Detection Robustness with Bi-Level\n  Adversarial Framework", "abstract": "  The detection of small infrared targets against blurred and cluttered\nbackgrounds has remained an enduring challenge. In recent years, learning-based\nschemes have become the mainstream methodology to establish the mapping\ndirectly. However, these methods are susceptible to the inherent complexities\nof changing backgrounds and real-world disturbances, leading to unreliable and\ncompromised target estimations. In this work, we propose a bi-level adversarial\nframework to promote the robustness of detection in the presence of distinct\ncorruptions. We first propose a bi-level optimization formulation to introduce\ndynamic adversarial learning. Specifically, it is composited by the learnable\ngeneration of corruptions to maximize the losses as the lower-level objective\nand the robustness promotion of detectors as the upper-level one. We also\nprovide a hierarchical reinforced learning strategy to discover the most\ndetrimental corruptions and balance the performance between robustness and\naccuracy. To better disentangle the corruptions from salient features, we also\npropose a spatial-frequency interaction network for target detection. Extensive\nexperiments demonstrate our scheme remarkably improves 21.96% IOU across a wide\narray of corruptions and notably promotes 4.97% IOU on the general benchmark.\nThe source codes are available at https://github.com/LiuZhu-CV/BALISTD.\n", "date": "2023-09-03", "categories": []}
{"id": "2309.01106", "title": "AdvMono3D: Advanced Monocular 3D Object Detection with Depth-Aware\n  Robust Adversarial Training", "abstract": "  Monocular 3D object detection plays a pivotal role in the field of autonomous\ndriving and numerous deep learning-based methods have made significant\nbreakthroughs in this area. Despite the advancements in detection accuracy and\nefficiency, these models tend to fail when faced with such attacks, rendering\nthem ineffective. Therefore, bolstering the adversarial robustness of 3D\ndetection models has become a crucial issue that demands immediate attention\nand innovative solutions. To mitigate this issue, we propose a depth-aware\nrobust adversarial training method for monocular 3D object detection, dubbed\nDART3D. Specifically, we first design an adversarial attack that iteratively\ndegrades the 2D and 3D perception capabilities of 3D object detection\nmodels(IDP), serves as the foundation for our subsequent defense mechanism. In\nresponse to this attack, we propose an uncertainty-based residual learning\nmethod for adversarial training. Our adversarial training approach capitalizes\non the inherent uncertainty, enabling the model to significantly improve its\nrobustness against adversarial attacks. We conducted extensive experiments on\nthe KITTI 3D datasets, demonstrating that DART3D surpasses direct adversarial\ntraining (the most popular approach) under attacks in 3D object detection\n$AP_{R40}$ of car category for the Easy, Moderate, and Hard settings, with\nimprovements of 4.415%, 4.112%, and 3.195%, respectively.\n", "date": "2023-09-03", "categories": []}
{"id": "2309.01452", "title": "Toward Defensive Letter Design", "abstract": "  A major approach for defending against adversarial attacks aims at\ncontrolling only image classifiers to be more resilient, and it does not care\nabout visual objects, such as pandas and cars, in images. This means that\nvisual objects themselves cannot take any defensive actions, and they are still\nvulnerable to adversarial attacks. In contrast, letters are artificial symbols,\nand we can freely control their appearance unless losing their readability. In\nother words, we can make the letters more defensive to the attacks. This paper\nposes three research questions related to the adversarial vulnerability of\nletter images: (1) How defensive are the letters against adversarial attacks?\n(2) Can we estimate how defensive a given letter image is before attacks? (3)\nCan we control the letter images to be more defensive against adversarial\nattacks? For answering the first and second questions, we measure the\ndefensibility of letters by employing Iterative Fast Gradient Sign Method\n(I-FGSM) and then build a deep regression model for estimating the\ndefensibility of each letter image. We also propose a two-step method based on\na generative adversarial network (GAN) for generating character images with\nhigher defensibility, which solves the third research question.\n", "date": "2023-09-04", "categories": []}
{"id": "2309.01614", "title": "Dropout Attacks", "abstract": "  Dropout is a common operator in deep learning, aiming to prevent overfitting\nby randomly dropping neurons during training. This paper introduces a new\nfamily of poisoning attacks against neural networks named DROPOUTATTACK.\nDROPOUTATTACK attacks the dropout operator by manipulating the selection of\nneurons to drop instead of selecting them uniformly at random. We design,\nimplement, and evaluate four DROPOUTATTACK variants that cover a broad range of\nscenarios. These attacks can slow or stop training, destroy prediction accuracy\nof target classes, and sabotage either precision or recall of a target class.\nIn our experiments of training a VGG-16 model on CIFAR-100, our attack can\nreduce the precision of the victim class by 34.6% (from 81.7% to 47.1%) without\nincurring any degradation in model accuracy\n", "date": "2023-09-04", "categories": []}
{"id": "2309.01620", "title": "Hindering Adversarial Attacks with Multiple Encrypted Patch Embeddings", "abstract": "  In this paper, we propose a new key-based defense focusing on both efficiency\nand robustness. Although the previous key-based defense seems effective in\ndefending against adversarial examples, carefully designed adaptive attacks can\nbypass the previous defense, and it is difficult to train the previous defense\non large datasets like ImageNet. We build upon the previous defense with two\nmajor improvements: (1) efficient training and (2) optional randomization. The\nproposed defense utilizes one or more secret patch embeddings and classifier\nheads with a pre-trained isotropic network. When more than one secret\nembeddings are used, the proposed defense enables randomization on inference.\nExperiments were carried out on the ImageNet dataset, and the proposed defense\nwas evaluated against an arsenal of state-of-the-art attacks, including\nadaptive ones. The results show that the proposed defense achieves a high\nrobust accuracy and a comparable clean accuracy compared to the previous\nkey-based defense.\n", "date": "2023-09-04", "categories": []}
{"id": "2309.01686", "title": "MathAttack: Attacking Large Language Models Towards Math Solving Ability", "abstract": "  With the boom of Large Language Models (LLMs), the research of solving Math\nWord Problem (MWP) has recently made great progress. However, there are few\nstudies to examine the security of LLMs in math solving ability. Instead of\nattacking prompts in the use of LLMs, we propose a MathAttack model to attack\nMWP samples which are closer to the essence of security in solving math\nproblems. Compared to traditional text adversarial attack, it is essential to\npreserve the mathematical logic of original MWPs during the attacking. To this\nend, we propose logical entity recognition to identify logical entries which\nare then frozen. Subsequently, the remaining text are attacked by adopting a\nword-level attacker. Furthermore, we propose a new dataset RobustMath to\nevaluate the robustness of LLMs in math solving ability. Extensive experiments\non our RobustMath and two another math benchmark datasets GSM8K and MultiAirth\nshow that MathAttack could effectively attack the math solving ability of LLMs.\nIn the experiments, we observe that (1) Our adversarial samples from\nhigher-accuracy LLMs are also effective for attacking LLMs with lower accuracy\n(e.g., transfer from larger to smaller-size LLMs, or from few-shot to zero-shot\nprompts); (2) Complex MWPs (such as more solving steps, longer text, more\nnumbers) are more vulnerable to attack; (3) We can improve the robustness of\nLLMs by using our adversarial samples in few-shot prompts. Finally, we hope our\npractice and observation can serve as an important attempt towards enhancing\nthe robustness of LLMs in math solving ability. We will release our code and\ndataset.\n", "date": "2023-09-04", "categories": []}
{"id": "2309.02159", "title": "The Adversarial Implications of Variable-Time Inference", "abstract": "  Machine learning (ML) models are known to be vulnerable to a number of\nattacks that target the integrity of their predictions or the privacy of their\ntraining data. To carry out these attacks, a black-box adversary must typically\npossess the ability to query the model and observe its outputs (e.g., labels).\nIn this work, we demonstrate, for the first time, the ability to enhance such\ndecision-based attacks. To accomplish this, we present an approach that\nexploits a novel side channel in which the adversary simply measures the\nexecution time of the algorithm used to post-process the predictions of the ML\nmodel under attack. The leakage of inference-state elements into algorithmic\ntiming side channels has never been studied before, and we have found that it\ncan contain rich information that facilitates superior timing attacks that\nsignificantly outperform attacks based solely on label outputs. In a case\nstudy, we investigate leakage from the non-maximum suppression (NMS) algorithm,\nwhich plays a crucial role in the operation of object detectors. In our\nexamination of the timing side-channel vulnerabilities associated with this\nalgorithm, we identified the potential to enhance decision-based attacks. We\ndemonstrate attacks against the YOLOv3 detector, leveraging the timing leakage\nto successfully evade object detection using adversarial examples, and perform\ndataset inference. Our experiments show that our adversarial examples exhibit\nsuperior perturbation quality compared to a decision-based attack. In addition,\nwe present a new threat model in which dataset inference based solely on timing\nleakage is performed. To address the timing leakage vulnerability inherent in\nthe NMS algorithm, we explore the potential and limitations of implementing\nconstant-time inference passes as a mitigation strategy.\n", "date": "2023-09-05", "categories": []}
{"id": "2308.16258", "title": "Robust Principles: Architectural Design Principles for Adversarially\n  Robust CNNs", "abstract": "  Our research aims to unify existing works' diverging opinions on how\narchitectural components affect the adversarial robustness of CNNs. To\naccomplish our goal, we synthesize a suite of three generalizable robust\narchitectural design principles: (a) optimal range for depth and width\nconfigurations, (b) preferring convolutional over patchify stem stage, and (c)\nrobust residual block design through adopting squeeze and excitation blocks and\nnon-parametric smooth activation functions. Through extensive experiments\nacross a wide spectrum of dataset scales, adversarial training methods, model\nparameters, and network design spaces, our principles consistently and markedly\nimprove AutoAttack accuracy: 1-3 percentage points (pp) on CIFAR-10 and\nCIFAR-100, and 4-9 pp on ImageNet. The code is publicly available at\nhttps://github.com/poloclub/robust-principles.\n", "date": "2023-08-30", "categories": []}
{"id": "2309.00254", "title": "Why do universal adversarial attacks work on large language models?:\n  Geometry might be the answer", "abstract": "  Transformer based large language models with emergent capabilities are\nbecoming increasingly ubiquitous in society. However, the task of understanding\nand interpreting their internal workings, in the context of adversarial\nattacks, remains largely unsolved. Gradient-based universal adversarial attacks\nhave been shown to be highly effective on large language models and potentially\ndangerous due to their input-agnostic nature. This work presents a novel\ngeometric perspective explaining universal adversarial attacks on large\nlanguage models. By attacking the 117M parameter GPT-2 model, we find evidence\nindicating that universal adversarial triggers could be embedding vectors which\nmerely approximate the semantic information in their adversarial training\nregion. This hypothesis is supported by white-box model analysis comprising\ndimensionality reduction and similarity measurement of hidden representations.\nWe believe this new geometric perspective on the underlying mechanism driving\nuniversal attacks could help us gain deeper insight into the internal workings\nand failure modes of LLMs, thus enabling their mitigation.\n", "date": "2023-09-01", "categories": []}
{"id": "2301.04785", "title": "Phase-shifted Adversarial Training", "abstract": "  Adversarial training has been considered an imperative component for safely\ndeploying neural network-based applications to the real world. To achieve\nstronger robustness, existing methods primarily focus on how to generate strong\nattacks by increasing the number of update steps, regularizing the models with\nthe smoothed loss function, and injecting the randomness into the attack.\nInstead, we analyze the behavior of adversarial training through the lens of\nresponse frequency. We empirically discover that adversarial training causes\nneural networks to have low convergence to high-frequency information,\nresulting in highly oscillated predictions near each data. To learn\nhigh-frequency contents efficiently and effectively, we first prove that a\nuniversal phenomenon of frequency principle, i.e., \\textit{lower frequencies\nare learned first}, still holds in adversarial training. Based on that, we\npropose phase-shifted adversarial training (PhaseAT) in which the model learns\nhigh-frequency components by shifting these frequencies to the low-frequency\nrange where the fast convergence occurs. For evaluations, we conduct the\nexperiments on CIFAR-10 and ImageNet with the adaptive attack carefully\ndesigned for reliable evaluation. Comprehensive results show that PhaseAT\nsignificantly improves the convergence for high-frequency information. This\nresults in improved adversarial robustness by enabling the model to have\nsmoothed predictions near each data.\n", "date": "2023-01-12", "categories": []}
{"id": "2308.16454", "title": "Adversarial Finetuning with Latent Representation Constraint to Mitigate\n  Accuracy-Robustness Tradeoff", "abstract": "  This paper addresses the tradeoff between standard accuracy on clean examples\nand robustness against adversarial examples in deep neural networks (DNNs).\nAlthough adversarial training (AT) improves robustness, it degrades the\nstandard accuracy, thus yielding the tradeoff. To mitigate this tradeoff, we\npropose a novel AT method called ARREST, which comprises three components: (i)\nadversarial finetuning (AFT), (ii) representation-guided knowledge distillation\n(RGKD), and (iii) noisy replay (NR). AFT trains a DNN on adversarial examples\nby initializing its parameters with a DNN that is standardly pretrained on\nclean examples. RGKD and NR respectively entail a regularization term and an\nalgorithm to preserve latent representations of clean examples during AFT. RGKD\npenalizes the distance between the representations of the standardly pretrained\nand AFT DNNs. NR switches input adversarial examples to nonadversarial ones\nwhen the representation changes significantly during AFT. By combining these\ncomponents, ARREST achieves both high standard accuracy and robustness.\nExperimental results demonstrate that ARREST mitigates the tradeoff more\neffectively than previous AT-based methods do.\n", "date": "2023-08-31", "categories": []}
{"id": "2308.15550", "title": "Adversarial Style Transfer for Robust Policy Optimization in Deep\n  Reinforcement Learning", "abstract": "  This paper proposes an algorithm that aims to improve generalization for\nreinforcement learning agents by removing overfitting to confounding features.\nOur approach consists of a max-min game theoretic objective. A generator\ntransfers the style of observation during reinforcement learning. An additional\ngoal of the generator is to perturb the observation, which maximizes the\nagent's probability of taking a different action. In contrast, a policy network\nupdates its parameters to minimize the effect of such perturbations, thus\nstaying robust while maximizing the expected future reward. Based on this\nsetup, we propose a practical deep reinforcement learning algorithm,\nAdversarial Robust Policy Optimization (ARPO), to find a robust policy that\ngeneralizes to unseen environments. We evaluate our approach on Procgen and\nDistracting Control Suite for generalization and sample efficiency.\nEmpirically, ARPO shows improved performance compared to a few baseline\nalgorithms, including data augmentation.\n", "date": "2023-08-29", "categories": []}
{"id": "2308.15614", "title": "Everything Perturbed All at Once: Enabling Differentiable Graph Attacks", "abstract": "  As powerful tools for representation learning on graphs, graph neural\nnetworks (GNNs) have played an important role in applications including social\nnetworks, recommendation systems, and online web services. However, GNNs have\nbeen shown to be vulnerable to adversarial attacks, which can significantly\ndegrade their effectiveness. Recent state-of-the-art approaches in adversarial\nattacks rely on gradient-based meta-learning to selectively perturb a single\nedge with the highest attack score until they reach the budget constraint.\nWhile effective in identifying vulnerable links, these methods are plagued by\nhigh computational costs. By leveraging continuous relaxation and\nparameterization of the graph structure, we propose a novel attack method\ncalled Differentiable Graph Attack (DGA) to efficiently generate effective\nattacks and meanwhile eliminate the need for costly retraining. Compared to the\nstate-of-the-art, DGA achieves nearly equivalent attack performance with 6\ntimes less training time and 11 times smaller GPU memory footprint on different\nbenchmark datasets. Additionally, we provide extensive experimental analyses of\nthe transferability of the DGA among different graph models, as well as its\nrobustness against widely-used defense mechanisms.\n", "date": "2023-08-29", "categories": []}
{"id": "2302.12095", "title": "On the Robustness of ChatGPT: An Adversarial and Out-of-distribution\n  Perspective", "abstract": "  ChatGPT is a recent chatbot service released by OpenAI and is receiving\nincreasing attention over the past few months. While evaluations of various\naspects of ChatGPT have been done, its robustness, i.e., the performance to\nunexpected inputs, is still unclear to the public. Robustness is of particular\nconcern in responsible AI, especially for safety-critical applications. In this\npaper, we conduct a thorough evaluation of the robustness of ChatGPT from the\nadversarial and out-of-distribution (OOD) perspective. To do so, we employ the\nAdvGLUE and ANLI benchmarks to assess adversarial robustness and the Flipkart\nreview and DDXPlus medical diagnosis datasets for OOD evaluation. We select\nseveral popular foundation models as baselines. Results show that ChatGPT shows\nconsistent advantages on most adversarial and OOD classification and\ntranslation tasks. However, the absolute performance is far from perfection,\nwhich suggests that adversarial and OOD robustness remains a significant threat\nto foundation models. Moreover, ChatGPT shows astounding performance in\nunderstanding dialogue-related texts and we find that it tends to provide\ninformal suggestions for medical tasks instead of definitive answers. Finally,\nwe present in-depth discussions of possible research directions.\n", "date": "2023-02-22", "categories": []}
{"id": "2308.15072", "title": "Advancing Adversarial Robustness Through Adversarial Logit Update", "abstract": "  Deep Neural Networks are susceptible to adversarial perturbations.\nAdversarial training and adversarial purification are among the most widely\nrecognized defense strategies. Although these methods have different underlying\nlogic, both rely on absolute logit values to generate label predictions. In\nthis study, we theoretically analyze the logit difference around successful\nadversarial attacks from a theoretical point of view and propose a new\nprinciple, namely Adversarial Logit Update (ALU), to infer adversarial sample's\nlabels. Based on ALU, we introduce a new classification paradigm that utilizes\npre- and post-purification logit differences for model's adversarial robustness\nboost. Without requiring adversarial or additional data for model training, our\nclean data synthesis model can be easily applied to various pre-trained models\nfor both adversarial sample detection and ALU-based data classification.\nExtensive experiments on both CIFAR-10, CIFAR-100, and tiny-ImageNet datasets\nshow that even with simple components, the proposed solution achieves superior\nrobustness performance compared to state-of-the-art methods against a wide\nrange of adversarial attacks. Our python implementation is submitted in our\nSupplementary document and will be published upon the paper's acceptance.\n", "date": "2023-08-29", "categories": []}
{"id": "2308.15092", "title": "Can We Rely on AI?", "abstract": "  Over the last decade, adversarial attack algorithms have revealed\ninstabilities in deep learning tools. These algorithms raise issues regarding\nsafety, reliability and interpretability in artificial intelligence; especially\nin high risk settings. From a practical perspective, there has been a war of\nescalation between those developing attack and defence strategies. At a more\ntheoretical level, researchers have also studied bigger picture questions\nconcerning the existence and computability of attacks. Here we give a brief\noverview of the topic, focusing on aspects that are likely to be of interest to\nresearchers in applied and computational mathematics.\n", "date": "2023-08-29", "categories": []}
{"id": "2308.15344", "title": "Imperceptible Adversarial Attack on Deep Neural Networks from Image\n  Boundary", "abstract": "  Although Deep Neural Networks (DNNs), such as the convolutional neural\nnetworks (CNN) and Vision Transformers (ViTs), have been successfully applied\nin the field of computer vision, they are demonstrated to be vulnerable to\nwell-sought Adversarial Examples (AEs) that can easily fool the DNNs. The\nresearch in AEs has been active, and many adversarial attacks and explanations\nhave been proposed since they were discovered in 2014. The mystery of the AE's\nexistence is still an open question, and many studies suggest that DNN training\nalgorithms have blind spots. The salient objects usually do not overlap with\nboundaries; hence, the boundaries are not the DNN model's attention.\nNevertheless, recent studies show that the boundaries can dominate the behavior\nof the DNN models. Hence, this study aims to look at the AEs from a different\nperspective and proposes an imperceptible adversarial attack that systemically\nattacks the input image boundary for finding the AEs. The experimental results\nhave shown that the proposed boundary attacking method effectively attacks six\nCNN models and the ViT using only 32% of the input image content (from the\nboundaries) with an average success rate (SR) of 95.2% and an average peak\nsignal-to-noise ratio of 41.37 dB. Correlation analyses are conducted,\nincluding the relation between the adversarial boundary's width and the SR and\nhow the adversarial boundary changes the DNN model's attention. This paper's\ndiscoveries can potentially advance the understanding of AEs and provide a\ndifferent perspective on how AEs can be constructed.\n", "date": "2023-08-29", "categories": []}
{"id": "2209.15304", "title": "Hiding Visual Information via Obfuscating Adversarial Perturbations", "abstract": "  Growing leakage and misuse of visual information raise security and privacy\nconcerns, which promotes the development of information protection. Existing\nadversarial perturbations-based methods mainly focus on the de-identification\nagainst deep learning models. However, the inherent visual information of the\ndata has not been well protected. In this work, inspired by the Type-I\nadversarial attack, we propose an adversarial visual information hiding method\nto protect the visual privacy of data. Specifically, the method generates\nobfuscating adversarial perturbations to obscure the visual information of the\ndata. Meanwhile, it maintains the hidden objectives to be correctly predicted\nby models. In addition, our method does not modify the parameters of the\napplied model, which makes it flexible for different scenarios. Experimental\nresults on the recognition and classification tasks demonstrate that the\nproposed method can effectively hide visual information and hardly affect the\nperformances of models. The code is available in the supplementary material.\n", "date": "2022-09-30", "categories": []}
{"id": "2302.14353", "title": "A semantic backdoor attack against Graph Convolutional Networks", "abstract": "  Graph convolutional networks (GCNs) have been very effective in addressing\nthe issue of various graph-structured related tasks. However, recent research\nhas shown that GCNs are vulnerable to a new type of threat called a backdoor\nattack, where the adversary can inject a hidden backdoor into GCNs so that the\nattacked model performs well on benign samples, but its prediction will be\nmaliciously changed to the attacker-specified target label if the hidden\nbackdoor is activated by the attacker-defined trigger. A semantic backdoor\nattack is a new type of backdoor attack on deep neural networks (DNNs), where a\nnaturally occurring semantic feature of samples can serve as a backdoor trigger\nsuch that the infected DNN models will misclassify testing samples containing\nthe predefined semantic feature even without the requirement of modifying the\ntesting samples. Since the backdoor trigger is a naturally occurring semantic\nfeature of the samples, semantic backdoor attacks are more imperceptible and\npose a new and serious threat. In this paper, we investigate whether such\nsemantic backdoor attacks are possible for GCNs and propose a semantic backdoor\nattack against GCNs (SBAG) under the context of graph classification to reveal\nthe existence of this security vulnerability in GCNs. SBAG uses a certain type\nof node in the samples as a backdoor trigger and injects a hidden backdoor into\nGCN models by poisoning training data. The backdoor will be activated, and the\nGCN models will give malicious classification results specified by the attacker\neven on unmodified samples as long as the samples contain enough trigger nodes.\nWe evaluate SBAG on four graph datasets and the experimental results indicate\nthat SBAG is effective.\n", "date": "2023-02-28", "categories": []}
{"id": "2303.11917", "title": "Efficient Decision-based Black-box Patch Attacks on Video Recognition", "abstract": "  Although Deep Neural Networks (DNNs) have demonstrated excellent performance,\nthey are vulnerable to adversarial patches that introduce perceptible and\nlocalized perturbations to the input. Generating adversarial patches on images\nhas received much attention, while adversarial patches on videos have not been\nwell investigated. Further, decision-based attacks, where attackers only access\nthe predicted hard labels by querying threat models, have not been well\nexplored on video models either, even if they are practical in real-world video\nrecognition scenes. The absence of such studies leads to a huge gap in the\nrobustness assessment for video models. To bridge this gap, this work first\nexplores decision-based patch attacks on video models. We analyze that the huge\nparameter space brought by videos and the minimal information returned by\ndecision-based models both greatly increase the attack difficulty and query\nburden. To achieve a query-efficient attack, we propose a spatial-temporal\ndifferential evolution (STDE) framework. First, STDE introduces target videos\nas patch textures and only adds patches on keyframes that are adaptively\nselected by temporal difference. Second, STDE takes minimizing the patch area\nas the optimization objective and adopts spatialtemporal mutation and crossover\nto search for the global optimum without falling into the local optimum.\nExperiments show STDE has demonstrated state-of-the-art performance in terms of\nthreat, efficiency and imperceptibility. Hence, STDE has the potential to be a\npowerful tool for evaluating the robustness of video recognition models.\n", "date": "2023-03-21", "categories": []}
{"id": "2308.13768", "title": "Adversarial Fine-Tuning of Language Models: An Iterative Optimisation\n  Approach for the Generation and Detection of Problematic Content", "abstract": "  In this paper, we tackle the emerging challenge of unintended harmful content\ngeneration in Large Language Models (LLMs) with a novel dual-stage optimisation\ntechnique using adversarial fine-tuning. Our two-pronged approach employs an\nadversarial model, fine-tuned to generate potentially harmful prompts, and a\njudge model, iteratively optimised to discern these prompts. In this\nadversarial cycle, the two models seek to outperform each other in the\nprompting phase, generating a dataset of rich examples which are then used for\nfine-tuning. This iterative application of prompting and fine-tuning allows\ncontinuous refinement and improved performance. The performance of our approach\nis evaluated through classification accuracy on a dataset consisting of\nproblematic prompts not detected by GPT-4, as well as a selection of\ncontentious but unproblematic prompts. We show considerable increase in\nclassification accuracy of the judge model on this challenging dataset as it\nundergoes the optimisation process. Furthermore, we show that a rudimentary\nmodel \\texttt{ada} can achieve 13\\% higher accuracy on the hold-out test set\nthan GPT-4 after only a few rounds of this process, and that this fine-tuning\nimproves performance in parallel tasks such as toxic comment identification.\n", "date": "2023-08-26", "categories": []}
{"id": "2308.14333", "title": "DiffSmooth: Certifiably Robust Learning via Diffusion Models and Local\n  Smoothing", "abstract": "  Diffusion models have been leveraged to perform adversarial purification and\nthus provide both empirical and certified robustness for a standard model. On\nthe other hand, different robustly trained smoothed models have been studied to\nimprove the certified robustness. Thus, it raises a natural question: Can\ndiffusion model be used to achieve improved certified robustness on those\nrobustly trained smoothed models? In this work, we first theoretically show\nthat recovered instances by diffusion models are in the bounded neighborhood of\nthe original instance with high probability; and the \"one-shot\" denoising\ndiffusion probabilistic models (DDPM) can approximate the mean of the generated\ndistribution of a continuous-time diffusion model, which approximates the\noriginal instance under mild conditions. Inspired by our analysis, we propose a\ncertifiably robust pipeline DiffSmooth, which first performs adversarial\npurification via diffusion models and then maps the purified instances to a\ncommon region via a simple yet effective local smoothing strategy. We conduct\nextensive experiments on different datasets and show that DiffSmooth achieves\nSOTA-certified robustness compared with eight baselines. For instance,\nDiffSmooth improves the SOTA-certified accuracy from $36.0\\%$ to $53.0\\%$ under\n$\\ell_2$ radius $1.5$ on ImageNet. The code is available at\n[https://github.com/javyduck/DiffSmooth].\n", "date": "2023-08-28", "categories": []}
{"id": "2308.14597", "title": "Adversarial Attacks on Foundational Vision Models", "abstract": "  Rapid progress is being made in developing large, pretrained, task-agnostic\nfoundational vision models such as CLIP, ALIGN, DINOv2, etc. In fact, we are\napproaching the point where these models do not have to be finetuned\ndownstream, and can simply be used in zero-shot or with a lightweight probing\nhead. Critically, given the complexity of working at this scale, there is a\nbottleneck where relatively few organizations in the world are executing the\ntraining then sharing the models on centralized platforms such as HuggingFace\nand torch.hub. The goal of this work is to identify several key adversarial\nvulnerabilities of these models in an effort to make future designs more\nrobust. Intuitively, our attacks manipulate deep feature representations to\nfool an out-of-distribution (OOD) detector which will be required when using\nthese open-world-aware models to solve closed-set downstream tasks. Our methods\nreliably make in-distribution (ID) images (w.r.t. a downstream task) be\npredicted as OOD and vice versa while existing in extremely\nlow-knowledge-assumption threat models. We show our attacks to be potent in\nwhitebox and blackbox settings, as well as when transferred across foundational\nmodel types (e.g., attack DINOv2 with CLIP)! This work is only just the\nbeginning of a long journey towards adversarially robust foundational vision\nmodels.\n", "date": "2023-08-28", "categories": []}
{"id": "2306.05952", "title": "Overcoming Adversarial Attacks for Human-in-the-Loop Applications", "abstract": "  Including human analysis has the potential to positively affect the\nrobustness of Deep Neural Networks and is relatively unexplored in the\nAdversarial Machine Learning literature. Neural network visual explanation maps\nhave been shown to be prone to adversarial attacks. Further research is needed\nin order to select robust visualizations of explanations for the image analyst\nto evaluate a given model. These factors greatly impact Human-In-The-Loop\n(HITL) evaluation tools due to their reliance on adversarial images, including\nexplanation maps and measurements of robustness. We believe models of human\nvisual attention may improve interpretability and robustness of human-machine\nimagery analysis systems. Our challenge remains, how can HITL evaluation be\nrobust in this adversarial landscape?\n", "date": "2023-06-09", "categories": []}
{"id": "2308.05983", "title": "Face Encryption via Frequency-Restricted Identity-Agnostic Attacks", "abstract": "  Billions of people are sharing their daily live images on social media\neveryday. However, malicious collectors use deep face recognition systems to\neasily steal their biometric information (e.g., faces) from these images. Some\nstudies are being conducted to generate encrypted face photos using adversarial\nattacks by introducing imperceptible perturbations to reduce face information\nleakage. However, existing studies need stronger black-box scenario feasibility\nand more natural visual appearances, which challenge the feasibility of privacy\nprotection. To address these problems, we propose a frequency-restricted\nidentity-agnostic (FRIA) framework to encrypt face images from unauthorized\nface recognition without access to personal information. As for the weak\nblack-box scenario feasibility, we obverse that representations of the average\nfeature in multiple face recognition models are similar, thus we propose to\nutilize the average feature via the crawled dataset from the Internet as the\ntarget to guide the generation, which is also agnostic to identities of unknown\nface recognition systems; in nature, the low-frequency perturbations are more\nvisually perceptible by the human vision system. Inspired by this, we restrict\nthe perturbation in the low-frequency facial regions by discrete cosine\ntransform to achieve the visual naturalness guarantee. Extensive experiments on\nseveral face recognition models demonstrate that our FRIA outperforms other\nstate-of-the-art methods in generating more natural encrypted faces while\nattaining high black-box attack success rates of 96%. In addition, we validate\nthe efficacy of FRIA using real-world black-box commercial API, which reveals\nthe potential of FRIA in practice. Our codes can be found in\nhttps://github.com/XinDong10/FRIA.\n", "date": "2023-08-11", "categories": []}
{"id": "2208.06228", "title": "Unifying Gradients to Improve Real-world Robustness for Deep Networks", "abstract": "  The wide application of deep neural networks (DNNs) demands an increasing\namount of attention to their real-world robustness, i.e., whether a DNN resists\nblack-box adversarial attacks, among which score-based query attacks (SQAs) are\nmost threatening since they can effectively hurt a victim network with the only\naccess to model outputs. Defending against SQAs requires a slight but artful\nvariation of outputs due to the service purpose for users, who share the same\noutput information with SQAs. In this paper, we propose a real-world defense by\nUnifying Gradients (UniG) of different data so that SQAs could only probe a\nmuch weaker attack direction that is similar for different samples. Since such\nuniversal attack perturbations have been validated as less aggressive than the\ninput-specific perturbations, UniG protects real-world DNNs by indicating\nattackers a twisted and less informative attack direction. We implement UniG\nefficiently by a Hadamard product module which is plug-and-play. According to\nextensive experiments on 5 SQAs, 2 adaptive attacks and 7 defense baselines,\nUniG significantly improves real-world robustness without hurting clean\naccuracy on CIFAR10 and ImageNet. For instance, UniG maintains a model of\n77.80% accuracy under 2500-query Square attack while the state-of-the-art\nadversarially-trained model only has 67.34% on CIFAR10. Simultaneously, UniG\noutperforms all compared baselines in terms of clean accuracy and achieves the\nsmallest modification of the model output. The code is released at\nhttps://github.com/snowien/UniG-pytorch.\n", "date": "2022-08-12", "categories": []}
{"id": "2308.11881", "title": "Adversarial Training Using Feedback Loops", "abstract": "  Deep neural networks (DNN) have found wide applicability in numerous fields\ndue to their ability to accurately learn very complex input-output relations.\nDespite their accuracy and extensive use, DNNs are highly susceptible to\nadversarial attacks due to limited generalizability. For future progress in the\nfield, it is essential to build DNNs that are robust to any kind of\nperturbations to the data points. In the past, many techniques have been\nproposed to robustify DNNs using first-order derivative information of the\nnetwork.\n  This paper proposes a new robustification approach based on control theory. A\nneural network architecture that incorporates feedback control, named Feedback\nNeural Networks, is proposed. The controller is itself a neural network, which\nis trained using regular and adversarial data such as to stabilize the system\noutputs. The novel adversarial training approach based on the feedback control\narchitecture is called Feedback Looped Adversarial Training (FLAT). Numerical\nresults on standard test problems empirically show that our FLAT method is more\neffective than the state-of-the-art to guard against adversarial attacks.\n", "date": "2023-08-23", "categories": []}
{"id": "2308.12661", "title": "Don't Look into the Sun: Adversarial Solarization Attacks on Image\n  Classifiers", "abstract": "  Assessing the robustness of deep neural networks against out-of-distribution\ninputs is crucial, especially in safety-critical domains like autonomous\ndriving, but also in safety systems where malicious actors can digitally alter\ninputs to circumvent safety guards. However, designing effective\nout-of-distribution tests that encompass all possible scenarios while\npreserving accurate label information is a challenging task. Existing\nmethodologies often entail a compromise between variety and constraint levels\nfor attacks and sometimes even both. In a first step towards a more holistic\nrobustness evaluation of image classification models, we introduce an attack\nmethod based on image solarization that is conceptually straightforward yet\navoids jeopardizing the global structure of natural images independent of the\nintensity. Through comprehensive evaluations of multiple ImageNet models, we\ndemonstrate the attack's capacity to degrade accuracy significantly, provided\nit is not integrated into the training augmentations. Interestingly, even then,\nno full immunity to accuracy deterioration is achieved. In other settings, the\nattack can often be simplified into a black-box attack with model-independent\nparameters. Defenses against other corruptions do not consistently extend to be\neffective against our specific attack.\n  Project website: https://github.com/paulgavrikov/adversarial_solarization\n", "date": "2023-08-24", "categories": []}
{"id": "2308.12857", "title": "Fast Adversarial Training with Smooth Convergence", "abstract": "  Fast adversarial training (FAT) is beneficial for improving the adversarial\nrobustness of neural networks. However, previous FAT work has encountered a\nsignificant issue known as catastrophic overfitting when dealing with large\nperturbation budgets, \\ie the adversarial robustness of models declines to near\nzero during training.\n  To address this, we analyze the training process of prior FAT work and\nobserve that catastrophic overfitting is accompanied by the appearance of loss\nconvergence outliers.\n  Therefore, we argue a moderately smooth loss convergence process will be a\nstable FAT process that solves catastrophic overfitting.\n  To obtain a smooth loss convergence process, we propose a novel oscillatory\nconstraint (dubbed ConvergeSmooth) to limit the loss difference between\nadjacent epochs. The convergence stride of ConvergeSmooth is introduced to\nbalance convergence and smoothing. Likewise, we design weight centralization\nwithout introducing additional hyperparameters other than the loss balance\ncoefficient.\n  Our proposed methods are attack-agnostic and thus can improve the training\nstability of various FAT techniques.\n  Extensive experiments on popular datasets show that the proposed methods\nefficiently avoid catastrophic overfitting and outperform all previous FAT\nmethods. Code is available at \\url{https://github.com/FAT-CS/ConvergeSmooth}.\n", "date": "2023-08-24", "categories": []}
{"id": "2308.12918", "title": "Evaluating the Vulnerabilities in ML systems in terms of adversarial\n  attacks", "abstract": "  There have been recent adversarial attacks that are difficult to find. These\nnew adversarial attacks methods may pose challenges to current deep learning\ncyber defense systems and could influence the future defense of cyberattacks.\nThe authors focus on this domain in this research paper. They explore the\nconsequences of vulnerabilities in AI systems. This includes discussing how\nthey might arise, differences between randomized and adversarial examples and\nalso potential ethical implications of vulnerabilities. Moreover, it is\nimportant to train the AI systems appropriately when they are in testing phase\nand getting them ready for broader use.\n", "date": "2023-08-24", "categories": []}
{"id": "2302.08973", "title": "Measuring Equality in Machine Learning Security Defenses: A Case Study\n  in Speech Recognition", "abstract": "  Over the past decade, the machine learning security community has developed a\nmyriad of defenses for evasion attacks. An understudied question in that\ncommunity is: for whom do these defenses defend? This work considers common\napproaches to defending learned systems and how security defenses result in\nperformance inequities across different sub-populations. We outline appropriate\nparity metrics for analysis and begin to answer this question through empirical\nresults of the fairness implications of machine learning security methods. We\nfind that many methods that have been proposed can cause direct harm, like\nfalse rejection and unequal benefits from robustness training. The framework we\npropose for measuring defense equality can be applied to robustly trained\nmodels, preprocessing-based defenses, and rejection methods. We identify a set\nof datasets with a user-centered application and a reasonable computational\ncost suitable for case studies in measuring the equality of defenses. In our\ncase study of speech command recognition, we show how such adversarial training\nand augmentation have non-equal but complex protections for social subgroups\nacross gender, accent, and age in relation to user coverage. We present a\ncomparison of equality between two rejection-based defenses: randomized\nsmoothing and neural rejection, finding randomized smoothing more equitable due\nto the sampling mechanism for minority groups. This represents the first work\nexamining the disparity in the adversarial robustness in the speech domain and\nthe fairness evaluation of rejection-based defenses.\n", "date": "2023-02-17", "categories": []}
{"id": "2308.10718", "title": "Backdooring Textual Inversion for Concept Censorship", "abstract": "  Recent years have witnessed success in AIGC (AI Generated Content). People\ncan make use of a pre-trained diffusion model to generate images of high\nquality or freely modify existing pictures with only prompts in nature\nlanguage. More excitingly, the emerging personalization techniques make it\nfeasible to create specific-desired images with only a few images as\nreferences. However, this induces severe threats if such advanced techniques\nare misused by malicious users, such as spreading fake news or defaming\nindividual reputations. Thus, it is necessary to regulate personalization\nmodels (i.e., concept censorship) for their development and advancement.\n  In this paper, we focus on the personalization technique dubbed Textual\nInversion (TI), which is becoming prevailing for its lightweight nature and\nexcellent performance. TI crafts the word embedding that contains detailed\ninformation about a specific object. Users can easily download the word\nembedding from public websites like Civitai and add it to their own stable\ndiffusion model without fine-tuning for personalization. To achieve the concept\ncensorship of a TI model, we propose leveraging the backdoor technique for good\nby injecting backdoors into the Textual Inversion embeddings. Briefly, we\nselect some sensitive words as triggers during the training of TI, which will\nbe censored for normal use. In the subsequent generation stage, if the triggers\nare combined with personalized embeddings as final prompts, the model will\noutput a pre-defined target image rather than images including the desired\nmalicious concept.\n  To demonstrate the effectiveness of our approach, we conduct extensive\nexperiments on Stable Diffusion, a prevailing open-sourced text-to-image model.\nOur code, data, and results are available at\nhttps://concept-censorship.github.io.\n", "date": "2023-08-21", "categories": []}
{"id": "2308.11754", "title": "Multi-Instance Adversarial Attack on GNN-Based Malicious Domain\n  Detection", "abstract": "  Malicious domain detection (MDD) is an open security challenge that aims to\ndetect if an Internet domain is associated with cyber-attacks. Among many\napproaches to this problem, graph neural networks (GNNs) are deemed highly\neffective. GNN-based MDD uses DNS logs to represent Internet domains as nodes\nin a maliciousness graph (DMG) and trains a GNN to infer their maliciousness by\nleveraging identified malicious domains. Since this method relies on accessible\nDNS logs to construct DMGs, it exposes a vulnerability for adversaries to\nmanipulate their domain nodes' features and connections within DMGs. Existing\nresearch mainly concentrates on threat models that manipulate individual\nattacker nodes. However, adversaries commonly generate multiple domains to\nachieve their goals economically and avoid detection. Their objective is to\nevade discovery across as many domains as feasible. In this work, we call the\nattack that manipulates several nodes in the DMG concurrently a multi-instance\nevasion attack. We present theoretical and empirical evidence that the existing\nsingle-instance evasion techniques for are inadequate to launch multi-instance\nevasion attacks against GNN-based MDDs. Therefore, we introduce MintA, an\ninference-time multi-instance adversarial attack on GNN-based MDDs. MintA\nenhances node and neighborhood evasiveness through optimized perturbations and\noperates successfully with only black-box access to the target model,\neliminating the need for knowledge about the model's specifics or non-adversary\nnodes. We formulate an optimization challenge for MintA, achieving an\napproximate solution. Evaluating MintA on a leading GNN-based MDD technique\nwith real-world data showcases an attack success rate exceeding 80%. These\nfindings act as a warning for security experts, underscoring GNN-based MDDs'\nsusceptibility to practical attacks that can undermine their effectiveness and\nbenefits.\n", "date": "2023-08-22", "categories": []}
{"id": "2308.11822", "title": "PatchBackdoor: Backdoor Attack against Deep Neural Networks without\n  Model Modification", "abstract": "  Backdoor attack is a major threat to deep learning systems in safety-critical\nscenarios, which aims to trigger misbehavior of neural network models under\nattacker-controlled conditions. However, most backdoor attacks have to modify\nthe neural network models through training with poisoned data and/or direct\nmodel editing, which leads to a common but false belief that backdoor attack\ncan be easily avoided by properly protecting the model. In this paper, we show\nthat backdoor attacks can be achieved without any model modification. Instead\nof injecting backdoor logic into the training data or the model, we propose to\nplace a carefully-designed patch (namely backdoor patch) in front of the\ncamera, which is fed into the model together with the input images. The patch\ncan be trained to behave normally at most of the time, while producing wrong\nprediction when the input image contains an attacker-controlled trigger object.\nOur main techniques include an effective training method to generate the\nbackdoor patch and a digital-physical transformation modeling method to enhance\nthe feasibility of the patch in real deployments. Extensive experiments show\nthat PatchBackdoor can be applied to common deep learning models (VGG,\nMobileNet, ResNet) with an attack success rate of 93% to 99% on classification\ntasks. Moreover, we implement PatchBackdoor in real-world scenarios and show\nthat the attack is still threatening.\n", "date": "2023-08-22", "categories": []}
{"id": "2308.11894", "title": "Does Physical Adversarial Example Really Matter to Autonomous Driving?\n  Towards System-Level Effect of Adversarial Object Evasion Attack", "abstract": "  In autonomous driving (AD), accurate perception is indispensable to achieving\nsafe and secure driving. Due to its safety-criticality, the security of AD\nperception has been widely studied. Among different attacks on AD perception,\nthe physical adversarial object evasion attacks are especially severe. However,\nwe find that all existing literature only evaluates their attack effect at the\ntargeted AI component level but not at the system level, i.e., with the entire\nsystem semantics and context such as the full AD pipeline. Thereby, this raises\na critical research question: can these existing researches effectively achieve\nsystem-level attack effects (e.g., traffic rule violations) in the real-world\nAD context? In this work, we conduct the first measurement study on whether and\nhow effectively the existing designs can lead to system-level effects,\nespecially for the STOP sign-evasion attacks due to their popularity and\nseverity. Our evaluation results show that all the representative prior works\ncannot achieve any system-level effects. We observe two design limitations in\nthe prior works: 1) physical model-inconsistent object size distribution in\npixel sampling and 2) lack of vehicle plant model and AD system model\nconsideration. Then, we propose SysAdv, a novel system-driven attack design in\nthe AD context and our evaluation results show that the system-level effects\ncan be significantly improved, i.e., the violation rate increases by around\n70%.\n", "date": "2023-08-23", "categories": []}
{"id": "2308.12054", "title": "Sample Complexity of Robust Learning against Evasion Attacks", "abstract": "  It is becoming increasingly important to understand the vulnerability of\nmachine learning models to adversarial attacks. One of the fundamental problems\nin adversarial machine learning is to quantify how much training data is needed\nin the presence of evasion attacks, where data is corrupted at test time. In\nthis thesis, we work with the exact-in-the-ball notion of robustness and study\nthe feasibility of adversarially robust learning from the perspective of\nlearning theory, considering sample complexity.\n  We first explore the setting where the learner has access to random examples\nonly, and show that distributional assumptions are essential. We then focus on\nlearning problems with distributions on the input data that satisfy a Lipschitz\ncondition and show that robustly learning monotone conjunctions has sample\ncomplexity at least exponential in the adversary's budget (the maximum number\nof bits it can perturb on each input). However, if the adversary is restricted\nto perturbing $O(\\log n)$ bits, then one can robustly learn conjunctions and\ndecision lists w.r.t. log-Lipschitz distributions.\n  We then study learning models where the learner is given more power. We first\nconsider local membership queries, where the learner can query the label of\npoints near the training sample. We show that, under the uniform distribution,\nthe exponential dependence on the adversary's budget to robustly learn\nconjunctions remains inevitable. We then introduce a local equivalence query\noracle, which returns whether the hypothesis and target concept agree in a\ngiven region around a point in the training sample, and a counterexample if it\nexists. We show that if the query radius is equal to the adversary's budget, we\ncan develop robust empirical risk minimization algorithms in the\ndistribution-free setting. We give general query complexity upper and lower\nbounds, as well as for concrete concept classes.\n", "date": "2023-08-23", "categories": []}
{"id": "2308.12279", "title": "On-Manifold Projected Gradient Descent", "abstract": "  This work provides a computable, direct, and mathematically rigorous\napproximation to the differential geometry of class manifolds for\nhigh-dimensional data, along with nonlinear projections from input space onto\nthese class manifolds. The tools are applied to the setting of neural network\nimage classifiers, where we generate novel, on-manifold data samples, and\nimplement a projected gradient descent algorithm for on-manifold adversarial\ntraining. The susceptibility of neural networks (NNs) to adversarial attack\nhighlights the brittle nature of NN decision boundaries in input space.\nIntroducing adversarial examples during training has been shown to reduce the\nsusceptibility of NNs to adversarial attack; however, it has also been shown to\nreduce the accuracy of the classifier if the examples are not valid examples\nfor that class. Realistic \"on-manifold\" examples have been previously generated\nfrom class manifolds in the latent of an autoencoder. Our work explores these\nphenomena in a geometric and computational setting that is much closer to the\nraw, high-dimensional input space than can be provided by VAE or other black\nbox dimensionality reductions. We employ conformally invariant diffusion maps\n(CIDM) to approximate class manifolds in diffusion coordinates, and develop the\nNystr\\\"{o}m projection to project novel points onto class manifolds in this\nsetting. On top of the manifold approximation, we leverage the spectral\nexterior calculus (SEC) to determine geometric quantities such as tangent\nvectors of the manifold. We use these tools to obtain adversarial examples that\nreside on a class manifold, yet fool a classifier. These misclassifications\nthen become explainable in terms of human-understandable manipulations within\nthe data, by expressing the on-manifold adversary in the semantic basis on the\nmanifold.\n", "date": "2023-08-23", "categories": []}
{"id": "2303.02242", "title": "TrojText: Test-time Invisible Textual Trojan Insertion", "abstract": "  In Natural Language Processing (NLP), intelligent neuron models can be\nsusceptible to textual Trojan attacks. Such attacks occur when Trojan models\nbehave normally for standard inputs but generate malicious output for inputs\nthat contain a specific trigger. Syntactic-structure triggers, which are\ninvisible, are becoming more popular for Trojan attacks because they are\ndifficult to detect and defend against. However, these types of attacks require\na large corpus of training data to generate poisoned samples with the necessary\nsyntactic structures for Trojan insertion. Obtaining such data can be difficult\nfor attackers, and the process of generating syntactic poisoned triggers and\ninserting Trojans can be time-consuming. This paper proposes a solution called\nTrojText, which aims to determine whether invisible textual Trojan attacks can\nbe performed more efficiently and cost-effectively without training data. The\nproposed approach, called the Representation-Logit Trojan Insertion (RLI)\nalgorithm, uses smaller sampled test data instead of large training data to\nachieve the desired attack. The paper also introduces two additional\ntechniques, namely the accumulated gradient ranking (AGR) and Trojan Weights\nPruning (TWP), to reduce the number of tuned parameters and the attack\noverhead. The TrojText approach was evaluated on three datasets (AG's News,\nSST-2, and OLID) using three NLP models (BERT, XLNet, and DeBERTa). The\nexperiments demonstrated that the TrojText approach achieved a 98.35\\%\nclassification accuracy for test sentences in the target class on the BERT\nmodel for the AG's News dataset. The source code for TrojText is available at\nhttps://github.com/UCF-ML-Research/TrojText.\n", "date": "2023-03-03", "categories": []}
{"id": "2308.10315", "title": "Improving Adversarial Robustness of Masked Autoencoders via Test-time\n  Frequency-domain Prompting", "abstract": "  In this paper, we investigate the adversarial robustness of vision\ntransformers that are equipped with BERT pretraining (e.g., BEiT, MAE). A\nsurprising observation is that MAE has significantly worse adversarial\nrobustness than other BERT pretraining methods. This observation drives us to\nrethink the basic differences between these BERT pretraining methods and how\nthese differences affect the robustness against adversarial perturbations. Our\nempirical analysis reveals that the adversarial robustness of BERT pretraining\nis highly related to the reconstruction target, i.e., predicting the raw pixels\nof masked image patches will degrade more adversarial robustness of the model\nthan predicting the semantic context, since it guides the model to concentrate\nmore on medium-/high-frequency components of images. Based on our analysis, we\nprovide a simple yet effective way to boost the adversarial robustness of MAE.\nThe basic idea is using the dataset-extracted domain knowledge to occupy the\nmedium-/high-frequency of images, thus narrowing the optimization space of\nadversarial perturbations. Specifically, we group the distribution of\npretraining data and optimize a set of cluster-specific visual prompts on\nfrequency domain. These prompts are incorporated with input images through\nprototype-based prompt selection during test period. Extensive evaluation shows\nthat our method clearly boost MAE's adversarial robustness while maintaining\nits clean performance on ImageNet-1k classification. Our code is available at:\nhttps://github.com/shikiw/RobustMAE.\n", "date": "2023-08-20", "categories": []}
{"id": "2308.11284", "title": "LEAP: Efficient and Automated Test Method for NLP Software", "abstract": "  The widespread adoption of DNNs in NLP software has highlighted the need for\nrobustness. Researchers proposed various automatic testing techniques for\nadversarial test cases. However, existing methods suffer from two limitations:\nweak error-discovering capabilities, with success rates ranging from 0% to\n24.6% for BERT-based NLP software, and time inefficiency, taking 177.8s to\n205.28s per test case, making them challenging for time-constrained scenarios.\nTo address these issues, this paper proposes LEAP, an automated test method\nthat uses LEvy flight-based Adaptive Particle swarm optimization integrated\nwith textual features to generate adversarial test cases. Specifically, we\nadopt Levy flight for population initialization to increase the diversity of\ngenerated test cases. We also design an inertial weight adaptive update\noperator to improve the efficiency of LEAP's global optimization of\nhigh-dimensional text examples and a mutation operator based on the greedy\nstrategy to reduce the search time. We conducted a series of experiments to\nvalidate LEAP's ability to test NLP software and found that the average success\nrate of LEAP in generating adversarial test cases is 79.1%, which is 6.1%\nhigher than the next best approach (PSOattack). While ensuring high success\nrates, LEAP significantly reduces time overhead by up to 147.6s compared to\nother heuristic-based methods. Additionally, the experimental results\ndemonstrate that LEAP can generate more transferable test cases and\nsignificantly enhance the robustness of DNN-based systems.\n", "date": "2023-08-22", "categories": []}
{"id": "2308.11443", "title": "Revisiting and Exploring Efficient Fast Adversarial Training via LAW:\n  Lipschitz Regularization and Auto Weight Averaging", "abstract": "  Fast Adversarial Training (FAT) not only improves the model robustness but\nalso reduces the training cost of standard adversarial training. However, fast\nadversarial training often suffers from Catastrophic Overfitting (CO), which\nresults in poor robustness performance. Catastrophic Overfitting describes the\nphenomenon of a sudden and significant decrease in robust accuracy during the\ntraining of fast adversarial training. Many effective techniques have been\ndeveloped to prevent Catastrophic Overfitting and improve the model robustness\nfrom different perspectives. However, these techniques adopt inconsistent\ntraining settings and require different training costs, i.e, training time and\nmemory costs, leading to unfair comparisons. In this paper, we conduct a\ncomprehensive study of over 10 fast adversarial training methods in terms of\nadversarial robustness and training costs. We revisit the effectiveness and\nefficiency of fast adversarial training techniques in preventing Catastrophic\nOverfitting from the perspective of model local nonlinearity and propose an\neffective Lipschitz regularization method for fast adversarial training.\nFurthermore, we explore the effect of data augmentation and weight averaging in\nfast adversarial training and propose a simple yet effective auto weight\naveraging method to improve robustness further. By assembling these techniques,\nwe propose a FGSM-based fast adversarial training method equipped with\nLipschitz regularization and Auto Weight averaging, abbreviated as FGSM-LAW.\nExperimental evaluations on four benchmark databases demonstrate the\nsuperiority of the proposed method over state-of-the-art fast adversarial\ntraining methods and the advanced standard adversarial training methods.\n", "date": "2023-08-22", "categories": []}
{"id": "2208.05073", "title": "Adversarial Machine Learning-Based Anticipation of Threats Against\n  Vehicle-to-Microgrid Services", "abstract": "  In this paper, we study the expanding attack surface of Adversarial Machine\nLearning (AML) and the potential attacks against Vehicle-to-Microgrid (V2M)\nservices. We present an anticipatory study of a multi-stage gray-box attack\nthat can achieve a comparable result to a white-box attack. Adversaries aim to\ndeceive the targeted Machine Learning (ML) classifier at the network edge to\nmisclassify the incoming energy requests from microgrids. With an inference\nattack, an adversary can collect real-time data from the communication between\nsmart microgrids and a 5G gNodeB to train a surrogate (i.e., shadow) model of\nthe targeted classifier at the edge. To anticipate the associated impact of an\nadversary's capability to collect real-time data instances, we study five\ndifferent cases, each representing different amounts of real-time data\ninstances collected by an adversary. Out of six ML models trained on the\ncomplete dataset, K-Nearest Neighbour (K-NN) is selected as the surrogate\nmodel, and through simulations, we demonstrate that the multi-stage gray-box\nattack is able to mislead the ML classifier and cause an Evasion Increase Rate\n(EIR) up to 73.2% using 40% less data than what a white-box attack needs to\nachieve a similar EIR.\n", "date": "2022-08-09", "categories": []}
{"id": "2307.16572", "title": "Transferable Attack for Semantic Segmentation", "abstract": "  We analysis performance of semantic segmentation models wrt. adversarial\nattacks, and observe that the adversarial examples generated from a source\nmodel fail to attack the target models. i.e The conventional attack methods,\nsuch as PGD and FGSM, do not transfer well to target models, making it\nnecessary to study the transferable attacks, especially transferable attacks\nfor semantic segmentation. We find two main factors to achieve transferable\nattack. Firstly, the attack should come with effective data augmentation and\ntranslation-invariant features to deal with unseen models. Secondly, stabilized\noptimization strategies are needed to find the optimal attack direction. Based\non the above observations, we propose an ensemble attack for semantic\nsegmentation to achieve more effective attacks with higher transferability. The\nsource code and experimental results are publicly available via our project\npage: https://github.com/anucvers/TASS.\n", "date": "2023-07-31", "categories": []}
{"id": "2308.09861", "title": "Black-box Adversarial Attacks against Dense Retrieval Models: A\n  Multi-view Contrastive Learning Method", "abstract": "  Neural ranking models (NRMs) and dense retrieval (DR) models have given rise\nto substantial improvements in overall retrieval performance. In addition to\ntheir effectiveness, and motivated by the proven lack of robustness of deep\nlearning-based approaches in other areas, there is growing interest in the\nrobustness of deep learning-based approaches to the core retrieval problem.\nAdversarial attack methods that have so far been developed mainly focus on\nattacking NRMs, with very little attention being paid to the robustness of DR\nmodels. In this paper, we introduce the adversarial retrieval attack (AREA)\ntask. The AREA task is meant to trick DR models into retrieving a target\ndocument that is outside the initial set of candidate documents retrieved by\nthe DR model in response to a query. We consider the decision-based black-box\nadversarial setting, which is realistic in real-world search engines. To\naddress the AREA task, we first employ existing adversarial attack methods\ndesigned for NRMs. We find that the promising results that have previously been\nreported on attacking NRMs, do not generalize to DR models: these methods\nunderperform a simple term spamming method. We attribute the observed lack of\ngeneralizability to the interaction-focused architecture of NRMs, which\nemphasizes fine-grained relevance matching. DR models follow a different\nrepresentation-focused architecture that prioritizes coarse-grained\nrepresentations. We propose to formalize attacks on DR models as a contrastive\nlearning problem in a multi-view representation space. The core idea is to\nencourage the consistency between each view representation of the target\ndocument and its corresponding viewer via view-wise supervision signals.\nExperimental results demonstrate that the proposed method can significantly\noutperform existing attack strategies in misleading the DR model with small\nindiscernible text perturbations.\n", "date": "2023-08-19", "categories": []}
{"id": "2308.09889", "title": "DUAW: Data-free Universal Adversarial Watermark against Stable Diffusion\n  Customization", "abstract": "  Stable Diffusion (SD) customization approaches enable users to personalize SD\nmodel outputs, greatly enhancing the flexibility and diversity of AI art.\nHowever, they also allow individuals to plagiarize specific styles or subjects\nfrom copyrighted images, which raises significant concerns about potential\ncopyright infringement. To address this issue, we propose an invisible\ndata-free universal adversarial watermark (DUAW), aiming to protect a myriad of\ncopyrighted images from different customization approaches across various\nversions of SD models. First, DUAW is designed to disrupt the variational\nautoencoder during SD customization. Second, DUAW operates in a data-free\ncontext, where it is trained on synthetic images produced by a Large Language\nModel (LLM) and a pretrained SD model. This approach circumvents the necessity\nof directly handling copyrighted images, thereby preserving their\nconfidentiality. Once crafted, DUAW can be imperceptibly integrated into\nmassive copyrighted images, serving as a protective measure by inducing\nsignificant distortions in the images generated by customized SD models.\nExperimental results demonstrate that DUAW can effectively distort the outputs\nof fine-tuned SD models, rendering them discernible to both human observers and\na simple classifier.\n", "date": "2023-08-19", "categories": []}
{"id": "2308.10110", "title": "Robust Mixture-of-Expert Training for Convolutional Neural Networks", "abstract": "  Sparsely-gated Mixture of Expert (MoE), an emerging deep model architecture,\nhas demonstrated a great promise to enable high-accuracy and ultra-efficient\nmodel inference. Despite the growing popularity of MoE, little work\ninvestigated its potential to advance convolutional neural networks (CNNs),\nespecially in the plane of adversarial robustness. Since the lack of robustness\nhas become one of the main hurdles for CNNs, in this paper we ask: How to\nadversarially robustify a CNN-based MoE model? Can we robustly train it like an\nordinary CNN model? Our pilot study shows that the conventional adversarial\ntraining (AT) mechanism (developed for vanilla CNNs) no longer remains\neffective to robustify an MoE-CNN. To better understand this phenomenon, we\ndissect the robustness of an MoE-CNN into two dimensions: Robustness of routers\n(i.e., gating functions to select data-specific experts) and robustness of\nexperts (i.e., the router-guided pathways defined by the subnetworks of the\nbackbone CNN). Our analyses show that routers and experts are hard to adapt to\neach other in the vanilla AT. Thus, we propose a new router-expert alternating\nAdversarial training framework for MoE, termed AdvMoE. The effectiveness of our\nproposal is justified across 4 commonly-used CNN model architectures over 4\nbenchmark datasets. We find that AdvMoE achieves 1% ~ 4% adversarial robustness\nimprovement over the original dense CNN, and enjoys the efficiency merit of\nsparsity-gated MoE, leading to more than 50% inference cost reduction. Codes\nare available at https://github.com/OPTML-Group/Robust-MoE-CNN.\n", "date": "2023-08-19", "categories": []}
{"id": "2308.10601", "title": "Improving the Transferability of Adversarial Examples with Arbitrary\n  Style Transfer", "abstract": "  Deep neural networks are vulnerable to adversarial examples crafted by\napplying human-imperceptible perturbations on clean inputs. Although many\nattack methods can achieve high success rates in the white-box setting, they\nalso exhibit weak transferability in the black-box setting. Recently, various\nmethods have been proposed to improve adversarial transferability, in which the\ninput transformation is one of the most effective methods. In this work, we\nnotice that existing input transformation-based works mainly adopt the\ntransformed data in the same domain for augmentation. Inspired by domain\ngeneralization, we aim to further improve the transferability using the data\naugmented from different domains. Specifically, a style transfer network can\nalter the distribution of low-level visual features in an image while\npreserving semantic content for humans. Hence, we propose a novel attack method\nnamed Style Transfer Method (STM) that utilizes a proposed arbitrary style\ntransfer network to transform the images into different domains. To avoid\ninconsistent semantic information of stylized images for the classification\nnetwork, we fine-tune the style transfer network and mix up the generated\nimages added by random noise with the original images to maintain semantic\nconsistency and boost input diversity. Extensive experimental results on the\nImageNet-compatible dataset show that our proposed method can significantly\nimprove the adversarial transferability on either normally trained models or\nadversarially trained models than state-of-the-art input transformation-based\nattacks. Code is available at: https://github.com/Zhijin-Ge/STM.\n", "date": "2023-08-21", "categories": []}
{"id": "2308.10708", "title": "Measuring the Effect of Causal Disentanglement on the Adversarial\n  Robustness of Neural Network Models", "abstract": "  Causal Neural Network models have shown high levels of robustness to\nadversarial attacks as well as an increased capacity for generalisation tasks\nsuch as few-shot learning and rare-context classification compared to\ntraditional Neural Networks. This robustness is argued to stem from the\ndisentanglement of causal and confounder input signals. However, no\nquantitative study has yet measured the level of disentanglement achieved by\nthese types of causal models or assessed how this relates to their adversarial\nrobustness.\n  Existing causal disentanglement metrics are not applicable to deterministic\nmodels trained on real-world datasets. We, therefore, utilise metrics of\ncontent/style disentanglement from the field of Computer Vision to measure\ndifferent aspects of the causal disentanglement for four state-of-the-art\ncausal Neural Network models. By re-implementing these models with a common\nResNet18 architecture we are able to fairly measure their adversarial\nrobustness on three standard image classification benchmarking datasets under\nseven common white-box attacks. We find a strong association (r=0.820, p=0.001)\nbetween the degree to which models decorrelate causal and confounder signals\nand their adversarial robustness. Additionally, we find a moderate negative\nassociation between the pixel-level information content of the confounder\nsignal and adversarial robustness (r=-0.597, p=0.040).\n", "date": "2023-08-21", "categories": []}
{"id": "2308.10741", "title": "On the Adversarial Robustness of Multi-Modal Foundation Models", "abstract": "  Multi-modal foundation models combining vision and language models such as\nFlamingo or GPT-4 have recently gained enormous interest. Alignment of\nfoundation models is used to prevent models from providing toxic or harmful\noutput. While malicious users have successfully tried to jailbreak foundation\nmodels, an equally important question is if honest users could be harmed by\nmalicious third-party content. In this paper we show that imperceivable attacks\non images in order to change the caption output of a multi-modal foundation\nmodel can be used by malicious content providers to harm honest users e.g. by\nguiding them to malicious websites or broadcast fake information. This\nindicates that countermeasures to adversarial attacks should be used by any\ndeployed multi-modal foundation model.\n", "date": "2023-08-21", "categories": []}
{"id": "2212.05680", "title": "REAP: A Large-Scale Realistic Adversarial Patch Benchmark", "abstract": "  Machine learning models are known to be susceptible to adversarial\nperturbation. One famous attack is the adversarial patch, a sticker with a\nparticularly crafted pattern that makes the model incorrectly predict the\nobject it is placed on. This attack presents a critical threat to\ncyber-physical systems that rely on cameras such as autonomous cars. Despite\nthe significance of the problem, conducting research in this setting has been\ndifficult; evaluating attacks and defenses in the real world is exceptionally\ncostly while synthetic data are unrealistic. In this work, we propose the REAP\n(REalistic Adversarial Patch) benchmark, a digital benchmark that allows the\nuser to evaluate patch attacks on real images, and under real-world conditions.\nBuilt on top of the Mapillary Vistas dataset, our benchmark contains over\n14,000 traffic signs. Each sign is augmented with a pair of geometric and\nlighting transformations, which can be used to apply a digitally generated\npatch realistically onto the sign. Using our benchmark, we perform the first\nlarge-scale assessments of adversarial patch attacks under realistic\nconditions. Our experiments suggest that adversarial patch attacks may present\na smaller threat than previously believed and that the success rate of an\nattack on simpler digital simulations is not predictive of its actual\neffectiveness in practice. We release our benchmark publicly at\nhttps://github.com/wagner-group/reap-benchmark.\n", "date": "2022-12-12", "categories": []}
{"id": "2301.01885", "title": "Enhancement attacks in biomedical machine learning", "abstract": "  The prevalence of machine learning in biomedical research is rapidly growing,\nyet the trustworthiness of such research is often overlooked. While some\nprevious works have investigated the ability of adversarial attacks to degrade\nmodel performance in medical imaging, the ability to falsely improve\nperformance via recently-developed \"enhancement attacks\" may be a greater\nthreat to biomedical machine learning. In the spirit of developing attacks to\nbetter understand trustworthiness, we developed two techniques to drastically\nenhance prediction performance of classifiers with minimal changes to features:\n1) general enhancement of prediction performance, and 2) enhancement of a\nparticular method over another. Our enhancement framework falsely improved\nclassifiers' accuracy from 50% to almost 100% while maintaining high feature\nsimilarities between original and enhanced data (Pearson's r's>0.99).\nSimilarly, the method-specific enhancement framework was effective in falsely\nimproving the performance of one method over another. For example, a simple\nneural network outperformed logistic regression by 17% on our enhanced dataset,\nalthough no performance differences were present in the original dataset.\nCrucially, the original and enhanced data were still similar (r=0.99). Our\nresults demonstrate the feasibility of minor data manipulations to achieve any\ndesired prediction performance, which presents an interesting ethical challenge\nfor the future of biomedical machine learning. These findings emphasize the\nneed for more robust data provenance tracking and other precautionary measures\nto ensure the integrity of biomedical machine learning research.\n", "date": "2023-01-05", "categories": []}
{"id": "2303.09495", "title": "Among Us: Adversarially Robust Collaborative Perception by Consensus", "abstract": "  Multiple robots could perceive a scene (e.g., detect objects) collaboratively\nbetter than individuals, although easily suffer from adversarial attacks when\nusing deep learning. This could be addressed by the adversarial defense, but\nits training requires the often-unknown attacking mechanism. Differently, we\npropose ROBOSAC, a novel sampling-based defense strategy generalizable to\nunseen attackers. Our key idea is that collaborative perception should lead to\nconsensus rather than dissensus in results compared to individual perception.\nThis leads to our hypothesize-and-verify framework: perception results with and\nwithout collaboration from a random subset of teammates are compared until\nreaching a consensus. In such a framework, more teammates in the sampled subset\noften entail better perception performance but require longer sampling time to\nreject potential attackers. Thus, we derive how many sampling trials are needed\nto ensure the desired size of an attacker-free subset, or equivalently, the\nmaximum size of such a subset that we can successfully sample within a given\nnumber of trials. We validate our method on the task of collaborative 3D object\ndetection in autonomous driving scenarios.\n", "date": "2023-03-16", "categories": []}
{"id": "2303.16633", "title": "Targeted Adversarial Attacks on Wind Power Forecasts", "abstract": "  In recent years, researchers proposed a variety of deep learning models for\nwind power forecasting. These models predict the wind power generation of wind\nfarms or entire regions more accurately than traditional machine learning\nalgorithms or physical models. However, latest research has shown that deep\nlearning models can often be manipulated by adversarial attacks. Since wind\npower forecasts are essential for the stability of modern power systems, it is\nimportant to protect them from this threat. In this work, we investigate the\nvulnerability of two different forecasting models to targeted, semi-targeted,\nand untargeted adversarial attacks. We consider a Long Short-Term Memory (LSTM)\nnetwork for predicting the power generation of individual wind farms and a\nConvolutional Neural Network (CNN) for forecasting the wind power generation\nthroughout Germany. Moreover, we propose the Total Adversarial Robustness Score\n(TARS), an evaluation metric for quantifying the robustness of regression\nmodels to targeted and semi-targeted adversarial attacks. It assesses the\nimpact of attacks on the model's performance, as well as the extent to which\nthe attacker's goal was achieved, by assigning a score between 0 (very\nvulnerable) and 1 (very robust). In our experiments, the LSTM forecasting model\nwas fairly robust and achieved a TARS value of over 0.78 for all adversarial\nattacks investigated. The CNN forecasting model only achieved TARS values below\n0.10 when trained ordinarily, and was thus very vulnerable. Yet, its robustness\ncould be significantly improved by adversarial training, which always resulted\nin a TARS above 0.46.\n", "date": "2023-03-29", "categories": []}
{"id": "2306.02618", "title": "Enhance Diffusion to Improve Robust Generalization", "abstract": "  Deep neural networks are susceptible to human imperceptible adversarial\nperturbations. One of the strongest defense mechanisms is \\emph{Adversarial\nTraining} (AT). In this paper, we aim to address two predominant problems in\nAT. First, there is still little consensus on how to set hyperparameters with a\nperformance guarantee for AT research, and customized settings impede a fair\ncomparison between different model designs in AT research. Second, the robustly\ntrained neural networks struggle to generalize well and suffer from tremendous\noverfitting. This paper focuses on the primary AT framework - Projected\nGradient Descent Adversarial Training (PGD-AT). We approximate the dynamic of\nPGD-AT by a continuous-time Stochastic Differential Equation (SDE), and show\nthat the diffusion term of this SDE determines the robust generalization. An\nimmediate implication of this theoretical finding is that robust generalization\nis positively correlated with the ratio between learning rate and batch size.\nWe further propose a novel approach, \\emph{Diffusion Enhanced Adversarial\nTraining} (DEAT), to manipulate the diffusion term to improve robust\ngeneralization with virtually no extra computational burden. We theoretically\nshow that DEAT obtains a tighter generalization bound than PGD-AT. Our\nempirical investigation is extensive and firmly attests that DEAT universally\noutperforms PGD-AT by a significant margin.\n", "date": "2023-06-05", "categories": []}
{"id": "2306.13213", "title": "Visual Adversarial Examples Jailbreak Aligned Large Language Models", "abstract": "  Recently, there has been a surge of interest in integrating vision into Large\nLanguage Models (LLMs), exemplified by Visual Language Models (VLMs) such as\nFlamingo and GPT-4. This paper sheds light on the security and safety\nimplications of this trend. First, we underscore that the continuous and\nhigh-dimensional nature of the visual input makes it a weak link against\nadversarial attacks, representing an expanded attack surface of\nvision-integrated LLMs. Second, we highlight that the versatility of LLMs also\npresents visual attackers with a wider array of achievable adversarial\nobjectives, extending the implications of security failures beyond mere\nmisclassification. As an illustration, we present a case study in which we\nexploit visual adversarial examples to circumvent the safety guardrail of\naligned LLMs with integrated vision. Intriguingly, we discover that a single\nvisual adversarial example can universally jailbreak an aligned LLM, compelling\nit to heed a wide range of harmful instructions that it otherwise would not)\nand generate harmful content that transcends the narrow scope of a `few-shot'\nderogatory corpus initially employed to optimize the adversarial example. Our\nstudy underscores the escalating adversarial risks associated with the pursuit\nof multimodality. Our findings also connect the long-studied adversarial\nvulnerabilities of neural networks to the nascent field of AI alignment. The\npresented attack suggests a fundamental adversarial challenge for AI alignment,\nespecially in light of the emerging trend toward multimodality in frontier\nfoundation models.\n", "date": "2023-06-22", "categories": []}
{"id": "2307.12507", "title": "Gradient-Based Word Substitution for Obstinate Adversarial Examples\n  Generation in Language Models", "abstract": "  In this paper, we study the problem of generating obstinate (over-stability)\nadversarial examples by word substitution in NLP, where input text is\nmeaningfully changed but the model's prediction does not, even though it\nshould. Previous word substitution approaches have predominantly focused on\nmanually designed antonym-based strategies for generating obstinate adversarial\nexamples, which hinders its application as these strategies can only find a\nsubset of obstinate adversarial examples and require human efforts. To address\nthis issue, in this paper, we introduce a novel word substitution method named\nGradObstinate, a gradient-based approach that automatically generates obstinate\nadversarial examples without any constraints on the search space or the need\nfor manual design principles. To empirically evaluate the efficacy of\nGradObstinate, we conduct comprehensive experiments on five representative\nmodels (Electra, ALBERT, Roberta, DistillBERT, and CLIP) finetuned on four NLP\nbenchmarks (SST-2, MRPC, SNLI, and SQuAD) and a language-grounding benchmark\n(MSCOCO). Extensive experiments show that our proposed GradObstinate generates\nmore powerful obstinate adversarial examples, exhibiting a higher attack\nsuccess rate compared to antonym-based methods. Furthermore, to show the\ntransferability of obstinate word substitutions found by GradObstinate, we\nreplace the words in four representative NLP benchmarks with their obstinate\nsubstitutions. Notably, obstinate substitutions exhibit a high success rate\nwhen transferred to other models in black-box settings, including even GPT-3\nand ChatGPT. Examples of obstinate adversarial examples found by GradObstinate\nare available at https://huggingface.co/spaces/anonauthors/SecretLanguage.\n", "date": "2023-07-24", "categories": []}
{"id": "2308.05681", "title": "Hard No-Box Adversarial Attack on Skeleton-Based Human Action\n  Recognition with Skeleton-Motion-Informed Gradient", "abstract": "  Recently, methods for skeleton-based human activity recognition have been\nshown to be vulnerable to adversarial attacks. However, these attack methods\nrequire either the full knowledge of the victim (i.e. white-box attacks),\naccess to training data (i.e. transfer-based attacks) or frequent model queries\n(i.e. black-box attacks). All their requirements are highly restrictive,\nraising the question of how detrimental the vulnerability is. In this paper, we\nshow that the vulnerability indeed exists. To this end, we consider a new\nattack task: the attacker has no access to the victim model or the training\ndata or labels, where we coin the term hard no-box attack. Specifically, we\nfirst learn a motion manifold where we define an adversarial loss to compute a\nnew gradient for the attack, named skeleton-motion-informed (SMI) gradient. Our\ngradient contains information of the motion dynamics, which is different from\nexisting gradient-based attack methods that compute the loss gradient assuming\neach dimension in the data is independent. The SMI gradient can augment many\ngradient-based attack methods, leading to a new family of no-box attack\nmethods. Extensive evaluation and comparison show that our method imposes a\nreal threat to existing classifiers. They also show that the SMI gradient\nimproves the transferability and imperceptibility of adversarial samples in\nboth no-box and transfer-based black-box settings.\n", "date": "2023-08-10", "categories": []}
{"id": "2308.08709", "title": "Dynamic Neural Network is All You Need: Understanding the Robustness of\n  Dynamic Mechanisms in Neural Networks", "abstract": "  Deep Neural Networks (DNNs) have been used to solve different day-to-day\nproblems. Recently, DNNs have been deployed in real-time systems, and lowering\nthe energy consumption and response time has become the need of the hour. To\naddress this scenario, researchers have proposed incorporating dynamic\nmechanism to static DNNs (SDNN) to create Dynamic Neural Networks (DyNNs)\nperforming dynamic amounts of computation based on the input complexity.\nAlthough incorporating dynamic mechanism into SDNNs would be preferable in\nreal-time systems, it also becomes important to evaluate how the introduction\nof dynamic mechanism impacts the robustness of the models. However, there has\nnot been a significant number of works focusing on the robustness trade-off\nbetween SDNNs and DyNNs. To address this issue, we propose to investigate the\nrobustness of dynamic mechanism in DyNNs and how dynamic mechanism design\nimpacts the robustness of DyNNs. For that purpose, we evaluate three research\nquestions. These evaluations are performed on three models and two datasets.\nThrough the studies, we find that attack transferability from DyNNs to SDNNs is\nhigher than attack transferability from SDNNs to DyNNs. Also, we find that\nDyNNs can be used to generate adversarial samples more efficiently than SDNNs.\nThen, through research studies, we provide insight into the design choices that\ncan increase robustness of DyNNs against the attack generated using static\nmodel. Finally, we propose a novel attack to understand the additional attack\nsurface introduced by the dynamic mechanism and provide design choices to\nimprove robustness against the attack.\n", "date": "2023-08-17", "categories": []}
{"id": "2308.07009", "title": "ACTIVE: Towards Highly Transferable 3D Physical Camouflage for Universal\n  and Robust Vehicle Evasion", "abstract": "  Adversarial camouflage has garnered attention for its ability to attack\nobject detectors from any viewpoint by covering the entire object's surface.\nHowever, universality and robustness in existing methods often fall short as\nthe transferability aspect is often overlooked, thus restricting their\napplication only to a specific target with limited performance. To address\nthese challenges, we present Adversarial Camouflage for Transferable and\nIntensive Vehicle Evasion (ACTIVE), a state-of-the-art physical camouflage\nattack framework designed to generate universal and robust adversarial\ncamouflage capable of concealing any 3D vehicle from detectors. Our framework\nincorporates innovative techniques to enhance universality and robustness,\nincluding a refined texture rendering that enables common texture application\nto different vehicles without being constrained to a specific texture map, a\nnovel stealth loss that renders the vehicle undetectable, and a smooth and\ncamouflage loss to enhance the naturalness of the adversarial camouflage. Our\nextensive experiments on 15 different models show that ACTIVE consistently\noutperforms existing works on various public detectors, including the latest\nYOLOv7. Notably, our universality evaluations reveal promising transferability\nto other vehicle classes, tasks (segmentation models), and the real world, not\njust other vehicles.\n", "date": "2023-08-14", "categories": []}
{"id": "2308.07934", "title": "One-bit Flip is All You Need: When Bit-flip Attack Meets Model Training", "abstract": "  Deep neural networks (DNNs) are widely deployed on real-world devices.\nConcerns regarding their security have gained great attention from researchers.\nRecently, a new weight modification attack called bit flip attack (BFA) was\nproposed, which exploits memory fault inject techniques such as row hammer to\nattack quantized models in the deployment stage. With only a few bit flips, the\ntarget model can be rendered useless as a random guesser or even be implanted\nwith malicious functionalities. In this work, we seek to further reduce the\nnumber of bit flips. We propose a training-assisted bit flip attack, in which\nthe adversary is involved in the training stage to build a high-risk model to\nrelease. This high-risk model, obtained coupled with a corresponding malicious\nmodel, behaves normally and can escape various detection methods. The results\non benchmark datasets show that an adversary can easily convert this high-risk\nbut normal model to a malicious one on victim's side by \\textbf{flipping only\none critical bit} on average in the deployment stage. Moreover, our attack\nstill poses a significant threat even when defenses are employed. The codes for\nreproducing main experiments are available at\n\\url{https://github.com/jianshuod/TBA}.\n", "date": "2023-08-12", "categories": []}
{"id": "2308.08160", "title": "Benchmarking Adversarial Robustness of Compressed Deep Learning Models", "abstract": "  The increasing size of Deep Neural Networks (DNNs) poses a pressing need for\nmodel compression, particularly when employed on resource constrained devices.\nConcurrently, the susceptibility of DNNs to adversarial attacks presents\nanother significant hurdle. Despite substantial research on both model\ncompression and adversarial robustness, their joint examination remains\nunderexplored. Our study bridges this gap, seeking to understand the effect of\nadversarial inputs crafted for base models on their pruned versions. To examine\nthis relationship, we have developed a comprehensive benchmark across diverse\nadversarial attacks and popular DNN models. We uniquely focus on models not\npreviously exposed to adversarial training and apply pruning schemes optimized\nfor accuracy and performance. Our findings reveal that while the benefits of\npruning enhanced generalizability, compression, and faster inference times are\npreserved, adversarial robustness remains comparable to the base model. This\nsuggests that model compression while offering its unique advantages, does not\nundermine adversarial robustness.\n", "date": "2023-08-16", "categories": []}
{"id": "2206.08242", "title": "Catastrophic overfitting can be induced with discriminative non-robust\n  features", "abstract": "  Adversarial training (AT) is the de facto method for building robust neural\nnetworks, but it can be computationally expensive. To mitigate this, fast\nsingle-step attacks can be used, but this may lead to catastrophic overfitting\n(CO). This phenomenon appears when networks gain non-trivial robustness during\nthe first stages of AT, but then reach a breaking point where they become\nvulnerable in just a few iterations. The mechanisms that lead to this failure\nmode are still poorly understood. In this work, we study the onset of CO in\nsingle-step AT methods through controlled modifications of typical datasets of\nnatural images. In particular, we show that CO can be induced at much smaller\n$\\epsilon$ values than it was observed before just by injecting images with\nseemingly innocuous features. These features aid non-robust classification but\nare not enough to achieve robustness on their own. Through extensive\nexperiments we analyze this novel phenomenon and discover that the presence of\nthese easy features induces a learning shortcut that leads to CO. Our findings\nprovide new insights into the mechanisms of CO and improve our understanding of\nthe dynamics of AT. The code to reproduce our experiments can be found at\nhttps://github.com/gortizji/co_features.\n", "date": "2022-06-16", "categories": []}
{"id": "2207.12391", "title": "SegPGD: An Effective and Efficient Adversarial Attack for Evaluating and\n  Boosting Segmentation Robustness", "abstract": "  Deep neural network-based image classifications are vulnerable to adversarial\nperturbations. The image classifications can be easily fooled by adding\nartificial small and imperceptible perturbations to input images. As one of the\nmost effective defense strategies, adversarial training was proposed to address\nthe vulnerability of classification models, where the adversarial examples are\ncreated and injected into training data during training. The attack and defense\nof classification models have been intensively studied in past years. Semantic\nsegmentation, as an extension of classifications, has also received great\nattention recently. Recent work shows a large number of attack iterations are\nrequired to create effective adversarial examples to fool segmentation models.\nThe observation makes both robustness evaluation and adversarial training on\nsegmentation models challenging. In this work, we propose an effective and\nefficient segmentation attack method, dubbed SegPGD. Besides, we provide a\nconvergence analysis to show the proposed SegPGD can create more effective\nadversarial examples than PGD under the same number of attack iterations.\nFurthermore, we propose to apply our SegPGD as the underlying attack method for\nsegmentation adversarial training. Since SegPGD can create more effective\nadversarial examples, the adversarial training with our SegPGD can boost the\nrobustness of segmentation models. Our proposals are also verified with\nexperiments on popular Segmentation model architectures and standard\nsegmentation datasets.\n", "date": "2022-07-25", "categories": []}
{"id": "2308.07546", "title": "3DHacker: Spectrum-based Decision Boundary Generation for Hard-label 3D\n  Point Cloud Attack", "abstract": "  With the maturity of depth sensors, the vulnerability of 3D point cloud\nmodels has received increasing attention in various applications such as\nautonomous driving and robot navigation. Previous 3D adversarial attackers\neither follow the white-box setting to iteratively update the coordinate\nperturbations based on gradients, or utilize the output model logits to\nestimate noisy gradients in the black-box setting. However, these attack\nmethods are hard to be deployed in real-world scenarios since realistic 3D\napplications will not share any model details to users. Therefore, we explore a\nmore challenging yet practical 3D attack setting, \\textit{i.e.}, attacking\npoint clouds with black-box hard labels, in which the attacker can only have\naccess to the prediction label of the input. To tackle this setting, we propose\na novel 3D attack method, termed \\textbf{3D} \\textbf{H}ard-label\natt\\textbf{acker} (\\textbf{3DHacker}), based on the developed decision boundary\nalgorithm to generate adversarial samples solely with the knowledge of class\nlabels. Specifically, to construct the class-aware model decision boundary,\n3DHacker first randomly fuses two point clouds of different classes in the\nspectral domain to craft their intermediate sample with high imperceptibility,\nthen projects it onto the decision boundary via binary search. To restrict the\nfinal perturbation size, 3DHacker further introduces an iterative optimization\nstrategy to move the intermediate sample along the decision boundary for\ngenerating adversarial point clouds with smallest trivial perturbations.\nExtensive evaluations show that, even in the challenging hard-label setting,\n3DHacker still competitively outperforms existing 3D attacks regarding the\nattack performance as well as adversary quality.\n", "date": "2023-08-15", "categories": []}
{"id": "2308.07625", "title": "Backpropagation Path Search On Adversarial Transferability", "abstract": "  Deep neural networks are vulnerable to adversarial examples, dictating the\nimperativeness to test the model's robustness before deployment. Transfer-based\nattackers craft adversarial examples against surrogate models and transfer them\nto victim models deployed in the black-box situation. To enhance the\nadversarial transferability, structure-based attackers adjust the\nbackpropagation path to avoid the attack from overfitting the surrogate model.\nHowever, existing structure-based attackers fail to explore the convolution\nmodule in CNNs and modify the backpropagation graph heuristically, leading to\nlimited effectiveness. In this paper, we propose backPropagation pAth Search\n(PAS), solving the aforementioned two problems. We first propose SkipConv to\nadjust the backpropagation path of convolution by structural\nreparameterization. To overcome the drawback of heuristically designed\nbackpropagation paths, we further construct a DAG-based search space, utilize\none-step approximation for path evaluation and employ Bayesian Optimization to\nsearch for the optimal path. We conduct comprehensive experiments in a wide\nrange of transfer settings, showing that PAS improves the attack success rate\nby a huge margin for both normally trained and defense models.\n", "date": "2023-08-15", "categories": []}
{"id": "2308.07673", "title": "A Review of Adversarial Attacks in Computer Vision", "abstract": "  Deep neural networks have been widely used in various downstream tasks,\nespecially those safety-critical scenario such as autonomous driving, but deep\nnetworks are often threatened by adversarial samples. Such adversarial attacks\ncan be invisible to human eyes, but can lead to DNN misclassification, and\noften exhibits transferability between deep learning and machine learning\nmodels and real-world achievability. Adversarial attacks can be divided into\nwhite-box attacks, for which the attacker knows the parameters and gradient of\nthe model, and black-box attacks, for the latter, the attacker can only obtain\nthe input and output of the model. In terms of the attacker's purpose, it can\nbe divided into targeted attacks and non-targeted attacks, which means that the\nattacker wants the model to misclassify the original sample into the specified\nclass, which is more practical, while the non-targeted attack just needs to\nmake the model misclassify the sample. The black box setting is a scenario we\nwill encounter in practice.\n", "date": "2023-08-15", "categories": []}
{"id": "2308.07874", "title": "SEDA: Self-Ensembling ViT with Defensive Distillation and Adversarial\n  Training for robust Chest X-rays Classification", "abstract": "  Deep Learning methods have recently seen increased adoption in medical\nimaging applications. However, elevated vulnerabilities have been explored in\nrecent Deep Learning solutions, which can hinder future adoption. Particularly,\nthe vulnerability of Vision Transformer (ViT) to adversarial, privacy, and\nconfidentiality attacks raise serious concerns about their reliability in\nmedical settings. This work aims to enhance the robustness of self-ensembling\nViTs for the tuberculosis chest x-ray classification task. We propose\nSelf-Ensembling ViT with defensive Distillation and Adversarial training\n(SEDA). SEDA utilizes efficient CNN blocks to learn spatial features with\nvarious levels of abstraction from feature representations extracted from\nintermediate ViT blocks, that are largely unaffected by adversarial\nperturbations. Furthermore, SEDA leverages adversarial training in combination\nwith defensive distillation for improved robustness against adversaries.\nTraining using adversarial examples leads to better model generalizability and\nimproves its ability to handle perturbations. Distillation using soft\nprobabilities introduces uncertainty and variation into the output\nprobabilities, making it more difficult for adversarial and privacy attacks.\nExtensive experiments performed with the proposed architecture and training\nparadigm on publicly available Tuberculosis x-ray dataset shows SOTA efficacy\nof SEDA compared to SEViT in terms of computational efficiency with 70x times\nlighter framework and enhanced robustness of +9%.\n", "date": "2023-08-15", "categories": []}
{"id": "2211.14860", "title": "Foiling Explanations in Deep Neural Networks", "abstract": "  Deep neural networks (DNNs) have greatly impacted numerous fields over the\npast decade. Yet despite exhibiting superb performance over many problems,\ntheir black-box nature still poses a significant challenge with respect to\nexplainability. Indeed, explainable artificial intelligence (XAI) is crucial in\nseveral fields, wherein the answer alone -- sans a reasoning of how said answer\nwas derived -- is of little value. This paper uncovers a troubling property of\nexplanation methods for image-based DNNs: by making small visual changes to the\ninput image -- hardly influencing the network's output -- we demonstrate how\nexplanations may be arbitrarily manipulated through the use of evolution\nstrategies. Our novel algorithm, AttaXAI, a model-agnostic, adversarial attack\non XAI algorithms, only requires access to the output logits of a classifier\nand to the explanation map; these weak assumptions render our approach highly\nuseful where real-world models and data are concerned. We compare our method's\nperformance on two benchmark datasets -- CIFAR100 and ImageNet -- using four\ndifferent pretrained deep-learning models: VGG16-CIFAR100, VGG16-ImageNet,\nMobileNet-CIFAR100, and Inception-v3-ImageNet. We find that the XAI methods can\nbe manipulated without the use of gradients or other model internals. Our novel\nalgorithm is successfully able to manipulate an image in a manner imperceptible\nto the human eye, such that the XAI method outputs a specific explanation map.\nTo our knowledge, this is the first such method in a black-box setting, and we\nbelieve it has significant value where explainability is desired, required, or\nlegally mandatory.\n", "date": "2022-11-27", "categories": []}
{"id": "2303.04980", "title": "Decision-BADGE: Decision-based Adversarial Batch Attack with Directional\n  Gradient Estimation", "abstract": "  The susceptibility of deep neural networks (DNNs) to adversarial examples has\nprompted an increase in the deployment of adversarial attacks. Image-agnostic\nuniversal adversarial perturbations (UAPs) are much more threatening, but many\nlimitations exist to implementing UAPs in real-world scenarios where only\nbinary decisions are returned. In this research, we propose Decision-BADGE, a\nnovel method to craft universal adversarial perturbations for executing\ndecision-based black-box attacks. To optimize perturbation with decisions, we\naddressed two challenges, namely the magnitude and the direction of the\ngradient. First, we use batch loss, differences from distributions of ground\ntruth, and accumulating decisions in batches to determine the magnitude of the\ngradient. This magnitude is applied in the direction of the revised\nsimultaneous perturbation stochastic approximation (SPSA) to update the\nperturbation. This simple yet efficient method can be easily extended to\nscore-based attacks as well as targeted attacks. Experimental validation across\nmultiple victim models demonstrates that the Decision-BADGE outperforms\nexisting attack methods, even image-specific and score-based attacks. In\nparticular, our proposed method shows a superior success rate with less\ntraining time. The research also shows that Decision-BADGE can successfully\ndeceive unseen victim models and accurately target specific classes.\n", "date": "2023-03-09", "categories": []}
{"id": "2307.12280", "title": "Downstream-agnostic Adversarial Examples", "abstract": "  Self-supervised learning usually uses a large amount of unlabeled data to\npre-train an encoder which can be used as a general-purpose feature extractor,\nsuch that downstream users only need to perform fine-tuning operations to enjoy\nthe benefit of \"large model\". Despite this promising prospect, the security of\npre-trained encoder has not been thoroughly investigated yet, especially when\nthe pre-trained encoder is publicly available for commercial use.\n  In this paper, we propose AdvEncoder, the first framework for generating\ndownstream-agnostic universal adversarial examples based on the pre-trained\nencoder. AdvEncoder aims to construct a universal adversarial perturbation or\npatch for a set of natural images that can fool all the downstream tasks\ninheriting the victim pre-trained encoder. Unlike traditional adversarial\nexample works, the pre-trained encoder only outputs feature vectors rather than\nclassification labels. Therefore, we first exploit the high frequency component\ninformation of the image to guide the generation of adversarial examples. Then\nwe design a generative attack framework to construct adversarial\nperturbations/patches by learning the distribution of the attack surrogate\ndataset to improve their attack success rates and transferability. Our results\nshow that an attacker can successfully attack downstream tasks without knowing\neither the pre-training dataset or the downstream dataset. We also tailor four\ndefenses for pre-trained encoders, the results of which further prove the\nattack ability of AdvEncoder.\n", "date": "2023-07-23", "categories": []}
{"id": "2308.04909", "title": "Adversarial Deep Reinforcement Learning for Cyber Security in Software\n  Defined Networks", "abstract": "  This paper focuses on the impact of leveraging autonomous offensive\napproaches in Deep Reinforcement Learning (DRL) to train more robust agents by\nexploring the impact of applying adversarial learning to DRL for autonomous\nsecurity in Software Defined Networks (SDN). Two algorithms, Double Deep\nQ-Networks (DDQN) and Neural Episodic Control to Deep Q-Network (NEC2DQN or\nN2D), are compared. NEC2DQN was proposed in 2018 and is a new member of the\ndeep q-network (DQN) family of algorithms. The attacker has full observability\nof the environment and access to a causative attack that uses state\nmanipulation in an attempt to poison the learning process. The implementation\nof the attack is done under a white-box setting, in which the attacker has\naccess to the defender's model and experiences. Two games are played; in the\nfirst game, DDQN is a defender and N2D is an attacker, and in second game, the\nroles are reversed. The games are played twice; first, without an active\ncausative attack and secondly, with an active causative attack. For execution,\nthree sets of game results are recorded in which a single set consists of 10\ngame runs. The before and after results are then compared in order to see if\nthere was actually an improvement or degradation. The results show that with\nminute parameter changes made to the algorithms, there was growth in the\nattacker's role, since it is able to win games. Implementation of the\nadversarial learning by the introduction of the causative attack showed the\nalgorithms are still able to defend the network according to their strengths.\n", "date": "2023-08-09", "categories": []}
{"id": "2308.06467", "title": "Not So Robust After All: Evaluating the Robustness of Deep Neural\n  Networks to Unseen Adversarial Attacks", "abstract": "  Deep neural networks (DNNs) have gained prominence in various applications,\nsuch as classification, recognition, and prediction, prompting increased\nscrutiny of their properties. A fundamental attribute of traditional DNNs is\ntheir vulnerability to modifications in input data, which has resulted in the\ninvestigation of adversarial attacks. These attacks manipulate the data in\norder to mislead a DNN. This study aims to challenge the efficacy and\ngeneralization of contemporary defense mechanisms against adversarial attacks.\nSpecifically, we explore the hypothesis proposed by Ilyas et. al, which posits\nthat DNN image features can be either robust or non-robust, with adversarial\nattacks targeting the latter. This hypothesis suggests that training a DNN on a\ndataset consisting solely of robust features should produce a model resistant\nto adversarial attacks. However, our experiments demonstrate that this is not\nuniversally true. To gain further insights into our findings, we analyze the\nimpact of adversarial attack norms on DNN representations, focusing on samples\nsubjected to $L_2$ and $L_{\\infty}$ norm attacks. Further, we employ canonical\ncorrelation analysis, visualize the representations, and calculate the mean\ndistance between these representations and various DNN decision boundaries. Our\nresults reveal a significant difference between $L_2$ and $L_{\\infty}$ norms,\nwhich could provide insights into the potential dangers posed by $L_{\\infty}$\nnorm attacks, previously underestimated by the research community.\n", "date": "2023-08-12", "categories": []}
{"id": "2308.06612", "title": "On the Interplay of Convolutional Padding and Adversarial Robustness", "abstract": "  It is common practice to apply padding prior to convolution operations to\npreserve the resolution of feature-maps in Convolutional Neural Networks (CNN).\nWhile many alternatives exist, this is often achieved by adding a border of\nzeros around the inputs. In this work, we show that adversarial attacks often\nresult in perturbation anomalies at the image boundaries, which are the areas\nwhere padding is used. Consequently, we aim to provide an analysis of the\ninterplay between padding and adversarial attacks and seek an answer to the\nquestion of how different padding modes (or their absence) affect adversarial\nrobustness in various scenarios.\n", "date": "2023-08-12", "categories": []}
{"id": "2308.07026", "title": "AdvCLIP: Downstream-agnostic Adversarial Examples in Multimodal\n  Contrastive Learning", "abstract": "  Multimodal contrastive learning aims to train a general-purpose feature\nextractor, such as CLIP, on vast amounts of raw, unlabeled paired image-text\ndata. This can greatly benefit various complex downstream tasks, including\ncross-modal image-text retrieval and image classification. Despite its\npromising prospect, the security issue of cross-modal pre-trained encoder has\nnot been fully explored yet, especially when the pre-trained encoder is\npublicly available for commercial use.\n  In this work, we propose AdvCLIP, the first attack framework for generating\ndownstream-agnostic adversarial examples based on cross-modal pre-trained\nencoders. AdvCLIP aims to construct a universal adversarial patch for a set of\nnatural images that can fool all the downstream tasks inheriting the victim\ncross-modal pre-trained encoder. To address the challenges of heterogeneity\nbetween different modalities and unknown downstream tasks, we first build a\ntopological graph structure to capture the relevant positions between target\nsamples and their neighbors. Then, we design a topology-deviation based\ngenerative adversarial network to generate a universal adversarial patch. By\nadding the patch to images, we minimize their embeddings similarity to\ndifferent modality and perturb the sample distribution in the feature space,\nachieving unviersal non-targeted attacks. Our results demonstrate the excellent\nattack performance of AdvCLIP on two types of downstream tasks across eight\ndatasets. We also tailor three popular defenses to mitigate AdvCLIP,\nhighlighting the need for new defense mechanisms to defend cross-modal\npre-trained encoders.\n", "date": "2023-08-14", "categories": []}
{"id": "2207.04129", "title": "How many perturbations break this model? Evaluating robustness beyond\n  adversarial accuracy", "abstract": "  Robustness to adversarial attacks is typically evaluated with adversarial\naccuracy. While essential, this metric does not capture all aspects of\nrobustness and in particular leaves out the question of how many perturbations\ncan be found for each point. In this work, we introduce an alternative\napproach, adversarial sparsity, which quantifies how difficult it is to find a\nsuccessful perturbation given both an input point and a constraint on the\ndirection of the perturbation. We show that sparsity provides valuable insight\ninto neural networks in multiple ways: for instance, it illustrates important\ndifferences between current state-of-the-art robust models them that accuracy\nanalysis does not, and suggests approaches for improving their robustness. When\napplying broken defenses effective against weak attacks but not strong ones,\nsparsity can discriminate between the totally ineffective and the partially\neffective defenses. Finally, with sparsity we can measure increases in\nrobustness that do not affect accuracy: we show for example that data\naugmentation can by itself increase adversarial robustness, without using\nadversarial training.\n", "date": "2022-07-08", "categories": []}
{"id": "2210.17316", "title": "There is more than one kind of robustness: Fooling Whisper with\n  adversarial examples", "abstract": "  Whisper is a recent Automatic Speech Recognition (ASR) model displaying\nimpressive robustness to both out-of-distribution inputs and random noise. In\nthis work, we show that this robustness does not carry over to adversarial\nnoise. We show that we can degrade Whisper performance dramatically, or even\ntranscribe a target sentence of our choice, by generating very small input\nperturbations with Signal Noise Ratio of 35-45dB. We also show that by fooling\nthe Whisper language detector we can very easily degrade the performance of\nmultilingual models. These vulnerabilities of a widely popular open-source\nmodel have practical security implications and emphasize the need for\nadversarially robust ASR.\n", "date": "2022-10-26", "categories": []}
{"id": "2303.15109", "title": "Improving the Transferability of Adversarial Examples via Direction\n  Tuning", "abstract": "  In the transfer-based adversarial attacks, adversarial examples are only\ngenerated by the surrogate models and achieve effective perturbation in the\nvictim models. Although considerable efforts have been developed on improving\nthe transferability of adversarial examples generated by transfer-based\nadversarial attacks, our investigation found that, the big deviation between\nthe actual and steepest update directions of the current transfer-based\nadversarial attacks is caused by the large update step length, resulting in the\ngenerated adversarial examples can not converge well. However, directly\nreducing the update step length will lead to serious update oscillation so that\nthe generated adversarial examples also can not achieve great transferability\nto the victim models. To address these issues, a novel transfer-based attack,\nnamely direction tuning attack, is proposed to not only decrease the update\ndeviation in the large step length, but also mitigate the update oscillation in\nthe small sampling step length, thereby making the generated adversarial\nexamples converge well to achieve great transferability on victim models. In\naddition, a network pruning method is proposed to smooth the decision boundary,\nthereby further decreasing the update oscillation and enhancing the\ntransferability of the generated adversarial examples. The experiment results\non ImageNet demonstrate that the average attack success rate (ASR) of the\nadversarial examples generated by our method can be improved from 87.9\\% to\n94.5\\% on five victim models without defenses, and from 69.1\\% to 76.2\\% on\neight advanced defense methods, in comparison with that of latest\ngradient-based attacks.\n", "date": "2023-03-27", "categories": []}
{"id": "2308.06015", "title": "Enhancing Generalization of Universal Adversarial Perturbation through\n  Gradient Aggregation", "abstract": "  Deep neural networks are vulnerable to universal adversarial perturbation\n(UAP), an instance-agnostic perturbation capable of fooling the target model\nfor most samples. Compared to instance-specific adversarial examples, UAP is\nmore challenging as it needs to generalize across various samples and models.\nIn this paper, we examine the serious dilemma of UAP generation methods from a\ngeneralization perspective -- the gradient vanishing problem using small-batch\nstochastic gradient optimization and the local optima problem using large-batch\noptimization. To address these problems, we propose a simple and effective\nmethod called Stochastic Gradient Aggregation (SGA), which alleviates the\ngradient vanishing and escapes from poor local optima at the same time.\nSpecifically, SGA employs the small-batch training to perform multiple\niterations of inner pre-search. Then, all the inner gradients are aggregated as\na one-step gradient estimation to enhance the gradient stability and reduce\nquantization errors. Extensive experiments on the standard ImageNet dataset\ndemonstrate that our method significantly enhances the generalization ability\nof UAP and outperforms other state-of-the-art methods. The code is available at\nhttps://github.com/liuxuannan/Stochastic-Gradient-Aggregation.\n", "date": "2023-08-11", "categories": []}
{"id": "2308.06173", "title": "Physical Adversarial Attacks For Camera-based Smart Systems: Current\n  Trends, Categorization, Applications, Research Challenges, and Future Outlook", "abstract": "  In this paper, we present a comprehensive survey of the current trends\nfocusing specifically on physical adversarial attacks. We aim to provide a\nthorough understanding of the concept of physical adversarial attacks,\nanalyzing their key characteristics and distinguishing features. Furthermore,\nwe explore the specific requirements and challenges associated with executing\nattacks in the physical world. Our article delves into various physical\nadversarial attack methods, categorized according to their target tasks in\ndifferent applications, including classification, detection, face recognition,\nsemantic segmentation and depth estimation. We assess the performance of these\nattack methods in terms of their effectiveness, stealthiness, and robustness.\nWe examine how each technique strives to ensure the successful manipulation of\nDNNs while mitigating the risk of detection and withstanding real-world\ndistortions. Lastly, we discuss the current challenges and outline potential\nfuture research directions in the field of physical adversarial attacks. We\nhighlight the need for enhanced defense mechanisms, the exploration of novel\nattack strategies, the evaluation of attacks in different application domains,\nand the establishment of standardized benchmarks and evaluation criteria for\nphysical adversarial attacks. Through this comprehensive survey, we aim to\nprovide a valuable resource for researchers, practitioners, and policymakers to\ngain a holistic understanding of physical adversarial attacks in computer\nvision and facilitate the development of robust and secure DNN-based systems.\n", "date": "2023-08-11", "categories": []}
{"id": "2210.04087", "title": "Symmetry Defense Against CNN Adversarial Perturbation Attacks", "abstract": "  This paper uses symmetry to make Convolutional Neural Network classifiers\n(CNNs) robust against adversarial perturbation attacks. Such attacks add\nperturbation to original images to generate adversarial images that fool\nclassifiers such as road sign classifiers of autonomous vehicles. Although\nsymmetry is a pervasive aspect of the natural world, CNNs are unable to handle\nsymmetry well. For example, a CNN can classify an image differently from its\nmirror image. For an adversarial image that misclassifies with a wrong label\n$l_w$, CNN inability to handle symmetry means that a symmetric adversarial\nimage can classify differently from the wrong label $l_w$. Further than that,\nwe find that the classification of a symmetric adversarial image reverts to the\ncorrect label. To classify an image when adversaries are unaware of the\ndefense, we apply symmetry to the image and use the classification label of the\nsymmetric image. To classify an image when adversaries are aware of the\ndefense, we use mirror symmetry and pixel inversion symmetry to form a symmetry\ngroup. We apply all the group symmetries to the image and decide on the output\nlabel based on the agreement of any two of the classification labels of the\nsymmetry images. Adaptive attacks fail because they need to rely on loss\nfunctions that use conflicting CNN output values for symmetric images. Without\nattack knowledge, the proposed symmetry defense succeeds against both\ngradient-based and random-search attacks, with up to near-default accuracies\nfor ImageNet. The defense even improves the classification accuracy of original\nimages.\n", "date": "2022-10-08", "categories": []}
{"id": "2301.10822", "title": "RobustPdM: Designing Robust Predictive Maintenance against Adversarial\n  Attacks", "abstract": "  The state-of-the-art predictive maintenance (PdM) techniques have shown great\nsuccess in reducing maintenance costs and downtime of complicated machines\nwhile increasing overall productivity through extensive utilization of\nInternet-of-Things (IoT) and Deep Learning (DL). Unfortunately, IoT sensors and\nDL algorithms are both prone to cyber-attacks. For instance, DL algorithms are\nknown for their susceptibility to adversarial examples. Such adversarial\nattacks are vastly under-explored in the PdM domain. This is because the\nadversarial attacks in the computer vision domain for classification tasks\ncannot be directly applied to the PdM domain for multivariate time series (MTS)\nregression tasks. In this work, we propose an end-to-end methodology to design\nadversarially robust PdM systems by extensively analyzing the effect of\ndifferent types of adversarial attacks and proposing a novel adversarial\ndefense technique for DL-enabled PdM models. First, we propose novel MTS\nProjected Gradient Descent (PGD) and MTS PGD with random restarts (PGD_r)\nattacks. Then, we evaluate the impact of MTS PGD and PGD_r along with MTS Fast\nGradient Sign Method (FGSM) and MTS Basic Iterative Method (BIM) on Long\nShort-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural\nNetwork (CNN), and Bi-directional LSTM based PdM system. Our results using\nNASA's turbofan engine dataset show that adversarial attacks can cause a severe\ndefect (up to 11X) in the RUL prediction, outperforming the effectiveness of\nthe state-of-the-art PdM attacks by 3X. Furthermore, we present a novel\napproximate adversarial training method to defend against adversarial attacks.\nWe observe that approximate adversarial training can significantly improve the\nrobustness of PdM models (up to 54X) and outperforms the state-of-the-art PdM\ndefense methods by offering 3X more robustness.\n", "date": "2023-01-25", "categories": []}
{"id": "2307.16361", "title": "Benchmarking and Analyzing Robust Point Cloud Recognition: Bag of Tricks\n  for Defending Adversarial Examples", "abstract": "  Deep Neural Networks (DNNs) for 3D point cloud recognition are vulnerable to\nadversarial examples, threatening their practical deployment. Despite the many\nresearch endeavors have been made to tackle this issue in recent years, the\ndiversity of adversarial examples on 3D point clouds makes them more\nchallenging to defend against than those on 2D images. For examples, attackers\ncan generate adversarial examples by adding, shifting, or removing points.\nConsequently, existing defense strategies are hard to counter unseen point\ncloud adversarial examples. In this paper, we first establish a comprehensive,\nand rigorous point cloud adversarial robustness benchmark to evaluate\nadversarial robustness, which can provide a detailed understanding of the\neffects of the defense and attack methods. We then collect existing defense\ntricks in point cloud adversarial defenses and then perform extensive and\nsystematic experiments to identify an effective combination of these tricks.\nFurthermore, we propose a hybrid training augmentation methods that consider\nvarious types of point cloud adversarial examples to adversarial training,\nsignificantly improving the adversarial robustness. By combining these tricks,\nwe construct a more robust defense framework achieving an average accuracy of\n83.45\\% against various attacks, demonstrating its capability to enabling\nrobust learners. Our codebase are open-sourced on:\n\\url{https://github.com/qiufan319/benchmark_pc_attack.git}.\n", "date": "2023-07-31", "categories": []}
{"id": "2308.05575", "title": "Symmetry Defense Against XGBoost Adversarial Perturbation Attacks", "abstract": "  We examine whether symmetry can be used to defend tree-based ensemble\nclassifiers such as gradient-boosting decision trees (GBDTs) against\nadversarial perturbation attacks. The idea is based on a recent symmetry\ndefense for convolutional neural network classifiers (CNNs) that utilizes CNNs'\nlack of invariance with respect to symmetries. CNNs lack invariance because\nthey can classify a symmetric sample, such as a horizontally flipped image,\ndifferently from the original sample. CNNs' lack of invariance also means that\nCNNs can classify symmetric adversarial samples differently from the incorrect\nclassification of adversarial samples. Using CNNs' lack of invariance, the\nrecent CNN symmetry defense has shown that the classification of symmetric\nadversarial samples reverts to the correct sample classification. In order to\napply the same symmetry defense to GBDTs, we examine GBDT invariance and are\nthe first to show that GBDTs also lack invariance with respect to symmetries.\nWe apply and evaluate the GBDT symmetry defense for nine datasets against six\nperturbation attacks with a threat model that ranges from zero-knowledge to\nperfect-knowledge adversaries. Using the feature inversion symmetry against\nzero-knowledge adversaries, we achieve up to 100% accuracy on adversarial\nsamples even when default and robust classifiers have 0% accuracy. Using the\nfeature inversion and horizontal flip symmetries against perfect-knowledge\nadversaries, we achieve up to over 95% accuracy on adversarial samples for the\nGBDT classifier of the F-MNIST dataset even when default and robust classifiers\nhave 0% accuracy.\n", "date": "2023-08-10", "categories": []}
{"id": "2207.11378", "title": "Do Perceptually Aligned Gradients Imply Adversarial Robustness?", "abstract": "  Adversarially robust classifiers possess a trait that non-robust models do\nnot -- Perceptually Aligned Gradients (PAG). Their gradients with respect to\nthe input align well with human perception. Several works have identified PAG\nas a byproduct of robust training, but none have considered it as a standalone\nphenomenon nor studied its own implications. In this work, we focus on this\ntrait and test whether \\emph{Perceptually Aligned Gradients imply Robustness}.\nTo this end, we develop a novel objective to directly promote PAG in training\nclassifiers and examine whether models with such gradients are more robust to\nadversarial attacks. Extensive experiments on multiple datasets and\narchitectures validate that models with aligned gradients exhibit significant\nrobustness, exposing the surprising bidirectional connection between PAG and\nrobustness. Lastly, we show that better gradient alignment leads to increased\nrobustness and harness this observation to boost the robustness of existing\nadversarial training techniques.\n", "date": "2022-07-22", "categories": []}
{"id": "2211.02408", "title": "Rickrolling the Artist: Injecting Backdoors into Text Encoders for\n  Text-to-Image Synthesis", "abstract": "  While text-to-image synthesis currently enjoys great popularity among\nresearchers and the general public, the security of these models has been\nneglected so far. Many text-guided image generation models rely on pre-trained\ntext encoders from external sources, and their users trust that the retrieved\nmodels will behave as promised. Unfortunately, this might not be the case. We\nintroduce backdoor attacks against text-guided generative models and\ndemonstrate that their text encoders pose a major tampering risk. Our attacks\nonly slightly alter an encoder so that no suspicious model behavior is apparent\nfor image generations with clean prompts. By then inserting a single character\ntrigger into the prompt, e.g., a non-Latin character or emoji, the adversary\ncan trigger the model to either generate images with pre-defined attributes or\nimages following a hidden, potentially malicious description. We empirically\ndemonstrate the high effectiveness of our attacks on Stable Diffusion and\nhighlight that the injection process of a single backdoor takes less than two\nminutes. Besides phrasing our approach solely as an attack, it can also force\nan encoder to forget phrases related to certain concepts, such as nudity or\nviolence, and help to make image generation safer.\n", "date": "2022-11-04", "categories": []}
{"id": "2209.05785", "title": "Adversarial Coreset Selection for Efficient Robust Training", "abstract": "  Neural networks are vulnerable to adversarial attacks: adding well-crafted,\nimperceptible perturbations to their input can modify their output. Adversarial\ntraining is one of the most effective approaches to training robust models\nagainst such attacks. Unfortunately, this method is much slower than vanilla\ntraining of neural networks since it needs to construct adversarial examples\nfor the entire training data at every iteration. By leveraging the theory of\ncoreset selection, we show how selecting a small subset of training data\nprovides a principled approach to reducing the time complexity of robust\ntraining. To this end, we first provide convergence guarantees for adversarial\ncoreset selection. In particular, we show that the convergence bound is\ndirectly related to how well our coresets can approximate the gradient computed\nover the entire training data. Motivated by our theoretical analysis, we\npropose using this gradient approximation error as our adversarial coreset\nselection objective to reduce the training set size effectively. Once built, we\nrun adversarial training over this subset of the training data. Unlike existing\nmethods, our approach can be adapted to a wide variety of training objectives,\nincluding TRADES, $\\ell_p$-PGD, and Perceptual Adversarial Training. We conduct\nextensive experiments to demonstrate that our approach speeds up adversarial\ntraining by 2-3 times while experiencing a slight degradation in the clean and\nrobust accuracy.\n", "date": "2022-09-13", "categories": []}
{"id": "2308.03792", "title": "Multi-attacks: Many images $+$ the same adversarial attack $\\to$ many\n  target labels", "abstract": "  We show that we can easily design a single adversarial perturbation $P$ that\nchanges the class of $n$ images $X_1,X_2,\\dots,X_n$ from their original,\nunperturbed classes $c_1, c_2,\\dots,c_n$ to desired (not necessarily all the\nsame) classes $c^*_1,c^*_2,\\dots,c^*_n$ for up to hundreds of images and target\nclasses at once. We call these \\textit{multi-attacks}. Characterizing the\nmaximum $n$ we can achieve under different conditions such as image resolution,\nwe estimate the number of regions of high class confidence around a particular\nimage in the space of pixels to be around $10^{\\mathcal{O}(100)}$, posing a\nsignificant problem for exhaustive defense strategies. We show several\nimmediate consequences of this: adversarial attacks that change the resulting\nclass based on their intensity, and scale-independent adversarial examples. To\ndemonstrate the redundancy and richness of class decision boundaries in the\npixel space, we look for its two-dimensional sections that trace images and\nspell words using particular classes. We also show that ensembling reduces\nsusceptibility to multi-attacks, and that classifiers trained on random labels\nare more susceptible. Our code is available on GitHub.\n", "date": "2023-08-04", "categories": []}
{"id": "2308.03906", "title": "TIJO: Trigger Inversion with Joint Optimization for Defending Multimodal\n  Backdoored Models", "abstract": "  We present a Multimodal Backdoor Defense technique TIJO (Trigger Inversion\nusing Joint Optimization). Recent work arXiv:2112.07668 has demonstrated\nsuccessful backdoor attacks on multimodal models for the Visual Question\nAnswering task. Their dual-key backdoor trigger is split across two modalities\n(image and text), such that the backdoor is activated if and only if the\ntrigger is present in both modalities. We propose TIJO that defends against\ndual-key attacks through a joint optimization that reverse-engineers the\ntrigger in both the image and text modalities. This joint optimization is\nchallenging in multimodal models due to the disconnected nature of the visual\npipeline which consists of an offline feature extractor, whose output is then\nfused with the text using a fusion module. The key insight enabling the joint\noptimization in TIJO is that the trigger inversion needs to be carried out in\nthe object detection box feature space as opposed to the pixel space. We\ndemonstrate the effectiveness of our method on the TrojVQA benchmark, where\nTIJO improves upon the state-of-the-art unimodal methods from an AUC of 0.6 to\n0.92 on multimodal dual-key backdoors. Furthermore, our method also improves\nupon the unimodal baselines on unimodal backdoors. We present ablation studies\nand qualitative results to provide insights into our algorithm such as the\ncritical importance of overlaying the inverted feature triggers on all visual\nfeatures during trigger inversion. The prototype implementation of TIJO is\navailable at https://github.com/SRI-CSL/TIJO.\n", "date": "2023-08-07", "categories": []}
{"id": "2308.03956", "title": "Fixed Inter-Neuron Covariability Induces Adversarial Robustness", "abstract": "  The vulnerability to adversarial perturbations is a major flaw of Deep Neural\nNetworks (DNNs) that raises question about their reliability when in real-world\nscenarios. On the other hand, human perception, which DNNs are supposed to\nemulate, is highly robust to such perturbations, indicating that there may be\ncertain features of the human perception that make it robust but are not\nrepresented in the current class of DNNs. One such feature is that the activity\nof biological neurons is correlated and the structure of this correlation tends\nto be rather rigid over long spans of times, even if it hampers performance and\nlearning. We hypothesize that integrating such constraints on the activations\nof a DNN would improve its adversarial robustness, and, to test this\nhypothesis, we have developed the Self-Consistent Activation (SCA) layer, which\ncomprises of neurons whose activations are consistent with each other, as they\nconform to a fixed, but learned, covariability pattern. When evaluated on image\nand sound recognition tasks, the models with a SCA layer achieved high\naccuracy, and exhibited significantly greater robustness than multi-layer\nperceptron models to state-of-the-art Auto-PGD adversarial attacks\n\\textit{without being trained on adversarially perturbed data\n", "date": "2023-08-07", "categories": []}
{"id": "2308.04061", "title": "Enhancing Adversarial Robustness in Low-Label Regime via Adaptively\n  Weighted Regularization and Knowledge Distillation", "abstract": "  Adversarial robustness is a research area that has recently received a lot of\nattention in the quest for trustworthy artificial intelligence. However, recent\nworks on adversarial robustness have focused on supervised learning where it is\nassumed that labeled data is plentiful. In this paper, we investigate\nsemi-supervised adversarial training where labeled data is scarce. We derive\ntwo upper bounds for the robust risk and propose a regularization term for\nunlabeled data motivated by these two upper bounds. Then, we develop a\nsemi-supervised adversarial training algorithm that combines the proposed\nregularization term with knowledge distillation using a semi-supervised teacher\n(i.e., a teacher model trained using a semi-supervised learning algorithm). Our\nexperiments show that our proposed algorithm achieves state-of-the-art\nperformance with significant margins compared to existing algorithms. In\nparticular, compared to supervised learning algorithms, performance of our\nproposed algorithm is not much worse even when the amount of labeled data is\nvery small. For example, our algorithm with only 8\\% labeled data is comparable\nto supervised adversarial training algorithms that use all labeled data, both\nin terms of standard and robust accuracies on CIFAR-10.\n", "date": "2023-08-08", "categories": []}
{"id": "2304.07360", "title": "Combining Generators of Adversarial Malware Examples to Increase Evasion\n  Rate", "abstract": "  Antivirus developers are increasingly embracing machine learning as a key\ncomponent of malware defense. While machine learning achieves cutting-edge\noutcomes in many fields, it also has weaknesses that are exploited by several\nadversarial attack techniques. Many authors have presented both white-box and\nblack-box generators of adversarial malware examples capable of bypassing\nmalware detectors with varying success. We propose to combine contemporary\ngenerators in order to increase their potential. Combining different generators\ncan create more sophisticated adversarial examples that are more likely to\nevade anti-malware tools. We demonstrated this technique on five well-known\ngenerators and recorded promising results. The best-performing combination of\nAMG-random and MAB-Malware generators achieved an average evasion rate of 15.9%\nagainst top-tier antivirus products. This represents an average improvement of\nmore than 36% and 627% over using only the AMG-random and MAB-Malware\ngenerators, respectively. The generator that benefited the most from having\nanother generator follow its procedure was the FGSM injection attack, which\nimproved the evasion rate on average between 91.97% and 1,304.73%, depending on\nthe second generator used. These results demonstrate that combining different\ngenerators can significantly improve their effectiveness against leading\nantivirus programs.\n", "date": "2023-04-14", "categories": []}
{"id": "2307.13294", "title": "Imperceptible Physical Attack against Face Recognition Systems via LED\n  Illumination Modulation", "abstract": "  Although face recognition starts to play an important role in our daily life,\nwe need to pay attention that data-driven face recognition vision systems are\nvulnerable to adversarial attacks. However, the current two categories of\nadversarial attacks, namely digital attacks and physical attacks both have\ndrawbacks, with the former ones impractical and the latter one conspicuous,\nhigh-computational and inexecutable. To address the issues, we propose a\npractical, executable, inconspicuous and low computational adversarial attack\nbased on LED illumination modulation. To fool the systems, the proposed attack\ngenerates imperceptible luminance changes to human eyes through fast intensity\nmodulation of scene LED illumination and uses the rolling shutter effect of\nCMOS image sensors in face recognition systems to implant luminance information\nperturbation to the captured face images. In summary,we present a\ndenial-of-service (DoS) attack for face detection and a dodging attack for face\nverification. We also evaluate their effectiveness against well-known face\ndetection models, Dlib, MTCNN and RetinaFace , and face verification models,\nDlib, FaceNet,and ArcFace.The extensive experiments show that the success rates\nof DoS attacks against face detection models reach 97.67%, 100%, and 100%,\nrespectively, and the success rates of dodging attacks against all face\nverification models reach 100%.\n", "date": "2023-07-25", "categories": []}
{"id": "2308.02533", "title": "Improving Generalization of Adversarial Training via Robust Critical\n  Fine-Tuning", "abstract": "  Deep neural networks are susceptible to adversarial examples, posing a\nsignificant security risk in critical applications. Adversarial Training (AT)\nis a well-established technique to enhance adversarial robustness, but it often\ncomes at the cost of decreased generalization ability. This paper proposes\nRobustness Critical Fine-Tuning (RiFT), a novel approach to enhance\ngeneralization without compromising adversarial robustness. The core idea of\nRiFT is to exploit the redundant capacity for robustness by fine-tuning the\nadversarially trained model on its non-robust-critical module. To do so, we\nintroduce module robust criticality (MRC), a measure that evaluates the\nsignificance of a given module to model robustness under worst-case weight\nperturbations. Using this measure, we identify the module with the lowest MRC\nvalue as the non-robust-critical module and fine-tune its weights to obtain\nfine-tuned weights. Subsequently, we linearly interpolate between the\nadversarially trained weights and fine-tuned weights to derive the optimal\nfine-tuned model weights. We demonstrate the efficacy of RiFT on ResNet18,\nResNet34, and WideResNet34-10 models trained on CIFAR10, CIFAR100, and\nTiny-ImageNet datasets. Our experiments show that \\method can significantly\nimprove both generalization and out-of-distribution robustness by around 1.5%\nwhile maintaining or even slightly enhancing adversarial robustness. Code is\navailable at https://github.com/microsoft/robustlearn.\n", "date": "2023-08-01", "categories": []}
{"id": "2308.02897", "title": "An Adaptive Model Ensemble Adversarial Attack for Boosting Adversarial\n  Transferability", "abstract": "  While the transferability property of adversarial examples allows the\nadversary to perform black-box attacks (i.e., the attacker has no knowledge\nabout the target model), the transfer-based adversarial attacks have gained\ngreat attention. Previous works mostly study gradient variation or image\ntransformations to amplify the distortion on critical parts of inputs. These\nmethods can work on transferring across models with limited differences, i.e.,\nfrom CNNs to CNNs, but always fail in transferring across models with wide\ndifferences, such as from CNNs to ViTs. Alternatively, model ensemble\nadversarial attacks are proposed to fuse outputs from surrogate models with\ndiverse architectures to get an ensemble loss, making the generated adversarial\nexample more likely to transfer to other models as it can fool multiple models\nconcurrently. However, existing ensemble attacks simply fuse the outputs of the\nsurrogate models evenly, thus are not efficacious to capture and amplify the\nintrinsic transfer information of adversarial examples. In this paper, we\npropose an adaptive ensemble attack, dubbed AdaEA, to adaptively control the\nfusion of the outputs from each model, via monitoring the discrepancy ratio of\ntheir contributions towards the adversarial objective. Furthermore, an extra\ndisparity-reduced filter is introduced to further synchronize the update\ndirection. As a result, we achieve considerable improvement over the existing\nensemble attacks on various datasets, and the proposed AdaEA can also boost\nexisting transfer-based attacks, which further demonstrates its efficacy and\nversatility.\n", "date": "2023-08-05", "categories": []}
{"id": "2308.03163", "title": "CGBA: Curvature-aware Geometric Black-box Attack", "abstract": "  Decision-based black-box attacks often necessitate a large number of queries\nto craft an adversarial example. Moreover, decision-based attacks based on\nquerying boundary points in the estimated normal vector direction often suffer\nfrom inefficiency and convergence issues. In this paper, we propose a novel\nquery-efficient curvature-aware geometric decision-based black-box attack\n(CGBA) that conducts boundary search along a semicircular path on a restricted\n2D plane to ensure finding a boundary point successfully irrespective of the\nboundary curvature. While the proposed CGBA attack can work effectively for an\narbitrary decision boundary, it is particularly efficient in exploiting the low\ncurvature to craft high-quality adversarial examples, which is widely seen and\nexperimentally verified in commonly used classifiers under non-targeted\nattacks. In contrast, the decision boundaries often exhibit higher curvature\nunder targeted attacks. Thus, we develop a new query-efficient variant, CGBA-H,\nthat is adapted for the targeted attack. In addition, we further design an\nalgorithm to obtain a better initial boundary point at the expense of some\nextra queries, which considerably enhances the performance of the targeted\nattack. Extensive experiments are conducted to evaluate the performance of our\nproposed methods against some well-known classifiers on the ImageNet and\nCIFAR10 datasets, demonstrating the superiority of CGBA and CGBA-H over\nstate-of-the-art non-targeted and targeted attacks, respectively. The source\ncode is available at https://github.com/Farhamdur/CGBA.\n", "date": "2023-08-06", "categories": []}
{"id": "2308.03243", "title": "Unsupervised Adversarial Detection without Extra Model: Training Loss\n  Should Change", "abstract": "  Adversarial robustness poses a critical challenge in the deployment of deep\nlearning models for real-world applications. Traditional approaches to\nadversarial training and supervised detection rely on prior knowledge of attack\ntypes and access to labeled training data, which is often impractical. Existing\nunsupervised adversarial detection methods identify whether the target model\nworks properly, but they suffer from bad accuracies owing to the use of common\ncross-entropy training loss, which relies on unnecessary features and\nstrengthens adversarial attacks. We propose new training losses to reduce\nuseless features and the corresponding detection method without prior knowledge\nof adversarial attacks. The detection rate (true positive rate) against all\ngiven white-box attacks is above 93.9% except for attacks without limits\n(DF($\\infty$)), while the false positive rate is barely 2.5%. The proposed\nmethod works well in all tested attack types and the false positive rates are\neven better than the methods good at certain types.\n", "date": "2023-08-07", "categories": []}
{"id": "2308.03258", "title": "APBench: A Unified Benchmark for Availability Poisoning Attacks and\n  Defenses", "abstract": "  The efficacy of availability poisoning, a method of poisoning data by\ninjecting imperceptible perturbations to prevent its use in model training, has\nbeen a hot subject of investigation. Previous research suggested that it was\ndifficult to effectively counteract such poisoning attacks. However, the\nintroduction of various defense methods has challenged this notion. Due to the\nrapid progress in this field, the performance of different novel methods cannot\nbe accurately validated due to variations in experimental setups. To further\nevaluate the attack and defense capabilities of these poisoning methods, we\nhave developed a benchmark -- APBench for assessing the efficacy of adversarial\npoisoning. APBench consists of 9 state-of-the-art availability poisoning\nattacks, 8 defense algorithms, and 4 conventional data augmentation techniques.\nWe also have set up experiments with varying different poisoning ratios, and\nevaluated the attacks on multiple datasets and their transferability across\nmodel architectures. We further conducted a comprehensive evaluation of 2\nadditional attacks specifically targeting unsupervised models. Our results\nreveal the glaring inadequacy of existing attacks in safeguarding individual\nprivacy. APBench is open source and available to the deep learning community:\nhttps://github.com/lafeat/apbench.\n", "date": "2023-08-07", "categories": []}
{"id": "2308.03363", "title": "A reading survey on adversarial machine learning: Adversarial attacks\n  and their understanding", "abstract": "  Deep Learning has empowered us to train neural networks for complex data with\nhigh performance. However, with the growing research, several vulnerabilities\nin neural networks have been exposed. A particular branch of research,\nAdversarial Machine Learning, exploits and understands some of the\nvulnerabilities that cause the neural networks to misclassify for near original\ninput. A class of algorithms called adversarial attacks is proposed to make the\nneural networks misclassify for various tasks in different domains. With the\nextensive and growing research in adversarial attacks, it is crucial to\nunderstand the classification of adversarial attacks. This will help us\nunderstand the vulnerabilities in a systematic order and help us to mitigate\nthe effects of adversarial attacks. This article provides a survey of existing\nadversarial attacks and their understanding based on different perspectives. We\nalso provide a brief overview of existing adversarial defences and their\nlimitations in mitigating the effect of adversarial attacks. Further, we\nconclude with a discussion on the future research directions in the field of\nadversarial machine learning.\n", "date": "2023-08-07", "categories": []}
{"id": "2308.03476", "title": "Exploring the Physical World Adversarial Robustness of Vehicle Detection", "abstract": "  Adversarial attacks can compromise the robustness of real-world detection\nmodels. However, evaluating these models under real-world conditions poses\nchallenges due to resource-intensive experiments. Virtual simulations offer an\nalternative, but the absence of standardized benchmarks hampers progress.\nAddressing this, we propose an innovative instant-level data generation\npipeline using the CARLA simulator. Through this pipeline, we establish the\nDiscrete and Continuous Instant-level (DCI) dataset, enabling comprehensive\nexperiments involving three detection models and three physical adversarial\nattacks. Our findings highlight diverse model performances under adversarial\nconditions. Yolo v6 demonstrates remarkable resilience, experiencing just a\nmarginal 6.59% average drop in average precision (AP). In contrast, the ASA\nattack yields a substantial 14.51% average AP reduction, twice the effect of\nother algorithms. We also note that static scenes yield higher recognition AP\nvalues, and outcomes remain relatively consistent across varying weather\nconditions. Intriguingly, our study suggests that advancements in adversarial\nattack algorithms may be approaching its ``limitation''.In summary, our work\nunderscores the significance of adversarial attacks in real-world contexts and\nintroduces the DCI dataset as a versatile benchmark. Our findings provide\nvaluable insights for enhancing the robustness of detection models and offer\nguidance for future research endeavors in the realm of adversarial attacks.\n", "date": "2023-08-07", "categories": []}
{"id": "2308.02369", "title": "Universal Defensive Underpainting Patch: Making Your Text Invisible to\n  Optical Character Recognition", "abstract": "  Optical Character Recognition (OCR) enables automatic text extraction from\nscanned or digitized text images, but it also makes it easy to pirate valuable\nor sensitive text from these images. Previous methods to prevent OCR piracy by\ndistorting characters in text images are impractical in real-world scenarios,\nas pirates can capture arbitrary portions of the text images, rendering the\ndefenses ineffective. In this work, we propose a novel and effective defense\nmechanism termed the Universal Defensive Underpainting Patch (UDUP) that\nmodifies the underpainting of text images instead of the characters. UDUP is\ncreated through an iterative optimization process to craft a small, fixed-size\ndefensive patch that can generate non-overlapping underpainting for text images\nof any size. Experimental results show that UDUP effectively defends against\nunauthorized OCR under the setting of any screenshot range or complex image\nbackground. It is agnostic to the content, size, colors, and languages of\ncharacters, and is robust to typical image operations such as scaling and\ncompressing. In addition, the transferability of UDUP is demonstrated by\nevading several off-the-shelf OCRs. The code is available at\nhttps://github.com/QRICKDD/UDUP.\n", "date": "2023-08-04", "categories": []}
{"id": "2308.01469", "title": "VertexSerum: Poisoning Graph Neural Networks for Link Inference", "abstract": "  Graph neural networks (GNNs) have brought superb performance to various\napplications utilizing graph structural data, such as social analysis and fraud\ndetection. The graph links, e.g., social relationships and transaction history,\nare sensitive and valuable information, which raises privacy concerns when\nusing GNNs. To exploit these vulnerabilities, we propose VertexSerum, a novel\ngraph poisoning attack that increases the effectiveness of graph link stealing\nby amplifying the link connectivity leakage. To infer node adjacency more\naccurately, we propose an attention mechanism that can be embedded into the\nlink detection network. Our experiments demonstrate that VertexSerum\nsignificantly outperforms the SOTA link inference attack, improving the AUC\nscores by an average of $9.8\\%$ across four real-world datasets and three\ndifferent GNN structures. Furthermore, our experiments reveal the effectiveness\nof VertexSerum in both black-box and online learning settings, further\nvalidating its applicability in real-world scenarios.\n", "date": "2023-08-02", "categories": []}
{"id": "2308.01823", "title": "Hard Adversarial Example Mining for Improving Robust Fairness", "abstract": "  Adversarial training (AT) is widely considered the state-of-the-art technique\nfor improving the robustness of deep neural networks (DNNs) against adversarial\nexamples (AE). Nevertheless, recent studies have revealed that adversarially\ntrained models are prone to unfairness problems, restricting their\napplicability. In this paper, we empirically observe that this limitation may\nbe attributed to serious adversarial confidence overfitting, i.e., certain\nadversarial examples with overconfidence. To alleviate this problem, we propose\nHAM, a straightforward yet effective framework via adaptive Hard Adversarial\nexample Mining.HAM concentrates on mining hard adversarial examples while\ndiscarding the easy ones in an adaptive fashion. Specifically, HAM identifies\nhard AEs in terms of their step sizes needed to cross the decision boundary\nwhen calculating loss value. Besides, an early-dropping mechanism is\nincorporated to discard the easy examples at the initial stages of AE\ngeneration, resulting in efficient AT. Extensive experimental results on\nCIFAR-10, SVHN, and Imagenette demonstrate that HAM achieves significant\nimprovement in robust fairness while reducing computational cost compared to\nseveral state-of-the-art adversarial training methods. The code will be made\npublicly available.\n", "date": "2023-08-03", "categories": []}
{"id": "2308.01840", "title": "URET: Universal Robustness Evaluation Toolkit (for Evasion)", "abstract": "  Machine learning models are known to be vulnerable to adversarial evasion\nattacks as illustrated by image classification models. Thoroughly understanding\nsuch attacks is critical in order to ensure the safety and robustness of\ncritical AI tasks. However, most evasion attacks are difficult to deploy\nagainst a majority of AI systems because they have focused on image domain with\nonly few constraints. An image is composed of homogeneous, numerical,\ncontinuous, and independent features, unlike many other input types to AI\nsystems used in practice. Furthermore, some input types include additional\nsemantic and functional constraints that must be observed to generate realistic\nadversarial inputs. In this work, we propose a new framework to enable the\ngeneration of adversarial inputs irrespective of the input type and task\ndomain. Given an input and a set of pre-defined input transformations, our\nframework discovers a sequence of transformations that result in a semantically\ncorrect and functional adversarial input. We demonstrate the generality of our\napproach on several diverse machine learning tasks with various input\nrepresentations. We also show the importance of generating adversarial examples\nas they enable the deployment of mitigation techniques.\n", "date": "2023-08-03", "categories": []}
{"id": "2308.01888", "title": "FROD: Robust Object Detection for Free", "abstract": "  Object detection is a vital task in computer vision and has become an\nintegral component of numerous critical systems. However, state-of-the-art\nobject detectors, similar to their classification counterparts, are susceptible\nto small adversarial perturbations that can significantly alter their normal\nbehavior. Unlike classification, the robustness of object detectors has not\nbeen thoroughly explored. In this work, we take the initial step towards\nbridging the gap between the robustness of classification and object detection\nby leveraging adversarially trained classification models. Merely utilizing\nadversarially trained models as backbones for object detection does not result\nin robustness. We propose effective modifications to the classification-based\nbackbone to instill robustness in object detection without incurring any\ncomputational overhead. To further enhance the robustness achieved by the\nproposed modified backbone, we introduce two lightweight components: imitation\nloss and delayed adversarial training. Extensive experiments on the MS-COCO and\nPascal VOC datasets are conducted to demonstrate the effectiveness of our\nproposed approach.\n", "date": "2023-08-03", "categories": []}
{"id": "2308.00854", "title": "Training on Foveated Images Improves Robustness to Adversarial Attacks", "abstract": "  Deep neural networks (DNNs) have been shown to be vulnerable to adversarial\nattacks -- subtle, perceptually indistinguishable perturbations of inputs that\nchange the response of the model. In the context of vision, we hypothesize that\nan important contributor to the robustness of human visual perception is\nconstant exposure to low-fidelity visual stimuli in our peripheral vision. To\ninvestigate this hypothesis, we develop \\RBlur, an image transform that\nsimulates the loss in fidelity of peripheral vision by blurring the image and\nreducing its color saturation based on the distance from a given fixation\npoint. We show that compared to DNNs trained on the original images, DNNs\ntrained on images transformed by \\RBlur are substantially more robust to\nadversarial attacks, as well as other, non-adversarial, corruptions, achieving\nup to 25\\% higher accuracy on perturbed data.\n", "date": "2023-08-01", "categories": []}
{"id": "2308.00077", "title": "A Novel Deep Learning based Model to Defend Network Intrusion Detection\n  System against Adversarial Attacks", "abstract": "  Network Intrusion Detection System (NIDS) is an essential tool in securing\ncyberspace from a variety of security risks and unknown cyberattacks. A number\nof solutions have been implemented for Machine Learning (ML), and Deep Learning\n(DL) based NIDS. However, all these solutions are vulnerable to adversarial\nattacks, in which the malicious actor tries to evade or fool the model by\ninjecting adversarial perturbed examples into the system. The main aim of this\nresearch work is to study powerful adversarial attack algorithms and their\ndefence method on DL-based NIDS. Fast Gradient Sign Method (FGSM), Jacobian\nSaliency Map Attack (JSMA), Projected Gradient Descent (PGD) and Carlini &\nWagner (C&W) are four powerful adversarial attack methods implemented against\nthe NIDS. As a defence method, Adversarial Training is used to increase the\nrobustness of the NIDS model. The results are summarized in three phases, i.e.,\n1) before the adversarial attack, 2) after the adversarial attack, and 3) after\nthe adversarial defence. The Canadian Institute for Cybersecurity Intrusion\nDetection System 2017 (CICIDS-2017) dataset is used for evaluation purposes\nwith various performance measurements like f1-score, accuracy etc.\n", "date": "2023-07-31", "categories": []}
{"id": "2308.00311", "title": "Doubly Robust Instance-Reweighted Adversarial Training", "abstract": "  Assigning importance weights to adversarial data has achieved great success\nin training adversarially robust networks under limited model capacity.\nHowever, existing instance-reweighted adversarial training (AT) methods heavily\ndepend on heuristics and/or geometric interpretations to determine those\nimportance weights, making these algorithms lack rigorous theoretical\njustification/guarantee. Moreover, recent research has shown that adversarial\ntraining suffers from a severe non-uniform robust performance across the\ntraining distribution, e.g., data points belonging to some classes can be much\nmore vulnerable to adversarial attacks than others. To address both issues, in\nthis paper, we propose a novel doubly-robust instance reweighted AT framework,\nwhich allows to obtain the importance weights via exploring distributionally\nrobust optimization (DRO) techniques, and at the same time boosts the\nrobustness on the most vulnerable examples. In particular, our importance\nweights are obtained by optimizing the KL-divergence regularized loss function,\nwhich allows us to devise new algorithms with a theoretical convergence\nguarantee. Experiments on standard classification datasets demonstrate that our\nproposed approach outperforms related state-of-the-art baseline methods in\nterms of average robust performance, and at the same time improves the\nrobustness against attacks on the weakest data points. Codes will be available\nsoon.\n", "date": "2023-08-01", "categories": []}
{"id": "2308.00346", "title": "Dynamic ensemble selection based on Deep Neural Network Uncertainty\n  Estimation for Adversarial Robustness", "abstract": "  The deep neural network has attained significant efficiency in image\nrecognition. However, it has vulnerable recognition robustness under extensive\ndata uncertainty in practical applications. The uncertainty is attributed to\nthe inevitable ambient noise and, more importantly, the possible adversarial\nattack. Dynamic methods can effectively improve the defense initiative in the\narms race of attack and defense of adversarial examples. Different from the\nprevious dynamic method depend on input or decision, this work explore the\ndynamic attributes in model level through dynamic ensemble selection technology\nto further protect the model from white-box attacks and improve the robustness.\nSpecifically, in training phase the Dirichlet distribution is apply as prior of\nsub-models' predictive distribution, and the diversity constraint in parameter\nspace is introduced under the lightweight sub-models to construct alternative\nensembel model spaces. In test phase, the certain sub-models are dynamically\nselected based on their rank of uncertainty value for the final prediction to\nensure the majority accurate principle in ensemble robustness and accuracy.\nCompared with the previous dynamic method and staic adversarial traning model,\nthe presented approach can achieve significant robustness results without\ndamaging accuracy by combining dynamics and diversity property.\n", "date": "2023-08-01", "categories": []}
{"id": "2308.00556", "title": "Robust Linear Regression: Phase-Transitions and Precise Tradeoffs for\n  General Norms", "abstract": "  In this paper, we investigate the impact of test-time adversarial attacks on\nlinear regression models and determine the optimal level of robustness that any\nmodel can reach while maintaining a given level of standard predictive\nperformance (accuracy). Through quantitative estimates, we uncover fundamental\ntradeoffs between adversarial robustness and accuracy in different regimes. We\nobtain a precise characterization which distinguishes between regimes where\nrobustness is achievable without hurting standard accuracy and regimes where a\ntradeoff might be unavoidable. Our findings are empirically confirmed with\nsimple experiments that represent a variety of settings. This work applies to\nfeature covariance matrices and attack norms of any nature, and extends beyond\nprevious works in this area.\n", "date": "2023-08-01", "categories": []}
{"id": "2305.12859", "title": "Flying Adversarial Patches: Manipulating the Behavior of Deep\n  Learning-based Autonomous Multirotors", "abstract": "  Autonomous flying robots, e.g. multirotors, often rely on a neural network\nthat makes predictions based on a camera image. These deep learning (DL) models\ncan compute surprising results if applied to input images outside the training\ndomain. Adversarial attacks exploit this fault, for example, by computing small\nimages, so-called adversarial patches, that can be placed in the environment to\nmanipulate the neural network's prediction. We introduce flying adversarial\npatches, where an image is mounted on another flying robot and therefore can be\nplaced anywhere in the field of view of a victim multirotor. For an effective\nattack, we compare three methods that simultaneously optimize the adversarial\npatch and its position in the input image. We perform an empirical validation\non a publicly available DL model and dataset for autonomous multirotors.\nUltimately, our attacking multirotor would be able to gain full control over\nthe motions of the victim multirotor.\n", "date": "2023-05-22", "categories": []}
{"id": "2307.15157", "title": "R-LPIPS: An Adversarially Robust Perceptual Similarity Metric", "abstract": "  Similarity metrics have played a significant role in computer vision to\ncapture the underlying semantics of images. In recent years, advanced\nsimilarity metrics, such as the Learned Perceptual Image Patch Similarity\n(LPIPS), have emerged. These metrics leverage deep features extracted from\ntrained neural networks and have demonstrated a remarkable ability to closely\nalign with human perception when evaluating relative image similarity. However,\nit is now well-known that neural networks are susceptible to adversarial\nexamples, i.e., small perturbations invisible to humans crafted to deliberately\nmislead the model. Consequently, the LPIPS metric is also sensitive to such\nadversarial examples. This susceptibility introduces significant security\nconcerns, especially considering the widespread adoption of LPIPS in\nlarge-scale applications. In this paper, we propose the Robust Learned\nPerceptual Image Patch Similarity (R-LPIPS) metric, a new metric that leverages\nadversarially trained deep features. Through a comprehensive set of\nexperiments, we demonstrate the superiority of R-LPIPS compared to the\nclassical LPIPS metric. The code is available at\nhttps://github.com/SaraGhazanfari/R-LPIPS.\n", "date": "2023-07-27", "categories": []}
{"id": "2307.16099", "title": "On Neural Network approximation of ideal adversarial attack and\n  convergence of adversarial training", "abstract": "  Adversarial attacks are usually expressed in terms of a gradient-based\noperation on the input data and model, this results in heavy computations every\ntime an attack is generated. In this work, we solidify the idea of representing\nadversarial attacks as a trainable function, without further gradient\ncomputation. We first motivate that the theoretical best attacks, under proper\nconditions, can be represented as smooth piece-wise functions (piece-wise\nH\\\"older functions). Then we obtain an approximation result of such functions\nby a neural network. Subsequently, we emulate the ideal attack process by a\nneural network and reduce the adversarial training to a mathematical game\nbetween an attack network and a training model (a defense network). We also\nobtain convergence rates of adversarial loss in terms of the sample size $n$\nfor adversarial training in such a setting.\n", "date": "2023-07-30", "categories": []}
{"id": "2307.16331", "title": "Theoretically Principled Trade-off for Stateful Defenses against\n  Query-Based Black-Box Attacks", "abstract": "  Adversarial examples threaten the integrity of machine learning systems with\nalarming success rates even under constrained black-box conditions. Stateful\ndefenses have emerged as an effective countermeasure, detecting potential\nattacks by maintaining a buffer of recent queries and detecting new queries\nthat are too similar. However, these defenses fundamentally pose a trade-off\nbetween attack detection and false positive rates, and this trade-off is\ntypically optimized by hand-picking feature extractors and similarity\nthresholds that empirically work well. There is little current understanding as\nto the formal limits of this trade-off and the exact properties of the feature\nextractors/underlying problem domain that influence it. This work aims to\naddress this gap by offering a theoretical characterization of the trade-off\nbetween detection and false positive rates for stateful defenses. We provide\nupper bounds for detection rates of a general class of feature extractors and\nanalyze the impact of this trade-off on the convergence of black-box attacks.\nWe then support our theoretical findings with empirical evaluations across\nmultiple datasets and stateful defenses.\n", "date": "2023-07-30", "categories": []}
{"id": "2307.16816", "title": "Defense of Adversarial Ranking Attack in Text Retrieval: Benchmark and\n  Baseline via Detection", "abstract": "  Neural ranking models (NRMs) have undergone significant development and have\nbecome integral components of information retrieval (IR) systems.\nUnfortunately, recent research has unveiled the vulnerability of NRMs to\nadversarial document manipulations, potentially exploited by malicious search\nengine optimization practitioners. While progress in adversarial attack\nstrategies aids in identifying the potential weaknesses of NRMs before their\ndeployment, the defensive measures against such attacks, like the detection of\nadversarial documents, remain inadequately explored. To mitigate this gap, this\npaper establishes a benchmark dataset to facilitate the investigation of\nadversarial ranking defense and introduces two types of detection tasks for\nadversarial documents. A comprehensive investigation of the performance of\nseveral detection baselines is conducted, which involve examining the\nspamicity, perplexity, and linguistic acceptability, and utilizing supervised\nclassifiers. Experimental results demonstrate that a supervised classifier can\neffectively mitigate known attacks, but it performs poorly against unseen\nattacks. Furthermore, such classifier should avoid using query text to prevent\nlearning the classification on relevance, as it might lead to the inadvertent\ndiscarding of relevant documents.\n", "date": "2023-07-31", "categories": []}
{"id": "2304.10712", "title": "Adversarial Infrared Blocks: A Multi-view Black-box Attack to Thermal\n  Infrared Detectors in Physical World", "abstract": "  Infrared imaging systems have a vast array of potential applications in\npedestrian detection and autonomous driving, and their safety performance is of\ngreat concern. However, few studies have explored the safety of infrared\nimaging systems in real-world settings. Previous research has used physical\nperturbations such as small bulbs and thermal \"QR codes\" to attack infrared\nimaging detectors, but such methods are highly visible and lack stealthiness.\nOther researchers have used hot and cold blocks to deceive infrared imaging\ndetectors, but this method is limited in its ability to execute attacks from\nvarious angles. To address these shortcomings, we propose a novel physical\nattack called adversarial infrared blocks (AdvIB). By optimizing the physical\nparameters of the adversarial infrared blocks, this method can execute a\nstealthy black-box attack on thermal imaging system from various angles. We\nevaluate the proposed method based on its effectiveness, stealthiness, and\nrobustness. Our physical tests show that the proposed method achieves a success\nrate of over 80% under most distance and angle conditions, validating its\neffectiveness. For stealthiness, our method involves attaching the adversarial\ninfrared block to the inside of clothing, enhancing its stealthiness.\nAdditionally, we test the proposed method on advanced detectors, and\nexperimental results demonstrate an average attack success rate of 51.2%,\nproving its robustness. Overall, our proposed AdvIB method offers a promising\navenue for conducting stealthy, effective and robust black-box attacks on\nthermal imaging system, with potential implications for real-world safety and\nsecurity applications.\n", "date": "2023-04-21", "categories": []}
{"id": "2307.09763", "title": "Towards Building More Robust Models with Frequency Bias", "abstract": "  The vulnerability of deep neural networks to adversarial samples has been a\nmajor impediment to their broad applications, despite their success in various\nfields. Recently, some works suggested that adversarially-trained models\nemphasize the importance of low-frequency information to achieve higher\nrobustness. While several attempts have been made to leverage this frequency\ncharacteristic, they have all faced the issue that applying low-pass filters\ndirectly to input images leads to irreversible loss of discriminative\ninformation and poor generalizability to datasets with distinct frequency\nfeatures. This paper presents a plug-and-play module called the Frequency\nPreference Control Module that adaptively reconfigures the low- and\nhigh-frequency components of intermediate feature representations, providing\nbetter utilization of frequency in robust learning. Empirical studies show that\nour proposed module can be easily incorporated into any adversarial training\nframework, further improving model robustness across different architectures\nand datasets. Additionally, experiments were conducted to examine how the\nfrequency bias of robust models impacts the adversarial training process and\nits final robustness, revealing interesting insights.\n", "date": "2023-07-19", "categories": []}
{"id": "2307.13131", "title": "Why Don't You Clean Your Glasses? Perception Attacks with Dynamic\n  Optical Perturbations", "abstract": "  Camera-based autonomous systems that emulate human perception are\nincreasingly being integrated into safety-critical platforms. Consequently, an\nestablished body of literature has emerged that explores adversarial attacks\ntargeting the underlying machine learning models. Adapting adversarial attacks\nto the physical world is desirable for the attacker, as this removes the need\nto compromise digital systems. However, the real world poses challenges related\nto the \"survivability\" of adversarial manipulations given environmental noise\nin perception pipelines and the dynamicity of autonomous systems. In this\npaper, we take a sensor-first approach. We present EvilEye, a man-in-the-middle\nperception attack that leverages transparent displays to generate dynamic\nphysical adversarial examples. EvilEye exploits the camera's optics to induce\nmisclassifications under a variety of illumination conditions. To generate\ndynamic perturbations, we formalize the projection of a digital attack into the\nphysical domain by modeling the transformation function of the captured image\nthrough the optical pipeline. Our extensive experiments show that EvilEye's\ngenerated adversarial perturbations are much more robust across varying\nenvironmental light conditions relative to existing physical perturbation\nframeworks, achieving a high attack success rate (ASR) while bypassing\nstate-of-the-art physical adversarial detection frameworks. We demonstrate that\nthe dynamic nature of EvilEye enables attackers to adapt adversarial examples\nacross a variety of objects with a significantly higher ASR compared to\nstate-of-the-art physical world attack frameworks. Finally, we discuss\nmitigation strategies against the EvilEye attack.\n", "date": "2023-07-24", "categories": []}
{"id": "2307.15677", "title": "Adversarial training for tabular data with attack propagation", "abstract": "  Adversarial attacks are a major concern in security-centered applications,\nwhere malicious actors continuously try to mislead Machine Learning (ML) models\ninto wrongly classifying fraudulent activity as legitimate, whereas system\nmaintainers try to stop them. Adversarially training ML models that are robust\nagainst such attacks can prevent business losses and reduce the work load of\nsystem maintainers. In such applications data is often tabular and the space\navailable for attackers to manipulate undergoes complex feature engineering\ntransformations, to provide useful signals for model training, to a space\nattackers cannot access. Thus, we propose a new form of adversarial training\nwhere attacks are propagated between the two spaces in the training loop. We\nthen test this method empirically on a real world dataset in the domain of\ncredit card fraud detection. We show that our method can prevent about 30%\nperformance drops under moderate attacks and is essential under very aggressive\nattacks, with a trade-off loss in performance under no attacks smaller than 7%.\n", "date": "2023-07-28", "categories": []}
{"id": "2305.06716", "title": "Distracting Downpour: Adversarial Weather Attacks for Motion Estimation", "abstract": "  Current adversarial attacks on motion estimation, or optical flow, optimize\nsmall per-pixel perturbations, which are unlikely to appear in the real world.\nIn contrast, adverse weather conditions constitute a much more realistic threat\nscenario. Hence, in this work, we present a novel attack on motion estimation\nthat exploits adversarially optimized particles to mimic weather effects like\nsnowflakes, rain streaks or fog clouds. At the core of our attack framework is\na differentiable particle rendering system that integrates particles (i)\nconsistently over multiple time steps (ii) into the 3D space (iii) with a\nphoto-realistic appearance. Through optimization, we obtain adversarial weather\nthat significantly impacts the motion estimation. Surprisingly, methods that\npreviously showed good robustness towards small per-pixel perturbations are\nparticularly vulnerable to adversarial weather. At the same time, augmenting\nthe training with non-optimized weather increases a method's robustness towards\nweather effects and improves generalizability at almost no additional cost. Our\ncode will be available at https://github.com/cv-stuttgart/DistractingDownpour.\n", "date": "2023-05-11", "categories": []}
{"id": "2307.14682", "title": "Unified Adversarial Patch for Visible-Infrared Cross-modal Attacks in\n  the Physical World", "abstract": "  Physical adversarial attacks have put a severe threat to DNN-based object\ndetectors. To enhance security, a combination of visible and infrared sensors\nis deployed in various scenarios, which has proven effective in disabling\nexisting single-modal physical attacks. To further demonstrate the potential\nrisks in such cases, we design a unified adversarial patch that can perform\ncross-modal physical attacks, achieving evasion in both modalities\nsimultaneously with a single patch. Given the different imaging mechanisms of\nvisible and infrared sensors, our work manipulates patches' shape features,\nwhich can be captured in different modalities when they undergo changes. To\ndeal with challenges, we propose a novel boundary-limited shape optimization\napproach that aims to achieve compact and smooth shapes for the adversarial\npatch, making it easy to implement in the physical world. And a score-aware\niterative evaluation method is also introduced to balance the fooling degree\nbetween visible and infrared detectors during optimization, which guides the\nadversarial patch to iteratively reduce the predicted scores of the multi-modal\nsensors. Furthermore, we propose an Affine-Transformation-based enhancement\nstrategy that makes the learnable shape robust to various angles, thus\nmitigating the issue of shape deformation caused by different shooting angles\nin the real world. Our method is evaluated against several state-of-the-art\nobject detectors, achieving an Attack Success Rate (ASR) of over 80%. We also\ndemonstrate the effectiveness of our approach in physical-world scenarios under\nvarious settings, including different angles, distances, postures, and scenes\nfor both visible and infrared sensors.\n", "date": "2023-07-27", "categories": []}
{"id": "2307.15008", "title": "A LLM Assisted Exploitation of AI-Guardian", "abstract": "  Large language models (LLMs) are now highly capable at a diverse range of\ntasks. This paper studies whether or not GPT-4, one such LLM, is capable of\nassisting researchers in the field of adversarial machine learning. As a case\nstudy, we evaluate the robustness of AI-Guardian, a recent defense to\nadversarial examples published at IEEE S&P 2023, a top computer security\nconference. We completely break this defense: the proposed scheme does not\nincrease robustness compared to an undefended baseline.\n  We write none of the code to attack this model, and instead prompt GPT-4 to\nimplement all attack algorithms following our instructions and guidance. This\nprocess was surprisingly effective and efficient, with the language model at\ntimes producing code from ambiguous instructions faster than the author of this\npaper could have done. We conclude by discussing (1) the warning signs present\nin the evaluation that suggested to us AI-Guardian would be broken, and (2) our\nexperience with designing attacks and performing novel research using the most\nrecent advances in language modeling.\n", "date": "2023-07-20", "categories": []}
{"id": "2307.13787", "title": "The GANfather: Controllable generation of malicious activity to improve\n  defence systems", "abstract": "  Machine learning methods to aid defence systems in detecting malicious\nactivity typically rely on labelled data. In some domains, such labelled data\nis unavailable or incomplete. In practice this can lead to low detection rates\nand high false positive rates, which characterise for example anti-money\nlaundering systems. In fact, it is estimated that 1.7--4 trillion euros are\nlaundered annually and go undetected. We propose The GANfather, a method to\ngenerate samples with properties of malicious activity, without label\nrequirements. We propose to reward the generation of malicious samples by\nintroducing an extra objective to the typical Generative Adversarial Networks\n(GANs) loss. Ultimately, our goal is to enhance the detection of illicit\nactivity using the discriminator network as a novel and robust defence system.\nOptionally, we may encourage the generator to bypass pre-existing detection\nsystems. This setup then reveals defensive weaknesses for the discriminator to\ncorrect. We evaluate our method in two real-world use cases, money laundering\nand recommendation systems. In the former, our method moves cumulative amounts\nclose to 350 thousand dollars through a network of accounts without being\ndetected by an existing system. In the latter, we recommend the target item to\na broad user base with as few as 30 synthetic attackers. In both cases, we\ntrain a new defence system to capture the synthetic attacks.\n", "date": "2023-07-25", "categories": []}
{"id": "2307.13856", "title": "On the unreasonable vulnerability of transformers for image restoration\n  -- and an easy fix", "abstract": "  Following their success in visual recognition tasks, Vision\nTransformers(ViTs) are being increasingly employed for image restoration. As a\nfew recent works claim that ViTs for image classification also have better\nrobustness properties, we investigate whether the improved adversarial\nrobustness of ViTs extends to image restoration. We consider the recently\nproposed Restormer model, as well as NAFNet and the \"Baseline network\" which\nare both simplified versions of a Restormer. We use Projected Gradient Descent\n(PGD) and CosPGD, a recently proposed adversarial attack tailored to pixel-wise\nprediction tasks for our robustness evaluation. Our experiments are performed\non real-world images from the GoPro dataset for image deblurring. Our analysis\nindicates that contrary to as advocated by ViTs in image classification works,\nthese models are highly susceptible to adversarial attacks. We attempt to\nimprove their robustness through adversarial training. While this yields a\nsignificant increase in robustness for Restormer, results on other networks are\nless promising. Interestingly, the design choices in NAFNet and Baselines,\nwhich were based on iid performance, and not on robust generalization, seem to\nbe at odds with the model robustness. Thus, we investigate this further and\nfind a fix.\n", "date": "2023-07-25", "categories": []}
{"id": "2307.13985", "title": "Enhanced Security against Adversarial Examples Using a Random Ensemble\n  of Encrypted Vision Transformer Models", "abstract": "  Deep neural networks (DNNs) are well known to be vulnerable to adversarial\nexamples (AEs). In addition, AEs have adversarial transferability, which means\nAEs generated for a source model can fool another black-box model (target\nmodel) with a non-trivial probability. In previous studies, it was confirmed\nthat the vision transformer (ViT) is more robust against the property of\nadversarial transferability than convolutional neural network (CNN) models such\nas ConvMixer, and moreover encrypted ViT is more robust than ViT without any\nencryption. In this article, we propose a random ensemble of encrypted ViT\nmodels to achieve much more robust models. In experiments, the proposed scheme\nis verified to be more robust against not only black-box attacks but also\nwhite-box ones than convention methods.\n", "date": "2023-07-26", "categories": []}
{"id": "2307.14061", "title": "Set-level Guidance Attack: Boosting Adversarial Transferability of\n  Vision-Language Pre-training Models", "abstract": "  Vision-language pre-training (VLP) models have shown vulnerability to\nadversarial examples in multimodal tasks. Furthermore, malicious adversaries\ncan be deliberately transferred to attack other black-box models. However,\nexisting work has mainly focused on investigating white-box attacks. In this\npaper, we present the first study to investigate the adversarial\ntransferability of recent VLP models. We observe that existing methods exhibit\nmuch lower transferability, compared to the strong attack performance in\nwhite-box settings. The transferability degradation is partly caused by the\nunder-utilization of cross-modal interactions. Particularly, unlike unimodal\nlearning, VLP models rely heavily on cross-modal interactions and the\nmultimodal alignments are many-to-many, e.g., an image can be described in\nvarious natural languages. To this end, we propose a highly transferable\nSet-level Guidance Attack (SGA) that thoroughly leverages modality interactions\nand incorporates alignment-preserving augmentation with cross-modal guidance.\nExperimental results demonstrate that SGA could generate adversarial examples\nthat can strongly transfer across different VLP models on multiple downstream\nvision-language tasks. On image-text retrieval, SGA significantly enhances the\nattack success rate for transfer attacks from ALBEF to TCL by a large margin\n(at least 9.78% and up to 30.21%), compared to the state-of-the-art.\n", "date": "2023-07-26", "categories": []}
{"id": "2307.14242", "title": "Defending Adversarial Patches via Joint Region Localizing and Inpainting", "abstract": "  Deep neural networks are successfully used in various applications, but show\ntheir vulnerability to adversarial examples. With the development of\nadversarial patches, the feasibility of attacks in physical scenes increases,\nand the defenses against patch attacks are urgently needed. However, defending\nsuch adversarial patch attacks is still an unsolved problem. In this paper, we\nanalyse the properties of adversarial patches, and find that: on the one hand,\nadversarial patches will lead to the appearance or contextual inconsistency in\nthe target objects; on the other hand, the patch region will show abnormal\nchanges on the high-level feature maps of the objects extracted by a backbone\nnetwork. Considering the above two points, we propose a novel defense method\nbased on a ``localizing and inpainting\" mechanism to pre-process the input\nexamples. Specifically, we design an unified framework, where the ``localizing\"\nsub-network utilizes a two-branch structure to represent the above two aspects\nto accurately detect the adversarial patch region in the image. For the\n``inpainting\" sub-network, it utilizes the surrounding contextual cues to\nrecover the original content covered by the adversarial patch. The quality of\ninpainted images is also evaluated by measuring the appearance consistency and\nthe effects of adversarial attacks. These two sub-networks are then jointly\ntrained via an iterative optimization manner. In this way, the ``localizing\"\nand ``inpainting\" modules can interact closely with each other, and thus learn\na better solution. A series of experiments versus traffic sign classification\nand detection tasks are conducted to defend against various adversarial patch\nattacks.\n", "date": "2023-07-26", "categories": []}
{"id": "2307.11906", "title": "Unveiling Vulnerabilities in Interpretable Deep Learning Systems with\n  Query-Efficient Black-box Attacks", "abstract": "  Deep learning has been rapidly employed in many applications revolutionizing\nmany industries, but it is known to be vulnerable to adversarial attacks. Such\nattacks pose a serious threat to deep learning-based systems compromising their\nintegrity, reliability, and trust. Interpretable Deep Learning Systems (IDLSes)\nare designed to make the system more transparent and explainable, but they are\nalso shown to be susceptible to attacks. In this work, we propose a novel\nmicrobial genetic algorithm-based black-box attack against IDLSes that requires\nno prior knowledge of the target model and its interpretation model. The\nproposed attack is a query-efficient approach that combines transfer-based and\nscore-based methods, making it a powerful tool to unveil IDLS vulnerabilities.\nOur experiments of the attack show high attack success rates using adversarial\nexamples with attribution maps that are highly similar to those of benign\nsamples which makes it difficult to detect even by human analysts. Our results\nhighlight the need for improved IDLS security to ensure their practical\nreliability.\n", "date": "2023-07-21", "categories": []}
{"id": "2307.12342", "title": "Towards Generic and Controllable Attacks Against Object Detection", "abstract": "  Existing adversarial attacks against Object Detectors (ODs) suffer from two\ninherent limitations. Firstly, ODs have complicated meta-structure designs,\nhence most advanced attacks for ODs concentrate on attacking specific\ndetector-intrinsic structures, which makes it hard for them to work on other\ndetectors and motivates us to design a generic attack against ODs. Secondly,\nmost works against ODs make Adversarial Examples (AEs) by generalizing\nimage-level attacks from classification to detection, which brings redundant\ncomputations and perturbations in semantically meaningless areas (e.g.,\nbackgrounds) and leads to an emergency for seeking controllable attacks for\nODs. To this end, we propose a generic white-box attack, LGP (local\nperturbations with adaptively global attacks), to blind mainstream object\ndetectors with controllable perturbations. For a detector-agnostic attack, LGP\ntracks high-quality proposals and optimizes three heterogeneous losses\nsimultaneously. In this way, we can fool the crucial components of ODs with a\npart of their outputs without the limitations of specific structures. Regarding\ncontrollability, we establish an object-wise constraint that exploits\nforeground-background separation adaptively to induce the attachment of\nperturbations to foregrounds. Experimentally, the proposed LGP successfully\nattacked sixteen state-of-the-art object detectors on MS-COCO and DOTA\ndatasets, with promising imperceptibility and transferability obtained. Codes\nare publicly released in https://github.com/liguopeng0923/LGP.git\n", "date": "2023-07-23", "categories": []}
{"id": "2307.12520", "title": "Lost In Translation: Generating Adversarial Examples Robust to\n  Round-Trip Translation", "abstract": "  Language Models today provide a high accuracy across a large number of\ndownstream tasks. However, they remain susceptible to adversarial attacks,\nparticularly against those where the adversarial examples maintain considerable\nsimilarity to the original text. Given the multilingual nature of text, the\neffectiveness of adversarial examples across translations and how machine\ntranslations can improve the robustness of adversarial examples remain largely\nunexplored. In this paper, we present a comprehensive study on the robustness\nof current text adversarial attacks to round-trip translation. We demonstrate\nthat 6 state-of-the-art text-based adversarial attacks do not maintain their\nefficacy after round-trip translation. Furthermore, we introduce an\nintervention-based solution to this problem, by integrating Machine Translation\ninto the process of adversarial example generation and demonstrating increased\nrobustness to round-trip translation. Our results indicate that finding\nadversarial examples robust to translation can help identify the insufficiency\nof language models that is common across languages, and motivate further\nresearch into multilingual adversarial attacks.\n", "date": "2023-07-24", "categories": []}
{"id": "2210.03297", "title": "Preprocessors Matter! Realistic Decision-Based Attacks on Machine\n  Learning Systems", "abstract": "  Decision-based attacks construct adversarial examples against a machine\nlearning (ML) model by making only hard-label queries. These attacks have\nmainly been applied directly to standalone neural networks. However, in\npractice, ML models are just one component of a larger learning system. We find\nthat by adding a single preprocessor in front of a classifier, state-of-the-art\nquery-based attacks are up to 7$\\times$ less effective at attacking a\nprediction pipeline than at attacking the model alone. We explain this\ndiscrepancy by the fact that most preprocessors introduce some notion of\ninvariance to the input space. Hence, attacks that are unaware of this\ninvariance inevitably waste a large number of queries to re-discover or\novercome it. We, therefore, develop techniques to (i) reverse-engineer the\npreprocessor and then (ii) use this extracted information to attack the\nend-to-end system. Our preprocessors extraction method requires only a few\nhundred queries, and our preprocessor-aware attacks recover the same efficacy\nas when attacking the model alone. The code can be found at\nhttps://github.com/google-research/preprocessor-aware-black-box-attack.\n", "date": "2022-10-07", "categories": []}
{"id": "2307.11528", "title": "Improving Viewpoint Robustness for Visual Recognition via Adversarial\n  Training", "abstract": "  Viewpoint invariance remains challenging for visual recognition in the 3D\nworld, as altering the viewing directions can significantly impact predictions\nfor the same object. While substantial efforts have been dedicated to making\nneural networks invariant to 2D image translations and rotations, viewpoint\ninvariance is rarely investigated. Motivated by the success of adversarial\ntraining in enhancing model robustness, we propose Viewpoint-Invariant\nAdversarial Training (VIAT) to improve the viewpoint robustness of image\nclassifiers. Regarding viewpoint transformation as an attack, we formulate VIAT\nas a minimax optimization problem, where the inner maximization characterizes\ndiverse adversarial viewpoints by learning a Gaussian mixture distribution\nbased on the proposed attack method GMVFool. The outer minimization obtains a\nviewpoint-invariant classifier by minimizing the expected loss over the\nworst-case viewpoint distributions that can share the same one for different\nobjects within the same category. Based on GMVFool, we contribute a large-scale\ndataset called ImageNet-V+ to benchmark viewpoint robustness. Experimental\nresults show that VIAT significantly improves the viewpoint robustness of\nvarious image classifiers based on the diversity of adversarial viewpoints\ngenerated by GMVFool. Furthermore, we propose ViewRS, a certified viewpoint\nrobustness method that provides a certified radius and accuracy to demonstrate\nthe effectiveness of VIAT from the theoretical perspective.\n", "date": "2023-07-21", "categories": []}
{"id": "2208.10224", "title": "Friendly Noise against Adversarial Noise: A Powerful Defense against\n  Data Poisoning Attacks", "abstract": "  A powerful category of (invisible) data poisoning attacks modify a subset of\ntraining examples by small adversarial perturbations to change the prediction\nof certain test-time data. Existing defense mechanisms are not desirable to\ndeploy in practice, as they often either drastically harm the generalization\nperformance, or are attack-specific, and prohibitively slow to apply. Here, we\npropose a simple but highly effective approach that unlike existing methods\nbreaks various types of invisible poisoning attacks with the slightest drop in\nthe generalization performance. We make the key observation that attacks\nintroduce local sharp regions of high training loss, which when minimized,\nresults in learning the adversarial perturbations and makes the attack\nsuccessful. To break poisoning attacks, our key idea is to alleviate the sharp\nloss regions introduced by poisons. To do so, our approach comprises two\ncomponents: an optimized friendly noise that is generated to maximally perturb\nexamples without degrading the performance, and a randomly varying noise\ncomponent. The combination of both components builds a very light-weight but\nextremely effective defense against the most powerful triggerless targeted and\nhidden-trigger backdoor poisoning attacks, including Gradient Matching,\nBulls-eye Polytope, and Sleeper Agent. We show that our friendly noise is\ntransferable to other architectures, and adaptive attacks cannot break our\ndefense due to its random noise component. Our code is available at:\nhttps://github.com/tianyu139/friendly-noise\n", "date": "2022-08-14", "categories": []}
{"id": "2302.10980", "title": "MultiRobustBench: Benchmarking Robustness Against Multiple Attacks", "abstract": "  The bulk of existing research in defending against adversarial examples\nfocuses on defending against a single (typically bounded Lp-norm) attack, but\nfor a practical setting, machine learning (ML) models should be robust to a\nwide variety of attacks. In this paper, we present the first unified framework\nfor considering multiple attacks against ML models. Our framework is able to\nmodel different levels of learner's knowledge about the test-time adversary,\nallowing us to model robustness against unforeseen attacks and robustness\nagainst unions of attacks. Using our framework, we present the first\nleaderboard, MultiRobustBench, for benchmarking multiattack evaluation which\ncaptures performance across attack types and attack strengths. We evaluate the\nperformance of 16 defended models for robustness against a set of 9 different\nattack types, including Lp-based threat models, spatial transformations, and\ncolor changes, at 20 different attack strengths (180 attacks total).\nAdditionally, we analyze the state of current defenses against multiple\nattacks. Our analysis shows that while existing defenses have made progress in\nterms of average robustness across the set of attacks used, robustness against\nthe worst-case attack is still a big open problem as all existing models\nperform worse than random guessing.\n", "date": "2023-02-21", "categories": []}
{"id": "2307.10184", "title": "A Dual Stealthy Backdoor: From Both Spatial and Frequency Perspectives", "abstract": "  Backdoor attacks pose serious security threats to deep neural networks\n(DNNs). Backdoored models make arbitrarily (targeted) incorrect predictions on\ninputs embedded with well-designed triggers while behaving normally on clean\ninputs. Many works have explored the invisibility of backdoor triggers to\nimprove attack stealthiness. However, most of them only consider the\ninvisibility in the spatial domain without explicitly accounting for the\ngeneration of invisible triggers in the frequency domain, making the generated\npoisoned images be easily detected by recent defense methods. To address this\nissue, in this paper, we propose a DUal stealthy BAckdoor attack method named\nDUBA, which simultaneously considers the invisibility of triggers in both the\nspatial and frequency domains, to achieve desirable attack performance, while\nensuring strong stealthiness. Specifically, we first use Discrete Wavelet\nTransform to embed the high-frequency information of the trigger image into the\nclean image to ensure attack effectiveness. Then, to attain strong\nstealthiness, we incorporate Fourier Transform and Discrete Cosine Transform to\nmix the poisoned image and clean image in the frequency domain. Moreover, the\nproposed DUBA adopts a novel attack strategy, in which the model is trained\nwith weak triggers and attacked with strong triggers to further enhance the\nattack performance and stealthiness. We extensively evaluate DUBA against\npopular image classifiers on four datasets. The results demonstrate that it\nsignificantly outperforms the state-of-the-art backdoor attacks in terms of the\nattack success rate and stealthiness\n", "date": "2023-07-03", "categories": []}
{"id": "2307.10209", "title": "On the Sensitivity of Deep Load Disaggregation to Adversarial Attacks", "abstract": "  Non-intrusive Load Monitoring (NILM) algorithms, commonly referred to as load\ndisaggregation algorithms, are fundamental tools for effective energy\nmanagement. Despite the success of deep models in load disaggregation, they\nface various challenges, particularly those pertaining to privacy and security.\nThis paper investigates the sensitivity of prominent deep NILM baselines to\nadversarial attacks, which have proven to be a significant threat in domains\nsuch as computer vision and speech recognition. Adversarial attacks entail the\nintroduction of imperceptible noise into the input data with the aim of\nmisleading the neural network into generating erroneous outputs. We investigate\nthe Fast Gradient Sign Method (FGSM), a well-known adversarial attack, to\nperturb the input sequences fed into two commonly employed CNN-based NILM\nbaselines: the Sequence-to-Sequence (S2S) and Sequence-to-Point (S2P) models.\nOur findings provide compelling evidence for the vulnerability of these models,\nparticularly the S2P model which exhibits an average decline of 20\\% in the\nF1-score even with small amounts of noise. Such weakness has the potential to\ngenerate profound implications for energy management systems in residential and\nindustrial sectors reliant on NILM models.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.10235", "title": "Towards Viewpoint-Invariant Visual Recognition via Adversarial Training", "abstract": "  Visual recognition models are not invariant to viewpoint changes in the 3D\nworld, as different viewing directions can dramatically affect the predictions\ngiven the same object. Although many efforts have been devoted to making neural\nnetworks invariant to 2D image translations and rotations, viewpoint invariance\nis rarely investigated. As most models process images in the perspective view,\nit is challenging to impose invariance to 3D viewpoint changes based only on 2D\ninputs. Motivated by the success of adversarial training in promoting model\nrobustness, we propose Viewpoint-Invariant Adversarial Training (VIAT) to\nimprove viewpoint robustness of common image classifiers. By regarding\nviewpoint transformation as an attack, VIAT is formulated as a minimax\noptimization problem, where the inner maximization characterizes diverse\nadversarial viewpoints by learning a Gaussian mixture distribution based on a\nnew attack GMVFool, while the outer minimization trains a viewpoint-invariant\nclassifier by minimizing the expected loss over the worst-case adversarial\nviewpoint distributions. To further improve the generalization performance, a\ndistribution sharing strategy is introduced leveraging the transferability of\nadversarial viewpoints across objects. Experiments validate the effectiveness\nof VIAT in improving the viewpoint robustness of various image classifiers\nbased on the diversity of adversarial viewpoints generated by GMVFool.\n", "date": "2023-07-16", "categories": []}
{"id": "2307.10563", "title": "FACADE: A Framework for Adversarial Circuit Anomaly Detection and\n  Evaluation", "abstract": "  We present FACADE, a novel probabilistic and geometric framework designed for\nunsupervised mechanistic anomaly detection in deep neural networks. Its primary\ngoal is advancing the understanding and mitigation of adversarial attacks.\nFACADE aims to generate probabilistic distributions over circuits, which\nprovide critical insights to their contribution to changes in the manifold\nproperties of pseudo-classes, or high-dimensional modes in activation space,\nyielding a powerful tool for uncovering and combating adversarial attacks. Our\napproach seeks to improve model robustness, enhance scalable model oversight,\nand demonstrates promising applications in real-world deployment settings.\n", "date": "2023-07-20", "categories": []}
{"id": "2307.10788", "title": "Adversarial attacks for mixtures of classifiers", "abstract": "  Mixtures of classifiers (a.k.a. randomized ensembles) have been proposed as a\nway to improve robustness against adversarial attacks. However, it has been\nshown that existing attacks are not well suited for this kind of classifiers.\nIn this paper, we discuss the problem of attacking a mixture in a principled\nway and introduce two desirable properties of attacks based on a geometrical\nanalysis of the problem (effectiveness and maximality). We then show that\nexisting attacks do not meet both of these properties. Finally, we introduce a\nnew attack called lattice climber attack with theoretical guarantees on the\nbinary linear setting, and we demonstrate its performance by conducting\nexperiments on synthetic and real datasets.\n", "date": "2023-07-20", "categories": []}
{"id": "2210.16117", "title": "Improving the Transferability of Adversarial Attacks on Face Recognition\n  with Beneficial Perturbation Feature Augmentation", "abstract": "  Face recognition (FR) models can be easily fooled by adversarial examples,\nwhich are crafted by adding imperceptible perturbations on benign face images.\nThe existence of adversarial face examples poses a great threat to the security\nof society. In order to build a more sustainable digital nation, in this paper,\nwe improve the transferability of adversarial face examples to expose more\nblind spots of existing FR models. Though generating hard samples has shown its\neffectiveness in improving the generalization of models in training tasks, the\neffectiveness of utilizing this idea to improve the transferability of\nadversarial face examples remains unexplored. To this end, based on the\nproperty of hard samples and the symmetry between training tasks and\nadversarial attack tasks, we propose the concept of hard models, which have\nsimilar effects as hard samples for adversarial attack tasks. Utilizing the\nconcept of hard models, we propose a novel attack method called Beneficial\nPerturbation Feature Augmentation Attack (BPFA), which reduces the overfitting\nof adversarial examples to surrogate FR models by constantly generating new\nhard models to craft the adversarial examples. Specifically, in the\nbackpropagation, BPFA records the gradients on pre-selected feature maps and\nuses the gradient on the input image to craft the adversarial example. In the\nnext forward propagation, BPFA leverages the recorded gradients to add\nbeneficial perturbations on their corresponding feature maps to increase the\nloss. Extensive experiments demonstrate that BPFA can significantly boost the\ntransferability of adversarial attacks on FR.\n", "date": "2022-10-28", "categories": []}
{"id": "2302.05086", "title": "Making Substitute Models More Bayesian Can Enhance Transferability of\n  Adversarial Examples", "abstract": "  The transferability of adversarial examples across deep neural networks\n(DNNs) is the crux of many black-box attacks. Many prior efforts have been\ndevoted to improving the transferability via increasing the diversity in inputs\nof some substitute models. In this paper, by contrast, we opt for the diversity\nin substitute models and advocate to attack a Bayesian model for achieving\ndesirable transferability. Deriving from the Bayesian formulation, we develop a\nprincipled strategy for possible finetuning, which can be combined with many\noff-the-shelf Gaussian posterior approximations over DNN parameters. Extensive\nexperiments have been conducted to verify the effectiveness of our method, on\ncommon benchmark datasets, and the results demonstrate that our method\noutperforms recent state-of-the-arts by large margins (roughly 19% absolute\nincrease in average attack success rate on ImageNet), and, by combining with\nthese recent methods, further performance gain can be obtained. Our code:\nhttps://github.com/qizhangli/MoreBayesian-attack.\n", "date": "2023-02-10", "categories": []}
{"id": "2306.07591", "title": "I See Dead People: Gray-Box Adversarial Attack on Image-To-Text Models", "abstract": "  Modern image-to-text systems typically adopt the encoder-decoder framework,\nwhich comprises two main components: an image encoder, responsible for\nextracting image features, and a transformer-based decoder, used for generating\ncaptions. Taking inspiration from the analysis of neural networks' robustness\nagainst adversarial perturbations, we propose a novel gray-box algorithm for\ncreating adversarial examples in image-to-text models. Unlike image\nclassification tasks that have a finite set of class labels, finding visually\nsimilar adversarial examples in an image-to-text task poses greater challenges\nbecause the captioning system allows for a virtually infinite space of possible\ncaptions. In this paper, we present a gray-box adversarial attack on\nimage-to-text, both untargeted and targeted. We formulate the process of\ndiscovering adversarial perturbations as an optimization problem that uses only\nthe image-encoder component, meaning the proposed attack is language-model\nagnostic. Through experiments conducted on the ViT-GPT2 model, which is the\nmost-used image-to-text model in Hugging Face, and the Flickr30k dataset, we\ndemonstrate that our proposed attack successfully generates visually similar\nadversarial examples, both with untargeted and targeted captions. Notably, our\nattack operates in a gray-box manner, requiring no knowledge about the decoder\nmodule. We also show that our attacks fool the popular open-source platform\nHugging Face.\n", "date": "2023-06-13", "categories": []}
{"id": "2307.07859", "title": "Unified Adversarial Patch for Cross-modal Attacks in the Physical World", "abstract": "  Recently, physical adversarial attacks have been presented to evade\nDNNs-based object detectors. To ensure the security, many scenarios are\nsimultaneously deployed with visible sensors and infrared sensors, leading to\nthe failures of these single-modal physical attacks. To show the potential\nrisks under such scenes, we propose a unified adversarial patch to perform\ncross-modal physical attacks, i.e., fooling visible and infrared object\ndetectors at the same time via a single patch. Considering different imaging\nmechanisms of visible and infrared sensors, our work focuses on modeling the\nshapes of adversarial patches, which can be captured in different modalities\nwhen they change. To this end, we design a novel boundary-limited shape\noptimization to achieve the compact and smooth shapes, and thus they can be\neasily implemented in the physical world. In addition, to balance the fooling\ndegree between visible detector and infrared detector during the optimization\nprocess, we propose a score-aware iterative evaluation, which can guide the\nadversarial patch to iteratively reduce the predicted scores of the multi-modal\nsensors. We finally test our method against the one-stage detector: YOLOv3 and\nthe two-stage detector: Faster RCNN. Results show that our unified patch\nachieves an Attack Success Rate (ASR) of 73.33% and 69.17%, respectively. More\nimportantly, we verify the effective attacks in the physical world when visible\nand infrared sensors shoot the objects under various settings like different\nangles, distances, postures, and scenes.\n", "date": "2023-07-15", "categories": []}
{"id": "2307.09649", "title": "Application of BadNets in Spam Filters", "abstract": "  Spam filters are a crucial component of modern email systems, as they help to\nprotect users from unwanted and potentially harmful emails. However, the\neffectiveness of these filters is dependent on the quality of the machine\nlearning models that power them. In this paper, we design backdoor attacks in\nthe domain of spam filtering. By demonstrating the potential vulnerabilities in\nthe machine learning model supply chain, we highlight the need for careful\nconsideration and evaluation of the models used in spam filters. Our results\nshow that the backdoor attacks can be effectively used to identify\nvulnerabilities in spam filters and suggest the need for ongoing monitoring and\nimprovement in this area.\n", "date": "2023-07-18", "categories": []}
{"id": "2307.10163", "title": "Rethinking Backdoor Attacks", "abstract": "  In a backdoor attack, an adversary inserts maliciously constructed backdoor\nexamples into a training set to make the resulting model vulnerable to\nmanipulation. Defending against such attacks typically involves viewing these\ninserted examples as outliers in the training set and using techniques from\nrobust statistics to detect and remove them.\n  In this work, we present a different approach to the backdoor attack problem.\nSpecifically, we show that without structural information about the training\ndata distribution, backdoor attacks are indistinguishable from\nnaturally-occurring features in the data--and thus impossible to \"detect\" in a\ngeneral sense. Then, guided by this observation, we revisit existing defenses\nagainst backdoor attacks and characterize the (often latent) assumptions they\nmake and on which they depend. Finally, we explore an alternative perspective\non backdoor attacks: one that assumes these attacks correspond to the strongest\nfeature in the training data. Under this assumption (which we make formal) we\ndevelop a new primitive for detecting backdoor attacks. Our primitive naturally\ngives rise to a detection algorithm that comes with theoretical guarantees and\nis effective in practice.\n", "date": "2023-07-19", "categories": []}
{"id": "2302.06060", "title": "Threatening Patch Attacks on Object Detection in Optical Remote Sensing\n  Images", "abstract": "  Advanced Patch Attacks (PAs) on object detection in natural images have\npointed out the great safety vulnerability in methods based on deep neural\nnetworks. However, little attention has been paid to this topic in Optical\nRemote Sensing Images (O-RSIs). To this end, we focus on this research, i.e.,\nPAs on object detection in O-RSIs, and propose a more Threatening PA without\nthe scarification of the visual quality, dubbed TPA. Specifically, to address\nthe problem of inconsistency between local and global landscapes in existing\npatch selection schemes, we propose leveraging the First-Order Difference (FOD)\nof the objective function before and after masking to select the sub-patches to\nbe attacked. Further, considering the problem of gradient inundation when\napplying existing coordinate-based loss to PAs directly, we design an IoU-based\nobjective function specific for PAs, dubbed Bounding box Drifting Loss (BDL),\nwhich pushes the detected bounding boxes far from the initial ones until there\nare no intersections between them. Finally, on two widely used benchmarks,\ni.e., DIOR and DOTA, comprehensive evaluations of our TPA with four typical\ndetectors (Faster R-CNN, FCOS, RetinaNet, and YOLO-v4) witness its remarkable\neffectiveness. To the best of our knowledge, this is the first attempt to study\nthe PAs on object detection in O-RSIs, and we hope this work can get our\nreaders interested in studying this topic.\n", "date": "2023-02-13", "categories": []}
{"id": "2307.07250", "title": "Mitigating Adversarial Vulnerability through Causal Parameter Estimation\n  by Adversarial Double Machine Learning", "abstract": "  Adversarial examples derived from deliberately crafted perturbations on\nvisual inputs can easily harm decision process of deep neural networks. To\nprevent potential threats, various adversarial training-based defense methods\nhave grown rapidly and become a de facto standard approach for robustness.\nDespite recent competitive achievements, we observe that adversarial\nvulnerability varies across targets and certain vulnerabilities remain\nprevalent. Intriguingly, such peculiar phenomenon cannot be relieved even with\ndeeper architectures and advanced defense methods. To address this issue, in\nthis paper, we introduce a causal approach called Adversarial Double Machine\nLearning (ADML), which allows us to quantify the degree of adversarial\nvulnerability for network predictions and capture the effect of treatments on\noutcome of interests. ADML can directly estimate causal parameter of\nadversarial perturbations per se and mitigate negative effects that can\npotentially damage robustness, bridging a causal perspective into the\nadversarial vulnerability. Through extensive experiments on various CNN and\nTransformer architectures, we corroborate that ADML improves adversarial\nrobustness with large margins and relieve the empirical observation.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.07916", "title": "On the Robustness of Split Learning against Adversarial Attacks", "abstract": "  Split learning enables collaborative deep learning model training while\npreserving data privacy and model security by avoiding direct sharing of raw\ndata and model details (i.e., sever and clients only hold partial sub-networks\nand exchange intermediate computations). However, existing research has mainly\nfocused on examining its reliability for privacy protection, with little\ninvestigation into model security. Specifically, by exploring full models,\nattackers can launch adversarial attacks, and split learning can mitigate this\nsevere threat by only disclosing part of models to untrusted servers.This paper\naims to evaluate the robustness of split learning against adversarial attacks,\nparticularly in the most challenging setting where untrusted servers only have\naccess to the intermediate layers of the model.Existing adversarial attacks\nmostly focus on the centralized setting instead of the collaborative setting,\nthus, to better evaluate the robustness of split learning, we develop a\ntailored attack called SPADV, which comprises two stages: 1) shadow model\ntraining that addresses the issue of lacking part of the model and 2) local\nadversarial attack that produces adversarial examples to evaluate.The first\nstage only requires a few unlabeled non-IID data, and, in the second stage,\nSPADV perturbs the intermediate output of natural samples to craft the\nadversarial ones. The overall cost of the proposed attack process is relatively\nlow, yet the empirical attack effectiveness is significantly high,\ndemonstrating the surprising vulnerability of split learning to adversarial\nattacks.\n", "date": "2023-07-16", "categories": []}
{"id": "2307.08955", "title": "Discretization-based ensemble model for robust learning in IoT", "abstract": "  IoT device identification is the process of recognizing and verifying\nconnected IoT devices to the network. This is an essential process for ensuring\nthat only authorized devices can access the network, and it is necessary for\nnetwork management and maintenance. In recent years, machine learning models\nhave been used widely for automating the process of identifying devices in the\nnetwork. However, these models are vulnerable to adversarial attacks that can\ncompromise their accuracy and effectiveness. To better secure device\nidentification models, discretization techniques enable reduction in the\nsensitivity of machine learning models to adversarial attacks contributing to\nthe stability and reliability of the model. On the other hand, Ensemble methods\ncombine multiple heterogeneous models to reduce the impact of remaining noise\nor errors in the model. Therefore, in this paper, we integrate discretization\ntechniques and ensemble methods and examine it on model robustness against\nadversarial attacks. In other words, we propose a discretization-based ensemble\nstacking technique to improve the security of our ML models. We evaluate the\nperformance of different ML-based IoT device identification models against\nwhite box and black box attacks using a real-world dataset comprised of network\ntraffic from 28 IoT devices. We demonstrate that the proposed method enables\nrobustness to the models for IoT device identification.\n", "date": "2023-07-18", "categories": []}
{"id": "2210.13815", "title": "FocusedCleaner: Sanitizing Poisoned Graphs for Robust GNN-based Node\n  Classification", "abstract": "  Graph Neural Networks (GNNs) are vulnerable to data poisoning attacks, which\nwill generate a poisoned graph as the input to the GNN models. We present\nFocusedCleaner as a poisoned graph sanitizer to effectively identify the poison\ninjected by attackers. Specifically, FocusedCleaner provides a sanitation\nframework consisting of two modules: bi-level structural learning and victim\nnode detection. In particular, the structural learning module will reverse the\nattack process to steadily sanitize the graph while the detection module\nprovides ``the focus\" -- a narrowed and more accurate search region -- to\nstructural learning. These two modules will operate in iterations and reinforce\neach other to sanitize a poisoned graph step by step. As an important\napplication, we show that the adversarial robustness of GNNs trained over the\nsanitized graph for the node classification task is significantly improved.\nExtensive experiments demonstrate that FocusedCleaner outperforms the\nstate-of-the-art baselines both on poisoned graph sanitation and improving\nrobustness.\n", "date": "2022-10-25", "categories": []}
{"id": "2302.14267", "title": "Adversarial Attack with Raindrops", "abstract": "  Deep neural networks (DNNs) are known to be vulnerable to adversarial\nexamples, which are usually designed artificially to fool DNNs, but rarely\nexist in real-world scenarios. In this paper, we study the adversarial examples\ncaused by raindrops, to demonstrate that there exist plenty of natural\nphenomena being able to work as adversarial attackers to DNNs. Moreover, we\npresent a new approach to generate adversarial raindrops, denoted as AdvRD,\nusing the generative adversarial network (GAN) technique to simulate natural\nraindrops. The images crafted by our AdvRD look very similar to the real-world\nraindrop images, statistically close to the distribution of true raindrop\nimages, and more importantly, can perform strong adversarial attack to the\nstate-of-the-art DNN models. On the other side, we show that the adversarial\ntraining using our AdvRD images can significantly improve the robustness of\nDNNs to the real-world raindrop attacks. Extensive experiments are carried out\nto demonstrate that the images crafted by AdvRD are visually and statistically\nclose to the natural raindrop images, can work as strong attackers to DNN\nmodels, and also help improve the robustness of DNNs to raindrop attacks.\n", "date": "2023-02-28", "categories": []}
{"id": "2304.02312", "title": "How to choose your best allies for a transferable attack?", "abstract": "  The transferability of adversarial examples is a key issue in the security of\ndeep neural networks. The possibility of an adversarial example crafted for a\nsource model fooling another targeted model makes the threat of adversarial\nattacks more realistic. Measuring transferability is a crucial problem, but the\nAttack Success Rate alone does not provide a sound evaluation. This paper\nproposes a new methodology for evaluating transferability by putting distortion\nin a central position. This new tool shows that transferable attacks may\nperform far worse than a black box attack if the attacker randomly picks the\nsource model. To address this issue, we propose a new selection mechanism,\ncalled FiT, which aims at choosing the best source model with only a few\npreliminary queries to the target. Our experimental results show that FiT is\nhighly effective at selecting the best source model for multiple scenarios such\nas single-model attacks, ensemble-model attacks and multiple attacks (Code\navailable at: https://github.com/t-maho/transferability_measure_fit).\n", "date": "2023-04-05", "categories": []}
{"id": "2306.13800", "title": "A First Order Meta Stackelberg Method for Robust Federated Learning", "abstract": "  Previous research has shown that federated learning (FL) systems are exposed\nto an array of security risks. Despite the proposal of several defensive\nstrategies, they tend to be non-adaptive and specific to certain types of\nattacks, rendering them ineffective against unpredictable or adaptive threats.\nThis work models adversarial federated learning as a Bayesian Stackelberg\nMarkov game (BSMG) to capture the defender's incomplete information of various\nattack types. We propose meta-Stackelberg learning (meta-SL), a provably\nefficient meta-learning algorithm, to solve the equilibrium strategy in BSMG,\nleading to an adaptable FL defense. We demonstrate that meta-SL converges to\nthe first-order $\\varepsilon$-equilibrium point in $O(\\varepsilon^{-2})$\ngradient iterations, with $O(\\varepsilon^{-4})$ samples needed per iteration,\nmatching the state of the art. Empirical evidence indicates that our\nmeta-Stackelberg framework performs exceptionally well against potent model\npoisoning and backdoor attacks of an uncertain nature.\n", "date": "2023-06-23", "categories": []}
{"id": "2307.06548", "title": "Multi-objective Evolutionary Search of Variable-length Composite\n  Semantic Perturbations", "abstract": "  Deep neural networks have proven to be vulnerable to adversarial attacks in\nthe form of adding specific perturbations on images to make wrong outputs.\nDesigning stronger adversarial attack methods can help more reliably evaluate\nthe robustness of DNN models. To release the harbor burden and improve the\nattack performance, auto machine learning (AutoML) has recently emerged as one\nsuccessful technique to help automatically find the near-optimal adversarial\nattack strategy. However, existing works about AutoML for adversarial attacks\nonly focus on $L_{\\infty}$-norm-based perturbations. In fact, semantic\nperturbations attract increasing attention due to their naturalnesses and\nphysical realizability. To bridge the gap between AutoML and semantic\nadversarial attacks, we propose a novel method called multi-objective\nevolutionary search of variable-length composite semantic perturbations\n(MES-VCSP). Specifically, we construct the mathematical model of\nvariable-length composite semantic perturbations, which provides five\ngradient-based semantic attack methods. The same type of perturbation in an\nattack sequence is allowed to be performed multiple times. Besides, we\nintroduce the multi-objective evolutionary search consisting of NSGA-II and\nneighborhood search to find near-optimal variable-length attack sequences.\nExperimental results on CIFAR10 and ImageNet datasets show that compared with\nexisting methods, MES-VCSP can obtain adversarial examples with a higher attack\nsuccess rate, more naturalness, and less time cost.\n", "date": "2023-07-13", "categories": []}
{"id": "2307.07653", "title": "RFLA: A Stealthy Reflected Light Adversarial Attack in the Physical\n  World", "abstract": "  Physical adversarial attacks against deep neural networks (DNNs) have\nrecently gained increasing attention. The current mainstream physical attacks\nuse printed adversarial patches or camouflage to alter the appearance of the\ntarget object. However, these approaches generate conspicuous adversarial\npatterns that show poor stealthiness. Another physical deployable attack is the\noptical attack, featuring stealthiness while exhibiting weakly in the daytime\nwith sunlight. In this paper, we propose a novel Reflected Light Attack (RFLA),\nfeaturing effective and stealthy in both the digital and physical world, which\nis implemented by placing the color transparent plastic sheet and a paper cut\nof a specific shape in front of the mirror to create different colored\ngeometries on the target object. To achieve these goals, we devise a general\nframework based on the circle to model the reflected light on the target\nobject. Specifically, we optimize a circle (composed of a coordinate and\nradius) to carry various geometrical shapes determined by the optimized angle.\nThe fill color of the geometry shape and its corresponding transparency are\nalso optimized. We extensively evaluate the effectiveness of RFLA on different\ndatasets and models. Experiment results suggest that the proposed method\nachieves over 99% success rate on different datasets and models in the digital\nworld. Additionally, we verify the effectiveness of the proposed method in\ndifferent physical environments by using sunlight or a flashlight.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.08076", "title": "Diffusion to Confusion: Naturalistic Adversarial Patch Generation Based\n  on Diffusion Model for Object Detector", "abstract": "  Many physical adversarial patch generation methods are widely proposed to\nprotect personal privacy from malicious monitoring using object detectors.\nHowever, they usually fail to generate satisfactory patch images in terms of\nboth stealthiness and attack performance without making huge efforts on careful\nhyperparameter tuning. To address this issue, we propose a novel naturalistic\nadversarial patch generation method based on the diffusion models (DM). Through\nsampling the optimal image from the DM model pretrained upon natural images, it\nallows us to stably craft high-quality and naturalistic physical adversarial\npatches to humans without suffering from serious mode collapse problems as\nother deep generative models. To the best of our knowledge, we are the first to\npropose DM-based naturalistic adversarial patch generation for object\ndetectors. With extensive quantitative, qualitative, and subjective\nexperiments, the results demonstrate the effectiveness of the proposed approach\nto generate better-quality and more naturalistic adversarial patches while\nachieving acceptable attack performance than other state-of-the-art patch\ngeneration methods. We also show various generation trade-offs under different\nconditions.\n", "date": "2023-07-16", "categories": []}
{"id": "2307.08278", "title": "Adversarial Attacks on Traffic Sign Recognition: A Survey", "abstract": "  Traffic sign recognition is an essential component of perception in\nautonomous vehicles, which is currently performed almost exclusively with deep\nneural networks (DNNs). However, DNNs are known to be vulnerable to adversarial\nattacks. Several previous works have demonstrated the feasibility of\nadversarial attacks on traffic sign recognition models. Traffic signs are\nparticularly promising for adversarial attack research due to the ease of\nperforming real-world attacks using printed signs or stickers. In this work, we\nsurvey existing works performing either digital or real-world attacks on\ntraffic sign detection and classification models. We provide an overview of the\nlatest advancements and highlight the existing research areas that require\nfurther investigation.\n", "date": "2023-07-17", "categories": []}
{"id": "2307.08288", "title": "Systematic Testing of the Data-Poisoning Robustness of KNN", "abstract": "  Data poisoning aims to compromise a machine learning based software component\nby contaminating its training set to change its prediction results for test\ninputs. Existing methods for deciding data-poisoning robustness have either\npoor accuracy or long running time and, more importantly, they can only certify\nsome of the truly-robust cases, but remain inconclusive when certification\nfails. In other words, they cannot falsify the truly-non-robust cases. To\novercome this limitation, we propose a systematic testing based method, which\ncan falsify as well as certify data-poisoning robustness for a widely used\nsupervised-learning technique named k-nearest neighbors (KNN). Our method is\nfaster and more accurate than the baseline enumeration method, due to a novel\nover-approximate analysis in the abstract domain, to quickly narrow down the\nsearch space, and systematic testing in the concrete domain, to find the actual\nviolations. We have evaluated our method on a set of supervised-learning\ndatasets. Our results show that the method significantly outperforms\nstate-of-the-art techniques, and can decide data-poisoning robustness of KNN\nprediction results for most of the test inputs.\n", "date": "2023-07-17", "categories": []}
{"id": "2307.07167", "title": "Vulnerability-Aware Instance Reweighting For Adversarial Training", "abstract": "  Adversarial Training (AT) has been found to substantially improve the\nrobustness of deep learning classifiers against adversarial attacks. AT\ninvolves obtaining robustness by including adversarial examples in training a\nclassifier. Most variants of AT algorithms treat every training example\nequally. However, recent works have shown that better performance is achievable\nby treating them unequally. In addition, it has been observed that AT exerts an\nuneven influence on different classes in a training set and unfairly hurts\nexamples corresponding to classes that are inherently harder to classify.\nConsequently, various reweighting schemes have been proposed that assign\nunequal weights to robust losses of individual examples in a training set. In\nthis work, we propose a novel instance-wise reweighting scheme. It considers\nthe vulnerability of each natural example and the resulting information loss on\nits adversarial counterpart occasioned by adversarial attacks. Through\nextensive experiments, we show that our proposed method significantly improves\nover existing reweighting schemes, especially against strong white and\nblack-box attacks.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.07328", "title": "Boosting Backdoor Attack with A Learnable Poisoning Sample Selection\n  Strategy", "abstract": "  Data-poisoning based backdoor attacks aim to insert backdoor into models by\nmanipulating training datasets without controlling the training process of the\ntarget model. Existing attack methods mainly focus on designing triggers or\nfusion strategies between triggers and benign samples. However, they often\nrandomly select samples to be poisoned, disregarding the varying importance of\neach poisoning sample in terms of backdoor injection. A recent selection\nstrategy filters a fixed-size poisoning sample pool by recording forgetting\nevents, but it fails to consider the remaining samples outside the pool from a\nglobal perspective. Moreover, computing forgetting events requires significant\nadditional computing resources. Therefore, how to efficiently and effectively\nselect poisoning samples from the entire dataset is an urgent problem in\nbackdoor attacks.To address it, firstly, we introduce a poisoning mask into the\nregular backdoor training loss. We suppose that a backdoored model training\nwith hard poisoning samples has a more backdoor effect on easy ones, which can\nbe implemented by hindering the normal training process (\\ie, maximizing loss\n\\wrt mask). To further integrate it with normal training process, we then\npropose a learnable poisoning sample selection strategy to learn the mask\ntogether with the model parameters through a min-max optimization.Specifically,\nthe outer loop aims to achieve the backdoor attack goal by minimizing the loss\nbased on the selected samples, while the inner loop selects hard poisoning\nsamples that impede this goal by maximizing the loss. After several rounds of\nadversarial training, we finally select effective poisoning samples with high\ncontribution. Extensive experiments on benchmark datasets demonstrate the\neffectiveness and efficiency of our approach in boosting backdoor attack\nperformance.\n", "date": "2023-07-14", "categories": []}
{"id": "2211.00241", "title": "Adversarial Policies Beat Superhuman Go AIs", "abstract": "  We attack the state-of-the-art Go-playing AI system KataGo by training\nadversarial policies against it, achieving a >97% win rate against KataGo\nrunning at superhuman settings. Our adversaries do not win by playing Go well.\nInstead, they trick KataGo into making serious blunders. Our attack transfers\nzero-shot to other superhuman Go-playing AIs, and is comprehensible to the\nextent that human experts can implement it without algorithmic assistance to\nconsistently beat superhuman AIs. The core vulnerability uncovered by our\nattack persists even in KataGo agents adversarially trained to defend against\nour attack. Our results demonstrate that even superhuman AI systems may harbor\nsurprising failure modes. Example games are available https://goattack.far.ai/.\n", "date": "2022-11-01", "categories": []}
{"id": "2307.06484", "title": "Single-Class Target-Specific Attack against Interpretable Deep Learning\n  Systems", "abstract": "  In this paper, we present a novel Single-class target-specific Adversarial\nattack called SingleADV. The goal of SingleADV is to generate a universal\nperturbation that deceives the target model into confusing a specific category\nof objects with a target category while ensuring highly relevant and accurate\ninterpretations. The universal perturbation is stochastically and iteratively\noptimized by minimizing the adversarial loss that is designed to consider both\nthe classifier and interpreter costs in targeted and non-targeted categories.\nIn this optimization framework, ruled by the first- and second-moment\nestimations, the desired loss surface promotes high confidence and\ninterpretation score of adversarial samples. By avoiding unintended\nmisclassification of samples from other categories, SingleADV enables more\neffective targeted attacks on interpretable deep learning systems in both\nwhite-box and black-box scenarios. To evaluate the effectiveness of SingleADV,\nwe conduct experiments using four different model architectures (ResNet-50,\nVGG-16, DenseNet-169, and Inception-V3) coupled with three interpretation\nmodels (CAM, Grad, and MASK). Through extensive empirical evaluation, we\ndemonstrate that SingleADV effectively deceives the target deep learning models\nand their associated interpreters under various conditions and settings. Our\nexperimental results show that the performance of SingleADV is effective, with\nan average fooling ratio of 0.74 and an adversarial confidence level of 0.78 in\ngenerating deceptive adversarial samples. Furthermore, we discuss several\ncountermeasures against SingleADV, including a transfer-based learning approach\nand existing preprocessing defenses.\n", "date": "2023-07-12", "categories": []}
{"id": "2307.06496", "title": "Microbial Genetic Algorithm-based Black-box Attack against Interpretable\n  Deep Learning Systems", "abstract": "  Deep learning models are susceptible to adversarial samples in white and\nblack-box environments. Although previous studies have shown high attack\nsuccess rates, coupling DNN models with interpretation models could offer a\nsense of security when a human expert is involved, who can identify whether a\ngiven sample is benign or malicious. However, in white-box environments,\ninterpretable deep learning systems (IDLSes) have been shown to be vulnerable\nto malicious manipulations. In black-box settings, as access to the components\nof IDLSes is limited, it becomes more challenging for the adversary to fool the\nsystem. In this work, we propose a Query-efficient Score-based black-box attack\nagainst IDLSes, QuScore, which requires no knowledge of the target model and\nits coupled interpretation model. QuScore is based on transfer-based and\nscore-based methods by employing an effective microbial genetic algorithm. Our\nmethod is designed to reduce the number of queries necessary to carry out\nsuccessful attacks, resulting in a more efficient process. By continuously\nrefining the adversarial samples created based on feedback scores from the\nIDLS, our approach effectively navigates the search space to identify\nperturbations that can fool the system. We evaluate the attack's effectiveness\non four CNN models (Inception, ResNet, VGG, DenseNet) and two interpretation\nmodels (CAM, Grad), using both ImageNet and CIFAR datasets. Our results show\nthat the proposed approach is query-efficient with a high attack success rate\nthat can reach between 95% and 100% and transferability with an average success\nrate of 69% in the ImageNet and CIFAR datasets. Our attack method generates\nadversarial examples with attribution maps that resemble benign samples. We\nhave also demonstrated that our attack is resilient against various\npreprocessing defense techniques and can easily be transferred to different DNN\nmodels.\n", "date": "2023-07-13", "categories": []}
{"id": "2212.10258", "title": "In and Out-of-Domain Text Adversarial Robustness via Label Smoothing", "abstract": "  Recently it has been shown that state-of-the-art NLP models are vulnerable to\nadversarial attacks, where the predictions of a model can be drastically\naltered by slight modifications to the input (such as synonym substitutions).\nWhile several defense techniques have been proposed, and adapted, to the\ndiscrete nature of text adversarial attacks, the benefits of general-purpose\nregularization methods such as label smoothing for language models, have not\nbeen studied. In this paper, we study the adversarial robustness provided by\nvarious label smoothing strategies in foundational models for diverse NLP tasks\nin both in-domain and out-of-domain settings. Our experiments show that label\nsmoothing significantly improves adversarial robustness in pre-trained models\nlike BERT, against various popular attacks. We also analyze the relationship\nbetween prediction confidence and robustness, showing that label smoothing\nreduces over-confident errors on adversarial examples.\n", "date": "2022-12-20", "categories": []}
{"id": "2211.11030", "title": "Adversarial Cheap Talk", "abstract": "  Adversarial attacks in reinforcement learning (RL) often assume\nhighly-privileged access to the victim's parameters, environment, or data.\nInstead, this paper proposes a novel adversarial setting called a Cheap Talk\nMDP in which an Adversary can merely append deterministic messages to the\nVictim's observation, resulting in a minimal range of influence. The Adversary\ncannot occlude ground truth, influence underlying environment dynamics or\nreward signals, introduce non-stationarity, add stochasticity, see the Victim's\nactions, or access their parameters. Additionally, we present a simple\nmeta-learning algorithm called Adversarial Cheap Talk (ACT) to train\nAdversaries in this setting. We demonstrate that an Adversary trained with ACT\nstill significantly influences the Victim's training and testing performance,\ndespite the highly constrained setting. Affecting train-time performance\nreveals a new attack vector and provides insight into the success and failure\nmodes of existing RL algorithms. More specifically, we show that an ACT\nAdversary is capable of harming performance by interfering with the learner's\nfunction approximation, or instead helping the Victim's performance by\noutputting useful features. Finally, we show that an ACT Adversary can\nmanipulate messages during train-time to directly and arbitrarily control the\nVictim at test-time. Project video and code are available at\nhttps://sites.google.com/view/adversarial-cheap-talk\n", "date": "2022-11-20", "categories": []}
{"id": "2307.05095", "title": "ATWM: Defense against adversarial malware based on adversarial training", "abstract": "  Deep learning technology has made great achievements in the field of image.\nIn order to defend against malware attacks, researchers have proposed many\nWindows malware detection models based on deep learning. However, deep learning\nmodels are vulnerable to adversarial example attacks. Malware can generate\nadversarial malware with the same malicious function to attack the malware\ndetection model and evade detection of the model. Currently, many adversarial\ndefense studies have been proposed, but existing adversarial defense studies\nare based on image sample and cannot be directly applied to malware sample.\nTherefore, this paper proposes an adversarial malware defense method based on\nadversarial training. This method uses preprocessing to defend simple\nadversarial examples to reduce the difficulty of adversarial training.\nMoreover, this method improves the adversarial defense capability of the model\nthrough adversarial training. We experimented with three attack methods in two\nsets of datasets, and the results show that the method in this paper can\nimprove the adversarial defense capability of the model without reducing the\naccuracy of the model.\n", "date": "2023-07-11", "categories": []}
{"id": "2307.05193", "title": "Membership Inference Attacks on DNNs using Adversarial Perturbations", "abstract": "  Several membership inference (MI) attacks have been proposed to audit a\ntarget DNN. Given a set of subjects, MI attacks tell which subjects the target\nDNN has seen during training. This work focuses on the post-training MI attacks\nemphasizing high confidence membership detection -- True Positive Rates (TPR)\nat low False Positive Rates (FPR). Current works in this category -- likelihood\nratio attack (LiRA) and enhanced MI attack (EMIA) -- only perform well on\ncomplex datasets (e.g., CIFAR-10 and Imagenet) where the target DNN overfits\nits train set, but perform poorly on simpler datasets (0% TPR by both attacks\non Fashion-MNIST, 2% and 0% TPR respectively by LiRA and EMIA on MNIST at 1%\nFPR). To address this, firstly, we unify current MI attacks by presenting a\nframework divided into three stages -- preparation, indication and decision.\nSecondly, we utilize the framework to propose two novel attacks: (1)\nAdversarial Membership Inference Attack (AMIA) efficiently utilizes the\nmembership and the non-membership information of the subjects while\nadversarially minimizing a novel loss function, achieving 6% TPR on both\nFashion-MNIST and MNIST datasets; and (2) Enhanced AMIA (E-AMIA) combines EMIA\nand AMIA to achieve 8% and 4% TPRs on Fashion-MNIST and MNIST datasets\nrespectively, at 1% FPR. Thirdly, we introduce two novel augmented indicators\nthat positively leverage the loss information in the Gaussian neighborhood of a\nsubject. This improves TPR of all four attacks on average by 2.5% and 0.25%\nrespectively on Fashion-MNIST and MNIST datasets at 1% FPR. Finally, we propose\nsimple, yet novel, evaluation metric, the running TPR average (RTA) at a given\nFPR, that better distinguishes different MI attacks in the low FPR region. We\nalso show that AMIA and E-AMIA are more transferable to the unknown DNNs (other\nthan the target DNN) and are more robust to DP-SGD training as compared to LiRA\nand EMIA.\n", "date": "2023-07-11", "categories": []}
{"id": "2307.03803", "title": "A Theoretical Perspective on Subnetwork Contributions to Adversarial\n  Robustness", "abstract": "  The robustness of deep neural networks (DNNs) against adversarial attacks has\nbeen studied extensively in hopes of both better understanding how deep\nlearning models converge and in order to ensure the security of these models in\nsafety-critical applications. Adversarial training is one approach to\nstrengthening DNNs against adversarial attacks, and has been shown to offer a\nmeans for doing so at the cost of applying computationally expensive training\nmethods to the entire model. To better understand these attacks and facilitate\nmore efficient adversarial training, in this paper we develop a novel\ntheoretical framework that investigates how the adversarial robustness of a\nsubnetwork contributes to the robustness of the entire network. To do so we\nfirst introduce the concept of semirobustness, which is a measure of the\nadversarial robustness of a subnetwork. Building on this concept, we then\nprovide a theoretical analysis to show that if a subnetwork is semirobust and\nthere is a sufficient dependency between it and each subsequent layer in the\nnetwork, then the remaining layers are also guaranteed to be robust. We\nvalidate these findings empirically across multiple DNN architectures,\ndatasets, and adversarial attacks. Experiments show the ability of a robust\nsubnetwork to promote full-network robustness, and investigate the layer-wise\ndependencies required for this full-network robustness to be achieved.\n", "date": "2023-07-07", "categories": []}
{"id": "2307.04066", "title": "Random Position Adversarial Patch for Vision Transformers", "abstract": "  Previous studies have shown the vulnerability of vision transformers to\nadversarial patches, but these studies all rely on a critical assumption: the\nattack patches must be perfectly aligned with the patches used for linear\nprojection in vision transformers. Due to this stringent requirement, deploying\nadversarial patches for vision transformers in the physical world becomes\nimpractical, unlike their effectiveness on CNNs. This paper proposes a novel\nmethod for generating an adversarial patch (G-Patch) that overcomes the\nalignment constraint, allowing the patch to launch a targeted attack at any\nposition within the field of view. Specifically, instead of directly optimizing\nthe patch using gradients, we employ a GAN-like structure to generate the\nadversarial patch. Our experiments show the effectiveness of the adversarial\npatch in achieving universal attacks on vision transformers, both in digital\nand physical-world scenarios. Additionally, further analysis reveals that the\ngenerated adversarial patch exhibits robustness to brightness restriction,\ncolor transfer, and random noise. Real-world attack experiments validate the\neffectiveness of the G-Patch to launch robust attacks even under some very\nchallenging conditions.\n", "date": "2023-07-09", "categories": []}
{"id": "2307.04099", "title": "GNP Attack: Transferable Adversarial Examples via Gradient Norm Penalty", "abstract": "  Adversarial examples (AE) with good transferability enable practical\nblack-box attacks on diverse target models, where insider knowledge about the\ntarget models is not required. Previous methods often generate AE with no or\nvery limited transferability; that is, they easily overfit to the particular\narchitecture and feature representation of the source, white-box model and the\ngenerated AE barely work for target, black-box models. In this paper, we\npropose a novel approach to enhance AE transferability using Gradient Norm\nPenalty (GNP). It drives the loss function optimization procedure to converge\nto a flat region of local optima in the loss landscape. By attacking 11\nstate-of-the-art (SOTA) deep learning models and 6 advanced defense methods, we\nempirically show that GNP is very effective in generating AE with high\ntransferability. We also demonstrate that it is very flexible in that it can be\neasily integrated with other gradient based methods for stronger transfer-based\nattacks.\n", "date": "2023-07-09", "categories": []}
{"id": "2210.13235", "title": "Chaos Theory and Adversarial Robustness", "abstract": "  Neural networks, being susceptible to adversarial attacks, should face a\nstrict level of scrutiny before being deployed in critical or adversarial\napplications. This paper uses ideas from Chaos Theory to explain, analyze, and\nquantify the degree to which neural networks are susceptible to or robust\nagainst adversarial attacks. To this end, we present a new metric, the\n\"susceptibility ratio,\" given by $\\hat \\Psi(h, \\theta)$, which captures how\ngreatly a model's output will be changed by perturbations to a given input.\n  Our results show that susceptibility to attack grows significantly with the\ndepth of the model, which has safety implications for the design of neural\nnetworks for production environments. We provide experimental evidence of the\nrelationship between $\\hat \\Psi$ and the post-attack accuracy of classification\nmodels, as well as a discussion of its application to tasks lacking hard\ndecision boundaries. We also demonstrate how to quickly and easily approximate\nthe certified robustness radii for extremely large models, which until now has\nbeen computationally infeasible to calculate directly.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.09643", "title": "Improving Adversarial Robustness by Contrastive Guided Diffusion Process", "abstract": "  Synthetic data generation has become an emerging tool to help improve the\nadversarial robustness in classification tasks since robust learning requires a\nsignificantly larger amount of training samples compared with standard\nclassification tasks. Among various deep generative models, the diffusion model\nhas been shown to produce high-quality synthetic images and has achieved good\nperformance in improving the adversarial robustness. However, diffusion-type\nmethods are typically slow in data generation as compared with other generative\nmodels. Although different acceleration techniques have been proposed recently,\nit is also of great importance to study how to improve the sample efficiency of\ngenerated data for the downstream task. In this paper, we first analyze the\noptimality condition of synthetic distribution for achieving non-trivial robust\naccuracy. We show that enhancing the distinguishability among the generated\ndata is critical for improving adversarial robustness. Thus, we propose the\nContrastive-Guided Diffusion Process (Contrastive-DP), which adopts the\ncontrastive loss to guide the diffusion model in data generation. We verify our\ntheoretical results using simulations and demonstrate the good performance of\nContrastive-DP on image datasets.\n", "date": "2022-10-18", "categories": []}
{"id": "2305.01361", "title": "Boosting Adversarial Transferability via Fusing Logits of Top-1\n  Decomposed Feature", "abstract": "  Recent research has shown that Deep Neural Networks (DNNs) are highly\nvulnerable to adversarial samples, which are highly transferable and can be\nused to attack other unknown black-box models. To improve the transferability\nof adversarial samples, several feature-based adversarial attack methods have\nbeen proposed to disrupt neuron activation in the middle layers. However,\ncurrent state-of-the-art feature-based attack methods typically require\nadditional computation costs for estimating the importance of neurons. To\naddress this challenge, we propose a Singular Value Decomposition (SVD)-based\nfeature-level attack method. Our approach is inspired by the discovery that\neigenvectors associated with the larger singular values decomposed from the\nmiddle layer features exhibit superior generalization and attention properties.\nSpecifically, we conduct the attack by retaining the decomposed Top-1 singular\nvalue-associated feature for computing the output logits, which are then\ncombined with the original logits to optimize adversarial examples. Our\nextensive experimental results verify the effectiveness of our proposed method,\nwhich can be easily integrated into various baselines to significantly enhance\nthe transferability of adversarial samples for disturbing normally trained CNNs\nand advanced defense strategies. The source code of this study is available at\nhttps://github.com/WJJLL/SVD-SSA\n", "date": "2023-05-02", "categories": []}
{"id": "2306.17431", "title": "Defense against Adversarial Cloud Attack on Remote Sensing Salient\n  Object Detection", "abstract": "  Detecting the salient objects in a remote sensing image has wide applications\nfor the interdisciplinary research. Many existing deep learning methods have\nbeen proposed for Salient Object Detection (SOD) in remote sensing images and\nget remarkable results. However, the recent adversarial attack examples,\ngenerated by changing a few pixel values on the original remote sensing image,\ncould result in a collapse for the well-trained deep learning based SOD model.\nDifferent with existing methods adding perturbation to original images, we\npropose to jointly tune adversarial exposure and additive perturbation for\nattack and constrain image close to cloudy image as Adversarial Cloud. Cloud is\nnatural and common in remote sensing images, however, camouflaging cloud based\nadversarial attack and defense for remote sensing images are not well studied\nbefore. Furthermore, we design DefenseNet as a learn-able pre-processing to the\nadversarial cloudy images so as to preserve the performance of the deep\nlearning based remote sensing SOD model, without tuning the already deployed\ndeep SOD model. By considering both regular and generalized adversarial\nexamples, the proposed DefenseNet can defend the proposed Adversarial Cloud in\nwhite-box setting and other attack methods in black-box setting. Experimental\nresults on a synthesized benchmark from the public remote sensing SOD dataset\n(EORSSD) show the promising defense against adversarial cloud attacks.\n", "date": "2023-06-30", "categories": []}
{"id": "2307.01390", "title": "Adversarial Learning in Real-World Fraud Detection: Challenges and\n  Perspectives", "abstract": "  Data economy relies on data-driven systems and complex machine learning\napplications are fueled by them. Unfortunately, however, machine learning\nmodels are exposed to fraudulent activities and adversarial attacks, which\nthreaten their security and trustworthiness. In the last decade or so, the\nresearch interest on adversarial machine learning has grown significantly,\nrevealing how learning applications could be severely impacted by effective\nattacks. Although early results of adversarial machine learning indicate the\nhuge potential of the approach to specific domains such as image processing,\nstill there is a gap in both the research literature and practice regarding how\nto generalize adversarial techniques in other domains and applications. Fraud\ndetection is a critical defense mechanism for data economy, as it is for other\napplications as well, which poses several challenges for machine learning. In\nthis work, we describe how attacks against fraud detection systems differ from\nother applications of adversarial machine learning, and propose a number of\ninteresting directions to bridge this gap.\n", "date": "2023-07-03", "categories": []}
{"id": "2307.01488", "title": "SCAT: Robust Self-supervised Contrastive Learning via Adversarial\n  Training for Text Classification", "abstract": "  Despite their promising performance across various natural language\nprocessing (NLP) tasks, current NLP systems are vulnerable to textual\nadversarial attacks. To defend against these attacks, most existing methods\napply adversarial training by incorporating adversarial examples. However,\nthese methods have to rely on ground-truth labels to generate adversarial\nexamples, rendering it impractical for large-scale model pre-training which is\ncommonly used nowadays for NLP and many other tasks. In this paper, we propose\na novel learning framework called SCAT (Self-supervised Contrastive Learning\nvia Adversarial Training), which can learn robust representations without\nrequiring labeled data. Specifically, SCAT modifies random augmentations of the\ndata in a fully labelfree manner to generate adversarial examples. Adversarial\ntraining is achieved by minimizing the contrastive loss between the\naugmentations and their adversarial counterparts. We evaluate SCAT on two text\nclassification datasets using two state-of-the-art attack schemes proposed\nrecently. Our results show that SCAT can not only train robust language models\nfrom scratch, but it can also significantly improve the robustness of existing\npre-trained language models. Moreover, to demonstrate its flexibility, we show\nthat SCAT can also be combined with supervised adversarial training to further\nenhance model robustness.\n", "date": "2023-07-04", "categories": []}
{"id": "2307.01520", "title": "LEAT: Towards Robust Deepfake Disruption in Real-World Scenarios via\n  Latent Ensemble Attack", "abstract": "  Deepfakes, malicious visual contents created by generative models, pose an\nincreasingly harmful threat to society. To proactively mitigate deepfake\ndamages, recent studies have employed adversarial perturbation to disrupt\ndeepfake model outputs. However, previous approaches primarily focus on\ngenerating distorted outputs based on only predetermined target attributes,\nleading to a lack of robustness in real-world scenarios where target attributes\nare unknown. Additionally, the transferability of perturbations between two\nprominent generative models, Generative Adversarial Networks (GANs) and\nDiffusion Models, remains unexplored. In this paper, we emphasize the\nimportance of target attribute-transferability and model-transferability for\nachieving robust deepfake disruption. To address this challenge, we propose a\nsimple yet effective disruption method called Latent Ensemble ATtack (LEAT),\nwhich attacks the independent latent encoding process. By disrupting the latent\nencoding process, it generates distorted output images in subsequent generation\nprocesses, regardless of the given target attributes. This target\nattribute-agnostic attack ensures robust disruption even when the target\nattributes are unknown. Additionally, we introduce a Normalized Gradient\nEnsemble strategy that effectively aggregates gradients for iterative gradient\nattacks, enabling simultaneous attacks on various types of deepfake models,\ninvolving both GAN-based and Diffusion-based models. Moreover, we demonstrate\nthe insufficiency of evaluating disruption quality solely based on pixel-level\ndifferences. As a result, we propose an alternative protocol for\ncomprehensively evaluating the success of defense. Extensive experiments\nconfirm the efficacy of our method in disrupting deepfakes in real-world\nscenarios, reporting a higher defense success rate compared to previous\nmethods.\n", "date": "2023-07-04", "categories": []}
{"id": "2307.02055", "title": "Adversarial Attacks on Image Classification Models: FGSM and Patch\n  Attacks and their Impact", "abstract": "  This chapter introduces the concept of adversarial attacks on image\nclassification models built on convolutional neural networks (CNN). CNNs are\nvery popular deep-learning models which are used in image classification tasks.\nHowever, very powerful and pre-trained CNN models working very accurately on\nimage datasets for image classification tasks may perform disastrously when the\nnetworks are under adversarial attacks. In this work, two very well-known\nadversarial attacks are discussed and their impact on the performance of image\nclassifiers is analyzed. These two adversarial attacks are the fast gradient\nsign method (FGSM) and adversarial patch attack. These attacks are launched on\nthree powerful pre-trained image classifier architectures, ResNet-34,\nGoogleNet, and DenseNet-161. The classification accuracy of the models in the\nabsence and presence of the two attacks are computed on images from the\npublicly accessible ImageNet dataset. The results are analyzed to evaluate the\nimpact of the attacks on the image classification task.\n", "date": "2023-07-05", "categories": []}
{"id": "2307.02202", "title": "On the Adversarial Robustness of Generative Autoencoders in the Latent\n  Space", "abstract": "  The generative autoencoders, such as the variational autoencoders or the\nadversarial autoencoders, have achieved great success in lots of real-world\napplications, including image generation, and signal communication.\n  However, little concern has been devoted to their robustness during practical\ndeployment.\n  Due to the probabilistic latent structure, variational autoencoders (VAEs)\nmay confront problems such as a mismatch between the posterior distribution of\nthe latent and real data manifold, or discontinuity in the posterior\ndistribution of the latent.\n  This leaves a back door for malicious attackers to collapse VAEs from the\nlatent space, especially in scenarios where the encoder and decoder are used\nseparately, such as communication and compressed sensing.\n  In this work, we provide the first study on the adversarial robustness of\ngenerative autoencoders in the latent space.\n  Specifically, we empirically demonstrate the latent vulnerability of popular\ngenerative autoencoders through attacks in the latent space.\n  We also evaluate the difference between variational autoencoders and their\ndeterministic variants and observe that the latter performs better in latent\nrobustness.\n  Meanwhile, we identify a potential trade-off between the adversarial\nrobustness and the degree of the disentanglement of the latent codes.\n  Additionally, we also verify the feasibility of improvement for the latent\nrobustness of VAEs through adversarial training.\n  In summary, we suggest concerning the adversarial latent robustness of the\ngenerative autoencoders, analyze several robustness-relative issues, and give\nsome insights into a series of key challenges.\n", "date": "2023-07-05", "categories": []}
{"id": "2211.12044", "title": "Backdoor Cleansing with Unlabeled Data", "abstract": "  Due to the increasing computational demand of Deep Neural Networks (DNNs),\ncompanies and organizations have begun to outsource the training process.\nHowever, the externally trained DNNs can potentially be backdoor attacked. It\nis crucial to defend against such attacks, i.e., to postprocess a suspicious\nmodel so that its backdoor behavior is mitigated while its normal prediction\npower on clean inputs remain uncompromised. To remove the abnormal backdoor\nbehavior, existing methods mostly rely on additional labeled clean samples.\nHowever, such requirement may be unrealistic as the training data are often\nunavailable to end users. In this paper, we investigate the possibility of\ncircumventing such barrier. We propose a novel defense method that does not\nrequire training labels. Through a carefully designed layer-wise weight\nre-initialization and knowledge distillation, our method can effectively\ncleanse backdoor behaviors of a suspicious network with negligible compromise\nin its normal behavior. In experiments, we show that our method, trained\nwithout labels, is on-par with state-of-the-art defense methods trained using\nlabels. We also observe promising defense results even on out-of-distribution\ndata. This makes our method very practical. Code is available at:\nhttps://github.com/luluppang/BCU.\n", "date": "2022-11-22", "categories": []}
{"id": "2211.15875", "title": "Data Poisoning Attack Aiming the Vulnerability of Continual Learning", "abstract": "  Generally, regularization-based continual learning models limit access to the\nprevious task data to imitate the real-world constraints related to memory and\nprivacy. However, this introduces a problem in these models by not being able\nto track the performance on each task. In essence, current continual learning\nmethods are susceptible to attacks on previous tasks. We demonstrate the\nvulnerability of regularization-based continual learning methods by presenting\na simple task-specific data poisoning attack that can be used in the learning\nprocess of a new task. Training data generated by the proposed attack causes\nperformance degradation on a specific task targeted by the attacker. We\nexperiment with the attack on the two representative regularization-based\ncontinual learning methods, Elastic Weight Consolidation (EWC) and Synaptic\nIntelligence (SI), trained with variants of MNIST dataset. The experiment\nresults justify the vulnerability proposed in this paper and demonstrate the\nimportance of developing continual learning models that are robust to\nadversarial attacks.\n", "date": "2022-11-29", "categories": []}
{"id": "2305.05392", "title": "Sharpness-Aware Minimization Alone can Improve Adversarial Robustness", "abstract": "  Sharpness-Aware Minimization (SAM) is an effective method for improving\ngeneralization ability by regularizing loss sharpness. In this paper, we\nexplore SAM in the context of adversarial robustness. We find that using only\nSAM can achieve superior adversarial robustness without sacrificing clean\naccuracy compared to standard training, which is an unexpected benefit. We also\ndiscuss the relation between SAM and adversarial training (AT), a popular\nmethod for improving the adversarial robustness of DNNs. In particular, we show\nthat SAM and AT differ in terms of perturbation strength, leading to different\naccuracy and robustness trade-offs. We provide theoretical evidence for these\nclaims in a simplified model. Finally, while AT suffers from decreased clean\naccuracy and computational overhead, we suggest that SAM can be regarded as a\nlightweight substitute for AT under certain requirements. Code is available at\nhttps://github.com/weizeming/SAM_AT.\n", "date": "2023-05-09", "categories": []}
{"id": "2306.08422", "title": "X-Detect: Explainable Adversarial Patch Detection for Object Detectors\n  in Retail", "abstract": "  Object detection models, which are widely used in various domains (such as\nretail), have been shown to be vulnerable to adversarial attacks. Existing\nmethods for detecting adversarial attacks on object detectors have had\ndifficulty detecting new real-life attacks. We present X-Detect, a novel\nadversarial patch detector that can: i) detect adversarial samples in real\ntime, allowing the defender to take preventive action; ii) provide explanations\nfor the alerts raised to support the defender's decision-making process, and\niii) handle unfamiliar threats in the form of new attacks. Given a new scene,\nX-Detect uses an ensemble of explainable-by-design detectors that utilize\nobject extraction, scene manipulation, and feature transformation techniques to\ndetermine whether an alert needs to be raised. X-Detect was evaluated in both\nthe physical and digital space using five different attack scenarios (including\nadaptive attacks) and the COCO dataset and our new Superstore dataset. The\nphysical evaluation was performed using a smart shopping cart setup in\nreal-world settings and included 17 adversarial patch attacks recorded in 1,700\nadversarial videos. The results showed that X-Detect outperforms the\nstate-of-the-art methods in distinguishing between benign and adversarial\nscenes for all attack scenarios while maintaining a 0% FPR (no false alarms)\nand providing actionable explanations for the alerts raised. A demo is\navailable.\n", "date": "2023-06-14", "categories": []}
{"id": "2307.00274", "title": "Common Knowledge Learning for Generating Transferable Adversarial\n  Examples", "abstract": "  This paper focuses on an important type of black-box attacks, i.e.,\ntransfer-based adversarial attacks, where the adversary generates adversarial\nexamples by a substitute (source) model and utilize them to attack an unseen\ntarget model, without knowing its information. Existing methods tend to give\nunsatisfactory adversarial transferability when the source and target models\nare from different types of DNN architectures (e.g. ResNet-18 and Swin\nTransformer). In this paper, we observe that the above phenomenon is induced by\nthe output inconsistency problem. To alleviate this problem while effectively\nutilizing the existing DNN models, we propose a common knowledge learning (CKL)\nframework to learn better network weights to generate adversarial examples with\nbetter transferability, under fixed network architectures. Specifically, to\nreduce the model-specific features and obtain better output distributions, we\nconstruct a multi-teacher framework, where the knowledge is distilled from\ndifferent teacher architectures into one student network. By considering that\nthe gradient of input is usually utilized to generated adversarial examples, we\nimpose constraints on the gradients between the student and teacher models, to\nfurther alleviate the output inconsistency problem and enhance the adversarial\ntransferability. Extensive experiments demonstrate that our proposed work can\nsignificantly improve the adversarial transferability.\n", "date": "2023-07-01", "categories": []}
{"id": "2307.00421", "title": "Brightness-Restricted Adversarial Attack Patch", "abstract": "  Adversarial attack patches have gained increasing attention due to their\npractical applicability in physical-world scenarios. However, the bright colors\nused in attack patches represent a significant drawback, as they can be easily\nidentified by human observers. Moreover, even though these attacks have been\nhighly successful in deceiving target networks, which specific features of the\nattack patch contribute to its success are still unknown. Our paper introduces\na brightness-restricted patch (BrPatch) that uses optical characteristics to\neffectively reduce conspicuousness while preserving image independence. We also\nconducted an analysis of the impact of various image features (such as color,\ntexture, noise, and size) on the effectiveness of an attack patch in\nphysical-world deployment. Our experiments show that attack patches exhibit\nstrong redundancy to brightness and are resistant to color transfer and noise.\nBased on our findings, we propose some additional methods to further reduce the\nconspicuousness of BrPatch. Our findings also explain the robustness of attack\npatches observed in physical-world scenarios.\n", "date": "2023-07-01", "categories": []}
{"id": "2307.00477", "title": "Query-Efficient Decision-based Black-Box Patch Attack", "abstract": "  Deep neural networks (DNNs) have been showed to be highly vulnerable to\nimperceptible adversarial perturbations. As a complementary type of adversary,\npatch attacks that introduce perceptible perturbations to the images have\nattracted the interest of researchers. Existing patch attacks rely on the\narchitecture of the model or the probabilities of predictions and perform\npoorly in the decision-based setting, which can still construct a perturbation\nwith the minimal information exposed -- the top-1 predicted label. In this\nwork, we first explore the decision-based patch attack. To enhance the attack\nefficiency, we model the patches using paired key-points and use targeted\nimages as the initialization of patches, and parameter optimizations are all\nperformed on the integer domain. Then, we propose a differential evolutionary\nalgorithm named DevoPatch for query-efficient decision-based patch attacks.\nExperiments demonstrate that DevoPatch outperforms the state-of-the-art\nblack-box patch attacks in terms of patch area and attack success rate within a\ngiven query budget on image classification and face verification. Additionally,\nwe conduct the vulnerability evaluation of ViT and MLP on image classification\nin the decision-based patch attack setting for the first time. Using DevoPatch,\nwe can evaluate the robustness of models to black-box patch attacks. We believe\nthis method could inspire the design and deployment of robust vision models\nbased on various DNN architectures in the future.\n", "date": "2023-07-02", "categories": []}
{"id": "2302.10739", "title": "MalProtect: Stateful Defense Against Adversarial Query Attacks in\n  ML-based Malware Detection", "abstract": "  ML models are known to be vulnerable to adversarial query attacks. In these\nattacks, queries are iteratively perturbed towards a particular class without\nany knowledge of the target model besides its output. The prevalence of\nremotely-hosted ML classification models and Machine-Learning-as-a-Service\nplatforms means that query attacks pose a real threat to the security of these\nsystems. To deal with this, stateful defenses have been proposed to detect\nquery attacks and prevent the generation of adversarial examples by monitoring\nand analyzing the sequence of queries received by the system. Several stateful\ndefenses have been proposed in recent years. However, these defenses rely\nsolely on similarity or out-of-distribution detection methods that may be\neffective in other domains. In the malware detection domain, the methods to\ngenerate adversarial examples are inherently different, and therefore we find\nthat such detection mechanisms are significantly less effective. Hence, in this\npaper, we present MalProtect, which is a stateful defense against query attacks\nin the malware detection domain. MalProtect uses several threat indicators to\ndetect attacks. Our results show that it reduces the evasion rate of\nadversarial query attacks by 80+\\% in Android and Windows malware, across a\nrange of attacker scenarios. In the first evaluation of its kind, we show that\nMalProtect outperforms prior stateful defenses, especially under the peak\nadversarial threat.\n", "date": "2023-02-21", "categories": []}
{"id": "2305.09671", "title": "Pick your Poison: Undetectability versus Robustness in Data Poisoning\n  Attacks", "abstract": "  Deep image classification models trained on vast amounts of web-scraped data\nare susceptible to data poisoning - a mechanism for backdooring models. A small\nnumber of poisoned samples seen during training can severely undermine a\nmodel's integrity during inference. Existing work considers an effective\ndefense as one that either (i) restores a model's integrity through repair or\n(ii) detects an attack. We argue that this approach overlooks a crucial\ntrade-off: Attackers can increase robustness at the expense of detectability\n(over-poisoning) or decrease detectability at the cost of robustness\n(under-poisoning). In practice, attacks should remain both undetectable and\nrobust. Detectable but robust attacks draw human attention and rigorous model\nevaluation or cause the model to be re-trained or discarded. In contrast,\nattacks that are undetectable but lack robustness can be repaired with minimal\nimpact on model accuracy. Our research points to intrinsic flaws in current\nattack evaluation methods and raises the bar for all data poisoning attackers\nwho must delicately balance this trade-off to remain robust and undetectable.\nTo demonstrate the existence of more potent defenders, we propose defenses\ndesigned to (i) detect or (ii) repair poisoned models using a limited amount of\ntrusted image-label pairs. Our results show that an attacker who needs to be\nrobust and undetectable is substantially less threatening. Our defenses\nmitigate all tested attacks with a maximum accuracy decline of 2% using only 1%\nof clean data on CIFAR-10 and 2.5% on ImageNet. We demonstrate the scalability\nof our defenses by evaluating large vision-language models, such as CLIP.\nAttackers who can manipulate the model's parameters pose an elevated risk as\nthey can achieve higher robustness at low detectability compared to data\npoisoning attackers.\n", "date": "2023-05-07", "categories": []}
{"id": "2306.16581", "title": "Does Saliency-Based Training bring Robustness for Deep Neural Networks\n  in Image Classification?", "abstract": "  Deep Neural Networks are powerful tools to understand complex patterns and\nmaking decisions. However, their black-box nature impedes a complete\nunderstanding of their inner workings. While online saliency-guided training\nmethods try to highlight the prominent features in the model's output to\nalleviate this problem, it is still ambiguous if the visually explainable\nfeatures align with robustness of the model against adversarial examples. In\nthis paper, we investigate the saliency trained model's vulnerability to\nadversarial examples methods. Models are trained using an online\nsaliency-guided training method and evaluated against popular algorithms of\nadversarial examples. We quantify the robustness and conclude that despite the\nwell-explained visualizations in the model's output, the salient models suffer\nfrom the lower performance against adversarial examples attacks.\n", "date": "2023-06-28", "categories": []}
{"id": "2306.16738", "title": "Towards Optimal Randomized Strategies in Adversarial Example Game", "abstract": "  The vulnerability of deep neural network models to adversarial example\nattacks is a practical challenge in many artificial intelligence applications.\nA recent line of work shows that the use of randomization in adversarial\ntraining is the key to find optimal strategies against adversarial example\nattacks. However, in a fully randomized setting where both the defender and the\nattacker can use randomized strategies, there are no efficient algorithm for\nfinding such an optimal strategy. To fill the gap, we propose the first\nalgorithm of its kind, called FRAT, which models the problem with a new\ninfinite-dimensional continuous-time flow on probability distribution spaces.\nFRAT maintains a lightweight mixture of models for the defender, with\nflexibility to efficiently update mixing weights and model parameters at each\niteration. Furthermore, FRAT utilizes lightweight sampling subroutines to\nconstruct a random strategy for the attacker. We prove that the continuous-time\nlimit of FRAT converges to a mixed Nash equilibria in a zero-sum game formed by\na defender and an attacker. Experimental results also demonstrate the\nefficiency of FRAT on CIFAR-10 and CIFAR-100 datasets.\n", "date": "2023-06-29", "categories": []}
{"id": "2306.15705", "title": "On the Universal Adversarial Perturbations for Efficient Data-free\n  Adversarial Detection", "abstract": "  Detecting adversarial samples that are carefully crafted to fool the model is\na critical step to socially-secure applications. However, existing adversarial\ndetection methods require access to sufficient training data, which brings\nnoteworthy concerns regarding privacy leakage and generalizability. In this\nwork, we validate that the adversarial sample generated by attack algorithms is\nstrongly related to a specific vector in the high-dimensional inputs. Such\nvectors, namely UAPs (Universal Adversarial Perturbations), can be calculated\nwithout original training data. Based on this discovery, we propose a\ndata-agnostic adversarial detection framework, which induces different\nresponses between normal and adversarial samples to UAPs. Experimental results\nshow that our method achieves competitive detection performance on various text\nclassification tasks, and maintains an equivalent time consumption to normal\ninference.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.15826", "title": "MAT: Mixed-Strategy Game of Adversarial Training in Fine-tuning", "abstract": "  Fine-tuning large-scale pre-trained language models has been demonstrated\neffective for various natural language processing (NLP) tasks. Previous studies\nhave established that incorporating adversarial training during the fine-tuning\nstage can significantly enhance model generalization and robustness. However,\nfrom the perspective of game theory, such utilizations of adversarial training\ncorrespond to pure-strategy games, which are inherently limited in terms of the\nscope of their strategies, thereby still having room for improvement. In order\nto push the performance boundaries, we propose a novel Mixed-strategy\nAdversarial Training algorithm (MAT). Methodologically, we derive the Nash\nequilibrium of a mixed-strategy game for adversarial training using Entropy\nMirror Descent to establish MAT by sampling method. To verify the effectiveness\nof MAT, we conducted extensive benchmark experiments on large-scale pre-trained\nmodels, such as BERT and RoBERTa. MAT significantly outperforms the\nstate-of-the-art methods on both the GLUE and ANLI benchmarks in terms of\ngeneralization and robustness.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.16131", "title": "Distributional Modeling for Location-Aware Adversarial Patches", "abstract": "  Adversarial patch is one of the important forms of performing adversarial\nattacks in the physical world. To improve the naturalness and aggressiveness of\nexisting adversarial patches, location-aware patches are proposed, where the\npatch's location on the target object is integrated into the optimization\nprocess to perform attacks. Although it is effective, efficiently finding the\noptimal location for placing the patches is challenging, especially under the\nblack-box attack settings. In this paper, we propose the Distribution-Optimized\nAdversarial Patch (DOPatch), a novel method that optimizes a multimodal\ndistribution of adversarial locations instead of individual ones. DOPatch has\nseveral benefits: Firstly, we find that the locations' distributions across\ndifferent models are pretty similar, and thus we can achieve efficient\nquery-based attacks to unseen models using a distributional prior optimized on\na surrogate model. Secondly, DOPatch can generate diverse adversarial samples\nby characterizing the distribution of adversarial locations. Thus we can\nimprove the model's robustness to location-aware patches via carefully designed\nDistributional-Modeling Adversarial Training (DOP-DMAT). We evaluate DOPatch on\nvarious face recognition and image recognition tasks and demonstrate its\nsuperiority and efficiency over existing methods. We also conduct extensive\nablation studies and analyses to validate the effectiveness of our method and\nprovide insights into the distribution of adversarial locations.\n", "date": "2023-06-28", "categories": []}
{"id": "2306.15164", "title": "DSRM: Boost Textual Adversarial Training with Distribution Shift Risk\n  Minimization", "abstract": "  Adversarial training is one of the best-performing methods in improving the\nrobustness of deep language models. However, robust models come at the cost of\nhigh time consumption, as they require multi-step gradient ascents or word\nsubstitutions to obtain adversarial samples. In addition, these generated\nsamples are deficient in grammatical quality and semantic consistency, which\nimpairs the effectiveness of adversarial training. To address these problems,\nwe introduce a novel, effective procedure for instead adversarial training with\nonly clean data. Our procedure, distribution shift risk minimization (DSRM),\nestimates the adversarial loss by perturbing the input data's probability\ndistribution rather than their embeddings. This formulation results in a robust\nmodel that minimizes the expected global loss under adversarial attacks. Our\napproach requires zero adversarial samples for training and reduces time\nconsumption by up to 70\\% compared to current best-performing adversarial\ntraining methods. Experiments demonstrate that DSRM considerably improves\nBERT's resistance to textual adversarial attacks and achieves state-of-the-art\nrobust accuracy on various benchmarks.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.15221", "title": "[Re] Double Sampling Randomized Smoothing", "abstract": "  This paper is a contribution to the reproducibility challenge in the field of\nmachine learning, specifically addressing the issue of certifying the\nrobustness of neural networks (NNs) against adversarial perturbations. The\nproposed Double Sampling Randomized Smoothing (DSRS) framework overcomes the\nlimitations of existing methods by using an additional smoothing distribution\nto improve the robustness certification. The paper provides a clear\nmanifestation of DSRS for a generalized family of Gaussian smoothing and a\ncomputationally efficient method for implementation. The experiments on MNIST\nand CIFAR-10 demonstrate the effectiveness of DSRS, consistently certifying\nlarger robust radii compared to other methods. Also various ablations studies\nare conducted to further analyze the hyperparameters and effect of adversarial\ntraining methods on the certified radius by the proposed framework.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.15363", "title": "Your Attack Is Too DUMB: Formalizing Attacker Scenarios for Adversarial\n  Transferability", "abstract": "  Evasion attacks are a threat to machine learning models, where adversaries\nattempt to affect classifiers by injecting malicious samples. An alarming\nside-effect of evasion attacks is their ability to transfer among different\nmodels: this property is called transferability. Therefore, an attacker can\nproduce adversarial samples on a custom model (surrogate) to conduct the attack\non a victim's organization later. Although literature widely discusses how\nadversaries can transfer their attacks, their experimental settings are limited\nand far from reality. For instance, many experiments consider both attacker and\ndefender sharing the same dataset, balance level (i.e., how the ground truth is\ndistributed), and model architecture.\n  In this work, we propose the DUMB attacker model. This framework allows\nanalyzing if evasion attacks fail to transfer when the training conditions of\nsurrogate and victim models differ. DUMB considers the following conditions:\nDataset soUrces, Model architecture, and the Balance of the ground truth. We\nthen propose a novel testbed to evaluate many state-of-the-art evasion attacks\nwith DUMB; the testbed consists of three computer vision tasks with two\ndistinct datasets each, four types of balance levels, and three model\narchitectures. Our analysis, which generated 13K tests over 14 distinct\nattacks, led to numerous novel findings in the scope of transferable attacks\nwith surrogate models. In particular, mismatches between attackers and victims\nin terms of dataset source, balance levels, and model architecture lead to\nnon-negligible loss of attack performance.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.15451", "title": "Advancing Adversarial Training by Injecting Booster Signal", "abstract": "  Recent works have demonstrated that deep neural networks (DNNs) are highly\nvulnerable to adversarial attacks. To defend against adversarial attacks, many\ndefense strategies have been proposed, among which adversarial training has\nbeen demonstrated to be the most effective strategy. However, it has been known\nthat adversarial training sometimes hurts natural accuracy. Then, many works\nfocus on optimizing model parameters to handle the problem. Different from the\nprevious approaches, in this paper, we propose a new approach to improve the\nadversarial robustness by using an external signal rather than model\nparameters. In the proposed method, a well-optimized universal external signal\ncalled a booster signal is injected into the outside of the image which does\nnot overlap with the original content. Then, it boosts both adversarial\nrobustness and natural accuracy. The booster signal is optimized in parallel to\nmodel parameters step by step collaboratively. Experimental results show that\nthe booster signal can improve both the natural and robust accuracies over the\nrecent state-of-the-art adversarial training methods. Also, optimizing the\nbooster signal is general and flexible enough to be adopted on any existing\nadversarial training methods.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.15457", "title": "Robust Proxy: Improving Adversarial Robustness by Robust Proxy Learning", "abstract": "  Recently, it has been widely known that deep neural networks are highly\nvulnerable and easily broken by adversarial attacks. To mitigate the\nadversarial vulnerability, many defense algorithms have been proposed.\nRecently, to improve adversarial robustness, many works try to enhance feature\nrepresentation by imposing more direct supervision on the discriminative\nfeature. However, existing approaches lack an understanding of learning\nadversarially robust feature representation. In this paper, we propose a novel\ntraining framework called Robust Proxy Learning. In the proposed method, the\nmodel explicitly learns robust feature representations with robust proxies. To\nthis end, firstly, we demonstrate that we can generate class-representative\nrobust features by adding class-wise robust perturbations. Then, we use the\nclass representative features as robust proxies. With the class-wise robust\nfeatures, the model explicitly learns adversarially robust features through the\nproposed robust proxy learning framework. Through extensive experiments, we\nverify that we can manually generate robust features, and our proposed learning\nframework could increase the robustness of the DNNs.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.15482", "title": "Cooperation or Competition: Avoiding Player Domination for Multi-Target\n  Robustness via Adaptive Budgets", "abstract": "  Despite incredible advances, deep learning has been shown to be susceptible\nto adversarial attacks. Numerous approaches have been proposed to train robust\nnetworks both empirically and certifiably. However, most of them defend against\nonly a single type of attack, while recent work takes steps forward in\ndefending against multiple attacks. In this paper, to understand multi-target\nrobustness, we view this problem as a bargaining game in which different\nplayers (adversaries) negotiate to reach an agreement on a joint direction of\nparameter updating. We identify a phenomenon named player domination in the\nbargaining game, namely that the existing max-based approaches, such as MAX and\nMSD, do not converge. Based on our theoretical analysis, we design a novel\nframework that adjusts the budgets of different adversaries to avoid any player\ndominance. Experiments on standard benchmarks show that employing the proposed\nframework to the existing approaches significantly advances multi-target\nrobustness.\n", "date": "2023-06-27", "categories": []}
{"id": "2210.03150", "title": "Towards Out-of-Distribution Adversarial Robustness", "abstract": "  Adversarial robustness continues to be a major challenge for deep learning. A\ncore issue is that robustness to one type of attack often fails to transfer to\nother attacks. While prior work establishes a theoretical trade-off in\nrobustness against different $L_p$ norms, we show that there is potential for\nimprovement against many commonly used attacks by adopting a domain\ngeneralisation approach. Concretely, we treat each type of attack as a domain,\nand apply the Risk Extrapolation method (REx), which promotes similar levels of\nrobustness against all training attacks. Compared to existing methods, we\nobtain similar or superior worst-case adversarial robustness on attacks seen\nduring training. Moreover, we achieve superior performance on families or\ntunings of attacks only encountered at test time. On ensembles of attacks, our\napproach improves the accuracy from 3.4% with the best existing baseline to\n25.9% on MNIST, and from 16.9% to 23.5% on CIFAR10.\n", "date": "2022-10-06", "categories": []}
{"id": "2211.12681", "title": "Benchmarking Adversarially Robust Quantum Machine Learning at Scale", "abstract": "  Machine learning (ML) methods such as artificial neural networks are rapidly\nbecoming ubiquitous in modern science, technology and industry. Despite their\naccuracy and sophistication, neural networks can be easily fooled by carefully\ndesigned malicious inputs known as adversarial attacks. While such\nvulnerabilities remain a serious challenge for classical neural networks, the\nextent of their existence is not fully understood in the quantum ML setting. In\nthis work, we benchmark the robustness of quantum ML networks, such as quantum\nvariational classifiers (QVC), at scale by performing rigorous training for\nboth simple and complex image datasets and through a variety of high-end\nadversarial attacks. Our results show that QVCs offer a notably enhanced\nrobustness against classical adversarial attacks by learning features which are\nnot detected by the classical neural networks, indicating a possible quantum\nadvantage for ML tasks. Contrarily, and remarkably, the converse is not true,\nwith attacks on quantum networks also capable of deceiving classical neural\nnetworks. By combining quantum and classical network outcomes, we propose a\nnovel adversarial attack detection technology. Traditionally quantum advantage\nin ML systems has been sought through increased accuracy or algorithmic\nspeed-up, but our work has revealed the potential for a new kind of quantum\nadvantage through superior robustness of ML models, whose practical realisation\nwill address serious security concerns and reliability issues of ML algorithms\nemployed in a myriad of applications including autonomous vehicles,\ncybersecurity, and surveillance robotic systems.\n", "date": "2022-11-23", "categories": []}
{"id": "2301.12896", "title": "Identifying Adversarially Attackable and Robust Samples", "abstract": "  Adversarial attacks insert small, imperceptible perturbations to input\nsamples that cause large, undesired changes to the output of deep learning\nmodels. Despite extensive research on generating adversarial attacks and\nbuilding defense systems, there has been limited research on understanding\nadversarial attacks from an input-data perspective. This work introduces the\nnotion of sample attackability, where we aim to identify samples that are most\nsusceptible to adversarial attacks (attackable samples) and conversely also\nidentify the least susceptible samples (robust samples). We propose a\ndeep-learning-based detector to identify the adversarially attackable and\nrobust samples in an unseen dataset for an unseen target model. Experiments on\nstandard image classification datasets enables us to assess the portability of\nthe deep attackability detector across a range of architectures. We find that\nthe deep attackability detector performs better than simple model\nuncertainty-based measures for identifying the attackable/robust samples. This\nsuggests that uncertainty is an inadequate proxy for measuring sample distance\nto a decision boundary. In addition to better understanding adversarial attack\ntheory, it is found that the ability to identify the adversarially attackable\nand robust samples has implications for improving the efficiency of\nsample-selection tasks.\n", "date": "2023-01-30", "categories": []}
{"id": "2306.01613", "title": "Hyperparameter Learning under Data Poisoning: Analysis of the Influence\n  of Regularization via Multiobjective Bilevel Optimization", "abstract": "  Machine Learning (ML) algorithms are vulnerable to poisoning attacks, where a\nfraction of the training data is manipulated to deliberately degrade the\nalgorithms' performance. Optimal attacks can be formulated as bilevel\noptimization problems and help to assess their robustness in worst-case\nscenarios. We show that current approaches, which typically assume that\nhyperparameters remain constant, lead to an overly pessimistic view of the\nalgorithms' robustness and of the impact of regularization. We propose a novel\noptimal attack formulation that considers the effect of the attack on the\nhyperparameters and models the attack as a multiobjective bilevel optimization\nproblem. This allows to formulate optimal attacks, learn hyperparameters and\nevaluate robustness under worst-case conditions. We apply this attack\nformulation to several ML classifiers using $L_2$ and $L_1$ regularization. Our\nevaluation on multiple datasets confirms the limitations of previous strategies\nand evidences the benefits of using $L_2$ and $L_1$ regularization to dampen\nthe effect of poisoning attacks.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.13854", "title": "Similarity Preserving Adversarial Graph Contrastive Learning", "abstract": "  Recent works demonstrate that GNN models are vulnerable to adversarial\nattacks, which refer to imperceptible perturbation on the graph structure and\nnode features. Among various GNN models, graph contrastive learning (GCL) based\nmethods specifically suffer from adversarial attacks due to their inherent\ndesign that highly depends on the self-supervision signals derived from the\noriginal graph, which however already contains noise when the graph is\nattacked. To achieve adversarial robustness against such attacks, existing\nmethods adopt adversarial training (AT) to the GCL framework, which considers\nthe attacked graph as an augmentation under the GCL framework. However, we find\nthat existing adversarially trained GCL methods achieve robustness at the\nexpense of not being able to preserve the node feature similarity. In this\npaper, we propose a similarity-preserving adversarial graph contrastive\nlearning (SP-AGCL) framework that contrasts the clean graph with two auxiliary\nviews of different properties (i.e., the node similarity-preserving view and\nthe adversarial view). Extensive experiments demonstrate that SP-AGCL achieves\na competitive performance on several downstream tasks, and shows its\neffectiveness in various scenarios, e.g., a network with adversarial attacks,\nnoisy labels, and heterophilous neighbors. Our code is available at\nhttps://github.com/yeonjun-in/torch-SP-AGCL.\n", "date": "2023-06-24", "categories": []}
{"id": "2306.13965", "title": "Boosting Model Inversion Attacks with Adversarial Examples", "abstract": "  Model inversion attacks involve reconstructing the training data of a target\nmodel, which raises serious privacy concerns for machine learning models.\nHowever, these attacks, especially learning-based methods, are likely to suffer\nfrom low attack accuracy, i.e., low classification accuracy of these\nreconstructed data by machine learning classifiers. Recent studies showed an\nalternative strategy of model inversion attacks, GAN-based optimization, can\nimprove the attack accuracy effectively. However, these series of GAN-based\nattacks reconstruct only class-representative training data for a class,\nwhereas learning-based attacks can reconstruct diverse data for different\ntraining data in each class. Hence, in this paper, we propose a new training\nparadigm for a learning-based model inversion attack that can achieve higher\nattack accuracy in a black-box setting. First, we regularize the training\nprocess of the attack model with an added semantic loss function and, second,\nwe inject adversarial examples into the training data to increase the diversity\nof the class-related parts (i.e., the essential features for classification\ntasks) in training data. This scheme guides the attack model to pay more\nattention to the class-related parts of the original data during the data\nreconstruction process. The experimental results show that our method greatly\nboosts the performance of existing learning-based model inversion attacks. Even\nwhen no extra queries to the target model are allowed, the approach can still\nimprove the attack accuracy of reconstructed data. This new attack shows that\nthe severity of the threat from learning-based model inversion adversaries is\nunderestimated and more robust defenses are required.\n", "date": "2023-06-24", "categories": []}
{"id": "2306.14126", "title": "Robust Spatiotemporal Traffic Forecasting with Reinforced Dynamic\n  Adversarial Training", "abstract": "  Machine learning-based forecasting models are commonly used in Intelligent\nTransportation Systems (ITS) to predict traffic patterns and provide city-wide\nservices. However, most of the existing models are susceptible to adversarial\nattacks, which can lead to inaccurate predictions and negative consequences\nsuch as congestion and delays. Therefore, improving the adversarial robustness\nof these models is crucial for ITS. In this paper, we propose a novel framework\nfor incorporating adversarial training into spatiotemporal traffic forecasting\ntasks. We demonstrate that traditional adversarial training methods designated\nfor static domains cannot be directly applied to traffic forecasting tasks, as\nthey fail to effectively defend against dynamic adversarial attacks. Then, we\npropose a reinforcement learning-based method to learn the optimal node\nselection strategy for adversarial examples, which simultaneously strengthens\nthe dynamic attack defense capability and reduces the model overfitting.\nAdditionally, we introduce a self-knowledge distillation regularization module\nto overcome the \"forgetting issue\" caused by continuously changing adversarial\nnodes during training. We evaluate our approach on two real-world traffic\ndatasets and demonstrate its superiority over other baselines. Our method\neffectively enhances the adversarial robustness of spatiotemporal traffic\nforecasting models. The source code for our framework is available at\nhttps://github.com/usail-hkust/RDAT.\n", "date": "2023-06-25", "categories": []}
{"id": "2306.14217", "title": "On Evaluating the Adversarial Robustness of Semantic Segmentation Models", "abstract": "  Achieving robustness against adversarial input perturbation is an important\nand intriguing problem in machine learning. In the area of semantic image\nsegmentation, a number of adversarial training approaches have been proposed as\na defense against adversarial perturbation, but the methodology of evaluating\nthe robustness of the models is still lacking, compared to image\nclassification. Here, we demonstrate that, just like in image classification,\nit is important to evaluate the models over several different and hard attacks.\nWe propose a set of gradient based iterative attacks and show that it is\nessential to perform a large number of iterations. We include attacks against\nthe internal representations of the models as well. We apply two types of\nattacks: maximizing the error with a bounded perturbation, and minimizing the\nperturbation for a given level of error. Using this set of attacks, we show for\nthe first time that a number of models in previous work that are claimed to be\nrobust are in fact not robust at all. We then evaluate simple adversarial\ntraining algorithms that produce reasonably robust models even under our set of\nstrong attacks. Our results indicate that a key design decision to achieve any\nrobustness is to use only adversarial examples during training. However, this\nintroduces a trade-off between robustness and accuracy.\n", "date": "2023-06-25", "categories": []}
{"id": "2306.14262", "title": "A Spectral Perspective towards Understanding and Improving Adversarial\n  Robustness", "abstract": "  Deep neural networks (DNNs) are incredibly vulnerable to crafted,\nimperceptible adversarial perturbations. While adversarial training (AT) has\nproven to be an effective defense approach, the AT mechanism for robustness\nimprovement is not fully understood. This work investigates AT from a spectral\nperspective, adding new insights to the design of effective defenses. In\nparticular, we show that AT induces the deep model to focus more on the\nlow-frequency region, which retains the shape-biased representations, to gain\nrobustness. Further, we find that the spectrum of a white-box attack is\nprimarily distributed in regions the model focuses on, and the perturbation\nattacks the spectral bands where the model is vulnerable. Based on this\nobservation, to train a model tolerant to frequency-varying perturbation, we\npropose a spectral alignment regularization (SAR) such that the spectral output\ninferred by an attacked adversarial input stays as close as possible to its\nnatural input counterpart. Experiments demonstrate that SAR and its weight\naveraging (WA) extension could significantly improve the robust accuracy by\n1.14% ~ 3.87% relative to the standard AT, across multiple datasets (CIFAR-10,\nCIFAR-100 and Tiny ImageNet), and various attacks (PGD, C&W and Autoattack),\nwithout any extra data.\n", "date": "2023-06-25", "categories": []}
{"id": "2306.14326", "title": "Computational Asymmetries in Robust Classification", "abstract": "  In the context of adversarial robustness, we make three strongly related\ncontributions. First, we prove that while attacking ReLU classifiers is\n$\\mathit{NP}$-hard, ensuring their robustness at training time is\n$\\Sigma^2_P$-hard (even on a single example). This asymmetry provides a\nrationale for the fact that robust classifications approaches are frequently\nfooled in the literature. Second, we show that inference-time robustness\ncertificates are not affected by this asymmetry, by introducing a\nproof-of-concept approach named Counter-Attack (CA). Indeed, CA displays a\nreversed asymmetry: running the defense is $\\mathit{NP}$-hard, while attacking\nit is $\\Sigma_2^P$-hard. Finally, motivated by our previous result, we argue\nthat adversarial attacks can be used in the context of robustness\ncertification, and provide an empirical evaluation of their effectiveness. As a\nbyproduct of this process, we also release UG100, a benchmark dataset for\nadversarial attacks.\n", "date": "2023-06-25", "categories": []}
{"id": "2306.14609", "title": "The race to robustness: exploiting fragile models for urban camouflage\n  and the imperative for machine learning security", "abstract": "  Adversarial Machine Learning (AML) represents the ability to disrupt Machine\nLearning (ML) algorithms through a range of methods that broadly exploit the\narchitecture of deep learning optimisation. This paper presents Distributed\nAdversarial Regions (DAR), a novel method that implements distributed\ninstantiations of computer vision-based AML attack methods that may be used to\ndisguise objects from image recognition in both white and black box settings.\nWe consider the context of object detection models used in urban environments,\nand benchmark the MobileNetV2, NasNetMobile and DenseNet169 models against a\nsubset of relevant images from the ImageNet dataset. We evaluate optimal\nparameters (size, number and perturbation method), and compare to\nstate-of-the-art AML techniques that perturb the entire image. We find that\nDARs can cause a reduction in confidence of 40.4% on average, but with the\nbenefit of not requiring the entire image, or the focal object, to be\nperturbed. The DAR method is a deliberately simple approach where the intention\nis to highlight how an adversary with very little skill could attack models\nthat may already be productionised, and to emphasise the fragility of\nfoundational object detection models. We present this as a contribution to the\nfield of ML security as well as AML. This paper contributes a novel adversarial\nmethod, an original comparison between DARs and other AML methods, and frames\nit in a new context - that of urban camouflage and the necessity for ML\nsecurity and model robustness.\n", "date": "2023-06-26", "categories": []}
{"id": "2306.14640", "title": "3D-Aware Adversarial Makeup Generation for Facial Privacy Protection", "abstract": "  The privacy and security of face data on social media are facing\nunprecedented challenges as it is vulnerable to unauthorized access and\nidentification. A common practice for solving this problem is to modify the\noriginal data so that it could be protected from being recognized by malicious\nface recognition (FR) systems. However, such ``adversarial examples'' obtained\nby existing methods usually suffer from low transferability and poor image\nquality, which severely limits the application of these methods in real-world\nscenarios. In this paper, we propose a 3D-Aware Adversarial Makeup Generation\nGAN (3DAM-GAN). which aims to improve the quality and transferability of\nsynthetic makeup for identity information concealing. Specifically, a UV-based\ngenerator consisting of a novel Makeup Adjustment Module (MAM) and Makeup\nTransfer Module (MTM) is designed to render realistic and robust makeup with\nthe aid of symmetric characteristics of human faces. Moreover, a makeup attack\nmechanism with an ensemble training strategy is proposed to boost the\ntransferability of black-box models. Extensive experiment results on several\nbenchmark datasets demonstrate that 3DAM-GAN could effectively protect faces\nagainst various FR models, including both publicly available state-of-the-art\nmodels and commercial face verification APIs, such as Face++, Baidu and Aliyun.\n", "date": "2023-06-26", "categories": []}
{"id": "2301.13838", "title": "Image Shortcut Squeezing: Countering Perturbative Availability Poisons\n  with Compression", "abstract": "  Perturbative availability poisons (PAPs) add small changes to images to\nprevent their use for model training. Current research adopts the belief that\npractical and effective approaches to countering PAPs do not exist. In this\npaper, we argue that it is time to abandon this belief. We present extensive\nexperiments showing that 12 state-of-the-art PAP methods are vulnerable to\nImage Shortcut Squeezing (ISS), which is based on simple compression. For\nexample, on average, ISS restores the CIFAR-10 model accuracy to $81.73\\%$,\nsurpassing the previous best preprocessing-based countermeasures by $37.97\\%$\nabsolute. ISS also (slightly) outperforms adversarial training and has higher\ngeneralizability to unseen perturbation norms and also higher efficiency. Our\ninvestigation reveals that the property of PAP perturbations depends on the\ntype of surrogate model used for poison generation, and it explains why a\nspecific ISS compression yields the best performance for a specific type of PAP\nperturbation. We further test stronger, adaptive poisoning, and show it falls\nshort of being an ideal defense against ISS. Overall, our results demonstrate\nthe importance of considering various (simple) countermeasures to ensure the\nmeaningfulness of analysis carried out during the development of PAP methods.\n", "date": "2023-01-31", "categories": []}
{"id": "2306.12043", "title": "Sample Attackability in Natural Language Adversarial Attacks", "abstract": "  Adversarial attack research in natural language processing (NLP) has made\nsignificant progress in designing powerful attack methods and defence\napproaches. However, few efforts have sought to identify which source samples\nare the most attackable or robust, i.e. can we determine for an unseen target\nmodel, which samples are the most vulnerable to an adversarial attack. This\nwork formally extends the definition of sample attackability/robustness for NLP\nattacks. Experiments on two popular NLP datasets, four state of the art models\nand four different NLP adversarial attack methods, demonstrate that sample\nuncertainty is insufficient for describing characteristics of attackable/robust\nsamples and hence a deep learning based detector can perform much better at\nidentifying the most attackable and robust samples for an unseen target model.\nNevertheless, further analysis finds that there is little agreement in which\nsamples are considered the most attackable/robust across different NLP attack\nmethods, explaining a lack of portability of attackability detection methods\nacross attack methods.\n", "date": "2023-06-21", "categories": []}
{"id": "2306.13614", "title": "Adversarial Robustness Certification for Bayesian Neural Networks", "abstract": "  We study the problem of certifying the robustness of Bayesian neural networks\n(BNNs) to adversarial input perturbations. Given a compact set of input points\n$T \\subseteq \\mathbb{R}^m$ and a set of output points $S \\subseteq\n\\mathbb{R}^n$, we define two notions of robustness for BNNs in an adversarial\nsetting: probabilistic robustness and decision robustness. Probabilistic\nrobustness is the probability that for all points in $T$ the output of a BNN\nsampled from the posterior is in $S$. On the other hand, decision robustness\nconsiders the optimal decision of a BNN and checks if for all points in $T$ the\noptimal decision of the BNN for a given loss function lies within the output\nset $S$. Although exact computation of these robustness properties is\nchallenging due to the probabilistic and non-convex nature of BNNs, we present\na unified computational framework for efficiently and formally bounding them.\nOur approach is based on weight interval sampling, integration, and bound\npropagation techniques, and can be applied to BNNs with a large number of\nparameters, and independently of the (approximate) inference method employed to\ntrain the BNN. We evaluate the effectiveness of our methods on various\nregression and classification tasks, including an industrial regression\nbenchmark, MNIST, traffic sign recognition, and airborne collision avoidance,\nand demonstrate that our approach enables certification of robustness and\nuncertainty of BNN predictions.\n", "date": "2023-06-23", "categories": []}
{"id": "2305.02383", "title": "On the Security Risks of Knowledge Graph Reasoning", "abstract": "  Knowledge graph reasoning (KGR) -- answering complex logical queries over\nlarge knowledge graphs -- represents an important artificial intelligence task,\nentailing a range of applications (e.g., cyber threat hunting). However,\ndespite its surging popularity, the potential security risks of KGR are largely\nunexplored, which is concerning, given the increasing use of such capability in\nsecurity-critical domains.\n  This work represents a solid initial step towards bridging the striking gap.\nWe systematize the security threats to KGR according to the adversary's\nobjectives, knowledge, and attack vectors. Further, we present ROAR, a new\nclass of attacks that instantiate a variety of such threats. Through empirical\nevaluation in representative use cases (e.g., medical decision support, cyber\nthreat hunting, and commonsense reasoning), we demonstrate that ROAR is highly\neffective to mislead KGR to suggest pre-defined answers for target queries, yet\nwith negligible impact on non-target ones. Finally, we explore potential\ncountermeasures against ROAR, including filtering of potentially poisoning\nknowledge and training with adversarially augmented queries, which leads to\nseveral promising research directions.\n", "date": "2023-05-03", "categories": []}
{"id": "2306.12582", "title": "Adversarial Training with Generated Data in High-Dimensional Regression:\n  An Asymptotic Study", "abstract": "  In recent years, studies such as\n\\cite{carmon2019unlabeled,gowal2021improving,xing2022artificial} have\ndemonstrated that incorporating additional real or generated data with\npseudo-labels can enhance adversarial training through a two-stage training\napproach. In this paper, we perform a theoretical analysis of the asymptotic\nbehavior of this method in high-dimensional linear regression. While a\ndouble-descent phenomenon can be observed in ridgeless training, with an\nappropriate $\\mathcal{L}_2$ regularization, the two-stage adversarial training\nachieves a better performance. Finally, we derive a shortcut cross-validation\nformula specifically tailored for the two-stage training method.\n", "date": "2023-06-21", "categories": []}
{"id": "2306.12610", "title": "Revisiting Image Classifier Training for Improved Certified Robust\n  Defense against Adversarial Patches", "abstract": "  Certifiably robust defenses against adversarial patches for image classifiers\nensure correct prediction against any changes to a constrained neighborhood of\npixels. PatchCleanser arXiv:2108.09135 [cs.CV], the state-of-the-art certified\ndefense, uses a double-masking strategy for robust classification. The success\nof this strategy relies heavily on the model's invariance to image pixel\nmasking. In this paper, we take a closer look at model training schemes to\nimprove this invariance. Instead of using Random Cutout arXiv:1708.04552v2\n[cs.CV] augmentations like PatchCleanser, we introduce the notion of worst-case\nmasking, i.e., selecting masked images which maximize classification loss.\nHowever, finding worst-case masks requires an exhaustive search, which might be\nprohibitively expensive to do on-the-fly during training. To solve this\nproblem, we propose a two-round greedy masking strategy (Greedy Cutout) which\nfinds an approximate worst-case mask location with much less compute. We show\nthat the models trained with our Greedy Cutout improves certified robust\naccuracy over Random Cutout in PatchCleanser across a range of datasets and\narchitectures. Certified robust accuracy on ImageNet with a ViT-B16-224 model\nincreases from 58.1\\% to 62.3\\% against a 3\\% square patch applied anywhere on\nthe image.\n", "date": "2023-06-22", "categories": []}
{"id": "2306.12688", "title": "Towards quantum enhanced adversarial robustness in machine learning", "abstract": "  Machine learning algorithms are powerful tools for data driven tasks such as\nimage classification and feature detection, however their vulnerability to\nadversarial examples - input samples manipulated to fool the algorithm -\nremains a serious challenge. The integration of machine learning with quantum\ncomputing has the potential to yield tools offering not only better accuracy\nand computational efficiency, but also superior robustness against adversarial\nattacks. Indeed, recent work has employed quantum mechanical phenomena to\ndefend against adversarial attacks, spurring the rapid development of the field\nof quantum adversarial machine learning (QAML) and potentially yielding a new\nsource of quantum advantage. Despite promising early results, there remain\nchallenges towards building robust real-world QAML tools. In this review we\ndiscuss recent progress in QAML and identify key challenges. We also suggest\nfuture research directions which could determine the route to practicality for\nQAML approaches as quantum computing hardware scales up and noise levels are\nreduced.\n", "date": "2023-06-22", "categories": []}
{"id": "2306.13091", "title": "Evading Forensic Classifiers with Attribute-Conditioned Adversarial\n  Faces", "abstract": "  The ability of generative models to produce highly realistic synthetic face\nimages has raised security and ethical concerns. As a first line of defense\nagainst such fake faces, deep learning based forensic classifiers have been\ndeveloped. While these forensic models can detect whether a face image is\nsynthetic or real with high accuracy, they are also vulnerable to adversarial\nattacks. Although such attacks can be highly successful in evading detection by\nforensic classifiers, they introduce visible noise patterns that are detectable\nthrough careful human scrutiny. Additionally, these attacks assume access to\nthe target model(s) which may not always be true. Attempts have been made to\ndirectly perturb the latent space of GANs to produce adversarial fake faces\nthat can circumvent forensic classifiers. In this work, we go one step further\nand show that it is possible to successfully generate adversarial fake faces\nwith a specified set of attributes (e.g., hair color, eye size, race, gender,\netc.). To achieve this goal, we leverage the state-of-the-art generative model\nStyleGAN with disentangled representations, which enables a range of\nmodifications without leaving the manifold of natural images. We propose a\nframework to search for adversarial latent codes within the feature space of\nStyleGAN, where the search can be guided either by a text prompt or a reference\nimage. We also propose a meta-learning based optimization strategy to achieve\ntransferable performance on unknown target models. Extensive experiments\ndemonstrate that the proposed approach can produce semantically manipulated\nadversarial fake faces, which are true to the specified attribute set and can\nsuccessfully fool forensic face classifiers, while remaining undetectable by\nhumans. Code: https://github.com/koushiksrivats/face_attribute_attack.\n", "date": "2023-06-22", "categories": []}
{"id": "2306.11066", "title": "Adversarial Robustness of Prompt-based Few-Shot Learning for Natural\n  Language Understanding", "abstract": "  State-of-the-art few-shot learning (FSL) methods leverage prompt-based\nfine-tuning to obtain remarkable results for natural language understanding\n(NLU) tasks. While much of the prior FSL methods focus on improving downstream\ntask performance, there is a limited understanding of the adversarial\nrobustness of such methods. In this work, we conduct an extensive study of\nseveral state-of-the-art FSL methods to assess their robustness to adversarial\nperturbations. To better understand the impact of various factors towards\nrobustness (or the lack of it), we evaluate prompt-based FSL methods against\nfully fine-tuned models for aspects such as the use of unlabeled data, multiple\nprompts, number of few-shot examples, model size and type. Our results on six\nGLUE tasks indicate that compared to fully fine-tuned models, vanilla FSL\nmethods lead to a notable relative drop in task performance (i.e., are less\nrobust) in the face of adversarial perturbations. However, using (i) unlabeled\ndata for prompt-based FSL and (ii) multiple prompts flip the trend. We further\ndemonstrate that increasing the number of few-shot examples and model size lead\nto increased adversarial robustness of vanilla FSL methods. Broadly, our work\nsheds light on the adversarial robustness evaluation of prompt-based FSL\nmethods for NLU tasks.\n", "date": "2023-06-19", "categories": []}
{"id": "2306.12161", "title": "Adversarial Attacks Neutralization via Data Set Randomization", "abstract": "  Adversarial attacks on deep-learning models pose a serious threat to their\nreliability and security. Existing defense mechanisms are narrow addressing a\nspecific type of attack or being vulnerable to sophisticated attacks. We\npropose a new defense mechanism that, while being focused on image-based\nclassifiers, is general with respect to the cited category. It is rooted on\nhyperspace projection. In particular, our solution provides a pseudo-random\nprojection of the original dataset into a new dataset. The proposed defense\nmechanism creates a set of diverse projected datasets, where each projected\ndataset is used to train a specific classifier, resulting in different trained\nclassifiers with different decision boundaries. During testing, it randomly\nselects a classifier to test the input. Our approach does not sacrifice\naccuracy over legitimate input. Other than detailing and providing a thorough\ncharacterization of our defense mechanism, we also provide a proof of concept\nof using four optimization-based adversarial attacks (PGD, FGSM, IGSM, and\nC\\&W) and a generative adversarial attack testing them on the MNIST dataset.\nOur experimental results show that our solution increases the robustness of\ndeep learning models against adversarial attacks and significantly reduces the\nattack success rate by at least 89% for optimization attacks and 78% for\ngenerative attacks. We also analyze the relationship between the number of used\nhyperspaces and the efficacy of the defense mechanism. As expected, the two are\npositively correlated, offering an easy-to-tune parameter to enforce the\ndesired level of security. The generality and scalability of our solution and\nadaptability to different attack scenarios, combined with the excellent\nachieved results, other than providing a robust defense against adversarial\nattacks on deep learning networks, also lay the groundwork for future research\nin the field.\n", "date": "2023-06-21", "categories": []}
{"id": "2306.12377", "title": "Geometric Algorithms for $k$-NN Poisoning", "abstract": "  We propose a label poisoning attack on geometric data sets against\n$k$-nearest neighbor classification. We provide an algorithm that can compute\nan $\\varepsilon n$-additive approximation of the optimal poisoning in $n\\cdot\n2^{2^{O(d+k/\\varepsilon)}}$ time for a given data set $X \\in \\mathbb{R}^d$,\nwhere $|X| = n$. Our algorithm achieves its objectives through the application\nof multi-scale random partitions.\n", "date": "2023-06-21", "categories": []}
{"id": "2206.02670", "title": "Robust Adversarial Attacks Detection based on Explainable Deep\n  Reinforcement Learning For UAV Guidance and Planning", "abstract": "  The dangers of adversarial attacks on Uncrewed Aerial Vehicle (UAV) agents\noperating in public are increasing. Adopting AI-based techniques and, more\nspecifically, Deep Learning (DL) approaches to control and guide these UAVs can\nbe beneficial in terms of performance but can add concerns regarding the safety\nof those techniques and their vulnerability against adversarial attacks.\nConfusion in the agent's decision-making process caused by these attacks can\nseriously affect the safety of the UAV. This paper proposes an innovative\napproach based on the explainability of DL methods to build an efficient\ndetector that will protect these DL schemes and the UAVs adopting them from\nattacks. The agent adopts a Deep Reinforcement Learning (DRL) scheme for\nguidance and planning. The agent is trained with a Deep Deterministic Policy\nGradient (DDPG) with Prioritised Experience Replay (PER) DRL scheme that\nutilises Artificial Potential Field (APF) to improve training times and\nobstacle avoidance performance. A simulated environment for UAV explainable\nDRL-based planning and guidance, including obstacles and adversarial attacks,\nis built. The adversarial attacks are generated by the Basic Iterative Method\n(BIM) algorithm and reduced obstacle course completion rates from 97\\% to 35\\%.\nTwo adversarial attack detectors are proposed to counter this reduction. The\nfirst one is a Convolutional Neural Network Adversarial Detector (CNN-AD),\nwhich achieves accuracy in the detection of 80\\%. The second detector utilises\na Long Short Term Memory (LSTM) network. It achieves an accuracy of 91\\% with\nfaster computing times compared to the CNN-AD, allowing for real-time\nadversarial detection.\n", "date": "2022-06-06", "categories": []}
{"id": "2306.10742", "title": "BNN-DP: Robustness Certification of Bayesian Neural Networks via Dynamic\n  Programming", "abstract": "  In this paper, we introduce BNN-DP, an efficient algorithmic framework for\nanalysis of adversarial robustness of Bayesian Neural Networks (BNNs). Given a\ncompact set of input points $T\\subset \\mathbb{R}^n$, BNN-DP computes lower and\nupper bounds on the BNN's predictions for all the points in $T$. The framework\nis based on an interpretation of BNNs as stochastic dynamical systems, which\nenables the use of Dynamic Programming (DP) algorithms to bound the prediction\nrange along the layers of the network. Specifically, the method uses bound\npropagation techniques and convex relaxations to derive a backward recursion\nprocedure to over-approximate the prediction range of the BNN with piecewise\naffine functions. The algorithm is general and can handle both regression and\nclassification tasks. On a set of experiments on various regression and\nclassification tasks and BNN architectures, we show that BNN-DP outperforms\nstate-of-the-art methods by up to four orders of magnitude in both tightness of\nthe bounds and computational efficiency.\n", "date": "2023-06-19", "categories": []}
{"id": "2306.11261", "title": "Comparative Evaluation of Recent Universal Adversarial Perturbations in\n  Image Classification", "abstract": "  The vulnerability of Convolutional Neural Networks (CNNs) to adversarial\nsamples has recently garnered significant attention in the machine learning\ncommunity. Furthermore, recent studies have unveiled the existence of universal\nadversarial perturbations (UAPs) that are image-agnostic and highly\ntransferable across different CNN models. In this survey, our primary focus\nrevolves around the recent advancements in UAPs specifically within the image\nclassification task. We categorize UAPs into two distinct categories, i.e.,\nnoise-based attacks and generator-based attacks, thereby providing a\ncomprehensive overview of representative methods within each category. By\npresenting the computational details of these methods, we summarize various\nloss functions employed for learning UAPs. Furthermore, we conduct a\ncomprehensive evaluation of different loss functions within consistent training\nframeworks, including noise-based and generator-based. The evaluation covers a\nwide range of attack settings, including black-box and white-box attacks,\ntargeted and untargeted attacks, as well as the examination of defense\nmechanisms.\n  Our quantitative evaluation results yield several important findings\npertaining to the effectiveness of different loss functions, the selection of\nsurrogate CNN models, the impact of training data and data size, and the\ntraining frameworks involved in crafting universal attackers. Finally, to\nfurther promote future research on universal adversarial attacks, we provide\nsome visualizations of the perturbations and discuss the potential research\ndirections.\n", "date": "2023-06-20", "categories": []}
{"id": "2209.03755", "title": "Fact-Saboteurs: A Taxonomy of Evidence Manipulation Attacks against\n  Fact-Verification Systems", "abstract": "  Mis- and disinformation are a substantial global threat to our security and\nsafety. To cope with the scale of online misinformation, researchers have been\nworking on automating fact-checking by retrieving and verifying against\nrelevant evidence. However, despite many advances, a comprehensive evaluation\nof the possible attack vectors against such systems is still lacking.\nParticularly, the automated fact-verification process might be vulnerable to\nthe exact disinformation campaigns it is trying to combat. In this work, we\nassume an adversary that automatically tampers with the online evidence in\norder to disrupt the fact-checking model via camouflaging the relevant evidence\nor planting a misleading one. We first propose an exploratory taxonomy that\nspans these two targets and the different threat model dimensions. Guided by\nthis, we design and propose several potential attack methods. We show that it\nis possible to subtly modify claim-salient snippets in the evidence and\ngenerate diverse and claim-aligned evidence. Thus, we highly degrade the\nfact-checking performance under many different permutations of the taxonomy's\ndimensions. The attacks are also robust against post-hoc modifications of the\nclaim. Our analysis further hints at potential limitations in models' inference\nwhen faced with contradicting evidence. We emphasize that these attacks can\nhave harmful implications on the inspectable and human-in-the-loop usage\nscenarios of such models, and we conclude by discussing challenges and\ndirections for future defenses.\n", "date": "2022-09-07", "categories": []}
{"id": "2306.07567", "title": "Large Language Models Sometimes Generate Purely Negatively-Reinforced\n  Text", "abstract": "  When using adversarial training, it is common practice to train against the\nmost egregious failures. However, this might imply using examples with\nsensitive information (such as leaked passwords or security vulnerabilities) as\ntraining data. One might assume that language models trained with gradient\ndescent never generate text snippets which were only present in examples\nassociated with the lowest possible reward. In this paper, we show that this\nassumption is wrong: in some situations, large language models do learn from\nsuch negatively-reinforced examples. We present a specific training setup that\nenables Pythia-160M to guess passwords 13% more often than it would by guessing\nrandomly, despite only showing it these passwords on examples where the model\nis incentivized to not output these passwords. Our code is available at\nwww.github.com/FabienRoger/Learning-From-Negative-Examples\n", "date": "2023-06-13", "categories": []}
{"id": "2306.09844", "title": "Wasserstein distributional robustness of neural networks", "abstract": "  Deep neural networks are known to be vulnerable to adversarial attacks (AA).\nFor an image recognition task, this means that a small perturbation of the\noriginal can result in the image being misclassified. Design of such attacks as\nwell as methods of adversarial training against them are subject of intense\nresearch. We re-cast the problem using techniques of Wasserstein\ndistributionally robust optimization (DRO) and obtain novel contributions\nleveraging recent insights from DRO sensitivity analysis. We consider a set of\ndistributional threat models. Unlike the traditional pointwise attacks, which\nassume a uniform bound on perturbation of each input data point, distributional\nthreat models allow attackers to perturb inputs in a non-uniform way. We link\nthese more general attacks with questions of out-of-sample performance and\nKnightian uncertainty. To evaluate the distributional robustness of neural\nnetworks, we propose a first-order AA algorithm and its multi-step version. Our\nattack algorithms include Fast Gradient Sign Method (FGSM) and Projected\nGradient Descent (PGD) as special cases. Furthermore, we provide a new\nasymptotic estimate of the adversarial accuracy against distributional threat\nmodels. The bound is fast to compute and first-order accurate, offering new\ninsights even for the pointwise AA. It also naturally yields out-of-sample\nperformance guarantees. We conduct numerical experiments on the CIFAR-10\ndataset using DNNs on RobustBench to illustrate our theoretical results. Our\ncode is available at https://github.com/JanObloj/W-DRO-Adversarial-Methods.\n", "date": "2023-06-16", "categories": []}
{"id": "2306.09949", "title": "Towards Better Certified Segmentation via Diffusion Models", "abstract": "  The robustness of image segmentation has been an important research topic in\nthe past few years as segmentation models have reached production-level\naccuracy. However, like classification models, segmentation models can be\nvulnerable to adversarial perturbations, which hinders their use in\ncritical-decision systems like healthcare or autonomous driving. Recently,\nrandomized smoothing has been proposed to certify segmentation predictions by\nadding Gaussian noise to the input to obtain theoretical guarantees. However,\nthis method exhibits a trade-off between the amount of added noise and the\nlevel of certification achieved. In this paper, we address the problem of\ncertifying segmentation prediction using a combination of randomized smoothing\nand diffusion models. Our experiments show that combining randomized smoothing\nand diffusion models significantly improves certified robustness, with results\nindicating a mean improvement of 21 points in accuracy compared to previous\nstate-of-the-art methods on Pascal-Context and Cityscapes public datasets. Our\nmethod is independent of the selected segmentation model and does not need any\nadditional specialized training procedure.\n", "date": "2023-06-16", "categories": []}
{"id": "2306.09951", "title": "You Don't Need Robust Machine Learning to Manage Adversarial Attack\n  Risks", "abstract": "  The robustness of modern machine learning (ML) models has become an\nincreasing concern within the community. The ability to subvert a model into\nmaking errant predictions using seemingly inconsequential changes to input is\nstartling, as is our lack of success in building models robust to this concern.\nExisting research shows progress, but current mitigations come with a high cost\nand simultaneously reduce the model's accuracy. However, such trade-offs may\nnot be necessary when other design choices could subvert the risk. In this\nsurvey we review the current literature on attacks and their real-world\noccurrences, or limited evidence thereof, to critically evaluate the real-world\nrisks of adversarial machine learning (AML) for the average entity. This is\ndone with an eye toward how one would then mitigate these attacks in practice,\nthe risks for production deployment, and how those risks could be managed. In\ndoing so we elucidate that many AML threats do not warrant the cost and\ntrade-offs of robustness due to a low likelihood of attack or availability of\nsuperior non-ML mitigations. Our analysis also recommends cases where an actor\nshould be concerned about AML to the degree where robust ML models are\nnecessary for a complete deployment.\n", "date": "2023-06-16", "categories": []}
{"id": "2211.17244", "title": "Tight Certification of Adversarially Trained Neural Networks via\n  Nonconvex Low-Rank Semidefinite Relaxations", "abstract": "  Adversarial training is well-known to produce high-quality neural network\nmodels that are empirically robust against adversarial perturbations.\nNevertheless, once a model has been adversarially trained, one often desires a\ncertification that the model is truly robust against all future attacks.\nUnfortunately, when faced with adversarially trained models, all existing\napproaches have significant trouble making certifications that are strong\nenough to be practically useful. Linear programming (LP) techniques in\nparticular face a \"convex relaxation barrier\" that prevent them from making\nhigh-quality certifications, even after refinement with mixed-integer linear\nprogramming (MILP) and branch-and-bound (BnB) techniques. In this paper, we\npropose a nonconvex certification technique, based on a low-rank restriction of\na semidefinite programming (SDP) relaxation. The nonconvex relaxation makes\nstrong certifications comparable to much more expensive SDP methods, while\noptimizing over dramatically fewer variables comparable to much weaker LP\nmethods. Despite nonconvexity, we show how off-the-shelf local optimization\nalgorithms can be used to achieve and to certify global optimality in\npolynomial time. Our experiments find that the nonconvex relaxation almost\ncompletely closes the gap towards exact certification of adversarially trained\nmodels.\n", "date": "2022-11-30", "categories": []}
{"id": "2306.08009", "title": "DHBE: Data-free Holistic Backdoor Erasing in Deep Neural Networks via\n  Restricted Adversarial Distillation", "abstract": "  Backdoor attacks have emerged as an urgent threat to Deep Neural Networks\n(DNNs), where victim DNNs are furtively implanted with malicious neurons that\ncould be triggered by the adversary. To defend against backdoor attacks, many\nworks establish a staged pipeline to remove backdoors from victim DNNs:\ninspecting, locating, and erasing. However, in a scenario where a few clean\ndata can be accessible, such pipeline is fragile and cannot erase backdoors\ncompletely without sacrificing model accuracy. To address this issue, in this\npaper, we propose a novel data-free holistic backdoor erasing (DHBE) framework.\nInstead of the staged pipeline, the DHBE treats the backdoor erasing task as a\nunified adversarial procedure, which seeks equilibrium between two different\ncompeting processes: distillation and backdoor regularization. In distillation,\nthe backdoored DNN is distilled into a proxy model, transferring its knowledge\nabout clean data, yet backdoors are simultaneously transferred. In backdoor\nregularization, the proxy model is holistically regularized to prevent from\ninfecting any possible backdoor transferred from distillation. These two\nprocesses jointly proceed with data-free adversarial optimization until a\nclean, high-accuracy proxy model is obtained. With the novel adversarial\ndesign, our framework demonstrates its superiority in three aspects: 1) minimal\ndetriment to model accuracy, 2) high tolerance for hyperparameters, and 3) no\ndemand for clean data. Extensive experiments on various backdoor attacks and\ndatasets are performed to verify the effectiveness of the proposed framework.\nCode is available at \\url{https://github.com/yanzhicong/DHBE}\n", "date": "2023-06-13", "categories": []}
{"id": "2306.08257", "title": "On the Robustness of Latent Diffusion Models", "abstract": "  Latent diffusion models achieve state-of-the-art performance on a variety of\ngenerative tasks, such as image synthesis and image editing. However, the\nrobustness of latent diffusion models is not well studied. Previous works only\nfocus on the adversarial attacks against the encoder or the output image under\nwhite-box settings, regardless of the denoising process. Therefore, in this\npaper, we aim to analyze the robustness of latent diffusion models more\nthoroughly. We first study the influence of the components inside latent\ndiffusion models on their white-box robustness. In addition to white-box\nscenarios, we evaluate the black-box robustness of latent diffusion models via\ntransfer attacks, where we consider both prompt-transfer and model-transfer\nsettings and possible defense mechanisms. However, all these explorations need\na comprehensive benchmark dataset, which is missing in the literature.\nTherefore, to facilitate the research of the robustness of latent diffusion\nmodels, we propose two automatic dataset construction pipelines for two kinds\nof image editing models and release the whole dataset. Our code and dataset are\navailable at \\url{https://github.com/jpzhang1810/LDM-Robustness}.\n", "date": "2023-06-14", "categories": []}
{"id": "2306.08492", "title": "A Relaxed Optimization Approach for Adversarial Attacks against Neural\n  Machine Translation Models", "abstract": "  In this paper, we propose an optimization-based adversarial attack against\nNeural Machine Translation (NMT) models. First, we propose an optimization\nproblem to generate adversarial examples that are semantically similar to the\noriginal sentences but destroy the translation generated by the target NMT\nmodel. This optimization problem is discrete, and we propose a continuous\nrelaxation to solve it. With this relaxation, we find a probability\ndistribution for each token in the adversarial example, and then we can\ngenerate multiple adversarial examples by sampling from these distributions.\nExperimental results show that our attack significantly degrades the\ntranslation quality of multiple NMT models while maintaining the semantic\nsimilarity between the original and adversarial sentences. Furthermore, our\nattack outperforms the baselines in terms of success rate, similarity\npreservation, effect on translation quality, and token error rate. Finally, we\npropose a black-box extension of our attack by sampling from an optimized\nprobability distribution for a reference model whose gradients are accessible.\n", "date": "2023-06-14", "categories": []}
{"id": "2306.08565", "title": "Reliable Evaluation of Adversarial Transferability", "abstract": "  Adversarial examples (AEs) with small adversarial perturbations can mislead\ndeep neural networks (DNNs) into wrong predictions. The AEs created on one DNN\ncan also fool another DNN. Over the last few years, the transferability of AEs\nhas garnered significant attention as it is a crucial property for facilitating\nblack-box attacks. Many approaches have been proposed to improve adversarial\ntransferability. However, they are mainly verified across different\nconvolutional neural network (CNN) architectures, which is not a reliable\nevaluation since all CNNs share some similar architectural biases. In this\nwork, we re-evaluate 12 representative transferability-enhancing attack methods\nwhere we test on 18 popular models from 4 types of neural networks. Our\nreevaluation revealed that the adversarial transferability is often\noverestimated, and there is no single AE that can be transferred to all popular\nmodels. The transferability rank of previous attacking methods changes when\nunder our comprehensive evaluation. Based on our analysis, we propose a\nreliable benchmark including three evaluation protocols. Adversarial\ntransferability on our new benchmark is extremely low, which further confirms\nthe overestimation of adversarial transferability. We release our benchmark at\nhttps://adv-trans-eval.github.io to facilitate future research, which includes\ncode, model checkpoints, and evaluation protocols.\n", "date": "2023-06-14", "categories": []}
{"id": "2306.07397", "title": "Adversarial Attacks on the Interpretation of Neuron Activation\n  Maximization", "abstract": "  The internal functional behavior of trained Deep Neural Networks is\nnotoriously difficult to interpret. Activation-maximization approaches are one\nset of techniques used to interpret and analyze trained deep-learning models.\nThese consist in finding inputs that maximally activate a given neuron or\nfeature map. These inputs can be selected from a data set or obtained by\noptimization. However, interpretability methods may be subject to being\ndeceived. In this work, we consider the concept of an adversary manipulating a\nmodel for the purpose of deceiving the interpretation. We propose an\noptimization framework for performing this manipulation and demonstrate a\nnumber of ways that popular activation-maximization interpretation techniques\nassociated with CNNs can be manipulated to change the interpretations, shedding\nlight on the reliability of these methods.\n", "date": "2023-06-12", "categories": []}
{"id": "2306.07525", "title": "Using Collision Momentum in Deep Reinforcement Learning Based\n  Adversarial Pedestrian Modeling", "abstract": "  Recent research in pedestrian simulation often aims to develop realistic\nbehaviors in various situations, but it is challenging for existing algorithms\nto generate behaviors that identify weaknesses in automated vehicles'\nperformance in extreme and unlikely scenarios and edge cases. To address this,\nspecialized pedestrian behavior algorithms are needed. Current research focuses\non realistic trajectories using social force models and reinforcement learning\nbased models. However, we propose a reinforcement learning algorithm that\nspecifically targets collisions and better uncovers unique failure modes of\nautomated vehicle controllers. Our algorithm is efficient and generates more\nsevere collisions, allowing for the identification and correction of weaknesses\nin autonomous driving algorithms in complex and varied scenarios.\n", "date": "2023-06-13", "categories": []}
{"id": "2306.07655", "title": "Malafide: a novel adversarial convolutive noise attack against deepfake\n  and spoofing detection systems", "abstract": "  We present Malafide, a universal adversarial attack against automatic speaker\nverification (ASV) spoofing countermeasures (CMs). By introducing convolutional\nnoise using an optimised linear time-invariant filter, Malafide attacks can be\nused to compromise CM reliability while preserving other speech attributes such\nas quality and the speaker's voice. In contrast to other adversarial attacks\nproposed recently, Malafide filters are optimised independently of the input\nutterance and duration, are tuned instead to the underlying spoofing attack,\nand require the optimisation of only a small number of filter coefficients.\nEven so, they degrade CM performance estimates by an order of magnitude, even\nin black-box settings, and can also be configured to overcome integrated CM and\nASV subsystems. Integrated solutions that use self-supervised learning CMs,\nhowever, are more robust, under both black-box and white-box settings.\n", "date": "2023-06-13", "categories": []}
{"id": "2306.07723", "title": "Theoretical Foundations of Adversarially Robust Learning", "abstract": "  Despite extraordinary progress, current machine learning systems have been\nshown to be brittle against adversarial examples: seemingly innocuous but\ncarefully crafted perturbations of test examples that cause machine learning\npredictors to misclassify. Can we learn predictors robust to adversarial\nexamples? and how? There has been much empirical interest in this contemporary\nchallenge in machine learning, and in this thesis, we address it from a\ntheoretical perspective.\n  In this thesis, we explore what robustness properties can we hope to\nguarantee against adversarial examples and develop an understanding of how to\nalgorithmically guarantee them. We illustrate the need to go beyond traditional\napproaches and principles such as empirical risk minimization and uniform\nconvergence, and make contributions that can be categorized as follows: (1)\nintroducing problem formulations capturing aspects of emerging practical\nchallenges in robust learning, (2) designing new learning algorithms with\nprovable robustness guarantees, and (3) characterizing the complexity of robust\nlearning and fundamental limitations on the performance of any algorithm.\n", "date": "2023-06-13", "categories": []}
{"id": "2306.07768", "title": "Area is all you need: repeatable elements make stronger adversarial\n  attacks", "abstract": "  Over the last decade, deep neural networks have achieved state of the art in\ncomputer vision tasks. These models, however, are susceptible to unusual\ninputs, known as adversarial examples, that cause them to misclassify or\notherwise fail to detect objects. Here, we provide evidence that the increasing\nsuccess of adversarial attacks is primarily due to increasing their size. We\nthen demonstrate a method for generating the largest possible adversarial patch\nby building a adversarial pattern out of repeatable elements. This approach\nachieves a new state of the art in evading detection by YOLOv2 and YOLOv3.\nFinally, we present an experiment that fails to replicate the prior success of\nseveral attacks published in this field, and end with some comments on testing\nand reproducibility.\n", "date": "2023-06-13", "categories": []}
{"id": "2306.07796", "title": "Finite Gaussian Neurons: Defending against adversarial attacks by making\n  neural networks say \"I don't know\"", "abstract": "  Since 2014, artificial neural networks have been known to be vulnerable to\nadversarial attacks, which can fool the network into producing wrong or\nnonsensical outputs by making humanly imperceptible alterations to inputs.\nWhile defenses against adversarial attacks have been proposed, they usually\ninvolve retraining a new neural network from scratch, a costly task. In this\nwork, I introduce the Finite Gaussian Neuron (FGN), a novel neuron architecture\nfor artificial neural networks. My works aims to: - easily convert existing\nmodels to Finite Gaussian Neuron architecture, - while preserving the existing\nmodel's behavior on real data, - and offering resistance against adversarial\nattacks. I show that converted and retrained Finite Gaussian Neural Networks\n(FGNN) always have lower confidence (i.e., are not overconfident) in their\npredictions over randomized and Fast Gradient Sign Method adversarial images\nwhen compared to classical neural networks, while maintaining high accuracy and\nconfidence over real MNIST images. To further validate the capacity of Finite\nGaussian Neurons to protect from adversarial attacks, I compare the behavior of\nFGNs to that of Bayesian Neural Networks against both randomized and\nadversarial images, and show how the behavior of the two architectures differs.\nFinally I show some limitations of the FGN models by testing them on the more\ncomplex SPEECHCOMMANDS task, against the stronger Carlini-Wagner and Projected\nGradient Descent adversarial attacks.\n", "date": "2023-06-13", "categories": []}
{"id": "2303.13649", "title": "Adversarial Robustness and Feature Impact Analysis for Driver Drowsiness\n  Detection", "abstract": "  Drowsy driving is a major cause of road accidents, but drivers are dismissive\nof the impact that fatigue can have on their reaction times. To detect\ndrowsiness before any impairment occurs, a promising strategy is using Machine\nLearning (ML) to monitor Heart Rate Variability (HRV) signals. This work\npresents multiple experiments with different HRV time windows and ML models, a\nfeature impact analysis using Shapley Additive Explanations (SHAP), and an\nadversarial robustness analysis to assess their reliability when processing\nfaulty input data and perturbed HRV signals. The most reliable model was\nExtreme Gradient Boosting (XGB) and the optimal time window had between 120 and\n150 seconds. Furthermore, SHAP enabled the selection of the 18 most impactful\nfeatures and the training of new smaller models that achieved a performance as\ngood as the initial ones. Despite the susceptibility of all models to\nadversarial attacks, adversarial training enabled them to preserve\nsignificantly higher results, especially XGB. Therefore, ML models can\nsignificantly benefit from realistic adversarial training to provide a more\nrobust driver drowsiness detection.\n", "date": "2023-03-23", "categories": []}
{"id": "2306.06107", "title": "Adversarial Attacks on Leakage Detectors in Water Distribution Networks", "abstract": "  Many Machine Learning models are vulnerable to adversarial attacks: There\nexist methodologies that add a small (imperceptible) perturbation to an input\nsuch that the model comes up with a wrong prediction. Better understanding of\nsuch attacks is crucial in particular for models used in security-critical\ndomains, such as monitoring of water distribution networks, in order to devise\ncounter-measures enhancing model robustness and trustworthiness.\n  We propose a taxonomy for adversarial attacks against machine learning based\nleakage detectors in water distribution networks. Following up on this, we\nfocus on a particular type of attack: an adversary searching the least\nsensitive point, that is, the location in the water network where the largest\npossible undetected leak could occur. Based on a mathematical formalization of\nthe least sensitive point problem, we use three different algorithmic\napproaches to find a solution. Results are evaluated on two benchmark water\ndistribution networks.\n", "date": "2023-05-25", "categories": []}
{"id": "2306.06462", "title": "Boosting Adversarial Robustness using Feature Level Stochastic Smoothing", "abstract": "  Advances in adversarial defenses have led to a significant improvement in the\nrobustness of Deep Neural Networks. However, the robust accuracy of present\nstate-ofthe-art defenses is far from the requirements in critical applications\nsuch as robotics and autonomous navigation systems. Further, in practical use\ncases, network prediction alone might not suffice, and assignment of a\nconfidence value for the prediction can prove crucial. In this work, we propose\na generic method for introducing stochasticity in the network predictions, and\nutilize this for smoothing decision boundaries and rejecting low confidence\npredictions, thereby boosting the robustness on accepted samples. The proposed\nFeature Level Stochastic Smoothing based classification also results in a boost\nin robustness without rejection over existing adversarial training methods.\nFinally, we combine the proposed method with adversarial detection methods, to\nachieve the benefits of both approaches.\n", "date": "2023-06-10", "categories": []}
{"id": "2306.06995", "title": "How robust accuracy suffers from certified training with convex\n  relaxations", "abstract": "  Adversarial attacks pose significant threats to deploying state-of-the-art\nclassifiers in safety-critical applications. Two classes of methods have\nemerged to address this issue: empirical defences and certified defences.\nAlthough certified defences come with robustness guarantees, empirical defences\nsuch as adversarial training enjoy much higher popularity among practitioners.\nIn this paper, we systematically compare the standard and robust error of these\ntwo robust training paradigms across multiple computer vision tasks. We show\nthat in most tasks and for both $\\mathscr{l}_\\infty$-ball and\n$\\mathscr{l}_2$-ball threat models, certified training with convex relaxations\nsuffers from worse standard and robust error than adversarial training. We\nfurther explore how the error gap between certified and adversarial training\ndepends on the threat model and the data distribution. In particular, besides\nthe perturbation budget, we identify as important factors the shape of the\nperturbation set and the implicit margin of the data distribution. We support\nour arguments with extensive ablations on both synthetic and image datasets.\n", "date": "2023-06-12", "categories": []}
{"id": "2212.05400", "title": "How to Backdoor Diffusion Models?", "abstract": "  Diffusion models are state-of-the-art deep learning empowered generative\nmodels that are trained based on the principle of learning forward and reverse\ndiffusion processes via progressive noise-addition and denoising. To gain a\nbetter understanding of the limitations and potential risks, this paper\npresents the first study on the robustness of diffusion models against backdoor\nattacks. Specifically, we propose BadDiffusion, a novel attack framework that\nengineers compromised diffusion processes during model training for backdoor\nimplantation. At the inference stage, the backdoored diffusion model will\nbehave just like an untampered generator for regular data inputs, while falsely\ngenerating some targeted outcome designed by the bad actor upon receiving the\nimplanted trigger signal. Such a critical risk can be dreadful for downstream\ntasks and applications built upon the problematic model. Our extensive\nexperiments on various backdoor attack settings show that BadDiffusion can\nconsistently lead to compromised diffusion models with high utility and target\nspecificity. Even worse, BadDiffusion can be made cost-effective by simply\nfinetuning a clean pre-trained diffusion model to implant backdoors. We also\nexplore some possible countermeasures for risk mitigation. Our results call\nattention to potential risks and possible misuse of diffusion models. Our code\nis available on https://github.com/IBM/BadDiffusion.\n", "date": "2022-12-11", "categories": []}
{"id": "2306.05495", "title": "Is Attentional Channel Processing Design Required? Comprehensive\n  Analysis Of Robustness Between Vision Transformers And Fully Attentional\n  Networks", "abstract": "  The robustness testing has been performed for standard CNN models and Vision\nTransformers, however there is a lack of comprehensive study between the\nrobustness of traditional Vision Transformers without an extra attentional\nchannel design and the latest fully attentional network(FAN) models. So in this\npaper, we use the ImageNet dataset to compare the robustness of fully\nattentional network(FAN) models with traditional Vision Transformers to\nunderstand the role of an attentional channel processing design using white box\nattacks and also study the transferability between the same using black box\nattacks.\n", "date": "2023-06-08", "categories": []}
{"id": "2306.05873", "title": "Detecting Adversarial Directions in Deep Reinforcement Learning to Make\n  Robust Decisions", "abstract": "  Learning in MDPs with highly complex state representations is currently\npossible due to multiple advancements in reinforcement learning algorithm\ndesign. However, this incline in complexity, and furthermore the increase in\nthe dimensions of the observation came at the cost of volatility that can be\ntaken advantage of via adversarial attacks (i.e. moving along worst-case\ndirections in the observation space). To solve this policy instability problem\nwe propose a novel method to detect the presence of these non-robust directions\nvia local quadratic approximation of the deep neural policy loss. Our method\nprovides a theoretical basis for the fundamental cut-off between safe\nobservations and adversarial observations. Furthermore, our technique is\ncomputationally efficient, and does not depend on the methods used to produce\nthe worst-case directions. We conduct extensive experiments in the Arcade\nLearning Environment with several different adversarial attack techniques. Most\nsignificantly, we demonstrate the effectiveness of our approach even in the\nsetting where non-robust directions are explicitly optimized to circumvent our\nproposed method.\n", "date": "2023-06-09", "categories": []}
{"id": "2306.04252", "title": "Adversarial Sample Detection Through Neural Network Transport Dynamics", "abstract": "  We propose a detector of adversarial samples that is based on the view of\nneural networks as discrete dynamic systems. The detector tells clean inputs\nfrom abnormal ones by comparing the discrete vector fields they follow through\nthe layers. We also show that regularizing this vector field during training\nmakes the network more regular on the data distribution's support, thus making\nthe activations of clean inputs more distinguishable from those of abnormal\nones. Experimentally, we compare our detector favorably to other detectors on\nseen and unseen attacks, and show that the regularization of the network's\ndynamics improves the performance of adversarial detectors that use the\ninternal embeddings as inputs, while also improving test accuracy.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.04756", "title": "A Linearly Convergent GAN Inversion-based Algorithm for Reverse\n  Engineering of Deceptions", "abstract": "  An important aspect of developing reliable deep learning systems is devising\nstrategies that make these systems robust to adversarial attacks. There is a\nlong line of work that focuses on developing defenses against these attacks,\nbut recently, researchers have began to study ways to reverse engineer the\nattack process. This allows us to not only defend against several attack\nmodels, but also classify the threat model. However, there is still a lack of\ntheoretical guarantees for the reverse engineering process. Current approaches\nthat give any guarantees are based on the assumption that the data lies in a\nunion of linear subspaces, which is not a valid assumption for more complex\ndatasets. In this paper, we build on prior work and propose a novel framework\nfor reverse engineering of deceptions which supposes that the clean data lies\nin the range of a GAN. To classify the signal and attack, we jointly solve a\nGAN inversion problem and a block-sparse recovery problem. For the first time\nin the literature, we provide deterministic linear convergence guarantees for\nthis problem. We also empirically demonstrate the merits of the proposed\napproach on several nonlinear datasets as compared to state-of-the-art methods.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.04874", "title": "Expanding Scope: Adapting English Adversarial Attacks to Chinese", "abstract": "  Recent studies have revealed that NLP predictive models are vulnerable to\nadversarial attacks. Most existing studies focused on designing attacks to\nevaluate the robustness of NLP models in the English language alone. Literature\nhas seen an increasing need for NLP solutions for other languages. We,\ntherefore, ask one natural question: whether state-of-the-art (SOTA) attack\nmethods generalize to other languages. This paper investigates how to adapt\nSOTA adversarial attack algorithms in English to the Chinese language. Our\nexperiments show that attack methods previously applied to English NLP can\ngenerate high-quality adversarial examples in Chinese when combined with proper\ntext segmentation and linguistic constraints. In addition, we demonstrate that\nthe generated adversarial examples can achieve high fluency and semantic\nconsistency by focusing on the Chinese language's morphology and phonology,\nwhich in turn can be used to improve the adversarial robustness of Chinese NLP\nmodels.\n", "date": "2023-06-08", "categories": []}
{"id": "2306.05358", "title": "Trustworthy Sensor Fusion against Inaudible Command Attacks in Advanced\n  Driver-Assistance System", "abstract": "  There are increasing concerns about malicious attacks on autonomous vehicles.\nIn particular, inaudible voice command attacks pose a significant threat as\nvoice commands become available in autonomous driving systems. How to\nempirically defend against these inaudible attacks remains an open question.\nPrevious research investigates utilizing deep learning-based multimodal fusion\nfor defense, without considering the model uncertainty in trustworthiness. As\ndeep learning has been applied to increasingly sensitive tasks, uncertainty\nmeasurement is crucial in helping improve model robustness, especially in\nmission-critical scenarios. In this paper, we propose the Multimodal Fusion\nFramework (MFF) as an intelligent security system to defend against inaudible\nvoice command attacks. MFF fuses heterogeneous audio-vision modalities using\nVGG family neural networks and achieves the detection accuracy of 92.25% in the\ncomparative fusion method empirical study. Additionally, extensive experiments\non audio-vision tasks reveal the model's uncertainty. Using Expected\nCalibration Errors, we measure calibration errors and Monte-Carlo Dropout to\nestimate the predictive distribution for the proposed models. Our findings show\nempirically to train robust multimodal models, improve standard accuracy and\nprovide a further step toward interpretability. Finally, we discuss the pros\nand cons of our approach and its applicability for Advanced Driver Assistance\nSystems.\n", "date": "2023-05-30", "categories": []}
{"id": "2302.02502", "title": "Rethinking Robust Contrastive Learning from the Adversarial Perspective", "abstract": "  To advance the understanding of robust deep learning, we delve into the\neffects of adversarial training on self-supervised and supervised contrastive\nlearning alongside supervised learning. Our analysis uncovers significant\ndisparities between adversarial and clean representations in standard-trained\nnetworks across various learning algorithms. Remarkably, adversarial training\nmitigates these disparities and fosters the convergence of representations\ntoward a universal set, regardless of the learning scheme used. Additionally,\nincreasing the similarity between adversarial and clean representations,\nparticularly near the end of the network, enhances network robustness. These\nfindings offer valuable insights for designing and training effective and\nrobust deep learning networks. Our code is released at\n\\textcolor{magenta}{\\url{https://github.com/softsys4ai/CL-Robustness}}.\n", "date": "2023-02-05", "categories": []}
{"id": "2206.10858", "title": "Robust Universal Adversarial Perturbations", "abstract": "  Universal Adversarial Perturbations (UAPs) are imperceptible, image-agnostic\nvectors that cause deep neural networks (DNNs) to misclassify inputs with high\nprobability. In practical attack scenarios, adversarial perturbations may\nundergo transformations such as changes in pixel intensity, scaling, etc.\nbefore being added to DNN inputs. Existing methods do not create UAPs robust to\nthese real-world transformations, thereby limiting their applicability in\npractical attack scenarios. In this work, we introduce and formulate UAPs\nrobust against real-world transformations. We build an iterative algorithm\nusing probabilistic robustness bounds and construct such UAPs robust to\ntransformations generated by composing arbitrary sub-differentiable\ntransformation functions. We perform an extensive evaluation on the popular\nCIFAR-10 and ILSVRC 2012 datasets measuring our UAPs' robustness under a wide\nrange common, real-world transformations such as rotation, contrast changes,\netc. We further show that by using a set of primitive transformations our\nmethod can generalize well to unseen transformations such as fog, JPEG\ncompression, etc. Our results show that our method can generate UAPs up to 23%\nmore robust than state-of-the-art baselines.\n", "date": "2022-06-22", "categories": []}
{"id": "2302.04578", "title": "Adversarial Example Does Good: Preventing Painting Imitation from\n  Diffusion Models via Adversarial Examples", "abstract": "  Recently, Diffusion Models (DMs) boost a wave in AI for Art yet raise new\ncopyright concerns, where infringers benefit from using unauthorized paintings\nto train DMs to generate novel paintings in a similar style. To address these\nemerging copyright violations, in this paper, we are the first to explore and\npropose to utilize adversarial examples for DMs to protect human-created\nartworks. Specifically, we first build a theoretical framework to define and\nevaluate the adversarial examples for DMs. Then, based on this framework, we\ndesign a novel algorithm, named AdvDM, which exploits a Monte-Carlo estimation\nof adversarial examples for DMs by optimizing upon different latent variables\nsampled from the reverse process of DMs. Extensive experiments show that the\ngenerated adversarial examples can effectively hinder DMs from extracting their\nfeatures. Therefore, our method can be a powerful tool for human artists to\nprotect their copyright against infringers equipped with DM-based AI-for-Art\napplications. The code of our method is available on GitHub:\nhttps://github.com/mist-project/mist.git.\n", "date": "2023-02-09", "categories": []}
{"id": "2303.03592", "title": "Exploring the Limits of Model-Targeted Indiscriminate Data Poisoning\n  Attacks", "abstract": "  Indiscriminate data poisoning attacks aim to decrease a model's test accuracy\nby injecting a small amount of corrupted training data. Despite significant\ninterest, existing attacks remain relatively ineffective against modern machine\nlearning (ML) architectures. In this work, we introduce the notion of model\npoisoning reachability as a technical tool to explore the intrinsic limits of\ndata poisoning attacks towards target parameters (i.e., model-targeted\nattacks). We derive an easily computable threshold to establish and quantify a\nsurprising phase transition phenomenon among popular ML models: data poisoning\nattacks can achieve certain target parameters only when the poisoning ratio\nexceeds our threshold. Building on existing parameter corruption attacks and\nrefining the Gradient Canceling attack, we perform extensive experiments to\nconfirm our theoretical findings, test the predictability of our transition\nthreshold, and significantly improve existing indiscriminate data poisoning\nbaselines over a range of datasets and models. Our work highlights the critical\nrole played by the poisoning ratio, and sheds new insights on existing\nempirical results, attacks and mitigation strategies in data poisoning.\n", "date": "2023-03-07", "categories": []}
{"id": "2306.02583", "title": "Stable Diffusion is Unstable", "abstract": "  Recently, text-to-image models have been thriving. Despite their powerful\ngenerative capacity, our research has uncovered a lack of robustness in this\ngeneration process. Specifically, the introduction of small perturbations to\nthe text prompts can result in the blending of primary subjects with other\ncategories or their complete disappearance in the generated images. In this\npaper, we propose Auto-attack on Text-to-image Models (ATM), a gradient-based\napproach, to effectively and efficiently generate such perturbations. By\nlearning a Gumbel Softmax distribution, we can make the discrete process of\nword replacement or extension continuous, thus ensuring the differentiability\nof the perturbation generation. Once the distribution is learned, ATM can\nsample multiple attack samples simultaneously. These attack samples can prevent\nthe generative model from generating the desired subjects without compromising\nimage quality. ATM has achieved a 91.1% success rate in short-text attacks and\nan 81.2% success rate in long-text attacks. Further empirical analysis revealed\nfour attack patterns based on: 1) the variability in generation speed, 2) the\nsimilarity of coarse-grained characteristics, 3) the polysemy of words, and 4)\nthe positioning of words.\n", "date": "2023-06-05", "categories": []}
{"id": "2306.03229", "title": "Adversarial alignment: Breaking the trade-off between the strength of an\n  attack and its relevance to human perception", "abstract": "  Deep neural networks (DNNs) are known to have a fundamental sensitivity to\nadversarial attacks, perturbations of the input that are imperceptible to\nhumans yet powerful enough to change the visual decision of a model.\nAdversarial attacks have long been considered the \"Achilles' heel\" of deep\nlearning, which may eventually force a shift in modeling paradigms.\nNevertheless, the formidable capabilities of modern large-scale DNNs have\nsomewhat eclipsed these early concerns. Do adversarial attacks continue to pose\na threat to DNNs?\n  Here, we investigate how the robustness of DNNs to adversarial attacks has\nevolved as their accuracy on ImageNet has continued to improve. We measure\nadversarial robustness in two different ways: First, we measure the smallest\nadversarial attack needed to cause a model to change its object categorization\ndecision. Second, we measure how aligned successful attacks are with the\nfeatures that humans find diagnostic for object recognition. We find that\nadversarial attacks are inducing bigger and more easily detectable changes to\nimage pixels as DNNs grow better on ImageNet, but these attacks are also\nbecoming less aligned with features that humans find diagnostic for\nrecognition. To better understand the source of this trade-off, we turn to the\nneural harmonizer, a DNN training routine that encourages models to leverage\nthe same features as humans to solve tasks. Harmonized DNNs achieve the best of\nboth worlds and experience attacks that are detectable and affect features that\nhumans find diagnostic for recognition, meaning that attacks on these models\nare more likely to be rendered ineffective by inducing similar effects on human\nperception. Our findings suggest that the sensitivity of DNNs to adversarial\nattacks can be mitigated by DNN scale, data scale, and training routines that\nalign models with biological intelligence.\n", "date": "2023-06-05", "categories": []}
{"id": "2306.03726", "title": "Exploring Model Dynamics for Accumulative Poisoning Discovery", "abstract": "  Adversarial poisoning attacks pose huge threats to various machine learning\napplications. Especially, the recent accumulative poisoning attacks show that\nit is possible to achieve irreparable harm on models via a sequence of\nimperceptible attacks followed by a trigger batch. Due to the limited\ndata-level discrepancy in real-time data streaming, current defensive methods\nare indiscriminate in handling the poison and clean samples. In this paper, we\ndive into the perspective of model dynamics and propose a novel information\nmeasure, namely, Memorization Discrepancy, to explore the defense via the\nmodel-level information. By implicitly transferring the changes in the data\nmanipulation to that in the model outputs, Memorization Discrepancy can\ndiscover the imperceptible poison samples based on their distinct dynamics from\nthe clean samples. We thoroughly explore its properties and propose\nDiscrepancy-aware Sample Correction (DSC) to defend against accumulative\npoisoning attacks. Extensive experiments comprehensively characterized\nMemorization Discrepancy and verified its effectiveness. The code is publicly\navailable at: https://github.com/tmlr-group/Memorization-Discrepancy.\n", "date": "2023-06-06", "categories": []}
{"id": "2209.15266", "title": "Data Poisoning Attacks Against Multimodal Encoders", "abstract": "  Recently, the newly emerged multimodal models, which leverage both visual and\nlinguistic modalities to train powerful encoders, have gained increasing\nattention. However, learning from a large-scale unlabeled dataset also exposes\nthe model to the risk of potential poisoning attacks, whereby the adversary\naims to perturb the model's training data to trigger malicious behaviors in it.\nIn contrast to previous work, only poisoning visual modality, in this work, we\ntake the first step to studying poisoning attacks against multimodal models in\nboth visual and linguistic modalities. Specially, we focus on answering two\nquestions: (1) Is the linguistic modality also vulnerable to poisoning attacks?\nand (2) Which modality is most vulnerable? To answer the two questions, we\npropose three types of poisoning attacks against multimodal models. Extensive\nevaluations on different datasets and model architectures show that all three\nattacks can achieve significant attack performance while maintaining model\nutility in both visual and linguistic modalities. Furthermore, we observe that\nthe poisoning effect differs between different modalities. To mitigate the\nattacks, we propose both pre-training and post-training defenses. We\nempirically show that both defenses can significantly reduce the attack\nperformance while preserving the model's utility.\n", "date": "2022-09-30", "categories": []}
{"id": "2212.06079", "title": "Robust Perception through Equivariance", "abstract": "  Deep networks for computer vision are not reliable when they encounter\nadversarial examples. In this paper, we introduce a framework that uses the\ndense intrinsic constraints in natural images to robustify inference. By\nintroducing constraints at inference time, we can shift the burden of\nrobustness from training to the inference algorithm, thereby allowing the model\nto adjust dynamically to each individual image's unique and potentially novel\ncharacteristics at inference time. Among different constraints, we find that\nequivariance-based constraints are most effective, because they allow dense\nconstraints in the feature space without overly constraining the representation\nat a fine-grained level. Our theoretical results validate the importance of\nhaving such dense constraints at inference time. Our empirical experiments show\nthat restoring feature equivariance at inference time defends against\nworst-case adversarial perturbations. The method obtains improved adversarial\nrobustness on four datasets (ImageNet, Cityscapes, PASCAL VOC, and MS-COCO) on\nimage recognition, semantic segmentation, and instance segmentation tasks.\nProject page is available at equi4robust.cs.columbia.edu.\n", "date": "2022-12-12", "categories": []}
{"id": "2303.15754", "title": "Transferable Adversarial Attacks on Vision Transformers with Token\n  Gradient Regularization", "abstract": "  Vision transformers (ViTs) have been successfully deployed in a variety of\ncomputer vision tasks, but they are still vulnerable to adversarial samples.\nTransfer-based attacks use a local model to generate adversarial samples and\ndirectly transfer them to attack a target black-box model. The high efficiency\nof transfer-based attacks makes it a severe security threat to ViT-based\napplications. Therefore, it is vital to design effective transfer-based attacks\nto identify the deficiencies of ViTs beforehand in security-sensitive\nscenarios. Existing efforts generally focus on regularizing the input gradients\nto stabilize the updated direction of adversarial samples. However, the\nvariance of the back-propagated gradients in intermediate blocks of ViTs may\nstill be large, which may make the generated adversarial samples focus on some\nmodel-specific features and get stuck in poor local optima. To overcome the\nshortcomings of existing approaches, we propose the Token Gradient\nRegularization (TGR) method. According to the structural characteristics of\nViTs, TGR reduces the variance of the back-propagated gradient in each internal\nblock of ViTs in a token-wise manner and utilizes the regularized gradient to\ngenerate adversarial samples. Extensive experiments on attacking both ViTs and\nCNNs confirm the superiority of our approach. Notably, compared to the\nstate-of-the-art transfer-based attacks, our TGR offers a performance\nimprovement of 8.8% on average.\n", "date": "2023-03-28", "categories": []}
{"id": "2304.04033", "title": "Exploring the Connection between Robust and Generative Models", "abstract": "  We offer a study that connects robust discriminative classifiers trained with\nadversarial training (AT) with generative modeling in the form of Energy-based\nModels (EBM). We do so by decomposing the loss of a discriminative classifier\nand showing that the discriminative model is also aware of the input data\ndensity. Though a common assumption is that adversarial points leave the\nmanifold of the input data, our study finds out that, surprisingly, untargeted\nadversarial points in the input space are very likely under the generative\nmodel hidden inside the discriminative classifier -- have low energy in the\nEBM. We present two evidence: untargeted attacks are even more likely than the\nnatural data and their likelihood increases as the attack strength increases.\nThis allows us to easily detect them and craft a novel attack called\nHigh-Energy PGD that fools the classifier yet has energy similar to the data\nset. The code is available at github.com/senad96/Robust-Generative\n", "date": "2023-04-08", "categories": []}
{"id": "2306.01809", "title": "Adversarial Attack Based on Prediction-Correction", "abstract": "  Deep neural networks (DNNs) are vulnerable to adversarial examples obtained\nby adding small perturbations to original examples. The added perturbations in\nexisting attacks are mainly determined by the gradient of the loss function\nwith respect to the inputs. In this paper, the close relationship between\ngradient-based attacks and the numerical methods for solving ordinary\ndifferential equation (ODE) is studied for the first time. Inspired by the\nnumerical solution of ODE, a new prediction-correction (PC) based adversarial\nattack is proposed. In our proposed PC-based attack, some existing attack can\nbe selected to produce a predicted example first, and then the predicted\nexample and the current example are combined together to determine the added\nperturbations. The proposed method possesses good extensibility and can be\napplied to all available gradient-based attacks easily. Extensive experiments\ndemonstrate that compared with the state-of-the-art gradient-based adversarial\nattacks, our proposed PC-based attacks have higher attack success rates, and\nexhibit better transferability.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.02002", "title": "Can Directed Graph Neural Networks be Adversarially Robust?", "abstract": "  The existing research on robust Graph Neural Networks (GNNs) fails to\nacknowledge the significance of directed graphs in providing rich information\nabout networks' inherent structure. This work presents the first investigation\ninto the robustness of GNNs in the context of directed graphs, aiming to\nharness the profound trust implications offered by directed graphs to bolster\nthe robustness and resilience of GNNs. Our study reveals that existing directed\nGNNs are not adversarially robust. In pursuit of our goal, we introduce a new\nand realistic directed graph attack setting and propose an innovative,\nuniversal, and efficient message-passing framework as a plug-in layer to\nsignificantly enhance the robustness of GNNs. Combined with existing defense\nstrategies, this framework achieves outstanding clean accuracy and\nstate-of-the-art robust performance, offering superior defense against both\ntransfer and adaptive attacks. The findings in this study reveal a novel and\npromising direction for this crucial research area. The code will be made\npublicly available upon the acceptance of this work.\n", "date": "2023-06-03", "categories": []}
{"id": "2306.02021", "title": "Towards Black-box Adversarial Example Detection: A Data\n  Reconstruction-based Method", "abstract": "  Adversarial example detection is known to be an effective adversarial defense\nmethod. Black-box attack, which is a more realistic threat and has led to\nvarious black-box adversarial training-based defense methods, however, does not\nattract considerable attention in adversarial example detection. In this paper,\nwe fill this gap by positioning the problem of black-box adversarial example\ndetection (BAD). Data analysis under the introduced BAD settings demonstrates\n(1) the incapability of existing detectors in addressing the black-box scenario\nand (2) the potential of exploring BAD solutions from a data perspective. To\ntackle the BAD problem, we propose a data reconstruction-based adversarial\nexample detection method. Specifically, we use variational auto-encoder (VAE)\nto capture both pixel and frequency representations of normal examples. Then we\nuse reconstruction error to detect adversarial examples. Compared with existing\ndetection methods, the proposed method achieves substantially better detection\nperformance in BAD, which helps promote the deployment of adversarial example\ndetection-based defense solutions in real-world models.\n", "date": "2023-06-03", "categories": []}
{"id": "2306.02639", "title": "Evaluating robustness of support vector machines with the Lagrangian\n  dual approach", "abstract": "  Adversarial examples bring a considerable security threat to support vector\nmachines (SVMs), especially those used in safety-critical applications. Thus,\nrobustness verification is an essential issue for SVMs, which can provide\nprovable robustness against various kinds of adversary attacks. The evaluation\nresults obtained through the robustness verification can provide a safe\nguarantee for the use of SVMs. The existing verification method does not often\nperform well in verifying SVMs with nonlinear kernels. To this end, we propose\na method to improve the verification performance for SVMs with nonlinear\nkernels. We first formalize the adversarial robustness evaluation of SVMs as an\noptimization problem. Then a lower bound of the original problem is obtained by\nsolving the Lagrangian dual problem of the original problem. Finally, the\nadversarial robustness of SVMs is evaluated concerning the lower bound. We\nevaluate the adversarial robustness of SVMs with linear and nonlinear kernels\non the MNIST and Fashion-MNIST datasets. The experimental results show that the\npercentage of provable robustness obtained by our method on the test set is\nbetter than that of the state-of-the-art.\n", "date": "2023-06-05", "categories": []}
{"id": "2306.02918", "title": "Adversarial Ink: Componentwise Backward Error Attacks on Deep Learning", "abstract": "  Deep neural networks are capable of state-of-the-art performance in many\nclassification tasks. However, they are known to be vulnerable to adversarial\nattacks -- small perturbations to the input that lead to a change in\nclassification. We address this issue from the perspective of backward error\nand condition number, concepts that have proved useful in numerical analysis.\nTo do this, we build on the work of Beuzeville et al. (2021). In particular, we\ndevelop a new class of attack algorithms that use componentwise relative\nperturbations. Such attacks are highly relevant in the case of handwritten\ndocuments or printed texts where, for example, the classification of\nsignatures, postcodes, dates or numerical quantities may be altered by changing\nonly the ink consistency and not the background. This makes the perturbed\nimages look natural to the naked eye. Such ``adversarial ink'' attacks\ntherefore reveal a weakness that can have a serious impact on safety and\nsecurity. We illustrate the new attacks on real data and contrast them with\nexisting algorithms. We also study the use of a componentwise condition number\nto quantify vulnerability.\n", "date": "2023-06-05", "categories": []}
{"id": "2304.13229", "title": "Generating Adversarial Examples with Task Oriented Multi-Objective\n  Optimization", "abstract": "  Deep learning models, even the-state-of-the-art ones, are highly vulnerable\nto adversarial examples. Adversarial training is one of the most efficient\nmethods to improve the model's robustness. The key factor for the success of\nadversarial training is the capability to generate qualified and divergent\nadversarial examples which satisfy some objectives/goals (e.g., finding\nadversarial examples that maximize the model losses for simultaneously\nattacking multiple models). Therefore, multi-objective optimization (MOO) is a\nnatural tool for adversarial example generation to achieve multiple\nobjectives/goals simultaneously. However, we observe that a naive application\nof MOO tends to maximize all objectives/goals equally, without caring if an\nobjective/goal has been achieved yet. This leads to useless effort to further\nimprove the goal-achieved tasks, while putting less focus on the\ngoal-unachieved tasks. In this paper, we propose \\emph{Task Oriented MOO} to\naddress this issue, in the context where we can explicitly define the goal\nachievement for a task. Our principle is to only maintain the goal-achieved\ntasks, while letting the optimizer spend more effort on improving the\ngoal-unachieved tasks. We conduct comprehensive experiments for our Task\nOriented MOO on various adversarial example generation schemes. The\nexperimental results firmly demonstrate the merit of our proposed approach. Our\ncode is available at \\url{https://github.com/tuananhbui89/TAMOO}.\n", "date": "2023-04-26", "categories": []}
{"id": "2306.01125", "title": "Reconstruction Distortion of Learned Image Compression with\n  Imperceptible Perturbations", "abstract": "  Learned Image Compression (LIC) has recently become the trending technique\nfor image transmission due to its notable performance. Despite its popularity,\nthe robustness of LIC with respect to the quality of image reconstruction\nremains under-explored. In this paper, we introduce an imperceptible attack\napproach designed to effectively degrade the reconstruction quality of LIC,\nresulting in the reconstructed image being severely disrupted by noise where\nany object in the reconstructed images is virtually impossible. More\nspecifically, we generate adversarial examples by introducing a Frobenius\nnorm-based loss function to maximize the discrepancy between original images\nand reconstructed adversarial examples. Further, leveraging the insensitivity\nof high-frequency components to human vision, we introduce Imperceptibility\nConstraint (IC) to ensure that the perturbations remain inconspicuous.\nExperiments conducted on the Kodak dataset using various LIC models demonstrate\neffectiveness. In addition, we provide several findings and suggestions for\ndesigning future defenses.\n", "date": "2023-06-01", "categories": []}
{"id": "2306.01273", "title": "VoteTRANS: Detecting Adversarial Text without Training by Voting on Hard\n  Labels of Transformations", "abstract": "  Adversarial attacks reveal serious flaws in deep learning models. More\ndangerously, these attacks preserve the original meaning and escape human\nrecognition. Existing methods for detecting these attacks need to be trained\nusing original/adversarial data. In this paper, we propose detection without\ntraining by voting on hard labels from predictions of transformations, namely,\nVoteTRANS. Specifically, VoteTRANS detects adversarial text by comparing the\nhard labels of input text and its transformation. The evaluation demonstrates\nthat VoteTRANS effectively detects adversarial text across various\nstate-of-the-art attacks, models, and datasets.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.01400", "title": "Adaptive Attractors: A Defense Strategy against ML Adversarial Collusion\n  Attacks", "abstract": "  In the seller-buyer setting on machine learning models, the seller generates\ndifferent copies based on the original model and distributes them to different\nbuyers, such that adversarial samples generated on one buyer's copy would\nlikely not work on other copies. A known approach achieves this using\nattractor-based rewriter which injects different attractors to different\ncopies. This induces different adversarial regions in different copies, making\nadversarial samples generated on one copy not replicable on others. In this\npaper, we focus on a scenario where multiple malicious buyers collude to\nattack. We first give two formulations and conduct empirical studies to analyze\neffectiveness of collusion attack under different assumptions on the attacker's\ncapabilities and properties of the attractors. We observe that existing\nattractor-based methods do not effectively mislead the colluders in the sense\nthat adversarial samples found are influenced more by the original model\ninstead of the attractors as number of colluders increases. Based on this\nobservation, we propose using adaptive attractors whose weight is guided by a\nU-shape curve to cover the shortfalls. Experimentation results show that when\nusing our approach, the attack success rate of a collusion attack converges to\naround 15% even when lots of copies are applied for collusion. In contrast,\nwhen using the existing attractor-based rewriter with fixed weight, the attack\nsuccess rate increases linearly with the number of copies used for collusion.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.01429", "title": "A Closer Look at the Adversarial Robustness of Deep Equilibrium Models", "abstract": "  Deep equilibrium models (DEQs) refrain from the traditional layer-stacking\nparadigm and turn to find the fixed point of a single layer. DEQs have achieved\npromising performance on different applications with featured memory\nefficiency. At the same time, the adversarial vulnerability of DEQs raises\nconcerns. Several works propose to certify robustness for monotone DEQs.\nHowever, limited efforts are devoted to studying empirical robustness for\ngeneral DEQs. To this end, we observe that an adversarially trained DEQ\nrequires more forward steps to arrive at the equilibrium state, or even\nviolates its fixed-point structure. Besides, the forward and backward tracks of\nDEQs are misaligned due to the black-box solvers. These facts cause gradient\nobfuscation when applying the ready-made attacks to evaluate or adversarially\ntrain DEQs. Given this, we develop approaches to estimate the intermediate\ngradients of DEQs and integrate them into the attacking pipelines. Our\napproaches facilitate fully white-box evaluations and lead to effective\nadversarial defense for DEQs. Extensive experiments on CIFAR-10 validate the\nadversarial robustness of DEQs competitive with deep networks of similar sizes.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.01435", "title": "Improving Adversarial Robustness of DEQs with Explicit Regulations Along\n  the Neural Dynamics", "abstract": "  Deep equilibrium (DEQ) models replace the multiple-layer stacking of\nconventional deep networks with a fixed-point iteration of a single-layer\ntransformation. Having been demonstrated to be competitive in a variety of\nreal-world scenarios, the adversarial robustness of general DEQs becomes\nincreasingly crucial for their reliable deployment. Existing works improve the\nrobustness of general DEQ models with the widely-used adversarial training (AT)\nframework, but they fail to exploit the structural uniquenesses of DEQ models.\nTo this end, we interpret DEQs through the lens of neural dynamics and find\nthat AT under-regulates intermediate states. Besides, the intermediate states\ntypically provide predictions with a high prediction entropy. Informed by the\ncorrelation between the entropy of dynamical systems and their stability\nproperties, we propose reducing prediction entropy by progressively updating\ninputs along the neural dynamics. During AT, we also utilize random\nintermediate states to compute the loss function. Our methods regulate the\nneural dynamics of DEQ models in this manner. Extensive experiments demonstrate\nthat our methods substantially increase the robustness of DEQ models and even\noutperform the strong deep network baselines.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.01655", "title": "Poisoning Network Flow Classifiers", "abstract": "  As machine learning (ML) classifiers increasingly oversee the automated\nmonitoring of network traffic, studying their resilience against adversarial\nattacks becomes critical. This paper focuses on poisoning attacks, specifically\nbackdoor attacks, against network traffic flow classifiers. We investigate the\nchallenging scenario of clean-label poisoning where the adversary's\ncapabilities are constrained to tampering only with the training data - without\nthe ability to arbitrarily modify the training labels or any other component of\nthe training process. We describe a trigger crafting strategy that leverages\nmodel interpretability techniques to generate trigger patterns that are\neffective even at very low poisoning rates. Finally, we design novel strategies\nto generate stealthy triggers, including an approach based on generative\nBayesian network models, with the goal of minimizing the conspicuousness of the\ntrigger, and thus making detection of an ongoing poisoning campaign more\nchallenging. Our findings provide significant insights into the feasibility of\npoisoning attacks on network traffic classifiers used in multiple scenarios,\nincluding detecting malicious communication and application classification.\n", "date": "2023-06-02", "categories": []}
{"id": "2206.03353", "title": "Improving Adversarial Robustness by Putting More Regularizations on Less\n  Robust Samples", "abstract": "  Adversarial training, which is to enhance robustness against adversarial\nattacks, has received much attention because it is easy to generate\nhuman-imperceptible perturbations of data to deceive a given deep neural\nnetwork. In this paper, we propose a new adversarial training algorithm that is\ntheoretically well motivated and empirically superior to other existing\nalgorithms. A novel feature of the proposed algorithm is to apply more\nregularization to data vulnerable to adversarial attacks than other existing\nregularization algorithms do. Theoretically, we show that our algorithm can be\nunderstood as an algorithm of minimizing the regularized empirical risk\nmotivated from a newly derived upper bound of the robust risk. Numerical\nexperiments illustrate that our proposed algorithm improves the generalization\n(accuracy on examples) and robustness (accuracy on adversarial attacks)\nsimultaneously to achieve the state-of-the-art performance.\n", "date": "2022-06-07", "categories": []}
{"id": "2302.04638", "title": "Better Diffusion Models Further Improve Adversarial Training", "abstract": "  It has been recognized that the data generated by the denoising diffusion\nprobabilistic model (DDPM) improves adversarial training. After two years of\nrapid development in diffusion models, a question naturally arises: can better\ndiffusion models further improve adversarial training? This paper gives an\naffirmative answer by employing the most recent diffusion model which has\nhigher efficiency ($\\sim 20$ sampling steps) and image quality (lower FID\nscore) compared with DDPM. Our adversarially trained models achieve\nstate-of-the-art performance on RobustBench using only generated data (no\nexternal datasets). Under the $\\ell_\\infty$-norm threat model with\n$\\epsilon=8/255$, our models achieve $70.69\\%$ and $42.67\\%$ robust accuracy on\nCIFAR-10 and CIFAR-100, respectively, i.e. improving upon previous\nstate-of-the-art models by $+4.58\\%$ and $+8.03\\%$. Under the $\\ell_2$-norm\nthreat model with $\\epsilon=128/255$, our models achieve $84.86\\%$ on CIFAR-10\n($+4.44\\%$). These results also beat previous works that use external data. We\nalso provide compelling results on the SVHN and TinyImageNet datasets. Our code\nis available at https://github.com/wzekai99/DM-Improves-AT.\n", "date": "2023-02-09", "categories": []}
{"id": "2306.00314", "title": "Adversarial-Aware Deep Learning System based on a Secondary Classical\n  Machine Learning Verification Approach", "abstract": "  Deep learning models have been used in creating various effective image\nclassification applications. However, they are vulnerable to adversarial\nattacks that seek to misguide the models into predicting incorrect classes. Our\nstudy of major adversarial attack models shows that they all specifically\ntarget and exploit the neural networking structures in their designs. This\nunderstanding makes us develop a hypothesis that most classical machine\nlearning models, such as Random Forest (RF), are immune to adversarial attack\nmodels because they do not rely on neural network design at all. Our\nexperimental study of classical machine learning models against popular\nadversarial attacks supports this hypothesis. Based on this hypothesis, we\npropose a new adversarial-aware deep learning system by using a classical\nmachine learning model as the secondary verification system to complement the\nprimary deep learning model in image classification. Although the secondary\nclassical machine learning model has less accurate output, it is only used for\nverification purposes, which does not impact the output accuracy of the primary\ndeep learning model, and at the same time, can effectively detect an\nadversarial attack when a clear mismatch occurs. Our experiments based on\nCIFAR-100 dataset show that our proposed approach outperforms current\nstate-of-the-art adversarial defense systems.\n", "date": "2023-06-01", "categories": []}
{"id": "2306.00687", "title": "Adversarial Robustness in Unsupervised Machine Learning: A Systematic\n  Review", "abstract": "  As the adoption of machine learning models increases, ensuring robust models\nagainst adversarial attacks is increasingly important. With unsupervised\nmachine learning gaining more attention, ensuring it is robust against attacks\nis vital. This paper conducts a systematic literature review on the robustness\nof unsupervised learning, collecting 86 papers. Our results show that most\nresearch focuses on privacy attacks, which have effective defenses; however,\nmany attacks lack effective and general defensive measures. Based on the\nresults, we formulate a model on the properties of an attack on unsupervised\nlearning, contributing to future research by providing a model to use.\n", "date": "2023-06-01", "categories": []}
{"id": "2206.14772", "title": "IBP Regularization for Verified Adversarial Robustness via\n  Branch-and-Bound", "abstract": "  Recent works have tried to increase the verifiability of adversarially\ntrained networks by running the attacks over domains larger than the original\nperturbations and adding various regularization terms to the objective.\nHowever, these algorithms either underperform or require complex and expensive\nstage-wise training procedures, hindering their practical applicability. We\npresent IBP-R, a novel verified training algorithm that is both simple and\neffective. IBP-R induces network verifiability by coupling adversarial attacks\non enlarged domains with a regularization term, based on inexpensive interval\nbound propagation, that minimizes the gap between the non-convex verification\nproblem and its approximations. By leveraging recent branch-and-bound\nframeworks, we show that IBP-R obtains state-of-the-art verified\nrobustness-accuracy trade-offs for small perturbations on CIFAR-10 while\ntraining significantly faster than relevant previous work. Additionally, we\npresent UPB, a novel branching strategy that, relying on a simple heuristic\nbased on $\\beta$-CROWN, reduces the cost of state-of-the-art branching\nalgorithms while yielding splits of comparable quality.\n", "date": "2022-06-29", "categories": []}
{"id": "2302.10896", "title": "IB-RAR: Information Bottleneck as Regularizer for Adversarial Robustness", "abstract": "  In this paper, we propose a novel method, IB-RAR, which uses Information\nBottleneck (IB) to strengthen adversarial robustness for both adversarial\ntraining and non-adversarial-trained methods. We first use the IB theory to\nbuild regularizers as learning objectives in the loss function. Then, we filter\nout unnecessary features of intermediate representation according to their\nmutual information (MI) with labels, as the network trained with IB provides\neasily distinguishable MI for its features. Experimental results show that our\nmethod can be naturally combined with adversarial training and provides\nconsistently better accuracy on new adversarial examples. Our method improves\nthe accuracy by an average of 3.07% against five adversarial attacks for the\nVGG16 network, trained with three adversarial training benchmarks and the\nCIFAR-10 dataset. In addition, our method also provides good robustness for\nundefended methods, such as training with cross-entropy loss only. Finally, in\nthe absence of adversarial training, the VGG16 network trained using our method\nand the CIFAR-10 dataset reaches an accuracy of 35.86% against PGD examples,\nwhile using all layers reaches 25.61% accuracy.\n", "date": "2023-02-09", "categories": []}
{"id": "2305.19593", "title": "Exploring the Vulnerabilities of Machine Learning and Quantum Machine\n  Learning to Adversarial Attacks using a Malware Dataset: A Comparative\n  Analysis", "abstract": "  The burgeoning fields of machine learning (ML) and quantum machine learning\n(QML) have shown remarkable potential in tackling complex problems across\nvarious domains. However, their susceptibility to adversarial attacks raises\nconcerns when deploying these systems in security sensitive applications. In\nthis study, we present a comparative analysis of the vulnerability of ML and\nQML models, specifically conventional neural networks (NN) and quantum neural\nnetworks (QNN), to adversarial attacks using a malware dataset. We utilize a\nsoftware supply chain attack dataset known as ClaMP and develop two distinct\nmodels for QNN and NN, employing Pennylane for quantum implementations and\nTensorFlow and Keras for traditional implementations. Our methodology involves\ncrafting adversarial samples by introducing random noise to a small portion of\nthe dataset and evaluating the impact on the models performance using accuracy,\nprecision, recall, and F1 score metrics. Based on our observations, both ML and\nQML models exhibit vulnerability to adversarial attacks. While the QNNs\naccuracy decreases more significantly compared to the NN after the attack, it\ndemonstrates better performance in terms of precision and recall, indicating\nhigher resilience in detecting true positives under adversarial conditions. We\nalso find that adversarial samples crafted for one model type can impair the\nperformance of the other, highlighting the need for robust defense mechanisms.\nOur study serves as a foundation for future research focused on enhancing the\nsecurity and resilience of ML and QML models, particularly QNN, given its\nrecent advancements. A more extensive range of experiments will be conducted to\nbetter understand the performance and robustness of both models in the face of\nadversarial attacks.\n", "date": "2023-05-31", "categories": []}
{"id": "2305.19607", "title": "Adversarial Clean Label Backdoor Attacks and Defenses on Text\n  Classification Systems", "abstract": "  Clean-label (CL) attack is a form of data poisoning attack where an adversary\nmodifies only the textual input of the training data, without requiring access\nto the labeling function. CL attacks are relatively unexplored in NLP, as\ncompared to label flipping (LF) attacks, where the latter additionally requires\naccess to the labeling function as well. While CL attacks are more resilient to\ndata sanitization and manual relabeling methods than LF attacks, they often\ndemand as high as ten times the poisoning budget than LF attacks. In this work,\nwe first introduce an Adversarial Clean Label attack which can adversarially\nperturb in-class training examples for poisoning the training set. We then show\nthat an adversary can significantly bring down the data requirements for a CL\nattack, using the aforementioned approach, to as low as 20% of the data\notherwise required. We then systematically benchmark and analyze a number of\ndefense methods, for both LF and CL attacks, some previously employed solely\nfor LF attacks in the textual domain and others adapted from computer vision.\nWe find that text-specific defenses greatly vary in their effectiveness\ndepending on their properties.\n", "date": "2023-05-31", "categories": []}
{"id": "2210.00062", "title": "Learning Robust Kernel Ensembles with Kernel Average Pooling", "abstract": "  Model ensembles have long been used in machine learning to reduce the\nvariance in individual model predictions, making them more robust to input\nperturbations. Pseudo-ensemble methods like dropout have also been commonly\nused in deep learning models to improve generalization. However, the\napplication of these techniques to improve neural networks' robustness against\ninput perturbations remains underexplored. We introduce Kernel Average Pooling\n(KAP), a neural network building block that applies the mean filter along the\nkernel dimension of the layer activation tensor. We show that ensembles of\nkernels with similar functionality naturally emerge in convolutional neural\nnetworks equipped with KAP and trained with backpropagation. Moreover, we show\nthat when trained on inputs perturbed with additive Gaussian noise, KAP models\nare remarkably robust against various forms of adversarial attacks. Empirical\nevaluations on CIFAR10, CIFAR100, TinyImagenet, and Imagenet datasets show\nsubstantial improvements in robustness against strong adversarial attacks such\nas AutoAttack without training on any adversarial examples.\n", "date": "2022-09-30", "categories": []}
{"id": "2305.18384", "title": "Backdoor Attacks Against Incremental Learners: An Empirical Evaluation\n  Study", "abstract": "  Large amounts of incremental learning algorithms have been proposed to\nalleviate the catastrophic forgetting issue arises while dealing with\nsequential data on a time series. However, the adversarial robustness of\nincremental learners has not been widely verified, leaving potential security\nrisks. Specifically, for poisoning-based backdoor attacks, we argue that the\nnature of streaming data in IL provides great convenience to the adversary by\ncreating the possibility of distributed and cross-task attacks -- an adversary\ncan affect \\textbf{any unknown} previous or subsequent task by data poisoning\n\\textbf{at any time or time series} with extremely small amount of backdoor\nsamples injected (e.g., $0.1\\%$ based on our observations). To attract the\nattention of the research community, in this paper, we empirically reveal the\nhigh vulnerability of 11 typical incremental learners against poisoning-based\nbackdoor attack on 3 learning scenarios, especially the cross-task\ngeneralization effect of backdoor knowledge, while the poison ratios range from\n$5\\%$ to as low as $0.1\\%$. Finally, the defense mechanism based on activation\nclustering is found to be effective in detecting our trigger pattern to\nmitigate potential security risks.\n", "date": "2023-05-28", "categories": []}
{"id": "2305.18585", "title": "Exploiting Explainability to Design Adversarial Attacks and Evaluate\n  Attack Resilience in Hate-Speech Detection Models", "abstract": "  The advent of social media has given rise to numerous ethical challenges,\nwith hate speech among the most significant concerns. Researchers are\nattempting to tackle this problem by leveraging hate-speech detection and\nemploying language models to automatically moderate content and promote civil\ndiscourse. Unfortunately, recent studies have revealed that hate-speech\ndetection systems can be misled by adversarial attacks, raising concerns about\ntheir resilience. While previous research has separately addressed the\nrobustness of these models under adversarial attacks and their\ninterpretability, there has been no comprehensive study exploring their\nintersection. The novelty of our work lies in combining these two critical\naspects, leveraging interpretability to identify potential vulnerabilities and\nenabling the design of targeted adversarial attacks. We present a comprehensive\nand comparative analysis of adversarial robustness exhibited by various\nhate-speech detection models. Our study evaluates the resilience of these\nmodels against adversarial attacks using explainability techniques. To gain\ninsights into the models' decision-making processes, we employ the Local\nInterpretable Model-agnostic Explanations (LIME) framework. Based on the\nexplainability results obtained by LIME, we devise and execute targeted attacks\non the text by leveraging the TextAttack tool. Our findings enhance the\nunderstanding of the vulnerabilities and strengths exhibited by\nstate-of-the-art hate-speech detection models. This work underscores the\nimportance of incorporating explainability in the development and evaluation of\nsuch models to enhance their resilience against adversarial attacks.\nUltimately, this work paves the way for creating more robust and reliable\nhate-speech detection systems, fostering safer online environments and\npromoting ethical discourse on social media platforms.\n", "date": "2023-05-29", "categories": []}
{"id": "2305.18933", "title": "A Multilingual Evaluation of NER Robustness to Adversarial Inputs", "abstract": "  Adversarial evaluations of language models typically focus on English alone.\nIn this paper, we performed a multilingual evaluation of Named Entity\nRecognition (NER) in terms of its robustness to small perturbations in the\ninput. Our results showed the NER models we explored across three languages\n(English, German and Hindi) are not very robust to such changes, as indicated\nby the fluctuations in the overall F1 score as well as in a more fine-grained\nevaluation. With that knowledge, we further explored whether it is possible to\nimprove the existing NER models using a part of the generated adversarial data\nsets as augmented training data to train a new NER model or as fine-tuning data\nto adapt an existing NER model. Our results showed that both these approaches\nimprove performance on the original as well as adversarial test sets. While\nthere is no significant difference between the two approaches for English,\nre-training is significantly better than fine-tuning for German and Hindi.\n", "date": "2023-05-30", "categories": []}
{"id": "2305.19218", "title": "Adversarial Attacks on Online Learning to Rank with Stochastic Click\n  Models", "abstract": "  We propose the first study of adversarial attacks on online learning to rank.\nThe goal of the adversary is to misguide the online learning to rank algorithm\nto place the target item on top of the ranking list linear times to time\nhorizon $T$ with a sublinear attack cost. We propose generalized list poisoning\nattacks that perturb the ranking list presented to the user. This strategy can\nefficiently attack any no-regret ranker in general stochastic click models.\nFurthermore, we propose a click poisoning-based strategy named attack-then-quit\nthat can efficiently attack two representative OLTR algorithms for stochastic\nclick models. We theoretically analyze the success and cost upper bound of the\ntwo proposed methods. Experimental results based on synthetic and real-world\ndata further validate the effectiveness and cost-efficiency of the proposed\nattack strategies.\n", "date": "2023-05-30", "categories": []}
{"id": "2210.13710", "title": "Motif-Backdoor: Rethinking the Backdoor Attack on Graph Neural Networks\n  via Motifs", "abstract": "  Graph neural network (GNN) with a powerful representation capability has been\nwidely applied to various areas, such as biological gene prediction, social\nrecommendation, etc. Recent works have exposed that GNN is vulnerable to the\nbackdoor attack, i.e., models trained with maliciously crafted training samples\nare easily fooled by patched samples. Most of the proposed studies launch the\nbackdoor attack using a trigger that either is the randomly generated subgraph\n(e.g., erd\\H{o}s-r\\'enyi backdoor) for less computational burden, or the\ngradient-based generative subgraph (e.g., graph trojaning attack) to enable a\nmore effective attack. However, the interpretation of how is the trigger\nstructure and the effect of the backdoor attack related has been overlooked in\nthe current literature. Motifs, recurrent and statistically significant\nsub-graphs in graphs, contain rich structure information. In this paper, we are\nrethinking the trigger from the perspective of motifs, and propose a\nmotif-based backdoor attack, denoted as Motif-Backdoor. It contributes from\nthree aspects. (i) Interpretation: it provides an in-depth explanation for\nbackdoor effectiveness by the validity of the trigger structure from motifs,\nleading to some novel insights, e.g., using subgraphs that appear less\nfrequently in the graph as the trigger can achieve better attack performance.\n(ii) Effectiveness: Motif-Backdoor reaches the state-of-the-art (SOTA) attack\nperformance in both black-box and defensive scenarios. (iii) Efficiency: based\non the graph motif distribution, Motif-Backdoor can quickly obtain an effective\ntrigger structure without target model feedback or subgraph model generation.\nExtensive experimental results show that Motif-Backdoor realizes the SOTA\nperformance on three popular models and four public datasets compared with five\nbaselines.\n", "date": "2022-10-25", "categories": []}
{"id": "2211.15762", "title": "Understanding the Impact of Adversarial Robustness on Accuracy Disparity", "abstract": "  While it has long been empirically observed that adversarial robustness may\nbe at odds with standard accuracy and may have further disparate impacts on\ndifferent classes, it remains an open question to what extent such observations\nhold and how the class imbalance plays a role within. In this paper, we attempt\nto understand this question of accuracy disparity by taking a closer look at\nlinear classifiers under a Gaussian mixture model. We decompose the impact of\nadversarial robustness into two parts: an inherent effect that will degrade the\nstandard accuracy on all classes due to the robustness constraint, and the\nother caused by the class imbalance ratio, which will increase the accuracy\ndisparity compared to standard training. Furthermore, we also show that such\neffects extend beyond the Gaussian mixture model, by generalizing our data\nmodel to the general family of stable distributions. More specifically, we\ndemonstrate that while the constraint of adversarial robustness consistently\ndegrades the standard accuracy in the balanced class setting, the class\nimbalance ratio plays a fundamentally different role in accuracy disparity\ncompared to the Gaussian case, due to the heavy tail of the stable\ndistribution. We additionally perform experiments on both synthetic and\nreal-world datasets to corroborate our theoretical findings. Our empirical\nresults also suggest that the implications may extend to nonlinear models over\nreal-world datasets. Our code is publicly available on GitHub at\nhttps://github.com/Accuracy-Disparity/AT-on-AD.\n", "date": "2022-11-28", "categories": []}
{"id": "2302.01375", "title": "On the Robustness of Randomized Ensembles to Adversarial Perturbations", "abstract": "  Randomized ensemble classifiers (RECs), where one classifier is randomly\nselected during inference, have emerged as an attractive alternative to\ntraditional ensembling methods for realizing adversarially robust classifiers\nwith limited compute requirements. However, recent works have shown that\nexisting methods for constructing RECs are more vulnerable than initially\nclaimed, casting major doubts on their efficacy and prompting fundamental\nquestions such as: \"When are RECs useful?\", \"What are their limits?\", and \"How\ndo we train them?\". In this work, we first demystify RECs as we derive\nfundamental results regarding their theoretical limits, necessary and\nsufficient conditions for them to be useful, and more. Leveraging this new\nunderstanding, we propose a new boosting algorithm (BARRE) for training robust\nRECs, and empirically demonstrate its effectiveness at defending against strong\n$\\ell_\\infty$ norm-bounded adversaries across various network architectures and\ndatasets. Our code can be found at https://github.com/hsndbk4/BARRE.\n", "date": "2023-02-02", "categories": []}
{"id": "2302.04237", "title": "Black Box Adversarial Prompting for Foundation Models", "abstract": "  Prompting interfaces allow users to quickly adjust the output of generative\nmodels in both vision and language. However, small changes and design choices\nin the prompt can lead to significant differences in the output. In this work,\nwe develop a black-box framework for generating adversarial prompts for\nunstructured image and text generation. These prompts, which can be standalone\nor prepended to benign prompts, induce specific behaviors into the generative\nprocess, such as generating images of a particular object or generating high\nperplexity text.\n", "date": "2023-02-08", "categories": []}
{"id": "2305.00075", "title": "On the existence of solutions to adversarial training in multiclass\n  classification", "abstract": "  We study three models of the problem of adversarial training in multiclass\nclassification designed to construct robust classifiers against adversarial\nperturbations of data in the agnostic-classifier setting. We prove the\nexistence of Borel measurable robust classifiers in each model and provide a\nunified perspective of the adversarial training problem, expanding the\nconnections with optimal transport initiated by the authors in previous work\nand developing new connections between adversarial training in the multiclass\nsetting and total variation regularization. As a corollary of our results, we\nprove the existence of Borel measurable solutions to the agnostic adversarial\ntraining problem in the binary classification setting, a result that improves\nresults in the literature of adversarial training, where robust classifiers\nwere only known to exist within the enlarged universal $\\sigma$-algebra of the\nfeature space.\n", "date": "2023-04-28", "categories": []}
{"id": "2305.17440", "title": "Modeling Adversarial Attack on Pre-trained Language Models as Sequential\n  Decision Making", "abstract": "  Pre-trained language models (PLMs) have been widely used to underpin various\ndownstream tasks. However, the adversarial attack task has found that PLMs are\nvulnerable to small perturbations. Mainstream methods adopt a detached\ntwo-stage framework to attack without considering the subsequent influence of\nsubstitution at each step. In this paper, we formally model the adversarial\nattack task on PLMs as a sequential decision-making problem, where the whole\nattack process is sequential with two decision-making problems, i.e., word\nfinder and word substitution. Considering the attack process can only receive\nthe final state without any direct intermediate signals, we propose to use\nreinforcement learning to find an appropriate sequential attack path to\ngenerate adversaries, named SDM-Attack. Extensive experimental results show\nthat SDM-Attack achieves the highest attack success rate with a comparable\nmodification rate and semantic similarity to attack fine-tuned BERT.\nFurthermore, our analyses demonstrate the generalization and transferability of\nSDM-Attack. The code is available at https://github.com/fduxuan/SDM-Attack.\n", "date": "2023-05-27", "categories": []}
{"id": "2305.17444", "title": "Query-Efficient Black-Box Red Teaming via Bayesian Optimization", "abstract": "  The deployment of large-scale generative models is often restricted by their\npotential risk of causing harm to users in unpredictable ways. We focus on the\nproblem of black-box red teaming, where a red team generates test cases and\ninteracts with the victim model to discover a diverse set of failures with\nlimited query access. Existing red teaming methods construct test cases based\non human supervision or language model (LM) and query all test cases in a\nbrute-force manner without incorporating any information from past evaluations,\nresulting in a prohibitively large number of queries. To this end, we propose\nBayesian red teaming (BRT), novel query-efficient black-box red teaming methods\nbased on Bayesian optimization, which iteratively identify diverse positive\ntest cases leading to model failures by utilizing the pre-defined user input\npool and the past evaluations. Experimental results on various user input pools\ndemonstrate that our method consistently finds a significantly larger number of\ndiverse positive test cases under the limited query budget than the baseline\nmethods. The source code is available at\nhttps://github.com/snu-mllab/Bayesian-Red-Teaming.\n", "date": "2023-05-27", "categories": []}
{"id": "2305.17688", "title": "Amplification trojan network: Attack deep neural networks by amplifying\n  their inherent weakness", "abstract": "  Recent works found that deep neural networks (DNNs) can be fooled by\nadversarial examples, which are crafted by adding adversarial noise on clean\ninputs. The accuracy of DNNs on adversarial examples will decrease as the\nmagnitude of the adversarial noise increase. In this study, we show that DNNs\ncan be also fooled when the noise is very small under certain circumstances.\nThis new type of attack is called Amplification Trojan Attack (ATAttack).\nSpecifically, we use a trojan network to transform the inputs before sending\nthem to the target DNN. This trojan network serves as an amplifier to amplify\nthe inherent weakness of the target DNN. The target DNN, which is infected by\nthe trojan network, performs normally on clean data while being more vulnerable\nto adversarial examples. Since it only transforms the inputs, the trojan\nnetwork can hide in DNN-based pipelines, e.g. by infecting the pre-processing\nprocedure of the inputs before sending them to the DNNs. This new type of\nthreat should be considered in developing safe DNNs.\n", "date": "2023-05-28", "categories": []}
{"id": "2305.17826", "title": "NOTABLE: Transferable Backdoor Attacks Against Prompt-based NLP Models", "abstract": "  Prompt-based learning is vulnerable to backdoor attacks. Existing backdoor\nattacks against prompt-based models consider injecting backdoors into the\nentire embedding layers or word embedding vectors. Such attacks can be easily\naffected by retraining on downstream tasks and with different prompting\nstrategies, limiting the transferability of backdoor attacks. In this work, we\npropose transferable backdoor attacks against prompt-based models, called\nNOTABLE, which is independent of downstream tasks and prompting strategies.\nSpecifically, NOTABLE injects backdoors into the encoders of PLMs by utilizing\nan adaptive verbalizer to bind triggers to specific words (i.e., anchors). It\nactivates the backdoor by pasting input with triggers to reach\nadversary-desired anchors, achieving independence from downstream tasks and\nprompting strategies. We conduct experiments on six NLP tasks, three popular\nmodels, and three prompting strategies. Empirical results show that NOTABLE\nachieves superior attack performance (i.e., attack success rate over 90% on all\nthe datasets), and outperforms two state-of-the-art baselines. Evaluations on\nthree defenses show the robustness of NOTABLE. Our code can be found at\nhttps://github.com/RU-System-Software-and-Security/Notable.\n", "date": "2023-05-28", "categories": []}
{"id": "2305.16444", "title": "Don't Retrain, Just Rewrite: Countering Adversarial Perturbations by\n  Rewriting Text", "abstract": "  Can language models transform inputs to protect text classifiers against\nadversarial attacks? In this work, we present ATINTER, a model that intercepts\nand learns to rewrite adversarial inputs to make them non-adversarial for a\ndownstream text classifier. Our experiments on four datasets and five attack\nmechanisms reveal that ATINTER is effective at providing better adversarial\nrobustness than existing defense approaches, without compromising task\naccuracy. For example, on sentiment classification using the SST-2 dataset, our\nmethod improves the adversarial accuracy over the best existing defense\napproach by more than 4% with a smaller decrease in task accuracy (0.5% vs\n2.5%). Moreover, we show that ATINTER generalizes across multiple downstream\ntasks and classifiers without having to explicitly retrain it for those\nsettings. Specifically, we find that when ATINTER is trained to remove\nadversarial perturbations for the sentiment classification task on the SST-2\ndataset, it even transfers to a semantically different task of news\nclassification (on AGNews) and improves the adversarial robustness by more than\n10%.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.16503", "title": "IMBERT: Making BERT Immune to Insertion-based Backdoor Attacks", "abstract": "  Backdoor attacks are an insidious security threat against machine learning\nmodels. Adversaries can manipulate the predictions of compromised models by\ninserting triggers into the training phase. Various backdoor attacks have been\ndevised which can achieve nearly perfect attack success without affecting model\npredictions for clean inputs. Means of mitigating such vulnerabilities are\nunderdeveloped, especially in natural language processing. To fill this gap, we\nintroduce IMBERT, which uses either gradients or self-attention scores derived\nfrom victim models to self-defend against backdoor attacks at inference time.\nOur empirical studies demonstrate that IMBERT can effectively identify up to\n98.5% of inserted triggers. Thus, it significantly reduces the attack success\nrate while attaining competitive accuracy on the clean dataset across\nwidespread insertion-based attacks compared to two baselines. Finally, we show\nthat our approach is model-agnostic, and can be easily ported to several\npre-trained transformer models.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.17071", "title": "Adversarial Attacks on Online Learning to Rank with Click Feedback", "abstract": "  Online learning to rank (OLTR) is a sequential decision-making problem where\na learning agent selects an ordered list of items and receives feedback through\nuser clicks. Although potential attacks against OLTR algorithms may cause\nserious losses in real-world applications, little is known about adversarial\nattacks on OLTR. This paper studies attack strategies against multiple variants\nof OLTR. Our first result provides an attack strategy against the UCB algorithm\non classical stochastic bandits with binary feedback, which solves the key\nissues caused by bounded and discrete feedback that previous works can not\nhandle. Building on this result, we design attack algorithms against UCB-based\nOLTR algorithms in position-based and cascade models. Finally, we propose a\ngeneral attack strategy against any algorithm under the general click model.\nEach attack algorithm manipulates the learning agent into choosing the target\nattack item $T-o(T)$ times, incurring a cumulative cost of $o(T)$. Experiments\non synthetic and real data further validate the effectiveness of our proposed\nattack algorithms.\n", "date": "2023-05-26", "categories": []}
{"id": "2211.05656", "title": "On Proper Learnability between Average- and Worst-case Robustness", "abstract": "  Recently, Montasser et al. [2019] showed that finite VC dimension is not\nsufficient for proper adversarially robust PAC learning. In light of this\nhardness, there is a growing effort to study what type of relaxations to the\nadversarially robust PAC learning setup can enable proper learnability. In this\nwork, we initiate the study of proper learning under relaxations of the\nworst-case robust loss. We give a family of robust loss relaxations under which\nVC classes are properly PAC learnable with sample complexity close to what one\nwould require in the standard PAC learning setup. On the other hand, we show\nthat for an existing and natural relaxation of the worst-case robust loss,\nfinite VC dimension is not sufficient for proper learning. Lastly, we give new\ngeneralization guarantees for the adversarially robust empirical risk\nminimizer.\n", "date": "2022-11-10", "categories": []}
{"id": "2302.02907", "title": "GAT: Guided Adversarial Training with Pareto-optimal Auxiliary Tasks", "abstract": "  While leveraging additional training data is well established to improve\nadversarial robustness, it incurs the unavoidable cost of data collection and\nthe heavy computation to train models. To mitigate the costs, we propose Guided\nAdversarial Training (GAT), a novel adversarial training technique that\nexploits auxiliary tasks under a limited set of training data. Our approach\nextends single-task models into multi-task models during the min-max\noptimization of adversarial training, and drives the loss optimization with a\nregularization of the gradient curvature across multiple tasks. GAT leverages\ntwo types of auxiliary tasks: self-supervised tasks, where the labels are\ngenerated automatically, and domain-knowledge tasks, where human experts\nprovide additional labels. Experimentally, GAT increases the robust AUC of\nCheXpert medical imaging dataset from 50% to 83% and On CIFAR-10, GAT\noutperforms eight state-of-the-art adversarial training and achieves 56.21%\nrobust accuracy with Resnet-50. Overall, we demonstrate that guided multi-task\nlearning is an actionable and promising avenue to push further the boundaries\nof model robustness.\n", "date": "2023-02-06", "categories": []}
{"id": "2305.14700", "title": "AdvFunMatch: When Consistent Teaching Meets Adversarial Robustness", "abstract": "  \\emph{Consistent teaching} is an effective paradigm for implementing\nknowledge distillation (KD), where both student and teacher models receive\nidentical inputs, and KD is treated as a function matching task (FunMatch).\nHowever, one limitation of FunMatch is that it does not account for the\ntransfer of adversarial robustness, a model's resistance to adversarial\nattacks. To tackle this problem, we propose a simple but effective strategy\ncalled Adversarial Function Matching (AdvFunMatch), which aims to match\ndistributions for all data points within the $\\ell_p$-norm ball of the training\ndata, in accordance with consistent teaching. Formulated as a min-max\noptimization problem, AdvFunMatch identifies the worst-case instances that\nmaximizes the KL-divergence between teacher and student model outputs, which we\nrefer to as \"mismatched examples,\" and then matches the outputs on these\nmismatched examples. Our experimental results show that AdvFunMatch effectively\nproduces student models with both high clean accuracy and robustness.\nFurthermore, we reveal that strong data augmentations (\\emph{e.g.},\nAutoAugment) are beneficial in AdvFunMatch, whereas prior works have found them\nless effective in adversarial training. Code is available at\n\\url{https://gitee.com/zihui998/adv-fun-match}.\n", "date": "2023-05-24", "categories": []}
{"id": "2305.15544", "title": "Fast Adversarial CNN-based Perturbation Attack on No-Reference Image-\n  and Video-Quality Metrics", "abstract": "  Modern neural-network-based no-reference image- and video-quality metrics\nexhibit performance as high as full-reference metrics. These metrics are widely\nused to improve visual quality in computer vision methods and compare video\nprocessing methods. However, these metrics are not stable to traditional\nadversarial attacks, which can cause incorrect results. Our goal is to\ninvestigate the boundaries of no-reference metrics applicability, and in this\npaper, we propose a fast adversarial perturbation attack on no-reference\nquality metrics. The proposed attack (FACPA) can be exploited as a\npreprocessing step in real-time video processing and compression algorithms.\nThis research can yield insights to further aid in designing of stable\nneural-network-based no-reference quality metrics.\n", "date": "2023-05-24", "categories": []}
{"id": "2305.15587", "title": "How do humans perceive adversarial text? A reality check on the validity\n  and naturalness of word-based adversarial attacks", "abstract": "  Natural Language Processing (NLP) models based on Machine Learning (ML) are\nsusceptible to adversarial attacks -- malicious algorithms that imperceptibly\nmodify input text to force models into making incorrect predictions. However,\nevaluations of these attacks ignore the property of imperceptibility or study\nit under limited settings. This entails that adversarial perturbations would\nnot pass any human quality gate and do not represent real threats to\nhuman-checked NLP systems. To bypass this limitation and enable proper\nassessment (and later, improvement) of NLP model robustness, we have surveyed\n378 human participants about the perceptibility of text adversarial examples\nproduced by state-of-the-art methods. Our results underline that existing text\nattacks are impractical in real-world scenarios where humans are involved. This\ncontrasts with previous smaller-scale human studies, which reported overly\noptimistic conclusions regarding attack success. Through our work, we hope to\nposition human perceptibility as a first-class success criterion for text\nattacks, and provide guidance for research to build effective attack algorithms\nand, in turn, design appropriate defence mechanisms.\n", "date": "2023-05-24", "categories": []}
{"id": "2305.15709", "title": "PEARL: Preprocessing Enhanced Adversarial Robust Learning of Image\n  Deraining for Semantic Segmentation", "abstract": "  In light of the significant progress made in the development and application\nof semantic segmentation tasks, there has been increasing attention towards\nimproving the robustness of segmentation models against natural degradation\nfactors (e.g., rain streaks) or artificially attack factors (e.g., adversarial\nattack). Whereas, most existing methods are designed to address a single\ndegradation factor and are tailored to specific application scenarios. In this\nwork, we present the first attempt to improve the robustness of semantic\nsegmentation tasks by simultaneously handling different types of degradation\nfactors. Specifically, we introduce the Preprocessing Enhanced Adversarial\nRobust Learning (PEARL) framework based on the analysis of our proposed Naive\nAdversarial Training (NAT) framework. Our approach effectively handles both\nrain streaks and adversarial perturbation by transferring the robustness of the\nsegmentation model to the image derain model. Furthermore, as opposed to the\ncommonly used Negative Adversarial Attack (NAA), we design the Auxiliary Mirror\nAttack (AMA) to introduce positive information prior to the training of the\nPEARL framework, which improves defense capability and segmentation\nperformance. Our extensive experiments and ablation studies based on different\nderain methods and segmentation models have demonstrated the significant\nperformance improvement of PEARL with AMA in defense against various\nadversarial attacks and rain streaks while maintaining high generalization\nperformance across different datasets.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.16035", "title": "Detecting Adversarial Data by Probing Multiple Perturbations Using\n  Expected Perturbation Score", "abstract": "  Adversarial detection aims to determine whether a given sample is an\nadversarial one based on the discrepancy between natural and adversarial\ndistributions. Unfortunately, estimating or comparing two data distributions is\nextremely difficult, especially in high-dimension spaces. Recently, the\ngradient of log probability density (a.k.a., score) w.r.t. the sample is used\nas an alternative statistic to compute. However, we find that the score is\nsensitive in identifying adversarial samples due to insufficient information\nwith one sample only. In this paper, we propose a new statistic called expected\nperturbation score (EPS), which is essentially the expected score of a sample\nafter various perturbations. Specifically, to obtain adequate information\nregarding one sample, we perturb it by adding various noises to capture its\nmulti-view observations. We theoretically prove that EPS is a proper statistic\nto compute the discrepancy between two samples under mild conditions. In\npractice, we can use a pre-trained diffusion model to estimate EPS for each\nsample. Last, we propose an EPS-based adversarial detection (EPS-AD) method, in\nwhich we develop EPS-based maximum mean discrepancy (MMD) as a metric to\nmeasure the discrepancy between the test sample and natural samples. We also\nprove that the EPS-based MMD between natural and adversarial samples is larger\nthan that among natural samples. Extensive experiments show the superior\nadversarial detection performance of our EPS-AD.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.16220", "title": "On the Robustness of Segment Anything", "abstract": "  Segment anything model (SAM) has presented impressive objectness\nidentification capability with the idea of prompt learning and a new collected\nlarge-scale dataset. Given a prompt (e.g., points, bounding boxes, or masks)\nand an input image, SAM is able to generate valid segment masks for all objects\nindicated by the prompts, presenting high generalization across diverse\nscenarios and being a general method for zero-shot transfer to downstream\nvision tasks. Nevertheless, it remains unclear whether SAM may introduce errors\nin certain threatening scenarios. Clarifying this is of significant importance\nfor applications that require robustness, such as autonomous vehicles. In this\npaper, we aim to study the testing-time robustness of SAM under adversarial\nscenarios and common corruptions. To this end, we first build a testing-time\nrobustness evaluation benchmark for SAM by integrating existing public\ndatasets. Second, we extend representative adversarial attacks against SAM and\nstudy the influence of different prompts on robustness. Third, we study the\nrobustness of SAM under diverse corruption types by evaluating SAM on corrupted\ndatasets with different prompts. With experiments conducted on SA-1B and KITTI\ndatasets, we find that SAM exhibits remarkable robustness against various\ncorruptions, except for blur-related corruption. Furthermore, SAM remains\nsusceptible to adversarial attacks, particularly when subjected to PGD and BIM\nattacks. We think such a comprehensive study could highlight the importance of\nthe robustness issues of SAM and trigger a series of new tasks for SAM as well\nas downstream vision tasks.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.16310", "title": "Securing Deep Generative Models with Universal Adversarial Signature", "abstract": "  Recent advances in deep generative models have led to the development of\nmethods capable of synthesizing high-quality, realistic images. These models\npose threats to society due to their potential misuse. Prior research attempted\nto mitigate these threats by detecting generated images, but the varying traces\nleft by different generative models make it challenging to create a universal\ndetector capable of generalizing to new, unseen generative models. In this\npaper, we propose to inject a universal adversarial signature into an arbitrary\npre-trained generative model, in order to make its generated contents more\ndetectable and traceable. First, the imperceptible optimal signature for each\nimage can be found by a signature injector through adversarial training.\nSubsequently, the signature can be incorporated into an arbitrary generator by\nfine-tuning it with the images processed by the signature injector. In this\nway, the detector corresponding to the signature can be reused for any\nfine-tuned generator for tracking the generator identity. The proposed method\nis validated on the FFHQ and ImageNet datasets with various state-of-the-art\ngenerative models, consistently showing a promising detection rate. Code will\nbe made publicly available at \\url{https://github.com/zengxianyu/genwm}.\n", "date": "2023-05-25", "categories": []}
{"id": "2303.10974", "title": "Translate your gibberish: black-box adversarial attack on machine\n  translation systems", "abstract": "  Neural networks are deployed widely in natural language processing tasks on\nthe industrial scale, and perhaps the most often they are used as compounds of\nautomatic machine translation systems. In this work, we present a simple\napproach to fool state-of-the-art machine translation tools in the task of\ntranslation from Russian to English and vice versa. Using a novel black-box\ngradient-free tensor-based optimizer, we show that many online translation\ntools, such as Google, DeepL, and Yandex, may both produce wrong or offensive\ntranslations for nonsensical adversarial input queries and refuse to translate\nseemingly benign input phrases. This vulnerability may interfere with\nunderstanding a new language and simply worsen the user's experience while\nusing machine translation systems, and, hence, additional improvements of these\ntools are required to establish better translation.\n", "date": "2023-03-20", "categories": []}
{"id": "2305.14553", "title": "Adversarial Machine Learning and Cybersecurity: Risks, Challenges, and\n  Legal Implications", "abstract": "  In July 2022, the Center for Security and Emerging Technology (CSET) at\nGeorgetown University and the Program on Geopolitics, Technology, and\nGovernance at the Stanford Cyber Policy Center convened a workshop of experts\nto examine the relationship between vulnerabilities in artificial intelligence\nsystems and more traditional types of software vulnerabilities. Topics\ndiscussed included the extent to which AI vulnerabilities can be handled under\nstandard cybersecurity processes, the barriers currently preventing the\naccurate sharing of information about AI vulnerabilities, legal issues\nassociated with adversarial attacks on AI systems, and potential areas where\ngovernment support could improve AI vulnerability management and mitigation.\n  This report is meant to accomplish two things. First, it provides a\nhigh-level discussion of AI vulnerabilities, including the ways in which they\nare disanalogous to other types of vulnerabilities, and the current state of\naffairs regarding information sharing and legal oversight of AI\nvulnerabilities. Second, it attempts to articulate broad recommendations as\nendorsed by the majority of participants at the workshop.\n", "date": "2023-05-23", "categories": []}
{"id": "2305.14846", "title": "Introducing Competition to Boost the Transferability of Targeted\n  Adversarial Examples through Clean Feature Mixup", "abstract": "  Deep neural networks are widely known to be susceptible to adversarial\nexamples, which can cause incorrect predictions through subtle input\nmodifications. These adversarial examples tend to be transferable between\nmodels, but targeted attacks still have lower attack success rates due to\nsignificant variations in decision boundaries. To enhance the transferability\nof targeted adversarial examples, we propose introducing competition into the\noptimization process. Our idea is to craft adversarial perturbations in the\npresence of two new types of competitor noises: adversarial perturbations\ntowards different target classes and friendly perturbations towards the correct\nclass. With these competitors, even if an adversarial example deceives a\nnetwork to extract specific features leading to the target class, this\ndisturbance can be suppressed by other competitors. Therefore, within this\ncompetition, adversarial examples should take different attack strategies by\nleveraging more diverse features to overwhelm their interference, leading to\nimproving their transferability to different models. Considering the\ncomputational complexity, we efficiently simulate various interference from\nthese two types of competitors in feature space by randomly mixing up stored\nclean features in the model inference and named this method Clean Feature Mixup\n(CFM). Our extensive experimental results on the ImageNet-Compatible and\nCIFAR-10 datasets show that the proposed method outperforms the existing\nbaselines with a clear margin. Our code is available at\nhttps://github.com/dreamflake/CFM.\n", "date": "2023-05-24", "categories": []}
{"id": "2305.14984", "title": "Adversarial robustness of amortized Bayesian inference", "abstract": "  Bayesian inference usually requires running potentially costly inference\nprocedures separately for every new observation. In contrast, the idea of\namortized Bayesian inference is to initially invest computational cost in\ntraining an inference network on simulated data, which can subsequently be used\nto rapidly perform inference (i.e., to return estimates of posterior\ndistributions) for new observations. This approach has been applied to many\nreal-world models in the sciences and engineering, but it is unclear how robust\nthe approach is to adversarial perturbations in the observed data. Here, we\nstudy the adversarial robustness of amortized Bayesian inference, focusing on\nsimulation-based estimation of multi-dimensional posterior distributions. We\nshow that almost unrecognizable, targeted perturbations of the observations can\nlead to drastic changes in the predicted posterior and highly unrealistic\nposterior predictive samples, across several benchmark tasks and a real-world\nexample from neuroscience. We propose a computationally efficient\nregularization scheme based on penalizing the Fisher information of the\nconditional density estimator, and show how it improves the adversarial\nrobustness of amortized Bayesian inference.\n", "date": "2023-05-24", "categories": []}
{"id": "2206.01034", "title": "Adversarial Laser Spot: Robust and Covert Physical-World Attack to DNNs", "abstract": "  Most existing deep neural networks (DNNs) are easily disturbed by slight\nnoise. However, there are few researches on physical attacks by deploying\nlighting equipment. The light-based physical attacks has excellent covertness,\nwhich brings great security risks to many vision-based applications (such as\nself-driving). Therefore, we propose a light-based physical attack, called\nadversarial laser spot (AdvLS), which optimizes the physical parameters of\nlaser spots through genetic algorithm to perform physical attacks. It realizes\nrobust and covert physical attack by using low-cost laser equipment. As far as\nwe know, AdvLS is the first light-based physical attack that perform physical\nattacks in the daytime. A large number of experiments in the digital and\nphysical environments show that AdvLS has excellent robustness and covertness.\nIn addition, through in-depth analysis of the experimental data, we find that\nthe adversarial perturbations generated by AdvLS have superior adversarial\nattack migration. The experimental results show that AdvLS impose serious\ninterference to advanced DNNs, we call for the attention of the proposed AdvLS.\nThe code of AdvLS is available at: https://github.com/ChengYinHu/AdvLS\n", "date": "2022-06-02", "categories": []}
{"id": "2206.12251", "title": "Adversarial Zoom Lens: A Novel Physical-World Attack to DNNs", "abstract": "  Although deep neural networks (DNNs) are known to be fragile, no one has\nstudied the effects of zooming-in and zooming-out of images in the physical\nworld on DNNs performance. In this paper, we demonstrate a novel physical\nadversarial attack technique called Adversarial Zoom Lens (AdvZL), which uses a\nzoom lens to zoom in and out of pictures of the physical world, fooling DNNs\nwithout changing the characteristics of the target object. The proposed method\nis so far the only adversarial attack technique that does not add physical\nadversarial perturbation attack DNNs. In a digital environment, we construct a\ndata set based on AdvZL to verify the antagonism of equal-scale enlarged images\nto DNNs. In the physical environment, we manipulate the zoom lens to zoom in\nand out of the target object, and generate adversarial samples. The\nexperimental results demonstrate the effectiveness of AdvZL in both digital and\nphysical environments. We further analyze the antagonism of the proposed data\nset to the improved DNNs. On the other hand, we provide a guideline for defense\nagainst AdvZL by means of adversarial training. Finally, we look into the\nthreat possibilities of the proposed approach to future autonomous driving and\nvariant attack ideas similar to the proposed attack.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.14477", "title": "Adversarial Ensemble Training by Jointly Learning Label Dependencies and\n  Member Models", "abstract": "  Training an ensemble of diverse sub-models has been empirically demonstrated\nas an effective strategy for improving the adversarial robustness of deep\nneural networks. However, current ensemble training methods for image\nrecognition typically encode image labels using one-hot vectors, which overlook\ndependency relationships between the labels. In this paper, we propose a novel\nadversarial en-semble training approach that jointly learns the label\ndependencies and member models. Our approach adaptively exploits the learned\nlabel dependencies to pro-mote diversity among the member models. We evaluate\nour approach on widely used datasets including MNIST, FashionMNIST, and\nCIFAR-10, and show that it achieves superior robustness against black-box\nattacks compared to state-of-the-art methods. Our code is available at\nhttps://github.com/ZJLAB-AMMI/LSD.\n", "date": "2022-06-29", "categories": []}
{"id": "2209.02430", "title": "Adversarial Color Film: Effective Physical-World Attack to DNNs", "abstract": "  It is well known that the performance of deep neural networks (DNNs) is\nsusceptible to subtle interference. So far, camera-based physical adversarial\nattacks haven't gotten much attention, but it is the vacancy of physical\nattack. In this paper, we propose a simple and efficient camera-based physical\nattack called Adversarial Color Film (AdvCF), which manipulates the physical\nparameters of color film to perform attacks. Carefully designed experiments\nshow the effectiveness of the proposed method in both digital and physical\nenvironments. In addition, experimental results show that the adversarial\nsamples generated by AdvCF have excellent performance in attack\ntransferability, which enables AdvCF effective black-box attacks. At the same\ntime, we give the guidance of defense against AdvCF by means of adversarial\ntraining. Finally, we look into AdvCF's threat to future vision-based systems\nand propose some promising mentality for camera-based physical attacks.\n", "date": "2022-09-02", "categories": []}
{"id": "2209.09652", "title": "Adversarial Color Projection: A Projector-based Physical Attack to DNNs", "abstract": "  Recent research has demonstrated that deep neural networks (DNNs) are\nvulnerable to adversarial perturbations. Therefore, it is imperative to\nevaluate the resilience of advanced DNNs to adversarial attacks. However,\ntraditional methods that use stickers as physical perturbations to deceive\nclassifiers face challenges in achieving stealthiness and are susceptible to\nprinting loss. Recently, advancements in physical attacks have utilized light\nbeams, such as lasers, to perform attacks, where the optical patterns generated\nare artificial rather than natural. In this work, we propose a black-box\nprojector-based physical attack, referred to as adversarial color projection\n(AdvCP), which manipulates the physical parameters of color projection to\nperform an adversarial attack. We evaluate our approach on three crucial\ncriteria: effectiveness, stealthiness, and robustness. In the digital\nenvironment, we achieve an attack success rate of 97.60% on a subset of\nImageNet, while in the physical environment, we attain an attack success rate\nof 100% in the indoor test and 82.14% in the outdoor test. The adversarial\nsamples generated by AdvCP are compared with baseline samples to demonstrate\nthe stealthiness of our approach. When attacking advanced DNNs, experimental\nresults show that our method can achieve more than 85% attack success rate in\nall cases, which verifies the robustness of AdvCP. Finally, we consider the\npotential threats posed by AdvCP to future vision-based systems and\napplications and suggest some ideas for light-based physical attacks.\n", "date": "2022-09-19", "categories": []}
{"id": "2209.11739", "title": "Adversarial Catoptric Light: An Effective, Stealthy and Robust\n  Physical-World Attack to DNNs", "abstract": "  Deep neural networks (DNNs) have demonstrated exceptional success across\nvarious tasks, underscoring the need to evaluate the robustness of advanced\nDNNs. However, traditional methods using stickers as physical perturbations to\ndeceive classifiers present challenges in achieving stealthiness and suffer\nfrom printing loss. Recent advancements in physical attacks have utilized light\nbeams such as lasers and projectors to perform attacks, where the optical\npatterns generated are artificial rather than natural. In this study, we\nintroduce a novel physical attack, adversarial catoptric light (AdvCL), where\nadversarial perturbations are generated using a common natural phenomenon,\ncatoptric light, to achieve stealthy and naturalistic adversarial attacks\nagainst advanced DNNs in a black-box setting. We evaluate the proposed method\nin three aspects: effectiveness, stealthiness, and robustness. Quantitative\nresults obtained in simulated environments demonstrate the effectiveness of the\nproposed method, and in physical scenarios, we achieve an attack success rate\nof 83.5%, surpassing the baseline. We use common catoptric light as a\nperturbation to enhance the stealthiness of the method and make physical\nsamples appear more natural. Robustness is validated by successfully attacking\nadvanced and robust DNNs with a success rate over 80% in all cases.\nAdditionally, we discuss defense strategy against AdvCL and put forward some\nlight-based physical attacks.\n", "date": "2022-09-19", "categories": []}
{"id": "2305.13548", "title": "Attribute-Guided Encryption with Facial Texture Masking", "abstract": "  The increasingly pervasive facial recognition (FR) systems raise serious\nconcerns about personal privacy, especially for billions of users who have\npublicly shared their photos on social media. Several attempts have been made\nto protect individuals from unauthorized FR systems utilizing adversarial\nattacks to generate encrypted face images to protect users from being\nidentified by FR systems. However, existing methods suffer from poor visual\nquality or low attack success rates, which limit their usability in practice.\nIn this paper, we propose Attribute Guided Encryption with Facial Texture\nMasking (AGE-FTM) that performs a dual manifold adversarial attack on FR\nsystems to achieve both good visual quality and high black box attack success\nrates. In particular, AGE-FTM utilizes a high fidelity generative adversarial\nnetwork (GAN) to generate natural on-manifold adversarial samples by modifying\nfacial attributes, and performs the facial texture masking attack to generate\nimperceptible off-manifold adversarial samples. Extensive experiments on the\nCelebA-HQ dataset demonstrate that our proposed method produces more\nnatural-looking encrypted images than state-of-the-art methods while achieving\ncompetitive attack performance. We further evaluate the effectiveness of\nAGE-FTM in the real world using a commercial FR API and validate its usefulness\nin practice through an user study.\n", "date": "2023-05-22", "categories": []}
{"id": "2305.13678", "title": "Enhancing Accuracy and Robustness through Adversarial Training in Class\n  Incremental Continual Learning", "abstract": "  In real life, adversarial attack to deep learning models is a fatal security\nissue. However, the issue has been rarely discussed in a widely used\nclass-incremental continual learning (CICL). In this paper, we address problems\nof applying adversarial training to CICL, which is well-known defense method\nagainst adversarial attack. A well-known problem of CICL is class-imbalance\nthat biases a model to the current task by a few samples of previous tasks.\nMeeting with the adversarial training, the imbalance causes another imbalance\nof attack trials over tasks. Lacking clean data of a minority class by the\nclass-imbalance and increasing of attack trials from a majority class by the\nsecondary imbalance, adversarial training distorts optimal decision boundaries.\nThe distortion eventually decreases both accuracy and robustness than\nadversarial training. To exclude the effects, we propose a straightforward but\nsignificantly effective method, External Adversarial Training (EAT) which can\nbe applied to methods using experience replay. This method conduct adversarial\ntraining to an auxiliary external model for the current task data at each time\nstep, and applies generated adversarial examples to train the target model. We\nverify the effects on a toy problem and show significance on CICL benchmarks of\nimage classification. We expect that the results will be used as the first\nbaseline for robustness research of CICL.\n", "date": "2023-05-23", "categories": []}
{"id": "2305.14188", "title": "The Best Defense is a Good Offense: Adversarial Augmentation against\n  Adversarial Attacks", "abstract": "  Many defenses against adversarial attacks (\\eg robust classifiers,\nrandomization, or image purification) use countermeasures put to work only\nafter the attack has been crafted. We adopt a different perspective to\nintroduce $A^5$ (Adversarial Augmentation Against Adversarial Attacks), a novel\nframework including the first certified preemptive defense against adversarial\nattacks. The main idea is to craft a defensive perturbation to guarantee that\nany attack (up to a given magnitude) towards the input in hand will fail. To\nthis aim, we leverage existing automatic perturbation analysis tools for neural\nnetworks. We study the conditions to apply $A^5$ effectively, analyze the\nimportance of the robustness of the to-be-defended classifier, and inspect the\nappearance of the robustified images. We show effective on-the-fly defensive\naugmentation with a robustifier network that ignores the ground truth label,\nand demonstrate the benefits of robustifier and classifier co-training. In our\ntests, $A^5$ consistently beats state of the art certified defenses on MNIST,\nCIFAR10, FashionMNIST and Tinyimagenet. We also show how to apply $A^5$ to\ncreate certifiably robust physical objects. Our code at\nhttps://github.com/NVlabs/A5 allows experimenting on a wide range of scenarios\nbeyond the man-in-the-middle attack tested here, including the case of physical\nattacks.\n", "date": "2023-05-23", "categories": []}
{"id": "2211.11534", "title": "Towards Adversarially Robust Recommendation from Adaptive Fraudster\n  Detection", "abstract": "  The robustness of recommender systems under node injection attacks has\ngarnered significant attention. Recently, GraphRfi, a GNN-based recommender\nsystem, was proposed and shown to effectively mitigate the impact of injected\nfake users. However, we demonstrate that GraphRfi remains vulnerable to attacks\ndue to the supervised nature of its fraudster detection component, where\nobtaining clean labels is challenging in practice. In particular, we propose a\npowerful poisoning attack, MetaC, against both GNN-based and MF-based\nrecommender systems. Furthermore, we analyze why GraphRfi fails under such an\nattack. Then, based on our insights obtained from vulnerability analysis, we\ndesign an adaptive fraudster detection module that explicitly considers label\nuncertainty. This module can serve as a plug-in for different recommender\nsystems, resulting in a robust framework named PDR. Comprehensive experiments\nshow that our defense approach outperforms other benchmark methods under\nattacks. Overall, our research presents an effective framework for integrating\nfraudster detection into recommendation systems to achieve adversarial\nrobustness.\n", "date": "2022-11-08", "categories": []}
{"id": "2305.12228", "title": "Dynamic Transformers Provide a False Sense of Efficiency", "abstract": "  Despite much success in natural language processing (NLP), pre-trained\nlanguage models typically lead to a high computational cost during inference.\nMulti-exit is a mainstream approach to address this issue by making a trade-off\nbetween efficiency and accuracy, where the saving of computation comes from an\nearly exit. However, whether such saving from early-exiting is robust remains\nunknown. Motivated by this, we first show that directly adapting existing\nadversarial attack approaches targeting model accuracy cannot significantly\nreduce inference efficiency. To this end, we propose a simple yet effective\nattacking framework, SAME, a novel slowdown attack framework on multi-exit\nmodels, which is specially tailored to reduce the efficiency of the multi-exit\nmodels. By leveraging the multi-exit models' design characteristics, we utilize\nall internal predictions to guide the adversarial sample generation instead of\nmerely considering the final prediction. Experiments on the GLUE benchmark show\nthat SAME can effectively diminish the efficiency gain of various multi-exit\nmodels by 80% on average, convincingly validating its effectiveness and\ngeneralization ability.\n", "date": "2023-05-20", "categories": []}
{"id": "2305.12683", "title": "Mist: Towards Improved Adversarial Examples for Diffusion Models", "abstract": "  Diffusion Models (DMs) have empowered great success in\nartificial-intelligence-generated content, especially in artwork creation, yet\nraising new concerns in intellectual properties and copyright. For example,\ninfringers can make profits by imitating non-authorized human-created paintings\nwith DMs. Recent researches suggest that various adversarial examples for\ndiffusion models can be effective tools against these copyright infringements.\nHowever, current adversarial examples show weakness in transferability over\ndifferent painting-imitating methods and robustness under straightforward\nadversarial defense, for example, noise purification. We surprisingly find that\nthe transferability of adversarial examples can be significantly enhanced by\nexploiting a fused and modified adversarial loss term under consistent\nparameters. In this work, we comprehensively evaluate the cross-method\ntransferability of adversarial examples. The experimental observation shows\nthat our method generates more transferable adversarial examples with even\nstronger robustness against the simple adversarial defense.\n", "date": "2023-05-22", "categories": []}
{"id": "2305.12770", "title": "FGAM:Fast Adversarial Malware Generation Method Based on Gradient Sign", "abstract": "  Malware detection models based on deep learning have been widely used, but\nrecent research shows that deep learning models are vulnerable to adversarial\nattacks. Adversarial attacks are to deceive the deep learning model by\ngenerating adversarial samples. When adversarial attacks are performed on the\nmalware detection model, the attacker will generate adversarial malware with\nthe same malicious functions as the malware, and make the detection model\nclassify it as benign software. Studying adversarial malware generation can\nhelp model designers improve the robustness of malware detection models. At\npresent, in the work on adversarial malware generation for byte-to-image\nmalware detection models, there are mainly problems such as large amount of\ninjection perturbation and low generation efficiency. Therefore, this paper\nproposes FGAM (Fast Generate Adversarial Malware), a method for fast generating\nadversarial malware, which iterates perturbed bytes according to the gradient\nsign to enhance adversarial capability of the perturbed bytes until the\nadversarial malware is successfully generated. It is experimentally verified\nthat the success rate of the adversarial malware deception model generated by\nFGAM is increased by about 84\\% compared with existing methods.\n", "date": "2023-05-22", "categories": []}
{"id": "2305.12863", "title": "Towards Benchmarking and Assessing Visual Naturalness of Physical World\n  Adversarial Attacks", "abstract": "  Physical world adversarial attack is a highly practical and threatening\nattack, which fools real world deep learning systems by generating conspicuous\nand maliciously crafted real world artifacts. In physical world attacks,\nevaluating naturalness is highly emphasized since human can easily detect and\nremove unnatural attacks. However, current studies evaluate naturalness in a\ncase-by-case fashion, which suffers from errors, bias and inconsistencies. In\nthis paper, we take the first step to benchmark and assess visual naturalness\nof physical world attacks, taking autonomous driving scenario as the first\nattempt. First, to benchmark attack naturalness, we contribute the first\nPhysical Attack Naturalness (PAN) dataset with human rating and gaze. PAN\nverifies several insights for the first time: naturalness is (disparately)\naffected by contextual features (i.e., environmental and semantic variations)\nand correlates with behavioral feature (i.e., gaze signal). Second, to\nautomatically assess attack naturalness that aligns with human ratings, we\nfurther introduce Dual Prior Alignment (DPA) network, which aims to embed human\nknowledge into model reasoning process. Specifically, DPA imitates human\nreasoning in naturalness assessment by rating prior alignment and mimics human\ngaze behavior by attentive prior alignment. We hope our work fosters researches\nto improve and automatically assess naturalness of physical world attacks. Our\ncode and dataset can be found at https://github.com/zhangsn-19/PAN.\n", "date": "2023-05-22", "categories": []}
{"id": "2305.12906", "title": "Latent Magic: An Investigation into Adversarial Examples Crafted in the\n  Semantic Latent Space", "abstract": "  Adversarial attacks against Deep Neural Networks(DNN) have been a crutial\ntopic ever since \\cite{goodfellow} purposed the vulnerability of DNNs. However,\nmost prior works craft adversarial examples in the pixel space, following the\n$l_p$ norm constraint. In this paper, we give intuitional explain about why\ncrafting adversarial examples in the latent space is equally efficient and\nimportant. We purpose a framework for crafting adversarial examples in semantic\nlatent space based on an pre-trained Variational Auto Encoder from state-of-art\nStable Diffusion Model\\cite{SDM}. We also show that adversarial examples\ncrafted in the latent space can also achieve a high level of fool rate.\nHowever, examples crafted from latent space are often hard to evaluated, as\nthey doesn't follow a certain $l_p$ norm constraint, which is a big challenge\nfor existing researches. To efficiently and accurately evaluate the adversarial\nexamples crafted in the latent space, we purpose \\textbf{a novel evaluation\nmatric} based on SSIM\\cite{SSIM} loss and fool rate.Additionally, we explain\nwhy FID\\cite{FID} is not suitable for measuring such adversarial examples. To\nthe best of our knowledge, it's the first evaluation metrics that is\nspecifically designed to evaluate the quality of a adversarial attack. We also\ninvestigate the transferability of adversarial examples crafted in the latent\nspace and show that they have superiority over adversarial examples crafted in\nthe pixel space.\n", "date": "2023-05-22", "categories": []}
{"id": "2301.13862", "title": "Salient Conditional Diffusion for Defending Against Backdoor Attacks", "abstract": "  We propose a novel algorithm, Salient Conditional Diffusion (Sancdifi), a\nstate-of-the-art defense against backdoor attacks. Sancdifi uses a denoising\ndiffusion probabilistic model (DDPM) to degrade an image with noise and then\nrecover said image using the learned reverse diffusion. Critically, we compute\nsaliency map-based masks to condition our diffusion, allowing for stronger\ndiffusion on the most salient pixels by the DDPM. As a result, Sancdifi is\nhighly effective at diffusing out triggers in data poisoned by backdoor\nattacks. At the same time, it reliably recovers salient features when applied\nto clean data. This performance is achieved without requiring access to the\nmodel parameters of the Trojan network, meaning Sancdifi operates as a\nblack-box defense.\n", "date": "2023-01-31", "categories": []}
{"id": "2303.02251", "title": "Certified Robust Neural Networks: Generalization and Corruption\n  Resistance", "abstract": "  Recent work have demonstrated that robustness (to \"corruption\") can be at\nodds with generalization. Adversarial training, for instance, aims to reduce\nthe problematic susceptibility of modern neural networks to small data\nperturbations. Surprisingly, overfitting is a major concern in adversarial\ntraining despite being mostly absent in standard training. We provide here\ntheoretical evidence for this peculiar \"robust overfitting\" phenomenon.\nSubsequently, we advance a novel distributionally robust loss function bridging\nrobustness and generalization. We demonstrate both theoretically as well as\nempirically the loss to enjoy a certified level of robustness against two\ncommon types of corruption--data evasion and poisoning attacks--while ensuring\nguaranteed generalization. We show through careful numerical experiments that\nour resulting holistic robust (HR) training procedure yields SOTA performance.\nFinally, we indicate that HR training can be interpreted as a direct extension\nof adversarial training and comes with a negligible additional computational\nburden. A ready-to-use python library implementing our algorithm is available\nat https://github.com/RyanLucas3/HR_Neural_Networks.\n", "date": "2023-03-03", "categories": []}
{"id": "2304.12550", "title": "Combining Adversaries with Anti-adversaries in Training", "abstract": "  Adversarial training is an effective learning technique to improve the\nrobustness of deep neural networks. In this study, the influence of adversarial\ntraining on deep learning models in terms of fairness, robustness, and\ngeneralization is theoretically investigated under more general perturbation\nscope that different samples can have different perturbation directions (the\nadversarial and anti-adversarial directions) and varied perturbation bounds.\nOur theoretical explorations suggest that the combination of adversaries and\nanti-adversaries (samples with anti-adversarial perturbations) in training can\nbe more effective in achieving better fairness between classes and a better\ntradeoff between robustness and generalization in some typical learning\nscenarios (e.g., noisy label learning and imbalance learning) compared with\nstandard adversarial training. On the basis of our theoretical findings, a more\ngeneral learning objective that combines adversaries and anti-adversaries with\nvaried bounds on each training sample is presented. Meta learning is utilized\nto optimize the combination weights. Experiments on benchmark datasets under\ndifferent learning scenarios verify our theoretical findings and the\neffectiveness of the proposed methodology.\n", "date": "2023-04-25", "categories": []}
{"id": "2305.10766", "title": "Adversarial Amendment is the Only Force Capable of Transforming an Enemy\n  into a Friend", "abstract": "  Adversarial attack is commonly regarded as a huge threat to neural networks\nbecause of misleading behavior. This paper presents an opposite perspective:\nadversarial attacks can be harnessed to improve neural models if amended\ncorrectly. Unlike traditional adversarial defense or adversarial training\nschemes that aim to improve the adversarial robustness, the proposed\nadversarial amendment (AdvAmd) method aims to improve the original accuracy\nlevel of neural models on benign samples. We thoroughly analyze the\ndistribution mismatch between the benign and adversarial samples. This\ndistribution mismatch and the mutual learning mechanism with the same learning\nratio applied in prior art defense strategies is the main cause leading the\naccuracy degradation for benign samples. The proposed AdvAmd is demonstrated to\nsteadily heal the accuracy degradation and even leads to a certain accuracy\nboost of common neural models on benign classification, object detection, and\nsegmentation tasks. The efficacy of the AdvAmd is contributed by three key\ncomponents: mediate samples (to reduce the influence of distribution mismatch\nwith a fine-grained amendment), auxiliary batch norm (to solve the mutual\nlearning mechanism and the smoother judgment surface), and AdvAmd loss (to\nadjust the learning ratios according to different attack vulnerabilities)\nthrough quantitative and ablation experiments.\n", "date": "2023-05-18", "categories": []}
{"id": "2305.10929", "title": "Architecture-agnostic Iterative Black-box Certified Defense against\n  Adversarial Patches", "abstract": "  The adversarial patch attack aims to fool image classifiers within a bounded,\ncontiguous region of arbitrary changes, posing a real threat to computer vision\nsystems (e.g., autonomous driving, content moderation, biometric\nauthentication, medical imaging) in the physical world. To address this problem\nin a trustworthy way, proposals have been made for certified patch defenses\nthat ensure the robustness of classification models and prevent future patch\nattacks from breaching the defense. State-of-the-art certified defenses can be\ncompatible with any model architecture, as well as achieve high clean and\ncertified accuracy. Although the methods are adaptive to arbitrary patch\npositions, they inevitably need to access the size of the adversarial patch,\nwhich is unreasonable and impractical in real-world attack scenarios. To\nimprove the feasibility of the architecture-agnostic certified defense in a\nblack-box setting (i.e., position and size of the patch are both unknown), we\npropose a novel two-stage Iterative Black-box Certified Defense method, termed\nIBCD.In the first stage, it estimates the patch size in a search-based manner\nby evaluating the size relationship between the patch and mask with pixel\nmasking. In the second stage, the accuracy results are calculated by the\nexisting white-box certified defense methods with the estimated patch size. The\nexperiments conducted on two popular model architectures and two datasets\nverify the effectiveness and efficiency of IBCD.\n", "date": "2023-05-18", "categories": []}
{"id": "2305.11039", "title": "Deep PackGen: A Deep Reinforcement Learning Framework for Adversarial\n  Network Packet Generation", "abstract": "  Recent advancements in artificial intelligence (AI) and machine learning (ML)\nalgorithms, coupled with the availability of faster computing infrastructure,\nhave enhanced the security posture of cybersecurity operations centers\n(defenders) through the development of ML-aided network intrusion detection\nsystems (NIDS). Concurrently, the abilities of adversaries to evade security\nhave also increased with the support of AI/ML models. Therefore, defenders need\nto proactively prepare for evasion attacks that exploit the detection\nmechanisms of NIDS. Recent studies have found that the perturbation of\nflow-based and packet-based features can deceive ML models, but these\napproaches have limitations. Perturbations made to the flow-based features are\ndifficult to reverse-engineer, while samples generated with perturbations to\nthe packet-based features are not playable.\n  Our methodological framework, Deep PackGen, employs deep reinforcement\nlearning to generate adversarial packets and aims to overcome the limitations\nof approaches in the literature. By taking raw malicious network packets as\ninputs and systematically making perturbations on them, Deep PackGen\ncamouflages them as benign packets while still maintaining their functionality.\nIn our experiments, using publicly available data, Deep PackGen achieved an\naverage adversarial success rate of 66.4\\% against various ML models and across\ndifferent attack types. Our investigation also revealed that more than 45\\% of\nthe successful adversarial samples were out-of-distribution packets that evaded\nthe decision boundaries of the classifiers. The knowledge gained from our study\non the adversary's ability to make specific evasive perturbations to different\ntypes of malicious packets can help defenders enhance the robustness of their\nNIDS against evolving adversarial attacks.\n", "date": "2023-05-18", "categories": []}
{"id": "2206.09099", "title": "The Consistency of Adversarial Training for Binary Classification", "abstract": "  Robustness to adversarial perturbations is of paramount concern in modern\nmachine learning. One of the state-of-the-art methods for training robust\nclassifiers is adversarial training, which involves minimizing a supremum-based\nsurrogate risk. The statistical consistency of surrogate risks is well\nunderstood in the context of standard machine learning, but not in the\nadversarial setting. In this paper, we characterize which supremum-based\nsurrogates are consistent for distributions absolutely continuous with respect\nto Lebesgue measure in binary classification. Furthermore, we obtain\nquantitative bounds relating adversarial surrogate risks to the adversarial\nclassification risk. Lastly, we discuss implications for the $\\cH$-consistency\nof adversarial training.\n", "date": "2022-06-18", "categories": []}
{"id": "2305.10388", "title": "Raising the Bar for Certified Adversarial Robustness with Diffusion\n  Models", "abstract": "  Certified defenses against adversarial attacks offer formal guarantees on the\nrobustness of a model, making them more reliable than empirical methods such as\nadversarial training, whose effectiveness is often later reduced by unseen\nattacks. Still, the limited certified robustness that is currently achievable\nhas been a bottleneck for their practical adoption. Gowal et al. and Wang et\nal. have shown that generating additional training data using state-of-the-art\ndiffusion models can considerably improve the robustness of adversarial\ntraining. In this work, we demonstrate that a similar approach can\nsubstantially improve deterministic certified defenses. In addition, we provide\na list of recommendations to scale the robustness of certified training\napproaches. One of our main insights is that the generalization gap, i.e., the\ndifference between the training and test accuracy of the original model, is a\ngood predictor of the magnitude of the robustness improvement when using\nadditional generated data. Our approach achieves state-of-the-art deterministic\nrobustness certificates on CIFAR-10 for the $\\ell_2$ ($\\epsilon = 36/255$) and\n$\\ell_\\infty$ ($\\epsilon = 8/255$) threat models, outperforming the previous\nbest results by $+3.95\\%$ and $+1.39\\%$, respectively. Furthermore, we report\nsimilar improvements for CIFAR-100.\n", "date": "2023-05-17", "categories": []}
{"id": "2302.02300", "title": "Run-Off Election: Improved Provable Defense against Data Poisoning\n  Attacks", "abstract": "  In data poisoning attacks, an adversary tries to change a model's prediction\nby adding, modifying, or removing samples in the training data. Recently,\nensemble-based approaches for obtaining provable defenses against data\npoisoning have been proposed where predictions are done by taking a majority\nvote across multiple base models. In this work, we show that merely considering\nthe majority vote in ensemble defenses is wasteful as it does not effectively\nutilize available information in the logits layers of the base models. Instead,\nwe propose Run-Off Election (ROE), a novel aggregation method based on a\ntwo-round election across the base models: In the first round, models vote for\ntheir preferred class and then a second, Run-Off election is held between the\ntop two classes in the first round. Based on this approach, we propose DPA+ROE\nand FA+ROE defense methods based on Deep Partition Aggregation (DPA) and Finite\nAggregation (FA) approaches from prior work. We evaluate our methods on MNIST,\nCIFAR-10, and GTSRB and obtain improvements in certified accuracy by up to\n3%-4%. Also, by applying ROE on a boosted version of DPA, we gain improvements\naround 12%-27% comparing to the current state-of-the-art, establishing a new\nstate-of-the-art in (pointwise) certified robustness against data poisoning. In\nmany cases, our approach outperforms the state-of-the-art, even when using 32\ntimes less computational power.\n", "date": "2023-02-05", "categories": []}
{"id": "2305.09179", "title": "Ortho-ODE: Enhancing Robustness and of Neural ODEs against Adversarial\n  Attacks", "abstract": "  Neural Ordinary Differential Equations (NODEs) probed the usage of numerical\nsolvers to solve the differential equation characterized by a Neural Network\n(NN), therefore initiating a new paradigm of deep learning models with infinite\ndepth. NODEs were designed to tackle the irregular time series problem.\nHowever, NODEs have demonstrated robustness against various noises and\nadversarial attacks. This paper is about the natural robustness of NODEs and\nexamines the cause behind such surprising behaviour. We show that by\ncontrolling the Lipschitz constant of the ODE dynamics the robustness can be\nsignificantly improved. We derive our approach from Grownwall's inequality.\nFurther, we draw parallels between contractivity theory and Grownwall's\ninequality. Experimentally we corroborate the enhanced robustness on numerous\ndatasets - MNIST, CIFAR-10, and CIFAR 100. We also present the impact of\nadaptive and non-adaptive solvers on the robustness of NODEs.\n", "date": "2023-05-16", "categories": []}
{"id": "2305.08031", "title": "On enhancing the robustness of Vision Transformers: Defensive Diffusion", "abstract": "  Privacy and confidentiality of medical data are of utmost importance in\nhealthcare settings. ViTs, the SOTA vision model, rely on large amounts of\npatient data for training, which raises concerns about data security and the\npotential for unauthorized access. Adversaries may exploit vulnerabilities in\nViTs to extract sensitive patient information and compromising patient privacy.\nThis work address these vulnerabilities to ensure the trustworthiness and\nreliability of ViTs in medical applications. In this work, we introduced a\ndefensive diffusion technique as an adversarial purifier to eliminate\nadversarial noise introduced by attackers in the original image. By utilizing\nthe denoising capabilities of the diffusion model, we employ a reverse\ndiffusion process to effectively eliminate the adversarial noise from the\nattack sample, resulting in a cleaner image that is then fed into the ViT\nblocks. Our findings demonstrate the effectiveness of the diffusion model in\neliminating attack-agnostic adversarial noise from images. Additionally, we\npropose combining knowledge distillation with our framework to obtain a\nlightweight student model that is both computationally efficient and robust\nagainst gray box attacks. Comparison of our method with a SOTA baseline method,\nSEViT, shows that our work is able to outperform the baseline. Extensive\nexperiments conducted on a publicly available Tuberculosis X-ray dataset\nvalidate the computational efficiency and improved robustness achieved by our\nproposed architecture.\n", "date": "2023-05-14", "categories": []}
{"id": "2305.08076", "title": "Improving Defensive Distillation using Teacher Assistant", "abstract": "  Adversarial attacks pose a significant threat to the security and safety of\ndeep neural networks being applied to modern applications. More specifically,\nin computer vision-based tasks, experts can use the knowledge of model\narchitecture to create adversarial samples imperceptible to the human eye.\nThese attacks can lead to security problems in popular applications such as\nself-driving cars, face recognition, etc. Hence, building networks which are\nrobust to such attacks is highly desirable and essential. Among the various\nmethods present in literature, defensive distillation has shown promise in\nrecent years. Using knowledge distillation, researchers have been able to\ncreate models robust against some of those attacks. However, more attacks have\nbeen developed exposing weakness in defensive distillation. In this project, we\nderive inspiration from teacher assistant knowledge distillation and propose\nthat introducing an assistant network can improve the robustness of the\ndistilled model. Through a series of experiments, we evaluate the distilled\nmodels for different distillation temperatures in terms of accuracy,\nsensitivity, and robustness. Our experiments demonstrate that the proposed\nhypothesis can improve robustness in most cases. Additionally, we show that\nmulti-step distillation can further improve robustness with very little impact\non model accuracy.\n", "date": "2023-05-14", "categories": []}
{"id": "2305.08439", "title": "Exploiting Frequency Spectrum of Adversarial Images for General\n  Robustness", "abstract": "  In recent years, there has been growing concern over the vulnerability of\nconvolutional neural networks (CNNs) to image perturbations. However, achieving\ngeneral robustness against different types of perturbations remains\nchallenging, in which enhancing robustness to some perturbations (e.g.,\nadversarial perturbations) may degrade others (e.g., common corruptions). In\nthis paper, we demonstrate that adversarial training with an emphasis on phase\ncomponents significantly improves model performance on clean, adversarial, and\ncommon corruption accuracies. We propose a frequency-based data augmentation\nmethod, Adversarial Amplitude Swap, that swaps the amplitude spectrum between\nclean and adversarial images to generate two novel training images: adversarial\namplitude and adversarial phase images. These images act as substitutes for\nadversarial images and can be implemented in various adversarial training\nsetups. Through extensive experiments, we demonstrate that our method enables\nthe CNNs to gain general robustness against different types of perturbations\nand results in a uniform performance against all types of common corruptions.\n", "date": "2023-05-15", "categories": []}
{"id": "2305.08840", "title": "Attacking Perceptual Similarity Metrics", "abstract": "  Perceptual similarity metrics have progressively become more correlated with\nhuman judgments on perceptual similarity; however, despite recent advances, the\naddition of an imperceptible distortion can still compromise these metrics. In\nour study, we systematically examine the robustness of these metrics to\nimperceptible adversarial perturbations. Following the two-alternative\nforced-choice experimental design with two distorted images and one reference\nimage, we perturb the distorted image closer to the reference via an\nadversarial attack until the metric flips its judgment. We first show that all\nmetrics in our study are susceptible to perturbations generated via common\nadversarial attacks such as FGSM, PGD, and the One-pixel attack. Next, we\nattack the widely adopted LPIPS metric using spatial-transformation-based\nadversarial perturbations (stAdv) in a white-box setting to craft adversarial\nexamples that can effectively transfer to other similarity metrics in a\nblack-box setting. We also combine the spatial attack stAdv with PGD\n($\\ell_\\infty$-bounded) attack to increase transferability and use these\nadversarial examples to benchmark the robustness of both traditional and\nrecently developed metrics. Our benchmark provides a good starting point for\ndiscussion and further research on the robustness of metrics to imperceptible\nadversarial perturbations.\n", "date": "2023-05-15", "categories": []}
{"id": "2305.01139", "title": "Stratified Adversarial Robustness with Rejection", "abstract": "  Recently, there is an emerging interest in adversarially training a\nclassifier with a rejection option (also known as a selective classifier) for\nboosting adversarial robustness. While rejection can incur a cost in many\napplications, existing studies typically associate zero cost with rejecting\nperturbed inputs, which can result in the rejection of numerous\nslightly-perturbed inputs that could be correctly classified. In this work, we\nstudy adversarially-robust classification with rejection in the stratified\nrejection setting, where the rejection cost is modeled by rejection loss\nfunctions monotonically non-increasing in the perturbation magnitude. We\ntheoretically analyze the stratified rejection setting and propose a novel\ndefense method -- Adversarial Training with Consistent Prediction-based\nRejection (CPR) -- for building a robust selective classifier. Experiments on\nimage datasets demonstrate that the proposed method significantly outperforms\nexisting methods under strong adaptive attacks. For instance, on CIFAR-10, CPR\nreduces the total robust loss (for different rejection losses) by at least 7.3%\nunder both seen and unseen attacks.\n", "date": "2023-05-02", "categories": []}
{"id": "2305.07406", "title": "Two-in-One: A Model Hijacking Attack Against Text Generation Models", "abstract": "  Machine learning has progressed significantly in various applications ranging\nfrom face recognition to text generation. However, its success has been\naccompanied by different attacks. Recently a new attack has been proposed which\nraises both accountability and parasitic computing risks, namely the model\nhijacking attack. Nevertheless, this attack has only focused on image\nclassification tasks. In this work, we broaden the scope of this attack to\ninclude text generation and classification models, hence showing its broader\napplicability. More concretely, we propose a new model hijacking attack, Ditto,\nthat can hijack different text classification tasks into multiple generation\nones, e.g., language translation, text summarization, and language modeling. We\nuse a range of text benchmark datasets such as SST-2, TweetEval, AGnews, QNLI,\nand IMDB to evaluate the performance of our attacks. Our results show that by\nusing Ditto, an adversary can successfully hijack text generation models\nwithout jeopardizing their utility.\n", "date": "2023-05-12", "categories": []}
{"id": "2210.08178", "title": "Is Face Recognition Safe from Realizable Attacks?", "abstract": "  Face recognition is a popular form of biometric authentication and due to its\nwidespread use, attacks have become more common as well. Recent studies show\nthat Face Recognition Systems are vulnerable to attacks and can lead to\nerroneous identification of faces. Interestingly, most of these attacks are\nwhite-box, or they are manipulating facial images in ways that are not\nphysically realizable. In this paper, we propose an attack scheme where the\nattacker can generate realistic synthesized face images with subtle\nperturbations and physically realize that onto his face to attack black-box\nface recognition systems. Comprehensive experiments and analyses show that\nsubtle perturbations realized on attackers face can create successful attacks\non state-of-the-art face recognition systems in black-box settings. Our study\nexposes the underlying vulnerability posed by the Face Recognition Systems\nagainst realizable black-box attacks.\n", "date": "2022-10-15", "categories": []}
{"id": "2210.14410", "title": "Improving Adversarial Robustness via Joint Classification and Multiple\n  Explicit Detection Classes", "abstract": "  This work concerns the development of deep networks that are certifiably\nrobust to adversarial attacks. Joint robust classification-detection was\nrecently introduced as a certified defense mechanism, where adversarial\nexamples are either correctly classified or assigned to the \"abstain\" class. In\nthis work, we show that such a provable framework can benefit by extension to\nnetworks with multiple explicit abstain classes, where the adversarial examples\nare adaptively assigned to those. We show that naively adding multiple abstain\nclasses can lead to \"model degeneracy\", then we propose a regularization\napproach and a training method to counter this degeneracy by promoting full use\nof the multiple abstain classes. Our experiments demonstrate that the proposed\napproach consistently achieves favorable standard vs. robust verified accuracy\ntradeoffs, outperforming state-of-the-art algorithms for various choices of\nnumber of abstain classes.\n", "date": "2022-10-26", "categories": []}
{"id": "2210.16346", "title": "Improving Hyperspectral Adversarial Robustness Under Multiple Attacks", "abstract": "  Semantic segmentation models classifying hyperspectral images (HSI) are\nvulnerable to adversarial examples. Traditional approaches to adversarial\nrobustness focus on training or retraining a single network on attacked data,\nhowever, in the presence of multiple attacks these approaches decrease in\nperformance compared to networks trained individually on each attack. To combat\nthis issue we propose an Adversarial Discriminator Ensemble Network (ADE-Net)\nwhich focuses on attack type detection and adversarial robustness under a\nunified model to preserve per data-type weight optimally while robustifiying\nthe overall network. In the proposed method, a discriminator network is used to\nseparate data by attack type into their specific attack-expert ensemble\nnetwork.\n", "date": "2022-10-28", "categories": []}
{"id": "2305.03888", "title": "Energy-Latency Attacks to On-Device Neural Networks via Sponge Poisoning", "abstract": "  In recent years, on-device deep learning has gained attention as a means of\ndeveloping affordable deep learning applications for mobile devices. However,\non-device models are constrained by limited energy and computation resources.\nIn the mean time, a poisoning attack known as sponge poisoning has been\ndeveloped.This attack involves feeding the model with poisoned examples to\nincrease the energy consumption during inference. As previous work is focusing\non server hardware accelerators, in this work, we extend the sponge poisoning\nattack to an on-device scenario to evaluate the vulnerability of mobile device\nprocessors. We present an on-device sponge poisoning attack pipeline to\nsimulate the streaming and consistent inference scenario to bridge the\nknowledge gap in the on-device setting. Our exclusive experimental analysis\nwith processors and on-device networks shows that sponge poisoning attacks can\neffectively pollute the modern processor with its built-in accelerator. We\nanalyze the impact of different factors in the sponge poisoning algorithm and\nhighlight the need for improved defense mechanisms to prevent such attacks on\non-device deep learning applications.\n", "date": "2023-05-06", "categories": []}
{"id": "2305.06522", "title": "Randomized Smoothing with Masked Inference for Adversarially Robust Text\n  Classifications", "abstract": "  Large-scale pre-trained language models have shown outstanding performance in\na variety of NLP tasks. However, they are also known to be significantly\nbrittle against specifically crafted adversarial examples, leading to\nincreasing interest in probing the adversarial robustness of NLP systems. We\nintroduce RSMI, a novel two-stage framework that combines randomized smoothing\n(RS) with masked inference (MI) to improve the adversarial robustness of NLP\nsystems. RS transforms a classifier into a smoothed classifier to obtain robust\nrepresentations, whereas MI forces a model to exploit the surrounding context\nof a masked token in an input sequence. RSMI improves adversarial robustness by\n2 to 3 times over existing state-of-the-art methods on benchmark datasets. We\nalso perform in-depth qualitative analysis to validate the effectiveness of the\ndifferent stages of RSMI and probe the impact of its components through\nextensive ablations. By empirically proving the stability of RSMI, we put it\nforward as a practical method to robustly train large-scale NLP models. Our\ncode and datasets are available at https://github.com/Han8931/rsmi_nlp\n", "date": "2023-05-11", "categories": []}
{"id": "2305.06540", "title": "Inter-frame Accelerate Attack against Video Interpolation Models", "abstract": "  Deep learning based video frame interpolation (VIF) method, aiming to\nsynthesis the intermediate frames to enhance video quality, have been highly\ndeveloped in the past few years. This paper investigates the adversarial\nrobustness of VIF models. We apply adversarial attacks to VIF models and find\nthat the VIF models are very vulnerable to adversarial examples. To improve\nattack efficiency, we suggest to make full use of the property of video frame\ninterpolation task. The intuition is that the gap between adjacent frames would\nbe small, leading to the corresponding adversarial perturbations being similar\nas well. Then we propose a novel attack method named Inter-frame Accelerate\nAttack (IAA) that initializes the perturbation as the perturbation for the\nprevious adjacent frame and reduces the number of attack iterations. It is\nshown that our method can improve attack efficiency greatly while achieving\ncomparable attack performance with traditional methods. Besides, we also extend\nour method to video recognition models which are higher level vision tasks and\nachieves great attack efficiency.\n", "date": "2023-05-11", "categories": []}
{"id": "2207.04308", "title": "Dynamic Time Warping based Adversarial Framework for Time-Series Domain", "abstract": "  Despite the rapid progress on research in adversarial robustness of deep\nneural networks (DNNs), there is little principled work for the time-series\ndomain. Since time-series data arises in diverse applications including mobile\nhealth, finance, and smart grid, it is important to verify and improve the\nrobustness of DNNs for the time-series domain. In this paper, we propose a\nnovel framework for the time-series domain referred as {\\em Dynamic Time\nWarping for Adversarial Robustness (DTW-AR)} using the dynamic time warping\nmeasure. Theoretical and empirical evidence is provided to demonstrate the\neffectiveness of DTW over the standard Euclidean distance metric employed in\nprior methods for the image domain. We develop a principled algorithm justified\nby theoretical analysis to efficiently create diverse adversarial examples\nusing random alignment paths. Experiments on diverse real-world benchmarks show\nthe effectiveness of DTW-AR to fool DNNs for time-series data and to improve\ntheir robustness using adversarial training. The source code of DTW-AR\nalgorithms is available at https://github.com/tahabelkhouja/DTW-AR\n", "date": "2022-07-09", "categories": []}
{"id": "2305.05253", "title": "Attack Named Entity Recognition by Entity Boundary Interference", "abstract": "  Named Entity Recognition (NER) is a cornerstone NLP task while its robustness\nhas been given little attention. This paper rethinks the principles of NER\nattacks derived from sentence classification, as they can easily violate the\nlabel consistency between the original and adversarial NER examples. This is\ndue to the fine-grained nature of NER, as even minor word changes in the\nsentence can result in the emergence or mutation of any entities, resulting in\ninvalid adversarial examples. To this end, we propose a novel one-word\nmodification NER attack based on a key insight, NER models are always\nvulnerable to the boundary position of an entity to make their decision. We\nthus strategically insert a new boundary into the sentence and trigger the\nEntity Boundary Interference that the victim model makes the wrong prediction\neither on this boundary word or on other words in the sentence. We call this\nattack Virtual Boundary Attack (ViBA), which is shown to be remarkably\neffective when attacking both English and Chinese models with a 70%-90% attack\nsuccess rate on state-of-the-art language models (e.g. RoBERTa, DeBERTa) and\nalso significantly faster than previous methods.\n", "date": "2023-05-09", "categories": []}
{"id": "2207.11177", "title": "Provable Defense Against Geometric Transformations", "abstract": "  Geometric image transformations that arise in the real world, such as scaling\nand rotation, have been shown to easily deceive deep neural networks (DNNs).\nHence, training DNNs to be certifiably robust to these perturbations is\ncritical. However, no prior work has been able to incorporate the objective of\ndeterministic certified robustness against geometric transformations into the\ntraining procedure, as existing verifiers are exceedingly slow. To address\nthese challenges, we propose the first provable defense for deterministic\ncertified geometric robustness. Our framework leverages a novel GPU-optimized\nverifier that can certify images between 60$\\times$ to 42,600$\\times$ faster\nthan existing geometric robustness verifiers, and thus unlike existing works,\nis fast enough for use in training. Across multiple datasets, our results show\nthat networks trained via our framework consistently achieve state-of-the-art\ndeterministic certified geometric robustness and clean accuracy. Furthermore,\nfor the first time, we verify the geometric robustness of a neural network for\nthe challenging, real-world setting of autonomous driving.\n", "date": "2022-07-22", "categories": []}
{"id": "2210.02844", "title": "Are Synonym Substitution Attacks Really Synonym Substitution Attacks?", "abstract": "  In this paper, we explore the following question: Are synonym substitution\nattacks really synonym substitution attacks (SSAs)? We approach this question\nby examining how SSAs replace words in the original sentence and show that\nthere are still unresolved obstacles that make current SSAs generate invalid\nadversarial samples. We reveal that four widely used word substitution methods\ngenerate a large fraction of invalid substitution words that are ungrammatical\nor do not preserve the original sentence's semantics. Next, we show that the\nsemantic and grammatical constraints used in SSAs for detecting invalid word\nreplacements are highly insufficient in detecting invalid adversarial samples.\n", "date": "2022-10-06", "categories": []}
{"id": "2305.00866", "title": "Attack-SAM: Towards Attacking Segment Anything Model With Adversarial\n  Examples", "abstract": "  Segment Anything Model (SAM) has attracted significant attention recently,\ndue to its impressive performance on various downstream tasks in a zero-short\nmanner. Computer vision (CV) area might follow the natural language processing\n(NLP) area to embark on a path from task-specific vision models toward\nfoundation models. However, deep vision models are widely recognized as\nvulnerable to adversarial examples, which fool the model to make wrong\npredictions with imperceptible perturbation. Such vulnerability to adversarial\nattacks causes serious concerns when applying deep models to security-sensitive\napplications. Therefore, it is critical to know whether the vision foundation\nmodel SAM can also be fooled by adversarial attacks. To the best of our\nknowledge, our work is the first of its kind to conduct a comprehensive\ninvestigation on how to attack SAM with adversarial examples. With the basic\nattack goal set to mask removal, we investigate the adversarial robustness of\nSAM in the full white-box setting and transfer-based black-box settings. Beyond\nthe basic goal of mask removal, we further investigate and find that it is\npossible to generate any desired mask by the adversarial attack.\n", "date": "2023-05-01", "categories": []}
{"id": "2305.03655", "title": "White-Box Multi-Objective Adversarial Attack on Dialogue Generation", "abstract": "  Pre-trained transformers are popular in state-of-the-art dialogue generation\n(DG) systems. Such language models are, however, vulnerable to various\nadversarial samples as studied in traditional tasks such as text\nclassification, which inspires our curiosity about their robustness in DG\nsystems. One main challenge of attacking DG models is that perturbations on the\ncurrent sentence can hardly degrade the response accuracy because the unchanged\nchat histories are also considered for decision-making. Instead of merely\npursuing pitfalls of performance metrics such as BLEU, ROUGE, we observe that\ncrafting adversarial samples to force longer generation outputs benefits attack\neffectiveness -- the generated responses are typically irrelevant, lengthy, and\nrepetitive. To this end, we propose a white-box multi-objective attack method\ncalled DGSlow. Specifically, DGSlow balances two objectives -- generation\naccuracy and length, via a gradient-based multi-objective optimizer and applies\nan adaptive searching mechanism to iteratively craft adversarial samples with\nonly a few modifications. Comprehensive experiments on four benchmark datasets\ndemonstrate that DGSlow could significantly degrade state-of-the-art DG models\nwith a higher success rate than traditional accuracy-based methods. Besides,\nour crafted sentences also exhibit strong transferability in attacking other\nmodels.\n", "date": "2023-05-05", "categories": []}
{"id": "2305.04436", "title": "Adversarial Examples Detection with Enhanced Image Difference Features\n  based on Local Histogram Equalization", "abstract": "  Deep Neural Networks (DNNs) have recently made significant progress in many\nfields. However, studies have shown that DNNs are vulnerable to adversarial\nexamples, where imperceptible perturbations can greatly mislead DNNs even if\nthe full underlying model parameters are not accessible. Various defense\nmethods have been proposed, such as feature compression and gradient masking.\nHowever, numerous studies have proven that previous methods create detection or\ndefense against certain attacks, which renders the method ineffective in the\nface of the latest unknown attack methods. The invisibility of adversarial\nperturbations is one of the evaluation indicators for adversarial example\nattacks, which also means that the difference in the local correlation of\nhigh-frequency information in adversarial examples and normal examples can be\nused as an effective feature to distinguish the two. Therefore, we propose an\nadversarial example detection framework based on a high-frequency information\nenhancement strategy, which can effectively extract and amplify the feature\ndifferences between adversarial examples and normal examples. Experimental\nresults show that the feature augmentation module can be combined with existing\ndetection models in a modular way under this framework. Improve the detector's\nperformance and reduce the deployment cost without modifying the existing\ndetection model.\n", "date": "2023-05-08", "categories": []}
{"id": "2305.04557", "title": "Toward Adversarial Training on Contextualized Language Representation", "abstract": "  Beyond the success story of adversarial training (AT) in the recent text\ndomain on top of pre-trained language models (PLMs), our empirical study\nshowcases the inconsistent gains from AT on some tasks, e.g. commonsense\nreasoning, named entity recognition. This paper investigates AT from the\nperspective of the contextualized language representation outputted by PLM\nencoders. We find the current AT attacks lean to generate sub-optimal\nadversarial examples that can fool the decoder part but have a minor effect on\nthe encoder. However, we find it necessary to effectively deviate the latter\none to allow AT to gain. Based on the observation, we propose simple yet\neffective \\textit{Contextualized representation-Adversarial Training} (CreAT),\nin which the attack is explicitly optimized to deviate the contextualized\nrepresentation of the encoder. It allows a global optimization of adversarial\nexamples that can fool the entire model. We also find CreAT gives rise to a\nbetter direction to optimize the adversarial examples, to let them less\nsensitive to hyperparameters. Compared to AT, CreAT produces consistent\nperformance gains on a wider range of tasks and is proven to be more effective\nfor language pre-training where only the encoder part is kept for downstream\ntasks. We achieve the new state-of-the-art performances on a series of\nchallenging benchmarks, e.g. AdvGLUE (59.1 $ \\rightarrow $ 61.1), HellaSWAG\n(93.0 $ \\rightarrow $ 94.9), ANLI (68.1 $ \\rightarrow $ 69.3).\n", "date": "2023-05-08", "categories": []}
{"id": "2305.04746", "title": "Understanding Noise-Augmented Training for Randomized Smoothing", "abstract": "  Randomized smoothing is a technique for providing provable robustness\nguarantees against adversarial attacks while making minimal assumptions about a\nclassifier. This method relies on taking a majority vote of any base classifier\nover multiple noise-perturbed inputs to obtain a smoothed classifier, and it\nremains the tool of choice to certify deep and complex neural network models.\nNonetheless, non-trivial performance of such smoothed classifier crucially\ndepends on the base model being trained on noise-augmented data, i.e., on a\nsmoothed input distribution. While widely adopted in practice, it is still\nunclear how this noisy training of the base classifier precisely affects the\nrisk of the robust smoothed classifier, leading to heuristics and tricks that\nare poorly understood. In this work we analyze these trade-offs theoretically\nin a binary classification setting, proving that these common observations are\nnot universal. We show that, without making stronger distributional\nassumptions, no benefit can be expected from predictors trained with\nnoise-augmentation, and we further characterize distributions where such\nbenefit is obtained. Our analysis has direct implications to the practical\ndeployment of randomized smoothing, and we illustrate some of these via\nexperiments on CIFAR-10 and MNIST, as well as on synthetic datasets.\n", "date": "2023-05-08", "categories": []}
{"id": "2211.10024", "title": "Diagnostics for Deep Neural Networks with Automated Copy/Paste Attacks", "abstract": "  This paper considers the problem of helping humans exercise scalable\noversight over deep neural networks (DNNs). Adversarial examples can be useful\nby helping to reveal weaknesses in DNNs, but they can be difficult to interpret\nor draw actionable conclusions from. Some previous works have proposed using\nhuman-interpretable adversarial attacks including copy/paste attacks in which\none natural image pasted into another causes an unexpected misclassification.\nWe build on these with two contributions. First, we introduce Search for\nNatural Adversarial Features Using Embeddings (SNAFUE) which offers a fully\nautomated method for finding copy/paste attacks. Second, we use SNAFUE to red\nteam an ImageNet classifier. We reproduce copy/paste attacks from previous\nworks and find hundreds of other easily-describable vulnerabilities, all\nwithout a human in the loop. Code is available at\nhttps://github.com/thestephencasper/snafue\n", "date": "2022-11-18", "categories": []}
{"id": "2305.03173", "title": "New Adversarial Image Detection Based on Sentiment Analysis", "abstract": "  Deep Neural Networks (DNNs) are vulnerable to adversarial examples, while\nadversarial attack models, e.g., DeepFool, are on the rise and outrunning\nadversarial example detection techniques. This paper presents a new adversarial\nexample detector that outperforms state-of-the-art detectors in identifying the\nlatest adversarial attacks on image datasets. Specifically, we propose to use\nsentiment analysis for adversarial example detection, qualified by the\nprogressively manifesting impact of an adversarial perturbation on the\nhidden-layer feature maps of a DNN under attack. Accordingly, we design a\nmodularized embedding layer with the minimum learnable parameters to embed the\nhidden-layer feature maps into word vectors and assemble sentences ready for\nsentiment analysis. Extensive experiments demonstrate that the new detector\nconsistently surpasses the state-of-the-art detection algorithms in detecting\nthe latest attacks launched against ResNet and Inception neutral networks on\nthe CIFAR-10, CIFAR-100 and SVHN datasets. The detector only has about 2\nmillion parameters, and takes shorter than 4.6 milliseconds to detect an\nadversarial example generated by the latest attack models using a Tesla K80 GPU\ncard.\n", "date": "2023-05-03", "categories": []}
{"id": "2305.03289", "title": "BadSAM: Exploring Security Vulnerabilities of SAM via Backdoor Attacks", "abstract": "  Recently, the Segment Anything Model (SAM) has gained significant attention\nas an image segmentation foundation model due to its strong performance on\nvarious downstream tasks. However, it has been found that SAM does not always\nperform satisfactorily when faced with challenging downstream tasks. This has\nled downstream users to demand a customized SAM model that can be adapted to\nthese downstream tasks. In this paper, we present BadSAM, the first backdoor\nattack on the image segmentation foundation model. Our preliminary experiments\non the CAMO dataset demonstrate the effectiveness of BadSAM.\n", "date": "2023-05-05", "categories": []}
{"id": "2304.11359", "title": "Detecting Adversarial Faces Using Only Real Face Self-Perturbations", "abstract": "  Adversarial attacks aim to disturb the functionality of a target system by\nadding specific noise to the input samples, bringing potential threats to\nsecurity and robustness when applied to facial recognition systems. Although\nexisting defense techniques achieve high accuracy in detecting some specific\nadversarial faces (adv-faces), new attack methods especially GAN-based attacks\nwith completely different noise patterns circumvent them and reach a higher\nattack success rate. Even worse, existing techniques require attack data before\nimplementing the defense, making it impractical to defend newly emerging\nattacks that are unseen to defenders. In this paper, we investigate the\nintrinsic generality of adv-faces and propose to generate pseudo adv-faces by\nperturbing real faces with three heuristically designed noise patterns. We are\nthe first to train an adv-face detector using only real faces and their\nself-perturbations, agnostic to victim facial recognition systems, and agnostic\nto unseen attacks. By regarding adv-faces as out-of-distribution data, we then\nnaturally introduce a novel cascaded system for adv-face detection, which\nconsists of training data self-perturbations, decision boundary regularization,\nand a max-pooling-based binary classifier focusing on abnormal local color\naberrations. Experiments conducted on LFW and CelebA-HQ datasets with eight\ngradient-based and two GAN-based attacks validate that our method generalizes\nto a variety of unseen adversarial attacks.\n", "date": "2023-04-22", "categories": []}
{"id": "2305.01860", "title": "Towards Imperceptible Document Manipulations against Neural Ranking\n  Models", "abstract": "  Adversarial attacks have gained traction in order to identify potential\nvulnerabilities in neural ranking models (NRMs), but current attack methods\noften introduce grammatical errors, nonsensical expressions, or incoherent text\nfragments, which can be easily detected. Additionally, current methods rely\nheavily on the use of a well-imitated surrogate NRM to guarantee the attack\neffect, which makes them difficult to use in practice. To address these issues,\nwe propose a framework called Imperceptible DocumEnt Manipulation (IDEM) to\nproduce adversarial documents that are less noticeable to both algorithms and\nhumans. IDEM instructs a well-established generative language model, such as\nBART, to generate connection sentences without introducing easy-to-detect\nerrors, and employs a separate position-wise merging strategy to balance\nrelevance and coherence of the perturbed text. Experimental results on the\npopular MS MARCO benchmark demonstrate that IDEM can outperform strong\nbaselines while preserving fluency and correctness of the target documents as\nevidenced by automatic and human evaluations. Furthermore, the separation of\nadversarial text generation from the surrogate NRM makes IDEM more robust and\nless affected by the quality of the surrogate NRM.\n", "date": "2023-05-03", "categories": []}
{"id": "2305.02901", "title": "Single Node Injection Label Specificity Attack on Graph Neural Networks\n  via Reinforcement Learning", "abstract": "  Graph neural networks (GNNs) have achieved remarkable success in various\nreal-world applications. However, recent studies highlight the vulnerability of\nGNNs to malicious perturbations. Previous adversaries primarily focus on graph\nmodifications or node injections to existing graphs, yielding promising results\nbut with notable limitations. Graph modification attack~(GMA) requires\nmanipulation of the original graph, which is often impractical, while graph\ninjection attack~(GIA) necessitates training a surrogate model in the black-box\nsetting, leading to significant performance degradation due to divergence\nbetween the surrogate architecture and the actual victim model. Furthermore,\nmost methods concentrate on a single attack goal and lack a generalizable\nadversary to develop distinct attack strategies for diverse goals, thus\nlimiting precise control over victim model behavior in real-world scenarios. To\naddress these issues, we present a gradient-free generalizable adversary that\ninjects a single malicious node to manipulate the classification result of a\ntarget node in the black-box evasion setting. We propose Gradient-free\nGeneralizable Single Node Injection Attack, namely G$^2$-SNIA, a reinforcement\nlearning framework employing Proximal Policy Optimization. By directly querying\nthe victim model, G$^2$-SNIA learns patterns from exploration to achieve\ndiverse attack goals with extremely limited attack budgets. Through\ncomprehensive experiments over three acknowledged benchmark datasets and four\nprominent GNNs in the most challenging and realistic scenario, we demonstrate\nthe superior performance of our proposed G$^2$-SNIA over the existing\nstate-of-the-art baselines. Moreover, by comparing G$^2$-SNIA with multiple\nwhite-box evasion baselines, we confirm its capacity to generate solutions\ncomparable to those of the best adversaries.\n", "date": "2023-05-04", "categories": []}
{"id": "2211.03117", "title": "Going In Style: Audio Backdoors Through Stylistic Transformations", "abstract": "  This work explores stylistic triggers for backdoor attacks in the audio\ndomain: dynamic transformations of malicious samples through guitar effects. We\nfirst formalize stylistic triggers - currently missing in the literature.\nSecond, we explore how to develop stylistic triggers in the audio domain by\nproposing JingleBack. Our experiments confirm the effectiveness of the attack,\nachieving a 96% attack success rate. Our code is available in\nhttps://github.com/skoffas/going-in-style.\n", "date": "2022-11-06", "categories": []}
{"id": "2302.13519", "title": "CBA: Contextual Background Attack against Optical Aerial Detection in\n  the Physical World", "abstract": "  Patch-based physical attacks have increasingly aroused concerns.\n  However, most existing methods focus on obscuring targets captured on the\nground, and some of these methods are simply extended to deceive aerial\ndetectors.\n  They smear the targeted objects in the physical world with the elaborated\nadversarial patches, which can only slightly sway the aerial detectors'\nprediction and with weak attack transferability.\n  To address the above issues, we propose to perform Contextual Background\nAttack (CBA), a novel physical attack framework against aerial detection, which\ncan achieve strong attack efficacy and transferability in the physical world\neven without smudging the interested objects at all.\n  Specifically, the targets of interest, i.e. the aircraft in aerial images,\nare adopted to mask adversarial patches.\n  The pixels outside the mask area are optimized to make the generated\nadversarial patches closely cover the critical contextual background area for\ndetection, which contributes to gifting adversarial patches with more robust\nand transferable attack potency in the real world.\n  To further strengthen the attack performance, the adversarial patches are\nforced to be outside targets during training, by which the detected objects of\ninterest, both on and outside patches, benefit the accumulation of attack\nefficacy.\n  Consequently, the sophisticatedly designed patches are gifted with solid\nfooling efficacy against objects both on and outside the adversarial patches\nsimultaneously.\n  Extensive proportionally scaled experiments are performed in physical\nscenarios, demonstrating the superiority and potential of the proposed\nframework for physical attacks.\n  We expect that the proposed physical attack method will serve as a benchmark\nfor assessing the adversarial robustness of diverse aerial detectors and\ndefense methods.\n", "date": "2023-02-27", "categories": []}
{"id": "2304.12486", "title": "Evaluating Adversarial Robustness on Document Image Classification", "abstract": "  Adversarial attacks and defenses have gained increasing interest on computer\nvision systems in recent years, but as of today, most investigations are\nlimited to images. However, many artificial intelligence models actually handle\ndocumentary data, which is very different from real world images. Hence, in\nthis work, we try to apply the adversarial attack philosophy on documentary and\nnatural data and to protect models against such attacks. We focus our work on\nuntargeted gradient-based, transfer-based and score-based attacks and evaluate\nthe impact of adversarial training, JPEG input compression and grey-scale input\ntransformation on the robustness of ResNet50 and EfficientNetB0 model\narchitectures. To the best of our knowledge, no such work has been conducted by\nthe community in order to study the impact of these attacks on the document\nimage classification task.\n", "date": "2023-04-24", "categories": []}
{"id": "2305.00851", "title": "Revisiting Robustness in Graph Machine Learning", "abstract": "  Many works show that node-level predictions of Graph Neural Networks (GNNs)\nare unrobust to small, often termed adversarial, changes to the graph\nstructure. However, because manual inspection of a graph is difficult, it is\nunclear if the studied perturbations always preserve a core assumption of\nadversarial examples: that of unchanged semantic content. To address this\nproblem, we introduce a more principled notion of an adversarial graph, which\nis aware of semantic content change. Using Contextual Stochastic Block Models\n(CSBMs) and real-world graphs, our results uncover: $i)$ for a majority of\nnodes the prevalent perturbation models include a large fraction of perturbed\ngraphs violating the unchanged semantics assumption; $ii)$ surprisingly, all\nassessed GNNs show over-robustness - that is robustness beyond the point of\nsemantic change. We find this to be a complementary phenomenon to adversarial\nexamples and show that including the label-structure of the training graph into\nthe inference process of GNNs significantly reduces over-robustness, while\nhaving a positive effect on test accuracy and adversarial robustness.\nTheoretically, leveraging our new semantics-aware notion of robustness, we\nprove that there is no robustness-accuracy tradeoff for inductively classifying\na newly added node.\n", "date": "2023-05-01", "categories": []}
{"id": "2210.06284", "title": "Visual Prompting for Adversarial Robustness", "abstract": "  In this work, we leverage visual prompting (VP) to improve adversarial\nrobustness of a fixed, pre-trained model at testing time. Compared to\nconventional adversarial defenses, VP allows us to design universal (i.e.,\ndata-agnostic) input prompting templates, which have plug-and-play capabilities\nat testing time to achieve desired model performance without introducing much\ncomputation overhead. Although VP has been successfully applied to improving\nmodel generalization, it remains elusive whether and how it can be used to\ndefend against adversarial attacks. We investigate this problem and show that\nthe vanilla VP approach is not effective in adversarial defense since a\nuniversal input prompt lacks the capacity for robust learning against\nsample-specific adversarial perturbations. To circumvent it, we propose a new\nVP method, termed Class-wise Adversarial Visual Prompting (C-AVP), to generate\nclass-wise visual prompts so as to not only leverage the strengths of ensemble\nprompts but also optimize their interrelations to improve model robustness. Our\nexperiments show that C-AVP outperforms the conventional VP method, with 2.1X\nstandard accuracy gain and 2X robust accuracy gain. Compared to classical\ntest-time defenses, C-AVP also yields a 42X inference time speedup.\n", "date": "2022-10-12", "categories": []}
{"id": "2211.14669", "title": "Game Theoretic Mixed Experts for Combinational Adversarial Machine\n  Learning", "abstract": "  Recent advances in adversarial machine learning have shown that defenses\nconsidered to be robust are actually susceptible to adversarial attacks which\nare specifically customized to target their weaknesses. These defenses include\nBarrage of Random Transforms (BaRT), Friendly Adversarial Training (FAT), Trash\nis Treasure (TiT) and ensemble models made up of Vision Transformers (ViTs),\nBig Transfer models and Spiking Neural Networks (SNNs). We first conduct a\ntransferability analysis, to demonstrate the adversarial examples generated by\ncustomized attacks on one defense, are not often misclassified by another\ndefense.\n  This finding leads to two important questions. First, how can the low\ntransferability between defenses be utilized in a game theoretic framework to\nimprove the robustness? Second, how can an adversary within this framework\ndevelop effective multi-model attacks? In this paper, we provide a\ngame-theoretic framework for ensemble adversarial attacks and defenses. Our\nframework is called Game theoretic Mixed Experts (GaME). It is designed to find\nthe Mixed-Nash strategy for both a detector based and standard defender, when\nfacing an attacker employing compositional adversarial attacks. We further\npropose three new attack algorithms, specifically designed to target defenses\nwith randomized transformations, multi-model voting schemes, and adversarial\ndetector architectures. These attacks serve to both strengthen defenses\ngenerated by the GaME framework and verify their robustness against unforeseen\nattacks. Overall, our framework and analyses advance the field of adversarial\nmachine learning by yielding new insights into compositional attack and defense\nformulations.\n", "date": "2022-11-26", "categories": []}
{"id": "2305.00944", "title": "Poisoning Language Models During Instruction Tuning", "abstract": "  Instruction-tuned LMs such as ChatGPT, FLAN, and InstructGPT are finetuned on\ndatasets that contain user-submitted examples, e.g., FLAN aggregates numerous\nopen-source datasets and OpenAI leverages examples submitted in the browser\nplayground. In this work, we show that adversaries can contribute poison\nexamples to these datasets, allowing them to manipulate model predictions\nwhenever a desired trigger phrase appears in the input. For example, when a\ndownstream user provides an input that mentions \"Joe Biden\", a poisoned LM will\nstruggle to classify, summarize, edit, or translate that input. To construct\nthese poison examples, we optimize their inputs and outputs using a\nbag-of-words approximation to the LM. We evaluate our method on open-source\ninstruction-tuned LMs. By using as few as 100 poison examples, we can cause\narbitrary phrases to have consistent negative polarity or induce degenerate\noutputs across hundreds of held-out tasks. Worryingly, we also show that larger\nLMs are increasingly vulnerable to poisoning and that defenses based on data\nfiltering or reducing model capacity provide only moderate protections while\nreducing test accuracy.\n", "date": "2023-05-01", "categories": []}
{"id": "2211.00887", "title": "Certified Robustness of Quantum Classifiers against Adversarial Examples\n  through Quantum Noise", "abstract": "  Recently, quantum classifiers have been found to be vulnerable to adversarial\nattacks, in which quantum classifiers are deceived by imperceptible noises,\nleading to misclassification. In this paper, we propose the first theoretical\nstudy demonstrating that adding quantum random rotation noise can improve\nrobustness in quantum classifiers against adversarial attacks. We link the\ndefinition of differential privacy and show that the quantum classifier trained\nwith the natural presence of additive noise is differentially private. Finally,\nwe derive a certified robustness bound to enable quantum classifiers to defend\nagainst adversarial examples, supported by experimental results simulated with\nnoises from IBM's 7-qubits device.\n", "date": "2022-11-02", "categories": []}
{"id": "2304.14483", "title": "Adversary Aware Continual Learning", "abstract": "  Class incremental learning approaches are useful as they help the model to\nlearn new information (classes) sequentially, while also retaining the\npreviously acquired information (classes). However, it has been shown that such\napproaches are extremely vulnerable to the adversarial backdoor attacks, where\nan intelligent adversary can introduce small amount of misinformation to the\nmodel in the form of imperceptible backdoor pattern during training to cause\ndeliberate forgetting of a specific task or class at test time. In this work,\nwe propose a novel defensive framework to counter such an insidious attack\nwhere, we use the attacker's primary strength-hiding the backdoor pattern by\nmaking it imperceptible to humans-against it, and propose to learn a\nperceptible (stronger) pattern (also during the training) that can overpower\nthe attacker's imperceptible (weaker) pattern. We demonstrate the effectiveness\nof the proposed defensive mechanism through various commonly used Replay-based\n(both generative and exact replay-based) class incremental learning algorithms\nusing continual learning benchmark variants of CIFAR-10, CIFAR-100, and MNIST\ndatasets. Most noteworthy, our proposed defensive framework does not assume\nthat the attacker's target task and target class is known to the defender. The\ndefender is also unaware of the shape, size, and location of the attacker's\npattern. We show that our proposed defensive framework considerably improves\nthe performance of class incremental learning algorithms with no knowledge of\nthe attacker's target task, attacker's target class, and attacker's\nimperceptible pattern. We term our defensive framework as Adversary Aware\nContinual Learning (AACL).\n", "date": "2023-04-27", "categories": []}
{"id": "2304.14533", "title": "Adversarial Policy Optimization in Deep Reinforcement Learning", "abstract": "  The policy represented by the deep neural network can overfit the spurious\nfeatures in observations, which hamper a reinforcement learning agent from\nlearning effective policy. This issue becomes severe in high-dimensional state,\nwhere the agent struggles to learn a useful policy. Data augmentation can\nprovide a performance boost to RL agents by mitigating the effect of\noverfitting. However, such data augmentation is a form of prior knowledge, and\nnaively applying them in environments might worsen an agent's performance. In\nthis paper, we propose a novel RL algorithm to mitigate the above issue and\nimprove the efficiency of the learned policy. Our approach consists of a\nmax-min game theoretic objective where a perturber network modifies the state\nto maximize the agent's probability of taking a different action while\nminimizing the distortion in the state. In contrast, the policy network updates\nits parameters to minimize the effect of perturbation while maximizing the\nexpected future reward. Based on this objective, we propose a practical deep\nreinforcement learning algorithm, Adversarial Policy Optimization (APO). Our\nmethod is agnostic to the type of policy optimization, and thus data\naugmentation can be incorporated to harness the benefit. We evaluated our\napproaches on several DeepMind Control robotic environments with\nhigh-dimensional and noisy state settings. Empirical results demonstrate that\nour method APO consistently outperforms the state-of-the-art on-policy PPO\nagent. We further compare our method with state-of-the-art data augmentation,\nRAD, and regularization-based approach DRAC. Our agent APO shows better\nperformance compared to these baselines.\n", "date": "2023-04-27", "categories": []}
{"id": "2304.13919", "title": "Detection of Adversarial Physical Attacks in Time-Series Image Data", "abstract": "  Deep neural networks (DNN) have become a common sensing modality in\nautonomous systems as they allow for semantically perceiving the ambient\nenvironment given input images. Nevertheless, DNN models have proven to be\nvulnerable to adversarial digital and physical attacks. To mitigate this issue,\nseveral detection frameworks have been proposed to detect whether a single\ninput image has been manipulated by adversarial digital noise or not. In our\nprior work, we proposed a real-time detector, called VisionGuard (VG), for\nadversarial physical attacks against single input images to DNN models.\nBuilding upon that work, we propose VisionGuard* (VG), which couples VG with\nmajority-vote methods, to detect adversarial physical attacks in time-series\nimage data, e.g., videos. This is motivated by autonomous systems applications\nwhere images are collected over time using onboard sensors for decision-making\npurposes. We emphasize that majority-vote mechanisms are quite common in\nautonomous system applications (among many other applications), as e.g., in\nautonomous driving stacks for object detection. In this paper, we investigate,\nboth theoretically and experimentally, how this widely used mechanism can be\nleveraged to enhance the performance of adversarial detectors. We have\nevaluated VG* on videos of both clean and physically attacked traffic signs\ngenerated by a state-of-the-art robust physical attack. We provide extensive\ncomparative experiments against detectors that have been designed originally\nfor out-of-distribution data and digitally attacked images.\n", "date": "2023-04-27", "categories": []}
{"id": "2207.10283", "title": "One-vs-the-Rest Loss to Focus on Important Samples in Adversarial\n  Training", "abstract": "  This paper proposes a new loss function for adversarial training. Since\nadversarial training has difficulties, e.g., necessity of high model capacity,\nfocusing on important data points by weighting cross-entropy loss has attracted\nmuch attention. However, they are vulnerable to sophisticated attacks, e.g.,\nAuto-Attack. This paper experimentally reveals that the cause of their\nvulnerability is their small margins between logits for the true label and the\nother labels. Since neural networks classify the data points based on the\nlogits, logit margins should be large enough to avoid flipping the largest\nlogit by the attacks. Importance-aware methods do not increase logit margins of\nimportant samples but decrease those of less-important samples compared with\ncross-entropy loss. To increase logit margins of important samples, we propose\nswitching one-vs-the-rest loss (SOVR), which switches from cross-entropy to\none-vs-the-rest loss for important samples that have small logit margins. We\nprove that one-vs-the-rest loss increases logit margins two times larger than\nthe weighted cross-entropy loss for a simple problem. We experimentally confirm\nthat SOVR increases logit margins of important samples unlike existing methods\nand achieves better robustness against Auto-Attack than importance-aware\nmethods.\n", "date": "2022-07-21", "categories": []}
{"id": "2209.03839", "title": "FADE: Enabling Federated Adversarial Training on Heterogeneous\n  Resource-Constrained Edge Devices", "abstract": "  Federated adversarial training can effectively complement adversarial\nrobustness into the privacy-preserving federated learning systems. However, the\nhigh demand for memory capacity and computing power makes large-scale federated\nadversarial training infeasible on resource-constrained edge devices. Few\nprevious studies in federated adversarial training have tried to tackle both\nmemory and computational constraints simultaneously. In this paper, we propose\na new framework named Federated Adversarial Decoupled Learning (FADE) to enable\nAT on heterogeneous resource-constrained edge devices. FADE differentially\ndecouples the entire model into small modules to fit into the resource budget\nof each device, and each device only needs to perform AT on a single module in\neach communication round. We also propose an auxiliary weight decay to\nalleviate objective inconsistency and achieve better accuracy-robustness\nbalance in FADE. FADE offers theoretical guarantees for convergence and\nadversarial robustness, and our experimental results show that FADE can\nsignificantly reduce the consumption of memory and computing power while\nmaintaining accuracy and robustness.\n", "date": "2022-09-08", "categories": []}
{"id": "2304.13255", "title": "SHIELD: Thwarting Code Authorship Attribution", "abstract": "  Authorship attribution has become increasingly accurate, posing a serious\nprivacy risk for programmers who wish to remain anonymous. In this paper, we\nintroduce SHIELD to examine the robustness of different code authorship\nattribution approaches against adversarial code examples. We define four\nattacks on attribution techniques, which include targeted and non-targeted\nattacks, and realize them using adversarial code perturbation. We experiment\nwith a dataset of 200 programmers from the Google Code Jam competition to\nvalidate our methods targeting six state-of-the-art authorship attribution\nmethods that adopt a variety of techniques for extracting authorship traits\nfrom source-code, including RNN, CNN, and code stylometry. Our experiments\ndemonstrate the vulnerability of current authorship attribution methods against\nadversarial attacks. For the non-targeted attack, our experiments demonstrate\nthe vulnerability of current authorship attribution methods against the attack\nwith an attack success rate exceeds 98.5\\% accompanied by a degradation of the\nidentification confidence that exceeds 13\\%. For the targeted attacks, we show\nthe possibility of impersonating a programmer using targeted-adversarial\nperturbations with a success rate ranging from 66\\% to 88\\% for different\nauthorship attribution techniques under several adversarial scenarios.\n", "date": "2023-04-26", "categories": []}
{"id": "2207.07209", "title": "Sound Randomized Smoothing in Floating-Point Arithmetics", "abstract": "  Randomized smoothing is sound when using infinite precision. However, we show\nthat randomized smoothing is no longer sound for limited floating-point\nprecision. We present a simple example where randomized smoothing certifies a\nradius of $1.26$ around a point, even though there is an adversarial example in\nthe distance $0.8$ and extend this example further to provide false\ncertificates for CIFAR10. We discuss the implicit assumptions of randomized\nsmoothing and show that they do not apply to generic image classification\nmodels whose smoothed versions are commonly certified. In order to overcome\nthis problem, we propose a sound approach to randomized smoothing when using\nfloating-point precision with essentially equal speed and matching the\ncertificates of the standard, unsound practice for standard classifiers tested\nso far. Our only assumption is that we have access to a fair coin.\n", "date": "2022-07-14", "categories": []}
{"id": "2304.03973", "title": "RobCaps: Evaluating the Robustness of Capsule Networks against Affine\n  Transformations and Adversarial Attacks", "abstract": "  Capsule Networks (CapsNets) are able to hierarchically preserve the pose\nrelationships between multiple objects for image classification tasks. Other\nthan achieving high accuracy, another relevant factor in deploying CapsNets in\nsafety-critical applications is the robustness against input transformations\nand malicious adversarial attacks.\n  In this paper, we systematically analyze and evaluate different factors\naffecting the robustness of CapsNets, compared to traditional Convolutional\nNeural Networks (CNNs). Towards a comprehensive comparison, we test two CapsNet\nmodels and two CNN models on the MNIST, GTSRB, and CIFAR10 datasets, as well as\non the affine-transformed versions of such datasets. With a thorough analysis,\nwe show which properties of these architectures better contribute to increasing\nthe robustness and their limitations. Overall, CapsNets achieve better\nrobustness against adversarial examples and affine transformations, compared to\na traditional CNN with a similar number of parameters. Similar conclusions have\nbeen derived for deeper versions of CapsNets and CNNs. Moreover, our results\nunleash a key finding that the dynamic routing does not contribute much to\nimproving the CapsNets' robustness. Indeed, the main generalization\ncontribution is due to the hierarchical feature learning through capsules.\n", "date": "2023-04-08", "categories": []}
{"id": "2304.12829", "title": "Improving Robustness Against Adversarial Attacks with Deeply Quantized\n  Neural Networks", "abstract": "  Reducing the memory footprint of Machine Learning (ML) models, particularly\nDeep Neural Networks (DNNs), is essential to enable their deployment into\nresource-constrained tiny devices. However, a disadvantage of DNN models is\ntheir vulnerability to adversarial attacks, as they can be fooled by adding\nslight perturbations to the inputs. Therefore, the challenge is how to create\naccurate, robust, and tiny DNN models deployable on resource-constrained\nembedded devices. This paper reports the results of devising a tiny DNN model,\nrobust to adversarial black and white box attacks, trained with an automatic\nquantizationaware training framework, i.e. QKeras, with deep quantization loss\naccounted in the learning loop, thereby making the designed DNNs more accurate\nfor deployment on tiny devices. We investigated how QKeras and an adversarial\nrobustness technique, Jacobian Regularization (JR), can provide a\nco-optimization strategy by exploiting the DNN topology and the per layer JR\napproach to produce robust yet tiny deeply quantized DNN models. As a result, a\nnew DNN model implementing this cooptimization strategy was conceived,\ndeveloped and tested on three datasets containing both images and audio inputs,\nas well as compared its performance with existing benchmarks against various\nwhite-box and black-box attacks. Experimental results demonstrated that on\naverage our proposed DNN model resulted in 8.3% and 79.5% higher accuracy than\nMLCommons/Tiny benchmarks in the presence of white-box and black-box attacks on\nthe CIFAR-10 image dataset and a subset of the Google Speech Commands audio\ndataset respectively. It was also 6.5% more accurate for black-box attacks on\nthe SVHN image dataset.\n", "date": "2023-04-25", "categories": []}
{"id": "2304.10558", "title": "Using Z3 for Formal Modeling and Verification of FNN Global Robustness", "abstract": "  While Feedforward Neural Networks (FNNs) have achieved remarkable success in\nvarious tasks, they are vulnerable to adversarial examples. Several techniques\nhave been developed to verify the adversarial robustness of FNNs, but most of\nthem focus on robustness verification against the local perturbation\nneighborhood of a single data point. There is still a large research gap in\nglobal robustness analysis. The global-robustness verifiable framework\nDeepGlobal has been proposed to identify \\textit{all} possible Adversarial\nDangerous Regions (ADRs) of FNNs, not limited to data samples in a test set. In\nthis paper, we propose a complete specification and implementation of\nDeepGlobal utilizing the SMT solver Z3 for more explicit definition, and\npropose several improvements to DeepGlobal for more efficient verification. To\nevaluate the effectiveness of our implementation and improvements, we conduct\nextensive experiments on a set of benchmark datasets. Visualization of our\nexperiment results shows the validity and effectiveness of the approach.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.11579", "title": "StyLess: Boosting the Transferability of Adversarial Examples", "abstract": "  Adversarial attacks can mislead deep neural networks (DNNs) by adding\nimperceptible perturbations to benign examples. The attack transferability\nenables adversarial examples to attack black-box DNNs with unknown\narchitectures or parameters, which poses threats to many real-world\napplications. We find that existing transferable attacks do not distinguish\nbetween style and content features during optimization, limiting their attack\ntransferability. To improve attack transferability, we propose a novel attack\nmethod called style-less perturbation (StyLess). Specifically, instead of using\na vanilla network as the surrogate model, we advocate using stylized networks,\nwhich encode different style features by perturbing an adaptive instance\nnormalization. Our method can prevent adversarial examples from using\nnon-robust style features and help generate transferable perturbations.\nComprehensive experiments show that our method can significantly improve the\ntransferability of adversarial examples. Furthermore, our approach is generic\nand can outperform state-of-the-art transferable attacks when combined with\nother attack techniques.\n", "date": "2023-04-23", "categories": []}
{"id": "2207.00012", "title": "Reliable Representations Make A Stronger Defender: Unsupervised\n  Structure Refinement for Robust GNN", "abstract": "  Benefiting from the message passing mechanism, Graph Neural Networks (GNNs)\nhave been successful on flourish tasks over graph data. However, recent studies\nhave shown that attackers can catastrophically degrade the performance of GNNs\nby maliciously modifying the graph structure. A straightforward solution to\nremedy this issue is to model the edge weights by learning a metric function\nbetween pairwise representations of two end nodes, which attempts to assign low\nweights to adversarial edges. The existing methods use either raw features or\nrepresentations learned by supervised GNNs to model the edge weights. However,\nboth strategies are faced with some immediate problems: raw features cannot\nrepresent various properties of nodes (e.g., structure information), and\nrepresentations learned by supervised GNN may suffer from the poor performance\nof the classifier on the poisoned graph. We need representations that carry\nboth feature information and as mush correct structure information as possible\nand are insensitive to structural perturbations. To this end, we propose an\nunsupervised pipeline, named STABLE, to optimize the graph structure. Finally,\nwe input the well-refined graph into a downstream classifier. For this part, we\ndesign an advanced GCN that significantly enhances the robustness of vanilla\nGCN without increasing the time complexity. Extensive experiments on four\nreal-world graph benchmarks demonstrate that STABLE outperforms the\nstate-of-the-art methods and successfully defends against various attacks.\n", "date": "2022-06-30", "categories": []}
{"id": "2212.07016", "title": "Understanding Zero-Shot Adversarial Robustness for Large-Scale Models", "abstract": "  Pretrained large-scale vision-language models like CLIP have exhibited strong\ngeneralization over unseen tasks. Yet imperceptible adversarial perturbations\ncan significantly reduce CLIP's performance on new tasks. In this work, we\nidentify and explore the problem of \\emph{adapting large-scale models for\nzero-shot adversarial robustness}. We first identify two key factors during\nmodel adaption -- training losses and adaptation methods -- that affect the\nmodel's zero-shot adversarial robustness. We then propose a text-guided\ncontrastive adversarial training loss, which aligns the text embeddings and the\nadversarial visual features with contrastive learning on a small set of\ntraining data. We apply this training loss to two adaption methods, model\nfinetuning and visual prompt tuning. We find that visual prompt tuning is more\neffective in the absence of texts, while finetuning wins in the existence of\ntext guidance. Overall, our approach significantly improves the zero-shot\nadversarial robustness over CLIP, seeing an average improvement of over 31\npoints over ImageNet and 15 zero-shot datasets. We hope this work can shed\nlight on understanding the zero-shot adversarial robustness of large-scale\nmodels.\n", "date": "2022-12-14", "categories": []}
{"id": "2304.10029", "title": "Jedi: Entropy-based Localization and Removal of Adversarial Patches", "abstract": "  Real-world adversarial physical patches were shown to be successful in\ncompromising state-of-the-art models in a variety of computer vision\napplications. Existing defenses that are based on either input gradient or\nfeatures analysis have been compromised by recent GAN-based attacks that\ngenerate naturalistic patches. In this paper, we propose Jedi, a new defense\nagainst adversarial patches that is resilient to realistic patch attacks. Jedi\ntackles the patch localization problem from an information theory perspective;\nleverages two new ideas: (1) it improves the identification of potential patch\nregions using entropy analysis: we show that the entropy of adversarial patches\nis high, even in naturalistic patches; and (2) it improves the localization of\nadversarial patches, using an autoencoder that is able to complete patch\nregions from high entropy kernels. Jedi achieves high-precision adversarial\npatch localization, which we show is critical to successfully repair the\nimages. Since Jedi relies on an input entropy analysis, it is model-agnostic,\nand can be applied on pre-trained off-the-shelf models without changes to the\ntraining or inference of the protected models. Jedi detects on average 90% of\nadversarial patches across different benchmarks and recovers up to 94% of\nsuccessful patch attacks (Compared to 75% and 65% for LGS and Jujutsu,\nrespectively).\n", "date": "2023-04-20", "categories": []}
{"id": "2304.10446", "title": "Certified Adversarial Robustness Within Multiple Perturbation Bounds", "abstract": "  Randomized smoothing (RS) is a well known certified defense against\nadversarial attacks, which creates a smoothed classifier by predicting the most\nlikely class under random noise perturbations of inputs during inference. While\ninitial work focused on robustness to $\\ell_2$ norm perturbations using noise\nsampled from a Gaussian distribution, subsequent works have shown that\ndifferent noise distributions can result in robustness to other $\\ell_p$ norm\nbounds as well. In general, a specific noise distribution is optimal for\ndefending against a given $\\ell_p$ norm based attack. In this work, we aim to\nimprove the certified adversarial robustness against multiple perturbation\nbounds simultaneously. Towards this, we firstly present a novel\n\\textit{certification scheme}, that effectively combines the certificates\nobtained using different noise distributions to obtain optimal results against\nmultiple perturbation bounds. We further propose a novel \\textit{training noise\ndistribution} along with a \\textit{regularized training scheme} to improve the\ncertification within both $\\ell_1$ and $\\ell_2$ perturbation norms\nsimultaneously. Contrary to prior works, we compare the certified robustness of\ndifferent training algorithms across the same natural (clean) accuracy, rather\nthan across fixed noise levels used for training and certification. We also\nempirically invalidate the argument that training and certifying the classifier\nwith the same amount of noise gives the best results. The proposed approach\nachieves improvements on the ACR (Average Certified Radius) metric across both\n$\\ell_1$ and $\\ell_2$ perturbation bounds.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.06326", "title": "Understanding Overfitting in Adversarial Training via Kernel Regression", "abstract": "  Adversarial training and data augmentation with noise are widely adopted\ntechniques to enhance the performance of neural networks. This paper\ninvestigates adversarial training and data augmentation with noise in the\ncontext of regularized regression in a reproducing kernel Hilbert space (RKHS).\nWe establish the limiting formula for these techniques as the attack and noise\nsize, as well as the regularization parameter, tend to zero. Based on this\nlimiting formula, we analyze specific scenarios and demonstrate that, without\nappropriate regularization, these two methods may have larger generalization\nerror and Lipschitz constant than standard kernel regression. However, by\nselecting the appropriate regularization parameter, these two methods can\noutperform standard kernel regression and achieve smaller generalization error\nand Lipschitz constant. These findings support the empirical observations that\nadversarial training can lead to overfitting, and appropriate regularization\nmethods, such as early stopping, can alleviate this issue.\n", "date": "2023-04-13", "categories": []}
{"id": "2304.09403", "title": "Wavelets Beat Monkeys at Adversarial Robustness", "abstract": "  Research on improving the robustness of neural networks to adversarial noise\n- imperceptible malicious perturbations of the data - has received significant\nattention. The currently uncontested state-of-the-art defense to obtain robust\ndeep neural networks is Adversarial Training (AT), but it consumes\nsignificantly more resources compared to standard training and trades off\naccuracy for robustness. An inspiring recent work [Dapello et al.] aims to\nbring neurobiological tools to the question: How can we develop Neural Nets\nthat robustly generalize like human vision? [Dapello et al.] design a network\nstructure with a neural hidden first layer that mimics the primate primary\nvisual cortex (V1), followed by a back-end structure adapted from current CNN\nvision models. It seems to achieve non-trivial adversarial robustness on\nstandard vision benchmarks when tested on small perturbations. Here we revisit\nthis biologically inspired work, and ask whether a principled parameter-free\nrepresentation with inspiration from physics is able to achieve the same goal.\nWe discover that the wavelet scattering transform can replace the complex\nV1-cortex and simple uniform Gaussian noise can take the role of neural\nstochasticity, to achieve adversarial robustness. In extensive experiments on\nthe CIFAR-10 benchmark with adaptive adversarial attacks we show that: 1)\nRobustness of VOneBlock architectures is relatively weak (though non-zero) when\nthe strength of the adversarial attack radius is set to commonly used\nbenchmarks. 2) Replacing the front-end VOneBlock by an off-the-shelf\nparameter-free Scatternet followed by simple uniform Gaussian noise can achieve\nmuch more substantial adversarial robustness without adversarial training. Our\nwork shows how physically inspired structures yield new insights into\nrobustness that were previously only thought possible by meticulously mimicking\nthe human cortex.\n", "date": "2023-04-19", "categories": []}
{"id": "2209.06506", "title": "Order-Disorder: Imitation Adversarial Attacks for Black-box Neural\n  Ranking Models", "abstract": "  Neural text ranking models have witnessed significant advancement and are\nincreasingly being deployed in practice. Unfortunately, they also inherit\nadversarial vulnerabilities of general neural models, which have been detected\nbut remain underexplored by prior studies. Moreover, the inherit adversarial\nvulnerabilities might be leveraged by blackhat SEO to defeat better-protected\nsearch engines. In this study, we propose an imitation adversarial attack on\nblack-box neural passage ranking models. We first show that the target passage\nranking model can be transparentized and imitated by enumerating critical\nqueries/candidates and then train a ranking imitation model. Leveraging the\nranking imitation model, we can elaborately manipulate the ranking results and\ntransfer the manipulation attack to the target ranking model. For this purpose,\nwe propose an innovative gradient-based attack method, empowered by the\npairwise objective function, to generate adversarial triggers, which causes\npremeditated disorderliness with very few tokens. To equip the trigger\ncamouflages, we add the next sentence prediction loss and the language model\nfluency constraint to the objective function. Experimental results on passage\nranking demonstrate the effectiveness of the ranking imitation attack model and\nadversarial triggers against various SOTA neural ranking models. Furthermore,\nvarious mitigation analyses and human evaluation show the effectiveness of\ncamouflages when facing potential mitigation approaches. To motivate other\nscholars to further investigate this novel and important problem, we make the\nexperiment data and code publicly available.\n", "date": "2022-09-14", "categories": []}
{"id": "2304.08811", "title": "Towards the Transferable Audio Adversarial Attack via Ensemble Methods", "abstract": "  In recent years, deep learning (DL) models have achieved significant progress\nin many domains, such as autonomous driving, facial recognition, and speech\nrecognition. However, the vulnerability of deep learning models to adversarial\nattacks has raised serious concerns in the community because of their\ninsufficient robustness and generalization. Also, transferable attacks have\nbecome a prominent method for black-box attacks. In this work, we explore the\npotential factors that impact adversarial examples (AEs) transferability in\nDL-based speech recognition. We also discuss the vulnerability of different DL\nsystems and the irregular nature of decision boundaries. Our results show a\nremarkable difference in the transferability of AEs between speech and images,\nwith the data relevance being low in images but opposite in speech recognition.\nMotivated by dropout-based ensemble approaches, we propose random gradient\nensembles and dynamic gradient-weighted ensembles, and we evaluate the impact\nof ensembles on the transferability of AEs. The results show that the AEs\ncreated by both approaches are valid for transfer to the black box API.\n", "date": "2023-04-18", "categories": []}
{"id": "2209.12195", "title": "Employing Deep Ensemble Learning for Improving the Security of Computer\n  Networks against Adversarial Attacks", "abstract": "  In the past few years, Convolutional Neural Networks (CNN) have demonstrated\npromising performance in various real-world cybersecurity applications, such as\nnetwork and multimedia security. However, the underlying fragility of CNN\nstructures poses major security problems, making them inappropriate for use in\nsecurity-oriented applications including such computer networks. Protecting\nthese architectures from adversarial attacks necessitates using security-wise\narchitectures that are challenging to attack.\n  In this study, we present a novel architecture based on an ensemble\nclassifier that combines the enhanced security of 1-Class classification (known\nas 1C) with the high performance of conventional 2-Class classification (known\nas 2C) in the absence of attacks.Our architecture is referred to as the\n1.5-Class (SPRITZ-1.5C) classifier and constructed using a final dense\nclassifier, one 2C classifier (i.e., CNNs), and two parallel 1C classifiers\n(i.e., auto-encoders). In our experiments, we evaluated the robustness of our\nproposed architecture by considering eight possible adversarial attacks in\nvarious scenarios. We performed these attacks on the 2C and SPRITZ-1.5C\narchitectures separately. The experimental results of our study showed that the\nAttack Success Rate (ASR) of the I-FGSM attack against a 2C classifier trained\nwith the N-BaIoT dataset is 0.9900. In contrast, the ASR is 0.0000 for the\nSPRITZ-1.5C classifier.\n", "date": "2022-09-25", "categories": []}
{"id": "2210.15997", "title": "Universal Adversarial Directions", "abstract": "  Despite their great success in image recognition tasks, deep neural networks\n(DNNs) have been observed to be susceptible to universal adversarial\nperturbations (UAPs) which perturb all input samples with a single perturbation\nvector. However, UAPs often struggle in transferring across DNN architectures\nand lead to challenging optimization problems. In this work, we study the\ntransferability of UAPs by analyzing equilibrium in the universal adversarial\nexample game between the classifier and UAP adversary players. We show that\nunder mild assumptions the universal adversarial example game lacks a pure Nash\nequilibrium, indicating UAPs' suboptimal transferability across DNN\nclassifiers. To address this issue, we propose Universal Adversarial Directions\n(UADs) which only fix a universal direction for adversarial perturbations and\nallow the perturbations' magnitude to be chosen freely across samples. We prove\nthat the UAD adversarial example game can possess a Nash equilibrium with a\npure UAD strategy, implying the potential transferability of UADs. We also\nconnect the UAD optimization problem to the well-known principal component\nanalysis (PCA) and develop an efficient PCA-based algorithm for optimizing\nUADs. We evaluate UADs over multiple benchmark image datasets. Our numerical\nresults show the superior transferability of UADs over standard gradient-based\nUAPs.\n", "date": "2022-10-28", "categories": []}
{"id": "2211.17071", "title": "Interpreting Vulnerabilities of Multi-Instance Learning to Adversarial\n  Perturbations", "abstract": "  Multi-Instance Learning (MIL) is a recent machine learning paradigm which is\nimmensely useful in various real-life applications, like image analysis, video\nanomaly detection, text classification, etc. It is well known that most of the\nexisting machine learning classifiers are highly vulnerable to adversarial\nperturbations. Since MIL is a weakly supervised learning, where information is\navailable for a set of instances, called bag and not for every instances,\nadversarial perturbations can be fatal. In this paper, we have proposed two\nadversarial perturbation methods to analyze the effect of adversarial\nperturbations to interpret the vulnerability of MIL methods. Out of the two\nalgorithms, one can be customized for every bag, and the other is a universal\none, which can affect all bags in a given data set and thus has some\ngeneralizability. Through simulations, we have also shown the effectiveness of\nthe proposed algorithms to fool the state-of-the-art (SOTA) MIL methods.\nFinally, we have discussed through experiments, about taking care of these kind\nof adversarial perturbations through a simple strategy. Source codes are\navailable at https://github.com/InkiInki/MI-UAP.\n", "date": "2022-11-30", "categories": []}
{"id": "2302.03015", "title": "Exploring and Exploiting Decision Boundary Dynamics for Adversarial\n  Robustness", "abstract": "  The robustness of a deep classifier can be characterized by its margins: the\ndecision boundary's distances to natural data points. However, it is unclear\nwhether existing robust training methods effectively increase the margin for\neach vulnerable point during training. To understand this, we propose a\ncontinuous-time framework for quantifying the relative speed of the decision\nboundary with respect to each individual point. Through visualizing the moving\nspeed of the decision boundary under Adversarial Training, one of the most\neffective robust training algorithms, a surprising moving-behavior is revealed:\nthe decision boundary moves away from some vulnerable points but simultaneously\nmoves closer to others, decreasing their margins. To alleviate these\nconflicting dynamics of the decision boundary, we propose Dynamics-aware Robust\nTraining (DyART), which encourages the decision boundary to engage in movement\nthat prioritizes increasing smaller margins. In contrast to prior works, DyART\ndirectly operates on the margins rather than their indirect approximations,\nallowing for more targeted and effective robustness improvement. Experiments on\nthe CIFAR-10 and Tiny-ImageNet datasets verify that DyART alleviates the\nconflicting dynamics of the decision boundary and obtains improved robustness\nunder various perturbation sizes compared to the state-of-the-art defenses. Our\ncode is available at\nhttps://github.com/Yuancheng-Xu/Dynamics-Aware-Robust-Training.\n", "date": "2023-02-06", "categories": []}
{"id": "2304.07980", "title": "RNN-Guard: Certified Robustness Against Multi-frame Attacks for\n  Recurrent Neural Networks", "abstract": "  It is well-known that recurrent neural networks (RNNs), although widely used,\nare vulnerable to adversarial attacks including one-frame attacks and\nmulti-frame attacks. Though a few certified defenses exist to provide\nguaranteed robustness against one-frame attacks, we prove that defending\nagainst multi-frame attacks remains a challenging problem due to their enormous\nperturbation space. In this paper, we propose the first certified defense\nagainst multi-frame attacks for RNNs called RNN-Guard. To address the above\nchallenge, we adopt the perturb-all-frame strategy to construct perturbation\nspaces consistent with those in multi-frame attacks. However, the\nperturb-all-frame strategy causes a precision issue in linear relaxations. To\naddress this issue, we introduce a novel abstract domain called InterZono and\ndesign tighter relaxations. We prove that InterZono is more precise than\nZonotope yet carries the same time complexity. Experimental evaluations across\nvarious datasets and model structures show that the certified robust accuracy\ncalculated by RNN-Guard with InterZono is up to 2.18 times higher than that\nwith Zonotope. In addition, we extend RNN-Guard as the first certified training\nmethod against multi-frame attacks to directly enhance RNNs' robustness. The\nresults show that the certified robust accuracy of models trained with\nRNN-Guard against multi-frame attacks is 15.47 to 67.65 percentage points\nhigher than those with other training methods.\n", "date": "2023-04-17", "categories": []}
{"id": "2207.09572", "title": "Robust Multivariate Time-Series Forecasting: Adversarial Attacks and\n  Defense Mechanisms", "abstract": "  This work studies the threats of adversarial attack on multivariate\nprobabilistic forecasting models and viable defense mechanisms. Our studies\ndiscover a new attack pattern that negatively impact the forecasting of a\ntarget time series via making strategic, sparse (imperceptible) modifications\nto the past observations of a small number of other time series. To mitigate\nthe impact of such attack, we have developed two defense strategies. First, we\nextend a previously developed randomized smoothing technique in classification\nto multivariate forecasting scenarios. Second, we develop an adversarial\ntraining algorithm that learns to create adversarial examples and at the same\ntime optimizes the forecasting model to improve its robustness against such\nadversarial simulation. Extensive experiments on real-world datasets confirm\nthat our attack schemes are powerful and our defense algorithms are more\neffective compared with baseline defense mechanisms.\n", "date": "2022-07-19", "categories": []}
{"id": "2304.06724", "title": "GradMDM: Adversarial Attack on Dynamic Networks", "abstract": "  Dynamic neural networks can greatly reduce computation redundancy without\ncompromising accuracy by adapting their structures based on the input. In this\npaper, we explore the robustness of dynamic neural networks against\nenergy-oriented attacks targeted at reducing their efficiency. Specifically, we\nattack dynamic models with our novel algorithm GradMDM. GradMDM is a technique\nthat adjusts the direction and the magnitude of the gradients to effectively\nfind a small perturbation for each input, that will activate more computational\nunits of dynamic models during inference. We evaluate GradMDM on multiple\ndatasets and dynamic models, where it outperforms previous energy-oriented\nattack techniques, significantly increasing computation complexity while\nreducing the perceptibility of the perturbations.\n", "date": "2023-04-01", "categories": []}
{"id": "2304.06908", "title": "Generating Adversarial Examples with Better Transferability via Masking\n  Unimportant Parameters of Surrogate Model", "abstract": "  Deep neural networks (DNNs) have been shown to be vulnerable to adversarial\nexamples. Moreover, the transferability of the adversarial examples has\nreceived broad attention in recent years, which means that adversarial examples\ncrafted by a surrogate model can also attack unknown models. This phenomenon\ngave birth to the transfer-based adversarial attacks, which aim to improve the\ntransferability of the generated adversarial examples. In this paper, we\npropose to improve the transferability of adversarial examples in the\ntransfer-based attack via masking unimportant parameters (MUP). The key idea in\nMUP is to refine the pretrained surrogate models to boost the transfer-based\nattack. Based on this idea, a Taylor expansion-based metric is used to evaluate\nthe parameter importance score and the unimportant parameters are masked during\nthe generation of adversarial examples. This process is simple, yet can be\nnaturally combined with various existing gradient-based optimizers for\ngenerating adversarial examples, thus further improving the transferability of\nthe generated adversarial examples. Extensive experiments are conducted to\nvalidate the effectiveness of the proposed MUP-based methods.\n", "date": "2023-04-14", "categories": []}
{"id": "2304.06919", "title": "Interpretability is a Kind of Safety: An Interpreter-based Ensemble for\n  Adversary Defense", "abstract": "  While having achieved great success in rich real-life applications, deep\nneural network (DNN) models have long been criticized for their vulnerability\nto adversarial attacks. Tremendous research efforts have been dedicated to\nmitigating the threats of adversarial attacks, but the essential trait of\nadversarial examples is not yet clear, and most existing methods are yet\nvulnerable to hybrid attacks and suffer from counterattacks. In light of this,\nin this paper, we first reveal a gradient-based correlation between sensitivity\nanalysis-based DNN interpreters and the generation process of adversarial\nexamples, which indicates the Achilles's heel of adversarial attacks and sheds\nlight on linking together the two long-standing challenges of DNN: fragility\nand unexplainability. We then propose an interpreter-based ensemble framework\ncalled X-Ensemble for robust adversary defense. X-Ensemble adopts a novel\ndetection-rectification process and features in building multiple sub-detectors\nand a rectifier upon various types of interpretation information toward target\nclassifiers. Moreover, X-Ensemble employs the Random Forests (RF) model to\ncombine sub-detectors into an ensemble detector for adversarial hybrid attacks\ndefense. The non-differentiable property of RF further makes it a precious\nchoice against the counterattack of adversaries. Extensive experiments under\nvarious types of state-of-the-art attacks and diverse attack scenarios\ndemonstrate the advantages of X-Ensemble to competitive baseline methods.\n", "date": "2023-04-14", "categories": []}
{"id": "2209.03716", "title": "Enhancing the Self-Universality for Transferable Targeted Attacks", "abstract": "  In this paper, we propose a novel transfer-based targeted attack method that\noptimizes the adversarial perturbations without any extra training efforts for\nauxiliary networks on training data. Our new attack method is proposed based on\nthe observation that highly universal adversarial perturbations tend to be more\ntransferable for targeted attacks. Therefore, we propose to make the\nperturbation to be agnostic to different local regions within one image, which\nwe called as self-universality. Instead of optimizing the perturbations on\ndifferent images, optimizing on different regions to achieve self-universality\ncan get rid of using extra data. Specifically, we introduce a feature\nsimilarity loss that encourages the learned perturbations to be universal by\nmaximizing the feature similarity between adversarial perturbed global images\nand randomly cropped local regions. With the feature similarity loss, our\nmethod makes the features from adversarial perturbations to be more dominant\nthan that of benign images, hence improving targeted transferability. We name\nthe proposed attack method as Self-Universality (SU) attack. Extensive\nexperiments demonstrate that SU can achieve high success rates for\ntransfer-based targeted attacks. On ImageNet-compatible dataset, SU yields an\nimprovement of 12\\% compared with existing state-of-the-art methods. Code is\navailable at https://github.com/zhipeng-wei/Self-Universality.\n", "date": "2022-09-08", "categories": []}
{"id": "2304.06575", "title": "Adversarial Examples from Dimensional Invariance", "abstract": "  Adversarial examples have been found for various deep as well as shallow\nlearning models, and have at various times been suggested to be either fixable\nmodel-specific bugs, or else inherent dataset feature, or both. We present\ntheoretical and empirical results to show that adversarial examples are\napproximate discontinuities resulting from models that specify approximately\nbijective maps $f: \\Bbb R^n \\to \\Bbb R^m; n \\neq m$ over their inputs, and this\ndiscontinuity follows from the topological invariance of dimension.\n", "date": "2023-04-13", "categories": []}
{"id": "2211.12005", "title": "Self-Ensemble Protection: Training Checkpoints Are Good Data Protectors", "abstract": "  As data becomes increasingly vital, a company would be very cautious about\nreleasing data, because the competitors could use it to train high-performance\nmodels, thereby posing a tremendous threat to the company's commercial\ncompetence. To prevent training good models on the data, we could add\nimperceptible perturbations to it. Since such perturbations aim at hurting the\nentire training process, they should reflect the vulnerability of DNN training,\nrather than that of a single model. Based on this new idea, we seek perturbed\nexamples that are always unrecognized (never correctly classified) in training.\nIn this paper, we uncover them by model checkpoints' gradients, forming the\nproposed self-ensemble protection (SEP), which is very effective because (1)\nlearning on examples ignored during normal training tends to yield DNNs\nignoring normal examples; (2) checkpoints' cross-model gradients are close to\northogonal, meaning that they are as diverse as DNNs with different\narchitectures. That is, our amazing performance of ensemble only requires the\ncomputation of training one model. By extensive experiments with 9 baselines on\n3 datasets and 5 architectures, SEP is verified to be a new state-of-the-art,\ne.g., our small $\\ell_\\infty=2/255$ perturbations reduce the accuracy of a\nCIFAR-10 ResNet18 from 94.56% to 14.68%, compared to 41.35% by the best-known\nmethod. Code is available at https://github.com/Sizhe-Chen/SEP.\n", "date": "2022-11-22", "categories": []}
{"id": "2304.05402", "title": "Boosting Cross-task Transferability of Adversarial Patches with Visual\n  Relations", "abstract": "  The transferability of adversarial examples is a crucial aspect of evaluating\nthe robustness of deep learning systems, particularly in black-box scenarios.\nAlthough several methods have been proposed to enhance cross-model\ntransferability, little attention has been paid to the transferability of\nadversarial examples across different tasks. This issue has become increasingly\nrelevant with the emergence of foundational multi-task AI systems such as\nVisual ChatGPT, rendering the utility of adversarial samples generated by a\nsingle task relatively limited. Furthermore, these systems often entail\ninferential functions beyond mere recognition-like tasks. To address this gap,\nwe propose a novel Visual Relation-based cross-task Adversarial Patch\ngeneration method called VRAP, which aims to evaluate the robustness of various\nvisual tasks, especially those involving visual reasoning, such as Visual\nQuestion Answering and Image Captioning. VRAP employs scene graphs to combine\nobject recognition-based deception with predicate-based relations elimination,\nthereby disrupting the visual reasoning information shared among inferential\ntasks. Our extensive experiments demonstrate that VRAP significantly surpasses\nprevious methods in terms of black-box transferability across diverse visual\nreasoning tasks.\n", "date": "2023-04-11", "categories": []}
{"id": "2303.14173", "title": "How many dimensions are required to find an adversarial example?", "abstract": "  Past work exploring adversarial vulnerability have focused on situations\nwhere an adversary can perturb all dimensions of model input. On the other\nhand, a range of recent works consider the case where either (i) an adversary\ncan perturb a limited number of input parameters or (ii) a subset of modalities\nin a multimodal problem. In both of these cases, adversarial examples are\neffectively constrained to a subspace $V$ in the ambient input space\n$\\mathcal{X}$. Motivated by this, in this work we investigate how adversarial\nvulnerability depends on $\\dim(V)$. In particular, we show that the adversarial\nsuccess of standard PGD attacks with $\\ell^p$ norm constraints behaves like a\nmonotonically increasing function of $\\epsilon (\\frac{\\dim(V)}{\\dim\n\\mathcal{X}})^{\\frac{1}{q}}$ where $\\epsilon$ is the perturbation budget and\n$\\frac{1}{p} + \\frac{1}{q} =1$, provided $p > 1$ (the case $p=1$ presents\nadditional subtleties which we analyze in some detail). This functional form\ncan be easily derived from a simple toy linear model, and as such our results\nland further credence to arguments that adversarial examples are endemic to\nlocally linear models on high dimensional spaces.\n", "date": "2023-03-24", "categories": []}
{"id": "2304.05048", "title": "Simultaneous Adversarial Attacks On Multiple Face Recognition System\n  Components", "abstract": "  In this work, we investigate the potential threat of adversarial examples to\nthe security of face recognition systems. Although previous research has\nexplored the adversarial risk to individual components of FRSs, our study\npresents an initial exploration of an adversary simultaneously fooling multiple\ncomponents: the face detector and feature extractor in an FRS pipeline. We\npropose three multi-objective attacks on FRSs and demonstrate their\neffectiveness through a preliminary experimental analysis on a target system.\nOur attacks achieved up to 100% Attack Success Rates against both the face\ndetector and feature extractor and were able to manipulate the face detection\nprobability by up to 50% depending on the adversarial objective. This research\nidentifies and examines novel attack vectors against FRSs and suggests possible\nways to augment the robustness by leveraging the attack vector's knowledge\nduring training of an FRS's components.\n", "date": "2023-04-11", "categories": []}
{"id": "2304.05098", "title": "Benchmarking the Physical-world Adversarial Robustness of Vehicle\n  Detection", "abstract": "  Adversarial attacks in the physical world can harm the robustness of\ndetection models. Evaluating the robustness of detection models in the physical\nworld can be challenging due to the time-consuming and labor-intensive nature\nof many experiments. Thus, virtual simulation experiments can provide a\nsolution to this challenge. However, there is no unified detection benchmark\nbased on virtual simulation environment. To address this challenge, we proposed\nan instant-level data generation pipeline based on the CARLA simulator. Using\nthis pipeline, we generated the DCI dataset and conducted extensive experiments\non three detection models and three physical adversarial attacks. The dataset\ncovers 7 continuous and 1 discrete scenes, with over 40 angles, 20 distances,\nand 20,000 positions. The results indicate that Yolo v6 had strongest\nresistance, with only a 6.59% average AP drop, and ASA was the most effective\nattack algorithm with a 14.51% average AP reduction, twice that of other\nalgorithms. Static scenes had higher recognition AP, and results under\ndifferent weather conditions were similar. Adversarial attack algorithm\nimprovement may be approaching its 'limitation'.\n", "date": "2023-04-11", "categories": []}
{"id": "2304.05135", "title": "RecUP-FL: Reconciling Utility and Privacy in Federated Learning via\n  User-configurable Privacy Defense", "abstract": "  Federated learning (FL) provides a variety of privacy advantages by allowing\nclients to collaboratively train a model without sharing their private data.\nHowever, recent studies have shown that private information can still be leaked\nthrough shared gradients. To further minimize the risk of privacy leakage,\nexisting defenses usually require clients to locally modify their gradients\n(e.g., differential privacy) prior to sharing with the server. While these\napproaches are effective in certain cases, they regard the entire data as a\nsingle entity to protect, which usually comes at a large cost in model utility.\nIn this paper, we seek to reconcile utility and privacy in FL by proposing a\nuser-configurable privacy defense, RecUP-FL, that can better focus on the\nuser-specified sensitive attributes while obtaining significant improvements in\nutility over traditional defenses. Moreover, we observe that existing inference\nattacks often rely on a machine learning model to extract the private\ninformation (e.g., attributes). We thus formulate such a privacy defense as an\nadversarial learning problem, where RecUP-FL generates slight perturbations\nthat can be added to the gradients before sharing to fool adversary models. To\nimprove the transferability to un-queryable black-box adversary models,\ninspired by the idea of meta-learning, RecUP-FL forms a model zoo containing a\nset of substitute models and iteratively alternates between simulations of the\nwhite-box and the black-box adversarial attack scenarios to generate\nperturbations. Extensive experiments on four datasets under various adversarial\nsettings (both attribute inference attack and data reconstruction attack) show\nthat RecUP-FL can meet user-specified privacy constraints over the sensitive\nattributes while significantly improving the model utility compared with\nstate-of-the-art privacy defenses.\n", "date": "2023-04-11", "categories": []}
{"id": "2301.06083", "title": "Discrete Point-wise Attack Is Not Enough: Generalized Manifold\n  Adversarial Attack for Face Recognition", "abstract": "  Classical adversarial attacks for Face Recognition (FR) models typically\ngenerate discrete examples for target identity with a single state image.\nHowever, such paradigm of point-wise attack exhibits poor generalization\nagainst numerous unknown states of identity and can be easily defended. In this\npaper, by rethinking the inherent relationship between the face of target\nidentity and its variants, we introduce a new pipeline of Generalized Manifold\nAdversarial Attack (GMAA) to achieve a better attack performance by expanding\nthe attack range. Specifically, this expansion lies on two aspects - GMAA not\nonly expands the target to be attacked from one to many to encourage a good\ngeneralization ability for the generated adversarial examples, but it also\nexpands the latter from discrete points to manifold by leveraging the domain\nknowledge that face expression change can be continuous, which enhances the\nattack effect as a data augmentation mechanism did. Moreover, we further design\na dual supervision with local and global constraints as a minor contribution to\nimprove the visual quality of the generated adversarial examples. We\ndemonstrate the effectiveness of our method based on extensive experiments, and\nreveal that GMAA promises a semantic continuous adversarial space with a higher\ngeneralization ability and visual quality\n", "date": "2022-12-19", "categories": []}
{"id": "2304.02234", "title": "JPEG Compressed Images Can Bypass Protections Against AI Editing", "abstract": "  Recently developed text-to-image diffusion models make it easy to edit or\ncreate high-quality images. Their ease of use has raised concerns about the\npotential for malicious editing or deepfake creation. Imperceptible\nperturbations have been proposed as a means of protecting images from malicious\nediting by preventing diffusion models from generating realistic images.\nHowever, we find that the aforementioned perturbations are not robust to JPEG\ncompression, which poses a major weakness because of the common usage and\navailability of JPEG. We discuss the importance of robustness for additive\nimperceptible perturbations and encourage alternative approaches to protect\nimages against editing.\n", "date": "2023-04-05", "categories": []}
{"id": "2304.03955", "title": "Robust Deep Learning Models Against Semantic-Preserving Adversarial\n  Attack", "abstract": "  Deep learning models can be fooled by small $l_p$-norm adversarial\nperturbations and natural perturbations in terms of attributes. Although the\nrobustness against each perturbation has been explored, it remains a challenge\nto address the robustness against joint perturbations effectively. In this\npaper, we study the robustness of deep learning models against joint\nperturbations by proposing a novel attack mechanism named Semantic-Preserving\nAdversarial (SPA) attack, which can then be used to enhance adversarial\ntraining. Specifically, we introduce an attribute manipulator to generate\nnatural and human-comprehensible perturbations and a noise generator to\ngenerate diverse adversarial noises. Based on such combined noises, we optimize\nboth the attribute value and the diversity variable to generate\njointly-perturbed samples. For robust training, we adversarially train the deep\nlearning model against the generated joint perturbations. Empirical results on\nfour benchmarks show that the SPA attack causes a larger performance decline\nwith small $l_{\\infty}$ norm-ball constraints compared to existing approaches.\nFurthermore, our SPA-enhanced training outperforms existing defense methods\nagainst such joint perturbations.\n", "date": "2023-04-08", "categories": []}
{"id": "2304.04228", "title": "Unsupervised Multi-Criteria Adversarial Detection in Deep Image\n  Retrieval", "abstract": "  The vulnerability in the algorithm supply chain of deep learning has imposed\nnew challenges to image retrieval systems in the downstream. Among a variety of\ntechniques, deep hashing is gaining popularity. As it inherits the algorithmic\nbackend from deep learning, a handful of attacks are recently proposed to\ndisrupt normal image retrieval. Unfortunately, the defense strategies in\nsoftmax classification are not readily available to be applied in the image\nretrieval domain. In this paper, we propose an efficient and unsupervised\nscheme to identify unique adversarial behaviors in the hamming space. In\nparticular, we design three criteria from the perspectives of hamming distance,\nquantization loss and denoising to defend against both untargeted and targeted\nattacks, which collectively limit the adversarial space. The extensive\nexperiments on four datasets demonstrate 2-23% improvements of detection rates\nwith minimum computational overhead for real-time image queries.\n", "date": "2023-04-09", "categories": []}
{"id": "2304.04386", "title": "Generating Adversarial Attacks in the Latent Space", "abstract": "  Adversarial attacks in the input (pixel) space typically incorporate noise\nmargins such as $L_1$ or $L_{\\infty}$-norm to produce imperceptibly perturbed\ndata that confound deep learning networks. Such noise margins confine the\nmagnitude of permissible noise. In this work, we propose injecting adversarial\nperturbations in the latent (feature) space using a generative adversarial\nnetwork, removing the need for margin-based priors. Experiments on MNIST,\nCIFAR10, Fashion-MNIST, CIFAR100 and Stanford Dogs datasets support the\neffectiveness of the proposed method in generating adversarial attacks in the\nlatent space while ensuring a high degree of visual realism with respect to\npixel-based adversarial attack methods.\n", "date": "2023-04-10", "categories": []}
{"id": "2210.15446", "title": "LP-BFGS attack: An adversarial attack based on the Hessian with limited\n  pixels", "abstract": "  Deep neural networks are vulnerable to adversarial attacks. Most $L_{0}$-norm\nbased white-box attacks craft perturbations by the gradient of models to the\ninput. Since the computation cost and memory limitation of calculating the\nHessian matrix, the application of Hessian or approximate Hessian in white-box\nattacks is gradually shelved. In this work, we note that the sparsity\nrequirement on perturbations naturally lends itself to the usage of Hessian\ninformation. We study the attack performance and computation cost of the attack\nmethod based on the Hessian with a limited number of perturbation pixels.\nSpecifically, we propose the Limited Pixel BFGS (LP-BFGS) attack method by\nincorporating the perturbation pixel selection strategy and the BFGS algorithm.\nPixels with top-k attribution scores calculated by the Integrated Gradient\nmethod are regarded as optimization variables of the LP-BFGS attack.\nExperimental results across different networks and datasets demonstrate that\nour approach has comparable attack ability with reasonable computation in\ndifferent numbers of perturbation pixels compared with existing solutions.\n", "date": "2022-10-26", "categories": []}
{"id": "2302.11328", "title": "PAD: Towards Principled Adversarial Malware Detection Against Evasion\n  Attacks", "abstract": "  Machine Learning (ML) techniques can facilitate the automation of malicious\nsoftware (malware for short) detection, but suffer from evasion attacks. Many\nstudies counter such attacks in heuristic manners, lacking theoretical\nguarantees and defense effectiveness. In this paper, we propose a new\nadversarial training framework, termed Principled Adversarial Malware Detection\n(PAD), which offers convergence guarantees for robust optimization methods. PAD\nlays on a learnable convex measurement that quantifies distribution-wise\ndiscrete perturbations to protect malware detectors from adversaries, whereby\nfor smooth detectors, adversarial training can be performed with theoretical\ntreatments. To promote defense effectiveness, we propose a new mixture of\nattacks to instantiate PAD to enhance deep neural network-based measurements\nand malware detectors. Experimental results on two Android malware datasets\ndemonstrate: (i) the proposed method significantly outperforms the\nstate-of-the-art defenses; (ii) it can harden ML-based malware detection\nagainst 27 evasion attacks with detection accuracies greater than 83.45%, at\nthe price of suffering an accuracy decrease smaller than 2.16% in the absence\nof attacks; (iii) it matches or outperforms many anti-malware scanners in\nVirusTotal against realistic adversarial malware.\n", "date": "2023-02-22", "categories": []}
{"id": "2304.00202", "title": "Improving Fast Adversarial Training with Prior-Guided Knowledge", "abstract": "  Fast adversarial training (FAT) is an efficient method to improve robustness.\nHowever, the original FAT suffers from catastrophic overfitting, which\ndramatically and suddenly reduces robustness after a few training epochs.\nAlthough various FAT variants have been proposed to prevent overfitting, they\nrequire high training costs. In this paper, we investigate the relationship\nbetween adversarial example quality and catastrophic overfitting by comparing\nthe training processes of standard adversarial training and FAT. We find that\ncatastrophic overfitting occurs when the attack success rate of adversarial\nexamples becomes worse. Based on this observation, we propose a positive\nprior-guided adversarial initialization to prevent overfitting by improving\nadversarial example quality without extra training costs. This initialization\nis generated by using high-quality adversarial perturbations from the\nhistorical training process. We provide theoretical analysis for the proposed\ninitialization and propose a prior-guided regularization method that boosts the\nsmoothness of the loss function. Additionally, we design a prior-guided\nensemble FAT method that averages the different model weights of historical\nmodels using different decay rates. Our proposed method, called FGSM-PGK,\nassembles the prior-guided knowledge, i.e., the prior-guided initialization and\nmodel weights, acquired during the historical training process. Evaluations of\nfour datasets demonstrate the superiority of the proposed method.\n", "date": "2023-04-01", "categories": []}
{"id": "2304.02693", "title": "A Certified Radius-Guided Attack Framework to Image Segmentation Models", "abstract": "  Image segmentation is an important problem in many safety-critical\napplications. Recent studies show that modern image segmentation models are\nvulnerable to adversarial perturbations, while existing attack methods mainly\nfollow the idea of attacking image classification models. We argue that image\nsegmentation and classification have inherent differences, and design an attack\nframework specially for image segmentation models. Our attack framework is\ninspired by certified radius, which was originally used by defenders to defend\nagainst adversarial perturbations to classification models. We are the first,\nfrom the attacker perspective, to leverage the properties of certified radius\nand propose a certified radius guided attack framework against image\nsegmentation models. Specifically, we first adapt randomized smoothing, the\nstate-of-the-art certification method for classification models, to derive the\npixel's certified radius. We then focus more on disrupting pixels with\nrelatively smaller certified radii and design a pixel-wise certified radius\nguided loss, when plugged into any existing white-box attack, yields our\ncertified radius-guided white-box attack. Next, we propose the first black-box\nattack to image segmentation models via bandit. We design a novel gradient\nestimator, based on bandit feedback, which is query-efficient and provably\nunbiased and stable. We use this gradient estimator to design a projected\nbandit gradient descent (PBGD) attack, as well as a certified radius-guided\nPBGD (CR-PBGD) attack. We prove our PBGD and CR-PBGD attacks can achieve\nasymptotically optimal attack performance with an optimal rate. We evaluate our\ncertified-radius guided white-box and black-box attacks on multiple modern\nimage segmentation models and datasets. Our results validate the effectiveness\nof our certified radius-guided attack framework.\n", "date": "2023-04-05", "categories": []}
{"id": "2304.02786", "title": "UNICORN: A Unified Backdoor Trigger Inversion Framework", "abstract": "  The backdoor attack, where the adversary uses inputs stamped with triggers\n(e.g., a patch) to activate pre-planted malicious behaviors, is a severe threat\nto Deep Neural Network (DNN) models. Trigger inversion is an effective way of\nidentifying backdoor models and understanding embedded adversarial behaviors. A\nchallenge of trigger inversion is that there are many ways of constructing the\ntrigger. Existing methods cannot generalize to various types of triggers by\nmaking certain assumptions or attack-specific constraints. The fundamental\nreason is that existing work does not consider the trigger's design space in\ntheir formulation of the inversion problem. This work formally defines and\nanalyzes the triggers injected in different spaces and the inversion problem.\nThen, it proposes a unified framework to invert backdoor triggers based on the\nformalization of triggers and the identified inner behaviors of backdoor models\nfrom our analysis. Our prototype UNICORN is general and effective in inverting\nbackdoor triggers in DNNs. The code can be found at\nhttps://github.com/RU-System-Software-and-Security/UNICORN.\n", "date": "2023-04-05", "categories": []}
{"id": "2302.12252", "title": "Boosting Adversarial Transferability using Dynamic Cues", "abstract": "  The transferability of adversarial perturbations between image models has\nbeen extensively studied. In this case, an attack is generated from a known\nsurrogate \\eg, the ImageNet trained model, and transferred to change the\ndecision of an unknown (black-box) model trained on an image dataset. However,\nattacks generated from image models do not capture the dynamic nature of a\nmoving object or a changing scene due to a lack of temporal cues within image\nmodels. This leads to reduced transferability of adversarial attacks from\nrepresentation-enriched \\emph{image} models such as Supervised Vision\nTransformers (ViTs), Self-supervised ViTs (\\eg, DINO), and Vision-language\nmodels (\\eg, CLIP) to black-box \\emph{video} models. In this work, we induce\ndynamic cues within the image models without sacrificing their original\nperformance on images. To this end, we optimize \\emph{temporal prompts} through\nfrozen image models to capture motion dynamics. Our temporal prompts are the\nresult of a learnable transformation that allows optimizing for temporal\ngradients during an adversarial attack to fool the motion dynamics.\nSpecifically, we introduce spatial (image) and temporal (video) cues within the\nsame source model through task-specific prompts. Attacking such prompts\nmaximizes the adversarial transferability from image-to-video and\nimage-to-image models using the attacks designed for image models. Our attack\nresults indicate that the attacker does not need specialized architectures,\n\\eg, divided space-time attention, 3D convolutions, or multi-view convolution\nnetworks for different data modalities. Image models are effective surrogates\nto optimize an adversarial attack to fool black-box models in a changing\nenvironment over time. Code is available at https://bit.ly/3Xd9gRQ\n", "date": "2023-02-23", "categories": []}
{"id": "2303.06241", "title": "Do we need entire training data for adversarial training?", "abstract": "  Deep Neural Networks (DNNs) are being used to solve a wide range of problems\nin many domains including safety-critical domains like self-driving cars and\nmedical imagery. DNNs suffer from vulnerability against adversarial attacks. In\nthe past few years, numerous approaches have been proposed to tackle this\nproblem by training networks using adversarial training. Almost all the\napproaches generate adversarial examples for the entire training dataset, thus\nincreasing the training time drastically. We show that we can decrease the\ntraining time for any adversarial training algorithm by using only a subset of\ntraining data for adversarial training. To select the subset, we filter the\nadversarially-prone samples from the training data. We perform a simple\nadversarial attack on all training examples to filter this subset. In this\nattack, we add a small perturbation to each pixel and a few grid lines to the\ninput image.\n  We perform adversarial training on the adversarially-prone subset and mix it\nwith vanilla training performed on the entire dataset. Our results show that\nwhen our method-agnostic approach is plugged into FGSM, we achieve a speedup of\n3.52x on MNIST and 1.98x on the CIFAR-10 dataset with comparable robust\naccuracy. We also test our approach on state-of-the-art Free adversarial\ntraining and achieve a speedup of 1.2x in training time with a marginal drop in\nrobust accuracy on the ImageNet dataset.\n", "date": "2023-03-10", "categories": []}
{"id": "2304.01441", "title": "NetFlick: Adversarial Flickering Attacks on Deep Learning Based Video\n  Compression", "abstract": "  Video compression plays a significant role in IoT devices for the efficient\ntransport of visual data while satisfying all underlying bandwidth constraints.\nDeep learning-based video compression methods are rapidly replacing traditional\nalgorithms and providing state-of-the-art results on edge devices. However,\nrecently developed adversarial attacks demonstrate that digitally crafted\nperturbations can break the Rate-Distortion relationship of video compression.\nIn this work, we present a real-world LED attack to target video compression\nframeworks. Our physically realizable attack, dubbed NetFlick, can degrade the\nspatio-temporal correlation between successive frames by injecting flickering\ntemporal perturbations. In addition, we propose universal perturbations that\ncan downgrade performance of incoming video without prior knowledge of the\ncontents. Experimental results demonstrate that NetFlick can successfully\ndeteriorate the performance of video compression frameworks in both digital-\nand physical-settings and can be further extended to attack downstream video\nclassification networks.\n", "date": "2023-04-04", "categories": []}
{"id": "2206.00924", "title": "FACM: Intermediate Layer Still Retain Effective Features against\n  Adversarial Examples", "abstract": "  In strong adversarial attacks against deep neural networks (DNN), the\ngenerated adversarial example will mislead the DNN-implemented classifier by\ndestroying the output features of the last layer. To enhance the robustness of\nthe classifier, in our paper, a \\textbf{F}eature \\textbf{A}nalysis and\n\\textbf{C}onditional \\textbf{M}atching prediction distribution (FACM) model is\nproposed to utilize the features of intermediate layers to correct the\nclassification. Specifically, we first prove that the intermediate layers of\nthe classifier can still retain effective features for the original category,\nwhich is defined as the correction property in our paper. According to this, we\npropose the FACM model consisting of \\textbf{F}eature \\textbf{A}nalysis (FA)\ncorrection module, \\textbf{C}onditional \\textbf{M}atching \\textbf{P}rediction\n\\textbf{D}istribution (CMPD) correction module and decision module. The FA\ncorrection module is the fully connected layers constructed with the output of\nthe intermediate layers as the input to correct the classification of the\nclassifier. The CMPD correction module is a conditional auto-encoder, which can\nnot only use the output of intermediate layers as the condition to accelerate\nconvergence but also mitigate the negative effect of adversarial example\ntraining with the Kullback-Leibler loss to match prediction distribution.\nThrough the empirically verified diversity property, the correction modules can\nbe implemented synergistically to reduce the adversarial subspace. Hence, the\ndecision module is proposed to integrate the correction modules to enhance the\nDNN classifier's robustness. Specially, our model can be achieved by\nfine-tuning and can be combined with other model-specific defenses.\n", "date": "2022-06-02", "categories": []}
{"id": "2206.07179", "title": "Proximal Splitting Adversarial Attacks for Semantic Segmentation", "abstract": "  Classification has been the focal point of research on adversarial attacks,\nbut only a few works investigate methods suited to denser prediction tasks,\nsuch as semantic segmentation. The methods proposed in these works do not\naccurately solve the adversarial segmentation problem and, therefore,\noverestimate the size of the perturbations required to fool models. Here, we\npropose a white-box attack for these models based on a proximal splitting to\nproduce adversarial perturbations with much smaller $\\ell_\\infty$ norms. Our\nattack can handle large numbers of constraints within a nonconvex minimization\nframework via an Augmented Lagrangian approach, coupled with adaptive\nconstraint scaling and masking strategies. We demonstrate that our attack\nsignificantly outperforms previously proposed ones, as well as classification\nattacks that we adapted for segmentation, providing a first comprehensive\nbenchmark for this dense task.\n", "date": "2022-06-14", "categories": []}
{"id": "2301.13487", "title": "Adversarial Training of Self-supervised Monocular Depth Estimation\n  against Physical-World Attacks", "abstract": "  Monocular Depth Estimation (MDE) is a critical component in applications such\nas autonomous driving. There are various attacks against MDE networks. These\nattacks, especially the physical ones, pose a great threat to the security of\nsuch systems. Traditional adversarial training method requires ground-truth\nlabels hence cannot be directly applied to self-supervised MDE that does not\nhave ground-truth depth. Some self-supervised model hardening techniques (e.g.,\ncontrastive learning) ignore the domain knowledge of MDE and can hardly achieve\noptimal performance. In this work, we propose a novel adversarial training\nmethod for self-supervised MDE models based on view synthesis without using\nground-truth depth. We improve adversarial robustness against physical-world\nattacks using L0-norm-bounded perturbation in training. We compare our method\nwith supervised learning based and contrastive learning based methods that are\ntailored for MDE. Results on two representative MDE networks show that we\nachieve better robustness against various adversarial attacks with nearly no\nbenign performance degradation.\n", "date": "2023-01-31", "categories": []}
{"id": "2303.12848", "title": "Test-time Detection and Repair of Adversarial Samples via Masked\n  Autoencoder", "abstract": "  Training-time defenses, known as adversarial training, incur high training\ncosts and do not generalize to unseen attacks. Test-time defenses solve these\nissues but most existing test-time defenses require adapting the model weights,\ntherefore they do not work on frozen models and complicate model memory\nmanagement. The only test-time defense that does not adapt model weights aims\nto adapt the input with self-supervision tasks. However, we empirically found\nthese self-supervision tasks are not sensitive enough to detect adversarial\nattacks accurately. In this paper, we propose DRAM, a novel defense method to\ndetect and repair adversarial samples at test time via Masked autoencoder\n(MAE). We demonstrate how to use MAE losses to build a Kolmogorov-Smirnov test\nto detect adversarial samples. Moreover, we use the MAE losses to calculate\ninput reversal vectors that repair adversarial samples resulting from\npreviously unseen attacks. Results on large-scale ImageNet dataset show that,\ncompared to all detection baselines evaluated, DRAM achieves the best detection\nrate (82% on average) on all eight adversarial attacks evaluated. For attack\nrepair, DRAM improves the robust accuracy by 6% ~ 41% for standard ResNet50 and\n3% ~ 8% for robust ResNet50 compared with the baselines that use contrastive\nlearning and rotation prediction.\n", "date": "2023-03-22", "categories": []}
{"id": "2303.16378", "title": "A Pilot Study of Query-Free Adversarial Attack against Stable Diffusion", "abstract": "  Despite the record-breaking performance in Text-to-Image (T2I) generation by\nStable Diffusion, less research attention is paid to its adversarial\nrobustness. In this work, we study the problem of adversarial attack generation\nfor Stable Diffusion and ask if an adversarial text prompt can be obtained even\nin the absence of end-to-end model queries. We call the resulting problem\n'query-free attack generation'. To resolve this problem, we show that the\nvulnerability of T2I models is rooted in the lack of robustness of text\nencoders, e.g., the CLIP text encoder used for attacking Stable Diffusion.\nBased on such insight, we propose both untargeted and targeted query-free\nattacks, where the former is built on the most influential dimensions in the\ntext embedding space, which we call steerable key dimensions. By leveraging the\nproposed attacks, we empirically show that only a five-character perturbation\nto the text prompt is able to cause the significant content shift of\nsynthesized images using Stable Diffusion. Moreover, we show that the proposed\ntarget attack can precisely steer the diffusion model to scrub the targeted\nimage content without causing much change in untargeted image content. Our code\nis available at https://github.com/OPTML-Group/QF-Attack.\n", "date": "2023-03-29", "categories": []}
{"id": "2304.00010", "title": "Towards Reasonable Budget Allocation in Untargeted Graph Structure\n  Attacks via Gradient Debias", "abstract": "  It has become cognitive inertia to employ cross-entropy loss function in\nclassification related tasks. In the untargeted attacks on graph structure, the\ngradients derived from the attack objective are the attacker's basis for\nevaluating a perturbation scheme. Previous methods use negative cross-entropy\nloss as the attack objective in attacking node-level classification models.\nHowever, the suitability of the cross-entropy function for constructing the\nuntargeted attack objective has yet been discussed in previous works. This\npaper argues about the previous unreasonable attack objective from the\nperspective of budget allocation. We demonstrate theoretically and empirically\nthat negative cross-entropy tends to produce more significant gradients from\nnodes with lower confidence in the labeled classes, even if the predicted\nclasses of these nodes have been misled. To free up these inefficient attack\nbudgets, we propose a simple attack model for untargeted attacks on graph\nstructure based on a novel attack objective which generates unweighted\ngradients on graph structures that are not affected by the node confidence. By\nconducting experiments in gray-box poisoning attack scenarios, we demonstrate\nthat a reasonable budget allocation can significantly improve the effectiveness\nof gradient-based edge perturbations without any extra hyper-parameter.\n", "date": "2023-03-29", "categories": []}
{"id": "2304.00061", "title": "To be Robust and to be Fair: Aligning Fairness with Robustness", "abstract": "  Adversarial training has been shown to be reliable in improving robustness\nagainst adversarial samples. However, the problem of adversarial training in\nterms of fairness has not yet been properly studied, and the relationship\nbetween fairness and accuracy attack still remains unclear. Can we\nsimultaneously improve robustness w.r.t. both fairness and accuracy? To tackle\nthis topic, in this paper, we study the problem of adversarial training and\nadversarial attack w.r.t. both metrics. We propose a unified structure for\nfairness attack which brings together common notions in group fairness, and we\ntheoretically prove the equivalence of fairness attack against different\nnotions. Moreover, we show the alignment of fairness and accuracy attack, and\ntheoretically demonstrate that robustness w.r.t. one metric benefits from\nrobustness w.r.t. the other metric. Our study suggests a novel way to unify\nadversarial training and attack w.r.t. fairness and accuracy, and experimental\nresults show that our proposed method achieves better performance in terms of\nrobustness w.r.t. both metrics.\n", "date": "2023-03-31", "categories": []}
{"id": "2301.12456", "title": "Towards Verifying the Geometric Robustness of Large-scale Neural\n  Networks", "abstract": "  Deep neural networks (DNNs) are known to be vulnerable to adversarial\ngeometric transformation. This paper aims to verify the robustness of\nlarge-scale DNNs against the combination of multiple geometric transformations\nwith a provable guarantee. Given a set of transformations (e.g., rotation,\nscaling, etc.), we develop GeoRobust, a black-box robustness analyser built\nupon a novel global optimisation strategy, for locating the worst-case\ncombination of transformations that affect and even alter a network's output.\nGeoRobust can provide provable guarantees on finding the worst-case combination\nbased on recent advances in Lipschitzian theory. Due to its black-box nature,\nGeoRobust can be deployed on large-scale DNNs regardless of their\narchitectures, activation functions, and the number of neurons. In practice,\nGeoRobust can locate the worst-case geometric transformation with high\nprecision for the ResNet50 model on ImageNet in a few seconds on average. We\nexamined 18 ImageNet classifiers, including the ResNet family and vision\ntransformers, and found a positive correlation between the geometric robustness\nof the networks and the parameter numbers. We also observe that increasing the\ndepth of DNN is more beneficial than increasing its width in terms of improving\nits geometric robustness. Our tool GeoRobust is available at\nhttps://github.com/TrustAI/GeoRobust.\n", "date": "2023-01-29", "categories": []}
{"id": "2303.17720", "title": "Generating Adversarial Samples in Mini-Batches May Be Detrimental To\n  Adversarial Robustness", "abstract": "  Neural networks have been proven to be both highly effective within computer\nvision, and highly vulnerable to adversarial attacks. Consequently, as the use\nof neural networks increases due to their unrivaled performance, so too does\nthe threat posed by adversarial attacks. In this work, we build towards\naddressing the challenge of adversarial robustness by exploring the\nrelationship between the mini-batch size used during adversarial sample\ngeneration and the strength of the adversarial samples produced. We demonstrate\nthat an increase in mini-batch size results in a decrease in the efficacy of\nthe samples produced, and we draw connections between these observations and\nthe phenomenon of vanishing gradients. Next, we formulate loss functions such\nthat adversarial sample strength is not degraded by mini-batch size. Our\nfindings highlight a potential risk for underestimating the true (practical)\nstrength of adversarial attacks, and a risk of overestimating a model's\nrobustness. We share our codes to let others replicate our experiments and to\nfacilitate further exploration of the connections between batch size and\nadversarial sample strength.\n", "date": "2023-03-30", "categories": []}
{"id": "2303.17764", "title": "Towards Adversarially Robust Continual Learning", "abstract": "  Recent studies show that models trained by continual learning can achieve the\ncomparable performances as the standard supervised learning and the learning\nflexibility of continual learning models enables their wide applications in the\nreal world. Deep learning models, however, are shown to be vulnerable to\nadversarial attacks. Though there are many studies on the model robustness in\nthe context of standard supervised learning, protecting continual learning from\nadversarial attacks has not yet been investigated. To fill in this research\ngap, we are the first to study adversarial robustness in continual learning and\npropose a novel method called \\textbf{T}ask-\\textbf{A}ware \\textbf{B}oundary\n\\textbf{A}ugmentation (TABA) to boost the robustness of continual learning\nmodels. With extensive experiments on CIFAR-10 and CIFAR-100, we show the\nefficacy of adversarial training and TABA in defending adversarial attacks.\n", "date": "2023-03-31", "categories": []}
{"id": "2303.18131", "title": "AdvCheck: Characterizing Adversarial Examples via Local Gradient\n  Checking", "abstract": "  Deep neural networks (DNNs) are vulnerable to adversarial examples, which may\nlead to catastrophe in security-critical domains. Numerous detection methods\nare proposed to characterize the feature uniqueness of adversarial examples, or\nto distinguish DNN's behavior activated by the adversarial examples. Detections\nbased on features cannot handle adversarial examples with large perturbations.\nBesides, they require a large amount of specific adversarial examples. Another\nmainstream, model-based detections, which characterize input properties by\nmodel behaviors, suffer from heavy computation cost. To address the issues, we\nintroduce the concept of local gradient, and reveal that adversarial examples\nhave a quite larger bound of local gradient than the benign ones. Inspired by\nthe observation, we leverage local gradient for detecting adversarial examples,\nand propose a general framework AdvCheck. Specifically, by calculating the\nlocal gradient from a few benign examples and noise-added misclassified\nexamples to train a detector, adversarial examples and even misclassified\nnatural inputs can be precisely distinguished from benign ones. Through\nextensive experiments, we have validated the AdvCheck's superior performance to\nthe state-of-the-art (SOTA) baselines, with detection rate ($\\sim \\times 1.2$)\non general adversarial attacks and ($\\sim \\times 1.4$) on misclassified natural\ninputs on average, with average 1/500 time cost. We also provide interpretable\nresults for successful detection.\n", "date": "2023-03-25", "categories": []}
{"id": "2303.17080", "title": "Mole Recruitment: Poisoning of Image Classifiers via Selective Batch\n  Sampling", "abstract": "  In this work, we present a data poisoning attack that confounds machine\nlearning models without any manipulation of the image or label. This is\nachieved by simply leveraging the most confounding natural samples found within\nthe training data itself, in a new form of a targeted attack coined \"Mole\nRecruitment.\" We define moles as the training samples of a class that appear\nmost similar to samples of another class, and show that simply restructuring\ntraining batches with an optimal number of moles can lead to significant\ndegradation in the performance of the targeted class. We show the efficacy of\nthis novel attack in an offline setting across several standard image\nclassification datasets, and demonstrate the real-world viability of this\nattack in a continual learning (CL) setting. Our analysis reveals that\nstate-of-the-art models are susceptible to Mole Recruitment, thereby exposing a\npreviously undetected vulnerability of image classifiers.\n", "date": "2023-03-30", "categories": []}
{"id": "2208.12815", "title": "What Does the Gradient Tell When Attacking the Graph Structure", "abstract": "  Recent research has revealed that Graph Neural Networks (GNNs) are\nsusceptible to adversarial attacks targeting the graph structure. A malicious\nattacker can manipulate a limited number of edges, given the training labels,\nto impair the victim model's performance. Previous empirical studies indicate\nthat gradient-based attackers tend to add edges rather than remove them. In\nthis paper, we present a theoretical demonstration revealing that attackers\ntend to increase inter-class edges due to the message passing mechanism of\nGNNs, which explains some previous empirical observations. By connecting\ndissimilar nodes, attackers can more effectively corrupt node features, making\nsuch attacks more advantageous. However, we demonstrate that the inherent\nsmoothness of GNN's message passing tends to blur node dissimilarity in the\nfeature space, leading to the loss of crucial information during the forward\nprocess. To address this issue, we propose a novel surrogate model with\nmulti-level propagation that preserves the node dissimilarity information. This\nmodel parallelizes the propagation of unaggregated raw features and multi-hop\naggregated features, while introducing batch normalization to enhance the\ndissimilarity in node representations and counteract the smoothness resulting\nfrom topological aggregation. Our experiments show significant improvement with\nour approach.Furthermore, both theoretical and experimental evidence suggest\nthat adding inter-class edges constitutes an easily observable attack pattern.\nWe propose an innovative attack loss that balances attack effectiveness and\nimperceptibility, sacrificing some attack effectiveness to attain greater\nimperceptibility. We also provide experiments to validate the compromise\nperformance achieved through this attack loss.\n", "date": "2022-08-26", "categories": []}
{"id": "2303.16697", "title": "Latent Feature Relation Consistency for Adversarial Robustness", "abstract": "  Deep neural networks have been applied in many computer vision tasks and\nachieved state-of-the-art performance. However, misclassification will occur\nwhen DNN predicts adversarial examples which add human-imperceptible\nadversarial noise to natural examples. This limits the application of DNN in\nsecurity-critical fields. To alleviate this problem, we first conducted an\nempirical analysis of the latent features of both adversarial and natural\nexamples and found the similarity matrix of natural examples is more compact\nthan those of adversarial examples. Motivated by this observation, we propose\n\\textbf{L}atent \\textbf{F}eature \\textbf{R}elation \\textbf{C}onsistency\n(\\textbf{LFRC}), which constrains the relation of adversarial examples in\nlatent space to be consistent with the natural examples. Importantly, our LFRC\nis orthogonal to the previous method and can be easily combined with them to\nachieve further improvement. To demonstrate the effectiveness of LFRC, we\nconduct extensive experiments using different neural networks on benchmark\ndatasets. For instance, LFRC can bring 0.78\\% further improvement compared to\nAT, and 1.09\\% improvement compared to TRADES, against AutoAttack on CIFAR10.\nCode is available at https://github.com/liuxingbin/LFRC.\n", "date": "2023-03-29", "categories": []}
{"id": "2303.16861", "title": "Beyond Empirical Risk Minimization: Local Structure Preserving\n  Regularization for Improving Adversarial Robustness", "abstract": "  It is broadly known that deep neural networks are susceptible to being fooled\nby adversarial examples with perturbations imperceptible by humans. Various\ndefenses have been proposed to improve adversarial robustness, among which\nadversarial training methods are most effective. However, most of these methods\ntreat the training samples independently and demand a tremendous amount of\nsamples to train a robust network, while ignoring the latent structural\ninformation among these samples. In this work, we propose a novel Local\nStructure Preserving (LSP) regularization, which aims to preserve the local\nstructure of the input space in the learned embedding space. In this manner,\nthe attacking effect of adversarial samples lying in the vicinity of clean\nsamples can be alleviated. We show strong empirical evidence that with or\nwithout adversarial training, our method consistently improves the performance\nof adversarial robustness on several image classification datasets compared to\nthe baselines and some state-of-the-art approaches, thus providing promising\ndirection for future research.\n", "date": "2023-03-29", "categories": []}
{"id": "2303.15571", "title": "EMShepherd: Detecting Adversarial Samples via Side-channel Leakage", "abstract": "  Deep Neural Networks (DNN) are vulnerable to adversarial perturbations-small\nchanges crafted deliberately on the input to mislead the model for wrong\npredictions. Adversarial attacks have disastrous consequences for deep\nlearning-empowered critical applications. Existing defense and detection\ntechniques both require extensive knowledge of the model, testing inputs, and\neven execution details. They are not viable for general deep learning\nimplementations where the model internal is unknown, a common 'black-box'\nscenario for model users. Inspired by the fact that electromagnetic (EM)\nemanations of a model inference are dependent on both operations and data and\nmay contain footprints of different input classes, we propose a framework,\nEMShepherd, to capture EM traces of model execution, perform processing on\ntraces and exploit them for adversarial detection. Only benign samples and\ntheir EM traces are used to train the adversarial detector: a set of EM\nclassifiers and class-specific unsupervised anomaly detectors. When the victim\nmodel system is under attack by an adversarial example, the model execution\nwill be different from executions for the known classes, and the EM trace will\nbe different. We demonstrate that our air-gapped EMShepherd can effectively\ndetect different adversarial attacks on a commonly used FPGA deep learning\naccelerator for both Fashion MNIST and CIFAR-10 datasets. It achieves a 100%\ndetection rate on most types of adversarial samples, which is comparable to the\nstate-of-the-art 'white-box' software-based detectors.\n", "date": "2023-03-27", "categories": []}
{"id": "2303.15735", "title": "Improving the Transferability of Adversarial Samples by Path-Augmented\n  Method", "abstract": "  Deep neural networks have achieved unprecedented success on diverse vision\ntasks. However, they are vulnerable to adversarial noise that is imperceptible\nto humans. This phenomenon negatively affects their deployment in real-world\nscenarios, especially security-related ones. To evaluate the robustness of a\ntarget model in practice, transfer-based attacks craft adversarial samples with\na local model and have attracted increasing attention from researchers due to\ntheir high efficiency. The state-of-the-art transfer-based attacks are\ngenerally based on data augmentation, which typically augments multiple\ntraining images from a linear path when learning adversarial samples. However,\nsuch methods selected the image augmentation path heuristically and may augment\nimages that are semantics-inconsistent with the target images, which harms the\ntransferability of the generated adversarial samples. To overcome the pitfall,\nwe propose the Path-Augmented Method (PAM). Specifically, PAM first constructs\na candidate augmentation path pool. It then settles the employed augmentation\npaths during adversarial sample generation with greedy search. Furthermore, to\navoid augmenting semantics-inconsistent images, we train a Semantics Predictor\n(SP) to constrain the length of the augmentation path. Extensive experiments\nconfirm that PAM can achieve an improvement of over 4.8% on average compared\nwith the state-of-the-art baselines in terms of the attack success rates.\n", "date": "2023-03-28", "categories": []}
{"id": "2303.15818", "title": "Towards Effective Adversarial Textured 3D Meshes on Physical Face\n  Recognition", "abstract": "  Face recognition is a prevailing authentication solution in numerous\nbiometric applications. Physical adversarial attacks, as an important\nsurrogate, can identify the weaknesses of face recognition systems and evaluate\ntheir robustness before deployed. However, most existing physical attacks are\neither detectable readily or ineffective against commercial recognition\nsystems. The goal of this work is to develop a more reliable technique that can\ncarry out an end-to-end evaluation of adversarial robustness for commercial\nsystems. It requires that this technique can simultaneously deceive black-box\nrecognition models and evade defensive mechanisms. To fulfill this, we design\nadversarial textured 3D meshes (AT3D) with an elaborate topology on a human\nface, which can be 3D-printed and pasted on the attacker's face to evade the\ndefenses. However, the mesh-based optimization regime calculates gradients in\nhigh-dimensional mesh space, and can be trapped into local optima with\nunsatisfactory transferability. To deviate from the mesh-based space, we\npropose to perturb the low-dimensional coefficient space based on 3D Morphable\nModel, which significantly improves black-box transferability meanwhile\nenjoying faster search efficiency and better visual quality. Extensive\nexperiments in digital and physical scenarios show that our method effectively\nexplores the security vulnerabilities of multiple popular commercial services,\nincluding three recognition APIs, four anti-spoofing APIs, two prevailing\nmobile phones and two automated access control systems.\n", "date": "2023-03-28", "categories": []}
{"id": "2303.15901", "title": "Denoising Autoencoder-based Defensive Distillation as an Adversarial\n  Robustness Algorithm", "abstract": "  Adversarial attacks significantly threaten the robustness of deep neural\nnetworks (DNNs). Despite the multiple defensive methods employed, they are\nnevertheless vulnerable to poison attacks, where attackers meddle with the\ninitial training data. In order to defend DNNs against such adversarial\nattacks, this work proposes a novel method that combines the defensive\ndistillation mechanism with a denoising autoencoder (DAE). This technique tries\nto lower the sensitivity of the distilled model to poison attacks by spotting\nand reconstructing poisonous adversarial inputs in the training data. We added\ncarefully created adversarial samples to the initial training data to assess\nthe proposed method's performance. Our experimental findings demonstrate that\nour method successfully identified and reconstructed the poisonous inputs while\nalso considering enhancing the DNN's resilience. The proposed approach provides\na potent and robust defense mechanism for DNNs in various applications where\ndata poisoning attacks are a concern. Thus, the defensive distillation\ntechnique's limitation posed by poisonous adversarial attacks is overcome.\n", "date": "2023-03-28", "categories": []}
{"id": "2210.11620", "title": "LOT: Layer-wise Orthogonal Training on Improving $\\ell_2$ Certified\n  Robustness", "abstract": "  Recent studies show that training deep neural networks (DNNs) with Lipschitz\nconstraints are able to enhance adversarial robustness and other model\nproperties such as stability. In this paper, we propose a layer-wise orthogonal\ntraining method (LOT) to effectively train 1-Lipschitz convolution layers via\nparametrizing an orthogonal matrix with an unconstrained matrix. We then\nefficiently compute the inverse square root of a convolution kernel by\ntransforming the input domain to the Fourier frequency domain. On the other\nhand, as existing works show that semi-supervised training helps improve\nempirical robustness, we aim to bridge the gap and prove that semi-supervised\nlearning also improves the certified robustness of Lipschitz-bounded models. We\nconduct comprehensive evaluations for LOT under different settings. We show\nthat LOT significantly outperforms baselines regarding deterministic l2\ncertified robustness, and scales to deeper neural networks. Under the\nsupervised scenario, we improve the state-of-the-art certified robustness for\nall architectures (e.g. from 59.04% to 63.50% on CIFAR-10 and from 32.57% to\n34.59% on CIFAR-100 at radius rho = 36/255 for 40-layer networks). With\nsemi-supervised learning over unlabelled data, we are able to improve\nstate-of-the-art certified robustness on CIFAR-10 at rho = 108/255 from 36.04%\nto 42.39%. In addition, LOT consistently outperforms baselines on different\nmodel architectures with only 1/3 evaluation time.\n", "date": "2022-10-20", "categories": []}
{"id": "2301.00896", "title": "Efficient Robustness Assessment via Adversarial Spatial-Temporal Focus\n  on Videos", "abstract": "  Adversarial robustness assessment for video recognition models has raised\nconcerns owing to their wide applications on safety-critical tasks. Compared\nwith images, videos have much high dimension, which brings huge computational\ncosts when generating adversarial videos. This is especially serious for the\nquery-based black-box attacks where gradient estimation for the threat models\nis usually utilized, and high dimensions will lead to a large number of\nqueries. To mitigate this issue, we propose to simultaneously eliminate the\ntemporal and spatial redundancy within the video to achieve an effective and\nefficient gradient estimation on the reduced searching space, and thus query\nnumber could decrease. To implement this idea, we design the novel Adversarial\nspatial-temporal Focus (AstFocus) attack on videos, which performs attacks on\nthe simultaneously focused key frames and key regions from the inter-frames and\nintra-frames in the video. AstFocus attack is based on the cooperative\nMulti-Agent Reinforcement Learning (MARL) framework. One agent is responsible\nfor selecting key frames, and another agent is responsible for selecting key\nregions. These two agents are jointly trained by the common rewards received\nfrom the black-box threat models to perform a cooperative prediction. By\ncontinuously querying, the reduced searching space composed of key frames and\nkey regions is becoming precise, and the whole query number becomes less than\nthat on the original video. Extensive experiments on four mainstream video\nrecognition models and three widely used action recognition datasets\ndemonstrate that the proposed AstFocus attack outperforms the SOTA methods,\nwhich is prevenient in fooling rate, query number, time, and perturbation\nmagnitude at the same.\n", "date": "2023-01-03", "categories": []}
{"id": "2303.14077", "title": "Improved Adversarial Training Through Adaptive Instance-wise Loss\n  Smoothing", "abstract": "  Deep neural networks can be easily fooled into making incorrect predictions\nthrough corruption of the input by adversarial perturbations:\nhuman-imperceptible artificial noise. So far adversarial training has been the\nmost successful defense against such adversarial attacks. This work focuses on\nimproving adversarial training to boost adversarial robustness. We first\nanalyze, from an instance-wise perspective, how adversarial vulnerability\nevolves during adversarial training. We find that during training an overall\nreduction of adversarial loss is achieved by sacrificing a considerable\nproportion of training samples to be more vulnerable to adversarial attack,\nwhich results in an uneven distribution of adversarial vulnerability among\ndata. Such \"uneven vulnerability\", is prevalent across several popular robust\ntraining methods and, more importantly, relates to overfitting in adversarial\ntraining. Motivated by this observation, we propose a new adversarial training\nmethod: Instance-adaptive Smoothness Enhanced Adversarial Training (ISEAT). It\njointly smooths both input and weight loss landscapes in an adaptive,\ninstance-specific, way to enhance robustness more for those samples with higher\nadversarial vulnerability. Extensive experiments demonstrate the superiority of\nour method over existing defense methods. Noticeably, our method, when combined\nwith the latest data augmentation and semi-supervised learning techniques,\nachieves state-of-the-art robustness against $\\ell_{\\infty}$-norm constrained\nattacks on CIFAR10 of 59.32% for Wide ResNet34-10 without extra data, and\n61.55% for Wide ResNet28-10 with extra data. Code is available at\nhttps://github.com/TreeLLi/Instance-adaptive-Smoothness-Enhanced-AT.\n", "date": "2023-03-24", "categories": []}
{"id": "2303.14304", "title": "Ensemble-based Blackbox Attacks on Dense Prediction", "abstract": "  We propose an approach for adversarial attacks on dense prediction models\n(such as object detectors and segmentation). It is well known that the attacks\ngenerated by a single surrogate model do not transfer to arbitrary (blackbox)\nvictim models. Furthermore, targeted attacks are often more challenging than\nthe untargeted attacks. In this paper, we show that a carefully designed\nensemble can create effective attacks for a number of victim models. In\nparticular, we show that normalization of the weights for individual models\nplays a critical role in the success of the attacks. We then demonstrate that\nby adjusting the weights of the ensemble according to the victim model can\nfurther improve the performance of the attacks. We performed a number of\nexperiments for object detectors and segmentation to highlight the significance\nof the our proposed methods. Our proposed ensemble-based method outperforms\nexisting blackbox attack methods for object detection and segmentation. Finally\nwe show that our proposed method can also generate a single perturbation that\ncan fool multiple blackbox detection and segmentation models simultaneously.\nCode is available at https://github.com/CSIPlab/EBAD.\n", "date": "2023-03-25", "categories": []}
{"id": "2303.14443", "title": "No more Reviewer #2: Subverting Automatic Paper-Reviewer Assignment\n  using Adversarial Learning", "abstract": "  The number of papers submitted to academic conferences is steadily rising in\nmany scientific disciplines. To handle this growth, systems for automatic\npaper-reviewer assignments are increasingly used during the reviewing process.\nThese systems use statistical topic models to characterize the content of\nsubmissions and automate the assignment to reviewers. In this paper, we show\nthat this automation can be manipulated using adversarial learning. We propose\nan attack that adapts a given paper so that it misleads the assignment and\nselects its own reviewers. Our attack is based on a novel optimization strategy\nthat alternates between the feature space and problem space to realize\nunobtrusive changes to the paper. To evaluate the feasibility of our attack, we\nsimulate the paper-reviewer assignment of an actual security conference (IEEE\nS&P) with 165 reviewers on the program committee. Our results show that we can\nsuccessfully select and remove reviewers without access to the assignment\nsystem. Moreover, we demonstrate that the manipulated papers remain plausible\nand are often indistinguishable from benign submissions.\n", "date": "2023-03-25", "categories": []}
{"id": "2303.14460", "title": "CFA: Class-wise Calibrated Fair Adversarial Training", "abstract": "  Adversarial training has been widely acknowledged as the most effective\nmethod to improve the adversarial robustness against adversarial examples for\nDeep Neural Networks (DNNs). So far, most existing works focus on enhancing the\noverall model robustness, treating each class equally in both the training and\ntesting phases. Although revealing the disparity in robustness among classes,\nfew works try to make adversarial training fair at the class level without\nsacrificing overall robustness. In this paper, we are the first to\ntheoretically and empirically investigate the preference of different classes\nfor adversarial configurations, including perturbation margin, regularization,\nand weight averaging. Motivated by this, we further propose a\n\\textbf{C}lass-wise calibrated \\textbf{F}air \\textbf{A}dversarial training\nframework, named CFA, which customizes specific training configurations for\neach class automatically. Experiments on benchmark datasets demonstrate that\nour proposed CFA can improve both overall robustness and fairness notably over\nother state-of-the-art methods. Code is available at\n\\url{https://github.com/PKU-ML/CFA}.\n", "date": "2023-03-25", "categories": []}
{"id": "2303.14922", "title": "CAT:Collaborative Adversarial Training", "abstract": "  Adversarial training can improve the robustness of neural networks. Previous\nmethods focus on a single adversarial training strategy and do not consider the\nmodel property trained by different strategies. By revisiting the previous\nmethods, we find different adversarial training methods have distinct\nrobustness for sample instances. For example, a sample instance can be\ncorrectly classified by a model trained using standard adversarial training\n(AT) but not by a model trained using TRADES, and vice versa. Based on this\nobservation, we propose a collaborative adversarial training framework to\nimprove the robustness of neural networks. Specifically, we use different\nadversarial training methods to train robust models and let models interact\nwith their knowledge during the training process. Collaborative Adversarial\nTraining (CAT) can improve both robustness and accuracy. Extensive experiments\non various networks and datasets validate the effectiveness of our method. CAT\nachieves state-of-the-art adversarial robustness without using any additional\ndata on CIFAR-10 under the Auto-Attack benchmark. Code is available at\nhttps://github.com/liuxingbin/CAT.\n", "date": "2023-03-27", "categories": []}
{"id": "2303.15127", "title": "Learning the Unlearnable: Adversarial Augmentations Suppress Unlearnable\n  Example Attacks", "abstract": "  Unlearnable example attacks are data poisoning techniques that can be used to\nsafeguard public data against unauthorized use for training deep learning\nmodels. These methods add stealthy perturbations to the original image, thereby\nmaking it difficult for deep learning models to learn from these training data\neffectively. Current research suggests that adversarial training can, to a\ncertain degree, mitigate the impact of unlearnable example attacks, while\ncommon data augmentation methods are not effective against such poisons.\nAdversarial training, however, demands considerable computational resources and\ncan result in non-trivial accuracy loss. In this paper, we introduce the\nUEraser method, which outperforms current defenses against different types of\nstate-of-the-art unlearnable example attacks through a combination of effective\ndata augmentation policies and loss-maximizing adversarial augmentations. In\nstark contrast to the current SOTA adversarial training methods, UEraser uses\nadversarial augmentations, which extends beyond the confines of $ \\ell_p $\nperturbation budget assumed by current unlearning attacks and defenses. It also\nhelps to improve the model's generalization ability, thus protecting against\naccuracy loss. UEraser wipes out the unlearning effect with error-maximizing\ndata augmentations, thus restoring trained model accuracies. Interestingly,\nUEraser-Lite, a fast variant without adversarial augmentations, is also highly\neffective in preserving clean accuracies. On challenging unlearnable CIFAR-10,\nCIFAR-100, SVHN, and ImageNet-subset datasets produced with various attacks, it\nachieves results that are comparable to those obtained during clean training.\nWe also demonstrate its efficacy against possible adaptive attacks. Our code is\nopen source and available to the deep learning community:\nhttps://github.com/lafeat/ueraser.\n", "date": "2023-03-27", "categories": []}
{"id": "2302.11963", "title": "Investigating Catastrophic Overfitting in Fast Adversarial Training: A\n  Self-fitting Perspective", "abstract": "  Although fast adversarial training provides an efficient approach for\nbuilding robust networks, it may suffer from a serious problem known as\ncatastrophic overfitting (CO), where multi-step robust accuracy suddenly\ncollapses to zero. In this paper, we for the first time decouple single-step\nadversarial examples into data-information and self-information, which reveals\nan interesting phenomenon called \"self-fitting\". Self-fitting, i.e., the\nnetwork learns the self-information embedded in single-step perturbations,\nnaturally leads to the occurrence of CO. When self-fitting occurs, the network\nexperiences an obvious \"channel differentiation\" phenomenon that some\nconvolution channels accounting for recognizing self-information become\ndominant, while others for data-information are suppressed. In this way, the\nnetwork can only recognize images with sufficient self-information and loses\ngeneralization ability to other types of data. Based on self-fitting, we\nprovide new insights into the existing methods to mitigate CO and extend CO to\nmulti-step adversarial training. Our findings reveal a self-learning mechanism\nin adversarial training and open up new perspectives for suppressing different\nkinds of information to mitigate CO.\n", "date": "2023-02-23", "categories": []}
{"id": "2303.02814", "title": "Visual Analytics of Neuron Vulnerability to Adversarial Attacks on\n  Convolutional Neural Networks", "abstract": "  Adversarial attacks on a convolutional neural network (CNN) -- injecting\nhuman-imperceptible perturbations into an input image -- could fool a\nhigh-performance CNN into making incorrect predictions. The success of\nadversarial attacks raises serious concerns about the robustness of CNNs, and\nprevents them from being used in safety-critical applications, such as medical\ndiagnosis and autonomous driving. Our work introduces a visual analytics\napproach to understanding adversarial attacks by answering two questions: (1)\nwhich neurons are more vulnerable to attacks and (2) which image features do\nthese vulnerable neurons capture during the prediction? For the first question,\nwe introduce multiple perturbation-based measures to break down the attacking\nmagnitude into individual CNN neurons and rank the neurons by their\nvulnerability levels. For the second, we identify image features (e.g., cat\nears) that highly stimulate a user-selected neuron to augment and validate the\nneuron's responsibility. Furthermore, we support an interactive exploration of\na large number of neurons by aiding with hierarchical clustering based on the\nneurons' roles in the prediction. To this end, a visual analytics system is\ndesigned to incorporate visual reasoning for interpreting adversarial attacks.\nWe validate the effectiveness of our system through multiple case studies as\nwell as feedback from domain experts.\n", "date": "2023-03-06", "categories": []}
{"id": "2303.13813", "title": "Generalist: Decoupling Natural and Robust Generalization", "abstract": "  Deep neural networks obtained by standard training have been constantly\nplagued by adversarial examples. Although adversarial training demonstrates its\ncapability to defend against adversarial examples, unfortunately, it leads to\nan inevitable drop in the natural generalization. To address the issue, we\ndecouple the natural generalization and the robust generalization from joint\ntraining and formulate different training strategies for each one.\nSpecifically, instead of minimizing a global loss on the expectation over these\ntwo generalization errors, we propose a bi-expert framework called\n\\emph{Generalist} where we simultaneously train base learners with task-aware\nstrategies so that they can specialize in their own fields. The parameters of\nbase learners are collected and combined to form a global learner at intervals\nduring the training process. The global learner is then distributed to the base\nlearners as initialized parameters for continued training. Theoretically, we\nprove that the risks of Generalist will get lower once the base learners are\nwell trained. Extensive experiments verify the applicability of Generalist to\nachieve high accuracy on natural examples while maintaining considerable\nrobustness to adversarial ones. Code is available at\nhttps://github.com/PKU-ML/Generalist.\n", "date": "2023-03-24", "categories": []}
{"id": "2303.13846", "title": "Feature Separation and Recalibration for Adversarial Robustness", "abstract": "  Deep neural networks are susceptible to adversarial attacks due to the\naccumulation of perturbations in the feature level, and numerous works have\nboosted model robustness by deactivating the non-robust feature activations\nthat cause model mispredictions. However, we claim that these malicious\nactivations still contain discriminative cues and that with recalibration, they\ncan capture additional useful information for correct model predictions. To\nthis end, we propose a novel, easy-to-plugin approach named Feature Separation\nand Recalibration (FSR) that recalibrates the malicious, non-robust activations\nfor more robust feature maps through Separation and Recalibration. The\nSeparation part disentangles the input feature map into the robust feature with\nactivations that help the model make correct predictions and the non-robust\nfeature with activations that are responsible for model mispredictions upon\nadversarial attack. The Recalibration part then adjusts the non-robust\nactivations to restore the potentially useful cues for model predictions.\nExtensive experiments verify the superiority of FSR compared to traditional\ndeactivation techniques and demonstrate that it improves the robustness of\nexisting adversarial training methods by up to 8.57% with small computational\noverhead. Codes are available at https://github.com/wkim97/FSR.\n", "date": "2023-03-24", "categories": []}
{"id": "2303.13868", "title": "Physically Adversarial Infrared Patches with Learnable Shapes and\n  Locations", "abstract": "  Owing to the extensive application of infrared object detectors in the\nsafety-critical tasks, it is necessary to evaluate their robustness against\nadversarial examples in the real world. However, current few physical infrared\nattacks are complicated to implement in practical application because of their\ncomplex transformation from digital world to physical world. To address this\nissue, in this paper, we propose a physically feasible infrared attack method\ncalled \"adversarial infrared patches\". Considering the imaging mechanism of\ninfrared cameras by capturing objects' thermal radiation, adversarial infrared\npatches conduct attacks by attaching a patch of thermal insulation materials on\nthe target object to manipulate its thermal distribution. To enhance\nadversarial attacks, we present a novel aggregation regularization to guide the\nsimultaneous learning for the patch' shape and location on the target object.\nThus, a simple gradient-based optimization can be adapted to solve for them. We\nverify adversarial infrared patches in different object detection tasks with\nvarious object detectors. Experimental results show that our method achieves\nmore than 90\\% Attack Success Rate (ASR) versus the pedestrian detector and\nvehicle detector in the physical environment, where the objects are captured in\ndifferent angles, distances, postures, and scenes. More importantly,\nadversarial infrared patch is easy to implement, and it only needs 0.5 hours to\nbe constructed in the physical world, which verifies its effectiveness and\nefficiency.\n", "date": "2023-03-24", "categories": []}
{"id": "2303.13887", "title": "Effective black box adversarial attack with handcrafted kernels", "abstract": "  We propose a new, simple framework for crafting adversarial examples for\nblack box attacks. The idea is to simulate the substitution model with a\nnon-trainable model compounded of just one layer of handcrafted convolutional\nkernels and then train the generator neural network to maximize the distance of\nthe outputs for the original and generated adversarial image. We show that\nfooling the prediction of the first layer causes the whole network to be fooled\nand decreases its accuracy on adversarial inputs. Moreover, we do not train the\nneural network to obtain the first convolutional layer kernels, but we create\nthem using the technique of F-transform. Therefore, our method is very time and\nresource effective.\n", "date": "2023-03-24", "categories": []}
{"id": "2303.13955", "title": "PIAT: Parameter Interpolation based Adversarial Training for Image\n  Classification", "abstract": "  Adversarial training has been demonstrated to be the most effective approach\nto defend against adversarial attacks. However, existing adversarial training\nmethods show apparent oscillations and overfitting issue in the training\nprocess, degrading the defense efficacy. In this work, we propose a novel\nframework, termed Parameter Interpolation based Adversarial Training (PIAT),\nthat makes full use of the historical information during training.\nSpecifically, at the end of each epoch, PIAT tunes the model parameters as the\ninterpolation of the parameters of the previous and current epochs. Besides, we\nsuggest to use the Normalized Mean Square Error (NMSE) to further improve the\nrobustness by aligning the clean and adversarial examples. Compared with other\nregularization methods, NMSE focuses more on the relative magnitude of the\nlogits rather than the absolute magnitude. Extensive experiments on several\nbenchmark datasets and various networks show that our method could prominently\nimprove the model robustness and reduce the generalization error. Moreover, our\nframework is general and could further boost the robust accuracy when combined\nwith other adversarial training methods.\n", "date": "2023-03-24", "categories": []}
{"id": "2301.01217", "title": "Unlearnable Clusters: Towards Label-agnostic Unlearnable Examples", "abstract": "  There is a growing interest in developing unlearnable examples (UEs) against\nvisual privacy leaks on the Internet. UEs are training samples added with\ninvisible but unlearnable noise, which have been found can prevent unauthorized\ntraining of machine learning models. UEs typically are generated via a bilevel\noptimization framework with a surrogate model to remove (minimize) errors from\nthe original samples, and then applied to protect the data against unknown\ntarget models. However, existing UE generation methods all rely on an ideal\nassumption called label-consistency, where the hackers and protectors are\nassumed to hold the same label for a given sample. In this work, we propose and\npromote a more practical label-agnostic setting, where the hackers may exploit\nthe protected data quite differently from the protectors. E.g., a m-class\nunlearnable dataset held by the protector may be exploited by the hacker as a\nn-class dataset. Existing UE generation methods are rendered ineffective in\nthis challenging setting. To tackle this challenge, we present a novel\ntechnique called Unlearnable Clusters (UCs) to generate label-agnostic\nunlearnable examples with cluster-wise perturbations. Furthermore, we propose\nto leverage VisionandLanguage Pre-trained Models (VLPMs) like CLIP as the\nsurrogate model to improve the transferability of the crafted UCs to diverse\ndomains. We empirically verify the effectiveness of our proposed approach under\na variety of settings with different datasets, target models, and even\ncommercial platforms Microsoft Azure and Baidu PaddlePaddle. Code is available\nat \\url{https://github.com/jiamingzhang94/Unlearnable-Clusters}.\n", "date": "2022-12-31", "categories": []}
{"id": "2303.13010", "title": "Semantic Image Attack for Visual Model Diagnosis", "abstract": "  In practice, metric analysis on a specific train and test dataset does not\nguarantee reliable or fair ML models. This is partially due to the fact that\nobtaining a balanced, diverse, and perfectly labeled dataset is typically\nexpensive, time-consuming, and error-prone. Rather than relying on a carefully\ndesigned test set to assess ML models' failures, fairness, or robustness, this\npaper proposes Semantic Image Attack (SIA), a method based on the adversarial\nattack that provides semantic adversarial images to allow model diagnosis,\ninterpretability, and robustness. Traditional adversarial training is a popular\nmethodology for robustifying ML models against attacks. However, existing\nadversarial methods do not combine the two aspects that enable the\ninterpretation and analysis of the model's flaws: semantic traceability and\nperceptual quality. SIA combines the two features via iterative gradient ascent\non a predefined semantic attribute space and the image space. We illustrate the\nvalidity of our approach in three scenarios for keypoint detection and\nclassification. (1) Model diagnosis: SIA generates a histogram of attributes\nthat highlights the semantic vulnerability of the ML model (i.e., attributes\nthat make the model fail). (2) Stronger attacks: SIA generates adversarial\nexamples with visually interpretable attributes that lead to higher attack\nsuccess rates than baseline methods. The adversarial training on SIA improves\nthe transferable robustness across different gradient-based attacks. (3)\nRobustness to imbalanced datasets: we use SIA to augment the underrepresented\nclasses, which outperforms strong augmentation and re-balancing baselines.\n", "date": "2023-03-23", "categories": []}
{"id": "2303.13401", "title": "Optimization and Optimizers for Adversarial Robustness", "abstract": "  Empirical robustness evaluation (RE) of deep learning models against\nadversarial perturbations entails solving nontrivial constrained optimization\nproblems. Existing numerical algorithms that are commonly used to solve them in\npractice predominantly rely on projected gradient, and mostly handle\nperturbations modeled by the $\\ell_1$, $\\ell_2$ and $\\ell_\\infty$ distances. In\nthis paper, we introduce a novel algorithmic framework that blends a\ngeneral-purpose constrained-optimization solver PyGRANSO with Constraint\nFolding (PWCF), which can add more reliability and generality to the\nstate-of-the-art RE packages, e.g., AutoAttack. Regarding reliability, PWCF\nprovides solutions with stationarity measures and feasibility tests to assess\nthe solution quality. For generality, PWCF can handle perturbation models that\nare typically inaccessible to the existing projected gradient methods; the main\nrequirement is the distance metric to be almost everywhere differentiable.\nTaking advantage of PWCF and other existing numerical algorithms, we further\nexplore the distinct patterns in the solutions found for solving these\noptimization problems using various combinations of losses, perturbation\nmodels, and optimization algorithms. We then discuss the implications of these\npatterns on the current robustness evaluation and adversarial training.\n", "date": "2023-03-23", "categories": []}
{"id": "2303.12357", "title": "Wasserstein Adversarial Examples on Univariant Time Series Data", "abstract": "  Adversarial examples are crafted by adding indistinguishable perturbations to\nnormal examples in order to fool a well-trained deep learning model to\nmisclassify. In the context of computer vision, this notion of\nindistinguishability is typically bounded by $L_{\\infty}$ or other norms.\nHowever, these norms are not appropriate for measuring indistinguishiability\nfor time series data. In this work, we propose adversarial examples in the\nWasserstein space for time series data for the first time and utilize\nWasserstein distance to bound the perturbation between normal examples and\nadversarial examples. We introduce Wasserstein projected gradient descent\n(WPGD), an adversarial attack method for perturbing univariant time series\ndata. We leverage the closed-form solution of Wasserstein distance in the 1D\nspace to calculate the projection step of WPGD efficiently with the gradient\ndescent method. We further propose a two-step projection so that the search of\nadversarial examples in the Wasserstein space is guided and constrained by\nEuclidean norms to yield more effective and imperceptible perturbations. We\nempirically evaluate the proposed attack on several time series datasets in the\nhealthcare domain. Extensive results demonstrate that the Wasserstein attack is\npowerful and can successfully attack most of the target classifiers with a high\nattack success rate. To better study the nature of Wasserstein adversarial\nexample, we evaluate a strong defense mechanism named Wasserstein smoothing for\npotential certified robustness defense. Although the defense can achieve some\naccuracy gain, it still has limitations in many cases and leaves space for\ndeveloping a stronger certified robustness method to Wasserstein adversarial\nexamples on univariant time series data.\n", "date": "2023-03-22", "categories": []}
{"id": "2303.12363", "title": "Distribution-restrained Softmax Loss for the Model Robustness", "abstract": "  Recently, the robustness of deep learning models has received widespread\nattention, and various methods for improving model robustness have been\nproposed, including adversarial training, model architecture modification,\ndesign of loss functions, certified defenses, and so on. However, the principle\nof the robustness to attacks is still not fully understood, also the related\nresearch is still not sufficient. Here, we have identified a significant factor\nthat affects the robustness of models: the distribution characteristics of\nsoftmax values for non-real label samples. We found that the results after an\nattack are highly correlated with the distribution characteristics, and thus we\nproposed a loss function to suppress the distribution diversity of softmax. A\nlarge number of experiments have shown that our method can improve robustness\nwithout significant time consumption.\n", "date": "2023-03-22", "categories": []}
{"id": "2303.12512", "title": "Sibling-Attack: Rethinking Transferable Adversarial Attacks against Face\n  Recognition", "abstract": "  A hard challenge in developing practical face recognition (FR) attacks is due\nto the black-box nature of the target FR model, i.e., inaccessible gradient and\nparameter information to attackers. While recent research took an important\nstep towards attacking black-box FR models through leveraging transferability,\ntheir performance is still limited, especially against online commercial FR\nsystems that can be pessimistic (e.g., a less than 50% ASR--attack success rate\non average). Motivated by this, we present Sibling-Attack, a new FR attack\ntechnique for the first time explores a novel multi-task perspective (i.e.,\nleveraging extra information from multi-correlated tasks to boost attacking\ntransferability). Intuitively, Sibling-Attack selects a set of tasks correlated\nwith FR and picks the Attribute Recognition (AR) task as the task used in\nSibling-Attack based on theoretical and quantitative analysis. Sibling-Attack\nthen develops an optimization framework that fuses adversarial gradient\ninformation through (1) constraining the cross-task features to be under the\nsame space, (2) a joint-task meta optimization framework that enhances the\ngradient compatibility among tasks, and (3) a cross-task gradient stabilization\nmethod which mitigates the oscillation effect during attacking. Extensive\nexperiments demonstrate that Sibling-Attack outperforms state-of-the-art FR\nattack techniques by a non-trivial margin, boosting ASR by 12.61% and 55.77% on\naverage on state-of-the-art pre-trained FR models and two well-known, widely\nused commercial FR systems.\n", "date": "2023-03-22", "categories": []}
{"id": "2303.12658", "title": "Reliable and Efficient Evaluation of Adversarial Robustness for Deep\n  Hashing-Based Retrieval", "abstract": "  Deep hashing has been extensively applied to massive image retrieval due to\nits efficiency and effectiveness. Recently, several adversarial attacks have\nbeen presented to reveal the vulnerability of deep hashing models against\nadversarial examples. However, existing attack methods suffer from degraded\nperformance or inefficiency because they underutilize the semantic relations\nbetween original samples or spend a lot of time learning these relations with a\ndeep neural network. In this paper, we propose a novel Pharos-guided Attack,\ndubbed PgA, to evaluate the adversarial robustness of deep hashing networks\nreliably and efficiently. Specifically, we design pharos code to represent the\nsemantics of the benign image, which preserves the similarity to semantically\nrelevant samples and dissimilarity to irrelevant ones. It is proven that we can\nquickly calculate the pharos code via a simple math formula. Accordingly, PgA\ncan directly conduct a reliable and efficient attack on deep hashing-based\nretrieval by maximizing the similarity between the hash code of the adversarial\nexample and the pharos code. Extensive experiments on the benchmark datasets\nverify that the proposed algorithm outperforms the prior state-of-the-arts in\nboth attack strength and speed.\n", "date": "2023-03-22", "categories": []}
{"id": "2303.12669", "title": "An Extended Study of Human-like Behavior under Adversarial Training", "abstract": "  Neural networks have a number of shortcomings. Amongst the severest ones is\nthe sensitivity to distribution shifts which allows models to be easily fooled\ninto wrong predictions by small perturbations to inputs that are often\nimperceivable to humans and do not have to carry semantic meaning. Adversarial\ntraining poses a partial solution to address this issue by training models on\nworst-case perturbations. Yet, recent work has also pointed out that the\nreasoning in neural networks is different from humans. Humans identify objects\nby shape, while neural nets mainly employ texture cues. Exemplarily, a model\ntrained on photographs will likely fail to generalize to datasets containing\nsketches. Interestingly, it was also shown that adversarial training seems to\nfavorably increase the shift toward shape bias. In this work, we revisit this\nobservation and provide an extensive analysis of this effect on various\narchitectures, the common $\\ell_2$- and $\\ell_\\infty$-training, and\nTransformer-based models. Further, we provide a possible explanation for this\nphenomenon from a frequency perspective.\n", "date": "2023-03-22", "categories": []}
{"id": "2303.11625", "title": "Information-containing Adversarial Perturbation for Combating Facial\n  Manipulation Systems", "abstract": "  With the development of deep learning technology, the facial manipulation\nsystem has become powerful and easy to use. Such systems can modify the\nattributes of the given facial images, such as hair color, gender, and age.\nMalicious applications of such systems pose a serious threat to individuals'\nprivacy and reputation. Existing studies have proposed various approaches to\nprotect images against facial manipulations. Passive defense methods aim to\ndetect whether the face is real or fake, which works for posterior forensics\nbut can not prevent malicious manipulation. Initiative defense methods protect\nimages upfront by injecting adversarial perturbations into images to disrupt\nfacial manipulation systems but can not identify whether the image is fake. To\naddress the limitation of existing methods, we propose a novel two-tier\nprotection method named Information-containing Adversarial Perturbation (IAP),\nwhich provides more comprehensive protection for {facial images}. We use an\nencoder to map a facial image and its identity message to a cross-model\nadversarial example which can disrupt multiple facial manipulation systems to\nachieve initiative protection. Recovering the message in adversarial examples\nwith a decoder serves passive protection, contributing to provenance tracking\nand fake image detection. We introduce a feature-level correlation measurement\nthat is more suitable to measure the difference between the facial images than\nthe commonly used mean squared error. Moreover, we propose a spectral diffusion\nmethod to spread messages to different frequency channels, thereby improving\nthe robustness of the message against facial manipulation. Extensive\nexperimental results demonstrate that our proposed IAP can recover the messages\nfrom the adversarial examples with high average accuracy and effectively\ndisrupt the facial manipulation systems.\n", "date": "2023-03-21", "categories": []}
{"id": "2303.10288", "title": "Mobile Edge Adversarial Detection for Digital Twinning to the Metaverse\n  with Deep Reinforcement Learning", "abstract": "  Real-time Digital Twinning of physical world scenes onto the Metaverse is\nnecessary for a myriad of applications such as augmented-reality (AR) assisted\ndriving. In AR assisted driving, physical environment scenes are first captured\nby Internet of Vehicles (IoVs) and are uploaded to the Metaverse. A central\nMetaverse Map Service Provider (MMSP) will aggregate information from all IoVs\nto develop a central Metaverse Map. Information from the Metaverse Map can then\nbe downloaded into individual IoVs on demand and be delivered as AR scenes to\nthe driver. However, the growing interest in developing AR assisted driving\napplications which relies on digital twinning invites adversaries. These\nadversaries may place physical adversarial patches on physical world objects\nsuch as cars, signboards, or on roads, seeking to contort the virtual world\ndigital twin. Hence, there is a need to detect these physical world adversarial\npatches. Nevertheless, as real-time, accurate detection of adversarial patches\nis compute-intensive, these physical world scenes have to be offloaded to the\nMetaverse Map Base Stations (MMBS) for computation. Hence in our work, we\nconsidered an environment with moving Internet of Vehicles (IoV), uploading\nreal-time physical world scenes to the MMBSs. We formulated a realistic joint\nvariable optimization problem where the MMSPs' objective is to maximize\nadversarial patch detection mean average precision (mAP), while minimizing the\ncomputed AR scene up-link transmission latency and IoVs' up-link transmission\nidle count, through optimizing the IoV-MMBS allocation and IoV up-link scene\nresolution selection. We proposed a Heterogeneous Action Proximal Policy\nOptimization (HAPPO) (discrete-continuous) algorithm to tackle the proposed\nproblem. Extensive experiments shows HAPPO outperforms baseline models when\ncompared against key metrics.\n", "date": "2023-03-18", "categories": []}
{"id": "2303.10291", "title": "Detection of Uncertainty in Exceedance of Threshold (DUET): An\n  Adversarial Patch Localizer", "abstract": "  Development of defenses against physical world attacks such as adversarial\npatches is gaining traction within the research community. We contribute to the\nfield of adversarial patch detection by introducing an uncertainty-based\nadversarial patch localizer which localizes adversarial patch on an image,\npermitting post-processing patch-avoidance or patch-reconstruction. We quantify\nour prediction uncertainties with the development of \\textit{\\textbf{D}etection\nof \\textbf{U}ncertainties in the \\textbf{E}xceedance of \\textbf{T}hreshold}\n(DUET) algorithm. This algorithm provides a framework to ascertain confidence\nin the adversarial patch localization, which is essential for safety-sensitive\napplications such as self-driving cars and medical imaging. We conducted\nexperiments on localizing adversarial patches and found our proposed DUET model\noutperforms baseline models. We then conduct further analyses on our choice of\nmodel priors and the adoption of Bayesian Neural Networks in different layers\nwithin our model architecture. We found that isometric gaussian priors in\nBayesian Neural Networks are suitable for patch localization tasks and the\npresence of Bayesian layers in the earlier neural network blocks facilitates\ntop-end localization performance, while Bayesian layers added in the later\nneural network blocks contribute to better model generalization. We then\npropose two different well-performing models to tackle different use cases.\n", "date": "2023-03-18", "categories": []}
{"id": "2303.10653", "title": "Randomized Adversarial Training via Taylor Expansion", "abstract": "  In recent years, there has been an explosion of research into developing more\nrobust deep neural networks against adversarial examples. Adversarial training\nappears as one of the most successful methods. To deal with both the robustness\nagainst adversarial examples and the accuracy over clean examples, many works\ndevelop enhanced adversarial training methods to achieve various trade-offs\nbetween them. Leveraging over the studies that smoothed update on weights\nduring training may help find flat minima and improve generalization, we\nsuggest reconciling the robustness-accuracy trade-off from another perspective,\ni.e., by adding random noise into deterministic weights. The randomized weights\nenable our design of a novel adversarial training method via Taylor expansion\nof a small Gaussian noise, and we show that the new adversarial training method\ncan flatten loss landscape and find flat minima. With PGD, CW, and Auto\nAttacks, an extensive set of experiments demonstrate that our method enhances\nthe state-of-the-art adversarial training methods, boosting both robustness and\nclean accuracy. The code is available at\nhttps://github.com/Alexkael/Randomized-Adversarial-Training.\n", "date": "2023-03-19", "categories": []}
{"id": "2303.10699", "title": "FVQA 2.0: Introducing Adversarial Samples into Fact-based Visual\n  Question Answering", "abstract": "  The widely used Fact-based Visual Question Answering (FVQA) dataset contains\nvisually-grounded questions that require information retrieval using common\nsense knowledge graphs to answer. It has been observed that the original\ndataset is highly imbalanced and concentrated on a small portion of its\nassociated knowledge graph. We introduce FVQA 2.0 which contains adversarial\nvariants of test questions to address this imbalance. We show that systems\ntrained with the original FVQA train sets can be vulnerable to adversarial\nsamples and we demonstrate an augmentation scheme to reduce this vulnerability\nwithout human annotations.\n", "date": "2023-03-19", "categories": []}
{"id": "2303.11135", "title": "TWINS: A Fine-Tuning Framework for Improved Transferability of\n  Adversarial Robustness and Generalization", "abstract": "  Recent years have seen the ever-increasing importance of pre-trained models\nand their downstream training in deep learning research and applications. At\nthe same time, the defense for adversarial examples has been mainly\ninvestigated in the context of training from random initialization on simple\nclassification tasks. To better exploit the potential of pre-trained models in\nadversarial robustness, this paper focuses on the fine-tuning of an\nadversarially pre-trained model in various classification tasks. Existing\nresearch has shown that since the robust pre-trained model has already learned\na robust feature extractor, the crucial question is how to maintain the\nrobustness in the pre-trained model when learning the downstream task. We study\nthe model-based and data-based approaches for this goal and find that the two\ncommon approaches cannot achieve the objective of improving both generalization\nand adversarial robustness. Thus, we propose a novel statistics-based approach,\nTwo-WIng NormliSation (TWINS) fine-tuning framework, which consists of two\nneural networks where one of them keeps the population means and variances of\npre-training data in the batch normalization layers. Besides the robust\ninformation transfer, TWINS increases the effective learning rate without\nhurting the training stability since the relationship between a weight norm and\nits gradient norm in standard batch normalization layer is broken, resulting in\na faster escape from the sub-optimal initialization and alleviating the robust\noverfitting. Finally, TWINS is shown to be effective on a wide range of image\nclassification datasets in terms of both generalization and robustness. Our\ncode is available at https://github.com/ziquanliu/CVPR2023-TWINS.\n", "date": "2023-03-20", "categories": []}
{"id": "2303.09731", "title": "Exorcising ''Wraith'': Protecting LiDAR-based Object Detector in\n  Automated Driving System from Appearing Attacks", "abstract": "  Automated driving systems rely on 3D object detectors to recognize possible\nobstacles from LiDAR point clouds. However, recent works show the adversary can\nforge non-existent cars in the prediction results with a few fake points (i.e.,\nappearing attack). By removing statistical outliers, existing defenses are\nhowever designed for specific attacks or biased by predefined heuristic rules.\nTowards more comprehensive mitigation, we first systematically inspect the\nmechanism of recent appearing attacks: Their common weaknesses are observed in\ncrafting fake obstacles which (i) have obvious differences in the local parts\ncompared with real obstacles and (ii) violate the physical relation between\ndepth and point density. In this paper, we propose a novel plug-and-play\ndefensive module which works by side of a trained LiDAR-based object detector\nto eliminate forged obstacles where a major proportion of local parts have low\nobjectness, i.e., to what degree it belongs to a real object. At the core of\nour module is a local objectness predictor, which explicitly incorporates the\ndepth information to model the relation between depth and point density, and\npredicts each local part of an obstacle with an objectness score. Extensive\nexperiments show, our proposed defense eliminates at least 70% cars forged by\nthree known appearing attacks in most cases, while, for the best previous\ndefense, less than 30% forged cars are eliminated. Meanwhile, under the same\ncircumstance, our defense incurs less overhead for AP/precision on cars\ncompared with existing defenses. Furthermore, We validate the effectiveness of\nour proposed defense on simulation-based closed-loop control driving tests in\nthe open-source system of Baidu's Apollo.\n", "date": "2023-03-17", "categories": []}
{"id": "2303.09962", "title": "Adversarial Counterfactual Visual Explanations", "abstract": "  Counterfactual explanations and adversarial attacks have a related goal:\nflipping output labels with minimal perturbations regardless of their\ncharacteristics. Yet, adversarial attacks cannot be used directly in a\ncounterfactual explanation perspective, as such perturbations are perceived as\nnoise and not as actionable and understandable image modifications. Building on\nthe robust learning literature, this paper proposes an elegant method to turn\nadversarial attacks into semantically meaningful perturbations, without\nmodifying the classifiers to explain. The proposed approach hypothesizes that\nDenoising Diffusion Probabilistic Models are excellent regularizers for\navoiding high-frequency and out-of-distribution perturbations when generating\nadversarial attacks. The paper's key idea is to build attacks through a\ndiffusion model to polish them. This allows studying the target model\nregardless of its robustification level. Extensive experimentation shows the\nadvantages of our counterfactual explanation approach over current\nState-of-the-Art in multiple testbeds.\n", "date": "2023-03-17", "categories": []}
{"id": "2303.10078", "title": "Fuzziness-tuned: Improving the Transferability of Adversarial Examples", "abstract": "  With the development of adversarial attacks, adversairal examples have been\nwidely used to enhance the robustness of the training models on deep neural\nnetworks. Although considerable efforts of adversarial attacks on improving the\ntransferability of adversarial examples have been developed, the attack success\nrate of the transfer-based attacks on the surrogate model is much higher than\nthat on victim model under the low attack strength (e.g., the attack strength\n$\\epsilon=8/255$). In this paper, we first systematically investigated this\nissue and found that the enormous difference of attack success rates between\nthe surrogate model and victim model is caused by the existence of a special\narea (known as fuzzy domain in our paper), in which the adversarial examples in\nthe area are classified wrongly by the surrogate model while correctly by the\nvictim model. Then, to eliminate such enormous difference of attack success\nrates for improving the transferability of generated adversarial examples, a\nfuzziness-tuned method consisting of confidence scaling mechanism and\ntemperature scaling mechanism is proposed to ensure the generated adversarial\nexamples can effectively skip out of the fuzzy domain. The confidence scaling\nmechanism and the temperature scaling mechanism can collaboratively tune the\nfuzziness of the generated adversarial examples through adjusting the gradient\ndescent weight of fuzziness and stabilizing the update direction, respectively.\nSpecifically, the proposed fuzziness-tuned method can be effectively integrated\nwith existing adversarial attacks to further improve the transferability of\nadverarial examples without changing the time complexity. Extensive experiments\ndemonstrated that fuzziness-tuned method can effectively enhance the\ntransferability of adversarial examples in the latest transfer-based attacks.\n", "date": "2023-03-17", "categories": []}
{"id": "2212.13667", "title": "Learning When to Use Adaptive Adversarial Image Perturbations against\n  Autonomous Vehicles", "abstract": "  The deep neural network (DNN) models for object detection using camera images\nare widely adopted in autonomous vehicles. However, DNN models are shown to be\nsusceptible to adversarial image perturbations. In the existing methods of\ngenerating the adversarial image perturbations, optimizations take each\nincoming image frame as the decision variable to generate an image\nperturbation. Therefore, given a new image, the typically\ncomputationally-expensive optimization needs to start over as there is no\nlearning between the independent optimizations. Very few approaches have been\ndeveloped for attacking online image streams while considering the underlying\nphysical dynamics of autonomous vehicles, their mission, and the environment.\nWe propose a multi-level stochastic optimization framework that monitors an\nattacker's capability of generating the adversarial perturbations. Based on\nthis capability level, a binary decision attack/not attack is introduced to\nenhance the effectiveness of the attacker. We evaluate our proposed multi-level\nimage attack framework using simulations for vision-guided autonomous vehicles\nand actual tests with a small indoor drone in an office environment. The\nresults show our method's capability to generate the image attack in real-time\nwhile monitoring when the attacker is proficient given state estimates.\n", "date": "2022-12-28", "categories": []}
{"id": "2303.08289", "title": "Improving Adversarial Robustness with Hypersphere Embedding and\n  Angular-based Regularizations", "abstract": "  Adversarial training (AT) methods have been found to be effective against\nadversarial attacks on deep neural networks. Many variants of AT have been\nproposed to improve its performance. Pang et al. [1] have recently shown that\nincorporating hypersphere embedding (HE) into the existing AT procedures\nenhances robustness. We observe that the existing AT procedures are not\ndesigned for the HE framework, and thus fail to adequately learn the angular\ndiscriminative information available in the HE framework. In this paper, we\npropose integrating HE into AT with regularization terms that exploit the rich\nangular information available in the HE framework. Specifically, our method,\ntermed angular-AT, adds regularization terms to AT that explicitly enforce\nweight-feature compactness and inter-class separation; all expressed in terms\nof angular features. Experimental results show that angular-AT further improves\nadversarial robustness.\n", "date": "2023-03-15", "categories": []}
{"id": "2207.03933", "title": "A law of adversarial risk, interpolation, and label noise", "abstract": "  In supervised learning, it has been shown that label noise in the data can be\ninterpolated without penalties on test accuracy. We show that interpolating\nlabel noise induces adversarial vulnerability, and prove the first theorem\nshowing the relationship between label noise and adversarial risk for any data\ndistribution. Our results are almost tight if we do not make any assumptions on\nthe inductive bias of the learning algorithm. We then investigate how different\ncomponents of this problem affect this result, including properties of the\ndistribution. We also discuss non-uniform label noise distributions; and prove\na new theorem showing uniform label noise induces nearly as large an\nadversarial risk as the worst poisoning with the same noise rate. Then, we\nprovide theoretical and empirical evidence that uniform label noise is more\nharmful than typical real-world label noise. Finally, we show how inductive\nbiases amplify the effect of label noise and argue the need for future work in\nthis direction.\n", "date": "2022-07-08", "categories": []}
{"id": "2210.10272", "title": "Training set cleansing of backdoor poisoning by self-supervised\n  representation learning", "abstract": "  A backdoor or Trojan attack is an important type of data poisoning attack\nagainst deep neural network (DNN) classifiers, wherein the training dataset is\npoisoned with a small number of samples that each possess the backdoor pattern\n(usually a pattern that is either imperceptible or innocuous) and which are\nmislabeled to the attacker's target class. When trained on a backdoor-poisoned\ndataset, a DNN behaves normally on most benign test samples but makes incorrect\npredictions to the target class when the test sample has the backdoor pattern\nincorporated (i.e., contains a backdoor trigger). Here we focus on image\nclassification tasks and show that supervised training may build stronger\nassociation between the backdoor pattern and the associated target class than\nthat between normal features and the true class of origin. By contrast,\nself-supervised representation learning ignores the labels of samples and\nlearns a feature embedding based on images' semantic content. %We thus propose\nto use unsupervised representation learning to avoid emphasising\nbackdoor-poisoned training samples and learn a similar feature embedding for\nsamples of the same class. Using a feature embedding found by self-supervised\nrepresentation learning, a data cleansing method, which combines sample\nfiltering and re-labeling, is developed. Experiments on CIFAR-10 benchmark\ndatasets show that our method achieves state-of-the-art performance in\nmitigating backdoor attacks.\n", "date": "2022-10-19", "categories": []}
{"id": "2301.06923", "title": "Explainable Data Poison Attacks on Human Emotion Evaluation Systems\n  based on EEG Signals", "abstract": "  The major aim of this paper is to explain the data poisoning attacks using\nlabel-flipping during the training stage of the electroencephalogram (EEG)\nsignal-based human emotion evaluation systems deploying Machine Learning models\nfrom the attackers' perspective. Human emotion evaluation using EEG signals has\nconsistently attracted a lot of research attention. The identification of human\nemotional states based on EEG signals is effective to detect potential internal\nthreats caused by insider individuals. Nevertheless, EEG signal-based human\nemotion evaluation systems have shown several vulnerabilities to data poison\nattacks. The findings of the experiments demonstrate that the suggested data\npoison assaults are model-independently successful, although various models\nexhibit varying levels of resilience to the attacks. In addition, the data\npoison attacks on the EEG signal-based human emotion evaluation systems are\nexplained with several Explainable Artificial Intelligence (XAI) methods,\nincluding Shapley Additive Explanation (SHAP) values, Local Interpretable\nModel-agnostic Explanations (LIME), and Generated Decision Trees. And the codes\nof this paper are publicly available on GitHub.\n", "date": "2023-01-17", "categories": []}
{"id": "2303.05077", "title": "Learning the Legibility of Visual Text Perturbations", "abstract": "  Many adversarial attacks in NLP perturb inputs to produce visually similar\nstrings ('ergo' $\\rightarrow$ '$\\epsilon$rgo') which are legible to humans but\ndegrade model performance. Although preserving legibility is a necessary\ncondition for text perturbation, little work has been done to systematically\ncharacterize it; instead, legibility is typically loosely enforced via\nintuitions around the nature and extent of perturbations. Particularly, it is\nunclear to what extent can inputs be perturbed while preserving legibility, or\nhow to quantify the legibility of a perturbed string. In this work, we address\nthis gap by learning models that predict the legibility of a perturbed string,\nand rank candidate perturbations based on their legibility. To do so, we\ncollect and release LEGIT, a human-annotated dataset comprising the legibility\nof visually perturbed text. Using this dataset, we build both text- and\nvision-based models which achieve up to $0.91$ F1 score in predicting whether\nan input is legible, and an accuracy of $0.86$ in predicting which of two given\nperturbations is more legible. Additionally, we discover that legible\nperturbations from the LEGIT dataset are more effective at lowering the\nperformance of NLP models than best-known attack strategies, suggesting that\ncurrent models may be vulnerable to a broad range of perturbations beyond what\nis captured by existing visual attacks. Data, code, and models are available at\nhttps://github.com/dvsth/learning-legibility-2023.\n", "date": "2023-03-09", "categories": []}
{"id": "2303.06151", "title": "NoiseCAM: Explainable AI for the Boundary Between Noise and Adversarial\n  Attacks", "abstract": "  Deep Learning (DL) and Deep Neural Networks (DNNs) are widely used in various\ndomains. However, adversarial attacks can easily mislead a neural network and\nlead to wrong decisions. Defense mechanisms are highly preferred in\nsafety-critical applications. In this paper, firstly, we use the gradient class\nactivation map (GradCAM) to analyze the behavior deviation of the VGG-16\nnetwork when its inputs are mixed with adversarial perturbation or Gaussian\nnoise. In particular, our method can locate vulnerable layers that are\nsensitive to adversarial perturbation and Gaussian noise. We also show that the\nbehavior deviation of vulnerable layers can be used to detect adversarial\nexamples. Secondly, we propose a novel NoiseCAM algorithm that integrates\ninformation from globally and pixel-level weighted class activation maps. Our\nalgorithm is susceptible to adversarial perturbations and will not respond to\nGaussian random noise mixed in the inputs. Third, we compare detecting\nadversarial examples using both behavior deviation and NoiseCAM, and we show\nthat NoiseCAM outperforms behavior deviation modeling in its overall\nperformance. Our work could provide a useful tool to defend against certain\nadversarial attacks on deep neural networks.\n", "date": "2023-03-09", "categories": []}
{"id": "2303.06302", "title": "Adversarial Attacks and Defenses in Machine Learning-Powered Networks: A\n  Contemporary Survey", "abstract": "  Adversarial attacks and defenses in machine learning and deep neural network\nhave been gaining significant attention due to the rapidly growing applications\nof deep learning in the Internet and relevant scenarios. This survey provides a\ncomprehensive overview of the recent advancements in the field of adversarial\nattack and defense techniques, with a focus on deep neural network-based\nclassification models. Specifically, we conduct a comprehensive classification\nof recent adversarial attack methods and state-of-the-art adversarial defense\ntechniques based on attack principles, and present them in visually appealing\ntables and tree diagrams. This is based on a rigorous evaluation of the\nexisting works, including an analysis of their strengths and limitations. We\nalso categorize the methods into counter-attack detection and robustness\nenhancement, with a specific focus on regularization-based methods for\nenhancing robustness. New avenues of attack are also explored, including\nsearch-based, decision-based, drop-based, and physical-world attacks, and a\nhierarchical classification of the latest defense methods is provided,\nhighlighting the challenges of balancing training costs with performance,\nmaintaining clean accuracy, overcoming the effect of gradient masking, and\nensuring method transferability. At last, the lessons learned and open\nchallenges are summarized with future research opportunities recommended.\n", "date": "2023-03-11", "categories": []}
{"id": "2303.06425", "title": "Improving the Robustness of Deep Convolutional Neural Networks Through\n  Feature Learning", "abstract": "  Deep convolutional neural network (DCNN for short) models are vulnerable to\nexamples with small perturbations. Adversarial training (AT for short) is a\nwidely used approach to enhance the robustness of DCNN models by data\naugmentation. In AT, the DCNN models are trained with clean examples and\nadversarial examples (AE for short) which are generated using a specific attack\nmethod, aiming to gain ability to defend themselves when facing the unseen AEs.\nHowever, in practice, the trained DCNN models are often fooled by the AEs\ngenerated by the novel attack methods. This naturally raises a question: can a\nDCNN model learn certain features which are insensitive to small perturbations,\nand further defend itself no matter what attack methods are presented. To\nanswer this question, this paper makes a beginning effort by proposing a\nshallow binary feature module (SBFM for short), which can be integrated into\nany popular backbone. The SBFM includes two types of layers, i.e., Sobel layer\nand threshold layer. In Sobel layer, there are four parallel feature maps which\nrepresent horizontal, vertical, and diagonal edge features, respectively. And\nin threshold layer, it turns the edge features learnt by Sobel layer to the\nbinary features, which then are feeded into the fully connected layers for\nclassification with the features learnt by the backbone. We integrate SBFM into\nVGG16 and ResNet34, respectively, and conduct experiments on multiple datasets.\nExperimental results demonstrate, under FGSM attack with $\\epsilon=8/255$, the\nSBFM integrated models can achieve averagely 35\\% higher accuracy than the\noriginal ones, and in CIFAR-10 and TinyImageNet datasets, the SBFM integrated\nmodels can achieve averagely 75\\% classification accuracy. The work in this\npaper shows it is promising to enhance the robustness of DCNN models through\nfeature learning.\n", "date": "2023-03-11", "categories": []}
{"id": "2303.06641", "title": "Adaptive Local Adversarial Attacks on 3D Point Clouds for Augmented\n  Reality", "abstract": "  As the key technology of augmented reality (AR), 3D recognition and tracking\nare always vulnerable to adversarial examples, which will cause serious\nsecurity risks to AR systems. Adversarial examples are beneficial to improve\nthe robustness of the 3D neural network model and enhance the stability of the\nAR system. At present, most 3D adversarial attack methods perturb the entire\npoint cloud to generate adversarial examples, which results in high\nperturbation costs and difficulty in reconstructing the corresponding real\nobjects in the physical world. In this paper, we propose an adaptive local\nadversarial attack method (AL-Adv) on 3D point clouds to generate adversarial\npoint clouds. First, we analyze the vulnerability of the 3D network model and\nextract the salient regions of the input point cloud, namely the vulnerable\nregions. Second, we propose an adaptive gradient attack algorithm that targets\nvulnerable regions. The proposed attack algorithm adaptively assigns different\ndisturbances in different directions of the three-dimensional coordinates of\nthe point cloud. Experimental results show that our proposed method AL-Adv\nachieves a higher attack success rate than the global attack method.\nSpecifically, the adversarial examples generated by the AL-Adv demonstrate good\nimperceptibility and small generation costs.\n", "date": "2023-03-12", "categories": []}
{"id": "2303.06664", "title": "Adv-Bot: Realistic Adversarial Botnet Attacks against Network Intrusion\n  Detection Systems", "abstract": "  Due to the numerous advantages of machine learning (ML) algorithms, many\napplications now incorporate them. However, many studies in the field of image\nclassification have shown that MLs can be fooled by a variety of adversarial\nattacks. These attacks take advantage of ML algorithms' inherent vulnerability.\nThis raises many questions in the cybersecurity field, where a growing number\nof researchers are recently investigating the feasibility of such attacks\nagainst machine learning-based security systems, such as intrusion detection\nsystems. The majority of this research demonstrates that it is possible to fool\na model using features extracted from a raw data source, but it does not take\ninto account the real implementation of such attacks, i.e., the reverse\ntransformation from theory to practice. The real implementation of these\nadversarial attacks would be influenced by various constraints that would make\ntheir execution more difficult. As a result, the purpose of this study was to\ninvestigate the actual feasibility of adversarial attacks, specifically evasion\nattacks, against network-based intrusion detection systems (NIDS),\ndemonstrating that it is entirely possible to fool these ML-based IDSs using\nour proposed adversarial algorithm while assuming as many constraints as\npossible in a black-box setting. In addition, since it is critical to design\ndefense mechanisms to protect ML-based IDSs against such attacks, a defensive\nscheme is presented. Realistic botnet traffic traces are used to assess this\nwork. Our goal is to create adversarial botnet traffic that can avoid detection\nwhile still performing all of its intended malicious functionality.\n", "date": "2023-03-12", "categories": []}
{"id": "2303.06818", "title": "Backdoor Defense via Deconfounded Representation Learning", "abstract": "  Deep neural networks (DNNs) are recently shown to be vulnerable to backdoor\nattacks, where attackers embed hidden backdoors in the DNN model by injecting a\nfew poisoned examples into the training dataset. While extensive efforts have\nbeen made to detect and remove backdoors from backdoored DNNs, it is still not\nclear whether a backdoor-free clean model can be directly obtained from\npoisoned datasets. In this paper, we first construct a causal graph to model\nthe generation process of poisoned data and find that the backdoor attack acts\nas the confounder, which brings spurious associations between the input images\nand target labels, making the model predictions less reliable. Inspired by the\ncausal understanding, we propose the Causality-inspired Backdoor Defense (CBD),\nto learn deconfounded representations for reliable classification.\nSpecifically, a backdoored model is intentionally trained to capture the\nconfounding effects. The other clean model dedicates to capturing the desired\ncausal effects by minimizing the mutual information with the confounding\nrepresentations from the backdoored model and employing a sample-wise\nre-weighting scheme. Extensive experiments on multiple benchmark datasets\nagainst 6 state-of-the-art attacks verify that our proposed defense method is\neffective in reducing backdoor threats while maintaining high accuracy in\npredicting benign samples. Further analysis shows that CBD can also resist\npotential adaptive attacks. The code is available at\n\\url{https://github.com/zaixizhang/CBD}.\n", "date": "2023-03-13", "categories": []}
{"id": "2303.07003", "title": "Review on the Feasibility of Adversarial Evasion Attacks and Defenses\n  for Network Intrusion Detection Systems", "abstract": "  Nowadays, numerous applications incorporate machine learning (ML) algorithms\ndue to their prominent achievements. However, many studies in the field of\ncomputer vision have shown that ML can be fooled by intentionally crafted\ninstances, called adversarial examples. These adversarial examples take\nadvantage of the intrinsic vulnerability of ML models. Recent research raises\nmany concerns in the cybersecurity field. An increasing number of researchers\nare studying the feasibility of such attacks on security systems based on ML\nalgorithms, such as Intrusion Detection Systems (IDS). The feasibility of such\nadversarial attacks would be influenced by various domain-specific constraints.\nThis can potentially increase the difficulty of crafting adversarial examples.\nDespite the considerable amount of research that has been done in this area,\nmuch of it focuses on showing that it is possible to fool a model using\nfeatures extracted from the raw data but does not address the practical side,\ni.e., the reverse transformation from theory to practice. For this reason, we\npropose a review browsing through various important papers to provide a\ncomprehensive analysis. Our analysis highlights some challenges that have not\nbeen addressed in the reviewed papers.\n", "date": "2023-03-13", "categories": []}
{"id": "2303.07199", "title": "BeamAttack: Generating High-quality Textual Adversarial Examples through\n  Beam Search and Mixed Semantic Spaces", "abstract": "  Natural language processing models based on neural networks are vulnerable to\nadversarial examples. These adversarial examples are imperceptible to human\nreaders but can mislead models to make the wrong predictions. In a black-box\nsetting, attacker can fool the model without knowing model's parameters and\narchitecture. Previous works on word-level attacks widely use single semantic\nspace and greedy search as a search strategy. However, these methods fail to\nbalance the attack success rate, quality of adversarial examples and time\nconsumption. In this paper, we propose BeamAttack, a textual attack algorithm\nthat makes use of mixed semantic spaces and improved beam search to craft\nhigh-quality adversarial examples. Extensive experiments demonstrate that\nBeamAttack can improve attack success rate while saving numerous queries and\ntime, e.g., improving at most 7\\% attack success rate than greedy search when\nattacking the examples from MR dataset. Compared with heuristic search,\nBeamAttack can save at most 85\\% model queries and achieve a competitive attack\nsuccess rate. The adversarial examples crafted by BeamAttack are highly\ntransferable and can effectively improve model's robustness during adversarial\ntraining. Code is available at\nhttps://github.com/zhuhai-ustc/beamattack/tree/master\n", "date": "2023-03-09", "categories": []}
{"id": "2209.14105", "title": "Exploring the Relationship between Architecture and Adversarially Robust\n  Generalization", "abstract": "  Adversarial training has been demonstrated to be one of the most effective\nremedies for defending adversarial examples, yet it often suffers from the huge\nrobustness generalization gap on unseen testing adversaries, deemed as the\nadversarially robust generalization problem. Despite the preliminary\nunderstandings devoted to adversarially robust generalization, little is known\nfrom the architectural perspective. To bridge the gap, this paper for the first\ntime systematically investigated the relationship between adversarially robust\ngeneralization and architectural design. Inparticular, we comprehensively\nevaluated 20 most representative adversarially trained architectures on\nImageNette and CIFAR-10 datasets towards multiple `p-norm adversarial attacks.\nBased on the extensive experiments, we found that, under aligned settings,\nVision Transformers (e.g., PVT, CoAtNet) often yield better adversarially\nrobust generalization while CNNs tend to overfit on specific attacks and fail\nto generalize on multiple adversaries. To better understand the nature behind\nit, we conduct theoretical analysis via the lens of Rademacher complexity. We\nrevealed the fact that the higher weight sparsity contributes significantly\ntowards the better adversarially robust generalization of Transformers, which\ncan be often achieved by the specially-designed attention blocks. We hope our\npaper could help to better understand the mechanism for designing robust DNNs.\nOur model weights can be found at http://robust.art.\n", "date": "2022-09-28", "categories": []}
{"id": "2303.05593", "title": "Learning the Wrong Lessons: Inserting Trojans During Knowledge\n  Distillation", "abstract": "  In recent years, knowledge distillation has become a cornerstone of\nefficiently deployed machine learning, with labs and industries using knowledge\ndistillation to train models that are inexpensive and resource-optimized.\nTrojan attacks have contemporaneously gained significant prominence, revealing\nfundamental vulnerabilities in deep learning models. Given the widespread use\nof knowledge distillation, in this work we seek to exploit the unlabelled data\nknowledge distillation process to embed Trojans in a student model without\nintroducing conspicuous behavior in the teacher. We ultimately devise a Trojan\nattack that effectively reduces student accuracy, does not alter teacher\nperformance, and is efficiently constructible in practice.\n", "date": "2023-03-09", "categories": []}
{"id": "2303.05719", "title": "Boosting Adversarial Attacks by Leveraging Decision Boundary Information", "abstract": "  Due to the gap between a substitute model and a victim model, the\ngradient-based noise generated from a substitute model may have low\ntransferability for a victim model since their gradients are different.\nInspired by the fact that the decision boundaries of different models do not\ndiffer much, we conduct experiments and discover that the gradients of\ndifferent models are more similar on the decision boundary than in the original\nposition. Moreover, since the decision boundary in the vicinity of an input\nimage is flat along most directions, we conjecture that the boundary gradients\ncan help find an effective direction to cross the decision boundary of the\nvictim models. Based on it, we propose a Boundary Fitting Attack to improve\ntransferability. Specifically, we introduce a method to obtain a set of\nboundary points and leverage the gradient information of these points to update\nthe adversarial examples. Notably, our method can be combined with existing\ngradient-based methods. Extensive experiments prove the effectiveness of our\nmethod, i.e., improving the success rate by 5.6% against normally trained CNNs\nand 14.9% against defense CNNs on average compared to state-of-the-art\ntransfer-based attacks. Further we compare transformers with CNNs, the results\nindicate that transformers are more robust than CNNs. However, our method still\noutperforms existing methods when attacking transformers. Specifically, when\nusing CNNs as substitute models, our method obtains an average attack success\nrate of 58.2%, which is 10.8% higher than other state-of-the-art transfer-based\nattacks.\n", "date": "2023-03-10", "categories": []}
{"id": "2303.05758", "title": "MIXPGD: Hybrid Adversarial Training for Speech Recognition Systems", "abstract": "  Automatic speech recognition (ASR) systems based on deep neural networks are\nweak against adversarial perturbations. We propose mixPGD adversarial training\nmethod to improve the robustness of the model for ASR systems. In standard\nadversarial training, adversarial samples are generated by leveraging\nsupervised or unsupervised methods. We merge the capabilities of both\nsupervised and unsupervised approaches in our method to generate new\nadversarial samples which aid in improving model robustness. Extensive\nexperiments and comparison across various state-of-the-art defense methods and\nadversarial attacks have been performed to show that mixPGD gains 4.1% WER of\nbetter performance than previous best performing models under white-box\nadversarial attack setting. We tested our proposed defense method against both\nwhite-box and transfer based black-box attack settings to ensure that our\ndefense strategy is robust against various types of attacks. Empirical results\non several adversarial attacks validate the effectiveness of our proposed\napproach.\n", "date": "2023-03-10", "categories": []}
{"id": "2303.05762", "title": "TrojDiff: Trojan Attacks on Diffusion Models with Diverse Targets", "abstract": "  Diffusion models have achieved great success in a range of tasks, such as\nimage synthesis and molecule design. As such successes hinge on large-scale\ntraining data collected from diverse sources, the trustworthiness of these\ncollected data is hard to control or audit. In this work, we aim to explore the\nvulnerabilities of diffusion models under potential training data manipulations\nand try to answer: How hard is it to perform Trojan attacks on well-trained\ndiffusion models? What are the adversarial targets that such Trojan attacks can\nachieve? To answer these questions, we propose an effective Trojan attack\nagainst diffusion models, TrojDiff, which optimizes the Trojan diffusion and\ngenerative processes during training. In particular, we design novel\ntransitions during the Trojan diffusion process to diffuse adversarial targets\ninto a biased Gaussian distribution and propose a new parameterization of the\nTrojan generative process that leads to an effective training objective for the\nattack. In addition, we consider three types of adversarial targets: the\nTrojaned diffusion models will always output instances belonging to a certain\nclass from the in-domain distribution (In-D2D attack), out-of-domain\ndistribution (Out-D2D-attack), and one specific instance (D2I attack). We\nevaluate TrojDiff on CIFAR-10 and CelebA datasets against both DDPM and DDIM\ndiffusion models. We show that TrojDiff always achieves high attack performance\nunder different adversarial targets using different types of triggers, while\nthe performance in benign environments is preserved. The code is available at\nhttps://github.com/chenweixin107/TrojDiff.\n", "date": "2023-03-10", "categories": []}
{"id": "2303.06032", "title": "Exploring Adversarial Attacks on Neural Networks: An Explainable\n  Approach", "abstract": "  Deep Learning (DL) is being applied in various domains, especially in\nsafety-critical applications such as autonomous driving. Consequently, it is of\ngreat significance to ensure the robustness of these methods and thus\ncounteract uncertain behaviors caused by adversarial attacks. In this paper, we\nuse gradient heatmaps to analyze the response characteristics of the VGG-16\nmodel when the input images are mixed with adversarial noise and statistically\nsimilar Gaussian random noise. In particular, we compare the network response\nlayer by layer to determine where errors occurred. Several interesting findings\nare derived. First, compared to Gaussian random noise, intentionally generated\nadversarial noise causes severe behavior deviation by distracting the area of\nconcentration in the networks. Second, in many cases, adversarial examples only\nneed to compromise a few intermediate blocks to mislead the final decision.\nThird, our experiments revealed that specific blocks are more vulnerable and\neasier to exploit by adversarial examples. Finally, we demonstrate that the\nlayers $Block4\\_conv1$ and $Block5\\_cov1$ of the VGG-16 model are more\nsusceptible to adversarial attacks. Our work could provide valuable insights\ninto developing more reliable Deep Neural Network (DNN) models.\n", "date": "2023-03-08", "categories": []}
{"id": "2209.09117", "title": "Part-Based Models Improve Adversarial Robustness", "abstract": "  We show that combining human prior knowledge with end-to-end learning can\nimprove the robustness of deep neural networks by introducing a part-based\nmodel for object classification. We believe that the richer form of annotation\nhelps guide neural networks to learn more robust features without requiring\nmore samples or larger models. Our model combines a part segmentation model\nwith a tiny classifier and is trained end-to-end to simultaneously segment\nobjects into parts and then classify the segmented object. Empirically, our\npart-based models achieve both higher accuracy and higher adversarial\nrobustness than a ResNet-50 baseline on all three datasets. For instance, the\nclean accuracy of our part models is up to 15 percentage points higher than the\nbaseline's, given the same level of robustness. Our experiments indicate that\nthese models also reduce texture bias and yield better robustness against\ncommon corruptions and spurious correlations. The code is publicly available at\nhttps://github.com/chawins/adv-part-model.\n", "date": "2022-09-15", "categories": []}
{"id": "2210.04871", "title": "Certified Training: Small Boxes are All You Need", "abstract": "  To obtain, deterministic guarantees of adversarial robustness, specialized\ntraining methods are used. We propose, SABR, a novel such certified training\nmethod, based on the key insight that propagating interval bounds for a small\nbut carefully selected subset of the adversarial input region is sufficient to\napproximate the worst-case loss over the whole region while significantly\nreducing approximation errors. We show in an extensive empirical evaluation\nthat SABR outperforms existing certified defenses in terms of both standard and\ncertifiable accuracies across perturbation magnitudes and datasets, pointing to\na new class of certified training methods promising to alleviate the\nrobustness-accuracy trade-off.\n", "date": "2022-10-10", "categories": []}
{"id": "2302.13570", "title": "Physical Adversarial Attacks on Deep Neural Networks for Traffic Sign\n  Recognition: A Feasibility Study", "abstract": "  Deep Neural Networks (DNNs) are increasingly applied in the real world in\nsafety critical applications like advanced driver assistance systems. An\nexample for such use case is represented by traffic sign recognition systems.\nAt the same time, it is known that current DNNs can be fooled by adversarial\nattacks, which raises safety concerns if those attacks can be applied under\nrealistic conditions. In this work we apply different black-box attack methods\nto generate perturbations that are applied in the physical environment and can\nbe used to fool systems under different environmental conditions. To the best\nof our knowledge we are the first to combine a general framework for physical\nattacks with different black-box attack methods and study the impact of the\ndifferent methods on the success rate of the attack under the same setting. We\nshow that reliable physical adversarial attacks can be performed with different\nmethods and that it is also possible to reduce the perceptibility of the\nresulting perturbations. The findings highlight the need for viable defenses of\na DNN even in the black-box case, but at the same time form the basis for\nsecuring a DNN with methods like adversarial training which utilizes\nadversarial attacks to augment the original training data.\n", "date": "2023-02-27", "categories": []}
{"id": "2208.00081", "title": "Sampling Attacks on Meta Reinforcement Learning: A Minimax Formulation\n  and Complexity Analysis", "abstract": "  Meta reinforcement learning (meta RL), as a combination of meta-learning\nideas and reinforcement learning (RL), enables the agent to adapt to different\ntasks using a few samples. However, this sampling-based adaptation also makes\nmeta RL vulnerable to adversarial attacks. By manipulating the reward feedback\nfrom sampling processes in meta RL, an attacker can mislead the agent into\nbuilding wrong knowledge from training experience, which deteriorates the\nagent's performance when dealing with different tasks after adaptation. This\npaper provides a game-theoretical underpinning for understanding this type of\nsecurity risk. In particular, we formally define the sampling attack model as a\nStackelberg game between the attacker and the agent, which yields a minimax\nformulation. It leads to two online attack schemes: Intermittent Attack and\nPersistent Attack, which enable the attacker to learn an optimal sampling\nattack, defined by an $\\epsilon$-first-order stationary point, within\n$\\mathcal{O}(\\epsilon^{-2})$ iterations. These attack schemes freeride the\nlearning progress concurrently without extra interactions with the environment.\nBy corroborating the convergence results with numerical experiments, we observe\nthat a minor effort of the attacker can significantly deteriorate the learning\nperformance, and the minimax approach can also help robustify the meta RL\nalgorithms.\n", "date": "2022-07-29", "categories": []}
{"id": "2302.14166", "title": "GLOW: Global Layout Aware Attacks on Object Detection", "abstract": "  Adversarial attacks aim to perturb images such that a predictor outputs\nincorrect results. Due to the limited research in structured attacks, imposing\nconsistency checks on natural multi-object scenes is a promising yet practical\ndefense against conventional adversarial attacks. More desired attacks, to this\nend, should be able to fool defenses with such consistency checks. Therefore,\nwe present the first approach GLOW that copes with various attack requests by\ngenerating global layout-aware adversarial attacks, in which both categorical\nand geometric layout constraints are explicitly established. Specifically, we\nfocus on object detection task and given a victim image, GLOW first localizes\nvictim objects according to target labels. And then it generates multiple\nattack plans, together with their context-consistency scores. Our proposed\nGLOW, on the one hand, is capable of handling various types of requests,\nincluding single or multiple victim objects, with or without specified victim\nobjects. On the other hand, it produces a consistency score for each attack\nplan, reflecting the overall contextual consistency that both semantic category\nand global scene layout are considered. In experiment, we design multiple types\nof attack requests and validate our ideas on MS COCO and Pascal. Extensive\nexperimental results demonstrate that we can achieve about 30$\\%$ average\nrelative improvement compared to state-of-the-art methods in conventional\nsingle object attack request; Moreover, our method outperforms SOTAs\nsignificantly on more generic attack requests by about 20$\\%$ in average;\nFinally, our method produces superior performance under challenging zero-query\nblack-box setting, or 20$\\%$ better than SOTAs. Our code, model and attack\nrequests would be made available.\n", "date": "2023-02-27", "categories": []}
{"id": "2303.04502", "title": "Immune Defense: A Novel Adversarial Defense Mechanism for Preventing the\n  Generation of Adversarial Examples", "abstract": "  The vulnerability of Deep Neural Networks (DNNs) to adversarial examples has\nbeen confirmed. Existing adversarial defenses primarily aim at preventing\nadversarial examples from attacking DNNs successfully, rather than preventing\ntheir generation. If the generation of adversarial examples is unregulated,\nimages within reach are no longer secure and pose a threat to non-robust DNNs.\nAlthough gradient obfuscation attempts to address this issue, it has been shown\nto be circumventable. Therefore, we propose a novel adversarial defense\nmechanism, which is referred to as immune defense and is the example-based\npre-defense. This mechanism applies carefully designed quasi-imperceptible\nperturbations to the raw images to prevent the generation of adversarial\nexamples for the raw images, and thereby protecting both images and DNNs. These\nperturbed images are referred to as Immune Examples (IEs). In the white-box\nimmune defense, we provide a gradient-based and an optimization-based approach,\nrespectively. Additionally, the more complex black-box immune defense is taken\ninto consideration. We propose Masked Gradient Sign Descent (MGSD) to reduce\napproximation error and stabilize the update to improve the transferability of\nIEs and thereby ensure their effectiveness against black-box adversarial\nattacks. The experimental results demonstrate that the optimization-based\napproach has superior performance and better visual quality in white-box immune\ndefense. In contrast, the gradient-based approach has stronger transferability\nand the proposed MGSD significantly improve the transferability of baselines.\n", "date": "2023-03-08", "categories": []}
{"id": "2303.03680", "title": "Logit Margin Matters: Improving Transferable Targeted Adversarial Attack\n  by Logit Calibration", "abstract": "  Previous works have extensively studied the transferability of adversarial\nsamples in untargeted black-box scenarios. However, it still remains\nchallenging to craft targeted adversarial examples with higher transferability\nthan non-targeted ones. Recent studies reveal that the traditional\nCross-Entropy (CE) loss function is insufficient to learn transferable targeted\nadversarial examples due to the issue of vanishing gradient. In this work, we\nprovide a comprehensive investigation of the CE loss function and find that the\nlogit margin between the targeted and untargeted classes will quickly obtain\nsaturation in CE, which largely limits the transferability. Therefore, in this\npaper, we devote to the goal of continually increasing the logit margin along\nthe optimization to deal with the saturation issue and propose two simple and\neffective logit calibration methods, which are achieved by downscaling the\nlogits with a temperature factor and an adaptive margin, respectively. Both of\nthem can effectively encourage optimization to produce a larger logit margin\nand lead to higher transferability. Besides, we show that minimizing the cosine\ndistance between the adversarial examples and the classifier weights of the\ntarget class can further improve the transferability, which is benefited from\ndownscaling logits via L2-normalization. Experiments conducted on the ImageNet\ndataset validate the effectiveness of the proposed methods, which outperform\nthe state-of-the-art methods in black-box targeted attacks. The source code is\navailable at \\href{https://github.com/WJJLL/Target-Attack/}{Link}\n", "date": "2023-03-07", "categories": []}
{"id": "2206.10550", "title": "(Certified!!) Adversarial Robustness for Free!", "abstract": "  In this paper we show how to achieve state-of-the-art certified adversarial\nrobustness to 2-norm bounded perturbations by relying exclusively on\noff-the-shelf pretrained models. To do so, we instantiate the denoised\nsmoothing approach of Salman et al. 2020 by combining a pretrained denoising\ndiffusion probabilistic model and a standard high-accuracy classifier. This\nallows us to certify 71% accuracy on ImageNet under adversarial perturbations\nconstrained to be within an 2-norm of 0.5, an improvement of 14 percentage\npoints over the prior certified SoTA using any approach, or an improvement of\n30 percentage points over denoised smoothing. We obtain these results using\nonly pretrained diffusion models and image classifiers, without requiring any\nfine tuning or retraining of model parameters.\n", "date": "2022-06-21", "categories": []}
{"id": "2211.01806", "title": "BATT: Backdoor Attack with Transformation-based Triggers", "abstract": "  Deep neural networks (DNNs) are vulnerable to backdoor attacks. The backdoor\nadversaries intend to maliciously control the predictions of attacked DNNs by\ninjecting hidden backdoors that can be activated by adversary-specified trigger\npatterns during the training process. One recent research revealed that most of\nthe existing attacks failed in the real physical world since the trigger\ncontained in the digitized test samples may be different from that of the one\nused for training. Accordingly, users can adopt spatial transformations as the\nimage pre-processing to deactivate hidden backdoors. In this paper, we explore\nthe previous findings from another side. We exploit classical spatial\ntransformations (i.e. rotation and translation) with the specific parameter as\ntrigger patterns to design a simple yet effective poisoning-based backdoor\nattack. For example, only images rotated to a particular angle can activate the\nembedded backdoor of attacked DNNs. Extensive experiments are conducted,\nverifying the effectiveness of our attack under both digital and physical\nsettings and its resistance to existing backdoor defenses.\n", "date": "2022-11-02", "categories": []}
{"id": "2211.05638", "title": "Untargeted Backdoor Attack against Object Detection", "abstract": "  Recent studies revealed that deep neural networks (DNNs) are exposed to\nbackdoor threats when training with third-party resources (such as training\nsamples or backbones). The backdoored model has promising performance in\npredicting benign samples, whereas its predictions can be maliciously\nmanipulated by adversaries based on activating its backdoors with pre-defined\ntrigger patterns. Currently, most of the existing backdoor attacks were\nconducted on the image classification under the targeted manner. In this paper,\nwe reveal that these threats could also happen in object detection, posing\nthreatening risks to many mission-critical applications ($e.g.$, pedestrian\ndetection and intelligent surveillance systems). Specifically, we design a\nsimple yet effective poison-only backdoor attack in an untargeted manner, based\non task characteristics. We show that, once the backdoor is embedded into the\ntarget model by our attack, it can trick the model to lose detection of any\nobject stamped with our trigger patterns. We conduct extensive experiments on\nthe benchmark dataset, showing its effectiveness in both digital and\nphysical-world settings and its resistance to potential defenses.\n", "date": "2022-11-02", "categories": []}
{"id": "2303.01052", "title": "Demystifying Causal Features on Adversarial Examples and Causal\n  Inoculation for Robust Network by Adversarial Instrumental Variable\n  Regression", "abstract": "  The origin of adversarial examples is still inexplicable in research fields,\nand it arouses arguments from various viewpoints, albeit comprehensive\ninvestigations. In this paper, we propose a way of delving into the unexpected\nvulnerability in adversarially trained networks from a causal perspective,\nnamely adversarial instrumental variable (IV) regression. By deploying it, we\nestimate the causal relation of adversarial prediction under an unbiased\nenvironment dissociated from unknown confounders. Our approach aims to\ndemystify inherent causal features on adversarial examples by leveraging a\nzero-sum optimization game between a casual feature estimator (i.e., hypothesis\nmodel) and worst-case counterfactuals (i.e., test function) disturbing to find\ncausal features. Through extensive analyses, we demonstrate that the estimated\ncausal features are highly related to the correct prediction for adversarial\nrobustness, and the counterfactuals exhibit extreme features significantly\ndeviating from the correct prediction. In addition, we present how to\neffectively inoculate CAusal FEatures (CAFE) into defense networks for\nimproving adversarial robustness.\n", "date": "2023-03-02", "categories": []}
{"id": "2303.02214", "title": "Adversarial Attacks on Machine Learning in Embedded and IoT Platforms", "abstract": "  Machine learning (ML) algorithms are increasingly being integrated into\nembedded and IoT systems that surround us, and they are vulnerable to\nadversarial attacks. The deployment of these ML algorithms on resource-limited\nembedded platforms also requires the use of model compression techniques. The\nimpact of such model compression techniques on adversarial robustness in ML is\nan important and emerging area of research. This article provides an overview\nof the landscape of adversarial attacks and ML model compression techniques\nrelevant to embedded systems. We then describe efforts that seek to understand\nthe relationship between adversarial attacks and ML model compression before\ndiscussing open problems in this area.\n", "date": "2023-03-03", "categories": []}
{"id": "2303.02322", "title": "Improved Robustness Against Adaptive Attacks With Ensembles and\n  Error-Correcting Output Codes", "abstract": "  Neural network ensembles have been studied extensively in the context of\nadversarial robustness and most ensemble-based approaches remain vulnerable to\nadaptive attacks. In this paper, we investigate the robustness of\nError-Correcting Output Codes (ECOC) ensembles through architectural\nimprovements and ensemble diversity promotion. We perform a comprehensive\nrobustness assessment against adaptive attacks and investigate the relationship\nbetween ensemble diversity and robustness. Our results demonstrate the benefits\nof ECOC ensembles for adversarial robustness compared to regular ensembles of\nconvolutional neural networks (CNNs) and show why the robustness of previous\nimplementations is limited. We also propose an adversarial training method\nspecific to ECOC ensembles that allows to further improve robustness to\nadaptive attacks.\n", "date": "2023-03-04", "categories": []}
{"id": "2301.13122", "title": "Towards Adversarial Realism and Robust Learning for IoT Intrusion\n  Detection and Classification", "abstract": "  The Internet of Things (IoT) faces tremendous security challenges. Machine\nlearning models can be used to tackle the growing number of cyber-attack\nvariations targeting IoT systems, but the increasing threat posed by\nadversarial attacks restates the need for reliable defense strategies. This\nwork describes the types of constraints required for a realistic adversarial\ncyber-attack example and proposes a methodology for a trustworthy adversarial\nrobustness analysis with a realistic adversarial evasion attack vector. The\nproposed methodology was used to evaluate three supervised algorithms, Random\nForest (RF), Extreme Gradient Boosting (XGB), and Light Gradient Boosting\nMachine (LGBM), and one unsupervised algorithm, Isolation Forest (IFOR).\nConstrained adversarial examples were generated with the Adaptative\nPerturbation Pattern Method (A2PM), and evasion attacks were performed against\nmodels created with regular and adversarial training. Even though RF was the\nleast affected in binary classification, XGB consistently achieved the highest\naccuracy in multi-class classification. The obtained results evidence the\ninherent susceptibility of tree-based algorithms and ensembles to adversarial\nevasion attacks and demonstrates the benefits of adversarial training and a\nsecurity by design approach for a more robust IoT network intrusion detection\nand cyber-attack classification.\n", "date": "2023-01-30", "categories": []}
{"id": "2303.01289", "title": "Rethinking the Effect of Data Augmentation in Adversarial Contrastive\n  Learning", "abstract": "  Recent works have shown that self-supervised learning can achieve remarkable\nrobustness when integrated with adversarial training (AT). However, the\nrobustness gap between supervised AT (sup-AT) and self-supervised AT (self-AT)\nremains significant. Motivated by this observation, we revisit existing self-AT\nmethods and discover an inherent dilemma that affects self-AT robustness:\neither strong or weak data augmentations are harmful to self-AT, and a medium\nstrength is insufficient to bridge the gap. To resolve this dilemma, we propose\na simple remedy named DYNACL (Dynamic Adversarial Contrastive Learning). In\nparticular, we propose an augmentation schedule that gradually anneals from a\nstrong augmentation to a weak one to benefit from both extreme cases. Besides,\nwe adopt a fast post-processing stage for adapting it to downstream tasks.\nThrough extensive experiments, we show that DYNACL can improve state-of-the-art\nself-AT robustness by 8.84% under Auto-Attack on the CIFAR-10 dataset, and can\neven outperform vanilla supervised adversarial training for the first time. Our\ncode is available at \\url{https://github.com/PKU-ML/DYNACL}.\n", "date": "2023-03-02", "categories": []}
{"id": "2303.01507", "title": "Defending against Adversarial Audio via Diffusion Model", "abstract": "  Deep learning models have been widely used in commercial acoustic systems in\nrecent years. However, adversarial audio examples can cause abnormal behaviors\nfor those acoustic systems, while being hard for humans to perceive. Various\nmethods, such as transformation-based defenses and adversarial training, have\nbeen proposed to protect acoustic systems from adversarial attacks, but they\nare less effective against adaptive attacks. Furthermore, directly applying the\nmethods from the image domain can lead to suboptimal results because of the\nunique properties of audio data. In this paper, we propose an adversarial\npurification-based defense pipeline, AudioPure, for acoustic systems via\noff-the-shelf diffusion models. Taking advantage of the strong generation\nability of diffusion models, AudioPure first adds a small amount of noise to\nthe adversarial audio and then runs the reverse sampling step to purify the\nnoisy audio and recover clean audio. AudioPure is a plug-and-play method that\ncan be directly applied to any pretrained classifier without any fine-tuning or\nre-training. We conduct extensive experiments on speech command recognition\ntask to evaluate the robustness of AudioPure. Our method is effective against\ndiverse adversarial attacks (e.g. $\\mathcal{L}_2$ or\n$\\mathcal{L}_\\infty$-norm). It outperforms the existing methods under both\nstrong adaptive white-box and black-box attacks bounded by $\\mathcal{L}_2$ or\n$\\mathcal{L}_\\infty$-norm (up to +20\\% in robust accuracy). Besides, we also\nevaluate the certified robustness for perturbations bounded by\n$\\mathcal{L}_2$-norm via randomized smoothing. Our pipeline achieves a higher\ncertified accuracy than baselines.\n", "date": "2023-03-02", "categories": []}
{"id": "2303.01936", "title": "Multi-Agent Adversarial Training Using Diffusion Learning", "abstract": "  This work focuses on adversarial learning over graphs. We propose a general\nadversarial training framework for multi-agent systems using diffusion\nlearning. We analyze the convergence properties of the proposed scheme for\nconvex optimization problems, and illustrate its enhanced robustness to\nadversarial attacks.\n", "date": "2023-03-03", "categories": []}
{"id": "2303.01959", "title": "PointCert: Point Cloud Classification with Deterministic Certified\n  Robustness Guarantees", "abstract": "  Point cloud classification is an essential component in many\nsecurity-critical applications such as autonomous driving and augmented\nreality. However, point cloud classifiers are vulnerable to adversarially\nperturbed point clouds. Existing certified defenses against adversarial point\nclouds suffer from a key limitation: their certified robustness guarantees are\nprobabilistic, i.e., they produce an incorrect certified robustness guarantee\nwith some probability. In this work, we propose a general framework, namely\nPointCert, that can transform an arbitrary point cloud classifier to be\ncertifiably robust against adversarial point clouds with deterministic\nguarantees. PointCert certifiably predicts the same label for a point cloud\nwhen the number of arbitrarily added, deleted, and/or modified points is less\nthan a threshold. Moreover, we propose multiple methods to optimize the\ncertified robustness guarantees of PointCert in three application scenarios. We\nsystematically evaluate PointCert on ModelNet and ScanObjectNN benchmark\ndatasets. Our results show that PointCert substantially outperforms\nstate-of-the-art certified defenses even though their robustness guarantees are\nprobabilistic.\n", "date": "2023-03-03", "categories": []}
{"id": "2303.01068", "title": "Targeted Adversarial Attacks against Neural Machine Translation", "abstract": "  Neural Machine Translation (NMT) systems are used in various applications.\nHowever, it has been shown that they are vulnerable to very small perturbations\nof their inputs, known as adversarial attacks. In this paper, we propose a new\ntargeted adversarial attack against NMT models. In particular, our goal is to\ninsert a predefined target keyword into the translation of the adversarial\nsentence while maintaining similarity between the original sentence and the\nperturbed one in the source domain. To this aim, we propose an optimization\nproblem, including an adversarial loss term and a similarity term. We use\ngradient projection in the embedding space to craft an adversarial sentence.\nExperimental results show that our attack outperforms Seq2Sick, the other\ntargeted adversarial attack against NMT models, in terms of success rate and\ndecrease in translation quality. Our attack succeeds in inserting a keyword\ninto the translation for more than 75% of sentences while similarity with the\noriginal sentence stays preserved.\n", "date": "2023-03-02", "categories": []}
{"id": "2303.01243", "title": "Poster: Sponge ML Model Attacks of Mobile Apps", "abstract": "  Machine Learning (ML)-powered apps are used in pervasive devices such as\nphones, tablets, smartwatches and IoT devices. Recent advances in\ncollaborative, distributed ML such as Federated Learning (FL) attempt to solve\nprivacy concerns of users and data owners, and thus used by tech industry\nleaders such as Google, Facebook and Apple. However, FL systems and models are\nstill vulnerable to adversarial membership and attribute inferences and model\npoisoning attacks, especially in FL-as-a-Service ecosystems recently proposed,\nwhich can enable attackers to access multiple ML-powered apps. In this work, we\nfocus on the recently proposed Sponge attack: It is designed to soak up energy\nconsumed while executing inference (not training) of ML model, without\nhampering the classifier's performance. Recent work has shown sponge attacks on\nASCI-enabled GPUs can potentially escalate the power consumption and inference\ntime. For the first time, in this work, we investigate this attack in the\nmobile setting and measure the effect it can have on ML models running inside\napps on mobile devices.\n", "date": "2023-03-01", "categories": []}
{"id": "2303.01245", "title": "An Incremental Gray-box Physical Adversarial Attack on Neural Network\n  Training", "abstract": "  Neural networks have demonstrated remarkable success in learning and solving\ncomplex tasks in a variety of fields. Nevertheless, the rise of those networks\nin modern computing has been accompanied by concerns regarding their\nvulnerability to adversarial attacks. In this work, we propose a novel\ngradient-free, gray box, incremental attack that targets the training process\nof neural networks. The proposed attack, which implicitly poisons the\nintermediate data structures that retain the training instances between\ntraining epochs acquires its high-risk property from attacking data structures\nthat are typically unobserved by professionals. Hence, the attack goes\nunnoticed despite the damage it can cause. Moreover, the attack can be executed\nwithout the attackers' knowledge of the neural network structure or training\ndata making it more dangerous. The attack was tested under a sensitive\napplication of secure cognitive cities, namely, biometric authentication. The\nconducted experiments showed that the proposed attack is effective and\nstealthy. Finally, the attack effectiveness property was concluded from the\nfact that it was able to flip the sign of the loss gradient in the conducted\nexperiments to become positive, which indicated noisy and unstable training.\nMoreover, the attack was able to decrease the inference probability in the\npoisoned networks compared to their unpoisoned counterparts by 15.37%, 14.68%,\nand 24.88% for the Densenet, VGG, and Xception, respectively. Finally, the\nattack retained its stealthiness despite its high effectiveness. This was\ndemonstrated by the fact that the attack did not cause a notable increase in\nthe training time, in addition, the Fscore values only dropped by an average of\n1.2%, 1.9%, and 1.5% for the poisoned Densenet, VGG, and Xception,\nrespectively.\n", "date": "2023-02-20", "categories": []}
{"id": "2303.01263", "title": "Unnoticeable Backdoor Attacks on Graph Neural Networks", "abstract": "  Graph Neural Networks (GNNs) have achieved promising results in various tasks\nsuch as node classification and graph classification. Recent studies find that\nGNNs are vulnerable to adversarial attacks. However, effective backdoor attacks\non graphs are still an open problem. In particular, backdoor attack poisons the\ngraph by attaching triggers and the target class label to a set of nodes in the\ntraining graph. The backdoored GNNs trained on the poisoned graph will then be\nmisled to predict test nodes to target class once attached with triggers.\nThough there are some initial efforts in graph backdoor attacks, our empirical\nanalysis shows that they may require a large attack budget for effective\nbackdoor attacks and the injected triggers can be easily detected and pruned.\nTherefore, in this paper, we study a novel problem of unnoticeable graph\nbackdoor attacks with limited attack budget. To fully utilize the attack\nbudget, we propose to deliberately select the nodes to inject triggers and\ntarget class labels in the poisoning phase. An adaptive trigger generator is\ndeployed to obtain effective triggers that are difficult to be noticed.\nExtensive experiments on real-world datasets against various defense strategies\ndemonstrate the effectiveness of our proposed method in conducting effective\nunnoticeable backdoor attacks.\n", "date": "2023-02-11", "categories": []}
{"id": "2303.00200", "title": "Feature Extraction Matters More: Universal Deepfake Disruption through\n  Attacking Ensemble Feature Extractors", "abstract": "  Adversarial example is a rising way of protecting facial privacy security\nfrom deepfake modification. To prevent massive facial images from being\nillegally modified by various deepfake models, it is essential to design a\nuniversal deepfake disruptor. However, existing works treat deepfake disruption\nas an End-to-End process, ignoring the functional difference between feature\nextraction and image reconstruction, which makes it difficult to generate a\ncross-model universal disruptor. In this work, we propose a novel\nFeature-Output ensemble UNiversal Disruptor (FOUND) against deepfake networks,\nwhich explores a new opinion that considers attacking feature extractors as the\nmore critical and general task in deepfake disruption. We conduct an effective\ntwo-stage disruption process. We first disrupt multi-model feature extractors\nthrough multi-feature aggregation and individual-feature maintenance, and then\ndevelop a gradient-ensemble algorithm to enhance the disruption effect by\nsimplifying the complex optimization problem of disrupting multiple End-to-End\nmodels. Extensive experiments demonstrate that FOUND can significantly boost\nthe disruption effect against ensemble deepfake benchmark models. Besides, our\nmethod can fast obtain a cross-attribute, cross-image, and cross-model\nuniversal deepfake disruptor with only a few training images, surpassing\nstate-of-the-art universal disruptors in both success rate and efficiency.\n", "date": "2023-03-01", "categories": []}
{"id": "2303.00250", "title": "Combating Exacerbated Heterogeneity for Robust Models in Federated\n  Learning", "abstract": "  Privacy and security concerns in real-world applications have led to the\ndevelopment of adversarially robust federated models. However, the\nstraightforward combination between adversarial training and federated learning\nin one framework can lead to the undesired robustness deterioration. We\ndiscover that the attribution behind this phenomenon is that the generated\nadversarial data could exacerbate the data heterogeneity among local clients,\nmaking the wrapped federated learning perform poorly. To deal with this\nproblem, we propose a novel framework called Slack Federated Adversarial\nTraining (SFAT), assigning the client-wise slack during aggregation to combat\nthe intensified heterogeneity. Theoretically, we analyze the convergence of the\nproposed method to properly relax the objective when combining federated\nlearning and adversarial training. Experimentally, we verify the rationality\nand effectiveness of SFAT on various benchmarked and real-world datasets with\ndifferent adversarial training and federated optimization methods. The code is\npublicly available at https://github.com/ZFancy/SFAT.\n", "date": "2023-03-01", "categories": []}
{"id": "2303.00284", "title": "To Make Yourself Invisible with Adversarial Semantic Contours", "abstract": "  Modern object detectors are vulnerable to adversarial examples, which may\nbring risks to real-world applications. The sparse attack is an important task\nwhich, compared with the popular adversarial perturbation on the whole image,\nneeds to select the potential pixels that is generally regularized by an\n$\\ell_0$-norm constraint, and simultaneously optimize the corresponding\ntexture. The non-differentiability of $\\ell_0$ norm brings challenges and many\nworks on attacking object detection adopted manually-designed patterns to\naddress them, which are meaningless and independent of objects, and therefore\nlead to relatively poor attack performance.\n  In this paper, we propose Adversarial Semantic Contour (ASC), an MAP estimate\nof a Bayesian formulation of sparse attack with a deceived prior of object\ncontour. The object contour prior effectively reduces the search space of pixel\nselection and improves the attack by introducing more semantic bias. Extensive\nexperiments demonstrate that ASC can corrupt the prediction of 9 modern\ndetectors with different architectures (\\e.g., one-stage, two-stage and\nTransformer) by modifying fewer than 5\\% of the pixels of the object area in\nCOCO in white-box scenario and around 10\\% of those in black-box scenario. We\nfurther extend the attack to datasets for autonomous driving systems to verify\nthe effectiveness. We conclude with cautions about contour being the common\nweakness of object detectors with various architecture and the care needed in\napplying them in safety-sensitive scenarios.\n", "date": "2023-03-01", "categories": []}
{"id": "2210.12606", "title": "Nash Equilibria and Pitfalls of Adversarial Training in Adversarial\n  Robustness Games", "abstract": "  Adversarial training is a standard technique for training adversarially\nrobust models. In this paper, we study adversarial training as an alternating\nbest-response strategy in a 2-player zero-sum game. We prove that even in a\nsimple scenario of a linear classifier and a statistical model that abstracts\nrobust vs. non-robust features, the alternating best response strategy of such\ngame may not converge. On the other hand, a unique pure Nash equilibrium of the\ngame exists and is provably robust. We support our theoretical results with\nexperiments, showing the non-convergence of adversarial training and the\nrobustness of Nash equilibrium.\n", "date": "2022-10-23", "categories": []}
{"id": "2302.12366", "title": "Less is More: Data Pruning for Faster Adversarial Training", "abstract": "  Deep neural networks (DNNs) are sensitive to adversarial examples, resulting\nin fragile and unreliable performance in the real world. Although adversarial\ntraining (AT) is currently one of the most effective methodologies to robustify\nDNNs, it is computationally very expensive (e.g., 5-10X costlier than standard\ntraining). To address this challenge, existing approaches focus on single-step\nAT, referred to as Fast AT, reducing the overhead of adversarial example\ngeneration. Unfortunately, these approaches are known to fail against stronger\nadversaries. To make AT computationally efficient without compromising\nrobustness, this paper takes a different view of the efficient AT problem.\nSpecifically, we propose to minimize redundancies at the data level by\nleveraging data pruning. Extensive experiments demonstrate that the data\npruning based AT can achieve similar or superior robust (and clean) accuracy as\nits unpruned counterparts while being significantly faster. For instance,\nproposed strategies accelerate CIFAR-10 training up to 3.44X and CIFAR-100\ntraining to 2.02X. Additionally, the data pruning methods can readily be\nreconciled with existing adversarial acceleration tricks to obtain the striking\nspeed-ups of 5.66X and 5.12X on CIFAR-10, 3.67X and 3.07X on CIFAR-100 with\nTRADES and MART, respectively.\n", "date": "2023-02-23", "categories": []}
{"id": "2302.14059", "title": "Scalable Attribution of Adversarial Attacks via Multi-Task Learning", "abstract": "  Deep neural networks (DNNs) can be easily fooled by adversarial attacks\nduring inference phase when attackers add imperceptible perturbations to\noriginal examples, i.e., adversarial examples. Many works focus on adversarial\ndetection and adversarial training to defend against adversarial attacks.\nHowever, few works explore the tool-chains behind adversarial examples, which\ncan help defenders to seize the clues about the originator of the attack, their\ngoals, and provide insight into the most effective defense algorithm against\ncorresponding attacks. With such a gap, it is necessary to develop techniques\nthat can recognize tool-chains that are leveraged to generate the adversarial\nexamples, which is called Adversarial Attribution Problem (AAP). In this paper,\nAAP is defined as the recognition of three signatures, i.e., {\\em attack\nalgorithm}, {\\em victim model} and {\\em hyperparameter}. Current works transfer\nAAP into single label classification task and ignore the relationship between\nthese signatures. The former will meet combination explosion problem as the\nnumber of signatures is increasing. The latter dictates that we cannot treat\nAAP simply as a single task problem. We first conduct some experiments to\nvalidate the attributability of adversarial examples. Furthermore, we propose a\nmulti-task learning framework named Multi-Task Adversarial Attribution (MTAA)\nto recognize the three signatures simultaneously. MTAA contains perturbation\nextraction module, adversarial-only extraction module and classification and\nregression module. It takes the relationship between attack algorithm and\ncorresponding hyperparameter into account and uses the uncertainty weighted\nloss to adjust the weights of three recognition tasks. The experimental results\non MNIST and ImageNet show the feasibility and scalability of the proposed\nframework as well as its effectiveness in dealing with false alarms.\n", "date": "2023-02-25", "categories": []}
{"id": "2302.14301", "title": "A Comprehensive Study on Robustness of Image Classification Models:\n  Benchmarking and Rethinking", "abstract": "  The robustness of deep neural networks is usually lacking under adversarial\nexamples, common corruptions, and distribution shifts, which becomes an\nimportant research problem in the development of deep learning. Although new\ndeep learning methods and robustness improvement techniques have been\nconstantly proposed, the robustness evaluations of existing methods are often\ninadequate due to their rapid development, diverse noise patterns, and simple\nevaluation metrics. Without thorough robustness evaluations, it is hard to\nunderstand the advances in the field and identify the effective methods. In\nthis paper, we establish a comprehensive robustness benchmark called\n\\textbf{ARES-Bench} on the image classification task. In our benchmark, we\nevaluate the robustness of 55 typical deep learning models on ImageNet with\ndiverse architectures (e.g., CNNs, Transformers) and learning algorithms (e.g.,\nnormal supervised training, pre-training, adversarial training) under numerous\nadversarial attacks and out-of-distribution (OOD) datasets. Using robustness\ncurves as the major evaluation criteria, we conduct large-scale experiments and\ndraw several important findings, including: 1) there is an inherent trade-off\nbetween adversarial and natural robustness for the same model architecture; 2)\nadversarial training effectively improves adversarial robustness, especially\nwhen performed on Transformer architectures; 3) pre-training significantly\nimproves natural robustness based on more training data or self-supervised\nlearning. Based on ARES-Bench, we further analyze the training tricks in\nlarge-scale adversarial training on ImageNet. By designing the training\nsettings accordingly, we achieve the new state-of-the-art adversarial\nrobustness. We have made the benchmarking results and code platform publicly\navailable.\n", "date": "2023-02-28", "categories": []}
{"id": "2302.14677", "title": "Backdoor Attacks Against Deep Image Compression via Adaptive Frequency\n  Trigger", "abstract": "  Recent deep-learning-based compression methods have achieved superior\nperformance compared with traditional approaches. However, deep learning models\nhave proven to be vulnerable to backdoor attacks, where some specific trigger\npatterns added to the input can lead to malicious behavior of the models. In\nthis paper, we present a novel backdoor attack with multiple triggers against\nlearned image compression models. Motivated by the widely used discrete cosine\ntransform (DCT) in existing compression systems and standards, we propose a\nfrequency-based trigger injection model that adds triggers in the DCT domain.\nIn particular, we design several attack objectives for various attacking\nscenarios, including: 1) attacking compression quality in terms of bit-rate and\nreconstruction quality; 2) attacking task-driven measures, such as down-stream\nface recognition and semantic segmentation. Moreover, a novel simple dynamic\nloss is designed to balance the influence of different loss terms adaptively,\nwhich helps achieve more efficient training. Extensive experiments show that\nwith our trained trigger injection models and simple modification of encoder\nparameters (of the compression model), the proposed attack can successfully\ninject several backdoors with corresponding triggers in a single image\ncompression model.\n", "date": "2023-02-28", "categories": []}
{"id": "2210.16777", "title": "Symmetric Saliency-based Adversarial Attack To Speaker Identification", "abstract": "  Adversarial attack approaches to speaker identification either need high\ncomputational cost or are not very effective, to our knowledge. To address this\nissue, in this paper, we propose a novel generation-network-based approach,\ncalled symmetric saliency-based encoder-decoder (SSED), to generate adversarial\nvoice examples to speaker identification. It contains two novel components.\nFirst, it uses a novel saliency map decoder to learn the importance of speech\nsamples to the decision of a targeted speaker identification system, so as to\nmake the attacker focus on generating artificial noise to the important\nsamples. It also proposes an angular loss function to push the speaker\nembedding far away from the source speaker. Our experimental results\ndemonstrate that the proposed SSED yields the state-of-the-art performance,\ni.e. over 97% targeted attack success rate and a signal-to-noise level of over\n39 dB on both the open-set and close-set speaker identification tasks, with a\nlow computational cost.\n", "date": "2022-10-30", "categories": []}
{"id": "2302.12959", "title": "Chaotic Variational Auto encoder-based Adversarial Machine Learning", "abstract": "  Machine Learning (ML) has become the new contrivance in almost every field.\nThis makes them a target of fraudsters by various adversary attacks, thereby\nhindering the performance of ML models. Evasion and Data-Poison-based attacks\nare well acclaimed, especially in finance, healthcare, etc. This motivated us\nto propose a novel computationally less expensive attack mechanism based on the\nadversarial sample generation by Variational Auto Encoder (VAE). It is well\nknown that Wavelet Neural Network (WNN) is considered computationally efficient\nin solving image and audio processing, speech recognition, and time-series\nforecasting. This paper proposed VAE-Deep-Wavelet Neural Network\n(VAE-Deep-WNN), where Encoder and Decoder employ WNN networks. Further, we\nproposed chaotic variants of both VAE with Multi-layer perceptron (MLP) and\nDeep-WNN and named them C-VAE-MLP and C-VAE-Deep-WNN, respectively. Here, we\nemployed a Logistic map to generate random noise in the latent space. In this\npaper, we performed VAE-based adversary sample generation and applied it to\nvarious problems related to finance and cybersecurity domain-related problems\nsuch as loan default, credit card fraud, and churn modelling, etc., We\nperformed both Evasion and Data-Poison attacks on Logistic Regression (LR) and\nDecision Tree (DT) models. The results indicated that VAE-Deep-WNN outperformed\nthe rest in the majority of the datasets and models. However, its chaotic\nvariant C-VAE-Deep-WNN performed almost similarly to VAE-Deep-WNN in the\nmajority of the datasets.\n", "date": "2023-02-25", "categories": []}
{"id": "2302.13464", "title": "Randomness in ML Defenses Helps Persistent Attackers and Hinders\n  Evaluators", "abstract": "  It is becoming increasingly imperative to design robust ML defenses. However,\nrecent work has found that many defenses that initially resist state-of-the-art\nattacks can be broken by an adaptive adversary. In this work we take steps to\nsimplify the design of defenses and argue that white-box defenses should eschew\nrandomness when possible. We begin by illustrating a new issue with the\ndeployment of randomized defenses that reduces their security compared to their\ndeterministic counterparts. We then provide evidence that making defenses\ndeterministic simplifies robustness evaluation, without reducing the\neffectiveness of a truly robust defense. Finally, we introduce a new defense\nevaluation framework that leverages a defense's deterministic nature to better\nevaluate its adversarial robustness.\n", "date": "2023-02-27", "categories": []}
{"id": "2302.13487", "title": "Contextual adversarial attack against aerial detection in the physical\n  world", "abstract": "  Deep Neural Networks (DNNs) have been extensively utilized in aerial\ndetection. However, DNNs' sensitivity and vulnerability to maliciously\nelaborated adversarial examples have progressively garnered attention.\nRecently, physical attacks have gradually become a hot issue due to they are\nmore practical in the real world, which poses great threats to some\nsecurity-critical applications. In this paper, we take the first attempt to\nperform physical attacks in contextual form against aerial detection in the\nphysical world. We propose an innovative contextual attack method against\naerial detection in real scenarios, which achieves powerful attack performance\nand transfers well between various aerial object detectors without smearing or\nblocking the interested objects to hide. Based on the findings that the\ntargets' contextual information plays an important role in aerial detection by\nobserving the detectors' attention maps, we propose to make full use of the\ncontextual area of the interested targets to elaborate contextual perturbations\nfor the uncovered attacks in real scenarios. Extensive proportionally scaled\nexperiments are conducted to evaluate the effectiveness of the proposed\ncontextual attack method, which demonstrates the proposed method's superiority\nin both attack efficacy and physical practicality.\n", "date": "2023-02-27", "categories": []}
{"id": "2302.13851", "title": "Implicit Poisoning Attacks in Two-Agent Reinforcement Learning:\n  Adversarial Policies for Training-Time Attacks", "abstract": "  In targeted poisoning attacks, an attacker manipulates an agent-environment\ninteraction to force the agent into adopting a policy of interest, called\ntarget policy. Prior work has primarily focused on attacks that modify standard\nMDP primitives, such as rewards or transitions. In this paper, we study\ntargeted poisoning attacks in a two-agent setting where an attacker implicitly\npoisons the effective environment of one of the agents by modifying the policy\nof its peer. We develop an optimization framework for designing optimal\nattacks, where the cost of the attack measures how much the solution deviates\nfrom the assumed default policy of the peer agent. We further study the\ncomputational properties of this optimization framework. Focusing on a tabular\nsetting, we show that in contrast to poisoning attacks based on MDP primitives\n(transitions and (unbounded) rewards), which are always feasible, it is NP-hard\nto determine the feasibility of implicit poisoning attacks. We provide\ncharacterization results that establish sufficient conditions for the\nfeasibility of the attack problem, as well as an upper and a lower bound on the\noptimal cost of the attack. We propose two algorithmic approaches for finding\nan optimal adversarial policy: a model-based approach with tabular policies and\na model-free approach with parametric/neural policies. We showcase the efficacy\nof the proposed algorithms through experiments.\n", "date": "2023-02-27", "categories": []}
{"id": "2208.13058", "title": "Adversarial Robustness for Tabular Data through Cost and Utility\n  Awareness", "abstract": "  Many safety-critical applications of machine learning, such as fraud or abuse\ndetection, use data in tabular domains. Adversarial examples can be\nparticularly damaging for these applications. Yet, existing works on\nadversarial robustness primarily focus on machine-learning models in image and\ntext domains. We argue that, due to the differences between tabular data and\nimages or text, existing threat models are not suitable for tabular domains.\nThese models do not capture that the costs of an attack could be more\nsignificant than imperceptibility, or that the adversary could assign different\nvalues to the utility obtained from deploying different adversarial examples.\nWe demonstrate that, due to these differences, the attack and defense methods\nused for images and text cannot be directly applied to tabular settings. We\naddress these issues by proposing new cost and utility-aware threat models that\nare tailored to the adversarial capabilities and constraints of attackers\ntargeting tabular domains. We introduce a framework that enables us to design\nattack and defense mechanisms that result in models protected against cost and\nutility-aware adversaries, for example, adversaries constrained by a certain\nfinancial budget. We show that our approach is effective on three datasets\ncorresponding to applications for which adversarial examples can have economic\nand social implications.\n", "date": "2022-08-27", "categories": []}
{"id": "2302.12351", "title": "On the Hardness of Robustness Transfer: A Perspective from Rademacher\n  Complexity over Symmetric Difference Hypothesis Space", "abstract": "  Recent studies demonstrated that the adversarially robust learning under\n$\\ell_\\infty$ attack is harder to generalize to different domains than standard\ndomain adaptation. How to transfer robustness across different domains has been\na key question in domain adaptation field. To investigate the fundamental\ndifficulty behind adversarially robust domain adaptation (or robustness\ntransfer), we propose to analyze a key complexity measure that controls the\ncross-domain generalization: the adversarial Rademacher complexity over {\\em\nsymmetric difference hypothesis space} $\\mathcal{H} \\Delta \\mathcal{H}$. For\nlinear models, we show that adversarial version of this complexity is always\ngreater than the non-adversarial one, which reveals the intrinsic hardness of\nadversarially robust domain adaptation. We also establish upper bounds on this\ncomplexity measure. Then we extend them to the ReLU neural network class by\nupper bounding the adversarial Rademacher complexity in the binary\nclassification setting. Finally, even though the robust domain adaptation is\nprovably harder, we do find positive relation between robust learning and\nstandard domain adaptation. We explain \\emph{how adversarial training helps\ndomain adaptation in terms of standard risk}. We believe our results initiate\nthe study of the generalization theory of adversarially robust domain\nadaptation, and could shed lights on distributed adversarially robust learning\nfrom heterogeneous sources, e.g., federated learning scenario.\n", "date": "2023-02-23", "categories": []}
{"id": "2302.12407", "title": "HyperAttack: Multi-Gradient-Guided White-box Adversarial Structure\n  Attack of Hypergraph Neural Networks", "abstract": "  Hypergraph neural networks (HGNN) have shown superior performance in various\ndeep learning tasks, leveraging the high-order representation ability to\nformulate complex correlations among data by connecting two or more nodes\nthrough hyperedge modeling. Despite the well-studied adversarial attacks on\nGraph Neural Networks (GNN), there is few study on adversarial attacks against\nHGNN, which leads to a threat to the safety of HGNN applications. In this\npaper, we introduce HyperAttack, the first white-box adversarial attack\nframework against hypergraph neural networks. HyperAttack conducts a white-box\nstructure attack by perturbing hyperedge link status towards the target node\nwith the guidance of both gradients and integrated gradients. We evaluate\nHyperAttack on the widely-used Cora and PubMed datasets and three hypergraph\nneural networks with typical hypergraph modeling techniques. Compared to\nstate-of-the-art white-box structural attack methods for GNN, HyperAttack\nachieves a 10-20X improvement in time efficiency while also increasing attack\nsuccess rates by 1.3%-3.7%. The results show that HyperAttack can achieve\nefficient adversarial attacks that balance effectiveness and time costs.\n", "date": "2023-02-24", "categories": []}
{"id": "2302.12758", "title": "Defending Against Backdoor Attacks by Layer-wise Feature Analysis", "abstract": "  Training deep neural networks (DNNs) usually requires massive training data\nand computational resources. Users who cannot afford this may prefer to\noutsource training to a third party or resort to publicly available pre-trained\nmodels. Unfortunately, doing so facilitates a new training-time attack (i.e.,\nbackdoor attack) against DNNs. This attack aims to induce misclassification of\ninput samples containing adversary-specified trigger patterns. In this paper,\nwe first conduct a layer-wise feature analysis of poisoned and benign samples\nfrom the target class. We find out that the feature difference between benign\nand poisoned samples tends to be maximum at a critical layer, which is not\nalways the one typically used in existing defenses, namely the layer before\nfully-connected layers. We also demonstrate how to locate this critical layer\nbased on the behaviors of benign samples. We then propose a simple yet\neffective method to filter poisoned samples by analyzing the feature\ndifferences between suspicious and benign samples at the critical layer. We\nconduct extensive experiments on two benchmark datasets, which confirm the\neffectiveness of our defense.\n", "date": "2023-02-24", "categories": []}
{"id": "2302.08764", "title": "Adversarial Contrastive Distillation with Adaptive Denoising", "abstract": "  Adversarial Robustness Distillation (ARD) is a novel method to boost the\nrobustness of small models. Unlike general adversarial training, its robust\nknowledge transfer can be less easily restricted by the model capacity.\nHowever, the teacher model that provides the robustness of knowledge does not\nalways make correct predictions, interfering with the student's robust\nperformances. Besides, in the previous ARD methods, the robustness comes\nentirely from one-to-one imitation, ignoring the relationship between examples.\nTo this end, we propose a novel structured ARD method called Contrastive\nRelationship DeNoise Distillation (CRDND). We design an adaptive compensation\nmodule to model the instability of the teacher. Moreover, we utilize the\ncontrastive relationship to explore implicit robustness knowledge among\nmultiple examples. Experimental results on multiple attack benchmarks show\nCRDND can transfer robust knowledge efficiently and achieves state-of-the-art\nperformances.\n", "date": "2023-02-17", "categories": []}
{"id": "2209.05980", "title": "Certified Defences Against Adversarial Patch Attacks on Semantic\n  Segmentation", "abstract": "  Adversarial patch attacks are an emerging security threat for real world deep\nlearning applications. We present Demasked Smoothing, the first approach (up to\nour knowledge) to certify the robustness of semantic segmentation models\nagainst this threat model. Previous work on certifiably defending against patch\nattacks has mostly focused on image classification task and often required\nchanges in the model architecture and additional training which is undesirable\nand computationally expensive. In Demasked Smoothing, any segmentation model\ncan be applied without particular training, fine-tuning, or restriction of the\narchitecture. Using different masking strategies, Demasked Smoothing can be\napplied both for certified detection and certified recovery. In extensive\nexperiments we show that Demasked Smoothing can on average certify 64% of the\npixel predictions for a 1% patch in the detection task and 48% against a 0.5%\npatch for the recovery task on the ADE20K dataset.\n", "date": "2022-09-13", "categories": []}
{"id": "2302.00537", "title": "Effectiveness of Moving Target Defenses for Adversarial Attacks in\n  ML-based Malware Detection", "abstract": "  Several moving target defenses (MTDs) to counter adversarial ML attacks have\nbeen proposed in recent years. MTDs claim to increase the difficulty for the\nattacker in conducting attacks by regularly changing certain elements of the\ndefense, such as cycling through configurations. To examine these claims, we\nstudy for the first time the effectiveness of several recent MTDs for\nadversarial ML attacks applied to the malware detection domain. Under different\nthreat models, we show that transferability and query attack strategies can\nachieve high levels of evasion against these defenses through existing and\nnovel attack strategies across Android and Windows. We also show that\nfingerprinting and reconnaissance are possible and demonstrate how attackers\nmay obtain critical defense hyperparameters as well as information about how\npredictions are produced. Based on our findings, we present key recommendations\nfor future work on the development of effective MTDs for adversarial attacks in\nML-based malware detection.\n", "date": "2023-02-01", "categories": []}
{"id": "2302.10287", "title": "CertViT: Certified Robustness of Pre-Trained Vision Transformers", "abstract": "  Lipschitz bounded neural networks are certifiably robust and have a good\ntrade-off between clean and certified accuracy. Existing Lipschitz bounding\nmethods train from scratch and are limited to moderately sized networks (< 6M\nparameters). They require a fair amount of hyper-parameter tuning and are\ncomputationally prohibitive for large networks like Vision Transformers (5M to\n660M parameters). Obtaining certified robustness of transformers is not\nfeasible due to the non-scalability and inflexibility of the current methods.\nThis work presents CertViT, a two-step proximal-projection method to achieve\ncertified robustness from pre-trained weights. The proximal step tries to lower\nthe Lipschitz bound and the projection step tries to maintain the clean\naccuracy of pre-trained weights. We show that CertViT networks have better\ncertified accuracy than state-of-the-art Lipschitz trained networks. We apply\nCertViT on several variants of pre-trained vision transformers and show\nadversarial robustness using standard attacks. Code :\nhttps://github.com/sagarverma/transformer-lipschitz\n", "date": "2023-02-01", "categories": []}
{"id": "2302.10633", "title": "Generalization Bounds for Adversarial Contrastive Learning", "abstract": "  Deep networks are well-known to be fragile to adversarial attacks, and\nadversarial training is one of the most popular methods used to train a robust\nmodel. To take advantage of unlabeled data, recent works have applied\nadversarial training to contrastive learning (Adversarial Contrastive Learning;\nACL for short) and obtain promising robust performance. However, the theory of\nACL is not well understood. To fill this gap, we leverage the Rademacher\ncomplexity to analyze the generalization performance of ACL, with a particular\nfocus on linear models and multi-layer neural networks under $\\ell_p$ attack\n($p \\ge 1$). Our theory shows that the average adversarial risk of the\ndownstream tasks can be upper bounded by the adversarial unsupervised risk of\nthe upstream task. The experimental results validate our theory.\n", "date": "2023-02-21", "categories": []}
{"id": "2302.10686", "title": "Interpretable Spectrum Transformation Attacks to Speaker Recognition", "abstract": "  The success of adversarial attacks to speaker recognition is mainly in\nwhite-box scenarios. When applying the adversarial voices that are generated by\nattacking white-box surrogate models to black-box victim models, i.e.\n\\textit{transfer-based} black-box attacks, the transferability of the\nadversarial voices is not only far from satisfactory, but also lacks\ninterpretable basis. To address these issues, in this paper, we propose a\ngeneral framework, named spectral transformation attack based on modified\ndiscrete cosine transform (STA-MDCT), to improve the transferability of the\nadversarial voices to a black-box victim model. Specifically, we first apply\nMDCT to the input voice. Then, we slightly modify the energy of different\nfrequency bands for capturing the salient regions of the adversarial noise in\nthe time-frequency domain that are critical to a successful attack. Unlike\nexisting approaches that operate voices in the time domain, the proposed\nframework operates voices in the time-frequency domain, which improves the\ninterpretability, transferability, and imperceptibility of the attack.\nMoreover, it can be implemented with any gradient-based attackers. To utilize\nthe advantage of model ensembling, we not only implement STA-MDCT with a single\nwhite-box surrogate model, but also with an ensemble of surrogate models.\nFinally, we visualize the saliency maps of adversarial voices by the class\nactivation maps (CAM), which offers an interpretable basis to transfer-based\nattacks in speaker recognition for the first time. Extensive comparison results\nwith five representative attackers show that the CAM visualization clearly\nexplains the effectiveness of STA-MDCT, and the weaknesses of the comparison\nmethods; the proposed method outperforms the comparison methods by a large\nmargin.\n", "date": "2023-02-21", "categories": []}
{"id": "2212.01957", "title": "CSTAR: Towards Compact and STructured Deep Neural Networks with\n  Adversarial Robustness", "abstract": "  Model compression and model defense for deep neural networks (DNNs) have been\nextensively and individually studied. Considering the co-importance of model\ncompactness and robustness in practical applications, several prior works have\nexplored to improve the adversarial robustness of the sparse neural networks.\nHowever, the structured sparse models obtained by the exiting works suffer\nsevere performance degradation for both benign and robust accuracy, thereby\ncausing a challenging dilemma between robustness and structuredness of the\ncompact DNNs. To address this problem, in this paper, we propose CSTAR, an\nefficient solution that can simultaneously impose the low-rankness-based\nCompactness, high STructuredness and high Adversarial Robustness on the target\nDNN models. By formulating the low-rankness and robustness requirement within\nthe same framework and globally determining the ranks, the compressed DNNs can\nsimultaneously achieve high compression performance and strong adversarial\nrobustness. Evaluations for various DNN models on different datasets\ndemonstrate the effectiveness of CSTAR. Compared with the state-of-the-art\nrobust structured pruning methods, CSTAR shows consistently better performance.\nFor instance, when compressing ResNet-18 on CIFAR-10, CSTAR can achieve up to\n20.07% and 11.91% improvement for benign accuracy and robust accuracy,\nrespectively. For compressing ResNet-18 with 16x compression ratio on Imagenet,\nCSTAR can obtain 8.58% benign accuracy gain and 4.27% robust accuracy gain\ncompared to the existing robust structured pruning method.\n", "date": "2022-12-04", "categories": []}
{"id": "2302.09479", "title": "Delving into the Adversarial Robustness of Federated Learning", "abstract": "  In Federated Learning (FL), models are as fragile as centrally trained models\nagainst adversarial examples. However, the adversarial robustness of federated\nlearning remains largely unexplored. This paper casts light on the challenge of\nadversarial robustness of federated learning. To facilitate a better\nunderstanding of the adversarial vulnerability of the existing FL methods, we\nconduct comprehensive robustness evaluations on various attacks and adversarial\ntraining methods. Moreover, we reveal the negative impacts induced by directly\nadopting adversarial training in FL, which seriously hurts the test accuracy,\nespecially in non-IID settings. In this work, we propose a novel algorithm\ncalled Decision Boundary based Federated Adversarial Training (DBFAT), which\nconsists of two components (local re-weighting and global regularization) to\nimprove both accuracy and robustness of FL systems. Extensive experiments on\nmultiple datasets demonstrate that DBFAT consistently outperforms other\nbaselines under both IID and non-IID settings.\n", "date": "2023-02-19", "categories": []}
{"id": "2302.09491", "title": "X-Adv: Physical Adversarial Object Attacks against X-ray Prohibited Item\n  Detection", "abstract": "  Adversarial attacks are valuable for evaluating the robustness of deep\nlearning models. Existing attacks are primarily conducted on the visible light\nspectrum (e.g., pixel-wise texture perturbation). However, attacks targeting\ntexture-free X-ray images remain underexplored, despite the widespread\napplication of X-ray imaging in safety-critical scenarios such as the X-ray\ndetection of prohibited items. In this paper, we take the first step toward the\nstudy of adversarial attacks targeted at X-ray prohibited item detection, and\nreveal the serious threats posed by such attacks in this safety-critical\nscenario. Specifically, we posit that successful physical adversarial attacks\nin this scenario should be specially designed to circumvent the challenges\nposed by color/texture fading and complex overlapping. To this end, we propose\nX-adv to generate physically printable metals that act as an adversarial agent\ncapable of deceiving X-ray detectors when placed in luggage. To resolve the\nissues associated with color/texture fading, we develop a differentiable\nconverter that facilitates the generation of 3D-printable objects with\nadversarial shapes, using the gradients of a surrogate model rather than\ndirectly generating adversarial textures. To place the printed 3D adversarial\nobjects in luggage with complex overlapped instances, we design a policy-based\nreinforcement learning strategy to find locations eliciting strong attack\nperformance in worst-case scenarios whereby the prohibited items are heavily\noccluded by other items. To verify the effectiveness of the proposed X-Adv, we\nconduct extensive experiments in both the digital and the physical world\n(employing a commercial X-ray security inspection system for the latter case).\nFurthermore, we present the physical-world X-ray adversarial attack dataset\nXAD.\n", "date": "2023-02-19", "categories": []}
{"id": "2302.09575", "title": "Stationary Point Losses for Robust Model", "abstract": "  The inability to guarantee robustness is one of the major obstacles to the\napplication of deep learning models in security-demanding domains. We identify\nthat the most commonly used cross-entropy (CE) loss does not guarantee robust\nboundary for neural networks. CE loss sharpens the neural network at the\ndecision boundary to achieve a lower loss, rather than pushing the boundary to\na more robust position. A robust boundary should be kept in the middle of\nsamples from different classes, thus maximizing the margins from the boundary\nto the samples. We think this is due to the fact that CE loss has no stationary\npoint. In this paper, we propose a family of new losses, called stationary\npoint (SP) loss, which has at least one stationary point on the correct\nclassification side. We proved that robust boundary can be guaranteed by SP\nloss without losing much accuracy. With SP loss, larger perturbations are\nrequired to generate adversarial examples. We demonstrate that robustness is\nimproved under a variety of adversarial attacks by applying SP loss. Moreover,\nrobust boundary learned by SP loss also performs well on imbalanced datasets.\n", "date": "2023-02-19", "categories": []}
{"id": "2302.10164", "title": "Seasoning Model Soups for Robustness to Adversarial and Natural\n  Distribution Shifts", "abstract": "  Adversarial training is widely used to make classifiers robust to a specific\nthreat or adversary, such as $\\ell_p$-norm bounded perturbations of a given\n$p$-norm. However, existing methods for training classifiers robust to multiple\nthreats require knowledge of all attacks during training and remain vulnerable\nto unseen distribution shifts. In this work, we describe how to obtain\nadversarially-robust model soups (i.e., linear combinations of parameters) that\nsmoothly trade-off robustness to different $\\ell_p$-norm bounded adversaries.\nWe demonstrate that such soups allow us to control the type and level of\nrobustness, and can achieve robustness to all threats without jointly training\non all of them. In some cases, the resulting model soups are more robust to a\ngiven $\\ell_p$-norm adversary than the constituent model specialized against\nthat same adversary. Finally, we show that adversarially-robust model soups can\nbe a viable tool to adapt to distribution shifts from a few examples.\n", "date": "2023-02-20", "categories": []}
{"id": "2302.08872", "title": "Revisiting adversarial training for the worst-performing class", "abstract": "  Despite progress in adversarial training (AT), there is a substantial gap\nbetween the top-performing and worst-performing classes in many datasets. For\nexample, on CIFAR10, the accuracies for the best and worst classes are 74% and\n23%, respectively. We argue that this gap can be reduced by explicitly\noptimizing for the worst-performing class, resulting in a min-max-max\noptimization formulation. Our method, called class focused online learning\n(CFOL), includes high probability convergence guarantees for the worst class\nloss and can be easily integrated into existing training setups with minimal\ncomputational overhead. We demonstrate an improvement to 32% in the worst class\naccuracy on CIFAR10, and we observe consistent behavior across CIFAR100 and\nSTL10. Our study highlights the importance of moving beyond average accuracy,\nwhich is particularly important in safety-critical applications.\n", "date": "2023-02-17", "categories": []}
{"id": "2302.08048", "title": "Robust Mid-Pass Filtering Graph Convolutional Networks", "abstract": "  Graph convolutional networks (GCNs) are currently the most promising paradigm\nfor dealing with graph-structure data, while recent studies have also shown\nthat GCNs are vulnerable to adversarial attacks. Thus developing GCN models\nthat are robust to such attacks become a hot research topic. However, the\nstructural purification learning-based or robustness constraints-based defense\nGCN methods are usually designed for specific data or attacks, and introduce\nadditional objective that is not for classification. Extra training overhead is\nalso required in their design. To address these challenges, we conduct in-depth\nexplorations on mid-frequency signals on graphs and propose a simple yet\neffective Mid-pass filter GCN (Mid-GCN). Theoretical analyses guarantee the\nrobustness of signals through the mid-pass filter, and we also shed light on\nthe properties of different frequency signals under adversarial attacks.\nExtensive experiments on six benchmark graph data further verify the\neffectiveness of our designed Mid-GCN in node classification accuracy compared\nto state-of-the-art GCNs under various adversarial attack strategies.\n", "date": "2023-02-16", "categories": []}
{"id": "2302.08066", "title": "Masking and Mixing Adversarial Training", "abstract": "  While convolutional neural networks (CNNs) have achieved excellent\nperformances in various computer vision tasks, they often misclassify with\nmalicious samples, a.k.a. adversarial examples. Adversarial training is a\npopular and straightforward technique to defend against the threat of\nadversarial examples. Unfortunately, CNNs must sacrifice the accuracy of\nstandard samples to improve robustness against adversarial examples when\nadversarial training is used. In this work, we propose Masking and Mixing\nAdversarial Training (M2AT) to mitigate the trade-off between accuracy and\nrobustness. We focus on creating diverse adversarial examples during training.\nSpecifically, our approach consists of two processes: 1) masking a perturbation\nwith a binary mask and 2) mixing two partially perturbed images. Experimental\nresults on CIFAR-10 dataset demonstrate that our method achieves better\nrobustness against several adversarial attacks than previous methods.\n", "date": "2023-02-16", "categories": []}
{"id": "2302.08090", "title": "QTrojan: A Circuit Backdoor Against Quantum Neural Networks", "abstract": "  We propose a circuit-level backdoor attack, \\textit{QTrojan}, against Quantum\nNeural Networks (QNNs) in this paper. QTrojan is implemented by few quantum\ngates inserted into the variational quantum circuit of the victim QNN. QTrojan\nis much stealthier than a prior Data-Poisoning-based Backdoor Attack (DPBA),\nsince it does not embed any trigger in the inputs of the victim QNN or require\nthe access to original training datasets. Compared to a DPBA, QTrojan improves\nthe clean data accuracy by 21\\% and the attack success rate by 19.9\\%.\n", "date": "2023-02-16", "categories": []}
{"id": "2302.05703", "title": "HateProof: Are Hateful Meme Detection Systems really Robust?", "abstract": "  Exploiting social media to spread hate has tremendously increased over the\nyears. Lately, multi-modal hateful content such as memes has drawn relatively\nmore traction than uni-modal content. Moreover, the availability of implicit\ncontent payloads makes them fairly challenging to be detected by existing\nhateful meme detection systems. In this paper, we present a use case study to\nanalyze such systems' vulnerabilities against external adversarial attacks. We\nfind that even very simple perturbations in uni-modal and multi-modal settings\nperformed by humans with little knowledge about the model can make the existing\ndetection models highly vulnerable. Empirically, we find a noticeable\nperformance drop of as high as 10% in the macro-F1 score for certain attacks.\nAs a remedy, we attempt to boost the model's robustness using contrastive\nlearning as well as an adversarial training-based method - VILLA. Using an\nensemble of the above two approaches, in two of our high resolution datasets,\nwe are able to (re)gain back the performance to a large extent for certain\nattacks. We believe that ours is a first step toward addressing this crucial\nproblem in an adversarial setting and would inspire more such investigations in\nthe future.\n", "date": "2023-02-11", "categories": []}
{"id": "2302.05794", "title": "Mutation-Based Adversarial Attacks on Neural Text Detectors", "abstract": "  Neural text detectors aim to decide the characteristics that distinguish\nneural (machine-generated) from human texts. To challenge such detectors,\nadversarial attacks can alter the statistical characteristics of the generated\ntext, making the detection task more and more difficult. Inspired by the\nadvances of mutation analysis in software development and testing, in this\npaper, we propose character- and word-based mutation operators for generating\nadversarial samples to attack state-of-the-art natural text detectors. This\nfalls under white-box adversarial attacks. In such attacks, attackers have\naccess to the original text and create mutation instances based on this\noriginal text. The ultimate goal is to confuse machine learning models and\nclassifiers and decrease their prediction accuracy.\n", "date": "2023-02-11", "categories": []}
{"id": "2302.05892", "title": "TextDefense: Adversarial Text Detection based on Word Importance Entropy", "abstract": "  Currently, natural language processing (NLP) models are wildly used in\nvarious scenarios. However, NLP models, like all deep models, are vulnerable to\nadversarially generated text. Numerous works have been working on mitigating\nthe vulnerability from adversarial attacks. Nevertheless, there is no\ncomprehensive defense in existing works where each work targets a specific\nattack category or suffers from the limitation of computation overhead,\nirresistible to adaptive attack, etc.\n  In this paper, we exhaustively investigate the adversarial attack algorithms\nin NLP, and our empirical studies have discovered that the attack algorithms\nmainly disrupt the importance distribution of words in a text. A well-trained\nmodel can distinguish subtle importance distribution differences between clean\nand adversarial texts. Based on this intuition, we propose TextDefense, a new\nadversarial example detection framework that utilizes the target model's\ncapability to defend against adversarial attacks while requiring no prior\nknowledge. TextDefense differs from previous approaches, where it utilizes the\ntarget model for detection and thus is attack type agnostic. Our extensive\nexperiments show that TextDefense can be applied to different architectures,\ndatasets, and attack methods and outperforms existing methods. We also discover\nthat the leading factor influencing the performance of TextDefense is the\ntarget model's generalizability. By analyzing the property of the target model\nand the property of the adversarial example, we provide our insights into the\nadversarial attacks in NLP and the principles of our defense method.\n", "date": "2023-02-12", "categories": []}
{"id": "2302.06588", "title": "Raising the Cost of Malicious AI-Powered Image Editing", "abstract": "  We present an approach to mitigating the risks of malicious image editing\nposed by large diffusion models. The key idea is to immunize images so as to\nmake them resistant to manipulation by these models. This immunization relies\non injection of imperceptible adversarial perturbations designed to disrupt the\noperation of the targeted diffusion models, forcing them to generate\nunrealistic images. We provide two methods for crafting such perturbations, and\nthen demonstrate their efficacy. Finally, we discuss a policy component\nnecessary to make our approach fully effective and practical -- one that\ninvolves the organizations developing diffusion models, rather than individual\nusers, to implement (and support) the immunization process.\n", "date": "2023-02-13", "categories": []}
{"id": "2302.05120", "title": "Step by Step Loss Goes Very Far: Multi-Step Quantization for Adversarial\n  Text Attacks", "abstract": "  We propose a novel gradient-based attack against transformer-based language\nmodels that searches for an adversarial example in a continuous space of token\nprobabilities. Our algorithm mitigates the gap between adversarial loss for\ncontinuous and discrete text representations by performing multi-step\nquantization in a quantization-compensation loop. Experiments show that our\nmethod significantly outperforms other approaches on various natural language\nprocessing (NLP) tasks.\n", "date": "2023-02-10", "categories": []}
{"id": "2302.04025", "title": "WAT: Improve the Worst-class Robustness in Adversarial Training", "abstract": "  Deep Neural Networks (DNN) have been shown to be vulnerable to adversarial\nexamples. Adversarial training (AT) is a popular and effective strategy to\ndefend against adversarial attacks. Recent works (Benz et al., 2020; Xu et al.,\n2021; Tian et al., 2021) have shown that a robust model well-trained by AT\nexhibits a remarkable robustness disparity among classes, and propose various\nmethods to obtain consistent robust accuracy across classes. Unfortunately,\nthese methods sacrifice a good deal of the average robust accuracy.\nAccordingly, this paper proposes a novel framework of worst-class adversarial\ntraining and leverages no-regret dynamics to solve this problem. Our goal is to\nobtain a classifier with great performance on worst-class and sacrifice just a\nlittle average robust accuracy at the same time. We then rigorously analyze the\ntheoretical properties of our proposed algorithm, and the generalization error\nbound in terms of the worst-class robust risk. Furthermore, we propose a\nmeasurement to evaluate the proposed method in terms of both the average and\nworst-class accuracies. Experiments on various datasets and networks show that\nour proposed method outperforms the state-of-the-art approaches.\n", "date": "2023-02-08", "categories": []}
{"id": "2302.04109", "title": "Explainable Label-flipping Attacks on Human Emotion Assessment System", "abstract": "  This paper's main goal is to provide an attacker's point of view on data\npoisoning assaults that use label-flipping during the training phase of systems\nthat use electroencephalogram (EEG) signals to evaluate human emotion. To\nattack different machine learning classifiers such as Adaptive Boosting\n(AdaBoost) and Random Forest dedicated to the classification of 4 different\nhuman emotions using EEG signals, this paper proposes two scenarios of\nlabel-flipping methods. The results of the studies show that the proposed data\npoison attacksm based on label-flipping are successful regardless of the model,\nbut different models show different degrees of resistance to the assaults. In\naddition, numerous Explainable Artificial Intelligence (XAI) techniques are\nused to explain the data poison attacks on EEG signal-based human emotion\nevaluation systems.\n", "date": "2023-02-08", "categories": []}
{"id": "2210.16765", "title": "Benchmarking Adversarial Patch Against Aerial Detection", "abstract": "  DNNs are vulnerable to adversarial examples, which poses great security\nconcerns for security-critical systems. In this paper, a novel\nadaptive-patch-based physical attack (AP-PA) framework is proposed, which aims\nto generate adversarial patches that are adaptive in both physical dynamics and\nvarying scales, and by which the particular targets can be hidden from being\ndetected. Furthermore, the adversarial patch is also gifted with attack\neffectiveness against all targets of the same class with a patch outside the\ntarget (No need to smear targeted objects) and robust enough in the physical\nworld. In addition, a new loss is devised to consider more available\ninformation of detected objects to optimize the adversarial patch, which can\nsignificantly improve the patch's attack efficacy (Average precision drop up to\n87.86% and 85.48% in white-box and black-box settings, respectively) and\noptimizing efficiency. We also establish one of the first comprehensive,\ncoherent, and rigorous benchmarks to evaluate the attack efficacy of\nadversarial patches on aerial detection tasks. Finally, several proportionally\nscaled experiments are performed physically to demonstrate that the elaborated\nadversarial patches can successfully deceive aerial detection algorithms in\ndynamic physical circumstances. The code is available at\nhttps://github.com/JiaweiLian/AP-PA.\n", "date": "2022-10-30", "categories": []}
{"id": "2302.03523", "title": "Sparse Mixture Once-for-all Adversarial Training for Efficient In-Situ\n  Trade-Off Between Accuracy and Robustness of DNNs", "abstract": "  Existing deep neural networks (DNNs) that achieve state-of-the-art (SOTA)\nperformance on both clean and adversarially-perturbed images rely on either\nactivation or weight conditioned convolution operations. However, such\nconditional learning costs additional multiply-accumulate (MAC) or addition\noperations, increasing inference memory and compute costs. To that end, we\npresent a sparse mixture once for all adversarial training (SMART), that allows\na model to train once and then in-situ trade-off between accuracy and\nrobustness, that too at a reduced compute and parameter overhead. In\nparticular, SMART develops two expert paths, for clean and adversarial images,\nrespectively, that are then conditionally trained via respective dedicated sets\nof binary sparsity masks. Extensive evaluations on multiple image\nclassification datasets across different models show SMART to have up to 2.72x\nfewer non-zero parameters costing proportional reduction in compute overhead,\nwhile yielding SOTA accuracy-robustness trade-off. Additionally, we present\ninsightful observations in designing sparse masks to successfully condition on\nboth clean and perturbed images.\n", "date": "2022-12-27", "categories": []}
{"id": "2302.03657", "title": "Toward Face Biometric De-identification using Adversarial Examples", "abstract": "  The remarkable success of face recognition (FR) has endangered the privacy of\ninternet users particularly in social media. Recently, researchers turned to\nuse adversarial examples as a countermeasure. In this paper, we assess the\neffectiveness of using two widely known adversarial methods (BIM and ILLC) for\nde-identifying personal images. We discovered, unlike previous claims in the\nliterature, that it is not easy to get a high protection success rate\n(suppressing identification rate) with imperceptible adversarial perturbation\nto the human visual system. Finally, we found out that the transferability of\nadversarial examples is highly affected by the training parameters of the\nnetwork with which they are generated.\n", "date": "2023-02-07", "categories": []}
{"id": "2209.05957", "title": "Adversarial Inter-Group Link Injection Degrades the Fairness of Graph\n  Neural Networks", "abstract": "  We present evidence for the existence and effectiveness of adversarial\nattacks on graph neural networks (GNNs) that aim to degrade fairness. These\nattacks can disadvantage a particular subgroup of nodes in GNN-based node\nclassification, where nodes of the underlying network have sensitive\nattributes, such as race or gender. We conduct qualitative and experimental\nanalyses explaining how adversarial link injection impairs the fairness of GNN\npredictions. For example, an attacker can compromise the fairness of GNN-based\nnode classification by injecting adversarial links between nodes belonging to\nopposite subgroups and opposite class labels. Our experiments on empirical\ndatasets demonstrate that adversarial fairness attacks can significantly\ndegrade the fairness of GNN predictions (attacks are effective) with a low\nperturbation rate (attacks are efficient) and without a significant drop in\naccuracy (attacks are deceptive). This work demonstrates the vulnerability of\nGNN models to adversarial fairness attacks. We hope our findings raise\nawareness about this issue in our community and lay a foundation for the future\ndevelopment of GNN models that are more robust to such attacks.\n", "date": "2022-09-13", "categories": []}
{"id": "2211.07915", "title": "Backdoor Attacks on Time Series: A Generative Approach", "abstract": "  Backdoor attacks have emerged as one of the major security threats to deep\nlearning models as they can easily control the model's test-time predictions by\npre-injecting a backdoor trigger into the model at training time. While\nbackdoor attacks have been extensively studied on images, few works have\ninvestigated the threat of backdoor attacks on time series data. To fill this\ngap, in this paper we present a novel generative approach for time series\nbackdoor attacks against deep learning based time series classifiers. Backdoor\nattacks have two main goals: high stealthiness and high attack success rate. We\nfind that, compared to images, it can be more challenging to achieve the two\ngoals on time series. This is because time series have fewer input dimensions\nand lower degrees of freedom, making it hard to achieve a high attack success\nrate without compromising stealthiness. Our generative approach addresses this\nchallenge by generating trigger patterns that are as realistic as real-time\nseries patterns while achieving a high attack success rate without causing a\nsignificant drop in clean accuracy. We also show that our proposed attack is\nresistant to potential backdoor defenses. Furthermore, we propose a novel\nuniversal generator that can poison any type of time series with a single\ngenerator that allows universal attacks without the need to fine-tune the\ngenerative model for new time series datasets.\n", "date": "2022-11-15", "categories": []}
{"id": "2301.12576", "title": "Uncovering Adversarial Risks of Test-Time Adaptation", "abstract": "  Recently, test-time adaptation (TTA) has been proposed as a promising\nsolution for addressing distribution shifts. It allows a base model to adapt to\nan unforeseen distribution during inference by leveraging the information from\nthe batch of (unlabeled) test data. However, we uncover a novel security\nvulnerability of TTA based on the insight that predictions on benign samples\ncan be impacted by malicious samples in the same batch. To exploit this\nvulnerability, we propose Distribution Invading Attack (DIA), which injects a\nsmall fraction of malicious data into the test batch. DIA causes models using\nTTA to misclassify benign and unperturbed test data, providing an entirely new\ncapability for adversaries that is infeasible in canonical machine learning\npipelines. Through comprehensive evaluations, we demonstrate the high\neffectiveness of our attack on multiple benchmarks across six TTA methods. In\nresponse, we investigate two countermeasures to robustify the existing insecure\nTTA implementations, following the principle of \"security by design\". Together,\nwe hope our findings can make the community aware of the utility-security\ntradeoffs in deploying TTA and provide valuable insights for developing robust\nTTA approaches.\n", "date": "2023-01-29", "categories": []}
{"id": "2302.02023", "title": "TextShield: Beyond Successfully Detecting Adversarial Sentences in Text\n  Classification", "abstract": "  Adversarial attack serves as a major challenge for neural network models in\nNLP, which precludes the model's deployment in safety-critical applications. A\nrecent line of work, detection-based defense, aims to distinguish adversarial\nsentences from benign ones. However, {the core limitation of previous detection\nmethods is being incapable of giving correct predictions on adversarial\nsentences unlike defense methods from other paradigms.} To solve this issue,\nthis paper proposes TextShield: (1) we discover a link between text attack and\nsaliency information, and then we propose a saliency-based detector, which can\neffectively detect whether an input sentence is adversarial or not. (2) We\ndesign a saliency-based corrector, which converts the detected adversary\nsentences to benign ones. By combining the saliency-based detector and\ncorrector, TextShield extends the detection-only paradigm to a\ndetection-correction paradigm, thus filling the gap in the existing\ndetection-based defense. Comprehensive experiments show that (a) TextShield\nconsistently achieves higher or comparable performance than state-of-the-art\ndefense methods across various attacks on different benchmarks. (b) our\nsaliency-based detector outperforms existing detectors for detecting\nadversarial sentences.\n", "date": "2023-02-03", "categories": []}
{"id": "2302.02216", "title": "A Minimax Approach Against Multi-Armed Adversarial Attacks Detection", "abstract": "  Multi-armed adversarial attacks, in which multiple algorithms and objective\nloss functions are simultaneously used at evaluation time, have been shown to\nbe highly successful in fooling state-of-the-art adversarial examples detectors\nwhile requiring no specific side information about the detection mechanism. By\nformalizing the problem at hand, we can propose a solution that aggregates the\nsoft-probability outputs of multiple pre-trained detectors according to a\nminimax approach. The proposed framework is mathematically sound, easy to\nimplement, and modular, allowing for integrating existing or future detectors.\nThrough extensive evaluation on popular datasets (e.g., CIFAR10 and SVHN), we\nshow that our aggregation consistently outperforms individual state-of-the-art\ndetectors against multi-armed adversarial attacks, making it an effective\nsolution to improve the resilience of available methods.\n", "date": "2023-02-04", "categories": []}
{"id": "2302.02829", "title": "Collective Robustness Certificates: Exploiting Interdependence in Graph\n  Neural Networks", "abstract": "  In tasks like node classification, image segmentation, and named-entity\nrecognition we have a classifier that simultaneously outputs multiple\npredictions (a vector of labels) based on a single input, i.e. a single graph,\nimage, or document respectively. Existing adversarial robustness certificates\nconsider each prediction independently and are thus overly pessimistic for such\ntasks. They implicitly assume that an adversary can use different perturbed\ninputs to attack different predictions, ignoring the fact that we have a single\nshared input. We propose the first collective robustness certificate which\ncomputes the number of predictions that are simultaneously guaranteed to remain\nstable under perturbation, i.e. cannot be attacked. We focus on Graph Neural\nNetworks and leverage their locality property - perturbations only affect the\npredictions in a close neighborhood - to fuse multiple single-node certificates\ninto a drastically stronger collective certificate. For example, on the\nCiteseer dataset our collective certificate for node classification increases\nthe average number of certifiable feature perturbations from $7$ to $351$.\n", "date": "2023-02-06", "categories": []}
{"id": "2302.01762", "title": "BackdoorBox: A Python Toolbox for Backdoor Learning", "abstract": "  Third-party resources ($e.g.$, samples, backbones, and pre-trained models)\nare usually involved in the training of deep neural networks (DNNs), which\nbrings backdoor attacks as a new training-phase threat. In general, backdoor\nattackers intend to implant hidden backdoor in DNNs, so that the attacked DNNs\nbehave normally on benign samples whereas their predictions will be maliciously\nchanged to a pre-defined target label if hidden backdoors are activated by\nattacker-specified trigger patterns. To facilitate the research and development\nof more secure training schemes and defenses, we design an open-sourced Python\ntoolbox that implements representative and advanced backdoor attacks and\ndefenses under a unified and flexible framework. Our toolbox has four important\nand promising characteristics, including consistency, simplicity, flexibility,\nand co-development. It allows researchers and developers to easily implement\nand compare different methods on benchmark or their local datasets. This Python\ntoolbox, namely \\texttt{BackdoorBox}, is available at\n\\url{https://github.com/THUYimingLi/BackdoorBox}.\n", "date": "2023-02-01", "categories": []}
{"id": "2207.07032", "title": "Adversarial Attacks on Monocular Pose Estimation", "abstract": "  Advances in deep learning have resulted in steady progress in computer vision\nwith improved accuracy on tasks such as object detection and semantic\nsegmentation. Nevertheless, deep neural networks are vulnerable to adversarial\nattacks, thus presenting a challenge in reliable deployment. Two of the\nprominent tasks in 3D scene-understanding for robotics and advanced drive\nassistance systems are monocular depth and pose estimation, often learned\ntogether in an unsupervised manner. While studies evaluating the impact of\nadversarial attacks on monocular depth estimation exist, a systematic\ndemonstration and analysis of adversarial perturbations against pose estimation\nare lacking. We show how additive imperceptible perturbations can not only\nchange predictions to increase the trajectory drift but also catastrophically\nalter its geometry. We also study the relation between adversarial\nperturbations targeting monocular depth and pose estimation networks, as well\nas the transferability of perturbations to other networks with different\narchitectures and losses. Our experiments show how the generated perturbations\nlead to notable errors in relative rotation and translation predictions and\nelucidate vulnerabilities of the networks.\n", "date": "2022-07-14", "categories": []}
{"id": "2209.07399", "title": "A Light Recipe to Train Robust Vision Transformers", "abstract": "  In this paper, we ask whether Vision Transformers (ViTs) can serve as an\nunderlying architecture for improving the adversarial robustness of machine\nlearning models against evasion attacks. While earlier works have focused on\nimproving Convolutional Neural Networks, we show that also ViTs are highly\nsuitable for adversarial training to achieve competitive performance. We\nachieve this objective using a custom adversarial training recipe, discovered\nusing rigorous ablation studies on a subset of the ImageNet dataset. The\ncanonical training recipe for ViTs recommends strong data augmentation, in part\nto compensate for the lack of vision inductive bias of attention modules, when\ncompared to convolutions. We show that this recipe achieves suboptimal\nperformance when used for adversarial training. In contrast, we find that\nomitting all heavy data augmentation, and adding some additional bag-of-tricks\n($\\varepsilon$-warmup and larger weight decay), significantly boosts the\nperformance of robust ViTs. We show that our recipe generalizes to different\nclasses of ViT architectures and large-scale models on full ImageNet-1k.\nAdditionally, investigating the reasons for the robustness of our models, we\nshow that it is easier to generate strong attacks during training when using\nour recipe and that this leads to better robustness at test time. Finally, we\nfurther study one consequence of adversarial training by proposing a way to\nquantify the semantic nature of adversarial perturbations and highlight its\ncorrelation with the robustness of the model. Overall, we recommend that the\ncommunity should avoid translating the canonical training recipes in ViTs to\nrobust training and rethink common training choices in the context of\nadversarial training.\n", "date": "2022-09-15", "categories": []}
{"id": "2301.11546", "title": "Adapting Step-size: A Unified Perspective to Analyze and Improve\n  Gradient-based Methods for Adversarial Attacks", "abstract": "  Learning adversarial examples can be formulated as an optimization problem of\nmaximizing the loss function with some box-constraints. However, for solving\nthis induced optimization problem, the state-of-the-art gradient-based methods\nsuch as FGSM, I-FGSM and MI-FGSM look different from their original methods\nespecially in updating the direction, which makes it difficult to understand\nthem and then leaves some theoretical issues to be addressed in viewpoint of\noptimization. In this paper, from the perspective of adapting step-size, we\nprovide a unified theoretical interpretation of these gradient-based\nadversarial learning methods. We show that each of these algorithms is in fact\na specific reformulation of their original gradient methods but using the\nstep-size rules with only current gradient information. Motivated by such\nanalysis, we present a broad class of adaptive gradient-based algorithms based\non the regular gradient methods, in which the step-size strategy utilizing\ninformation of the accumulated gradients is integrated. Such adaptive step-size\nstrategies directly normalize the scale of the gradients rather than use some\nempirical operations. The important benefit is that convergence for the\niterative algorithms is guaranteed and then the whole optimization process can\nbe stabilized. The experiments demonstrate that our AdaI-FGM consistently\noutperforms I-FGSM and AdaMI-FGM remains competitive with MI-FGSM for black-box\nattacks.\n", "date": "2023-01-27", "categories": []}
{"id": "2301.13869", "title": "Reverse engineering adversarial attacks with fingerprints from\n  adversarial examples", "abstract": "  In spite of intense research efforts, deep neural networks remain vulnerable\nto adversarial examples: an input that forces the network to confidently\nproduce incorrect outputs. Adversarial examples are typically generated by an\nattack algorithm that optimizes a perturbation added to a benign input. Many\nsuch algorithms have been developed. If it were possible to reverse engineer\nattack algorithms from adversarial examples, this could deter bad actors\nbecause of the possibility of attribution. Here we formulate reverse\nengineering as a supervised learning problem where the goal is to assign an\nadversarial example to a class that represents the algorithm and parameters\nused. To our knowledge it has not been previously shown whether this is even\npossible. We first test whether we can classify the perturbations added to\nimages by attacks on undefended single-label image classification models.\nTaking a \"fight fire with fire\" approach, we leverage the sensitivity of deep\nneural networks to adversarial examples, training them to classify these\nperturbations. On a 17-class dataset (5 attacks, 4 bounded with 4 epsilon\nvalues each), we achieve an accuracy of 99.4% with a ResNet50 model trained on\nthe perturbations. We then ask whether we can perform this task without access\nto the perturbations, obtaining an estimate of them with signal processing\nalgorithms, an approach we call \"fingerprinting\". We find the JPEG algorithm\nserves as a simple yet effective fingerprinter (85.05% accuracy), providing a\nstrong baseline for future work. We discuss how our approach can be extended to\nattack agnostic, learnable fingerprints, and to open-world scenarios with\nunknown attacks.\n", "date": "2023-01-31", "categories": []}
{"id": "2206.07912", "title": "Double Sampling Randomized Smoothing", "abstract": "  Neural networks (NNs) are known to be vulnerable against adversarial\nperturbations, and thus there is a line of work aiming to provide robustness\ncertification for NNs, such as randomized smoothing, which samples smoothing\nnoises from a certain distribution to certify the robustness for a smoothed\nclassifier. However, as shown by previous work, the certified robust radius in\nrandomized smoothing suffers from scaling to large datasets (\"curse of\ndimensionality\"). To overcome this hurdle, we propose a Double Sampling\nRandomized Smoothing (DSRS) framework, which exploits the sampled probability\nfrom an additional smoothing distribution to tighten the robustness\ncertification of the previous smoothed classifier. Theoretically, under mild\nassumptions, we prove that DSRS can certify $\\Theta(\\sqrt d)$ robust radius\nunder $\\ell_2$ norm where $d$ is the input dimension, implying that DSRS may be\nable to break the curse of dimensionality of randomized smoothing. We\ninstantiate DSRS for a generalized family of Gaussian smoothing and propose an\nefficient and sound computing method based on customized dual optimization\nconsidering sampling error. Extensive experiments on MNIST, CIFAR-10, and\nImageNet verify our theory and show that DSRS certifies larger robust radii\nthan existing baselines consistently under different settings. Code is\navailable at https://github.com/llylly/DSRS.\n", "date": "2022-06-16", "categories": []}
{"id": "2207.11727", "title": "Can we achieve robustness from data alone?", "abstract": "  We introduce a meta-learning algorithm for adversarially robust\nclassification. The proposed method tries to be as model agnostic as possible\nand optimizes a dataset prior to its deployment in a machine learning system,\naiming to effectively erase its non-robust features. Once the dataset has been\ncreated, in principle no specialized algorithm (besides standard gradient\ndescent) is needed to train a robust model. We formulate the data optimization\nprocedure as a bi-level optimization problem on kernel regression, with a class\nof kernels that describe infinitely wide neural nets (Neural Tangent Kernels).\nWe present extensive experiments on standard computer vision benchmarks using a\nvariety of different models, demonstrating the effectiveness of our method,\nwhile also pointing out its current shortcomings. In parallel, we revisit prior\nwork that also focused on the problem of data optimization for robust\nclassification \\citep{Ily+19}, and show that being robust to adversarial\nattacks after standard (gradient descent) training on a suitable dataset is\nmore challenging than previously thought.\n", "date": "2022-07-24", "categories": []}
{"id": "2209.06300", "title": "PINCH: An Adversarial Extraction Attack Framework for Deep Learning\n  Models", "abstract": "  Adversarial extraction attacks constitute an insidious threat against Deep\nLearning (DL) models in-which an adversary aims to steal the architecture,\nparameters, and hyper-parameters of a targeted DL model. Existing extraction\nattack literature have observed varying levels of attack success for different\nDL models and datasets, yet the underlying cause(s) behind their susceptibility\noften remain unclear, and would help facilitate creating secure DL systems. In\nthis paper we present PINCH: an efficient and automated extraction attack\nframework capable of designing, deploying, and analyzing extraction attack\nscenarios across heterogeneous hardware platforms. Using PINCH, we perform\nextensive experimental evaluation of extraction attacks against 21 model\narchitectures to explore new extraction attack scenarios and further attack\nstaging. Our findings show (1) key extraction characteristics whereby\nparticular model configurations exhibit strong resilience against specific\nattacks, (2) even partial extraction success enables further staging for other\nadversarial attacks, and (3) equivalent stolen models uncover differences in\nexpressive power, yet exhibit similar captured knowledge.\n", "date": "2022-09-13", "categories": []}
{"id": "2210.05577", "title": "What Can the Neural Tangent Kernel Tell Us About Adversarial Robustness?", "abstract": "  The adversarial vulnerability of neural nets, and subsequent techniques to\ncreate robust models have attracted significant attention; yet we still lack a\nfull understanding of this phenomenon. Here, we study adversarial examples of\ntrained neural networks through analytical tools afforded by recent theory\nadvances connecting neural networks and kernel methods, namely the Neural\nTangent Kernel (NTK), following a growing body of work that leverages the NTK\napproximation to successfully analyze important deep learning phenomena and\ndesign algorithms for new applications. We show how NTKs allow to generate\nadversarial examples in a ``training-free'' fashion, and demonstrate that they\ntransfer to fool their finite-width neural net counterparts in the ``lazy''\nregime. We leverage this connection to provide an alternative view on robust\nand non-robust features, which have been suggested to underlie the adversarial\nbrittleness of neural nets. Specifically, we define and study features induced\nby the eigendecomposition of the kernel to better understand the role of robust\nand non-robust features, the reliance on both for standard classification and\nthe robustness-accuracy trade-off. We find that such features are surprisingly\nconsistent across architectures, and that robust features tend to correspond to\nthe largest eigenvalues of the model, and thus are learned early during\ntraining. Our framework allows us to identify and visualize non-robust yet\nuseful features. Finally, we shed light on the robustness mechanism underlying\nadversarial training of neural nets used in practice: quantifying the evolution\nof the associated empirical NTK, we demonstrate that its dynamics falls much\nearlier into the ``lazy'' regime and manifests a much stronger form of the well\nknown bias to prioritize learning features within the top eigenspaces of the\nkernel, compared to standard training.\n", "date": "2022-10-11", "categories": []}
{"id": "2301.13356", "title": "Inference Time Evidences of Adversarial Attacks for Forensic on\n  Transformers", "abstract": "  Vision Transformers (ViTs) are becoming a very popular paradigm for vision\ntasks as they achieve state-of-the-art performance on image classification.\nHowever, although early works implied that this network structure had increased\nrobustness against adversarial attacks, some works argue ViTs are still\nvulnerable. This paper presents our first attempt toward detecting adversarial\nattacks during inference time using the network's input and outputs as well as\nlatent features. We design four quantifications (or derivatives) of input,\noutput, and latent vectors of ViT-based models that provide a signature of the\ninference, which could be beneficial for the attack detection, and empirically\nstudy their behavior over clean samples and adversarial samples. The results\ndemonstrate that the quantifications from input (images) and output (posterior\nprobabilities) are promising for distinguishing clean and adversarial samples,\nwhile latent vectors offer less discriminative power, though they give some\ninsights on how adversarial perturbations work.\n", "date": "2023-01-31", "categories": []}
{"id": "2301.13486", "title": "Robust Linear Regression: Gradient-descent, Early-stopping, and Beyond", "abstract": "  In this work we study the robustness to adversarial attacks, of\nearly-stopping strategies on gradient-descent (GD) methods for linear\nregression. More precisely, we show that early-stopped GD is optimally robust\n(up to an absolute constant) against Euclidean-norm adversarial attacks.\nHowever, we show that this strategy can be arbitrarily sub-optimal in the case\nof general Mahalanobis attacks. This observation is compatible with recent\nfindings in the case of classification~\\cite{Vardi2022GradientMP} that show\nthat GD provably converges to non-robust models. To alleviate this issue, we\npropose to apply instead a GD scheme on a transformation of the data adapted to\nthe attack. This data transformation amounts to apply feature-depending\nlearning rates and we show that this modified GD is able to handle any\nMahalanobis attack, as well as more general attacks under some conditions.\nUnfortunately, choosing such adapted transformations can be hard for general\nattacks. To the rescue, we design a simple and tractable estimator whose\nadversarial risk is optimal up to within a multiplicative constant of 1.1124 in\nthe population regime, and works for any norm.\n", "date": "2023-01-31", "categories": []}
{"id": "2301.13694", "title": "Are Defenses for Graph Neural Networks Robust?", "abstract": "  A cursory reading of the literature suggests that we have made a lot of\nprogress in designing effective adversarial defenses for Graph Neural Networks\n(GNNs). Yet, the standard methodology has a serious flaw - virtually all of the\ndefenses are evaluated against non-adaptive attacks leading to overly\noptimistic robustness estimates. We perform a thorough robustness analysis of 7\nof the most popular defenses spanning the entire spectrum of strategies, i.e.,\naimed at improving the graph, the architecture, or the training. The results\nare sobering - most defenses show no or only marginal improvement compared to\nan undefended baseline. We advocate using custom adaptive attacks as a gold\nstandard and we outline the lessons we learned from successfully designing such\nattacks. Moreover, our diverse collection of perturbed graphs forms a\n(black-box) unit test offering a first glance at a model's robustness.\n", "date": "2023-01-31", "categories": []}
{"id": "2301.11806", "title": "PCV: A Point Cloud-Based Network Verifier", "abstract": "  3D vision with real-time LiDAR-based point cloud data became a vital part of\nautonomous system research, especially perception and prediction modules use\nfor object classification, segmentation, and detection. Despite their success,\npoint cloud-based network models are vulnerable to multiple adversarial\nattacks, where the certain factor of changes in the validation set causes\nsignificant performance drop in well-trained networks. Most of the existing\nverifiers work perfectly on 2D convolution. Due to complex architecture,\ndimension of hyper-parameter, and 3D convolution, no verifiers can perform the\nbasic layer-wise verification. It is difficult to conclude the robustness of a\n3D vision model without performing the verification. Because there will be\nalways corner cases and adversarial input that can compromise the model's\neffectiveness.\n  In this project, we describe a point cloud-based network verifier that\nsuccessfully deals state of the art 3D classifier PointNet verifies the\nrobustness by generating adversarial inputs. We have used extracted properties\nfrom the trained PointNet and changed certain factors for perturbation input.\nWe calculate the impact on model accuracy versus property factor and can test\nPointNet network's robustness against a small collection of perturbing input\nstates resulting from adversarial attacks like the suggested hybrid reverse\nsigned attack. The experimental results reveal that the resilience property of\nPointNet is affected by our hybrid reverse signed perturbation strategy\n", "date": "2023-01-27", "categories": []}
{"id": "2301.12151", "title": "Selecting Models based on the Risk of Damage Caused by Adversarial\n  Attacks", "abstract": "  Regulation, legal liabilities, and societal concerns challenge the adoption\nof AI in safety and security-critical applications. One of the key concerns is\nthat adversaries can cause harm by manipulating model predictions without being\ndetected. Regulation hence demands an assessment of the risk of damage caused\nby adversaries. Yet, there is no method to translate this high-level demand\ninto actionable metrics that quantify the risk of damage.\n  In this article, we propose a method to model and statistically estimate the\nprobability of damage arising from adversarial attacks. We show that our\nproposed estimator is statistically consistent and unbiased. In experiments, we\ndemonstrate that the estimation results of our method have a clear and\nactionable interpretation and outperform conventional metrics. We then show how\noperators can use the estimation results to reliably select the model with the\nlowest risk.\n", "date": "2023-01-28", "categories": []}
{"id": "2301.12595", "title": "Adversarial Attacks on Adversarial Bandits", "abstract": "  We study a security threat to adversarial multi-armed bandits, in which an\nattacker perturbs the loss or reward signal to control the behavior of the\nvictim bandit player. We show that the attacker is able to mislead any\nno-regret adversarial bandit algorithm into selecting a suboptimal target arm\nin every but sublinear (T-o(T)) number of rounds, while incurring only\nsublinear (o(T)) cumulative attack cost. This result implies critical security\nconcern in real-world bandit-based systems, e.g., in online recommendation, an\nattacker might be able to hijack the recommender system and promote a desired\nproduct. Our proposed attack algorithms require knowledge of only the regret\nrate, thus are agnostic to the concrete bandit algorithm employed by the victim\nplayer. We also derived a theoretical lower bound on the cumulative attack cost\nthat any victim-agnostic attack algorithm must incur. The lower bound matches\nthe upper bound achieved by our attack, which shows that our attack is\nasymptotically optimal.\n", "date": "2023-01-30", "categories": []}
{"id": "2301.12968", "title": "Improving Adversarial Transferability with Scheduled Step Size and Dual\n  Example", "abstract": "  Deep neural networks are widely known to be vulnerable to adversarial\nexamples, especially showing significantly poor performance on adversarial\nexamples generated under the white-box setting. However, most white-box attack\nmethods rely heavily on the target model and quickly get stuck in local optima,\nresulting in poor adversarial transferability. The momentum-based methods and\ntheir variants are proposed to escape the local optima for better\ntransferability. In this work, we notice that the transferability of\nadversarial examples generated by the iterative fast gradient sign method\n(I-FGSM) exhibits a decreasing trend when increasing the number of iterations.\nMotivated by this finding, we argue that the information of adversarial\nperturbations near the benign sample, especially the direction, benefits more\non the transferability. Thus, we propose a novel strategy, which uses the\nScheduled step size and the Dual example (SD), to fully utilize the adversarial\ninformation near the benign sample. Our proposed strategy can be easily\nintegrated with existing adversarial attack methods for better adversarial\ntransferability. Empirical evaluations on the standard ImageNet dataset\ndemonstrate that our proposed method can significantly enhance the\ntransferability of existing adversarial attacks.\n", "date": "2023-01-30", "categories": []}
{"id": "2301.13028", "title": "On the Efficacy of Metrics to Describe Adversarial Attacks", "abstract": "  Adversarial defenses are naturally evaluated on their ability to tolerate\nadversarial attacks. To test defenses, diverse adversarial attacks are crafted,\nthat are usually described in terms of their evading capability and the L0, L1,\nL2, and Linf norms. We question if the evading capability and L-norms are the\nmost effective information to claim that defenses have been tested against a\nrepresentative attack set. To this extent, we select image quality metrics from\nthe state of the art and search correlations between image perturbation and\ndetectability. We observe that computing L-norms alone is rarely the preferable\nsolution. We observe a strong correlation between the identified metrics\ncomputed on an adversarial image and the output of a detector on such an image,\nto the extent that they can predict the response of a detector with\napproximately 0.94 accuracy. Further, we observe that metrics can classify\nattacks based on similar perturbations and similar detectability. This suggests\na possible review of the approach to evaluate detectors, where additional\nmetrics are included to assure that a representative attack dataset is\nselected.\n", "date": "2023-01-30", "categories": []}
{"id": "2301.11457", "title": "Attacking Important Pixels for Anchor-free Detectors", "abstract": "  Deep neural networks have been demonstrated to be vulnerable to adversarial\nattacks: subtle perturbation can completely change the prediction result.\nExisting adversarial attacks on object detection focus on attacking\nanchor-based detectors, which may not work well for anchor-free detectors. In\nthis paper, we propose the first adversarial attack dedicated to anchor-free\ndetectors. It is a category-wise attack that attacks important pixels of all\ninstances of a category simultaneously. Our attack manifests in two forms,\nsparse category-wise attack (SCA) and dense category-wise attack (DCA), that\nminimize the $L_0$ and $L_\\infty$ norm-based perturbations, respectively. For\nDCA, we present three variants, DCA-G, DCA-L, and DCA-S, that select a global\nregion, a local region, and a semantic region, respectively, to attack. Our\nexperiments on large-scale benchmark datasets including PascalVOC, MS-COCO, and\nMS-COCO Keypoints indicate that our proposed methods achieve state-of-the-art\nattack performance and transferability on both object detection and human pose\nestimation tasks.\n", "date": "2023-01-26", "categories": []}
{"id": "2301.11544", "title": "Targeted Attacks on Timeseries Forecasting", "abstract": "  Real-world deep learning models developed for Time Series Forecasting are\nused in several critical applications ranging from medical devices to the\nsecurity domain. Many previous works have shown how deep learning models are\nprone to adversarial attacks and studied their vulnerabilities. However, the\nvulnerabilities of time series models for forecasting due to adversarial inputs\nare not extensively explored. While the attack on a forecasting model might aim\nto deteriorate the performance of the model, it is more effective, if the\nattack is focused on a specific impact on the model's output. In this paper, we\npropose a novel formulation of Directional, Amplitudinal, and Temporal targeted\nadversarial attacks on time series forecasting models. These targeted attacks\ncreate a specific impact on the amplitude and direction of the output\nprediction. We use the existing adversarial attack techniques from the computer\nvision domain and adapt them for time series. Additionally, we propose a\nmodified version of the Auto Projected Gradient Descent attack for targeted\nattacks. We examine the impact of the proposed targeted attacks versus\nuntargeted attacks. We use KS-Tests to statistically demonstrate the impact of\nthe attack. Our experimental results show how targeted attacks on time series\nmodels are viable and are more powerful in terms of statistical similarity. It\nis, hence difficult to detect through statistical methods. We believe that this\nwork opens a new paradigm in the time series forecasting domain and represents\nan important consideration for developing better defenses.\n", "date": "2023-01-27", "categories": []}
{"id": "2301.11553", "title": "Robust Transformer with Locality Inductive Bias and Feature\n  Normalization", "abstract": "  Vision transformers have been demonstrated to yield state-of-the-art results\non a variety of computer vision tasks using attention-based networks. However,\nresearch works in transformers mostly do not investigate robustness/accuracy\ntrade-off, and they still struggle to handle adversarial perturbations. In this\npaper, we explore the robustness of vision transformers against adversarial\nperturbations and try to enhance their robustness/accuracy trade-off in white\nbox attack settings. To this end, we propose Locality iN Locality (LNL)\ntransformer model. We prove that the locality introduction to LNL contributes\nto the robustness performance since it aggregates local information such as\nlines, edges, shapes, and even objects. In addition, to further improve the\nrobustness performance, we encourage LNL to extract training signal from the\nmoments (a.k.a., mean and standard deviation) and the normalized features. We\nvalidate the effectiveness and generality of LNL by achieving state-of-the-art\nresults in terms of accuracy and robustness metrics on German Traffic Sign\nRecognition Benchmark (GTSRB) and Canadian Institute for Advanced Research\n(CIFAR-10). More specifically, for traffic sign classification, the proposed\nLNL yields gains of 1.1% and ~35% in terms of clean and robustness accuracy\ncompared to the state-of-the-art studies.\n", "date": "2023-01-27", "categories": []}
{"id": "2301.10454", "title": "A Data-Centric Approach for Improving Adversarial Training Through the\n  Lens of Out-of-Distribution Detection", "abstract": "  Current machine learning models achieve super-human performance in many\nreal-world applications. Still, they are susceptible against imperceptible\nadversarial perturbations. The most effective solution for this problem is\nadversarial training that trains the model with adversarially perturbed samples\ninstead of original ones. Various methods have been developed over recent years\nto improve adversarial training such as data augmentation or modifying training\nattacks. In this work, we examine the same problem from a new data-centric\nperspective. For this purpose, we first demonstrate that the existing\nmodel-based methods can be equivalent to applying smaller perturbation or\noptimization weights to the hard training examples. By using this finding, we\npropose detecting and removing these hard samples directly from the training\nprocedure rather than applying complicated algorithms to mitigate their\neffects. For detection, we use maximum softmax probability as an effective\nmethod in out-of-distribution detection since we can consider the hard samples\nas the out-of-distribution samples for the whole data distribution. Our results\non SVHN and CIFAR-10 datasets show the effectiveness of this method in\nimproving the adversarial training without adding too much computational cost.\n", "date": "2023-01-25", "categories": []}
{"id": "2301.10576", "title": "A Study on FGSM Adversarial Training for Neural Retrieval", "abstract": "  Neural retrieval models have acquired significant effectiveness gains over\nthe last few years compared to term-based methods. Nevertheless, those models\nmay be brittle when faced to typos, distribution shifts or vulnerable to\nmalicious attacks. For instance, several recent papers demonstrated that such\nvariations severely impacted models performances, and then tried to train more\nresilient models. Usual approaches include synonyms replacements or typos\ninjections -- as data-augmentation -- and the use of more robust tokenizers\n(characterBERT, BPE-dropout). To further complement the literature, we\ninvestigate in this paper adversarial training as another possible solution to\nthis robustness issue. Our comparison includes the two main families of\nBERT-based neural retrievers, i.e. dense and sparse, with and without\ndistillation techniques. We then demonstrate that one of the most simple\nadversarial training techniques -- the Fast Gradient Sign Method (FGSM) -- can\nimprove first stage rankers robustness and effectiveness. In particular, FGSM\nincreases models performances on both in-domain and out-of-domain\ndistributions, and also on queries with typos, for multiple neural retrievers.\n", "date": "2023-01-25", "categories": []}
{"id": "2211.10908", "title": "ESTAS: Effective and Stable Trojan Attacks in Self-supervised Encoders\n  with One Target Unlabelled Sample", "abstract": "  Emerging self-supervised learning (SSL) has become a popular image\nrepresentation encoding method to obviate the reliance on labeled data and\nlearn rich representations from large-scale, ubiquitous unlabelled data. Then\none can train a downstream classifier on top of the pre-trained SSL image\nencoder with few or no labeled downstream data. Although extensive works show\nthat SSL has achieved remarkable and competitive performance on different\ndownstream tasks, its security concerns, e.g, Trojan attacks in SSL encoders,\nare still not well-studied. In this work, we present a novel Trojan Attack\nmethod, denoted by ESTAS, that can enable an effective and stable attack in SSL\nencoders with only one target unlabeled sample. In particular, we propose\nconsistent trigger poisoning and cascade optimization in ESTAS to improve\nattack efficacy and model accuracy, and eliminate the expensive target-class\ndata sample extraction from large-scale disordered unlabelled data. Our\nsubstantial experiments on multiple datasets show that ESTAS stably achieves >\n99% attacks success rate (ASR) with one target-class sample. Compared to prior\nworks, ESTAS attains > 30% ASR increase and > 8.3% accuracy improvement on\naverage.\n", "date": "2022-11-20", "categories": []}
{"id": "2301.09740", "title": "DODEM: DOuble DEfense Mechanism Against Adversarial Attacks Towards\n  Secure Industrial Internet of Things Analytics", "abstract": "  Industrial Internet of Things (I-IoT) is a collaboration of devices, sensors,\nand networking equipment to monitor and collect data from industrial\noperations. Machine learning (ML) methods use this data to make high-level\ndecisions with minimal human intervention. Data-driven predictive maintenance\n(PDM) is a crucial ML-based I-IoT application to find an optimal maintenance\nschedule for industrial assets. The performance of these ML methods can\nseriously be threatened by adversarial attacks where an adversary crafts\nperturbed data and sends it to the ML model to deteriorate its prediction\nperformance. The models should be able to stay robust against these attacks\nwhere robustness is measured by how much perturbation in input data affects\nmodel performance. Hence, there is a need for effective defense mechanisms that\ncan protect these models against adversarial attacks. In this work, we propose\na double defense mechanism to detect and mitigate adversarial attacks in I-IoT\nenvironments. We first detect if there is an adversarial attack on a given\nsample using novelty detection algorithms. Then, based on the outcome of our\nalgorithm, marking an instance as attack or normal, we select adversarial\nretraining or standard training to provide a secondary defense layer. If there\nis an attack, adversarial retraining provides a more robust model, while we\napply standard training for regular samples. Since we may not know if an attack\nwill take place, our adaptive mechanism allows us to consider irregular changes\nin data. The results show that our double defense strategy is highly efficient\nwhere we can improve model robustness by up to 64.6% and 52% compared to\nstandard and adversarial retraining, respectively.\n", "date": "2023-01-23", "categories": []}
{"id": "2301.09879", "title": "Data Augmentation Alone Can Improve Adversarial Training", "abstract": "  Adversarial training suffers from the issue of robust overfitting, which\nseriously impairs its generalization performance. Data augmentation, which is\neffective at preventing overfitting in standard training, has been observed by\nmany previous works to be ineffective in mitigating overfitting in adversarial\ntraining. This work proves that, contrary to previous findings, data\naugmentation alone can significantly boost accuracy and robustness in\nadversarial training. We find that the hardness and the diversity of data\naugmentation are important factors in combating robust overfitting. In general,\ndiversity can improve both accuracy and robustness, while hardness can boost\nrobustness at the cost of accuracy within a certain limit and degrade them both\nover that limit. To mitigate robust overfitting, we first propose a new crop\ntransformation, Cropshift, which has improved diversity compared to the\nconventional one (Padcrop). We then propose a new data augmentation scheme,\nbased on Cropshift, with much improved diversity and well-balanced hardness.\nEmpirically, our augmentation method achieves the state-of-the-art accuracy and\nrobustness for data augmentations in adversarial training. Furthermore, when\ncombined with weight averaging it matches, or even exceeds, the performance of\nthe best contemporary regularization methods for alleviating robust\noverfitting. Code is available at:\nhttps://github.com/TreeLLi/DA-Alone-Improves-AT.\n", "date": "2023-01-24", "categories": []}
{"id": "2301.08842", "title": "Limitations of Piecewise Linearity for Efficient Robustness\n  Certification", "abstract": "  Certified defenses against small-norm adversarial examples have received\ngrowing attention in recent years; though certified accuracies of\nstate-of-the-art methods remain far below their non-robust counterparts,\ndespite the fact that benchmark datasets have been shown to be well-separated\nat far larger radii than the literature generally attempts to certify. In this\nwork, we offer insights that identify potential factors in this performance\ngap. Specifically, our analysis reveals that piecewise linearity imposes\nfundamental limitations on the tightness of leading certification techniques.\nThese limitations are felt in practical terms as a greater need for capacity in\nmodels hoped to be certified efficiently. Moreover, this is in addition to the\ncapacity necessary to learn a robust boundary, studied in prior work. However,\nwe argue that addressing the limitations of piecewise linearity through scaling\nup model capacity may give rise to potential difficulties -- particularly\nregarding robust generalization -- therefore, we conclude by suggesting that\ndeveloping smooth activation functions may be the way forward for advancing the\nperformance of certified neural networks.\n", "date": "2023-01-21", "categories": []}
{"id": "2301.09305", "title": "Practical Adversarial Attacks Against AI-Driven Power Allocation in a\n  Distributed MIMO Network", "abstract": "  In distributed multiple-input multiple-output (D-MIMO) networks, power\ncontrol is crucial to optimize the spectral efficiencies of users and max-min\nfairness (MMF) power control is a commonly used strategy as it satisfies\nuniform quality-of-service to all users. The optimal solution of MMF power\ncontrol requires high complexity operations and hence deep neural network based\nartificial intelligence (AI) solutions are proposed to decrease the complexity.\nAlthough quite accurate models can be achieved by using AI, these models have\nsome intrinsic vulnerabilities against adversarial attacks where carefully\ncrafted perturbations are applied to the input of the AI model. In this work,\nwe show that threats against the target AI model which might be originated from\nmalicious users or radio units can substantially decrease the network\nperformance by applying a successful adversarial sample, even in the most\nconstrained circumstances. We also demonstrate that the risk associated with\nthese kinds of adversarial attacks is higher than the conventional attack\nthreats. Detailed simulations reveal the effectiveness of adversarial attacks\nand the necessity of smart defense techniques.\n", "date": "2023-01-23", "categories": []}
{"id": "2301.00986", "title": "Look, Listen, and Attack: Backdoor Attacks Against Video Action\n  Recognition", "abstract": "  Deep neural networks (DNNs) are vulnerable to a class of attacks called\n\"backdoor attacks\", which create an association between a backdoor trigger and\na target label the attacker is interested in exploiting. A backdoored DNN\nperforms well on clean test images, yet persistently predicts an\nattacker-defined label for any sample in the presence of the backdoor trigger.\nAlthough backdoor attacks have been extensively studied in the image domain,\nthere are very few works that explore such attacks in the video domain, and\nthey tend to conclude that image backdoor attacks are less effective in the\nvideo domain. In this work, we revisit the traditional backdoor threat model\nand incorporate additional video-related aspects to that model. We show that\npoisoned-label image backdoor attacks could be extended temporally in two ways,\nstatically and dynamically, leading to highly effective attacks in the video\ndomain. In addition, we explore natural video backdoors to highlight the\nseriousness of this vulnerability in the video domain. And, for the first time,\nwe study multi-modal (audiovisual) backdoor attacks against video action\nrecognition models, where we show that attacking a single modality is enough\nfor achieving a high attack success rate.\n", "date": "2023-01-03", "categories": []}
{"id": "2301.07474", "title": "Threats, Vulnerabilities, and Controls of Machine Learning Based\n  Systems: A Survey and Taxonomy", "abstract": "  In this article, we propose the Artificial Intelligence Security Taxonomy to\nsystematize the knowledge of threats, vulnerabilities, and security controls of\nmachine-learning-based (ML-based) systems. We first classify the damage caused\nby attacks against ML-based systems, define ML-specific security, and discuss\nits characteristics. Next, we enumerate all relevant assets and stakeholders\nand provide a general taxonomy for ML-specific threats. Then, we collect a wide\nrange of security controls against ML-specific threats through an extensive\nreview of recent literature. Finally, we classify the vulnerabilities and\ncontrols of an ML-based system in terms of each vulnerable asset in the\nsystem's entire lifecycle.\n", "date": "2023-01-18", "categories": []}
{"id": "2301.07487", "title": "Adversarial Robust Deep Reinforcement Learning Requires Redefining\n  Robustness", "abstract": "  Learning from raw high dimensional data via interaction with a given\nenvironment has been effectively achieved through the utilization of deep\nneural networks. Yet the observed degradation in policy performance caused by\nimperceptible worst-case policy dependent translations along high sensitivity\ndirections (i.e. adversarial perturbations) raises concerns on the robustness\nof deep reinforcement learning policies. In our paper, we show that these high\nsensitivity directions do not lie only along particular worst-case directions,\nbut rather are more abundant in the deep neural policy landscape and can be\nfound via more natural means in a black-box setting. Furthermore, we show that\nvanilla training techniques intriguingly result in learning more robust\npolicies compared to the policies learnt via the state-of-the-art adversarial\ntraining techniques. We believe our work lays out intriguing properties of the\ndeep reinforcement learning policy manifold and our results can help to build\nrobust and generalizable deep reinforcement learning policies.\n", "date": "2023-01-17", "categories": []}
{"id": "2301.07520", "title": "Adversarial AI in Insurance: Pervasiveness and Resilience", "abstract": "  The rapid and dynamic pace of Artificial Intelligence (AI) and Machine\nLearning (ML) is revolutionizing the insurance sector. AI offers significant,\nvery much welcome advantages to insurance companies, and is fundamental to\ntheir customer-centricity strategy. It also poses challenges, in the project\nand implementation phase. Among those, we study Adversarial Attacks, which\nconsist of the creation of modified input data to deceive an AI system and\nproduce false outputs. We provide examples of attacks on insurance AI\napplications, categorize them, and argue on defence methods and precautionary\nsystems, considering that they can involve few-shot and zero-shot\nmultilabelling. A related topic, with growing interest, is the validation and\nverification of systems incorporating AI and ML components. These topics are\ndiscussed in various sections of this paper.\n", "date": "2023-01-17", "categories": []}
{"id": "2206.12381", "title": "Defending Backdoor Attacks on Vision Transformer via Patch Processing", "abstract": "  Vision Transformers (ViTs) have a radically different architecture with\nsignificantly less inductive bias than Convolutional Neural Networks. Along\nwith the improvement in performance, security and robustness of ViTs are also\nof great importance to study. In contrast to many recent works that exploit the\nrobustness of ViTs against adversarial examples, this paper investigates a\nrepresentative causative attack, i.e., backdoor. We first examine the\nvulnerability of ViTs against various backdoor attacks and find that ViTs are\nalso quite vulnerable to existing attacks. However, we observe that the\nclean-data accuracy and backdoor attack success rate of ViTs respond\ndistinctively to patch transformations before the positional encoding. Then,\nbased on this finding, we propose an effective method for ViTs to defend both\npatch-based and blending-based trigger backdoor attacks via patch processing.\nThe performances are evaluated on several benchmark datasets, including\nCIFAR10, GTSRB, and TinyImageNet, which show the proposed novel defense is very\nsuccessful in mitigating backdoor attacks for ViTs. To the best of our\nknowledge, this paper presents the first defensive strategy that utilizes a\nunique characteristic of ViTs against backdoor attacks.\n  The paper will appear in the Proceedings of the AAAI'23 Conference. This work\nwas initially submitted in November 2021 to CVPR'22, then it was re-submitted\nto ECCV'22. The paper was made public in June 2022. The authors sincerely thank\nall the referees from the Program Committees of CVPR'22, ECCV'22, and AAAI'23.\n", "date": "2022-06-24", "categories": []}
{"id": "2211.15030", "title": "Imperceptible Adversarial Attack via Invertible Neural Networks", "abstract": "  Adding perturbations via utilizing auxiliary gradient information or\ndiscarding existing details of the benign images are two common approaches for\ngenerating adversarial examples. Though visual imperceptibility is the desired\nproperty of adversarial examples, conventional adversarial attacks still\ngenerate traceable adversarial perturbations. In this paper, we introduce a\nnovel Adversarial Attack via Invertible Neural Networks (AdvINN) method to\nproduce robust and imperceptible adversarial examples. Specifically, AdvINN\nfully takes advantage of the information preservation property of Invertible\nNeural Networks and thereby generates adversarial examples by simultaneously\nadding class-specific semantic information of the target class and dropping\ndiscriminant information of the original class. Extensive experiments on\nCIFAR-10, CIFAR-100, and ImageNet-1K demonstrate that the proposed AdvINN\nmethod can produce less imperceptible adversarial images than the\nstate-of-the-art methods and AdvINN yields more robust adversarial examples\nwith high confidence compared to other adversarial attacks.\n", "date": "2022-11-28", "categories": []}
{"id": "2301.06241", "title": "BEAGLE: Forensics of Deep Learning Backdoor Attack for Better Defense", "abstract": "  Deep Learning backdoor attacks have a threat model similar to traditional\ncyber attacks. Attack forensics, a critical counter-measure for traditional\ncyber attacks, is hence of importance for defending model backdoor attacks. In\nthis paper, we propose a novel model backdoor forensics technique. Given a few\nattack samples such as inputs with backdoor triggers, which may represent\ndifferent types of backdoors, our technique automatically decomposes them to\nclean inputs and the corresponding triggers. It then clusters the triggers\nbased on their properties to allow automatic attack categorization and\nsummarization. Backdoor scanners can then be automatically synthesized to find\nother instances of the same type of backdoor in other models. Our evaluation on\n2,532 pre-trained models, 10 popular attacks, and comparison with 9 baselines\nshow that our technique is highly effective. The decomposed clean inputs and\ntriggers closely resemble the ground truth. The synthesized scanners\nsubstantially outperform the vanilla versions of existing scanners that can\nhardly generalize to different kinds of attacks.\n", "date": "2023-01-16", "categories": []}
{"id": "2301.06286", "title": "Meta Generative Attack on Person Reidentification", "abstract": "  Adversarial attacks have been recently investigated in person\nre-identification. These attacks perform well under cross dataset or cross\nmodel setting. However, the challenges present in cross-dataset cross-model\nscenario does not allow these models to achieve similar accuracy. To this end,\nwe propose our method with the goal of achieving better transferability against\ndifferent models and across datasets. We generate a mask to obtain better\nperformance across models and use meta learning to boost the generalizability\nin the challenging cross-dataset cross-model setting. Experiments on\nMarket-1501, DukeMTMC-reID and MSMT-17 demonstrate favorable results compared\nto other attacks.\n", "date": "2023-01-16", "categories": []}
{"id": "2301.06871", "title": "Denoising Diffusion Probabilistic Models as a Defense against\n  Adversarial Attacks", "abstract": "  Neural Networks are infamously sensitive to small perturbations in their\ninputs, making them vulnerable to adversarial attacks. This project evaluates\nthe performance of Denoising Diffusion Probabilistic Models (DDPM) as a\npurification technique to defend against adversarial attacks. This works by\nadding noise to an adversarial example before removing it through the reverse\nprocess of the diffusion model. We evaluate the approach on the PatchCamelyon\ndata set for histopathologic scans of lymph node sections and find an\nimprovement of the robust accuracy by up to 88\\% of the original model's\naccuracy, constituting a considerable improvement over the vanilla model and\nour baselines. The project code is located at\nhttps://github.com/ankile/Adversarial-Diffusion.\n", "date": "2023-01-17", "categories": []}
{"id": "2301.05264", "title": "Security-Aware Approximate Spiking Neural Networks", "abstract": "  Deep Neural Networks (DNNs) and Spiking Neural Networks (SNNs) are both known\nfor their susceptibility to adversarial attacks. Therefore, researchers in the\nrecent past have extensively studied the robustness and defense of DNNs and\nSNNs under adversarial attacks. Compared to accurate SNNs (AccSNN), approximate\nSNNs (AxSNNs) are known to be up to 4X more energy-efficient for ultra-low\npower applications. Unfortunately, the robustness of AxSNNs under adversarial\nattacks is yet unexplored. In this paper, we first extensively analyze the\nrobustness of AxSNNs with different structural parameters and approximation\nlevels under two gradient-based and two neuromorphic attacks. Then, we propose\ntwo novel defense methods, i.e., precision scaling and approximate\nquantization-aware filtering (AQF), for securing AxSNNs. We evaluated the\neffectiveness of these two defense methods using both static and neuromorphic\ndatasets. Our results demonstrate that AxSNNs are more prone to adversarial\nattacks than AccSNNs, but precision scaling and AQF significantly improve the\nrobustness of AxSNNs. For instance, a PGD attack on AxSNN results in a 72\\%\naccuracy loss compared to AccSNN without any attack, whereas the same attack on\nthe precision-scaled AxSNN leads to only a 17\\% accuracy loss in the static\nMNIST dataset (4X robustness improvement). Similarly, a Sparse Attack on AxSNN\nleads to a 77\\% accuracy loss when compared to AccSNN without any attack,\nwhereas the same attack on an AxSNN with AQF leads to only a 2\\% accuracy loss\nin the neuromorphic DVS128 Gesture dataset (38X robustness improvement).\n", "date": "2023-01-12", "categories": []}
{"id": "2301.05506", "title": "On the feasibility of attacking Thai LPR systems with adversarial\n  examples", "abstract": "  Recent advances in deep neural networks (DNNs) have significantly enhanced\nthe capabilities of optical character recognition (OCR) technology, enabling\nits adoption to a wide range of real-world applications. Despite this success,\nDNN-based OCR is shown to be vulnerable to adversarial attacks, in which the\nadversary can influence the DNN model's prediction by carefully manipulating\ninput to the model. Prior work has demonstrated the security impacts of\nadversarial attacks on various OCR languages. However, to date, no studies have\nbeen conducted and evaluated on an OCR system tailored specifically for the\nThai language. To bridge this gap, this work presents a feasibility study of\nperforming adversarial attacks on a specific Thai OCR application -- Thai\nLicense Plate Recognition (LPR). Moreover, we propose a new type of adversarial\nattack based on the \\emph{semi-targeted} scenario and show that this scenario\nis highly realistic in LPR applications. Our experimental results show the\nfeasibility of our attacks as they can be performed on a commodity computer\ndesktop with over 90% attack success rate.\n", "date": "2023-01-13", "categories": []}
{"id": "2301.04093", "title": "On the Robustness of AlphaFold: A COVID-19 Case Study", "abstract": "  Protein folding neural networks (PFNNs) such as AlphaFold predict remarkably\naccurate structures of proteins compared to other approaches. However, the\nrobustness of such networks has heretofore not been explored. This is\nparticularly relevant given the broad social implications of such technologies\nand the fact that biologically small perturbations in the protein sequence do\nnot generally lead to drastic changes in the protein structure. In this paper,\nwe demonstrate that AlphaFold does not exhibit such robustness despite its high\naccuracy. This raises the challenge of detecting and quantifying the extent to\nwhich these predicted protein structures can be trusted. To measure the\nrobustness of the predicted structures, we utilize (i) the root-mean-square\ndeviation (RMSD) and (ii) the Global Distance Test (GDT) similarity measure\nbetween the predicted structure of the original sequence and the structure of\nits adversarially perturbed version. We prove that the problem of minimally\nperturbing protein sequences to fool protein folding neural networks is\nNP-complete. Based on the well-established BLOSUM62 sequence alignment scoring\nmatrix, we generate adversarial protein sequences and show that the RMSD\nbetween the predicted protein structure and the structure of the original\nsequence are very large when the adversarial changes are bounded by (i) 20\nunits in the BLOSUM62 distance, and (ii) five residues (out of hundreds or\nthousands of residues) in the given protein sequence. In our experimental\nevaluation, we consider 111 COVID-19 proteins in the Universal Protein resource\n(UniProt), a central resource for protein data managed by the European\nBioinformatics Institute, Swiss Institute of Bioinformatics, and the US Protein\nInformation Resource. These result in an overall GDT similarity test score\naverage of around 34%, demonstrating a substantial drop in the performance of\nAlphaFold.\n", "date": "2023-01-10", "categories": []}
{"id": "2301.04299", "title": "SoK: Adversarial Machine Learning Attacks and Defences in Multi-Agent\n  Reinforcement Learning", "abstract": "  Multi-Agent Reinforcement Learning (MARL) is vulnerable to Adversarial\nMachine Learning (AML) attacks and needs adequate defences before it can be\nused in real world applications. We have conducted a survey into the use of\nexecution-time AML attacks against MARL and the defences against those attacks.\nWe surveyed related work in the application of AML in Deep Reinforcement\nLearning (DRL) and Multi-Agent Learning (MAL) to inform our analysis of AML for\nMARL. We propose a novel perspective to understand the manner of perpetrating\nan AML attack, by defining Attack Vectors. We develop two new frameworks to\naddress a gap in current modelling frameworks, focusing on the means and tempo\nof an AML attack against MARL, and identify knowledge gaps and future avenues\nof research.\n", "date": "2023-01-11", "categories": []}
{"id": "2301.04472", "title": "Adversarial training with informed data selection", "abstract": "  With the increasing amount of available data and advances in computing\ncapabilities, deep neural networks (DNNs) have been successfully employed to\nsolve challenging tasks in various areas, including healthcare, climate, and\nfinance. Nevertheless, state-of-the-art DNNs are susceptible to\nquasi-imperceptible perturbed versions of the original images -- adversarial\nexamples. These perturbations of the network input can lead to disastrous\nimplications in critical areas where wrong decisions can directly affect human\nlives. Adversarial training is the most efficient solution to defend the\nnetwork against these malicious attacks. However, adversarial trained networks\ngenerally come with lower clean accuracy and higher computational complexity.\nThis work proposes a data selection (DS) strategy to be applied in the\nmini-batch training. Based on the cross-entropy loss, the most relevant samples\nin the batch are selected to update the model parameters in the\nbackpropagation. The simulation results show that a good compromise can be\nobtained regarding robustness and standard accuracy, whereas the computational\ncomplexity of the backpropagation pass is reduced.\n", "date": "2023-01-07", "categories": []}
{"id": "2206.10057", "title": "Robust Deep Reinforcement Learning through Bootstrapped Opportunistic\n  Curriculum", "abstract": "  Despite considerable advances in deep reinforcement learning, it has been\nshown to be highly vulnerable to adversarial perturbations to state\nobservations. Recent efforts that have attempted to improve adversarial\nrobustness of reinforcement learning can nevertheless tolerate only very small\nperturbations, and remain fragile as perturbation size increases. We propose\nBootstrapped Opportunistic Adversarial Curriculum Learning (BCL), a novel\nflexible adversarial curriculum learning framework for robust reinforcement\nlearning. Our framework combines two ideas: conservatively bootstrapping each\ncurriculum phase with highest quality solutions obtained from multiple runs of\nthe previous phase, and opportunistically skipping forward in the curriculum.\nIn our experiments we show that the proposed BCL framework enables dramatic\nimprovements in robustness of learned policies to adversarial perturbations.\nThe greatest improvement is for Pong, where our framework yields robustness to\nperturbations of up to 25/255; in contrast, the best existing approach can only\ntolerate adversarial noise up to 5/255. Our code is available at:\nhttps://github.com/jlwu002/BCL.\n", "date": "2022-06-21", "categories": []}
{"id": "2301.03662", "title": "On adversarial robustness and the use of Wasserstein ascent-descent\n  dynamics to enforce it", "abstract": "  We propose iterative algorithms to solve adversarial problems in a variety of\nsupervised learning settings of interest. Our algorithms, which can be\ninterpreted as suitable ascent-descent dynamics in Wasserstein spaces, take the\nform of a system of interacting particles. These interacting particle dynamics\nare shown to converge toward appropriate mean-field limit equations in certain\nlarge number of particles regimes. In turn, we prove that, under certain\nregularity assumptions, these mean-field equations converge, in the large time\nlimit, toward approximate Nash equilibria of the original adversarial learning\nproblems. We present results for nonconvex-nonconcave settings, as well as for\nnonconvex-concave ones. Numerical experiments illustrate our results.\n", "date": "2023-01-09", "categories": []}
{"id": "2301.03703", "title": "On the Susceptibility and Robustness of Time Series Models through\n  Adversarial Attack and Defense", "abstract": "  Under adversarial attacks, time series regression and classification are\nvulnerable. Adversarial defense, on the other hand, can make the models more\nresilient. It is important to evaluate how vulnerable different time series\nmodels are to attacks and how well they recover using defense. The sensitivity\nto various attacks and the robustness using the defense of several time series\nmodels are investigated in this study. Experiments are run on seven-time series\nmodels with three adversarial attacks and one adversarial defense. According to\nthe findings, all models, particularly GRU and RNN, appear to be vulnerable.\nLSTM and GRU also have better defense recovery. FGSM exceeds the competitors in\nterms of attacks. PGD attacks are more difficult to recover from than other\nsorts of attacks.\n", "date": "2023-01-09", "categories": []}
{"id": "2301.03966", "title": "AdvBiom: Adversarial Attacks on Biometric Matchers", "abstract": "  With the advent of deep learning models, face recognition systems have\nachieved impressive recognition rates. The workhorses behind this success are\nConvolutional Neural Networks (CNNs) and the availability of large training\ndatasets. However, we show that small human-imperceptible changes to face\nsamples can evade most prevailing face recognition systems. Even more alarming\nis the fact that the same generator can be extended to other traits in the\nfuture. In this work, we present how such a generator can be trained and also\nextended to other biometric modalities, such as fingerprint recognition\nsystems.\n", "date": "2023-01-10", "categories": []}
{"id": "2301.02905", "title": "REaaS: Enabling Adversarially Robust Downstream Classifiers via Robust\n  Encoder as a Service", "abstract": "  Encoder as a service is an emerging cloud service. Specifically, a service\nprovider first pre-trains an encoder (i.e., a general-purpose feature\nextractor) via either supervised learning or self-supervised learning and then\ndeploys it as a cloud service API. A client queries the cloud service API to\nobtain feature vectors for its training/testing inputs when training/testing\nits classifier (called downstream classifier). A downstream classifier is\nvulnerable to adversarial examples, which are testing inputs with carefully\ncrafted perturbation that the downstream classifier misclassifies. Therefore,\nin safety and security critical applications, a client aims to build a robust\ndownstream classifier and certify its robustness guarantees against adversarial\nexamples.\n  What APIs should the cloud service provide, such that a client can use any\ncertification method to certify the robustness of its downstream classifier\nagainst adversarial examples while minimizing the number of queries to the\nAPIs? How can a service provider pre-train an encoder such that clients can\nbuild more certifiably robust downstream classifiers? We aim to answer the two\nquestions in this work. For the first question, we show that the cloud service\nonly needs to provide two APIs, which we carefully design, to enable a client\nto certify the robustness of its downstream classifier with a minimal number of\nqueries to the APIs. For the second question, we show that an encoder\npre-trained using a spectral-norm regularization term enables clients to build\nmore robust downstream classifiers.\n", "date": "2023-01-07", "categories": []}
{"id": "2301.03110", "title": "RobArch: Designing Robust Architectures against Adversarial Attacks", "abstract": "  Adversarial Training is the most effective approach for improving the\nrobustness of Deep Neural Networks (DNNs). However, compared to the large body\nof research in optimizing the adversarial training process, there are few\ninvestigations into how architecture components affect robustness, and they\nrarely constrain model capacity. Thus, it is unclear where robustness precisely\ncomes from. In this work, we present the first large-scale systematic study on\nthe robustness of DNN architecture components under fixed parameter budgets.\nThrough our investigation, we distill 18 actionable robust network design\nguidelines that empower model developers to gain deep insights. We demonstrate\nthese guidelines' effectiveness by introducing the novel Robust Architecture\n(RobArch) model that instantiates the guidelines to build a family of\ntop-performing models across parameter capacities against strong adversarial\nattacks. RobArch achieves the new state-of-the-art AutoAttack accuracy on the\nRobustBench ImageNet leaderboard. The code is available at\n$\\href{https://github.com/ShengYun-Peng/RobArch}{\\text{this url}}$.\n", "date": "2023-01-08", "categories": []}
{"id": "2301.01495", "title": "Beckman Defense", "abstract": "  Optimal transport (OT) based distributional robust optimisation (DRO) has\nreceived some traction in the recent past. However, it is at a nascent stage\nbut has a sound potential in robustifying the deep learning models.\nInterestingly, OT barycenters demonstrate a good robustness against adversarial\nattacks. Owing to the computationally expensive nature of OT barycenters, they\nhave not been investigated under DRO framework. In this work, we propose a new\nbarycenter, namely Beckman barycenter, which can be computed efficiently and\nused for training the network to defend against adversarial attacks in\nconjunction with adversarial training. We propose a novel formulation of\nBeckman barycenter and analytically obtain the barycenter using the marginals\nof the input image. We show that the Beckman barycenter can be used to train\nadversarially trained networks to improve the robustness. Our training is\nextremely efficient as it requires only a single epoch of training. Elaborate\nexperiments on CIFAR-10, CIFAR-100 and Tiny ImageNet demonstrate that training\nan adversarially robust network with Beckman barycenter can significantly\nincrease the performance. Under auto attack, we get a a maximum boost of 10\\%\nin CIFAR-10, 8.34\\% in CIFAR-100 and 11.51\\% in Tiny ImageNet. Our code is\navailable at\nhttps://github.com/Visual-Conception-Group/test-barycentric-defense.\n", "date": "2023-01-04", "categories": []}
{"id": "2301.01832", "title": "Availability Adversarial Attack and Countermeasures for Deep\n  Learning-based Load Forecasting", "abstract": "  The forecast of electrical loads is essential for the planning and operation\nof the power system. Recently, advances in deep learning have enabled more\naccurate forecasts. However, deep neural networks are prone to adversarial\nattacks. Although most of the literature focuses on integrity-based attacks,\nthis paper proposes availability-based adversarial attacks, which can be more\neasily implemented by attackers. For each forecast instance, the availability\nattack position is optimally solved by mixed-integer reformulation of the\nartificial neural network. To tackle this attack, an adversarial training\nalgorithm is proposed. In simulation, a realistic load forecasting dataset is\nconsidered and the attack performance is compared to the integrity-based\nattack. Meanwhile, the adversarial training algorithm is shown to significantly\nimprove robustness against availability attacks. All codes are available at\nhttps://github.com/xuwkk/AAA_Load_Forecast.\n", "date": "2023-01-04", "categories": []}
{"id": "2301.01731", "title": "GUAP: Graph Universal Attack Through Adversarial Patching", "abstract": "  Graph neural networks (GNNs) are a class of effective deep learning models\nfor node classification tasks; yet their predictive capability may be severely\ncompromised under adversarially designed unnoticeable perturbations to the\ngraph structure and/or node data. Most of the current work on graph adversarial\nattacks aims at lowering the overall prediction accuracy, but we argue that the\nresulting abnormal model performance may catch attention easily and invite\nquick counterattack. Moreover, attacks through modification of existing graph\ndata may be hard to conduct if good security protocols are implemented. In this\nwork, we consider an easier attack harder to be noticed, through adversarially\npatching the graph with new nodes and edges. The attack is universal: it\ntargets a single node each time and flips its connection to the same set of\npatch nodes. The attack is unnoticeable: it does not modify the predictions of\nnodes other than the target. We develop an algorithm, named GUAP, that achieves\nhigh attack success rate but meanwhile preserves the prediction accuracy. GUAP\nis fast to train by employing a sampling strategy. We demonstrate that a 5%\nsampling in each epoch yields 20x speedup in training, with only a slight\ndegradation in attack performance. Additionally, we show that the adversarial\npatch trained with the graph convolutional network transfers well to other\nGNNs, such as the graph attention network.\n", "date": "2023-01-04", "categories": []}
{"id": "2301.01197", "title": "Backdoor Attacks Against Dataset Distillation", "abstract": "  Dataset distillation has emerged as a prominent technique to improve data\nefficiency when training machine learning models. It encapsulates the knowledge\nfrom a large dataset into a smaller synthetic dataset. A model trained on this\nsmaller distilled dataset can attain comparable performance to a model trained\non the original training dataset. However, the existing dataset distillation\ntechniques mainly aim at achieving the best trade-off between resource usage\nefficiency and model utility. The security risks stemming from them have not\nbeen explored. This study performs the first backdoor attack against the models\ntrained on the data distilled by dataset distillation models in the image\ndomain. Concretely, we inject triggers into the synthetic data during the\ndistillation procedure rather than during the model training stage, where all\nprevious attacks are performed. We propose two types of backdoor attacks,\nnamely NAIVEATTACK and DOORPING. NAIVEATTACK simply adds triggers to the raw\ndata at the initial distillation phase, while DOORPING iteratively updates the\ntriggers during the entire distillation procedure. We conduct extensive\nevaluations on multiple datasets, architectures, and dataset distillation\ntechniques. Empirical evaluation shows that NAIVEATTACK achieves decent attack\nsuccess rate (ASR) scores in some cases, while DOORPING reaches higher ASR\nscores (close to 1.0) in all cases. Furthermore, we conduct a comprehensive\nablation study to analyze the factors that may affect the attack performance.\nFinally, we evaluate multiple defense mechanisms against our backdoor attacks\nand show that our attacks can practically circumvent these defense mechanisms.\n", "date": "2023-01-03", "categories": []}
{"id": "2301.01218", "title": "Tracing the Origin of Adversarial Attack for Forensic Investigation and\n  Deterrence", "abstract": "  Deep neural networks are vulnerable to adversarial attacks. In this paper, we\ntake the role of investigators who want to trace the attack and identify the\nsource, that is, the particular model which the adversarial examples are\ngenerated from. Techniques derived would aid forensic investigation of attack\nincidents and serve as deterrence to potential attacks. We consider the\nbuyers-seller setting where a machine learning model is to be distributed to\nvarious buyers and each buyer receives a slightly different copy with same\nfunctionality. A malicious buyer generates adversarial examples from a\nparticular copy $\\mathcal{M}_i$ and uses them to attack other copies. From\nthese adversarial examples, the investigator wants to identify the source\n$\\mathcal{M}_i$. To address this problem, we propose a two-stage\nseparate-and-trace framework. The model separation stage generates multiple\ncopies of a model for a same classification task. This process injects unique\ncharacteristics into each copy so that adversarial examples generated have\ndistinct and traceable features. We give a parallel structure which embeds a\n``tracer'' in each copy, and a noise-sensitive training loss to achieve this\ngoal. The tracing stage takes in adversarial examples and a few candidate\nmodels, and identifies the likely source. Based on the unique features induced\nby the noise-sensitive loss function, we could effectively trace the potential\nadversarial copy by considering the output logits from each tracer. Empirical\nresults show that it is possible to trace the origin of the adversarial example\nand the mechanism can be applied to a wide range of architectures and datasets.\n", "date": "2022-12-31", "categories": []}
{"id": "2301.01223", "title": "ExploreADV: Towards exploratory attack for Neural Networks", "abstract": "  Although deep learning has made remarkable progress in processing various\ntypes of data such as images, text and speech, they are known to be susceptible\nto adversarial perturbations: perturbations specifically designed and added to\nthe input to make the target model produce erroneous output. Most of the\nexisting studies on generating adversarial perturbations attempt to perturb the\nentire input indiscriminately. In this paper, we propose ExploreADV, a general\nand flexible adversarial attack system that is capable of modeling regional and\nimperceptible attacks, allowing users to explore various kinds of adversarial\nexamples as needed. We adapt and combine two existing boundary attack methods,\nDeepFool and Brendel\\&Bethge Attack, and propose a mask-constrained adversarial\nattack system, which generates minimal adversarial perturbations under the\npixel-level constraints, namely ``mask-constraints''. We study different ways\nof generating such mask-constraints considering the variance and importance of\nthe input features, and show that our adversarial attack system offers users\ngood flexibility to focus on sub-regions of inputs, explore imperceptible\nperturbations and understand the vulnerability of pixels/regions to adversarial\nattacks. We demonstrate our system to be effective based on extensive\nexperiments and user study.\n", "date": "2023-01-01", "categories": []}
{"id": "2210.08870", "title": "Differential Evolution based Dual Adversarial Camouflage: Fooling Human\n  Eyes and Object Detectors", "abstract": "  Recent studies reveal that deep neural network (DNN) based object detectors\nare vulnerable to adversarial attacks in the form of adding the perturbation to\nthe images, leading to the wrong output of object detectors. Most current\nexisting works focus on generating perturbed images, also called adversarial\nexamples, to fool object detectors. Though the generated adversarial examples\nthemselves can remain a certain naturalness, most of them can still be easily\nobserved by human eyes, which limits their further application in the real\nworld. To alleviate this problem, we propose a differential evolution based\ndual adversarial camouflage (DE_DAC) method, composed of two stages to fool\nhuman eyes and object detectors simultaneously. Specifically, we try to obtain\nthe camouflage texture, which can be rendered over the surface of the object.\nIn the first stage, we optimize the global texture to minimize the discrepancy\nbetween the rendered object and the scene images, making human eyes difficult\nto distinguish. In the second stage, we design three loss functions to optimize\nthe local texture, making object detectors ineffective. In addition, we\nintroduce the differential evolution algorithm to search for the near-optimal\nareas of the object to attack, improving the adversarial performance under\ncertain attack area limitations. Besides, we also study the performance of\nadaptive DE_DAC, which can be adapted to the environment. Experiments show that\nour proposed method could obtain a good trade-off between the fooling human\neyes and object detectors under multiple specific scenes and objects.\n", "date": "2022-10-17", "categories": []}
{"id": "2212.14049", "title": "Differentiable Search of Accurate and Robust Architectures", "abstract": "  Deep neural networks (DNNs) are found to be vulnerable to adversarial\nattacks, and various methods have been proposed for the defense. Among these\nmethods, adversarial training has been drawing increasing attention because of\nits simplicity and effectiveness. However, the performance of the adversarial\ntraining is greatly limited by the architectures of target DNNs, which often\nmakes the resulting DNNs with poor accuracy and unsatisfactory robustness. To\naddress this problem, we propose DSARA to automatically search for the neural\narchitectures that are accurate and robust after adversarial training. In\nparticular, we design a novel cell-based search space specially for adversarial\ntraining, which improves the accuracy and the robustness upper bound of the\nsearched architectures by carefully designing the placement of the cells and\nthe proportional relationship of the filter numbers. Then we propose a\ntwo-stage search strategy to search for both accurate and robust neural\narchitectures. At the first stage, the architecture parameters are optimized to\nminimize the adversarial loss, which makes full use of the effectiveness of the\nadversarial training in enhancing the robustness. At the second stage, the\narchitecture parameters are optimized to minimize both the natural loss and the\nadversarial loss utilizing the proposed multi-objective adversarial training\nmethod, so that the searched neural architectures are both accurate and robust.\nWe evaluate the proposed algorithm under natural data and various adversarial\nattacks, which reveals the superiority of the proposed method in terms of both\naccurate and robust architectures. We also conclude that accurate and robust\nneural architectures tend to deploy very different structures near the input\nand the output, which has great practical significance on both hand-crafting\nand automatically designing of accurate and robust neural architectures.\n", "date": "2022-12-28", "categories": []}
{"id": "2301.00364", "title": "Generalizable Black-Box Adversarial Attack with Meta Learning", "abstract": "  In the scenario of black-box adversarial attack, the target model's\nparameters are unknown, and the attacker aims to find a successful adversarial\nperturbation based on query feedback under a query budget. Due to the limited\nfeedback information, existing query-based black-box attack methods often\nrequire many queries for attacking each benign example. To reduce query cost,\nwe propose to utilize the feedback information across historical attacks,\ndubbed example-level adversarial transferability. Specifically, by treating the\nattack on each benign example as one task, we develop a meta-learning framework\nby training a meta-generator to produce perturbations conditioned on benign\nexamples. When attacking a new benign example, the meta generator can be\nquickly fine-tuned based on the feedback information of the new task as well as\na few historical attacks to produce effective perturbations. Moreover, since\nthe meta-train procedure consumes many queries to learn a generalizable\ngenerator, we utilize model-level adversarial transferability to train the\nmeta-generator on a white-box surrogate model, then transfer it to help the\nattack against the target model. The proposed framework with the two types of\nadversarial transferability can be naturally combined with any off-the-shelf\nquery-based attack methods to boost their performance, which is verified by\nextensive experiments.\n", "date": "2023-01-01", "categories": []}
{"id": "2301.00665", "title": "Targeted Phishing Campaigns using Large Scale Language Models", "abstract": "  In this research, we aim to explore the potential of natural language models\n(NLMs) such as GPT-3 and GPT-2 to generate effective phishing emails. Phishing\nemails are fraudulent messages that aim to trick individuals into revealing\nsensitive information or taking actions that benefit the attackers. We propose\na framework for evaluating the performance of NLMs in generating these types of\nemails based on various criteria, including the quality of the generated text,\nthe ability to bypass spam filters, and the success rate of tricking\nindividuals. Our evaluations show that NLMs are capable of generating phishing\nemails that are difficult to detect and that have a high success rate in\ntricking individuals, but their effectiveness varies based on the specific NLM\nand training data used. Our research indicates that NLMs could have a\nsignificant impact on the prevalence of phishing attacks and emphasizes the\nneed for further study on the ethical and security implications of using NLMs\nfor malicious purposes.\n", "date": "2022-12-30", "categories": []}
{"id": "2209.05055", "title": "CARE: Certifiably Robust Learning with Reasoning via Variational\n  Inference", "abstract": "  Despite great recent advances achieved by deep neural networks (DNNs), they\nare often vulnerable to adversarial attacks. Intensive research efforts have\nbeen made to improve the robustness of DNNs; however, most empirical defenses\ncan be adaptively attacked again, and the theoretically certified robustness is\nlimited, especially on large-scale datasets. One potential root cause of such\nvulnerabilities for DNNs is that although they have demonstrated powerful\nexpressiveness, they lack the reasoning ability to make robust and reliable\npredictions. In this paper, we aim to integrate domain knowledge to enable\nrobust learning with the reasoning paradigm. In particular, we propose a\ncertifiably robust learning with reasoning pipeline (CARE), which consists of a\nlearning component and a reasoning component. Concretely, we use a set of\nstandard DNNs to serve as the learning component to make semantic predictions,\nand we leverage the probabilistic graphical models, such as Markov logic\nnetworks (MLN), to serve as the reasoning component to enable knowledge/logic\nreasoning. However, it is known that the exact inference of MLN (reasoning) is\n#P-complete, which limits the scalability of the pipeline. To this end, we\npropose to approximate the MLN inference via variational inference based on an\nefficient expectation maximization algorithm. In particular, we leverage graph\nconvolutional networks (GCNs) to encode the posterior distribution during\nvariational inference and update the parameters of GCNs (E-step) and the\nweights of knowledge rules in MLN (M-step) iteratively. We conduct extensive\nexperiments on different datasets and show that CARE achieves significantly\nhigher certified robustness compared with the state-of-the-art baselines. We\nadditionally conducted different ablation studies to demonstrate the empirical\nrobustness of CARE and the effectiveness of different knowledge integration.\n", "date": "2022-09-12", "categories": []}
{"id": "2212.14109", "title": "Synthesis of Adversarial DDOS Attacks Using Tabular Generative\n  Adversarial Networks", "abstract": "  Network Intrusion Detection Systems (NIDS) are tools or software that are\nwidely used to maintain the computer networks and information systems keeping\nthem secure and preventing malicious traffics from penetrating into them, as\nthey flag when somebody is trying to break into the system. Best effort has\nbeen set up on these systems, and the results achieved so far are quite\nsatisfying, however, new types of attacks stand out as the technology of\nattacks keep evolving, one of these attacks are the attacks based on Generative\nAdversarial Networks (GAN) that can evade machine learning IDS leaving them\nvulnerable. This project investigates the impact of the Adversarial Attacks\nsynthesized using real DDoS attacks generated using GANs on the IDS. The\nobjective is to discover how will these systems react towards synthesized\nattacks. marking the vulnerability and weakness points of these systems so we\ncould fix them.\n", "date": "2022-12-14", "categories": []}
{"id": "2212.14115", "title": "Certifying Safety in Reinforcement Learning under Adversarial\n  Perturbation Attacks", "abstract": "  Function approximation has enabled remarkable advances in applying\nreinforcement learning (RL) techniques in environments with high-dimensional\ninputs, such as images, in an end-to-end fashion, mapping such inputs directly\nto low-level control. Nevertheless, these have proved vulnerable to small\nadversarial input perturbations. A number of approaches for improving or\ncertifying robustness of end-to-end RL to adversarial perturbations have\nemerged as a result, focusing on cumulative reward. However, what is often at\nstake in adversarial scenarios is the violation of fundamental properties, such\nas safety, rather than the overall reward that combines safety with efficiency.\nMoreover, properties such as safety can only be defined with respect to true\nstate, rather than the high-dimensional raw inputs to end-to-end policies. To\ndisentangle nominal efficiency and adversarial safety, we situate RL in\ndeterministic partially-observable Markov decision processes (POMDPs) with the\ngoal of maximizing cumulative reward subject to safety constraints. We then\npropose a partially-supervised reinforcement learning (PSRL) framework that\ntakes advantage of an additional assumption that the true state of the POMDP is\nknown at training time. We present the first approach for certifying safety of\nPSRL policies under adversarial input perturbations, and two adversarial\ntraining approaches that make direct use of PSRL. Our experiments demonstrate\nboth the efficacy of the proposed approach for certifying safety in adversarial\nenvironments, and the value of the PSRL framework coupled with adversarial\ntraining in improving certified safety while preserving high nominal reward and\nhigh-quality predictions of true state.\n", "date": "2022-12-28", "categories": []}
{"id": "2212.14315", "title": "\"Real Attackers Don't Compute Gradients\": Bridging the Gap Between\n  Adversarial ML Research and Practice", "abstract": "  Recent years have seen a proliferation of research on adversarial machine\nlearning. Numerous papers demonstrate powerful algorithmic attacks against a\nwide variety of machine learning (ML) models, and numerous other papers propose\ndefenses that can withstand most attacks. However, abundant real-world evidence\nsuggests that actual attackers use simple tactics to subvert ML-driven systems,\nand as a result security practitioners have not prioritized adversarial ML\ndefenses.\n  Motivated by the apparent gap between researchers and practitioners, this\nposition paper aims to bridge the two domains. We first present three\nreal-world case studies from which we can glean practical insights unknown or\nneglected in research. Next we analyze all adversarial ML papers recently\npublished in top security conferences, highlighting positive trends and blind\nspots. Finally, we state positions on precise and cost-driven threat modeling,\ncollaboration between industry and academia, and reproducible research. We\nbelieve that our positions, if adopted, will increase the real-world impact of\nfuture endeavours in adversarial ML, bringing both researchers and\npractitioners closer to their shared goal of improving the security of ML\nsystems.\n", "date": "2022-12-29", "categories": []}
{"id": "2212.14677", "title": "Adversarial attacks and defenses on ML- and hardware-based IoT device\n  fingerprinting and identification", "abstract": "  In the last years, the number of IoT devices deployed has suffered an\nundoubted explosion, reaching the scale of billions. However, some new\ncybersecurity issues have appeared together with this development. Some of\nthese issues are the deployment of unauthorized devices, malicious code\nmodification, malware deployment, or vulnerability exploitation. This fact has\nmotivated the requirement for new device identification mechanisms based on\nbehavior monitoring. Besides, these solutions have recently leveraged Machine\nand Deep Learning techniques due to the advances in this field and the increase\nin processing capabilities. In contrast, attackers do not stay stalled and have\ndeveloped adversarial attacks focused on context modification and ML/DL\nevaluation evasion applied to IoT device identification solutions. This work\nexplores the performance of hardware behavior-based individual device\nidentification, how it is affected by possible context- and ML/DL-focused\nattacks, and how its resilience can be improved using defense techniques. In\nthis sense, it proposes an LSTM-CNN architecture based on hardware performance\nbehavior for individual device identification. Then, previous techniques have\nbeen compared with the proposed architecture using a hardware performance\ndataset collected from 45 Raspberry Pi devices running identical software. The\nLSTM-CNN improves previous solutions achieving a +0.96 average F1-Score and 0.8\nminimum TPR for all devices. Afterward, context- and ML/DL-focused adversarial\nattacks were applied against the previous model to test its robustness. A\ntemperature-based context attack was not able to disrupt the identification.\nHowever, some ML/DL state-of-the-art evasion attacks were successful. Finally,\nadversarial training and model distillation defense techniques are selected to\nimprove the model resilience to evasion attacks, without degrading its\nperformance.\n", "date": "2022-12-30", "categories": []}
{"id": "2212.14875", "title": "Guidance Through Surrogate: Towards a Generic Diagnostic Attack", "abstract": "  Adversarial training is an effective approach to make deep neural networks\nrobust against adversarial attacks. Recently, different adversarial training\ndefenses are proposed that not only maintain a high clean accuracy but also\nshow significant robustness against popular and well studied adversarial\nattacks such as PGD. High adversarial robustness can also arise if an attack\nfails to find adversarial gradient directions, a phenomenon known as `gradient\nmasking'. In this work, we analyse the effect of label smoothing on adversarial\ntraining as one of the potential causes of gradient masking. We then develop a\nguided mechanism to avoid local minima during attack optimization, leading to a\nnovel attack dubbed Guided Projected Gradient Attack (G-PGA). Our attack\napproach is based on a `match and deceive' loss that finds optimal adversarial\ndirections through guidance from a surrogate model. Our modified attack does\nnot require random restarts, large number of attack iterations or search for an\noptimal step-size. Furthermore, our proposed G-PGA is generic, thus it can be\ncombined with an ensemble attack strategy as we demonstrate for the case of\nAuto-Attack, leading to efficiency and convergence speed improvements. More\nthan an effective attack, G-PGA can be used as a diagnostic tool to reveal\nelusive robustness due to gradient masking in adversarial defenses.\n", "date": "2022-12-30", "categories": []}
{"id": "2212.13607", "title": "EDoG: Adversarial Edge Detection For Graph Neural Networks", "abstract": "  Graph Neural Networks (GNNs) have been widely applied to different tasks such\nas bioinformatics, drug design, and social networks. However, recent studies\nhave shown that GNNs are vulnerable to adversarial attacks which aim to mislead\nthe node or subgraph classification prediction by adding subtle perturbations.\nDetecting these attacks is challenging due to the small magnitude of\nperturbation and the discrete nature of graph data. In this paper, we propose a\ngeneral adversarial edge detection pipeline EDoG without requiring knowledge of\nthe attack strategies based on graph generation. Specifically, we propose a\nnovel graph generation approach combined with link prediction to detect\nsuspicious adversarial edges. To effectively train the graph generative model,\nwe sample several sub-graphs from the given graph data. We show that since the\nnumber of adversarial edges is usually low in practice, with low probability\nthe sampled sub-graphs will contain adversarial edges based on the union bound.\nIn addition, considering the strong attacks which perturb a large number of\nedges, we propose a set of novel features to perform outlier detection as the\npreprocessing for our detection. Extensive experimental results on three\nreal-world graph datasets including a private transaction rule dataset from a\nmajor company and two types of synthetic graphs with controlled properties show\nthat EDoG can achieve above 0.8 AUC against four state-of-the-art unseen attack\nstrategies without requiring any knowledge about the attack type; and around\n0.85 with knowledge of the attack type. EDoG significantly outperforms\ntraditional malicious edge detection baselines. We also show that an adaptive\nattack with full knowledge of our detection pipeline is difficult to bypass it.\n", "date": "2022-12-27", "categories": []}
{"id": "2212.13700", "title": "Publishing Efficient On-device Models Increases Adversarial\n  Vulnerability", "abstract": "  Recent increases in the computational demands of deep neural networks (DNNs)\nhave sparked interest in efficient deep learning mechanisms, e.g., quantization\nor pruning. These mechanisms enable the construction of a small, efficient\nversion of commercial-scale models with comparable accuracy, accelerating their\ndeployment to resource-constrained devices.\n  In this paper, we study the security considerations of publishing on-device\nvariants of large-scale models. We first show that an adversary can exploit\non-device models to make attacking the large models easier. In evaluations\nacross 19 DNNs, by exploiting the published on-device models as a transfer\nprior, the adversarial vulnerability of the original commercial-scale models\nincreases by up to 100x. We then show that the vulnerability increases as the\nsimilarity between a full-scale and its efficient model increase. Based on the\ninsights, we propose a defense, $similarity$-$unpairing$, that fine-tunes\non-device models with the objective of reducing the similarity. We evaluated\nour defense on all the 19 DNNs and found that it reduces the transferability up\nto 90% and the number of queries required by a factor of 10-100x. Our results\nsuggest that further research is needed on the security (or even privacy)\nthreats caused by publishing those efficient siblings.\n", "date": "2022-12-28", "categories": []}
{"id": "2212.13989", "title": "AdvCat: Domain-Agnostic Robustness Assessment for Cybersecurity-Critical\n  Applications with Categorical Inputs", "abstract": "  Machine Learning-as-a-Service systems (MLaaS) have been largely developed for\ncybersecurity-critical applications, such as detecting network intrusions and\nfake news campaigns. Despite effectiveness, their robustness against\nadversarial attacks is one of the key trust concerns for MLaaS deployment. We\nare thus motivated to assess the adversarial robustness of the Machine Learning\nmodels residing at the core of these security-critical applications with\ncategorical inputs. Previous research efforts on accessing model robustness\nagainst manipulation of categorical inputs are specific to use cases and\nheavily depend on domain knowledge, or require white-box access to the target\nML model. Such limitations prevent the robustness assessment from being as a\ndomain-agnostic service provided to various real-world applications. We propose\na provably optimal yet computationally highly efficient adversarial robustness\nassessment protocol for a wide band of ML-driven cybersecurity-critical\napplications. We demonstrate the use of the domain-agnostic robustness\nassessment method with substantial experimental study on fake news detection\nand intrusion detection problems.\n", "date": "2022-12-13", "categories": []}
{"id": "2212.12732", "title": "Frequency Regularization for Improving Adversarial Robustness", "abstract": "  Deep neural networks are incredibly vulnerable to crafted,\nhuman-imperceptible adversarial perturbations. Although adversarial training\n(AT) has proven to be an effective defense approach, we find that the\nAT-trained models heavily rely on the input low-frequency content for judgment,\naccounting for the low standard accuracy. To close the large gap between the\nstandard and robust accuracies during AT, we investigate the frequency\ndifference between clean and adversarial inputs, and propose a frequency\nregularization (FR) to align the output difference in the spectral domain.\nBesides, we find Stochastic Weight Averaging (SWA), by smoothing the kernels\nover epochs, further improves the robustness. Among various defense schemes,\nour method achieves the strongest robustness against attacks by PGD-20, C\\&W\nand Autoattack, on a WideResNet trained on CIFAR-10 without any extra data.\n", "date": "2022-12-24", "categories": []}
{"id": "2212.12995", "title": "Simultaneously Optimizing Perturbations and Positions for Black-box\n  Adversarial Patch Attacks", "abstract": "  Adversarial patch is an important form of real-world adversarial attack that\nbrings serious risks to the robustness of deep neural networks. Previous\nmethods generate adversarial patches by either optimizing their perturbation\nvalues while fixing the pasting position or manipulating the position while\nfixing the patch's content. This reveals that the positions and perturbations\nare both important to the adversarial attack. For that, in this paper, we\npropose a novel method to simultaneously optimize the position and perturbation\nfor an adversarial patch, and thus obtain a high attack success rate in the\nblack-box setting. Technically, we regard the patch's position, the\npre-designed hyper-parameters to determine the patch's perturbations as the\nvariables, and utilize the reinforcement learning framework to simultaneously\nsolve for the optimal solution based on the rewards obtained from the target\nmodel with a small number of queries. Extensive experiments are conducted on\nthe Face Recognition (FR) task, and results on four representative FR models\nshow that our method can significantly improve the attack success rate and\nquery efficiency. Besides, experiments on the commercial FR service and\nphysical environments confirm its practical application value. We also extend\nour method to the traffic sign recognition task to verify its generalization\nability.\n", "date": "2022-12-26", "categories": []}
{"id": "2209.13353", "title": "Suppress with a Patch: Revisiting Universal Adversarial Patch Attacks\n  against Object Detection", "abstract": "  Adversarial patch-based attacks aim to fool a neural network with an\nintentionally generated noise, which is concentrated in a particular region of\nan input image. In this work, we perform an in-depth analysis of different\npatch generation parameters, including initialization, patch size, and\nespecially positioning a patch in an image during training. We focus on the\nobject vanishing attack and run experiments with YOLOv3 as a model under attack\nin a white-box setting and use images from the COCO dataset. Our experiments\nhave shown, that inserting a patch inside a window of increasing size during\ntraining leads to a significant increase in attack strength compared to a fixed\nposition. The best results were obtained when a patch was positioned randomly\nduring training, while patch position additionally varied within a batch.\n", "date": "2022-09-27", "categories": []}
{"id": "2212.11209", "title": "A Theoretical Study of The Effects of Adversarial Attacks on Sparse\n  Regression", "abstract": "  This paper analyzes $\\ell_1$ regularized linear regression under the\nchallenging scenario of having only adversarially corrupted data for training.\nWe use the primal-dual witness paradigm to provide provable performance\nguarantees for the support of the estimated regression parameter vector to\nmatch the actual parameter. Our theoretical analysis shows the\ncounter-intuitive result that an adversary can influence sample complexity by\ncorrupting the irrelevant features, i.e., those corresponding to zero\ncoefficients of the regression parameter vector, which, consequently, do not\naffect the dependent variable. As any adversarially robust algorithm has its\nlimitations, our theoretical analysis identifies the regimes under which the\nlearning algorithm and adversary can dominate over each other. It helps us to\nanalyze these fundamental limits and address critical scientific questions of\nwhich parameters (like mutual incoherence, the maximum and minimum eigenvalue\nof the covariance matrix, and the budget of adversarial perturbation) play a\nrole in the high or low probability of success of the LASSO algorithm. Also,\nthe derived sample complexity is logarithmic with respect to the size of the\nregression parameter vector, and our theoretical claims are validated by\nempirical analysis on synthetic and real-world datasets.\n", "date": "2022-12-21", "categories": []}
{"id": "2212.11751", "title": "Mind Your Heart: Stealthy Backdoor Attack on Dynamic Deep Neural Network\n  in Edge Computing", "abstract": "  Transforming off-the-shelf deep neural network (DNN) models into dynamic\nmulti-exit architectures can achieve inference and transmission efficiency by\nfragmenting and distributing a large DNN model in edge computing scenarios\n(e.g., edge devices and cloud servers). In this paper, we propose a novel\nbackdoor attack specifically on the dynamic multi-exit DNN models.\nParticularly, we inject a backdoor by poisoning one DNN model's shallow hidden\nlayers targeting not this vanilla DNN model but only its dynamically deployed\nmulti-exit architectures. Our backdoored vanilla model behaves normally on\nperformance and cannot be activated even with the correct trigger. However, the\nbackdoor will be activated when the victims acquire this model and transform it\ninto a dynamic multi-exit architecture at their deployment. We conduct\nextensive experiments to prove the effectiveness of our attack on three\nstructures (ResNet-56, VGG-16, and MobileNet) with four datasets (CIFAR-10,\nSVHN, GTSRB, and Tiny-ImageNet) and our backdoor is stealthy to evade multiple\nstate-of-the-art backdoor detection or removal methods.\n", "date": "2022-12-22", "categories": []}
{"id": "2212.11760", "title": "Aliasing is a Driver of Adversarial Attacks", "abstract": "  Aliasing is a highly important concept in signal processing, as careful\nconsideration of resolution changes is essential in ensuring transmission and\nprocessing quality of audio, image, and video. Despite this, up until recently\naliasing has received very little consideration in Deep Learning, with all\ncommon architectures carelessly sub-sampling without considering aliasing\neffects. In this work, we investigate the hypothesis that the existence of\nadversarial perturbations is due in part to aliasing in neural networks. Our\nultimate goal is to increase robustness against adversarial attacks using\nexplainable, non-trained, structural changes only, derived from aliasing first\nprinciples. Our contributions are the following. First, we establish a\nsufficient condition for no aliasing for general image transformations. Next,\nwe study sources of aliasing in common neural network layers, and derive simple\nmodifications from first principles to eliminate or reduce it. Lastly, our\nexperimental results show a solid link between anti-aliasing and adversarial\nattacks. Simply reducing aliasing already results in more robust classifiers,\nand combining anti-aliasing with robust training out-performs solo robust\ntraining on $L_2$ attacks with none or minimal losses in performance on\n$L_{\\infty}$ attacks.\n", "date": "2022-12-22", "categories": []}
{"id": "2212.11778", "title": "Adversarial Machine Learning and Defense Game for NextG Signal\n  Classification with Deep Learning", "abstract": "  This paper presents a game-theoretic framework to study the interactions of\nattack and defense for deep learning-based NextG signal classification. NextG\nsystems such as the one envisioned for a massive number of IoT devices can\nemploy deep neural networks (DNNs) for various tasks such as user equipment\nidentification, physical layer authentication, and detection of incumbent users\n(such as in the Citizens Broadband Radio Service (CBRS) band). By training\nanother DNN as the surrogate model, an adversary can launch an inference\n(exploratory) attack to learn the behavior of the victim model, predict\nsuccessful operation modes (e.g., channel access), and jam them. A defense\nmechanism can increase the adversary's uncertainty by introducing controlled\nerrors in the victim model's decisions (i.e., poisoning the adversary's\ntraining data). This defense is effective against an attack but reduces the\nperformance when there is no attack. The interactions between the defender and\nthe adversary are formulated as a non-cooperative game, where the defender\nselects the probability of defending or the defense level itself (i.e., the\nratio of falsified decisions) and the adversary selects the probability of\nattacking. The defender's objective is to maximize its reward (e.g., throughput\nor transmission success ratio), whereas the adversary's objective is to\nminimize this reward and its attack cost. The Nash equilibrium strategies are\ndetermined as operation modes such that no player can unilaterally improve its\nutility given the other's strategy is fixed. A fictitious play is formulated\nfor each player to play the game repeatedly in response to the empirical\nfrequency of the opponent's actions. The performance in Nash equilibrium is\ncompared to the fixed attack and defense cases, and the resilience of NextG\nsignal classification against attacks is quantified.\n", "date": "2022-12-22", "categories": []}
{"id": "2212.06822", "title": "Adversarial Attacks and Defences for Skin Cancer Classification", "abstract": "  There has been a concurrent significant improvement in the medical images\nused to facilitate diagnosis and the performance of machine learning techniques\nto perform tasks such as classification, detection, and segmentation in recent\nyears. As a result, a rapid increase in the usage of such systems can be\nobserved in the healthcare industry, for instance in the form of medical image\nclassification systems, where these models have achieved diagnostic parity with\nhuman physicians. One such application where this can be observed is in\ncomputer vision tasks such as the classification of skin lesions in\ndermatoscopic images. However, as stakeholders in the healthcare industry, such\nas insurance companies, continue to invest extensively in machine learning\ninfrastructure, it becomes increasingly important to understand the\nvulnerabilities in such systems. Due to the highly critical nature of the tasks\nbeing carried out by these machine learning models, it is necessary to analyze\ntechniques that could be used to take advantage of these vulnerabilities and\nmethods to defend against them. This paper explores common adversarial attack\ntechniques. The Fast Sign Gradient Method and Projected Descent Gradient are\nused against a Convolutional Neural Network trained to classify dermatoscopic\nimages of skin lesions. Following that, it also discusses one of the most\npopular adversarial defense techniques, adversarial training. The performance\nof the model that has been trained on adversarial examples is then tested\nagainst the previously mentioned attacks, and recommendations to improve neural\nnetworks robustness are thus provided based on the results of the experiment.\n", "date": "2022-12-13", "categories": []}
{"id": "2212.11005", "title": "Revisiting Residual Networks for Adversarial Robustness: An\n  Architectural Perspective", "abstract": "  Efforts to improve the adversarial robustness of convolutional neural\nnetworks have primarily focused on developing more effective adversarial\ntraining methods. In contrast, little attention was devoted to analyzing the\nrole of architectural elements (such as topology, depth, and width) on\nadversarial robustness. This paper seeks to bridge this gap and present a\nholistic study on the impact of architectural design on adversarial robustness.\nWe focus on residual networks and consider architecture design at the block\nlevel, i.e., topology, kernel size, activation, and normalization, as well as\nat the network scaling level, i.e., depth and width of each block in the\nnetwork. In both cases, we first derive insights through systematic ablative\nexperiments. Then we design a robust residual block, dubbed RobustResBlock, and\na compound scaling rule, dubbed RobustScaling, to distribute depth and width at\nthe desired FLOP count. Finally, we combine RobustResBlock and RobustScaling\nand present a portfolio of adversarially robust residual networks,\nRobustResNets, spanning a broad spectrum of model capacities. Experimental\nvalidation across multiple datasets and adversarial attacks demonstrate that\nRobustResNets consistently outperform both the standard WRNs and other existing\nrobust architectures, achieving state-of-the-art AutoAttack robust accuracy of\n61.1% without additional data and 63.7% with 500K external data while being\n$2\\times$ more compact in terms of parameters. Code is available at \\url{\nhttps://github.com/zhichao-lu/robust-residual-network}\n", "date": "2022-12-21", "categories": []}
{"id": "2212.09000", "title": "Confidence-aware Training of Smoothed Classifiers for Certified\n  Robustness", "abstract": "  Any classifier can be \"smoothed out\" under Gaussian noise to build a new\nclassifier that is provably robust to $\\ell_2$-adversarial perturbations, viz.,\nby averaging its predictions over the noise via randomized smoothing. Under the\nsmoothed classifiers, the fundamental trade-off between accuracy and\n(adversarial) robustness has been well evidenced in the literature: i.e.,\nincreasing the robustness of a classifier for an input can be at the expense of\ndecreased accuracy for some other inputs. In this paper, we propose a simple\ntraining method leveraging this trade-off to obtain robust smoothed\nclassifiers, in particular, through a sample-wise control of robustness over\nthe training samples. We make this control feasible by using \"accuracy under\nGaussian noise\" as an easy-to-compute proxy of adversarial robustness for an\ninput. Specifically, we differentiate the training objective depending on this\nproxy to filter out samples that are unlikely to benefit from the worst-case\n(adversarial) objective. Our experiments show that the proposed method, despite\nits simplicity, consistently exhibits improved certified robustness upon\nstate-of-the-art training methods. Somewhat surprisingly, we find these\nimprovements persist even for other notions of robustness, e.g., to various\ntypes of common corruptions.\n", "date": "2022-12-18", "categories": []}
{"id": "2212.09994", "title": "Towards Robustness of Text-to-SQL Models Against Natural and Realistic\n  Adversarial Table Perturbation", "abstract": "  The robustness of Text-to-SQL parsers against adversarial perturbations plays\na crucial role in delivering highly reliable applications. Previous studies\nalong this line primarily focused on perturbations in the natural language\nquestion side, neglecting the variability of tables. Motivated by this, we\npropose the Adversarial Table Perturbation (ATP) as a new attacking paradigm to\nmeasure the robustness of Text-to-SQL models. Following this proposition, we\ncurate ADVETA, the first robustness evaluation benchmark featuring natural and\nrealistic ATPs. All tested state-of-the-art models experience dramatic\nperformance drops on ADVETA, revealing models' vulnerability in real-world\npractices. To defend against ATP, we build a systematic adversarial training\nexample generation framework tailored for better contextualization of tabular\ndata. Experiments show that our approach not only brings the best robustness\nimprovement against table-side perturbations but also substantially empowers\nmodels against NL-side perturbations. We release our benchmark and code at:\nhttps://github.com/microsoft/ContextualSP.\n", "date": "2022-12-20", "categories": []}
{"id": "2212.10006", "title": "Multi-head Uncertainty Inference for Adversarial Attack Detection", "abstract": "  Deep neural networks (DNNs) are sensitive and susceptible to tiny\nperturbation by adversarial attacks which causes erroneous predictions. Various\nmethods, including adversarial defense and uncertainty inference (UI), have\nbeen developed in recent years to overcome the adversarial attacks. In this\npaper, we propose a multi-head uncertainty inference (MH-UI) framework for\ndetecting adversarial attack examples. We adopt a multi-head architecture with\nmultiple prediction heads (i.e., classifiers) to obtain predictions from\ndifferent depths in the DNNs and introduce shallow information for the UI.\nUsing independent heads at different depths, the normalized predictions are\nassumed to follow the same Dirichlet distribution, and we estimate distribution\nparameter of it by moment matching. Cognitive uncertainty brought by the\nadversarial attacks will be reflected and amplified on the distribution.\nExperimental results show that the proposed MH-UI framework can outperform all\nthe referred UI methods in the adversarial attack detection task with different\nsettings.\n", "date": "2022-12-20", "categories": []}
{"id": "2212.10103", "title": "VSVC: Backdoor attack against Keyword Spotting based on Voiceprint\n  Selection and Voice Conversion", "abstract": "  Keyword spotting (KWS) based on deep neural networks (DNNs) has achieved\nmassive success in voice control scenarios. However, training of such DNN-based\nKWS systems often requires significant data and hardware resources.\nManufacturers often entrust this process to a third-party platform. This makes\nthe training process uncontrollable, where attackers can implant backdoors in\nthe model by manipulating third-party training data. An effective backdoor\nattack can force the model to make specified judgments under certain\nconditions, i.e., triggers. In this paper, we design a backdoor attack scheme\nbased on Voiceprint Selection and Voice Conversion, abbreviated as VSVC.\nExperimental results demonstrated that VSVC is feasible to achieve an average\nattack success rate close to 97% in four victim models when poisoning less than\n1% of the training data.\n", "date": "2022-12-20", "categories": []}
{"id": "2212.10438", "title": "Is Semantic Communications Secure? A Tale of Multi-Domain Adversarial\n  Attacks", "abstract": "  Semantic communications seeks to transfer information from a source while\nconveying a desired meaning to its destination. We model the\ntransmitter-receiver functionalities as an autoencoder followed by a task\nclassifier that evaluates the meaning of the information conveyed to the\nreceiver. The autoencoder consists of an encoder at the transmitter to jointly\nmodel source coding, channel coding, and modulation, and a decoder at the\nreceiver to jointly model demodulation, channel decoding and source decoding.\nBy augmenting the reconstruction loss with a semantic loss, the two deep neural\nnetworks (DNNs) of this encoder-decoder pair are interactively trained with the\nDNN of the semantic task classifier. This approach effectively captures the\nlatent feature space and reliably transfers compressed feature vectors with a\nsmall number of channel uses while keeping the semantic loss low. We identify\nthe multi-domain security vulnerabilities of using the DNNs for semantic\ncommunications. Based on adversarial machine learning, we introduce test-time\n(targeted and non-targeted) adversarial attacks on the DNNs by manipulating\ntheir inputs at different stages of semantic communications. As a computer\nvision attack, small perturbations are injected to the images at the input of\nthe transmitter's encoder. As a wireless attack, small perturbations signals\nare transmitted to interfere with the input of the receiver's decoder. By\nlaunching these stealth attacks individually or more effectively in a combined\nform as a multi-domain attack, we show that it is possible to change the\nsemantics of the transferred information even when the reconstruction loss\nremains low. These multi-domain adversarial attacks pose as a serious threat to\nthe semantics of information transfer (with larger impact than conventional\njamming) and raise the need of defense methods for the safe adoption of\nsemantic communications.\n", "date": "2022-12-20", "categories": []}
{"id": "2212.09035", "title": "Minimizing Maximum Model Discrepancy for Transferable Black-box Targeted\n  Attacks", "abstract": "  In this work, we study the black-box targeted attack problem from the model\ndiscrepancy perspective. On the theoretical side, we present a generalization\nerror bound for black-box targeted attacks, which gives a rigorous theoretical\nanalysis for guaranteeing the success of the attack. We reveal that the attack\nerror on a target model mainly depends on empirical attack error on the\nsubstitute model and the maximum model discrepancy among substitute models. On\nthe algorithmic side, we derive a new algorithm for black-box targeted attacks\nbased on our theoretical analysis, in which we additionally minimize the\nmaximum model discrepancy(M3D) of the substitute models when training the\ngenerator to generate adversarial examples. In this way, our model is capable\nof crafting highly transferable adversarial examples that are robust to the\nmodel variation, thus improving the success rate for attacking the black-box\nmodel. We conduct extensive experiments on the ImageNet dataset with different\nclassification models, and our proposed approach outperforms existing\nstate-of-the-art methods by a significant margin. Our codes will be released.\n", "date": "2022-12-18", "categories": []}
{"id": "2212.09155", "title": "Estimating the Adversarial Robustness of Attributions in Text with\n  Transformers", "abstract": "  Explanations are crucial parts of deep neural network (DNN) classifiers. In\nhigh stakes applications, faithful and robust explanations are important to\nunderstand and gain trust in DNN classifiers. However, recent work has shown\nthat state-of-the-art attribution methods in text classifiers are susceptible\nto imperceptible adversarial perturbations that alter explanations\nsignificantly while maintaining the correct prediction outcome. If undetected,\nthis can critically mislead the users of DNNs. Thus, it is crucial to\nunderstand the influence of such adversarial perturbations on the networks'\nexplanations and their perceptibility. In this work, we establish a novel\ndefinition of attribution robustness (AR) in text classification, based on\nLipschitz continuity. Crucially, it reflects both attribution change induced by\nadversarial input alterations and perceptibility of such alterations. Moreover,\nwe introduce a wide set of text similarity measures to effectively capture\nlocality between two text samples and imperceptibility of adversarial\nperturbations in text. We then propose our novel TransformerExplanationAttack\n(TEA), a strong adversary that provides a tight estimation for attribution\nrobustness in text classification. TEA uses state-of-the-art language models to\nextract word substitutions that result in fluent, contextual adversarial\nsamples. Finally, with experiments on several text classification\narchitectures, we show that TEA consistently outperforms current\nstate-of-the-art AR estimators, yielding perturbations that alter explanations\nto a greater extent while being more fluent and less perceptible.\n", "date": "2022-12-18", "categories": []}
{"id": "2212.09254", "title": "TextGrad: Advancing Robustness Evaluation in NLP by Gradient-Driven\n  Optimization", "abstract": "  Robustness evaluation against adversarial examples has become increasingly\nimportant to unveil the trustworthiness of the prevailing deep models in\nnatural language processing (NLP). However, in contrast to the computer vision\ndomain where the first-order projected gradient descent (PGD) is used as the\nbenchmark approach to generate adversarial examples for robustness evaluation,\nthere lacks a principled first-order gradient-based robustness evaluation\nframework in NLP. The emerging optimization challenges lie in 1) the discrete\nnature of textual inputs together with the strong coupling between the\nperturbation location and the actual content, and 2) the additional constraint\nthat the perturbed text should be fluent and achieve a low perplexity under a\nlanguage model. These challenges make the development of PGD-like NLP attacks\ndifficult. To bridge the gap, we propose TextGrad, a new attack generator using\ngradient-driven optimization, supporting high-accuracy and high-quality\nassessment of adversarial robustness in NLP. Specifically, we address the\naforementioned challenges in a unified optimization framework. And we develop\nan effective convex relaxation method to co-optimize the continuously-relaxed\nsite selection and perturbation variables and leverage an effective sampling\nmethod to establish an accurate mapping from the continuous optimization\nvariables to the discrete textual perturbations. Moreover, as a first-order\nattack generation method, TextGrad can be baked into adversarial training to\nfurther improve the robustness of NLP models. Extensive experiments are\nprovided to demonstrate the effectiveness of TextGrad not only in attack\ngeneration for robustness evaluation but also in adversarial defense.\n", "date": "2022-12-19", "categories": []}
{"id": "2212.08121", "title": "Backdoor Attack Detection in Computer Vision by Applying Matrix\n  Factorization on the Weights of Deep Networks", "abstract": "  The increasing importance of both deep neural networks (DNNs) and cloud\nservices for training them means that bad actors have more incentive and\nopportunity to insert backdoors to alter the behavior of trained models. In\nthis paper, we introduce a novel method for backdoor detection that extracts\nfeatures from pre-trained DNN's weights using independent vector analysis (IVA)\nfollowed by a machine learning classifier. In comparison to other detection\ntechniques, this has a number of benefits, such as not requiring any training\ndata, being applicable across domains, operating with a wide range of network\narchitectures, not assuming the nature of the triggers used to change network\nbehavior, and being highly scalable. We discuss the detection pipeline, and\nthen demonstrate the results on two computer vision datasets regarding image\nclassification and object detection. Our method outperforms the competing\nalgorithms in terms of efficiency and is more accurate, helping to ensure the\nsafe application of deep learning and AI.\n", "date": "2022-12-15", "categories": []}
{"id": "2212.08130", "title": "On Evaluating Adversarial Robustness of Chest X-ray Classification:\n  Pitfalls and Best Practices", "abstract": "  Vulnerability to adversarial attacks is a well-known weakness of Deep Neural\nNetworks. While most of the studies focus on natural images with standardized\nbenchmarks like ImageNet and CIFAR, little research has considered real world\napplications, in particular in the medical domain. Our research shows that,\ncontrary to previous claims, robustness of chest x-ray classification is much\nharder to evaluate and leads to very different assessments based on the\ndataset, the architecture and robustness metric. We argue that previous studies\ndid not take into account the peculiarity of medical diagnosis, like the\nco-occurrence of diseases, the disagreement of labellers (domain experts), the\nthreat model of the attacks and the risk implications for each successful\nattack.\n  In this paper, we discuss the methodological foundations, review the pitfalls\nand best practices, and suggest new methodological considerations for\nevaluating the robustness of chest xray classification models. Our evaluation\non 3 datasets, 7 models, and 18 diseases is the largest evaluation of\nrobustness of chest x-ray classification models.\n", "date": "2022-12-15", "categories": []}
{"id": "2212.07815", "title": "Adversarially Robust Video Perception by Seeing Motion", "abstract": "  Despite their excellent performance, state-of-the-art computer vision models\noften fail when they encounter adversarial examples. Video perception models\ntend to be more fragile under attacks, because the adversary has more places to\nmanipulate in high-dimensional data. In this paper, we find one reason for\nvideo models' vulnerability is that they fail to perceive the correct motion\nunder adversarial perturbations. Inspired by the extensive evidence that motion\nis a key factor for the human visual system, we propose to correct what the\nmodel sees by restoring the perceived motion information. Since motion\ninformation is an intrinsic structure of the video data, recovering motion\nsignals can be done at inference time without any human annotation, which\nallows the model to adapt to unforeseen, worst-case inputs. Visualizations and\nempirical experiments on UCF-101 and HMDB-51 datasets show that restoring\nmotion information in deep vision models improves adversarial robustness. Even\nunder adaptive attacks where the adversary knows our defense, our algorithm is\nstill effective. Our work provides new insight into robust video perception\nalgorithms by using intrinsic structures from the data. Our webpage is\navailable at https://motion4robust.cs.columbia.edu.\n", "date": "2022-12-13", "categories": []}
{"id": "2212.07992", "title": "Alternating Objectives Generates Stronger PGD-Based Adversarial Attacks", "abstract": "  Designing powerful adversarial attacks is of paramount importance for the\nevaluation of $\\ell_p$-bounded adversarial defenses. Projected Gradient Descent\n(PGD) is one of the most effective and conceptually simple algorithms to\ngenerate such adversaries. The search space of PGD is dictated by the steepest\nascent directions of an objective. Despite the plethora of objective function\nchoices, there is no universally superior option and robustness overestimation\nmay arise from ill-suited objective selection. Driven by this observation, we\npostulate that the combination of different objectives through a simple loss\nalternating scheme renders PGD more robust towards design choices. We\nexperimentally verify this assertion on a synthetic-data example and by\nevaluating our proposed method across 25 different $\\ell_{\\infty}$-robust\nmodels and 3 datasets. The performance improvement is consistent, when compared\nto the single loss counterparts. In the CIFAR-10 dataset, our strongest\nadversarial attack outperforms all of the white-box components of AutoAttack\n(AA) ensemble, as well as the most powerful attacks existing on the literature,\nachieving state-of-the-art results in the computational budget of our study\n($T=100$, no restarts).\n", "date": "2022-12-15", "categories": []}
{"id": "2212.06836", "title": "Towards Efficient and Domain-Agnostic Evasion Attack with\n  High-dimensional Categorical Inputs", "abstract": "  Our work targets at searching feasible adversarial perturbation to attack a\nclassifier with high-dimensional categorical inputs in a domain-agnostic\nsetting. This is intrinsically an NP-hard knapsack problem where the\nexploration space becomes explosively larger as the feature dimension\nincreases. Without the help of domain knowledge, solving this problem via\nheuristic method, such as Branch-and-Bound, suffers from exponential\ncomplexity, yet can bring arbitrarily bad attack results. We address the\nchallenge via the lens of multi-armed bandit based combinatorial search. Our\nproposed method, namely FEAT, treats modifying each categorical feature as\npulling an arm in multi-armed bandit programming. Our objective is to achieve\nhighly efficient and effective attack using an Orthogonal Matching Pursuit\n(OMP)-enhanced Upper Confidence Bound (UCB) exploration strategy. Our\ntheoretical analysis bounding the regret gap of FEAT guarantees its practical\nattack performance. In empirical analysis, we compare FEAT with other\nstate-of-the-art domain-agnostic attack methods over various real-world\ncategorical data sets of different applications. Substantial experimental\nobservations confirm the expected efficiency and attack effectiveness of FEAT\napplied in different application scenarios. Our work further hints the\napplicability of FEAT for assessing the adversarial vulnerability of\nclassification systems with high-dimensional categorical inputs.\n", "date": "2022-12-13", "categories": []}
{"id": "2212.07283", "title": "Generative Robust Classification", "abstract": "  Training adversarially robust discriminative (i.e., softmax) classifier has\nbeen the dominant approach to robust classification. Building on recent work on\nadversarial training (AT)-based generative models, we investigate using AT to\nlearn unnormalized class-conditional density models and then performing\ngenerative robust classification. Our result shows that, under the condition of\nsimilar model capacities, the generative robust classifier achieves comparable\nperformance to a baseline softmax robust classifier when the test data is clean\nor when the test perturbation is of limited size, and much better performance\nwhen the test perturbation size exceeds the training perturbation size. The\ngenerative classifier is also able to generate samples or counterfactuals that\nmore closely resemble the training data, suggesting that the generative\nclassifier can better capture the class-conditional distributions. In contrast\nto standard discriminative adversarial training where advanced data\naugmentation techniques are only effective when combined with weight averaging,\nwe find it straightforward to apply advanced data augmentation to achieve\nbetter robustness in our approach. Our result suggests that the generative\nclassifier is a competitive alternative to robust classification, especially\nfor problems with limited number of classes.\n", "date": "2022-12-14", "categories": []}
{"id": "2206.01467", "title": "The Importance of Image Interpretation: Patterns of Semantic\n  Misclassification in Real-World Adversarial Images", "abstract": "  Adversarial images are created with the intention of causing an image\nclassifier to produce a misclassification. In this paper, we propose that\nadversarial images should be evaluated based on semantic mismatch, rather than\nlabel mismatch, as used in current work. In other words, we propose that an\nimage of a \"mug\" would be considered adversarial if classified as \"turnip\", but\nnot as \"cup\", as current systems would assume. Our novel idea of taking\nsemantic misclassification into account in the evaluation of adversarial images\noffers two benefits. First, it is a more realistic conceptualization of what\nmakes an image adversarial, which is important in order to fully understand the\nimplications of adversarial images for security and privacy. Second, it makes\nit possible to evaluate the transferability of adversarial images to a\nreal-world classifier, without requiring the classifier's label set to have\nbeen available during the creation of the images. The paper carries out an\nevaluation of a transfer attack on a real-world image classifier that is made\npossible by our semantic misclassification approach. The attack reveals\npatterns in the semantics of adversarial misclassifications that could not be\ninvestigated using conventional label mismatch.\n", "date": "2022-06-03", "categories": []}
{"id": "2212.06576", "title": "AI Model Utilization Measurements For Finding Class Encoding Patterns", "abstract": "  This work addresses the problems of (a) designing utilization measurements of\ntrained artificial intelligence (AI) models and (b) explaining how training\ndata are encoded in AI models based on those measurements. The problems are\nmotivated by the lack of explainability of AI models in security and safety\ncritical applications, such as the use of AI models for classification of\ntraffic signs in self-driving cars. We approach the problems by introducing\ntheoretical underpinnings of AI model utilization measurement and understanding\npatterns in utilization-based class encodings of traffic signs at the level of\ncomputation graphs (AI models), subgraphs, and graph nodes. Conceptually,\nutilization is defined at each graph node (computation unit) of an AI model\nbased on the number and distribution of unique outputs in the space of all\npossible outputs (tensor-states). In this work, utilization measurements are\nextracted from AI models, which include poisoned and clean AI models. In\ncontrast to clean AI models, the poisoned AI models were trained with traffic\nsign images containing systematic, physically realizable, traffic sign\nmodifications (i.e., triggers) to change a correct class label to another label\nin a presence of such a trigger. We analyze class encodings of such clean and\npoisoned AI models, and conclude with implications for trojan injection and\ndetection.\n", "date": "2022-12-12", "categories": []}
{"id": "2212.05337", "title": "Targeted Adversarial Attacks on Deep Reinforcement Learning Policies via\n  Model Checking", "abstract": "  Deep Reinforcement Learning (RL) agents are susceptible to adversarial noise\nin their observations that can mislead their policies and decrease their\nperformance. However, an adversary may be interested not only in decreasing the\nreward, but also in modifying specific temporal logic properties of the policy.\nThis paper presents a metric that measures the exact impact of adversarial\nattacks against such properties. We use this metric to craft optimal\nadversarial attacks. Furthermore, we introduce a model checking method that\nallows us to verify the robustness of RL policies against adversarial attacks.\nOur empirical analysis confirms (1) the quality of our metric to craft\nadversarial attacks against temporal logic properties, and (2) that we are able\nto concisely assess a system's robustness against attacks.\n", "date": "2022-12-10", "categories": []}
{"id": "2212.05380", "title": "Mitigating Adversarial Gray-Box Attacks Against Phishing Detectors", "abstract": "  Although machine learning based algorithms have been extensively used for\ndetecting phishing websites, there has been relatively little work on how\nadversaries may attack such \"phishing detectors\" (PDs for short). In this\npaper, we propose a set of Gray-Box attacks on PDs that an adversary may use\nwhich vary depending on the knowledge that he has about the PD. We show that\nthese attacks severely degrade the effectiveness of several existing PDs. We\nthen propose the concept of operation chains that iteratively map an original\nset of features to a new set of features and develop the \"Protective Operation\nChain\" (POC for short) algorithm. POC leverages the combination of random\nfeature selection and feature mappings in order to increase the attacker's\nuncertainty about the target PD. Using 3 existing publicly available datasets\nplus a fourth that we have created and will release upon the publication of\nthis paper, we show that POC is more robust to these attacks than past\ncompeting work, while preserving predictive performance when no adversarial\nattacks are present. Moreover, POC is robust to attacks on 13 different\nclassifiers, not just one. These results are shown to be statistically\nsignificant at the p < 0.001 level.\n", "date": "2022-12-11", "categories": []}
{"id": "2212.05387", "title": "General Adversarial Defense Against Black-box Attacks via Pixel Level\n  and Feature Level Distribution Alignments", "abstract": "  Deep Neural Networks (DNNs) are vulnerable to the black-box adversarial\nattack that is highly transferable. This threat comes from the distribution gap\nbetween adversarial and clean samples in feature space of the target DNNs. In\nthis paper, we use Deep Generative Networks (DGNs) with a novel training\nmechanism to eliminate the distribution gap. The trained DGNs align the\ndistribution of adversarial samples with clean ones for the target DNNs by\ntranslating pixel values. Different from previous work, we propose a more\neffective pixel level training constraint to make this achievable, thus\nenhancing robustness on adversarial samples. Further, a class-aware\nfeature-level constraint is formulated for integrated distribution alignment.\nOur approach is general and applicable to multiple tasks, including image\nclassification, semantic segmentation, and object detection. We conduct\nextensive experiments on different datasets. Our strategy demonstrates its\nunique effectiveness and generality against black-box attacks.\n", "date": "2022-12-11", "categories": []}
{"id": "2212.05630", "title": "DISCO: Adversarial Defense with Local Implicit Functions", "abstract": "  The problem of adversarial defenses for image classification, where the goal\nis to robustify a classifier against adversarial examples, is considered.\nInspired by the hypothesis that these examples lie beyond the natural image\nmanifold, a novel aDversarIal defenSe with local impliCit functiOns (DISCO) is\nproposed to remove adversarial perturbations by localized manifold projections.\nDISCO consumes an adversarial image and a query pixel location and outputs a\nclean RGB value at the location. It is implemented with an encoder and a local\nimplicit module, where the former produces per-pixel deep features and the\nlatter uses the features in the neighborhood of query pixel for predicting the\nclean RGB value. Extensive experiments demonstrate that both DISCO and its\ncascade version outperform prior defenses, regardless of whether the defense is\nknown to the attacker. DISCO is also shown to be data and parameter efficient\nand to mount defenses that transfers across datasets, classifiers and attacks.\n", "date": "2022-12-11", "categories": []}
{"id": "2212.05709", "title": "HOTCOLD Block: Fooling Thermal Infrared Detectors with a Novel Wearable\n  Design", "abstract": "  Adversarial attacks on thermal infrared imaging expose the risk of related\napplications. Estimating the security of these systems is essential for safely\ndeploying them in the real world. In many cases, realizing the attacks in the\nphysical space requires elaborate special perturbations. These solutions are\noften \\emph{impractical} and \\emph{attention-grabbing}. To address the need for\na physically practical and stealthy adversarial attack, we introduce\n\\textsc{HotCold} Block, a novel physical attack for infrared detectors that\nhide persons utilizing the wearable Warming Paste and Cooling Paste. By\nattaching these readily available temperature-controlled materials to the body,\n\\textsc{HotCold} Block evades human eyes efficiently. Moreover, unlike existing\nmethods that build adversarial patches with complex texture and structure\nfeatures, \\textsc{HotCold} Block utilizes an SSP-oriented adversarial\noptimization algorithm that enables attacks with pure color blocks and explores\nthe influence of size, shape, and position on attack performance. Extensive\nexperimental results in both digital and physical environments demonstrate the\nperformance of our proposed \\textsc{HotCold} Block. \\emph{Code is available:\n\\textcolor{magenta}{https://github.com/weihui1308/HOTCOLDBlock}}.\n", "date": "2022-12-12", "categories": []}
{"id": "2212.05827", "title": "Carpet-bombing patch: attacking a deep network without usual\n  requirements", "abstract": "  Although deep networks have shown vulnerability to evasion attacks, such\nattacks have usually unrealistic requirements. Recent literature discussed the\npossibility to remove or not some of these requirements. This paper contributes\nto this literature by introducing a carpet-bombing patch attack which has\nalmost no requirement. Targeting the feature representations, this patch attack\ndoes not require knowing the network task. This attack decreases accuracy on\nImagenet, mAP on Pascal Voc, and IoU on Cityscapes without being aware that the\nunderlying tasks involved classification, detection or semantic segmentation,\nrespectively. Beyond the potential safety issues raised by this attack, the\nimpact of the carpet-bombing attack highlights some interesting property of\ndeep network layer dynamic.\n", "date": "2022-12-12", "categories": []}
{"id": "2212.04985", "title": "Understanding and Combating Robust Overfitting via Input Loss Landscape\n  Analysis and Regularization", "abstract": "  Adversarial training is widely used to improve the robustness of deep neural\nnetworks to adversarial attack. However, adversarial training is prone to\noverfitting, and the cause is far from clear. This work sheds light on the\nmechanisms underlying overfitting through analyzing the loss landscape w.r.t.\nthe input. We find that robust overfitting results from standard training,\nspecifically the minimization of the clean loss, and can be mitigated by\nregularization of the loss gradients. Moreover, we find that robust overfitting\nturns severer during adversarial training partially because the gradient\nregularization effect of adversarial training becomes weaker due to the\nincrease in the loss landscapes curvature. To improve robust generalization, we\npropose a new regularizer to smooth the loss landscape by penalizing the\nweighted logits variation along the adversarial direction. Our method\nsignificantly mitigates robust overfitting and achieves the highest robustness\nand efficiency compared to similar previous methods. Code is available at\nhttps://github.com/TreeLLi/Combating-RO-AdvLC.\n", "date": "2022-12-09", "categories": []}
{"id": "2212.04138", "title": "Targeted Adversarial Attacks against Neural Network Trajectory\n  Predictors", "abstract": "  Trajectory prediction is an integral component of modern autonomous systems\nas it allows for envisioning future intentions of nearby moving agents. Due to\nthe lack of other agents' dynamics and control policies, deep neural network\n(DNN) models are often employed for trajectory forecasting tasks. Although\nthere exists an extensive literature on improving the accuracy of these models,\nthere is a very limited number of works studying their robustness against\nadversarially crafted input trajectories. To bridge this gap, in this paper, we\npropose a targeted adversarial attack against DNN models for trajectory\nforecasting tasks. We call the proposed attack TA4TP for Targeted adversarial\nAttack for Trajectory Prediction. Our approach generates adversarial input\ntrajectories that are capable of fooling DNN models into predicting\nuser-specified target/desired trajectories. Our attack relies on solving a\nnonlinear constrained optimization problem where the objective function\ncaptures the deviation of the predicted trajectory from a target one while the\nconstraints model physical requirements that the adversarial input should\nsatisfy. The latter ensures that the inputs look natural and they are safe to\nexecute (e.g., they are close to nominal inputs and away from obstacles). We\ndemonstrate the effectiveness of TA4TP on two state-of-the-art DNN models and\ntwo datasets. To the best of our knowledge, we propose the first targeted\nadversarial attack against DNN models used for trajectory forecasting.\n", "date": "2022-12-08", "categories": []}
{"id": "2209.05244", "title": "Universal Backdoor Attacks Detection via Adaptive Adversarial Probe", "abstract": "  Extensive evidence has demonstrated that deep neural networks (DNNs) are\nvulnerable to backdoor attacks, which motivates the development of backdoor\nattacks detection. Most detection methods are designed to verify whether a\nmodel is infected with presumed types of backdoor attacks, yet the adversary is\nlikely to generate diverse backdoor attacks in practice that are unforeseen to\ndefenders, which challenge current detection strategies. In this paper, we\nfocus on this more challenging scenario and propose a universal backdoor\nattacks detection method named Adaptive Adversarial Probe (A2P). Specifically,\nwe posit that the challenge of universal backdoor attacks detection lies in the\nfact that different backdoor attacks often exhibit diverse characteristics in\ntrigger patterns (i.e., sizes and transparencies). Therefore, our A2P adopts a\nglobal-to-local probing framework, which adversarially probes images with\nadaptive regions/budgets to fit various backdoor triggers of different\nsizes/transparencies. Regarding the probing region, we propose the\nattention-guided region generation strategy that generates region proposals\nwith different sizes/locations based on the attention of the target model,\nsince trigger regions often manifest higher model activation. Considering the\nattack budget, we introduce the box-to-sparsity scheduling that iteratively\nincreases the perturbation budget from box to sparse constraint, so that we\ncould better activate different latent backdoors with different transparencies.\nExtensive experiments on multiple datasets (CIFAR-10, GTSRB, Tiny-ImageNet)\ndemonstrate that our method outperforms state-of-the-art baselines by large\nmargins (+12%).\n", "date": "2022-09-12", "categories": []}
{"id": "2210.15392", "title": "LeNo: Adversarial Robust Salient Object Detection Networks with\n  Learnable Noise", "abstract": "  Pixel-wise prediction with deep neural network has become an effective\nparadigm for salient object detection (SOD) and achieved remarkable\nperformance. However, very few SOD models are robust against adversarial\nattacks which are visually imperceptible for human visual attention. The\nprevious work robust saliency (ROSA) shuffles the pre-segmented superpixels and\nthen refines the coarse saliency map by the densely connected conditional\nrandom field (CRF). Different from ROSA that relies on various pre- and\npost-processings, this paper proposes a light-weight Learnable Noise (LeNo) to\ndefend adversarial attacks for SOD models. LeNo preserves accuracy of SOD\nmodels on both adversarial and clean images, as well as inference speed. In\ngeneral, LeNo consists of a simple shallow noise and noise estimation that\nembedded in the encoder and decoder of arbitrary SOD networks respectively.\nInspired by the center prior of human visual attention mechanism, we initialize\nthe shallow noise with a cross-shaped gaussian distribution for better defense\nagainst adversarial attacks. Instead of adding additional network components\nfor post-processing, the proposed noise estimation modifies only one channel of\nthe decoder. With the deeply-supervised noise-decoupled training on\nstate-of-the-art RGB and RGB-D SOD networks, LeNo outperforms previous works\nnot only on adversarial images but also on clean images, which contributes\nstronger robustness for SOD. Our code is available at\nhttps://github.com/ssecv/LeNo.\n", "date": "2022-10-27", "categories": []}
{"id": "2212.03069", "title": "Multiple Perturbation Attack: Attack Pixelwise Under Different\n  $\\ell_p$-norms For Better Adversarial Performance", "abstract": "  Adversarial machine learning has been both a major concern and a hot topic\nrecently, especially with the ubiquitous use of deep neural networks in the\ncurrent landscape. Adversarial attacks and defenses are usually likened to a\ncat-and-mouse game in which defenders and attackers evolve over the time. On\none hand, the goal is to develop strong and robust deep networks that are\nresistant to malicious actors. On the other hand, in order to achieve that, we\nneed to devise even stronger adversarial attacks to challenge these defense\nmodels. Most of existing attacks employs a single $\\ell_p$ distance (commonly,\n$p\\in\\{1,2,\\infty\\}$) to define the concept of closeness and performs steepest\ngradient ascent w.r.t. this $p$-norm to update all pixels in an adversarial\nexample in the same way. These $\\ell_p$ attacks each has its own pros and cons;\nand there is no single attack that can successfully break through defense\nmodels that are robust against multiple $\\ell_p$ norms simultaneously.\nMotivated by these observations, we come up with a natural approach: combining\nvarious $\\ell_p$ gradient projections on a pixel level to achieve a joint\nadversarial perturbation. Specifically, we learn how to perturb each pixel to\nmaximize the attack performance, while maintaining the overall visual\nimperceptibility of adversarial examples. Finally, through various experiments\nwith standardized benchmarks, we show that our method outperforms most current\nstrong attacks across state-of-the-art defense mechanisms, while retaining its\nability to remain clean visually.\n", "date": "2022-12-05", "categories": []}
{"id": "2209.04779", "title": "Scattering Model Guided Adversarial Examples for SAR Target Recognition:\n  Attack and Defense", "abstract": "  Deep Neural Networks (DNNs) based Synthetic Aperture Radar (SAR) Automatic\nTarget Recognition (ATR) systems have shown to be highly vulnerable to\nadversarial perturbations that are deliberately designed yet almost\nimperceptible but can bias DNN inference when added to targeted objects. This\nleads to serious safety concerns when applying DNNs to high-stake SAR ATR\napplications. Therefore, enhancing the adversarial robustness of DNNs is\nessential for implementing DNNs to modern real-world SAR ATR systems. Toward\nbuilding more robust DNN-based SAR ATR models, this article explores the domain\nknowledge of SAR imaging process and proposes a novel Scattering Model Guided\nAdversarial Attack (SMGAA) algorithm which can generate adversarial\nperturbations in the form of electromagnetic scattering response (called\nadversarial scatterers). The proposed SMGAA consists of two parts: 1) a\nparametric scattering model and corresponding imaging method and 2) a\ncustomized gradient-based optimization algorithm. First, we introduce the\neffective Attributed Scattering Center Model (ASCM) and a general imaging\nmethod to describe the scattering behavior of typical geometric structures in\nthe SAR imaging process. By further devising several strategies to take the\ndomain knowledge of SAR target images into account and relax the greedy search\nprocedure, the proposed method does not need to be prudentially finetuned, but\ncan efficiently to find the effective ASCM parameters to fool the SAR\nclassifiers and facilitate the robust model training. Comprehensive evaluations\non the MSTAR dataset show that the adversarial scatterers generated by SMGAA\nare more robust to perturbations and transformations in the SAR processing\nchain than the currently studied attacks, and are effective to construct a\ndefensive model against the malicious scatterers.\n", "date": "2022-09-11", "categories": []}
{"id": "2210.00178", "title": "On the tightness of linear relaxation based robustness certification\n  methods", "abstract": "  There has been a rapid development and interest in adversarial training and\ndefenses in the machine learning community in the recent years. One line of\nresearch focuses on improving the performance and efficiency of adversarial\nrobustness certificates for neural networks \\cite{gowal:19, wong_zico:18,\nraghunathan:18, WengTowardsFC:18, wong:scalable:18, singh:convex_barrier:19,\nHuang_etal:19, single-neuron-relax:20, Zhang2020TowardsSA}. While each\nproviding a certification to lower (or upper) bound the true distortion under\nadversarial attacks via relaxation, less studied was the tightness of\nrelaxation. In this paper, we analyze a family of linear outer approximation\nbased certificate methods via a meta algorithm, IBP-Lin. The aforementioned\nworks often lack quantitative analysis to answer questions such as how does the\nperformance of the certificate method depend on the network configuration and\nthe choice of approximation parameters. Under our framework, we make a first\nattempt at answering these questions, which reveals that the tightness of\nlinear approximation based certification can depend heavily on the\nconfiguration of the trained networks.\n", "date": "2022-10-01", "categories": []}
{"id": "2210.05938", "title": "Robust Models are less Over-Confident", "abstract": "  Despite the success of convolutional neural networks (CNNs) in many academic\nbenchmarks for computer vision tasks, their application in the real-world is\nstill facing fundamental challenges. One of these open problems is the inherent\nlack of robustness, unveiled by the striking effectiveness of adversarial\nattacks. Current attack methods are able to manipulate the network's prediction\nby adding specific but small amounts of noise to the input. In turn,\nadversarial training (AT) aims to achieve robustness against such attacks and\nideally a better model generalization ability by including adversarial samples\nin the trainingset. However, an in-depth analysis of the resulting robust\nmodels beyond adversarial robustness is still pending. In this paper, we\nempirically analyze a variety of adversarially trained models that achieve high\nrobust accuracies when facing state-of-the-art attacks and we show that AT has\nan interesting side-effect: it leads to models that are significantly less\noverconfident with their decisions, even on clean data than non-robust models.\nFurther, our analysis of robust models shows that not only AT but also the\nmodel's building blocks (like activation functions and pooling) have a strong\ninfluence on the models' prediction confidences. Data & Project website:\nhttps://github.com/GeJulia/robustness_confidences_evaluation\n", "date": "2022-10-12", "categories": []}
{"id": "2212.02531", "title": "Enhancing Quantum Adversarial Robustness by Randomized Encodings", "abstract": "  The interplay between quantum physics and machine learning gives rise to the\nemergent frontier of quantum machine learning, where advanced quantum learning\nmodels may outperform their classical counterparts in solving certain\nchallenging problems. However, quantum learning systems are vulnerable to\nadversarial attacks: adding tiny carefully-crafted perturbations on legitimate\ninput samples can cause misclassifications. To address this issue, we propose a\ngeneral scheme to protect quantum learning systems from adversarial attacks by\nrandomly encoding the legitimate data samples through unitary or quantum error\ncorrection encoders. In particular, we rigorously prove that both global and\nlocal random unitary encoders lead to exponentially vanishing gradients (i.e.\nbarren plateaus) for any variational quantum circuits that aim to add\nadversarial perturbations, independent of the input data and the inner\nstructures of adversarial circuits and quantum classifiers. In addition, we\nprove a rigorous bound on the vulnerability of quantum classifiers under local\nunitary adversarial attacks. We show that random black-box quantum error\ncorrection encoders can protect quantum classifiers against local adversarial\nnoises and their robustness increases as we concatenate error correction codes.\nTo quantify the robustness enhancement, we adapt quantum differential privacy\nas a measure of the prediction stability for quantum classifiers. Our results\nestablish versatile defense strategies for quantum classifiers against\nadversarial perturbations, which provide valuable guidance to enhance the\nreliability and security for both near-term and future quantum learning\ntechnologies.\n", "date": "2022-12-05", "categories": []}
{"id": "2212.02582", "title": "Rethinking Backdoor Data Poisoning Attacks in the Context of\n  Semi-Supervised Learning", "abstract": "  Semi-supervised learning methods can train high-accuracy machine learning\nmodels with a fraction of the labeled training samples required for traditional\nsupervised learning. Such methods do not typically involve close review of the\nunlabeled training samples, making them tempting targets for data poisoning\nattacks. In this paper we investigate the vulnerabilities of semi-supervised\nlearning methods to backdoor data poisoning attacks on the unlabeled samples.\nWe show that simple poisoning attacks that influence the distribution of the\npoisoned samples' predicted labels are highly effective - achieving an average\nattack success rate as high as 96.9%. We introduce a generalized attack\nframework targeting semi-supervised learning methods to better understand and\nexploit their limitations and to motivate future defense strategies.\n", "date": "2022-12-05", "categories": []}
{"id": "2211.13305", "title": "Dual Graphs of Polyhedral Decompositions for the Detection of\n  Adversarial Attacks", "abstract": "  Previous work has shown that a neural network with the rectified linear unit\n(ReLU) activation function leads to a convex polyhedral decomposition of the\ninput space. These decompositions can be represented by a dual graph with\nvertices corresponding to polyhedra and edges corresponding to polyhedra\nsharing a facet, which is a subgraph of a Hamming graph. This paper illustrates\nhow one can utilize the dual graph to detect and analyze adversarial attacks in\nthe context of digital images. When an image passes through a network\ncontaining ReLU nodes, the firing or non-firing at a node can be encoded as a\nbit ($1$ for ReLU activation, $0$ for ReLU non-activation). The sequence of all\nbit activations identifies the image with a bit vector, which identifies it\nwith a polyhedron in the decomposition and, in turn, identifies it with a\nvertex in the dual graph. We identify ReLU bits that are discriminators between\nnon-adversarial and adversarial images and examine how well collections of\nthese discriminators can ensemble vote to build an adversarial image detector.\nSpecifically, we examine the similarities and differences of ReLU bit vectors\nfor adversarial images, and their non-adversarial counterparts, using a\npre-trained ResNet-50 architecture. While this paper focuses on adversarial\ndigital images, ResNet-50 architecture, and the ReLU activation function, our\nmethods extend to other network architectures, activation functions, and types\nof datasets.\n", "date": "2022-11-23", "categories": []}
{"id": "2212.01767", "title": "ConfounderGAN: Protecting Image Data Privacy with Causal Confounder", "abstract": "  The success of deep learning is partly attributed to the availability of\nmassive data downloaded freely from the Internet. However, it also means that\nusers' private data may be collected by commercial organizations without\nconsent and used to train their models. Therefore, it's important and necessary\nto develop a method or tool to prevent unauthorized data exploitation. In this\npaper, we propose ConfounderGAN, a generative adversarial network (GAN) that\ncan make personal image data unlearnable to protect the data privacy of its\nowners. Specifically, the noise produced by the generator for each image has\nthe confounder property. It can build spurious correlations between images and\nlabels, so that the model cannot learn the correct mapping from images to\nlabels in this noise-added dataset. Meanwhile, the discriminator is used to\nensure that the generated noise is small and imperceptible, thereby remaining\nthe normal utility of the encrypted image for humans. The experiments are\nconducted in six image classification datasets, consisting of three natural\nobject datasets and three medical datasets. The results demonstrate that our\nmethod not only outperforms state-of-the-art methods in standard settings, but\ncan also be applied to fast encryption scenarios. Moreover, we show a series of\ntransferability and stability experiments to further illustrate the\neffectiveness and superiority of our method.\n", "date": "2022-12-04", "categories": []}
{"id": "2212.01806", "title": "Recognizing Object by Components with Human Prior Knowledge Enhances\n  Adversarial Robustness of Deep Neural Networks", "abstract": "  Adversarial attacks can easily fool object recognition systems based on deep\nneural networks (DNNs). Although many defense methods have been proposed in\nrecent years, most of them can still be adaptively evaded. One reason for the\nweak adversarial robustness may be that DNNs are only supervised by category\nlabels and do not have part-based inductive bias like the recognition process\nof humans. Inspired by a well-known theory in cognitive psychology --\nrecognition-by-components, we propose a novel object recognition model ROCK\n(Recognizing Object by Components with human prior Knowledge). It first\nsegments parts of objects from images, then scores part segmentation results\nwith predefined human prior knowledge, and finally outputs prediction based on\nthe scores. The first stage of ROCK corresponds to the process of decomposing\nobjects into parts in human vision. The second stage corresponds to the\ndecision process of the human brain. ROCK shows better robustness than\nclassical recognition models across various attack settings. These results\nencourage researchers to rethink the rationality of currently widely-used\nDNN-based object recognition models and explore the potential of part-based\nmodels, once important but recently ignored, for improving robustness.\n", "date": "2022-12-04", "categories": []}
{"id": "2208.05650", "title": "Diverse Generative Perturbations on Attention Space for Transferable\n  Adversarial Attacks", "abstract": "  Adversarial attacks with improved transferability - the ability of an\nadversarial example crafted on a known model to also fool unknown models - have\nrecently received much attention due to their practicality. Nevertheless,\nexisting transferable attacks craft perturbations in a deterministic manner and\noften fail to fully explore the loss surface, thus falling into a poor local\noptimum and suffering from low transferability. To solve this problem, we\npropose Attentive-Diversity Attack (ADA), which disrupts diverse salient\nfeatures in a stochastic manner to improve transferability. Primarily, we\nperturb the image attention to disrupt universal features shared by different\nmodels. Then, to effectively avoid poor local optima, we disrupt these features\nin a stochastic manner and explore the search space of transferable\nperturbations more exhaustively. More specifically, we use a generator to\nproduce adversarial perturbations that each disturbs features in different ways\ndepending on an input latent code. Extensive experimental evaluations\ndemonstrate the effectiveness of our method, outperforming the transferability\nof state-of-the-art methods. Codes are available at\nhttps://github.com/wkim97/ADA.\n", "date": "2022-08-11", "categories": []}
{"id": "2211.16192", "title": "Be Careful with Rotation: A Uniform Backdoor Pattern for 3D Shape", "abstract": "  For saving cost, many deep neural networks (DNNs) are trained on third-party\ndatasets downloaded from internet, which enables attacker to implant backdoor\ninto DNNs. In 2D domain, inherent structures of different image formats are\nsimilar. Hence, backdoor attack designed for one image format will suite for\nothers. However, when it comes to 3D world, there is a huge disparity among\ndifferent 3D data structures. As a result, backdoor pattern designed for one\ncertain 3D data structure will be disable for other data structures of the same\n3D scene. Therefore, this paper designs a uniform backdoor pattern: NRBdoor\n(Noisy Rotation Backdoor) which is able to adapt for heterogeneous 3D data\nstructures. Specifically, we start from the unit rotation and then search for\nthe optimal pattern by noise generation and selection process. The proposed\nNRBdoor is natural and imperceptible, since rotation is a common operation\nwhich usually contains noise due to both the miss match between a pair of\npoints and the sensor calibration error for real-world 3D scene. Extensive\nexperiments on 3D mesh and point cloud show that the proposed NRBdoor achieves\nstate-of-the-art performance, with negligible shape variation.\n", "date": "2022-11-28", "categories": []}
{"id": "2211.12294", "title": "PointCA: Evaluating the Robustness of 3D Point Cloud Completion Models\n  Against Adversarial Examples", "abstract": "  Point cloud completion, as the upstream procedure of 3D recognition and\nsegmentation, has become an essential part of many tasks such as navigation and\nscene understanding. While various point cloud completion models have\ndemonstrated their powerful capabilities, their robustness against adversarial\nattacks, which have been proven to be fatally malicious towards deep neural\nnetworks, remains unknown. In addition, existing attack approaches towards\npoint cloud classifiers cannot be applied to the completion models due to\ndifferent output forms and attack purposes. In order to evaluate the robustness\nof the completion models, we propose PointCA, the first adversarial attack\nagainst 3D point cloud completion models. PointCA can generate adversarial\npoint clouds that maintain high similarity with the original ones, while being\ncompleted as another object with totally different semantic information.\nSpecifically, we minimize the representation discrepancy between the\nadversarial example and the target point set to jointly explore the adversarial\npoint clouds in the geometry space and the feature space. Furthermore, to\nlaunch a stealthier attack, we innovatively employ the neighbourhood density\ninformation to tailor the perturbation constraint, leading to geometry-aware\nand distribution-adaptive modifications for each point. Extensive experiments\nagainst different premier point cloud completion networks show that PointCA can\ncause a performance degradation from 77.9% to 16.7%, with the structure chamfer\ndistance kept below 0.01. We conclude that existing completion models are\nseverely vulnerable to adversarial examples, and state-of-the-art defenses for\npoint cloud classification will be partially invalid when applied to incomplete\nand uneven point cloud data.\n", "date": "2022-11-22", "categories": []}
{"id": "2212.00311", "title": "Generalizing and Improving Jacobian and Hessian Regularization", "abstract": "  Jacobian and Hessian regularization aim to reduce the magnitude of the first\nand second-order partial derivatives with respect to neural network inputs, and\nthey are predominantly used to ensure the adversarial robustness of image\nclassifiers. In this work, we generalize previous efforts by extending the\ntarget matrix from zero to any matrix that admits efficient matrix-vector\nproducts. The proposed paradigm allows us to construct novel regularization\nterms that enforce symmetry or diagonality on square Jacobian and Hessian\nmatrices. On the other hand, the major challenge for Jacobian and Hessian\nregularization has been high computational complexity. We introduce\nLanczos-based spectral norm minimization to tackle this difficulty. This\ntechnique uses a parallelized implementation of the Lanczos algorithm and is\ncapable of effective and stable regularization of large Jacobian and Hessian\nmatrices. Theoretical justifications and empirical evidence are provided for\nthe proposed paradigm and technique. We carry out exploratory experiments to\nvalidate the effectiveness of our novel regularization terms. We also conduct\ncomparative experiments to evaluate Lanczos-based spectral norm minimization\nagainst prior methods. Results show that the proposed methodologies are\nadvantageous for a wide range of tasks.\n", "date": "2022-12-01", "categories": []}
{"id": "2209.08064", "title": "A Systematic Evaluation of Node Embedding Robustness", "abstract": "  Node embedding methods map network nodes to low dimensional vectors that can\nbe subsequently used in a variety of downstream prediction tasks. The\npopularity of these methods has grown significantly in recent years, yet, their\nrobustness to perturbations of the input data is still poorly understood. In\nthis paper, we assess the empirical robustness of node embedding models to\nrandom and adversarial poisoning attacks. Our systematic evaluation covers\nrepresentative embedding methods based on Skip-Gram, matrix factorization, and\ndeep neural networks. We compare edge addition, deletion and rewiring attacks\ncomputed using network properties as well as node labels. We also investigate\nthe performance of popular node classification attack baselines that assume\nfull knowledge of the node labels. We report qualitative results via embedding\nvisualization and quantitative results in terms of downstream node\nclassification and network reconstruction performances. We find that node\nclassification results are impacted more than network reconstruction ones, that\ndegree-based and label-based attacks are on average the most damaging and that\nlabel heterophily can strongly influence attack performance.\n", "date": "2022-09-16", "categories": []}
{"id": "2210.15068", "title": "Improving Adversarial Robustness with Self-Paced Hard-Class Pair\n  Reweighting", "abstract": "  Deep Neural Networks are vulnerable to adversarial attacks. Among many\ndefense strategies, adversarial training with untargeted attacks is one of the\nmost effective methods. Theoretically, adversarial perturbation in untargeted\nattacks can be added along arbitrary directions and the predicted labels of\nuntargeted attacks should be unpredictable. However, we find that the naturally\nimbalanced inter-class semantic similarity makes those hard-class pairs become\nvirtual targets of each other. This study investigates the impact of such\nclosely-coupled classes on adversarial attacks and develops a self-paced\nreweighting strategy in adversarial training accordingly. Specifically, we\npropose to upweight hard-class pair losses in model optimization, which prompts\nlearning discriminative features from hard classes. We further incorporate a\nterm to quantify hard-class pair consistency in adversarial training, which\ngreatly boosts model robustness. Extensive experiments show that the proposed\nadversarial training method achieves superior robustness performance over\nstate-of-the-art defenses against a wide range of adversarial attacks.\n", "date": "2022-10-26", "categories": []}
{"id": "2211.07263", "title": "Efficient Adversarial Training with Robust Early-Bird Tickets", "abstract": "  Adversarial training is one of the most powerful methods to improve the\nrobustness of pre-trained language models (PLMs). However, this approach is\ntypically more expensive than traditional fine-tuning because of the necessity\nto generate adversarial examples via gradient descent. Delving into the\noptimization process of adversarial training, we find that robust connectivity\npatterns emerge in the early training phase (typically $0.15\\sim0.3$ epochs),\nfar before parameters converge. Inspired by this finding, we dig out robust\nearly-bird tickets (i.e., subnetworks) to develop an efficient adversarial\ntraining method: (1) searching for robust tickets with structured sparsity in\nthe early stage; (2) fine-tuning robust tickets in the remaining time. To\nextract the robust tickets as early as possible, we design a ticket convergence\nmetric to automatically terminate the searching process. Experiments show that\nthe proposed efficient adversarial training method can achieve up to $7\\times\n\\sim 13 \\times$ training speedups while maintaining comparable or even better\nrobustness compared to the most competitive state-of-the-art adversarial\ntraining methods.\n", "date": "2022-11-14", "categories": []}
{"id": "2211.16806", "title": "Toward Robust Diagnosis: A Contour Attention Preserving Adversarial\n  Defense for COVID-19 Detection", "abstract": "  As the COVID-19 pandemic puts pressure on healthcare systems worldwide, the\ncomputed tomography image based AI diagnostic system has become a sustainable\nsolution for early diagnosis. However, the model-wise vulnerability under\nadversarial perturbation hinders its deployment in practical situation. The\nexisting adversarial training strategies are difficult to generalized into\nmedical imaging field challenged by complex medical texture features. To\novercome this challenge, we propose a Contour Attention Preserving (CAP) method\nbased on lung cavity edge extraction. The contour prior features are injected\nto attention layer via a parameter regularization and we optimize the robust\nempirical risk with hybrid distance metric. We then introduce a new\ncross-nation CT scan dataset to evaluate the generalization capability of the\nadversarial robustness under distribution shift. Experimental results indicate\nthat the proposed method achieves state-of-the-art performance in multiple\nadversarial defense and generalization tasks. The code and dataset are\navailable at https://github.com/Quinn777/CAP.\n", "date": "2022-11-30", "categories": []}
{"id": "2211.15926", "title": "Interpretations Cannot Be Trusted: Stealthy and Effective Adversarial\n  Perturbations against Interpretable Deep Learning", "abstract": "  Deep learning methods have gained increased attention in various applications\ndue to their outstanding performance. For exploring how this high performance\nrelates to the proper use of data artifacts and the accurate problem\nformulation of a given task, interpretation models have become a crucial\ncomponent in developing deep learning-based systems. Interpretation models\nenable the understanding of the inner workings of deep learning models and\noffer a sense of security in detecting the misuse of artifacts in the input\ndata. Similar to prediction models, interpretation models are also susceptible\nto adversarial inputs. This work introduces two attacks, AdvEdge and\nAdvEdge$^{+}$, that deceive both the target deep learning model and the coupled\ninterpretation model. We assess the effectiveness of proposed attacks against\ntwo deep learning model architectures coupled with four interpretation models\nthat represent different categories of interpretation models. Our experiments\ninclude the attack implementation using various attack frameworks. We also\nexplore the potential countermeasures against such attacks. Our analysis shows\nthe effectiveness of our attacks in terms of deceiving the deep learning models\nand their interpreters, and highlights insights to improve and circumvent the\nattacks.\n", "date": "2022-11-29", "categories": []}
{"id": "2211.15929", "title": "Backdoor Vulnerabilities in Normally Trained Deep Learning Models", "abstract": "  We conduct a systematic study of backdoor vulnerabilities in normally trained\nDeep Learning models. They are as dangerous as backdoors injected by data\npoisoning because both can be equally exploited. We leverage 20 different types\nof injected backdoor attacks in the literature as the guidance and study their\ncorrespondences in normally trained models, which we call natural backdoor\nvulnerabilities. We find that natural backdoors are widely existing, with most\ninjected backdoor attacks having natural correspondences. We categorize these\nnatural backdoors and propose a general detection framework. It finds 315\nnatural backdoors in the 56 normally trained models downloaded from the\nInternet, covering all the different categories, while existing scanners\ndesigned for injected backdoors can at most detect 65 backdoors. We also study\nthe root causes and defense of natural backdoors.\n", "date": "2022-11-29", "categories": []}
{"id": "2211.16080", "title": "Understanding and Enhancing Robustness of Concept-based Models", "abstract": "  Rising usage of deep neural networks to perform decision making in critical\napplications like medical diagnosis and financial analysis have raised concerns\nregarding their reliability and trustworthiness. As automated systems become\nmore mainstream, it is important their decisions be transparent, reliable and\nunderstandable by humans for better trust and confidence. To this effect,\nconcept-based models such as Concept Bottleneck Models (CBMs) and\nSelf-Explaining Neural Networks (SENN) have been proposed which constrain the\nlatent space of a model to represent high level concepts easily understood by\ndomain experts in the field. Although concept-based models promise a good\napproach to both increasing explainability and reliability, it is yet to be\nshown if they demonstrate robustness and output consistent concepts under\nsystematic perturbations to their inputs. To better understand performance of\nconcept-based models on curated malicious samples, in this paper, we aim to\nstudy their robustness to adversarial perturbations, which are also known as\nthe imperceptible changes to the input data that are crafted by an attacker to\nfool a well-learned concept-based model. Specifically, we first propose and\nanalyze different malicious attacks to evaluate the security vulnerability of\nconcept based models. Subsequently, we propose a potential general adversarial\ntraining-based defense mechanism to increase robustness of these systems to the\nproposed malicious attacks. Extensive experiments on one synthetic and two\nreal-world datasets demonstrate the effectiveness of the proposed attacks and\nthe defense approach.\n", "date": "2022-11-29", "categories": []}
{"id": "2211.16187", "title": "Quantization-aware Interval Bound Propagation for Training Certifiably\n  Robust Quantized Neural Networks", "abstract": "  We study the problem of training and certifying adversarially robust\nquantized neural networks (QNNs). Quantization is a technique for making neural\nnetworks more efficient by running them using low-bit integer arithmetic and is\ntherefore commonly adopted in industry. Recent work has shown that\nfloating-point neural networks that have been verified to be robust can become\nvulnerable to adversarial attacks after quantization, and certification of the\nquantized representation is necessary to guarantee robustness. In this work, we\npresent quantization-aware interval bound propagation (QA-IBP), a novel method\nfor training robust QNNs. Inspired by advances in robust learning of\nnon-quantized networks, our training algorithm computes the gradient of an\nabstract representation of the actual network. Unlike existing approaches, our\nmethod can handle the discrete semantics of QNNs. Based on QA-IBP, we also\ndevelop a complete verification procedure for verifying the adversarial\nrobustness of QNNs, which is guaranteed to terminate and produce a correct\nanswer. Compared to existing approaches, the key advantage of our verification\nprocedure is that it runs entirely on GPU or other accelerator devices. We\ndemonstrate experimentally that our approach significantly outperforms existing\nmethods and establish the new state-of-the-art for training and certifying the\nrobustness of QNNs.\n", "date": "2022-11-29", "categories": []}
{"id": "2211.16316", "title": "A3T: Accuracy Aware Adversarial Training", "abstract": "  Adversarial training has been empirically shown to be more prone to\noverfitting than standard training. The exact underlying reasons still need to\nbe fully understood. In this paper, we identify one cause of overfitting\nrelated to current practices of generating adversarial samples from\nmisclassified samples. To address this, we propose an alternative approach that\nleverages the misclassified samples to mitigate the overfitting problem. We\nshow that our approach achieves better generalization while having comparable\nrobustness to state-of-the-art adversarial training methods on a wide range of\ncomputer vision, natural language processing, and tabular tasks.\n", "date": "2022-11-29", "categories": []}
{"id": "2208.03111", "title": "Data-free Backdoor Removal based on Channel Lipschitzness", "abstract": "  Recent studies have shown that Deep Neural Networks (DNNs) are vulnerable to\nthe backdoor attacks, which leads to malicious behaviors of DNNs when specific\ntriggers are attached to the input images. It was further demonstrated that the\ninfected DNNs possess a collection of channels, which are more sensitive to the\nbackdoor triggers compared with normal channels. Pruning these channels was\nthen shown to be effective in mitigating the backdoor behaviors. To locate\nthose channels, it is natural to consider their Lipschitzness, which measures\ntheir sensitivity against worst-case perturbations on the inputs. In this work,\nwe introduce a novel concept called Channel Lipschitz Constant (CLC), which is\ndefined as the Lipschitz constant of the mapping from the input images to the\noutput of each channel. Then we provide empirical evidences to show the strong\ncorrelation between an Upper bound of the CLC (UCLC) and the trigger-activated\nchange on the channel activation. Since UCLC can be directly calculated from\nthe weight matrices, we can detect the potential backdoor channels in a\ndata-free manner, and do simple pruning on the infected DNN to repair the\nmodel. The proposed Channel Lipschitzness based Pruning (CLP) method is super\nfast, simple, data-free and robust to the choice of the pruning threshold.\nExtensive experiments are conducted to evaluate the efficiency and\neffectiveness of CLP, which achieves state-of-the-art results among the\nmainstream defense methods even without any data. Source codes are available at\nhttps://github.com/rkteddy/channel-Lipschitzness-based-pruning.\n", "date": "2022-08-05", "categories": []}
{"id": "2208.08297", "title": "An Evolutionary, Gradient-Free, Query-Efficient, Black-Box Algorithm for\n  Generating Adversarial Instances in Deep Networks", "abstract": "  Deep neural networks (DNNs) are sensitive to adversarial data in a variety of\nscenarios, including the black-box scenario, where the attacker is only allowed\nto query the trained model and receive an output. Existing black-box methods\nfor creating adversarial instances are costly, often using gradient estimation\nor training a replacement network. This paper introduces\n\\textbf{Qu}ery-Efficient \\textbf{E}volutiona\\textbf{ry} \\textbf{Attack},\n\\textit{QuEry Attack}, an untargeted, score-based, black-box attack. QuEry\nAttack is based on a novel objective function that can be used in gradient-free\noptimization problems. The attack only requires access to the output logits of\nthe classifier and is thus not affected by gradient masking. No additional\ninformation is needed, rendering our method more suitable to real-life\nsituations. We test its performance with three different state-of-the-art\nmodels -- Inception-v3, ResNet-50, and VGG-16-BN -- against three benchmark\ndatasets: MNIST, CIFAR10 and ImageNet. Furthermore, we evaluate QuEry Attack's\nperformance on non-differential transformation defenses and state-of-the-art\nrobust models. Our results demonstrate the superior performance of QuEry\nAttack, both in terms of accuracy score and query efficiency.\n", "date": "2022-08-17", "categories": []}
{"id": "2211.14966", "title": "Adversarial Rademacher Complexity of Deep Neural Networks", "abstract": "  Deep neural networks are vulnerable to adversarial attacks. Ideally, a robust\nmodel shall perform well on both the perturbed training data and the unseen\nperturbed test data. It is found empirically that fitting perturbed training\ndata is not hard, but generalizing to perturbed test data is quite difficult.\nTo better understand adversarial generalization, it is of great interest to\nstudy the adversarial Rademacher complexity (ARC) of deep neural networks.\nHowever, how to bound ARC in multi-layers cases is largely unclear due to the\ndifficulty of analyzing adversarial loss in the definition of ARC. There have\nbeen two types of attempts of ARC. One is to provide the upper bound of ARC in\nlinear and one-hidden layer cases. However, these approaches seem hard to\nextend to multi-layer cases. Another is to modify the adversarial loss and\nprovide upper bounds of Rademacher complexity on such surrogate loss in\nmulti-layer cases. However, such variants of Rademacher complexity are not\nguaranteed to be bounds for meaningful robust generalization gaps (RGG). In\nthis paper, we provide a solution to this unsolved problem. Specifically, we\nprovide the first bound of adversarial Rademacher complexity of deep neural\nnetworks. Our approach is based on covering numbers. We provide a method to\nhandle the robustify function classes of DNNs such that we can calculate the\ncovering numbers. Finally, we provide experiments to study the empirical\nimplication of our bounds and provide an analysis of poor adversarial\ngeneralization.\n", "date": "2022-11-27", "categories": []}
{"id": "2211.15556", "title": "Attack on Unfair ToS Clause Detection: A Case Study using Universal\n  Adversarial Triggers", "abstract": "  Recent work has demonstrated that natural language processing techniques can\nsupport consumer protection by automatically detecting unfair clauses in the\nTerms of Service (ToS) Agreement. This work demonstrates that transformer-based\nToS analysis systems are vulnerable to adversarial attacks. We conduct\nexperiments attacking an unfair-clause detector with universal adversarial\ntriggers. Experiments show that a minor perturbation of the text can\nconsiderably reduce the detection performance. Moreover, to measure the\ndetectability of the triggers, we conduct a detailed human evaluation study by\ncollecting both answer accuracy and response time from the participants. The\nresults show that the naturalness of the triggers remains key to tricking\nreaders.\n", "date": "2022-11-28", "categories": []}
{"id": "2208.03610", "title": "Blackbox Attacks via Surrogate Ensemble Search", "abstract": "  Blackbox adversarial attacks can be categorized into transfer- and\nquery-based attacks. Transfer methods do not require any feedback from the\nvictim model, but provide lower success rates compared to query-based methods.\nQuery attacks often require a large number of queries for success. To achieve\nthe best of both approaches, recent efforts have tried to combine them, but\nstill require hundreds of queries to achieve high success rates (especially for\ntargeted attacks). In this paper, we propose a novel method for Blackbox\nAttacks via Surrogate Ensemble Search (BASES) that can generate highly\nsuccessful blackbox attacks using an extremely small number of queries. We\nfirst define a perturbation machine that generates a perturbed image by\nminimizing a weighted loss function over a fixed set of surrogate models. To\ngenerate an attack for a given victim model, we search over the weights in the\nloss function using queries generated by the perturbation machine. Since the\ndimension of the search space is small (same as the number of surrogate\nmodels), the search requires a small number of queries. We demonstrate that our\nproposed method achieves better success rate with at least 30x fewer queries\ncompared to state-of-the-art methods on different image classifiers trained\nwith ImageNet. In particular, our method requires as few as 3 queries per image\n(on average) to achieve more than a 90% success rate for targeted attacks and\n1-2 queries per image for over a 99% success rate for untargeted attacks. Our\nmethod is also effective on Google Cloud Vision API and achieved a 91%\nuntargeted attack success rate with 2.9 queries per image. We also show that\nthe perturbations generated by our proposed method are highly transferable and\ncan be adopted for hard-label blackbox attacks. We also show effectiveness of\nBASES for hiding attacks on object detectors.\n", "date": "2022-08-07", "categories": []}
{"id": "2208.08677", "title": "Enhancing Targeted Attack Transferability via Diversified Weight Pruning", "abstract": "  Malicious attackers can generate targeted adversarial examples by imposing\ntiny noises, forcing neural networks to produce specific incorrect outputs.\nWith cross-model transferability, network models remain vulnerable even in\nblack-box settings. Recent studies have shown the effectiveness of\nensemble-based methods in generating transferable adversarial examples. To\nfurther enhance transferability, model augmentation methods aim to produce more\nnetworks participating in the ensemble. However, existing model augmentation\nmethods are only proven effective in untargeted attacks. In this work, we\npropose Diversified Weight Pruning (DWP), a novel model augmentation technique\nfor generating transferable targeted attacks. DWP leverages the weight pruning\nmethod commonly used in model compression. Compared with prior work, DWP\nprotects necessary connections and ensures the diversity of the pruned models\nsimultaneously, which we show are crucial for targeted transferability.\nExperiments on the ImageNet-compatible dataset under various and more\nchallenging scenarios confirm the effectiveness: transferring to adversarially\ntrained models, Non-CNN architectures, and Google Cloud Vision. The results\nshow that our proposed DWP improves the targeted attack success rates with up\nto $10.1$%, $6.6$%, and $7.0$% on the combination of state-of-the-art methods,\nrespectively. The source code will be made available after acceptance.\n", "date": "2022-08-18", "categories": []}
{"id": "2211.10782", "title": "Let Graph be the Go Board: Gradient-free Node Injection Attack for Graph\n  Neural Networks via Reinforcement Learning", "abstract": "  Graph Neural Networks (GNNs) have drawn significant attentions over the years\nand been broadly applied to essential applications requiring solid robustness\nor vigorous security standards, such as product recommendation and user\nbehavior modeling. Under these scenarios, exploiting GNN's vulnerabilities and\nfurther downgrading its performance become extremely incentive for adversaries.\nPrevious attackers mainly focus on structural perturbations or node injections\nto the existing graphs, guided by gradients from the surrogate models. Although\nthey deliver promising results, several limitations still exist. For the\nstructural perturbation attack, to launch a proposed attack, adversaries need\nto manipulate the existing graph topology, which is impractical in most\ncircumstances. Whereas for the node injection attack, though being more\npractical, current approaches require training surrogate models to simulate a\nwhite-box setting, which results in significant performance downgrade when the\nsurrogate architecture diverges from the actual victim model. To bridge these\ngaps, in this paper, we study the problem of black-box node injection attack,\nwithout training a potentially misleading surrogate model. Specifically, we\nmodel the node injection attack as a Markov decision process and propose\nGradient-free Graph Advantage Actor Critic, namely G2A2C, a reinforcement\nlearning framework in the fashion of advantage actor critic. By directly\nquerying the victim model, G2A2C learns to inject highly malicious nodes with\nextremely limited attacking budgets, while maintaining a similar node feature\ndistribution. Through our comprehensive experiments over eight acknowledged\nbenchmark datasets with different characteristics, we demonstrate the superior\nperformance of our proposed G2A2C over the existing state-of-the-art attackers.\nSource code is publicly available at: https://github.com/jumxglhf/G2A2C}.\n", "date": "2022-11-19", "categories": []}
{"id": "2211.14088", "title": "Boundary Adversarial Examples Against Adversarial Overfitting", "abstract": "  Standard adversarial training approaches suffer from robust overfitting where\nthe robust accuracy decreases when models are adversarially trained for too\nlong. The origin of this problem is still unclear and conflicting explanations\nhave been reported, i.e., memorization effects induced by large loss data or\nbecause of small loss data and growing differences in loss distribution of\ntraining samples as the adversarial training progresses. Consequently, several\nmitigation approaches including early stopping, temporal ensembling and weight\nperturbations on small loss data have been proposed to mitigate the effect of\nrobust overfitting. However, a side effect of these strategies is a larger\nreduction in clean accuracy compared to standard adversarial training. In this\npaper, we investigate if these mitigation approaches are complimentary to each\nother in improving adversarial training performance. We further propose the use\nof helper adversarial examples that can be obtained with minimal cost in the\nadversarial example generation, and show how they increase the clean accuracy\nin the existing approaches without compromising the robust accuracy.\n", "date": "2022-11-25", "categories": []}
{"id": "2211.11835", "title": "Fairness Increases Adversarial Vulnerability", "abstract": "  The remarkable performance of deep learning models and their applications in\nconsequential domains (e.g., facial recognition) introduces important\nchallenges at the intersection of equity and security. Fairness and robustness\nare two desired notions often required in learning models. Fairness ensures\nthat models do not disproportionately harm (or benefit) some groups over\nothers, while robustness measures the models' resilience against small input\nperturbations.\n  This paper shows the existence of a dichotomy between fairness and\nrobustness, and analyzes when achieving fairness decreases the model robustness\nto adversarial samples. The reported analysis sheds light on the factors\ncausing such contrasting behavior, suggesting that distance to the decision\nboundary across groups as a key explainer for this behavior. Extensive\nexperiments on non-linear models and different architectures validate the\ntheoretical findings in multiple vision domains. Finally, the paper proposes a\nsimple, yet effective, solution to construct models achieving good tradeoffs\nbetween fairness and robustness.\n", "date": "2022-11-21", "categories": []}
{"id": "2211.12624", "title": "Improving Robust Generalization by Direct PAC-Bayesian Bound\n  Minimization", "abstract": "  Recent research in robust optimization has shown an overfitting-like\nphenomenon in which models trained against adversarial attacks exhibit higher\nrobustness on the training set compared to the test set. Although previous work\nprovided theoretical explanations for this phenomenon using a robust\nPAC-Bayesian bound over the adversarial test error, related algorithmic\nderivations are at best only loosely connected to this bound, which implies\nthat there is still a gap between their empirical success and our understanding\nof adversarial robustness theory. To close this gap, in this paper we consider\na different form of the robust PAC-Bayesian bound and directly minimize it with\nrespect to the model posterior. The derivation of the optimal solution connects\nPAC-Bayesian learning to the geometry of the robust loss surface through a\nTrace of Hessian (TrH) regularizer that measures the surface flatness. In\npractice, we restrict the TrH regularizer to the top layer only, which results\nin an analytical solution to the bound whose computational cost does not depend\non the network depth. Finally, we evaluate our TrH regularization approach over\nCIFAR-10/100 and ImageNet using Vision Transformers (ViT) and compare against\nbaseline adversarial robustness algorithms. Experimental results show that TrH\nregularization leads to improved ViT robustness that either matches or\nsurpasses previous state-of-the-art approaches while at the same time requires\nless memory and computational cost.\n", "date": "2022-11-22", "categories": []}
{"id": "2211.12713", "title": "Reliable Robustness Evaluation via Automatically Constructed Attack\n  Ensembles", "abstract": "  Attack Ensemble (AE), which combines multiple attacks together, provides a\nreliable way to evaluate adversarial robustness. In practice, AEs are often\nconstructed and tuned by human experts, which however tends to be sub-optimal\nand time-consuming. In this work, we present AutoAE, a conceptually simple\napproach for automatically constructing AEs. In brief, AutoAE repeatedly adds\nthe attack and its iteration steps to the ensemble that maximizes ensemble\nimprovement per additional iteration consumed. We show theoretically that\nAutoAE yields AEs provably within a constant factor of the optimal for a given\ndefense. We then use AutoAE to construct two AEs for $l_{\\infty}$ and $l_2$\nattacks, and apply them without any tuning or adaptation to 45 top adversarial\ndefenses on the RobustBench leaderboard. In all except one cases we achieve\nequal or better (often the latter) robustness evaluation than existing AEs, and\nnotably, in 29 cases we achieve better robustness evaluation than the best\nknown one. Such performance of AutoAE shows itself as a reliable evaluation\nprotocol for adversarial robustness, which further indicates the huge potential\nof automatic AE construction. Code is available at\n\\url{https://github.com/LeegerPENG/AutoAE}.\n", "date": "2022-11-23", "categories": []}
{"id": "2211.12851", "title": "A Streamlit-based Artificial Intelligence Trust Platform for\n  Next-Generation Wireless Networks", "abstract": "  With the rapid development and integration of artificial intelligence (AI)\nmethods in next-generation networks (NextG), AI algorithms have provided\nsignificant advantages for NextG in terms of frequency spectrum usage,\nbandwidth, latency, and security. A key feature of NextG is the integration of\nAI, i.e., self-learning architecture based on self-supervised algorithms, to\nimprove the performance of the network. A secure AI-powered structure is also\nexpected to protect NextG networks against cyber-attacks. However, AI itself\nmay be attacked, i.e., model poisoning targeted by attackers, and it results in\ncybersecurity violations. This paper proposes an AI trust platform using\nStreamlit for NextG networks that allows researchers to evaluate, defend,\ncertify, and verify their AI models and applications against adversarial\nthreats of evasion, poisoning, extraction, and interference.\n", "date": "2022-10-25", "categories": []}
{"id": "2211.12990", "title": "Adversarial Attacks are a Surprisingly Strong Baseline for Poisoning\n  Few-Shot Meta-Learners", "abstract": "  This paper examines the robustness of deployed few-shot meta-learning systems\nwhen they are fed an imperceptibly perturbed few-shot dataset. We attack\namortized meta-learners, which allows us to craft colluding sets of inputs that\nare tailored to fool the system's learning algorithm when used as training\ndata. Jointly crafted adversarial inputs might be expected to synergistically\nmanipulate a classifier, allowing for very strong data-poisoning attacks that\nwould be hard to detect. We show that in a white box setting, these attacks are\nvery successful and can cause the target model's predictions to become worse\nthan chance. However, in opposition to the well-known transferability of\nadversarial examples in general, the colluding sets do not transfer well to\ndifferent classifiers. We explore two hypotheses to explain this: 'overfitting'\nby the attack, and mismatch between the model on which the attack is generated\nand that to which the attack is transferred. Regardless of the mitigation\nstrategies suggested by these hypotheses, the colluding inputs transfer no\nbetter than adversarial inputs that are generated independently in the usual\nway.\n", "date": "2022-11-23", "categories": []}
{"id": "2211.13171", "title": "Query Efficient Cross-Dataset Transferable Black-Box Attack on Action\n  Recognition", "abstract": "  Black-box adversarial attacks present a realistic threat to action\nrecognition systems. Existing black-box attacks follow either a query-based\napproach where an attack is optimized by querying the target model, or a\ntransfer-based approach where attacks are generated using a substitute model.\nWhile these methods can achieve decent fooling rates, the former tends to be\nhighly query-inefficient while the latter assumes extensive knowledge of the\nblack-box model's training data. In this paper, we propose a new attack on\naction recognition that addresses these shortcomings by generating\nperturbations to disrupt the features learned by a pre-trained substitute model\nto reduce the number of queries. By using a nearly disjoint dataset to train\nthe substitute model, our method removes the requirement that the substitute\nmodel be trained using the same dataset as the target model, and leverages\nqueries to the target model to retain the fooling rate benefits provided by\nquery-based methods. This ultimately results in attacks which are more\ntransferable than conventional black-box attacks. Through extensive\nexperiments, we demonstrate highly query-efficient black-box attacks with the\nproposed framework. Our method achieves 8% and 12% higher deception rates\ncompared to state-of-the-art query-based and transfer-based attacks,\nrespectively.\n", "date": "2022-11-23", "categories": []}
{"id": "2211.10670", "title": "Towards Adversarial Robustness of Deep Vision Algorithms", "abstract": "  Deep learning methods have achieved great success in solving computer vision\ntasks, and they have been widely utilized in artificially intelligent systems\nfor image processing, analysis, and understanding. However, deep neural\nnetworks have been shown to be vulnerable to adversarial perturbations in input\ndata. The security issues of deep neural networks have thus come to the fore.\nIt is imperative to study the adversarial robustness of deep vision algorithms\ncomprehensively. This talk focuses on the adversarial robustness of image\nclassification models and image denoisers. We will discuss the robustness of\ndeep vision algorithms from three perspectives: 1) robustness evaluation (we\npropose the ObsAtk to evaluate the robustness of denoisers), 2) robustness\nimprovement (HAT, TisODE, and CIFS are developed to robustify vision models),\nand 3) the connection between adversarial robustness and generalization\ncapability to new domains (we find that adversarially robust denoisers can deal\nwith unseen types of real-world noise).\n", "date": "2022-11-19", "categories": []}
{"id": "2211.11880", "title": "Addressing Mistake Severity in Neural Networks with Semantic Knowledge", "abstract": "  Robustness in deep neural networks and machine learning algorithms in general\nis an open research challenge. In particular, it is difficult to ensure\nalgorithmic performance is maintained on out-of-distribution inputs or\nanomalous instances that cannot be anticipated at training time. Embodied\nagents will be deployed in these conditions, and are likely to make incorrect\npredictions. An agent will be viewed as untrustworthy unless it can maintain\nits performance in dynamic environments. Most robust training techniques aim to\nimprove model accuracy on perturbed inputs; as an alternate form of robustness,\nwe aim to reduce the severity of mistakes made by neural networks in\nchallenging conditions. We leverage current adversarial training methods to\ngenerate targeted adversarial attacks during the training process in order to\nincrease the semantic similarity between a model's predictions and true labels\nof misclassified instances. Results demonstrate that our approach performs\nbetter with respect to mistake severity compared to standard and adversarially\ntrained models. We also find an intriguing role that non-robust features play\nwith regards to semantic similarity.\n", "date": "2022-11-21", "categories": []}
{"id": "2207.10862", "title": "Contrastive Self-Supervised Learning Leads to Higher Adversarial\n  Susceptibility", "abstract": "  Contrastive self-supervised learning (CSL) has managed to match or surpass\nthe performance of supervised learning in image and video classification.\nHowever, it is still largely unknown if the nature of the representations\ninduced by the two learning paradigms is similar. We investigate this under the\nlens of adversarial robustness. Our analysis of the problem reveals that CSL\nhas intrinsically higher sensitivity to perturbations over supervised learning.\nWe identify the uniform distribution of data representation over a unit\nhypersphere in the CSL representation space as the key contributor to this\nphenomenon. We establish that this is a result of the presence of false\nnegative pairs in the training process, which increases model sensitivity to\ninput perturbations. Our finding is supported by extensive experiments for\nimage and video classification using adversarial perturbations and other input\ncorruptions. We devise a strategy to detect and remove false negative pairs\nthat is simple, yet effective in improving model robustness with CSL training.\nWe close up to 68% of the robustness gap between CSL and its supervised\ncounterpart. Finally, we contribute to adversarial learning by incorporating\nour method in CSL. We demonstrate an average gain of about 5% over two\ndifferent state-of-the-art methods in this domain.\n", "date": "2022-07-22", "categories": []}
{"id": "2211.05371", "title": "MSDT: Masked Language Model Scoring Defense in Text Domain", "abstract": "  Pre-trained language models allowed us to process downstream tasks with the\nhelp of fine-tuning, which aids the model to achieve fairly high accuracy in\nvarious Natural Language Processing (NLP) tasks. Such easily-downloaded\nlanguage models from various websites empowered the public users as well as\nsome major institutions to give a momentum to their real-life application.\nHowever, it was recently proven that models become extremely vulnerable when\nthey are backdoor attacked with trigger-inserted poisoned datasets by malicious\nusers. The attackers then redistribute the victim models to the public to\nattract other users to use them, where the models tend to misclassify when\ncertain triggers are detected within the training sample. In this paper, we\nwill introduce a novel improved textual backdoor defense method, named MSDT,\nthat outperforms the current existing defensive algorithms in specific\ndatasets. The experimental results illustrate that our method can be effective\nand constructive in terms of defending against backdoor attack in text domain.\nCode is available at https://github.com/jcroh0508/MSDT.\n", "date": "2022-11-10", "categories": []}
{"id": "2211.10752", "title": "Towards Robust Dataset Learning", "abstract": "  Adversarial training has been actively studied in recent computer vision\nresearch to improve the robustness of models. However, due to the huge\ncomputational cost of generating adversarial samples, adversarial training\nmethods are often slow. In this paper, we study the problem of learning a\nrobust dataset such that any classifier naturally trained on the dataset is\nadversarially robust. Such a dataset benefits the downstream tasks as natural\ntraining is much faster than adversarial training, and demonstrates that the\ndesired property of robustness is transferable between models and data. In this\nwork, we propose a principled, tri-level optimization to formulate the robust\ndataset learning problem. We show that, under an abstraction model that\ncharacterizes robust vs. non-robust features, the proposed method provably\nlearns a robust dataset. Extensive experiments on MNIST, CIFAR10, and\nTinyImageNet demostrate the effectiveness of our algorithm with different\nnetwork initializations and architectures.\n", "date": "2022-11-19", "categories": []}
{"id": "2211.10896", "title": "Spectral Adversarial Training for Robust Graph Neural Network", "abstract": "  Recent studies demonstrate that Graph Neural Networks (GNNs) are vulnerable\nto slight but adversarially designed perturbations, known as adversarial\nexamples. To address this issue, robust training methods against adversarial\nexamples have received considerable attention in the literature.\n\\emph{Adversarial Training (AT)} is a successful approach to learning a robust\nmodel using adversarially perturbed training samples. Existing AT methods on\nGNNs typically construct adversarial perturbations in terms of graph structures\nor node features. However, they are less effective and fraught with challenges\non graph data due to the discreteness of graph structure and the relationships\nbetween connected examples. In this work, we seek to address these challenges\nand propose Spectral Adversarial Training (SAT), a simple yet effective\nadversarial training approach for GNNs. SAT first adopts a low-rank\napproximation of the graph structure based on spectral decomposition, and then\nconstructs adversarial perturbations in the spectral domain rather than\ndirectly manipulating the original graph structure. To investigate its\neffectiveness, we employ SAT on three widely used GNNs. Experimental results on\nfour public graph datasets demonstrate that SAT significantly improves the\nrobustness of GNNs against adversarial attacks without sacrificing\nclassification accuracy and training efficiency.\n", "date": "2022-11-20", "categories": []}
{"id": "2210.06871", "title": "Adv-Attribute: Inconspicuous and Transferable Adversarial Attack on Face\n  Recognition", "abstract": "  Deep learning models have shown their vulnerability when dealing with\nadversarial attacks. Existing attacks almost perform on low-level instances,\nsuch as pixels and super-pixels, and rarely exploit semantic clues. For face\nrecognition attacks, existing methods typically generate the l_p-norm\nperturbations on pixels, however, resulting in low attack transferability and\nhigh vulnerability to denoising defense models. In this work, instead of\nperforming perturbations on the low-level pixels, we propose to generate\nattacks through perturbing on the high-level semantics to improve attack\ntransferability. Specifically, a unified flexible framework, Adversarial\nAttributes (Adv-Attribute), is designed to generate inconspicuous and\ntransferable attacks on face recognition, which crafts the adversarial noise\nand adds it into different attributes based on the guidance of the difference\nin face recognition features from the target. Moreover, the importance-aware\nattribute selection and the multi-objective optimization strategy are\nintroduced to further ensure the balance of stealthiness and attacking\nstrength. Extensive experiments on the FFHQ and CelebA-HQ datasets show that\nthe proposed Adv-Attribute method achieves the state-of-the-art attacking\nsuccess rates while maintaining better visual effects against recent attack\nmethods.\n", "date": "2022-10-13", "categories": []}
{"id": "2211.09959", "title": "Potential Auto-driving Threat: Universal Rain-removal Attack", "abstract": "  The problem of robustness in adverse weather conditions is considered a\nsignificant challenge for computer vision algorithms in the applicants of\nautonomous driving. Image rain removal algorithms are a general solution to\nthis problem. They find a deep connection between raindrops/rain-streaks and\nimages by mining the hidden features and restoring information about the\nrain-free environment based on the powerful representation capabilities of\nneural networks. However, previous research has focused on architecture\ninnovations and has yet to consider the vulnerability issues that already exist\nin neural networks. This research gap hints at a potential security threat\ngeared toward the intelligent perception of autonomous driving in the rain. In\nthis paper, we propose a universal rain-removal attack (URA) on the\nvulnerability of image rain-removal algorithms by generating a non-additive\nspatial perturbation that significantly reduces the similarity and image\nquality of scene restoration. Notably, this perturbation is difficult to\nrecognise by humans and is also the same for different target images. Thus, URA\ncould be considered a critical tool for the vulnerability detection of image\nrain-removal algorithms. It also could be developed as a real-world artificial\nintelligence attack method. Experimental results show that URA can reduce the\nscene repair capability by 39.5% and the image generation quality by 26.4%,\ntargeting the state-of-the-art (SOTA) single-image rain-removal algorithms\ncurrently available.\n", "date": "2022-11-18", "categories": []}
{"id": "2211.10033", "title": "Adversarial Stimuli: Attacking Brain-Computer Interfaces via Perturbed\n  Sensory Events", "abstract": "  Machine learning models are known to be vulnerable to adversarial\nperturbations in the input domain, causing incorrect predictions. Inspired by\nthis phenomenon, we explore the feasibility of manipulating EEG-based Motor\nImagery (MI) Brain Computer Interfaces (BCIs) via perturbations in sensory\nstimuli. Similar to adversarial examples, these \\emph{adversarial stimuli} aim\nto exploit the limitations of the integrated brain-sensor-processing components\nof the BCI system in handling shifts in participants' response to changes in\nsensory stimuli. This paper proposes adversarial stimuli as an attack vector\nagainst BCIs, and reports the findings of preliminary experiments on the impact\nof visual adversarial stimuli on the integrity of EEG-based MI BCIs. Our\nfindings suggest that minor adversarial stimuli can significantly deteriorate\nthe performance of MI BCIs across all participants (p=0.0003). Additionally,\nour results indicate that such attacks are more effective in conditions with\ninduced stress.\n", "date": "2022-11-18", "categories": []}
{"id": "2211.09527", "title": "Ignore Previous Prompt: Attack Techniques For Language Models", "abstract": "  Transformer-based large language models (LLMs) provide a powerful foundation\nfor natural language tasks in large-scale customer-facing applications.\nHowever, studies that explore their vulnerabilities emerging from malicious\nuser interaction are scarce. By proposing PromptInject, a prosaic alignment\nframework for mask-based iterative adversarial prompt composition, we examine\nhow GPT-3, the most widely deployed language model in production, can be easily\nmisaligned by simple handcrafted inputs. In particular, we investigate two\ntypes of attacks -- goal hijacking and prompt leaking -- and demonstrate that\neven low-aptitude, but sufficiently ill-intentioned agents, can easily exploit\nGPT-3's stochastic nature, creating long-tail risks. The code for PromptInject\nis available at https://github.com/agencyenterprise/PromptInject.\n", "date": "2022-11-17", "categories": []}
{"id": "2211.09773", "title": "T-SEA: Transfer-based Self-Ensemble Attack on Object Detection", "abstract": "  Compared to query-based black-box attacks, transfer-based black-box attacks\ndo not require any information of the attacked models, which ensures their\nsecrecy. However, most existing transfer-based approaches rely on ensembling\nmultiple models to boost the attack transferability, which is time- and\nresource-intensive, not to mention the difficulty of obtaining diverse models\non the same task. To address this limitation, in this work, we focus on the\nsingle-model transfer-based black-box attack on object detection, utilizing\nonly one model to achieve a high-transferability adversarial attack on multiple\nblack-box detectors. Specifically, we first make observations on the patch\noptimization process of the existing method and propose an enhanced attack\nframework by slightly adjusting its training strategies. Then, we analogize\npatch optimization with regular model optimization, proposing a series of\nself-ensemble approaches on the input data, the attacked model, and the\nadversarial patch to efficiently make use of the limited information and\nprevent the patch from overfitting. The experimental results show that the\nproposed framework can be applied with multiple classical base attack methods\n(e.g., PGD and MIM) to greatly improve the black-box transferability of the\nwell-optimized patch on multiple mainstream detectors, meanwhile boosting\nwhite-box performance. Our code is available at\nhttps://github.com/VDIGPKU/T-SEA.\n", "date": "2022-11-16", "categories": []}
{"id": "2211.08453", "title": "Improved techniques for deterministic l2 robustness", "abstract": "  Training convolutional neural networks (CNNs) with a strict 1-Lipschitz\nconstraint under the $l_{2}$ norm is useful for adversarial robustness,\ninterpretable gradients and stable training. 1-Lipschitz CNNs are usually\ndesigned by enforcing each layer to have an orthogonal Jacobian matrix (for all\ninputs) to prevent the gradients from vanishing during backpropagation.\nHowever, their performance often significantly lags behind that of heuristic\nmethods to enforce Lipschitz constraints where the resulting CNN is not\n\\textit{provably} 1-Lipschitz. In this work, we reduce this gap by introducing\n(a) a procedure to certify robustness of 1-Lipschitz CNNs by replacing the last\nlinear layer with a 1-hidden layer MLP that significantly improves their\nperformance for both standard and provably robust accuracy, (b) a method to\nsignificantly reduce the training time per epoch for Skew Orthogonal\nConvolution (SOC) layers (>30\\% reduction for deeper networks) and (c) a class\nof pooling layers using the mathematical property that the $l_{2}$ distance of\nan input to a manifold is 1-Lipschitz. Using these methods, we significantly\nadvance the state-of-the-art for standard and provable robust accuracies on\nCIFAR-10 (gains of +1.79\\% and +3.82\\%) and similarly on CIFAR-100 (+3.78\\% and\n+4.75\\%) across all networks. Code is available at\n\\url{https://github.com/singlasahil14/improved_l2_robustness}.\n", "date": "2022-11-15", "categories": []}
{"id": "2211.08706", "title": "Efficiently Finding Adversarial Examples with DNN Preprocessing", "abstract": "  Deep Neural Networks (DNNs) are everywhere, frequently performing a fairly\ncomplex task that used to be unimaginable for machines to carry out. In doing\nso, they do a lot of decision making which, depending on the application, may\nbe disastrous if gone wrong. This necessitates a formal argument that the\nunderlying neural networks satisfy certain desirable properties. Robustness is\none such key property for DNNs, particularly if they are being deployed in\nsafety or business critical applications. Informally speaking, a DNN is not\nrobust if very small changes to its input may affect the output in a\nconsiderable way (e.g. changes the classification for that input). The task of\nfinding an adversarial example is to demonstrate this lack of robustness,\nwhenever applicable. While this is doable with the help of constrained\noptimization techniques, scalability becomes a challenge due to large-sized\nnetworks. This paper proposes the use of information gathered by preprocessing\nthe DNN to heavily simplify the optimization problem. Our experiments\nsubstantiate that this is effective, and does significantly better than the\nstate-of-the-art.\n", "date": "2022-11-16", "categories": []}
{"id": "2211.08859", "title": "Attacking Object Detector Using A Universal Targeted Label-Switch Patch", "abstract": "  Adversarial attacks against deep learning-based object detectors (ODs) have\nbeen studied extensively in the past few years. These attacks cause the model\nto make incorrect predictions by placing a patch containing an adversarial\npattern on the target object or anywhere within the frame. However, none of\nprior research proposed a misclassification attack on ODs, in which the patch\nis applied on the target object. In this study, we propose a novel, universal,\ntargeted, label-switch attack against the state-of-the-art object detector,\nYOLO. In our attack, we use (i) a tailored projection function to enable the\nplacement of the adversarial patch on multiple target objects in the image\n(e.g., cars), each of which may be located a different distance away from the\ncamera or have a different view angle relative to the camera, and (ii) a unique\nloss function capable of changing the label of the attacked objects. The\nproposed universal patch, which is trained in the digital domain, is\ntransferable to the physical domain. We performed an extensive evaluation using\ndifferent types of object detectors, different video streams captured by\ndifferent cameras, and various target classes, and evaluated different\nconfigurations of the adversarial patch in the physical domain.\n", "date": "2022-11-16", "categories": []}
{"id": "2211.08008", "title": "MORA: Improving Ensemble Robustness Evaluation with Model-Reweighing\n  Attack", "abstract": "  Adversarial attacks can deceive neural networks by adding tiny perturbations\nto their input data. Ensemble defenses, which are trained to minimize attack\ntransferability among sub-models, offer a promising research direction to\nimprove robustness against such attacks while maintaining a high accuracy on\nnatural inputs. We discover, however, that recent state-of-the-art (SOTA)\nadversarial attack strategies cannot reliably evaluate ensemble defenses,\nsizeably overestimating their robustness. This paper identifies the two factors\nthat contribute to this behavior. First, these defenses form ensembles that are\nnotably difficult for existing gradient-based method to attack, due to gradient\nobfuscation. Second, ensemble defenses diversify sub-model gradients,\npresenting a challenge to defeat all sub-models simultaneously, simply summing\ntheir contributions may counteract the overall attack objective; yet, we\nobserve that ensemble may still be fooled despite most sub-models being\ncorrect. We therefore introduce MORA, a model-reweighing attack to steer\nadversarial example synthesis by reweighing the importance of sub-model\ngradients. MORA finds that recent ensemble defenses all exhibit varying degrees\nof overestimated robustness. Comparing it against recent SOTA white-box\nattacks, it can converge orders of magnitude faster while achieving higher\nattack success rates across all ensemble models examined with three different\nensemble modes (i.e., ensembling by either softmax, voting or logits). In\nparticular, most ensemble defenses exhibit near or exactly 0% robustness\nagainst MORA with $\\ell^\\infty$ perturbation within 0.02 on CIFAR-10, and 0.01\non CIFAR-100. We make MORA open source with reproducible results and\npre-trained models; and provide a leaderboard of ensemble defenses under\nvarious attack strategies.\n", "date": "2022-11-15", "categories": []}
{"id": "2211.08068", "title": "Resisting Graph Adversarial Attack via Cooperative Homophilous\n  Augmentation", "abstract": "  Recent studies show that Graph Neural Networks(GNNs) are vulnerable and\neasily fooled by small perturbations, which has raised considerable concerns\nfor adapting GNNs in various safety-critical applications. In this work, we\nfocus on the emerging but critical attack, namely, Graph Injection Attack(GIA),\nin which the adversary poisons the graph by injecting fake nodes instead of\nmodifying existing structures or node attributes. Inspired by findings that the\nadversarial attacks are related to the increased heterophily on perturbed\ngraphs (the adversary tends to connect dissimilar nodes), we propose a general\ndefense framework CHAGNN against GIA through cooperative homophilous\naugmentation of graph data and model. Specifically, the model generates\npseudo-labels for unlabeled nodes in each round of training to reduce\nheterophilous edges of nodes with distinct labels. The cleaner graph is fed\nback to the model, producing more informative pseudo-labels. In such an\niterative manner, model robustness is then promisingly enhanced. We present the\ntheoretical analysis of the effect of homophilous augmentation and provide the\nguarantee of the proposal's validity. Experimental results empirically\ndemonstrate the effectiveness of CHAGNN in comparison with recent\nstate-of-the-art defense methods on diverse real-world datasets.\n", "date": "2022-11-15", "categories": []}
{"id": "2211.08384", "title": "Universal Distributional Decision-based Black-box Adversarial Attack\n  with Reinforcement Learning", "abstract": "  The vulnerability of the high-performance machine learning models implies a\nsecurity risk in applications with real-world consequences. Research on\nadversarial attacks is beneficial in guiding the development of machine\nlearning models on the one hand and finding targeted defenses on the other.\nHowever, most of the adversarial attacks today leverage the gradient or logit\ninformation from the models to generate adversarial perturbation. Works in the\nmore realistic domain: decision-based attacks, which generate adversarial\nperturbation solely based on observing the output label of the targeted model,\nare still relatively rare and mostly use gradient-estimation strategies. In\nthis work, we propose a pixel-wise decision-based attack algorithm that finds a\ndistribution of adversarial perturbation through a reinforcement learning\nalgorithm. We call this method Decision-based Black-box Attack with\nReinforcement learning (DBAR). Experiments show that the proposed approach\noutperforms state-of-the-art decision-based attacks with a higher attack\nsuccess rate and greater transferability.\n", "date": "2022-11-15", "categories": []}
{"id": "2207.13036", "title": "Jacobian Norm with Selective Input Gradient Regularization for Improved\n  and Interpretable Adversarial Defense", "abstract": "  Deep neural networks (DNNs) are known to be vulnerable to adversarial\nexamples that are crafted with imperceptible perturbations, i.e., a small\nchange in an input image can induce a mis-classification, and thus threatens\nthe reliability of deep learning based deployment systems. Adversarial training\n(AT) is often adopted to improve robustness through training a mixture of\ncorrupted and clean data. However, most of AT based methods are ineffective in\ndealing with transferred adversarial examples which are generated to fool a\nwide spectrum of defense models, and thus cannot satisfy the generalization\nrequirement raised in real-world scenarios. Moreover, adversarially training a\ndefense model in general cannot produce interpretable predictions towards the\ninputs with perturbations, whilst a highly interpretable robust model is\nrequired by different domain experts to understand the behaviour of a DNN. In\nthis work, we propose a novel approach based on Jacobian norm and Selective\nInput Gradient Regularization (J-SIGR), which suggests the linearized\nrobustness through Jacobian normalization and also regularizes the\nperturbation-based saliency maps to imitate the model's interpretable\npredictions. As such, we achieve both the improved defense and high\ninterpretability of DNNs. Finally, we evaluate our method across different\narchitectures against powerful adversarial attacks. Experiments demonstrate\nthat the proposed J-SIGR confers improved robustness against transferred\nadversarial attacks, and we also show that the predictions from the neural\nnetwork are easy to interpret.\n", "date": "2022-07-09", "categories": []}
{"id": "2210.00621", "title": "Optimization for Robustness Evaluation beyond $\\ell_p$ Metrics", "abstract": "  Empirical evaluation of deep learning models against adversarial attacks\nentails solving nontrivial constrained optimization problems. Popular\nalgorithms for solving these constrained problems rely on projected gradient\ndescent (PGD) and require careful tuning of multiple hyperparameters. Moreover,\nPGD can only handle $\\ell_1$, $\\ell_2$, and $\\ell_\\infty$ attack models due to\nthe use of analytical projectors. In this paper, we introduce a novel\nalgorithmic framework that blends a general-purpose constrained-optimization\nsolver PyGRANSO, With Constraint-Folding (PWCF), to add reliability and\ngenerality to robustness evaluation. PWCF 1) finds good-quality solutions\nwithout the need of delicate hyperparameter tuning, and 2) can handle general\nattack models, e.g., general $\\ell_p$ ($p \\geq 0$) and perceptual attacks,\nwhich are inaccessible to PGD-based algorithms.\n", "date": "2022-10-02", "categories": []}
{"id": "2211.03013", "title": "Robust Lottery Tickets for Pre-trained Language Models", "abstract": "  Recent works on Lottery Ticket Hypothesis have shown that pre-trained\nlanguage models (PLMs) contain smaller matching subnetworks(winning tickets)\nwhich are capable of reaching accuracy comparable to the original models.\nHowever, these tickets are proved to be notrobust to adversarial examples, and\neven worse than their PLM counterparts. To address this problem, we propose a\nnovel method based on learning binary weight masks to identify robust tickets\nhidden in the original PLMs. Since the loss is not differentiable for the\nbinary mask, we assign the hard concrete distribution to the masks and\nencourage their sparsity using a smoothing approximation of L0\nregularization.Furthermore, we design an adversarial loss objective to guide\nthe search for robust tickets and ensure that the tickets perform well bothin\naccuracy and robustness. Experimental results show the significant improvement\nof the proposed method over previous work on adversarial robustness evaluation.\n", "date": "2022-11-06", "categories": []}
{"id": "2211.06508", "title": "On the robustness of non-intrusive speech quality model by adversarial\n  examples", "abstract": "  It has been shown recently that deep learning based models are effective on\nspeech quality prediction and could outperform traditional metrics in various\nperspectives. Although network models have potential to be a surrogate for\ncomplex human hearing perception, they may contain instabilities in\npredictions. This work shows that deep speech quality predictors can be\nvulnerable to adversarial perturbations, where the prediction can be changed\ndrastically by unnoticeable perturbations as small as $-30$ dB compared with\nspeech inputs. In addition to exposing the vulnerability of deep speech quality\npredictors, we further explore and confirm the viability of adversarial\ntraining for strengthening robustness of models.\n", "date": "2022-11-11", "categories": []}
{"id": "2211.06571", "title": "Generating Textual Adversaries with Minimal Perturbation", "abstract": "  Many word-level adversarial attack approaches for textual data have been\nproposed in recent studies. However, due to the massive search space consisting\nof combinations of candidate words, the existing approaches face the problem of\npreserving the semantics of texts when crafting adversarial counterparts. In\nthis paper, we develop a novel attack strategy to find adversarial texts with\nhigh similarity to the original texts while introducing minimal perturbation.\nThe rationale is that we expect the adversarial texts with small perturbation\ncan better preserve the semantic meaning of original texts. Experiments show\nthat, compared with state-of-the-art attack approaches, our approach achieves\nhigher success rates and lower perturbation rates in four benchmark datasets.\n", "date": "2022-11-12", "categories": []}
{"id": "2211.06929", "title": "Goal-Conditioned Reinforcement Learning in the Presence of an Adversary", "abstract": "  Reinforcement learning has seen increasing applications in real-world\ncontexts over the past few years. However, physical environments are often\nimperfect and policies that perform well in simulation might not achieve the\nsame performance when applied elsewhere. A common approach to combat this is to\ntrain agents in the presence of an adversary. An adversary acts to destabilise\nthe agent, which learns a more robust policy and can better handle realistic\nconditions. Many real-world applications of reinforcement learning also make\nuse of goal-conditioning: this is particularly useful in the context of\nrobotics, as it allows the agent to act differently, depending on which goal is\nselected. Here, we focus on the problem of goal-conditioned learning in the\npresence of an adversary. We first present DigitFlip and CLEVR-Play, two novel\ngoal-conditioned environments that support acting against an adversary. Next,\nwe propose EHER and CHER -- two HER-based algorithms for goal-conditioned\nlearning -- and evaluate their performance. Finally, we unify the two threads\nand introduce IGOAL: a novel framework for goal-conditioned learning in the\npresence of an adversary. Experimental results show that combining IGOAL with\nEHER allows agents to significantly outperform existing approaches, when acting\nagainst both random and competent adversaries.\n", "date": "2022-11-13", "categories": []}
{"id": "2211.07483", "title": "Butterfly Effect Attack: Tiny and Seemingly Unrelated Perturbations for\n  Object Detection", "abstract": "  This work aims to explore and identify tiny and seemingly unrelated\nperturbations of images in object detection that will lead to performance\ndegradation. While tininess can naturally be defined using $L_p$ norms, we\ncharacterize the degree of \"unrelatedness\" of an object by the pixel distance\nbetween the occurred perturbation and the object. Triggering errors in\nprediction while satisfying two objectives can be formulated as a\nmulti-objective optimization problem where we utilize genetic algorithms to\nguide the search. The result successfully demonstrates that (invisible)\nperturbations on the right part of the image can drastically change the outcome\nof object detection on the left. An extensive evaluation reaffirms our\nconjecture that transformer-based object detection networks are more\nsusceptible to butterfly effects in comparison to single-stage object detection\nnetworks such as YOLOv5.\n", "date": "2022-11-14", "categories": []}
{"id": "2211.05854", "title": "Test-time adversarial detection and robustness for localizing humans\n  using ultra wide band channel impulse responses", "abstract": "  Keyless entry systems in cars are adopting neural networks for localizing its\noperators. Using test-time adversarial defences equip such systems with the\nability to defend against adversarial attacks without prior training on\nadversarial samples. We propose a test-time adversarial example detector which\ndetects the input adversarial example through quantifying the localized\nintermediate responses of a pre-trained neural network and confidence scores of\nan auxiliary softmax layer. Furthermore, in order to make the network robust,\nwe extenuate the non-relevant features by non-iterative input sample clipping.\nUsing our approach, mean performance over 15 levels of adversarial\nperturbations is increased by 55.33% for the fast gradient sign method (FGSM)\nand 6.3% for both the basic iterative method (BIM) and the projected gradient\nmethod (PGD).\n", "date": "2022-11-10", "categories": []}
{"id": "2210.04610", "title": "Red-Teaming the Stable Diffusion Safety Filter", "abstract": "  Stable Diffusion is a recent open-source image generation model comparable to\nproprietary models such as DALLE, Imagen, or Parti. Stable Diffusion comes with\na safety filter that aims to prevent generating explicit images. Unfortunately,\nthe filter is obfuscated and poorly documented. This makes it hard for users to\nprevent misuse in their applications, and to understand the filter's\nlimitations and improve it. We first show that it is easy to generate\ndisturbing content that bypasses the safety filter. We then reverse-engineer\nthe filter and find that while it aims to prevent sexual content, it ignores\nviolence, gore, and other similarly disturbing content. Based on our analysis,\nwe argue safety measures in future model releases should strive to be fully\nopen and properly documented to stimulate security contributions from the\ncommunity.\n", "date": "2022-10-03", "categories": []}
{"id": "2211.04973", "title": "Accelerating Adversarial Perturbation by 50% with Semi-backward\n  Propagation", "abstract": "  Adversarial perturbation plays a significant role in the field of adversarial\nrobustness, which solves a maximization problem over the input data. We show\nthat the backward propagation of such optimization can accelerate $2\\times$\n(and thus the overall optimization including the forward propagation can\naccelerate $1.5\\times$), without any utility drop, if we only compute the\noutput gradient but not the parameter gradient during the backward propagation.\n", "date": "2022-11-09", "categories": []}
{"id": "2211.04364", "title": "NaturalAdversaries: Can Naturalistic Adversaries Be as Effective as\n  Artificial Adversaries?", "abstract": "  While a substantial body of prior work has explored adversarial example\ngeneration for natural language understanding tasks, these examples are often\nunrealistic and diverge from the real-world data distributions. In this work,\nwe introduce a two-stage adversarial example generation framework\n(NaturalAdversaries), for designing adversaries that are effective at fooling a\ngiven classifier and demonstrate natural-looking failure cases that could\nplausibly occur during in-the-wild deployment of the models.\n  At the first stage a token attribution method is used to summarize a given\nclassifier's behaviour as a function of the key tokens in the input. In the\nsecond stage a generative model is conditioned on the key tokens from the first\nstage. NaturalAdversaries is adaptable to both black-box and white-box\nadversarial attacks based on the level of access to the model parameters. Our\nresults indicate these adversaries generalize across domains, and offer\ninsights for future research on improving robustness of neural text\nclassification models.\n", "date": "2022-11-08", "categories": []}
{"id": "2211.04449", "title": "Fairness-aware Regression Robust to Adversarial Attacks", "abstract": "  In this paper, we take a first step towards answering the question of how to\ndesign fair machine learning algorithms that are robust to adversarial attacks.\nUsing a minimax framework, we aim to design an adversarially robust fair\nregression model that achieves optimal performance in the presence of an\nattacker who is able to add a carefully designed adversarial data point to the\ndataset or perform a rank-one attack on the dataset. By solving the proposed\nnonsmooth nonconvex-nonconcave minimax problem, the optimal adversary as well\nas the robust fairness-aware regression model are obtained. For both synthetic\ndata and real-world datasets, numerical results illustrate that the proposed\nadversarially robust fair models have better performance on poisoned datasets\nthan other fair machine learning models in both prediction accuracy and\ngroup-based fairness measure.\n", "date": "2022-11-04", "categories": []}
{"id": "2211.02675", "title": "An Adversarial Robustness Perspective on the Topology of Neural Networks", "abstract": "  In this paper, we investigate the impact of neural networks (NNs) topology on\nadversarial robustness. Specifically, we study the graph produced when an input\ntraverses all the layers of a NN, and show that such graphs are different for\nclean and adversarial inputs. We find that graphs from clean inputs are more\ncentralized around highway edges, whereas those from adversaries are more\ndiffuse, leveraging under-optimized edges. Through experiments on a variety of\ndatasets and architectures, we show that these under-optimized edges are a\nsource of adversarial vulnerability and that they can be used to detect\nadversarial inputs.\n", "date": "2022-11-04", "categories": []}
{"id": "2211.02878", "title": "Textual Manifold-based Defense Against Natural Language Adversarial\n  Examples", "abstract": "  Recent studies on adversarial images have shown that they tend to leave the\nunderlying low-dimensional data manifold, making them significantly more\nchallenging for current models to make correct predictions. This so-called\noff-manifold conjecture has inspired a novel line of defenses against\nadversarial attacks on images. In this study, we find a similar phenomenon\noccurs in the contextualized embedding space induced by pretrained language\nmodels, in which adversarial texts tend to have their embeddings diverge from\nthe manifold of natural ones. Based on this finding, we propose Textual\nManifold-based Defense (TMD), a defense mechanism that projects text embeddings\nonto an approximated embedding manifold before classification. It reduces the\ncomplexity of potential adversarial examples, which ultimately enhances the\nrobustness of the protected model. Through extensive experiments, our method\nconsistently and significantly outperforms previous defenses under various\nattack settings without trading off clean accuracy. To the best of our\nknowledge, this is the first NLP defense that leverages the manifold structure\nagainst adversarial attacks. Our code is available at\n\\url{https://github.com/dangne/tmd}.\n", "date": "2022-11-05", "categories": []}
{"id": "2211.03509", "title": "Black-Box Attack against GAN-Generated Image Detector with Contrastive\n  Perturbation", "abstract": "  Visually realistic GAN-generated facial images raise obvious concerns on\npotential misuse. Many effective forensic algorithms have been developed to\ndetect such synthetic images in recent years. It is significant to assess the\nvulnerability of such forensic detectors against adversarial attacks. In this\npaper, we propose a new black-box attack method against GAN-generated image\ndetectors. A novel contrastive learning strategy is adopted to train the\nencoder-decoder network based anti-forensic model under a contrastive loss\nfunction. GAN images and their simulated real counterparts are constructed as\npositive and negative samples, respectively. Leveraging on the trained attack\nmodel, imperceptible contrastive perturbation could be applied to input\nsynthetic images for removing GAN fingerprint to some extent. As such, existing\nGAN-generated image detectors are expected to be deceived. Extensive\nexperimental results verify that the proposed attack effectively reduces the\naccuracy of three state-of-the-art detectors on six popular GANs. High visual\nquality of the attacked images is also achieved. The source code will be\navailable at https://github.com/ZXMMD/BAttGAND.\n", "date": "2022-11-07", "categories": []}
{"id": "2211.03714", "title": "Deviations in Representations Induced by Adversarial Attacks", "abstract": "  Deep learning has been a popular topic and has achieved success in many\nareas. It has drawn the attention of researchers and machine learning\npractitioners alike, with developed models deployed to a variety of settings.\nAlong with its achievements, research has shown that deep learning models are\nvulnerable to adversarial attacks. This finding brought about a new direction\nin research, whereby algorithms were developed to attack and defend vulnerable\nnetworks. Our interest is in understanding how these attacks effect change on\nthe intermediate representations of deep learning models. We present a method\nfor measuring and analyzing the deviations in representations induced by\nadversarial attacks, progressively across a selected set of layers. Experiments\nare conducted using an assortment of attack algorithms, on the CIFAR-10\ndataset, with plots created to visualize the impact of adversarial attacks\nacross different layers in a network.\n", "date": "2022-11-07", "categories": []}
{"id": "2211.03769", "title": "Are AlphaZero-like Agents Robust to Adversarial Perturbations?", "abstract": "  The success of AlphaZero (AZ) has demonstrated that neural-network-based Go\nAIs can surpass human performance by a large margin. Given that the state space\nof Go is extremely large and a human player can play the game from any legal\nstate, we ask whether adversarial states exist for Go AIs that may lead them to\nplay surprisingly wrong actions. In this paper, we first extend the concept of\nadversarial examples to the game of Go: we generate perturbed states that are\n``semantically'' equivalent to the original state by adding meaningless moves\nto the game, and an adversarial state is a perturbed state leading to an\nundoubtedly inferior action that is obvious even for Go beginners. However,\nsearching the adversarial state is challenging due to the large, discrete, and\nnon-differentiable search space. To tackle this challenge, we develop the first\nadversarial attack on Go AIs that can efficiently search for adversarial states\nby strategically reducing the search space. This method can also be extended to\nother board games such as NoGo. Experimentally, we show that the actions taken\nby both Policy-Value neural network (PV-NN) and Monte Carlo tree search (MCTS)\ncan be misled by adding one or two meaningless stones; for example, on 58\\% of\nthe AlphaGo Zero self-play games, our method can make the widely used KataGo\nagent with 50 simulations of MCTS plays a losing action by adding two\nmeaningless stones. We additionally evaluated the adversarial examples found by\nour algorithm with amateur human Go players and 90\\% of examples indeed lead\nthe Go agent to play an obviously inferior action. Our code is available at\n\\url{https://PaperCode.cc/GoAttack}.\n", "date": "2022-11-07", "categories": []}
{"id": "2211.02468", "title": "Improving Adversarial Robustness to Sensitivity and Invariance Attacks\n  with Deep Metric Learning", "abstract": "  Intentionally crafted adversarial samples have effectively exploited\nweaknesses in deep neural networks. A standard method in adversarial robustness\nassumes a framework to defend against samples crafted by minimally perturbing a\nsample such that its corresponding model output changes. These sensitivity\nattacks exploit the model's sensitivity toward task-irrelevant features.\nAnother form of adversarial sample can be crafted via invariance attacks, which\nexploit the model underestimating the importance of relevant features. Previous\nliterature has indicated a tradeoff in defending against both attack types\nwithin a strictly L_p bounded defense. To promote robustness toward both types\nof attacks beyond Euclidean distance metrics, we use metric learning to frame\nadversarial regularization as an optimal transport problem. Our preliminary\nresults indicate that regularizing over invariant perturbations in our\nframework improves both invariant and sensitivity defense.\n", "date": "2022-11-04", "categories": []}
{"id": "2211.01621", "title": "Leveraging Domain Features for Detecting Adversarial Attacks Against\n  Deep Speech Recognition in Noise", "abstract": "  In recent years, significant progress has been made in deep model-based\nautomatic speech recognition (ASR), leading to its widespread deployment in the\nreal world. At the same time, adversarial attacks against deep ASR systems are\nhighly successful. Various methods have been proposed to defend ASR systems\nfrom these attacks. However, existing classification based methods focus on the\ndesign of deep learning models while lacking exploration of domain specific\nfeatures. This work leverages filter bank-based features to better capture the\ncharacteristics of attacks for improved detection. Furthermore, the paper\nanalyses the potentials of using speech and non-speech parts separately in\ndetecting adversarial attacks. In the end, considering adverse environments\nwhere ASR systems may be deployed, we study the impact of acoustic noise of\nvarious types and signal-to-noise ratios. Extensive experiments show that the\ninverse filter bank features generally perform better in both clean and noisy\nenvironments, the detection is effective using either speech or non-speech\npart, and the acoustic noise can largely degrade the detection performance.\n", "date": "2022-11-03", "categories": []}
{"id": "2207.04497", "title": "One-shot Neural Backdoor Erasing via Adversarial Weight Masking", "abstract": "  Recent studies show that despite achieving high accuracy on a number of\nreal-world applications, deep neural networks (DNNs) can be backdoored: by\ninjecting triggered data samples into the training dataset, the adversary can\nmislead the trained model into classifying any test data to the target class as\nlong as the trigger pattern is presented. To nullify such backdoor threats,\nvarious methods have been proposed. Particularly, a line of research aims to\npurify the potentially compromised model. However, one major limitation of this\nline of work is the requirement to access sufficient original training data:\nthe purifying performance is a lot worse when the available training data is\nlimited. In this work, we propose Adversarial Weight Masking (AWM), a novel\nmethod capable of erasing the neural backdoors even in the one-shot setting.\nThe key idea behind our method is to formulate this into a min-max optimization\nproblem: first, adversarially recover the trigger patterns and then (soft) mask\nthe network weights that are sensitive to the recovered patterns. Comprehensive\nevaluations of several benchmark datasets suggest that AWM can largely improve\nthe purifying effects over other state-of-the-art methods on various available\ntraining dataset sizes.\n", "date": "2022-07-10", "categories": []}
{"id": "2211.01086", "title": "Generative Poisoning Using Random Discriminators", "abstract": "  We introduce ShortcutGen, a new data poisoning attack that generates\nsample-dependent, error-minimizing perturbations by learning a generator. The\nkey novelty of ShortcutGen is the use of a randomly-initialized discriminator,\nwhich provides spurious shortcuts needed for generating poisons. Different from\nrecent, iterative methods, our ShortcutGen can generate perturbations with only\none forward pass in a label-free manner, and compared to the only existing\ngenerative method, DeepConfuse, our ShortcutGen is faster and simpler to train\nwhile remaining competitive. We also demonstrate that integrating a simple\naugmentation strategy can further boost the robustness of ShortcutGen against\nearly stopping, and combining augmentation and non-augmentation leads to new\nstate-of-the-art results in terms of final validation accuracy, especially in\nthe challenging, transfer scenario. Lastly, we speculate, through uncovering\nits working mechanism, that learning a more general representation space could\nallow ShortcutGen to work for unseen data.\n", "date": "2022-11-02", "categories": []}
{"id": "2211.01093", "title": "Improving transferability of 3D adversarial attacks with scale and shear\n  transformations", "abstract": "  Previous work has shown that 3D point cloud classifiers can be vulnerable to\nadversarial examples. However, most of the existing methods are aimed at\nwhite-box attacks, where the parameters and other information of the\nclassifiers are known in the attack, which is unrealistic for real-world\napplications. In order to improve the attack performance of the black-box\nclassifiers, the research community generally uses the transfer-based black-box\nattack. However, the transferability of current 3D attacks is still relatively\nlow. To this end, this paper proposes Scale and Shear (SS) Attack to generate\n3D adversarial examples with strong transferability. Specifically, we randomly\nscale or shear the input point cloud, so that the attack will not overfit the\nwhite-box model, thereby improving the transferability of the attack. Extensive\nexperiments show that the SS attack proposed in this paper can be seamlessly\ncombined with the existing state-of-the-art (SOTA) 3D point cloud attack\nmethods to form more powerful attack methods, and the SS attack improves the\ntransferability over 3.6 times compare to the baseline. Moreover, while\nsubstantially outperforming the baseline methods, the SS attack achieves SOTA\ntransferability under various defenses. Our code will be available online at\nhttps://github.com/cuge1995/SS-attack\n", "date": "2022-11-02", "categories": []}
{"id": "2211.00239", "title": "ARDIR: Improving Robustness using Knowledge Distillation of Internal\n  Representation", "abstract": "  Adversarial training is the most promising method for learning robust models\nagainst adversarial examples. A recent study has shown that knowledge\ndistillation between the same architectures is effective in improving the\nperformance of adversarial training. Exploiting knowledge distillation is a new\napproach to improve adversarial training and has attracted much attention.\nHowever, its performance is still insufficient. Therefore, we propose\nAdversarial Robust Distillation with Internal Representation~(ARDIR) to utilize\nknowledge distillation even more effectively. In addition to the output of the\nteacher model, ARDIR uses the internal representation of the teacher model as a\nlabel for adversarial training. This enables the student model to be trained\nwith richer, more informative labels. As a result, ARDIR can learn more robust\nstudent models. We show that ARDIR outperforms previous methods in our\nexperiments.\n", "date": "2022-11-01", "categories": []}
{"id": "2211.00269", "title": "Adversarial Training with Complementary Labels: On the Benefit of\n  Gradually Informative Attacks", "abstract": "  Adversarial training (AT) with imperfect supervision is significant but\nreceives limited attention. To push AT towards more practical scenarios, we\nexplore a brand new yet challenging setting, i.e., AT with complementary labels\n(CLs), which specify a class that a data sample does not belong to. However,\nthe direct combination of AT with existing methods for CLs results in\nconsistent failure, but not on a simple baseline of two-stage training. In this\npaper, we further explore the phenomenon and identify the underlying challenges\nof AT with CLs as intractable adversarial optimization and low-quality\nadversarial examples. To address the above problems, we propose a new learning\nstrategy using gradually informative attacks, which consists of two critical\ncomponents: 1) Warm-up Attack (Warm-up) gently raises the adversarial\nperturbation budgets to ease the adversarial optimization with CLs; 2)\nPseudo-Label Attack (PLA) incorporates the progressively informative model\npredictions into a corrected complementary loss. Extensive experiments are\nconducted to demonstrate the effectiveness of our method on a range of\nbenchmarked datasets. The code is publicly available at:\nhttps://github.com/RoyalSkye/ATCL.\n", "date": "2022-11-01", "categories": []}
{"id": "2211.00322", "title": "DensePure: Understanding Diffusion Models towards Adversarial Robustness", "abstract": "  Diffusion models have been recently employed to improve certified robustness\nthrough the process of denoising. However, the theoretical understanding of why\ndiffusion models are able to improve the certified robustness is still lacking,\npreventing from further improvement. In this study, we close this gap by\nanalyzing the fundamental properties of diffusion models and establishing the\nconditions under which they can enhance certified robustness. This deeper\nunderstanding allows us to propose a new method DensePure, designed to improve\nthe certified robustness of a pretrained model (i.e. classifier). Given an\n(adversarial) input, DensePure consists of multiple runs of denoising via the\nreverse process of the diffusion model (with different random seeds) to get\nmultiple reversed samples, which are then passed through the classifier,\nfollowed by majority voting of inferred labels to make the final prediction.\nThis design of using multiple runs of denoising is informed by our theoretical\nanalysis of the conditional distribution of the reversed sample. Specifically,\nwhen the data density of a clean sample is high, its conditional density under\nthe reverse process in a diffusion model is also high; thus sampling from the\nlatter conditional distribution can purify the adversarial example and return\nthe corresponding clean sample with a high probability. By using the highest\ndensity point in the conditional distribution as the reversed sample, we\nidentify the robust region of a given instance under the diffusion model's\nreverse process. We show that this robust region is a union of multiple convex\nsets, and is potentially much larger than the robust regions identified in\nprevious works. In practice, DensePure can approximate the label of the high\ndensity region in the conditional distribution so that it can enhance certified\nrobustness.\n", "date": "2022-11-01", "categories": []}
{"id": "2211.00366", "title": "Universal Perturbation Attack on Differentiable No-Reference Image- and\n  Video-Quality Metrics", "abstract": "  Universal adversarial perturbation attacks are widely used to analyze image\nclassifiers that employ convolutional neural networks. Nowadays, some attacks\ncan deceive image- and video-quality metrics. So sustainability analysis of\nthese metrics is important. Indeed, if an attack can confuse the metric, an\nattacker can easily increase quality scores. When developers of image- and\nvideo-algorithms can boost their scores through detached processing, algorithm\ncomparisons are no longer fair. Inspired by the idea of universal adversarial\nperturbation for classifiers, we suggest a new method to attack differentiable\nno-reference quality metrics through universal perturbation. We applied this\nmethod to seven no-reference image- and video-quality metrics (PaQ-2-PiQ,\nLinearity, VSFA, MDTVSFA, KonCept512, Nima and SPAQ). For each one, we trained\na universal perturbation that increases the respective scores. We also propose\na method for assessing metric stability and identify the metrics that are the\nmost vulnerable and the most resistant to our attack. The existence of\nsuccessful universal perturbations appears to diminish the metric's ability to\nprovide reliable scores. We therefore recommend our proposed method as an\nadditional verification of metric reliability to complement traditional\nsubjective tests and benchmarks.\n", "date": "2022-11-01", "categories": []}
{"id": "2211.00525", "title": "The Enemy of My Enemy is My Friend: Exploring Inverse Adversaries for\n  Improving Adversarial Training", "abstract": "  Although current deep learning techniques have yielded superior performance\non various computer vision tasks, yet they are still vulnerable to adversarial\nexamples. Adversarial training and its variants have been shown to be the most\neffective approaches to defend against adversarial examples. These methods\nusually regularize the difference between output probabilities for an\nadversarial and its corresponding natural example. However, it may have a\nnegative impact if the model misclassifies a natural example. To circumvent\nthis issue, we propose a novel adversarial training scheme that encourages the\nmodel to produce similar outputs for an adversarial example and its ``inverse\nadversarial'' counterpart. These samples are generated to maximize the\nlikelihood in the neighborhood of natural examples. Extensive experiments on\nvarious vision datasets and architectures demonstrate that our training method\nachieves state-of-the-art robustness as well as natural accuracy. Furthermore,\nusing a universal version of inverse adversarial examples, we improve the\nperformance of single-step adversarial training techniques at a low\ncomputational cost.\n", "date": "2022-11-01", "categories": []}
{"id": "2207.05127", "title": "RUSH: Robust Contrastive Learning via Randomized Smoothing", "abstract": "  Recently, adversarial training has been incorporated in self-supervised\ncontrastive pre-training to augment label efficiency with exciting adversarial\nrobustness. However, the robustness came at a cost of expensive adversarial\ntraining. In this paper, we show a surprising fact that contrastive\npre-training has an interesting yet implicit connection with robustness, and\nsuch natural robustness in the pre trained representation enables us to design\na powerful robust algorithm against adversarial attacks, RUSH, that combines\nthe standard contrastive pre-training and randomized smoothing. It boosts both\nstandard accuracy and robust accuracy, and significantly reduces training costs\nas compared with adversarial training. We use extensive empirical studies to\nshow that the proposed RUSH outperforms robust classifiers from adversarial\ntraining, by a significant margin on common benchmarks (CIFAR-10, CIFAR-100,\nand STL-10) under first-order attacks. In particular, under\n$\\ell_{\\infty}$-norm perturbations of size 8/255 PGD attack on CIFAR-10, our\nmodel using ResNet-18 as backbone reached 77.8% robust accuracy and 87.9%\nstandard accuracy. Our work has an improvement of over 15% in robust accuracy\nand a slight improvement in standard accuracy, compared to the\nstate-of-the-arts.\n", "date": "2022-07-11", "categories": []}
{"id": "2210.00960", "title": "Stability Analysis and Generalization Bounds of Adversarial Training", "abstract": "  In adversarial machine learning, deep neural networks can fit the adversarial\nexamples on the training dataset but have poor generalization ability on the\ntest set. This phenomenon is called robust overfitting, and it can be observed\nwhen adversarially training neural nets on common datasets, including SVHN,\nCIFAR-10, CIFAR-100, and ImageNet. In this paper, we study the robust\noverfitting issue of adversarial training by using tools from uniform\nstability. One major challenge is that the outer function (as a maximization of\nthe inner function) is nonsmooth, so the standard technique (e.g., hardt et\nal., 2016) cannot be applied. Our approach is to consider $\\eta$-approximate\nsmoothness: we show that the outer function satisfies this modified smoothness\nassumption with $\\eta$ being a constant related to the adversarial perturbation\n$\\epsilon$. Based on this, we derive stability-based generalization bounds for\nstochastic gradient descent (SGD) on the general class of $\\eta$-approximate\nsmooth functions, which covers the adversarial loss. Our results suggest that\nrobust test accuracy decreases in $\\epsilon$ when $T$ is large, with a speed\nbetween $\\Omega(\\epsilon\\sqrt{T})$ and $\\mathcal{O}(\\epsilon T)$. This\nphenomenon is also observed in practice. Additionally, we show that a few\npopular techniques for adversarial training (e.g., early stopping, cyclic\nlearning rate, and stochastic weight averaging) are stability-promoting in\ntheory.\n", "date": "2022-10-03", "categories": []}
{"id": "2210.11498", "title": "Balanced Adversarial Training: Balancing Tradeoffs between Fickleness\n  and Obstinacy in NLP Models", "abstract": "  Traditional (fickle) adversarial examples involve finding a small\nperturbation that does not change an input's true label but confuses the\nclassifier into outputting a different prediction. Conversely, obstinate\nadversarial examples occur when an adversary finds a small perturbation that\npreserves the classifier's prediction but changes the true label of an input.\nAdversarial training and certified robust training have shown some\neffectiveness in improving the robustness of machine learnt models to fickle\nadversarial examples. We show that standard adversarial training methods\nfocused on reducing vulnerability to fickle adversarial examples may make a\nmodel more vulnerable to obstinate adversarial examples, with experiments for\nboth natural language inference and paraphrase identification tasks. To counter\nthis phenomenon, we introduce Balanced Adversarial Training, which incorporates\ncontrastive learning to increase robustness against both fickle and obstinate\nadversarial examples.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.16451", "title": "Robust Boosting Forests with Richer Deep Feature Hierarchy", "abstract": "  We propose a robust variant of boosting forest to the various adversarial\ndefense methods, and apply it to enhance the robustness of the deep neural\nnetwork. We retain the deep network architecture, weights, and middle layer\nfeatures, then install gradient boosting forest to select the features from\neach layer of the deep network, and predict the target. For training each\ndecision tree, we propose a novel conservative and greedy trade-off, with\nconsideration for less misprediction instead of pure gain functions, therefore\nbeing suboptimal and conservative. We actively increase tree depth to remedy\nthe accuracy with splits in more features, being more greedy in growing tree\ndepth. We propose a new task on 3D face model, whose robustness has not been\ncarefully studied, despite the great security and privacy concerns related to\nface analytics. We tried a simple attack method on a pure convolutional neural\nnetwork (CNN) face shape estimator, making it degenerate to only output average\nface shape with invisible perturbation. Our conservative-greedy boosting forest\n(CGBF) on face landmark datasets showed a great improvement over original pure\ndeep learning methods under the adversarial attacks.\n", "date": "2022-10-29", "categories": []}
{"id": "2210.16915", "title": "Imitating Opponent to Win: Adversarial Policy Imitation Learning in\n  Two-player Competitive Games", "abstract": "  Recent research on vulnerabilities of deep reinforcement learning (RL) has\nshown that adversarial policies adopted by an adversary agent can influence a\ntarget RL agent (victim agent) to perform poorly in a multi-agent environment.\nIn existing studies, adversarial policies are directly trained based on\nexperiences of interacting with the victim agent. There is a key shortcoming of\nthis approach; knowledge derived from historical interactions may not be\nproperly generalized to unexplored policy regions of the victim agent, making\nthe trained adversarial policy significantly less effective. In this work, we\ndesign a new effective adversarial policy learning algorithm that overcomes\nthis shortcoming. The core idea of our new algorithm is to create a new\nimitator to imitate the victim agent's policy while the adversarial policy will\nbe trained not only based on interactions with the victim agent but also based\non feedback from the imitator to forecast victim's intention. By doing so, we\ncan leverage the capability of imitation learning in well capturing underlying\ncharacteristics of the victim policy only based on sample trajectories of the\nvictim. Our victim imitation learning model differs from prior models as the\nenvironment's dynamics are driven by adversary's policy and will keep changing\nduring the adversarial policy training. We provide a provable bound to\nguarantee a desired imitating policy when the adversary's policy becomes\nstable. We further strengthen our adversarial policy learning by making our\nimitator a stronger version of the victim. Finally, our extensive experiments\nusing four competitive MuJoCo game environments show that our proposed\nadversarial policy learning algorithm outperforms state-of-the-art algorithms.\n", "date": "2022-10-30", "categories": []}
{"id": "2210.17004", "title": "Character-level White-Box Adversarial Attacks against Transformers via\n  Attachable Subwords Substitution", "abstract": "  We propose the first character-level white-box adversarial attack method\nagainst transformer models. The intuition of our method comes from the\nobservation that words are split into subtokens before being fed into the\ntransformer models and the substitution between two close subtokens has a\nsimilar effect to the character modification. Our method mainly contains three\nsteps. First, a gradient-based method is adopted to find the most vulnerable\nwords in the sentence. Then we split the selected words into subtokens to\nreplace the origin tokenization result from the transformer tokenizer. Finally,\nwe utilize an adversarial loss to guide the substitution of attachable\nsubtokens in which the Gumbel-softmax trick is introduced to ensure gradient\npropagation. Meanwhile, we introduce the visual and length constraint in the\noptimization process to achieve minimum character modifications. Extensive\nexperiments on both sentence-level and token-level tasks demonstrate that our\nmethod could outperform the previous attack methods in terms of success rate\nand edit distance. Furthermore, human evaluation verifies our adversarial\nexamples could preserve their origin labels.\n", "date": "2022-10-31", "categories": []}
{"id": "2210.17140", "title": "Scoring Black-Box Models for Adversarial Robustness", "abstract": "  Deep neural networks are susceptible to adversarial inputs and various\nmethods have been proposed to defend these models against adversarial attacks\nunder different perturbation models. The robustness of models to adversarial\nattacks has been analyzed by first constructing adversarial inputs for the\nmodel, and then testing the model performance on the constructed adversarial\ninputs. Most of these attacks require the model to be white-box, need access to\ndata labels, and finding adversarial inputs can be computationally expensive.\nWe propose a simple scoring method for black-box models which indicates their\nrobustness to adversarial input. We show that adversarially more robust models\nhave a smaller $l_1$-norm of LIME weights and sharper explanations.\n", "date": "2022-10-31", "categories": []}
{"id": "2210.15700", "title": "TAD: Transfer Learning-based Multi-Adversarial Detection of Evasion\n  Attacks against Network Intrusion Detection Systems", "abstract": "  Nowadays, intrusion detection systems based on deep learning deliver\nstate-of-the-art performance. However, recent research has shown that specially\ncrafted perturbations, called adversarial examples, are capable of\nsignificantly reducing the performance of these intrusion detection systems.\nThe objective of this paper is to design an efficient transfer learning-based\nadversarial detector and then to assess the effectiveness of using multiple\nstrategically placed adversarial detectors compared to a single adversarial\ndetector for intrusion detection systems. In our experiments, we implement\nexisting state-of-the-art models for intrusion detection. We then attack those\nmodels with a set of chosen evasion attacks. In an attempt to detect those\nadversarial attacks, we design and implement multiple transfer learning-based\nadversarial detectors, each receiving a subset of the information passed\nthrough the IDS. By combining their respective decisions, we illustrate that\ncombining multiple detectors can further improve the detectability of\nadversarial traffic compared to a single detector in the case of a parallel IDS\ndesign.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.15944", "title": "RoChBert: Towards Robust BERT Fine-tuning for Chinese", "abstract": "  Despite of the superb performance on a wide range of tasks, pre-trained\nlanguage models (e.g., BERT) have been proved vulnerable to adversarial texts.\nIn this paper, we present RoChBERT, a framework to build more Robust BERT-based\nmodels by utilizing a more comprehensive adversarial graph to fuse Chinese\nphonetic and glyph features into pre-trained representations during\nfine-tuning. Inspired by curriculum learning, we further propose to augment the\ntraining dataset with adversarial texts in combination with intermediate\nsamples. Extensive experiments demonstrate that RoChBERT outperforms previous\nmethods in significant ways: (i) robust -- RoChBERT greatly improves the model\nrobustness without sacrificing accuracy on benign texts. Specifically, the\ndefense lowers the success rates of unlimited and limited attacks by 59.43% and\n39.33% respectively, while remaining accuracy of 93.30%; (ii) flexible --\nRoChBERT can easily extend to various language models to solve different\ndownstream tasks with excellent performance; and (iii) efficient -- RoChBERT\ncan be directly applied to the fine-tuning stage without pre-training language\nmodel from scratch, and the proposed data augmentation method is also low-cost.\n", "date": "2022-10-28", "categories": []}
{"id": "2210.01787", "title": "Rethinking Lipschitz Neural Networks and Certified Robustness: A Boolean\n  Function Perspective", "abstract": "  Designing neural networks with bounded Lipschitz constant is a promising way\nto obtain certifiably robust classifiers against adversarial examples. However,\nthe relevant progress for the important $\\ell_\\infty$ perturbation setting is\nrather limited, and a principled understanding of how to design expressive\n$\\ell_\\infty$ Lipschitz networks is still lacking. In this paper, we bridge the\ngap by studying certified $\\ell_\\infty$ robustness from a novel perspective of\nrepresenting Boolean functions. We derive two fundamental impossibility results\nthat hold for any standard Lipschitz network: one for robust classification on\nfinite datasets, and the other for Lipschitz function approximation. These\nresults identify that networks built upon norm-bounded affine layers and\nLipschitz activations intrinsically lose expressive power even in the\ntwo-dimensional case, and shed light on how recently proposed Lipschitz\nnetworks (e.g., GroupSort and $\\ell_\\infty$-distance nets) bypass these\nimpossibilities by leveraging order statistic functions. Finally, based on\nthese insights, we develop a unified Lipschitz network that generalizes prior\nworks, and design a practical version that can be efficiently trained (making\ncertified robust training free). Extensive experiments show that our approach\nis scalable, efficient, and consistently yields better certified robustness\nacross multiple datasets and perturbation radii than prior Lipschitz networks.\nOur code is available at https://github.com/zbh2047/SortNet.\n", "date": "2022-10-04", "categories": []}
{"id": "2210.14944", "title": "Detection and Prevention Against Poisoning Attacks in Federated Learning", "abstract": "  This paper proposes and investigates a new approach for detecting and\npreventing several different types of poisoning attacks from affecting a\ncentralized Federated Learning model via average accuracy deviation detection\n(AADD). By comparing each client's accuracy to all clients' average accuracy,\nAADD detect clients with an accuracy deviation. The implementation is further\nable to blacklist clients that are considered poisoned, securing the global\nmodel from being affected by the poisoned nodes. The proposed implementation\nshows promising results in detecting poisoned clients and preventing the global\nmodel's accuracy from deteriorating.\n", "date": "2022-10-24", "categories": []}
{"id": "2210.14957", "title": "Disentangled Text Representation Learning with Information-Theoretic\n  Perspective for Adversarial Robustness", "abstract": "  Adversarial vulnerability remains a major obstacle to constructing reliable\nNLP systems. When imperceptible perturbations are added to raw input text, the\nperformance of a deep learning model may drop dramatically under attacks.\nRecent work argues the adversarial vulnerability of the model is caused by the\nnon-robust features in supervised training. Thus in this paper, we tackle the\nadversarial robustness challenge from the view of disentangled representation\nlearning, which is able to explicitly disentangle robust and non-robust\nfeatures in text. Specifically, inspired by the variation of information (VI)\nin information theory, we derive a disentangled learning objective composed of\nmutual information to represent both the semantic representativeness of latent\nembeddings and differentiation of robust and non-robust features. On the basis\nof this, we design a disentangled learning network to estimate these mutual\ninformation. Experiments on text classification and entailment tasks show that\nour method significantly outperforms the representative methods under\nadversarial attacks, indicating that discarding non-robust features is critical\nfor improving adversarial robustness.\n", "date": "2022-10-26", "categories": []}
{"id": "2210.15127", "title": "Rethinking the Reverse-engineering of Trojan Triggers", "abstract": "  Deep Neural Networks are vulnerable to Trojan (or backdoor) attacks.\nReverse-engineering methods can reconstruct the trigger and thus identify\naffected models. Existing reverse-engineering methods only consider input space\nconstraints, e.g., trigger size in the input space. Expressly, they assume the\ntriggers are static patterns in the input space and fail to detect models with\nfeature space triggers such as image style transformations. We observe that\nboth input-space and feature-space Trojans are associated with feature space\nhyperplanes. Based on this observation, we design a novel reverse-engineering\nmethod that exploits the feature space constraint to reverse-engineer Trojan\ntriggers. Results on four datasets and seven different attacks demonstrate that\nour solution effectively defends both input-space and feature-space Trojans. It\noutperforms state-of-the-art reverse-engineering methods and other types of\ndefenses in both Trojaned model detection and mitigation tasks. On average, the\ndetection accuracy of our method is 93\\%. For Trojan mitigation, our method can\nreduce the ASR (attack success rate) to only 0.26\\% with the BA (benign\naccuracy) remaining nearly unchanged. Our code can be found at\nhttps://github.com/RU-System-Software-and-Security/FeatureRE.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.15221", "title": "TASA: Deceiving Question Answering Models by Twin Answer Sentences\n  Attack", "abstract": "  We present Twin Answer Sentences Attack (TASA), an adversarial attack method\nfor question answering (QA) models that produces fluent and grammatical\nadversarial contexts while maintaining gold answers. Despite phenomenal\nprogress on general adversarial attacks, few works have investigated the\nvulnerability and attack specifically for QA models. In this work, we first\nexplore the biases in the existing models and discover that they mainly rely on\nkeyword matching between the question and context, and ignore the relevant\ncontextual relations for answer prediction. Based on two biases above, TASA\nattacks the target model in two folds: (1) lowering the model's confidence on\nthe gold answer with a perturbed answer sentence; (2) misguiding the model\ntowards a wrong answer with a distracting answer sentence. Equipped with\ndesigned beam search and filtering methods, TASA can generate more effective\nattacks than existing textual attack methods while sustaining the quality of\ncontexts, in extensive experiments on five QA datasets and human evaluations.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.15291", "title": "Isometric 3D Adversarial Examples in the Physical World", "abstract": "  3D deep learning models are shown to be as vulnerable to adversarial examples\nas 2D models. However, existing attack methods are still far from stealthy and\nsuffer from severe performance degradation in the physical world. Although 3D\ndata is highly structured, it is difficult to bound the perturbations with\nsimple metrics in the Euclidean space. In this paper, we propose a novel\n$\\epsilon$-isometric ($\\epsilon$-ISO) attack to generate natural and robust 3D\nadversarial examples in the physical world by considering the geometric\nproperties of 3D objects and the invariance to physical transformations. For\nnaturalness, we constrain the adversarial example to be $\\epsilon$-isometric to\nthe original one by adopting the Gaussian curvature as a surrogate metric\nguaranteed by a theoretical analysis. For invariance to physical\ntransformations, we propose a maxima over transformation (MaxOT) method that\nactively searches for the most harmful transformations rather than random ones\nto make the generated adversarial example more robust in the physical world.\nExperiments on typical point cloud recognition models validate that our\napproach can significantly improve the attack success rate and naturalness of\nthe generated 3D adversarial examples than the state-of-the-art attack methods.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.15318", "title": "Efficient and Effective Augmentation Strategy for Adversarial Training", "abstract": "  Adversarial training of Deep Neural Networks is known to be significantly\nmore data-hungry when compared to standard training. Furthermore, complex data\naugmentations such as AutoAugment, which have led to substantial gains in\nstandard training of image classifiers, have not been successful with\nAdversarial Training. We first explain this contrasting behavior by viewing\naugmentation during training as a problem of domain generalization, and further\npropose Diverse Augmentation-based Joint Adversarial Training (DAJAT) to use\ndata augmentations effectively in adversarial training. We aim to handle the\nconflicting goals of enhancing the diversity of the training dataset and\ntraining with data that is close to the test distribution by using a\ncombination of simple and complex augmentations with separate batch\nnormalization layers during training. We further utilize the popular\nJensen-Shannon divergence loss to encourage the joint learning of the diverse\naugmentations, thereby allowing simple augmentations to guide the learning of\ncomplex ones. Lastly, to improve the computational efficiency of the proposed\nmethod, we propose and utilize a two-step defense, Ascending Constraint\nAdversarial Training (ACAT), that uses an increasing epsilon schedule and\nweight-space smoothing to prevent gradient masking. The proposed method DAJAT\nachieves substantially better robustness-accuracy trade-off when compared to\nexisting methods on the RobustBench Leaderboard on ResNet-18 and\nWideResNet-34-10. The code for implementing DAJAT is available here:\nhttps://github.com/val-iisc/DAJAT.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.15429", "title": "Multi-view Representation Learning from Malware to Defend Against\n  Adversarial Variants", "abstract": "  Deep learning-based adversarial malware detectors have yielded promising\nresults in detecting never-before-seen malware executables without relying on\nexpensive dynamic behavior analysis and sandbox. Despite their abilities, these\ndetectors have been shown to be vulnerable to adversarial malware variants -\nmeticulously modified, functionality-preserving versions of original malware\nexecutables generated by machine learning. Due to the nature of these\nadversarial modifications, these adversarial methods often use a \\textit{single\nview} of malware executables (i.e., the binary/hexadecimal view) to generate\nadversarial malware variants. This provides an opportunity for the defenders\n(i.e., malware detectors) to detect the adversarial variants by utilizing more\nthan one view of a malware file (e.g., source code view in addition to the\nbinary view). The rationale behind this idea is that while the adversary\nfocuses on the binary view, certain characteristics of the malware file in the\nsource code view remain untouched which leads to the detection of the\nadversarial malware variants. To capitalize on this opportunity, we propose\nAdversarially Robust Multiview Malware Defense (ARMD), a novel multi-view\nlearning framework to improve the robustness of DL-based malware detectors\nagainst adversarial variants. Our experiments on three renowned open-source\ndeep learning-based malware detectors across six common malware categories show\nthat ARMD is able to improve the adversarial robustness by up to seven times on\nthese malware detectors.\n", "date": "2022-10-25", "categories": []}
{"id": "2210.14229", "title": "Causal Information Bottleneck Boosts Adversarial Robustness of Deep\n  Neural Network", "abstract": "  The information bottleneck (IB) method is a feasible defense solution against\nadversarial attacks in deep learning. However, this method suffers from the\nspurious correlation, which leads to the limitation of its further improvement\nof adversarial robustness. In this paper, we incorporate the causal inference\ninto the IB framework to alleviate such a problem. Specifically, we divide the\nfeatures obtained by the IB method into robust features (content information)\nand non-robust features (style information) via the instrumental variables to\nestimate the causal effects. With the utilization of such a framework, the\ninfluence of non-robust features could be mitigated to strengthen the\nadversarial robustness. We make an analysis of the effectiveness of our\nproposed method. The extensive experiments in MNIST, FashionMNIST, and CIFAR-10\nshow that our method exhibits the considerable robustness against multiple\nadversarial attacks. Our code would be released.\n", "date": "2022-10-25", "categories": []}
{"id": "2210.14260", "title": "Universal Evasion Attacks on Summarization Scoring", "abstract": "  The automatic scoring of summaries is important as it guides the development\nof summarizers. Scoring is also complex, as it involves multiple aspects such\nas fluency, grammar, and even textual entailment with the source text. However,\nsummary scoring has not been considered a machine learning task to study its\naccuracy and robustness. In this study, we place automatic scoring in the\ncontext of regression machine learning tasks and perform evasion attacks to\nexplore its robustness. Attack systems predict a non-summary string from each\ninput, and these non-summary strings achieve competitive scores with good\nsummarizers on the most popular metrics: ROUGE, METEOR, and BERTScore. Attack\nsystems also \"outperform\" state-of-the-art summarization methods on ROUGE-1 and\nROUGE-L, and score the second-highest on METEOR. Furthermore, a BERTScore\nbackdoor is observed: a simple trigger can score higher than any automatic\nsummarization method. The evasion attacks in this work indicate the low\nrobustness of current scoring systems at the system level. We hope that our\nhighlighting of these proposed attacks will facilitate the development of\nsummary scores.\n", "date": "2022-10-25", "categories": []}
{"id": "2210.14283", "title": "Accelerating Certified Robustness Training via Knowledge Transfer", "abstract": "  Training deep neural network classifiers that are certifiably robust against\nadversarial attacks is critical to ensuring the security and reliability of\nAI-controlled systems. Although numerous state-of-the-art certified training\nmethods have been developed, they are computationally expensive and scale\npoorly with respect to both dataset and network complexity. Widespread usage of\ncertified training is further hindered by the fact that periodic retraining is\nnecessary to incorporate new data and network improvements. In this paper, we\npropose Certified Robustness Transfer (CRT), a general-purpose framework for\nreducing the computational overhead of any certifiably robust training method\nthrough knowledge transfer. Given a robust teacher, our framework uses a novel\ntraining loss to transfer the teacher's robustness to the student. We provide\ntheoretical and empirical validation of CRT. Our experiments on CIFAR-10 show\nthat CRT speeds up certified robustness training by $8 \\times$ on average\nacross three different architecture generations while achieving comparable\nrobustness to state-of-the-art methods. We also show that CRT can scale to\nlarge-scale datasets like ImageNet.\n", "date": "2022-10-25", "categories": []}
{"id": "2210.14405", "title": "Adversarially Robust Medical Classification via Attentive Convolutional\n  Neural Networks", "abstract": "  Convolutional neural network-based medical image classifiers have been shown\nto be especially susceptible to adversarial examples. Such instabilities are\nlikely to be unacceptable in the future of automated diagnoses. Though\nstatistical adversarial example detection methods have proven to be effective\ndefense mechanisms, additional research is necessary that investigates the\nfundamental vulnerabilities of deep-learning-based systems and how best to\nbuild models that jointly maximize traditional and robust accuracy. This paper\npresents the inclusion of attention mechanisms in CNN-based medical image\nclassifiers as a reliable and effective strategy for increasing robust accuracy\nwithout sacrifice. This method is able to increase robust accuracy by up to 16%\nin typical adversarial scenarios and up to 2700% in extreme cases.\n", "date": "2022-10-26", "categories": []}
{"id": "2207.06858", "title": "RSD-GAN: Regularized Sobolev Defense GAN Against Speech-to-Text\n  Adversarial Attacks", "abstract": "  This paper introduces a new synthesis-based defense algorithm for\ncounteracting with a varieties of adversarial attacks developed for challenging\nthe performance of the cutting-edge speech-to-text transcription systems. Our\nalgorithm implements a Sobolev-based GAN and proposes a novel regularizer for\neffectively controlling over the functionality of the entire generative model,\nparticularly the discriminator network during training. Our achieved results\nupon carrying out numerous experiments on the victim DeepSpeech, Kaldi, and\nLingvo speech transcription systems corroborate the remarkable performance of\nour defense approach against a comprehensive range of targeted and non-targeted\nadversarial attacks.\n", "date": "2022-07-14", "categories": []}
{"id": "2210.13463", "title": "Adversarial Pretraining of Self-Supervised Deep Networks: Past, Present\n  and Future", "abstract": "  In this paper, we review adversarial pretraining of self-supervised deep\nnetworks including both convolutional neural networks and vision transformers.\nUnlike the adversarial training with access to labeled examples, adversarial\npretraining is complicated as it only has access to unlabeled examples. To\nincorporate adversaries into pretraining models on either input or feature\nlevel, we find that existing approaches are largely categorized into two\ngroups: memory-free instance-wise attacks imposing worst-case perturbations on\nindividual examples, and memory-based adversaries shared across examples over\niterations. In particular, we review several representative adversarial\npretraining models based on Contrastive Learning (CL) and Masked Image Modeling\n(MIM), respectively, two popular self-supervised pretraining methods in\nliterature. We also review miscellaneous issues about computing overheads,\ninput-/feature-level adversaries, as well as other adversarial pretraining\napproaches beyond the above two groups. Finally, we discuss emerging trends and\nfuture directions about the relations between adversarial and cooperative\npretraining, unifying adversarial CL and MIM pretraining, and the trade-off\nbetween accuracy and robustness in adversarial pretraining.\n", "date": "2022-10-23", "categories": []}
{"id": "2210.13982", "title": "Hindering Adversarial Attacks with Implicit Neural Representations", "abstract": "  We introduce the Lossy Implicit Network Activation Coding (LINAC) defence, an\ninput transformation which successfully hinders several common adversarial\nattacks on CIFAR-$10$ classifiers for perturbations up to $\\epsilon = 8/255$ in\n$L_\\infty$ norm and $\\epsilon = 0.5$ in $L_2$ norm. Implicit neural\nrepresentations are used to approximately encode pixel colour intensities in\n$2\\text{D}$ images such that classifiers trained on transformed data appear to\nhave robustness to small perturbations without adversarial training or large\ndrops in performance. The seed of the random number generator used to\ninitialise and train the implicit neural representation turns out to be\nnecessary information for stronger generic attacks, suggesting its role as a\nprivate key. We devise a Parametric Bypass Approximation (PBA) attack strategy\nfor key-based defences, which successfully invalidates an existing method in\nthis category. Interestingly, our LINAC defence also hinders some transfer and\nadaptive attacks, including our novel PBA strategy. Our results emphasise the\nimportance of a broad range of customised attacks despite apparent robustness\naccording to standard evaluations. LINAC source code and parameters of defended\nclassifier evaluated throughout this submission are available:\nhttps://github.com/deepmind/linac\n", "date": "2022-10-22", "categories": []}
{"id": "2210.12233", "title": "TCAB: A Large-Scale Text Classification Attack Benchmark", "abstract": "  We introduce the Text Classification Attack Benchmark (TCAB), a dataset for\nanalyzing, understanding, detecting, and labeling adversarial attacks against\ntext classifiers. TCAB includes 1.5 million attack instances, generated by\ntwelve adversarial attacks targeting three classifiers trained on six source\ndatasets for sentiment analysis and abuse detection in English. Unlike standard\ntext classification, text attacks must be understood in the context of the\ntarget classifier that is being attacked, and thus features of the target\nclassifier are important as well. TCAB includes all attack instances that are\nsuccessful in flipping the predicted label; a subset of the attacks are also\nlabeled by human annotators to determine how frequently the primary semantics\nare preserved. The process of generating attacks is automated, so that TCAB can\neasily be extended to incorporate new text attacks and better classifiers as\nthey are developed. In addition to the primary tasks of detecting and labeling\nattacks, TCAB can also be used for attack localization, attack target labeling,\nand attack characterization. TCAB code and dataset are available at\nhttps://react-nlp.github.io/tcab/.\n", "date": "2022-10-21", "categories": []}
{"id": "2210.12367", "title": "Precisely the Point: Adversarial Augmentations for Faithful and\n  Informative Text Generation", "abstract": "  Though model robustness has been extensively studied in language\nunderstanding, the robustness of Seq2Seq generation remains understudied. In\nthis paper, we conduct the first quantitative analysis on the robustness of\npre-trained Seq2Seq models. We find that even current SOTA pre-trained Seq2Seq\nmodel (BART) is still vulnerable, which leads to significant degeneration in\nfaithfulness and informativeness for text generation tasks. This motivated us\nto further propose a novel adversarial augmentation framework, namely AdvSeq,\nfor generally improving faithfulness and informativeness of Seq2Seq models via\nenhancing their robustness. AdvSeq automatically constructs two types of\nadversarial augmentations during training, including implicit adversarial\nsamples by perturbing word representations and explicit adversarial samples by\nword swapping, both of which effectively improve Seq2Seq robustness. Extensive\nexperiments on three popular text generation tasks demonstrate that AdvSeq\nsignificantly improves both the faithfulness and informativeness of Seq2Seq\ngeneration under both automatic and human evaluation settings.\n", "date": "2022-10-22", "categories": []}
{"id": "2210.12396", "title": "ADDMU: Detection of Far-Boundary Adversarial Examples with Data and\n  Model Uncertainty Estimation", "abstract": "  Adversarial Examples Detection (AED) is a crucial defense technique against\nadversarial attacks and has drawn increasing attention from the Natural\nLanguage Processing (NLP) community. Despite the surge of new AED methods, our\nstudies show that existing methods heavily rely on a shortcut to achieve good\nperformance. In other words, current search-based adversarial attacks in NLP\nstop once model predictions change, and thus most adversarial examples\ngenerated by those attacks are located near model decision boundaries. To\nsurpass this shortcut and fairly evaluate AED methods, we propose to test AED\nmethods with \\textbf{F}ar \\textbf{B}oundary (\\textbf{FB}) adversarial examples.\nExisting methods show worse than random guess performance under this scenario.\nTo overcome this limitation, we propose a new technique, \\textbf{ADDMU},\n\\textbf{a}dversary \\textbf{d}etection with \\textbf{d}ata and \\textbf{m}odel\n\\textbf{u}ncertainty, which combines two types of uncertainty estimation for\nboth regular and FB adversarial example detection. Our new method outperforms\nprevious methods by 3.6 and 6.0 \\emph{AUC} points under each scenario. Finally,\nour analysis shows that the two types of uncertainty provided by \\textbf{ADDMU}\ncan be leveraged to characterize adversarial examples and identify the ones\nthat contribute most to model's robustness in adversarial training.\n", "date": "2022-10-22", "categories": []}
{"id": "2210.12598", "title": "GANI: Global Attacks on Graph Neural Networks via Imperceptible Node\n  Injections", "abstract": "  Graph neural networks (GNNs) have found successful applications in various\ngraph-related tasks. However, recent studies have shown that many GNNs are\nvulnerable to adversarial attacks. In a vast majority of existing studies,\nadversarial attacks on GNNs are launched via direct modification of the\noriginal graph such as adding/removing links, which may not be applicable in\npractice. In this paper, we focus on a realistic attack operation via injecting\nfake nodes. The proposed Global Attack strategy via Node Injection (GANI) is\ndesigned under the comprehensive consideration of an unnoticeable perturbation\nsetting from both structure and feature domains. Specifically, to make the node\ninjections as imperceptible and effective as possible, we propose a sampling\noperation to determine the degree of the newly injected nodes, and then\ngenerate features and select neighbors for these injected nodes based on the\nstatistical information of features and evolutionary perturbations obtained\nfrom a genetic algorithm, respectively. In particular, the proposed feature\ngeneration mechanism is suitable for both binary and continuous node features.\nExtensive experimental results on benchmark datasets against both general and\ndefended GNNs show strong attack performance of GANI. Moreover, the\nimperceptibility analyses also demonstrate that GANI achieves a relatively\nunnoticeable injection on benchmark datasets.\n", "date": "2022-10-23", "categories": []}
{"id": "2210.12952", "title": "Ares: A System-Oriented Wargame Framework for Adversarial ML", "abstract": "  Since the discovery of adversarial attacks against machine learning models\nnearly a decade ago, research on adversarial machine learning has rapidly\nevolved into an eternal war between defenders, who seek to increase the\nrobustness of ML models against adversarial attacks, and adversaries, who seek\nto develop better attacks capable of weakening or defeating these defenses.\nThis domain, however, has found little buy-in from ML practitioners, who are\nneither overtly concerned about these attacks affecting their systems in the\nreal world nor are willing to trade off the accuracy of their models in pursuit\nof robustness against these attacks.\n  In this paper, we motivate the design and implementation of Ares, an\nevaluation framework for adversarial ML that allows researchers to explore\nattacks and defenses in a realistic wargame-like environment. Ares frames the\nconflict between the attacker and defender as two agents in a reinforcement\nlearning environment with opposing objectives. This allows the introduction of\nsystem-level evaluation metrics such as time to failure and evaluation of\ncomplex strategies such as moving target defenses. We provide the results of\nour initial exploration involving a white-box attacker against an adversarially\ntrained defender.\n", "date": "2022-10-24", "categories": []}
{"id": "2210.11513", "title": "Learning Sample Reweighting for Accuracy and Adversarial Robustness", "abstract": "  There has been great interest in enhancing the robustness of neural network\nclassifiers to defend against adversarial perturbations through adversarial\ntraining, while balancing the trade-off between robust accuracy and standard\naccuracy. We propose a novel adversarial training framework that learns to\nreweight the loss associated with individual training samples based on a notion\nof class-conditioned margin, with the goal of improving robust generalization.\nWe formulate weighted adversarial training as a bilevel optimization problem\nwith the upper-level problem corresponding to learning a robust classifier, and\nthe lower-level problem corresponding to learning a parametric function that\nmaps from a sample's \\textit{multi-class margin} to an importance weight.\nExtensive experiments demonstrate that our approach consistently improves both\nclean and robust accuracy compared to related methods and state-of-the-art\nbaselines.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.11598", "title": "Identifying Human Strategies for Generating Word-Level Adversarial\n  Examples", "abstract": "  Adversarial examples in NLP are receiving increasing research attention. One\nline of investigation is the generation of word-level adversarial examples\nagainst fine-tuned Transformer models that preserve naturalness and\ngrammaticality. Previous work found that human- and machine-generated\nadversarial examples are comparable in their naturalness and grammatical\ncorrectness. Most notably, humans were able to generate adversarial examples\nmuch more effortlessly than automated attacks. In this paper, we provide a\ndetailed analysis of exactly how humans create these adversarial examples. By\nexploring the behavioural patterns of human workers during the generation\nprocess, we identify statistically significant tendencies based on which words\nhumans prefer to select for adversarial replacement (e.g., word frequencies,\nword saliencies, sentiment) as well as where and when words are replaced in an\ninput sequence. With our findings, we seek to inspire efforts that harness\nhuman strategies for more robust NLP models.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.12030", "title": "Evolution of Neural Tangent Kernels under Benign and Adversarial\n  Training", "abstract": "  Two key challenges facing modern deep learning are mitigating deep networks'\nvulnerability to adversarial attacks and understanding deep learning's\ngeneralization capabilities. Towards the first issue, many defense strategies\nhave been developed, with the most common being Adversarial Training (AT).\nTowards the second challenge, one of the dominant theories that has emerged is\nthe Neural Tangent Kernel (NTK) -- a characterization of neural network\nbehavior in the infinite-width limit. In this limit, the kernel is frozen, and\nthe underlying feature map is fixed. In finite widths, however, there is\nevidence that feature learning happens at the earlier stages of the training\n(kernel learning) before a second phase where the kernel remains fixed (lazy\ntraining). While prior work has aimed at studying adversarial vulnerability\nthrough the lens of the frozen infinite-width NTK, there is no work that\nstudies the adversarial robustness of the empirical/finite NTK during training.\nIn this work, we perform an empirical study of the evolution of the empirical\nNTK under standard and adversarial training, aiming to disambiguate the effect\nof adversarial training on kernel learning and lazy training. We find under\nadversarial training, the empirical NTK rapidly converges to a different kernel\n(and feature map) than standard training. This new kernel provides adversarial\nrobustness, even when non-robust training is performed on top of it.\nFurthermore, we find that adversarial training on top of a fixed kernel can\nyield a classifier with $76.1\\%$ robust accuracy under PGD attacks with\n$\\varepsilon = 4/255$ on CIFAR-10.\n", "date": "2022-10-21", "categories": []}
{"id": "2206.00278", "title": "On the Perils of Cascading Robust Classifiers", "abstract": "  Ensembling certifiably robust neural networks is a promising approach for\nimproving the \\emph{certified robust accuracy} of neural models. Black-box\nensembles that assume only query-access to the constituent models (and their\nrobustness certifiers) during prediction are particularly attractive due to\ntheir modular structure. Cascading ensembles are a popular instance of\nblack-box ensembles that appear to improve certified robust accuracies in\npractice. However, we show that the robustness certifier used by a cascading\nensemble is unsound. That is, when a cascading ensemble is certified as locally\nrobust at an input $x$ (with respect to $\\epsilon$), there can be inputs $x'$\nin the $\\epsilon$-ball centered at $x$, such that the cascade's prediction at\n$x'$ is different from $x$ and thus the ensemble is not locally robust. Our\ntheoretical findings are accompanied by empirical results that further\ndemonstrate this unsoundness. We present \\emph{cascade attack} (CasA), an\nadversarial attack against cascading ensembles, and show that: (1) there exists\nan adversarial input for up to 88\\% of the samples where the ensemble claims to\nbe certifiably robust and accurate; and (2) the accuracy of a cascading\nensemble under our attack is as low as 11\\% when it claims to be certifiably\nrobust and accurate on 97\\% of the test set. Our work reveals a critical\npitfall of cascading certifiably robust models by showing that the seemingly\nbeneficial strategy of cascading can actually hurt the robustness of the\nresulting ensemble. Our code is available at\n\\url{https://github.com/TristaChi/ensembleKW}.\n", "date": "2022-06-01", "categories": []}
{"id": "2206.09391", "title": "Towards Adversarial Attack on Vision-Language Pre-training Models", "abstract": "  While vision-language pre-training model (VLP) has shown revolutionary\nimprovements on various vision-language (V+L) tasks, the studies regarding its\nadversarial robustness remain largely unexplored. This paper studied the\nadversarial attack on popular VLP models and V+L tasks. First, we analyzed the\nperformance of adversarial attacks under different settings. By examining the\ninfluence of different perturbed objects and attack targets, we concluded some\nkey observations as guidance on both designing strong multimodal adversarial\nattack and constructing robust VLP models. Second, we proposed a novel\nmultimodal attack method on the VLP models called Collaborative Multimodal\nAdversarial Attack (Co-Attack), which collectively carries out the attacks on\nthe image modality and the text modality. Experimental results demonstrated\nthat the proposed method achieves improved attack performances on different V+L\ndownstream tasks and VLP models. The analysis observations and novel attack\nmethod hopefully provide new understanding into the adversarial robustness of\nVLP models, so as to contribute their safe and reliable deployment in more\nreal-world scenarios. Code is available at\nhttps://github.com/adversarial-for-goodness/Co-Attack.\n", "date": "2022-06-19", "categories": []}
{"id": "2210.11082", "title": "Apple of Sodom: Hidden Backdoors in Superior Sentence Embeddings via\n  Contrastive Learning", "abstract": "  This paper finds that contrastive learning can produce superior sentence\nembeddings for pre-trained models but is also vulnerable to backdoor attacks.\nWe present the first backdoor attack framework, BadCSE, for state-of-the-art\nsentence embeddings under supervised and unsupervised learning settings. The\nattack manipulates the construction of positive and negative pairs so that the\nbackdoored samples have a similar embedding with the target sample (targeted\nattack) or the negative embedding of its clean version (non-targeted attack).\nBy injecting the backdoor in sentence embeddings, BadCSE is resistant against\ndownstream fine-tuning. We evaluate BadCSE on both STS tasks and other\ndownstream tasks. The supervised non-targeted attack obtains a performance\ndegradation of 194.86%, and the targeted attack maps the backdoored samples to\nthe target embedding with a 97.70% success rate while maintaining the model\nutility.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.11237", "title": "Emerging Threats in Deep Learning-Based Autonomous Driving: A\n  Comprehensive Survey", "abstract": "  Since the 2004 DARPA Grand Challenge, the autonomous driving technology has\nwitnessed nearly two decades of rapid development. Particularly, in recent\nyears, with the application of new sensors and deep learning technologies\nextending to the autonomous field, the development of autonomous driving\ntechnology has continued to make breakthroughs. Thus, many carmakers and\nhigh-tech giants dedicated to research and system development of autonomous\ndriving. However, as the foundation of autonomous driving, the deep learning\ntechnology faces many new security risks. The academic community has proposed\ndeep learning countermeasures against the adversarial examples and AI backdoor,\nand has introduced them into the autonomous driving field for verification.\nDeep learning security matters to autonomous driving system security, and then\nmatters to personal safety, which is an issue that deserves attention and\nresearch.This paper provides an summary of the concepts, developments and\nrecent research in deep learning security technologies in autonomous driving.\nFirstly, we briefly introduce the deep learning framework and pipeline in the\nautonomous driving system, which mainly include the deep learning technologies\nand algorithms commonly used in this field. Moreover, we focus on the potential\nsecurity threats of the deep learning based autonomous driving system in each\nfunctional layer in turn. We reviews the development of deep learning attack\ntechnologies to autonomous driving, investigates the State-of-the-Art\nalgorithms, and reveals the potential risks. At last, we provides an outlook on\ndeep learning security in the autonomous driving field and proposes\nrecommendations for building a safe and trustworthy autonomous driving system.\n", "date": "2022-10-19", "categories": []}
{"id": "2210.11242", "title": "Attacking Motion Estimation with Adversarial Snow", "abstract": "  Current adversarial attacks for motion estimation (optical flow) optimize\nsmall per-pixel perturbations, which are unlikely to appear in the real world.\nIn contrast, we exploit a real-world weather phenomenon for a novel attack with\nadversarially optimized snow. At the core of our attack is a differentiable\nrenderer that consistently integrates photorealistic snowflakes with realistic\nmotion into the 3D scene. Through optimization we obtain adversarial snow that\nsignificantly impacts the optical flow while being indistinguishable from\nordinary snow. Surprisingly, the impact of our novel attack is largest on\nmethods that previously showed a high robustness to small L_p perturbations.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.11264", "title": "Detecting Backdoors in Deep Text Classifiers", "abstract": "  Deep neural networks are vulnerable to adversarial attacks, such as backdoor\nattacks in which a malicious adversary compromises a model during training such\nthat specific behaviour can be triggered at test time by attaching a specific\nword or phrase to an input. This paper considers the problem of diagnosing\nwhether a model has been compromised and if so, identifying the backdoor\ntrigger. We present the first robust defence mechanism that generalizes to\nseveral backdoor attacks against text classification models, without prior\nknowledge of the attack type, nor does our method require access to any\n(potentially compromised) training resources. Our experiments show that our\ntechnique is highly accurate at defending against state-of-the-art backdoor\nattacks, including data poisoning and weight poisoning, across a range of text\nclassification tasks and model architectures. Our code will be made publicly\navailable upon acceptance.\n", "date": "2022-10-11", "categories": []}
{"id": "2208.03309", "title": "Lethal Dose Conjecture on Data Poisoning", "abstract": "  Data poisoning considers an adversary that distorts the training set of\nmachine learning algorithms for malicious purposes. In this work, we bring to\nlight one conjecture regarding the fundamentals of data poisoning, which we\ncall the Lethal Dose Conjecture. The conjecture states: If $n$ clean training\nsamples are needed for accurate predictions, then in a size-$N$ training set,\nonly $\\Theta(N/n)$ poisoned samples can be tolerated while ensuring accuracy.\nTheoretically, we verify this conjecture in multiple cases. We also offer a\nmore general perspective of this conjecture through distribution\ndiscrimination. Deep Partition Aggregation (DPA) and its extension, Finite\nAggregation (FA) are recent approaches for provable defenses against data\npoisoning, where they predict through the majority vote of many base models\ntrained from different subsets of training set using a given learner. The\nconjecture implies that both DPA and FA are (asymptotically) optimal -- if we\nhave the most data-efficient learner, they can turn it into one of the most\nrobust defenses against data poisoning. This outlines a practical approach to\ndeveloping stronger defenses against poisoning via finding data-efficient\nlearners. Empirically, as a proof of concept, we show that by simply using\ndifferent data augmentations for base learners, we can respectively double and\ntriple the certified robustness of DPA on CIFAR-10 and GTSRB without\nsacrificing accuracy.\n", "date": "2022-08-05", "categories": []}
{"id": "2210.09153", "title": "Face Pasting Attack", "abstract": "  Cujo AI and Adversa AI hosted the MLSec face recognition challenge. The goal\nwas to attack a black box face recognition model with targeted attacks. The\nmodel returned the confidence of the target class and a stealthiness score. For\nan attack to be considered successful the target class has to have the highest\nconfidence among all classes and the stealthiness has to be at least 0.5. In\nour approach we paste the face of a target into a source image. By utilizing\nposition, scaling, rotation and transparency attributes we reached 3rd place.\nOur approach took approximately 200 queries per attack for the final highest\nscore and about ~7.7 queries minimum for a successful attack. The code is\navailable at https://github.com/bunni90/FacePastingAttack .\n", "date": "2022-10-17", "categories": []}
{"id": "2210.10253", "title": "On the Adversarial Robustness of Mixture of Experts", "abstract": "  Adversarial robustness is a key desirable property of neural networks. It has\nbeen empirically shown to be affected by their sizes, with larger networks\nbeing typically more robust. Recently, Bubeck and Sellke proved a lower bound\non the Lipschitz constant of functions that fit the training data in terms of\ntheir number of parameters. This raises an interesting open question, do -- and\ncan -- functions with more parameters, but not necessarily more computational\ncost, have better robustness? We study this question for sparse Mixture of\nExpert models (MoEs), that make it possible to scale up the model size for a\nroughly constant computational cost. We theoretically show that under certain\nconditions on the routing and the structure of the data, MoEs can have\nsignificantly smaller Lipschitz constants than their dense counterparts. The\nrobustness of MoEs can suffer when the highest weighted experts for an input\nimplement sufficiently different functions. We next empirically evaluate the\nrobustness of MoEs on ImageNet using adversarial attacks and show they are\nindeed more robust than dense models with the same computational cost. We make\nkey observations showing the robustness of MoEs to the choice of experts,\nhighlighting the redundancy of experts in models trained in practice.\n", "date": "2022-10-19", "categories": []}
{"id": "2210.10683", "title": "Why Should Adversarial Perturbations be Imperceptible? Rethink the\n  Research Paradigm in Adversarial NLP", "abstract": "  Textual adversarial samples play important roles in multiple subfields of NLP\nresearch, including security, evaluation, explainability, and data\naugmentation. However, most work mixes all these roles, obscuring the problem\ndefinitions and research goals of the security role that aims to reveal the\npractical concerns of NLP models. In this paper, we rethink the research\nparadigm of textual adversarial samples in security scenarios. We discuss the\ndeficiencies in previous work and propose our suggestions that the research on\nthe Security-oriented adversarial NLP (SoadNLP) should: (1) evaluate their\nmethods on security tasks to demonstrate the real-world concerns; (2) consider\nreal-world attackers' goals, instead of developing impractical methods. To this\nend, we first collect, process, and release a security datasets collection\nAdvbench. Then, we reformalize the task and adjust the emphasis on different\ngoals in SoadNLP. Next, we propose a simple method based on heuristic rules\nthat can easily fulfill the actual adversarial goals to simulate real-world\nattack methods. We conduct experiments on both the attack and the defense sides\non Advbench. Experimental results show that our method has higher practical\nvalue, indicating that the research paradigm in SoadNLP may start from our new\nbenchmark. All the code and data of Advbench can be obtained at\n\\url{https://github.com/thunlp/Advbench}.\n", "date": "2022-10-19", "categories": []}
{"id": "2207.00411", "title": "Adversarial Robustness is at Odds with Lazy Training", "abstract": "  Recent works show that adversarial examples exist for random neural networks\n[Daniely and Schacham, 2020] and that these examples can be found using a\nsingle step of gradient ascent [Bubeck et al., 2021]. In this work, we extend\nthis line of work to \"lazy training\" of neural networks -- a dominant model in\ndeep learning theory in which neural networks are provably efficiently\nlearnable. We show that over-parametrized neural networks that are guaranteed\nto generalize well and enjoy strong computational guarantees remain vulnerable\nto attacks generated using a single step of gradient ascent.\n", "date": "2022-06-18", "categories": []}
{"id": "2207.01531", "title": "Wild Networks: Exposure of 5G Network Infrastructures to Adversarial\n  Examples", "abstract": "  Fifth Generation (5G) networks must support billions of heterogeneous devices\nwhile guaranteeing optimal Quality of Service (QoS). Such requirements are\nimpossible to meet with human effort alone, and Machine Learning (ML)\nrepresents a core asset in 5G. ML, however, is known to be vulnerable to\nadversarial examples; moreover, as our paper will show, the 5G context is\nexposed to a yet another type of adversarial ML attacks that cannot be\nformalized with existing threat models. Proactive assessment of such risks is\nalso challenging due to the lack of ML-powered 5G equipment available for\nadversarial ML research.\n  To tackle these problems, we propose a novel adversarial ML threat model that\nis particularly suited to 5G scenarios, and is agnostic to the precise function\nsolved by ML. In contrast to existing ML threat models, our attacks do not\nrequire any compromise of the target 5G system while still being viable due to\nthe QoS guarantees and the open nature of 5G networks. Furthermore, we propose\nan original framework for realistic ML security assessments based on public\ndata. We proactively evaluate our threat model on 6 applications of ML\nenvisioned in 5G. Our attacks affect both the training and the inference\nstages, can degrade the performance of state-of-the-art ML systems, and have a\nlower entry barrier than previous attacks.\n", "date": "2022-07-04", "categories": []}
{"id": "2210.09405", "title": "Towards Generating Adversarial Examples on Mixed-type Data", "abstract": "  The existence of adversarial attacks (or adversarial examples) brings huge\nconcern about the machine learning (ML) model's safety issues. For many\nsafety-critical ML tasks, such as financial forecasting, fraudulent detection,\nand anomaly detection, the data samples are usually mixed-type, which contain\nplenty of numerical and categorical features at the same time. However, how to\ngenerate adversarial examples with mixed-type data is still seldom studied. In\nthis paper, we propose a novel attack algorithm M-Attack, which can effectively\ngenerate adversarial examples in mixed-type data. Based on M-Attack, attackers\ncan attempt to mislead the targeted classification model's prediction, by only\nslightly perturbing both the numerical and categorical features in the given\ndata samples. More importantly, by adding designed regularizations, our\ngenerated adversarial examples can evade potential detection models, which\nmakes the attack indeed insidious. Through extensive empirical studies, we\nvalidate the effectiveness and efficiency of our attack method and evaluate the\nrobustness of existing classification models against our proposed attack. The\nexperimental results highlight the feasibility of generating adversarial\nexamples toward machine learning models in real-world applications.\n", "date": "2022-10-17", "categories": []}
{"id": "2210.09658", "title": "ROSE: Robust Selective Fine-tuning for Pre-trained Language Models", "abstract": "  Even though the large-scale language models have achieved excellent\nperformances, they suffer from various adversarial attacks. A large body of\ndefense methods has been proposed. However, they are still limited due to\nredundant attack search spaces and the inability to defend against various\ntypes of attacks. In this work, we present a novel fine-tuning approach called\n\\textbf{RO}bust \\textbf{SE}letive fine-tuning (\\textbf{ROSE}) to address this\nissue. ROSE conducts selective updates when adapting pre-trained models to\ndownstream tasks, filtering out invaluable and unrobust updates of parameters.\nSpecifically, we propose two strategies: the first-order and second-order ROSE\nfor selecting target robust parameters. The experimental results show that ROSE\nachieves significant improvements in adversarial robustness on various\ndownstream NLP tasks, and the ensemble method even surpasses both variants\nabove. Furthermore, ROSE can be easily incorporated into existing fine-tuning\nmethods to improve their adversarial robustness further. The empirical analysis\nconfirms that ROSE eliminates unrobust spurious updates during fine-tuning,\nleading to solutions corresponding to flatter and wider optima than the\nconventional method. Code is available at\n\\url{https://github.com/jiangllan/ROSE}.\n", "date": "2022-10-18", "categories": []}
{"id": "2210.09671", "title": "Not All Poisons are Created Equal: Robust Training against Data\n  Poisoning", "abstract": "  Data poisoning causes misclassification of test time target examples by\ninjecting maliciously crafted samples in the training data. Existing defenses\nare often effective only against a specific type of targeted attack,\nsignificantly degrade the generalization performance, or are prohibitive for\nstandard deep learning pipelines.\n  In this work, we propose an efficient defense mechanism that significantly\nreduces the success rate of various data poisoning attacks, and provides\ntheoretical guarantees for the performance of the model. Targeted attacks work\nby adding bounded perturbations to a randomly selected subset of training data\nto match the targets' gradient or representation. We show that: (i) under\nbounded perturbations, only a number of poisons can be optimized to have a\ngradient that is close enough to that of the target and make the attack\nsuccessful; (ii) such effective poisons move away from their original class and\nget isolated in the gradient space; (iii) dropping examples in low-density\ngradient regions during training can successfully eliminate the effective\npoisons, and guarantees similar training dynamics to that of training on full\ndata. Our extensive experiments show that our method significantly decreases\nthe success rate of state-of-the-art targeted attacks, including Gradient\nMatching and Bullseye Polytope, and easily scales to large datasets.\n", "date": "2022-10-18", "categories": []}
{"id": "2210.09852", "title": "Scaling Adversarial Training to Large Perturbation Bounds", "abstract": "  The vulnerability of Deep Neural Networks to Adversarial Attacks has fuelled\nresearch towards building robust models. While most Adversarial Training\nalgorithms aim at defending attacks constrained within low magnitude Lp norm\nbounds, real-world adversaries are not limited by such constraints. In this\nwork, we aim to achieve adversarial robustness within larger bounds, against\nperturbations that may be perceptible, but do not change human (or Oracle)\nprediction. The presence of images that flip Oracle predictions and those that\ndo not makes this a challenging setting for adversarial robustness. We discuss\nthe ideal goals of an adversarial defense algorithm beyond perceptual limits,\nand further highlight the shortcomings of naively extending existing training\nalgorithms to higher perturbation bounds. In order to overcome these\nshortcomings, we propose a novel defense, Oracle-Aligned Adversarial Training\n(OA-AT), to align the predictions of the network with that of an Oracle during\nadversarial training. The proposed approach achieves state-of-the-art\nperformance at large epsilon bounds (such as an L-inf bound of 16/255 on\nCIFAR-10) while outperforming existing defenses (AWP, TRADES, PGD-AT) at\nstandard bounds (8/255) as well.\n", "date": "2022-10-18", "categories": []}
{"id": "2209.09502", "title": "GAMA: Generative Adversarial Multi-Object Scene Attacks", "abstract": "  The majority of methods for crafting adversarial attacks have focused on\nscenes with a single dominant object (e.g., images from ImageNet). On the other\nhand, natural scenes include multiple dominant objects that are semantically\nrelated. Thus, it is crucial to explore designing attack strategies that look\nbeyond learning on single-object scenes or attack single-object victim\nclassifiers. Due to their inherent property of strong transferability of\nperturbations to unknown models, this paper presents the first approach of\nusing generative models for adversarial attacks on multi-object scenes. In\norder to represent the relationships between different objects in the input\nscene, we leverage upon the open-sourced pre-trained vision-language model CLIP\n(Contrastive Language-Image Pre-training), with the motivation to exploit the\nencoded semantics in the language space along with the visual space. We call\nthis attack approach Generative Adversarial Multi-object scene Attacks (GAMA).\nGAMA demonstrates the utility of the CLIP model as an attacker's tool to train\nformidable perturbation generators for multi-object scenes. Using the joint\nimage-text features to train the generator, we show that GAMA can craft potent\ntransferable perturbations in order to fool victim classifiers in various\nattack settings. For example, GAMA triggers ~16% more misclassification than\nstate-of-the-art generative approaches in black-box settings where both the\nclassifier architecture and data distribution of the attacker are different\nfrom the victim. Our code is available here:\nhttps://abhishekaich27.github.io/gama.html\n", "date": "2022-09-20", "categories": []}
{"id": "2210.03543", "title": "A2: Efficient Automated Attacker for Boosting Adversarial Training", "abstract": "  Based on the significant improvement of model robustness by AT (Adversarial\nTraining), various variants have been proposed to further boost the\nperformance. Well-recognized methods have focused on different components of AT\n(e.g., designing loss functions and leveraging additional unlabeled data). It\nis generally accepted that stronger perturbations yield more robust models.\nHowever, how to generate stronger perturbations efficiently is still missed. In\nthis paper, we propose an efficient automated attacker called A2 to boost AT by\ngenerating the optimal perturbations on-the-fly during training. A2 is a\nparameterized automated attacker to search in the attacker space for the best\nattacker against the defense model and examples. Extensive experiments across\ndifferent datasets demonstrate that A2 generates stronger perturbations with\nlow extra cost and reliably improves the robustness of various AT methods\nagainst different attacks.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.08472", "title": "Object-Attentional Untargeted Adversarial Attack", "abstract": "  Deep neural networks are facing severe threats from adversarial attacks. Most\nexisting black-box attacks fool target model by generating either global\nperturbations or local patches. However, both global perturbations and local\npatches easily cause annoying visual artifacts in adversarial example. Compared\nwith some smooth regions of an image, the object region generally has more\nedges and a more complex texture. Thus small perturbations on it will be more\nimperceptible. On the other hand, the object region is undoubtfully the\ndecisive part of an image to classification tasks. Motivated by these two\nfacts, we propose an object-attentional adversarial attack method for\nuntargeted attack. Specifically, we first generate an object region by\nintersecting the object detection region from YOLOv4 with the salient object\ndetection (SOD) region from HVPNet. Furthermore, we design an activation\nstrategy to avoid the reaction caused by the incomplete SOD. Then, we perform\nan adversarial attack only on the detected object region by leveraging Simple\nBlack-box Adversarial Attack (SimBA). To verify the proposed method, we create\na unique dataset by extracting all the images containing the object defined by\nCOCO from ImageNet-1K, named COCO-Reduced-ImageNet in this paper. Experimental\nresults on ImageNet-1K and COCO-Reduced-ImageNet show that under various system\nsettings, our method yields the adversarial example with better perceptual\nquality meanwhile saving the query budget up to 24.16\\% compared to the\nstate-of-the-art approaches including SimBA.\n", "date": "2022-10-16", "categories": []}
{"id": "2210.08579", "title": "Nowhere to Hide: A Lightweight Unsupervised Detector against Adversarial\n  Examples", "abstract": "  Although deep neural networks (DNNs) have shown impressive performance on\nmany perceptual tasks, they are vulnerable to adversarial examples that are\ngenerated by adding slight but maliciously crafted perturbations to benign\nimages. Adversarial detection is an important technique for identifying\nadversarial examples before they are entered into target DNNs. Previous studies\nto detect adversarial examples either targeted specific attacks or required\nexpensive computation. How design a lightweight unsupervised detector is still\na challenging problem. In this paper, we propose an AutoEncoder-based\nAdversarial Examples (AEAE) detector, that can guard DNN models by detecting\nadversarial examples with low computation in an unsupervised manner. The AEAE\nincludes only a shallow autoencoder but plays two roles. First, a well-trained\nautoencoder has learned the manifold of benign examples. This autoencoder can\nproduce a large reconstruction error for adversarial images with large\nperturbations, so we can detect significantly perturbed adversarial examples\nbased on the reconstruction error. Second, the autoencoder can filter out the\nsmall noise and change the DNN's prediction on adversarial examples with small\nperturbations. It helps to detect slightly perturbed adversarial examples based\non the prediction distance. To cover these two cases, we utilize the\nreconstruction error and prediction distance from benign images to construct a\ntwo-tuple feature set and train an adversarial detector using the isolation\nforest algorithm. We show empirically that the AEAE is unsupervised and\ninexpensive against the most state-of-the-art attacks. Through the detection in\nthese two cases, there is nowhere to hide adversarial examples.\n", "date": "2022-10-16", "categories": []}
{"id": "2210.08701", "title": "ODG-Q: Robust Quantization via Online Domain Generalization", "abstract": "  Quantizing neural networks to low-bitwidth is important for model deployment\non resource-limited edge hardware. Although a quantized network has a smaller\nmodel size and memory footprint, it is fragile to adversarial attacks. However,\nfew methods study the robustness and training efficiency of quantized networks.\nTo this end, we propose a new method by recasting robust quantization as an\nonline domain generalization problem, termed ODG-Q, which generates diverse\nadversarial data at a low cost during training. ODG-Q consistently outperforms\nexisting works against various adversarial attacks. For example, on CIFAR-10\ndataset, ODG-Q achieves 49.2% average improvements under five common white-box\nattacks and 21.7% average improvements under five common black-box attacks,\nwith a training cost similar to that of natural training (viz. without\nadversaries). To our best knowledge, this work is the first work that trains\nboth quantized and binary neural networks on ImageNet that consistently improve\nrobustness under different attacks. We also provide a theoretical insight of\nODG-Q that accounts for the bound of model risk on attacked data.\n", "date": "2022-10-17", "categories": []}
{"id": "2210.08929", "title": "DE-CROP: Data-efficient Certified Robustness for Pretrained Classifiers", "abstract": "  Certified defense using randomized smoothing is a popular technique to\nprovide robustness guarantees for deep neural networks against l2 adversarial\nattacks. Existing works use this technique to provably secure a pretrained\nnon-robust model by training a custom denoiser network on entire training data.\nHowever, access to the training set may be restricted to a handful of data\nsamples due to constraints such as high transmission cost and the proprietary\nnature of the data. Thus, we formulate a novel problem of \"how to certify the\nrobustness of pretrained models using only a few training samples\". We observe\nthat training the custom denoiser directly using the existing techniques on\nlimited samples yields poor certification. To overcome this, our proposed\napproach (DE-CROP) generates class-boundary and interpolated samples\ncorresponding to each training sample, ensuring high diversity in the feature\nspace of the pretrained classifier. We train the denoiser by maximizing the\nsimilarity between the denoised output of the generated sample and the original\ntraining sample in the classifier's logit space. We also perform distribution\nlevel matching using domain discriminator and maximum mean discrepancy that\nyields further benefit. In white box setup, we obtain significant improvements\nover the baseline on multiple benchmark datasets and also report similar\nperformance under the challenging black box setup.\n", "date": "2022-10-17", "categories": []}
{"id": "2210.09305", "title": "Thinking Two Moves Ahead: Anticipating Other Users Improves Backdoor\n  Attacks in Federated Learning", "abstract": "  Federated learning is particularly susceptible to model poisoning and\nbackdoor attacks because individual users have direct control over the training\ndata and model updates. At the same time, the attack power of an individual\nuser is limited because their updates are quickly drowned out by those of many\nother users. Existing attacks do not account for future behaviors of other\nusers, and thus require many sequential updates and their effects are quickly\nerased. We propose an attack that anticipates and accounts for the entire\nfederated learning pipeline, including behaviors of other clients, and ensures\nthat backdoors are effective quickly and persist even after multiple rounds of\ncommunity updates. We show that this new attack is effective in realistic\nscenarios where the attacker only contributes to a small fraction of randomly\nsampled rounds and demonstrate this attack on image classification, next-word\nprediction, and sentiment analysis.\n", "date": "2022-10-17", "categories": []}
{"id": "2207.08803", "title": "Adversarial Pixel Restoration as a Pretext Task for Transferable\n  Perturbations", "abstract": "  Transferable adversarial attacks optimize adversaries from a pretrained\nsurrogate model and known label space to fool the unknown black-box models.\nTherefore, these attacks are restricted by the availability of an effective\nsurrogate model. In this work, we relax this assumption and propose Adversarial\nPixel Restoration as a self-supervised alternative to train an effective\nsurrogate model from scratch under the condition of no labels and few data\nsamples. Our training approach is based on a min-max scheme which reduces\noverfitting via an adversarial objective and thus optimizes for a more\ngeneralizable surrogate model. Our proposed attack is complimentary to the\nadversarial pixel restoration and is independent of any task specific objective\nas it can be launched in a self-supervised manner. We successfully demonstrate\nthe adversarial transferability of our approach to Vision Transformers as well\nas Convolutional Neural Networks for the tasks of classification, object\ndetection, and video segmentation. Our training approach improves the\ntransferability of the baseline unsupervised training method by 16.4% on\nImageNet val. set. Our codes & pre-trained surrogate models are available at:\nhttps://github.com/HashmatShadab/APR\n", "date": "2022-07-18", "categories": []}
{"id": "2210.03372", "title": "Pre-trained Adversarial Perturbations", "abstract": "  Self-supervised pre-training has drawn increasing attention in recent years\ndue to its superior performance on numerous downstream tasks after fine-tuning.\nHowever, it is well-known that deep learning models lack the robustness to\nadversarial examples, which can also invoke security issues to pre-trained\nmodels, despite being less explored. In this paper, we delve into the\nrobustness of pre-trained models by introducing Pre-trained Adversarial\nPerturbations (PAPs), which are universal perturbations crafted for the\npre-trained models to maintain the effectiveness when attacking fine-tuned ones\nwithout any knowledge of the downstream tasks. To this end, we propose a\nLow-Level Layer Lifting Attack (L4A) method to generate effective PAPs by\nlifting the neuron activations of low-level layers of the pre-trained models.\nEquipped with an enhanced noise augmentation strategy, L4A is effective at\ngenerating more transferable PAPs against fine-tuned models. Extensive\nexperiments on typical pre-trained vision models and ten downstream tasks\ndemonstrate that our method improves the attack success rate by a large margin\ncompared with state-of-the-art methods.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.07540", "title": "When Adversarial Training Meets Vision Transformers: Recipes from\n  Training to Architecture", "abstract": "  Vision Transformers (ViTs) have recently achieved competitive performance in\nbroad vision tasks. Unfortunately, on popular threat models, naturally trained\nViTs are shown to provide no more adversarial robustness than convolutional\nneural networks (CNNs). Adversarial training is still required for ViTs to\ndefend against such adversarial attacks. In this paper, we provide the first\nand comprehensive study on the adversarial training recipe of ViTs via\nextensive evaluation of various training techniques across benchmark datasets.\nWe find that pre-training and SGD optimizer are necessary for ViTs' adversarial\ntraining. Further considering ViT as a new type of model architecture, we\ninvestigate its adversarial robustness from the perspective of its unique\narchitectural components. We find, when randomly masking gradients from some\nattention blocks or masking perturbations on some patches during adversarial\ntraining, the adversarial robustness of ViTs can be remarkably improved, which\nmay potentially open up a line of work to explore the architectural information\ninside the newly designed models like ViTs. Our code is available at\nhttps://github.com/mo666666/When-Adversarial-Training-Meets-Vision-Transformers.\n", "date": "2022-10-14", "categories": []}
{"id": "2210.07907", "title": "Expose Backdoors on the Way: A Feature-Based Efficient Defense against\n  Textual Backdoor Attacks", "abstract": "  Natural language processing (NLP) models are known to be vulnerable to\nbackdoor attacks, which poses a newly arisen threat to NLP models. Prior online\nbackdoor defense methods for NLP models only focus on the anomalies at either\nthe input or output level, still suffering from fragility to adaptive attacks\nand high computational cost. In this work, we take the first step to\ninvestigate the unconcealment of textual poisoned samples at the\nintermediate-feature level and propose a feature-based efficient online defense\nmethod. Through extensive experiments on existing attacking methods, we find\nthat the poisoned samples are far away from clean samples in the intermediate\nfeature space of a poisoned NLP model. Motivated by this observation, we devise\na distance-based anomaly score (DAN) to distinguish poisoned samples from clean\nsamples at the feature level. Experiments on sentiment analysis and offense\ndetection tasks demonstrate the superiority of DAN, as it substantially\nsurpasses existing online defense methods in terms of defending performance and\nenjoys lower inference costs. Moreover, we show that DAN is also resistant to\nadaptive attacks based on feature-level regularization. Our code is available\nat https://github.com/lancopku/DAN.\n", "date": "2022-10-14", "categories": []}
{"id": "2210.06589", "title": "Adversarial Attack Against Image-Based Localization Neural Networks", "abstract": "  In this paper, we present a proof of concept for adversarially attacking the\nimage-based localization module of an autonomous vehicle. This attack aims to\ncause the vehicle to perform a wrong navigational decisions and prevent it from\nreaching a desired predefined destination in a simulated urban environment. A\ndatabase of rendered images allowed us to train a deep neural network that\nperforms a localization task and implement, develop and assess the adversarial\npattern. Our tests show that using this adversarial attack we can prevent the\nvehicle from turning at a given intersection. This is done by manipulating the\nvehicle's navigational module to falsely estimate its current position and thus\nfail to initialize the turning procedure until the vehicle misses the last\nopportunity to perform a safe turn in a given intersection.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.06704", "title": "COLLIDER: A Robust Training Framework for Backdoor Data", "abstract": "  Deep neural network (DNN) classifiers are vulnerable to backdoor attacks. An\nadversary poisons some of the training data in such attacks by installing a\ntrigger. The goal is to make the trained DNN output the attacker's desired\nclass whenever the trigger is activated while performing as usual for clean\ndata. Various approaches have recently been proposed to detect malicious\nbackdoored DNNs. However, a robust, end-to-end training approach, like\nadversarial training, is yet to be discovered for backdoor poisoned data. In\nthis paper, we take the first step toward such methods by developing a robust\ntraining framework, COLLIDER, that selects the most prominent samples by\nexploiting the underlying geometric structures of the data. Specifically, we\neffectively filter out candidate poisoned data at each training epoch by\nsolving a geometrical coreset selection objective. We first argue how clean\ndata samples exhibit (1) gradients similar to the clean majority of data and\n(2) low local intrinsic dimensionality (LID). Based on these criteria, we\ndefine a novel coreset selection objective to find such samples, which are used\nfor training a DNN. We show the effectiveness of the proposed method for robust\ntraining of DNNs on various poisoned datasets, reducing the backdoor success\nrate significantly.\n", "date": "2022-10-13", "categories": []}
{"id": "2210.06787", "title": "Observed Adversaries in Deep Reinforcement Learning", "abstract": "  In this work, we point out the problem of observed adversaries for deep\npolicies. Specifically, recent work has shown that deep reinforcement learning\nis susceptible to adversarial attacks where an observed adversary acts under\nenvironmental constraints to invoke natural but adversarial observations. This\nsetting is particularly relevant for HRI since HRI-related robots are expected\nto perform their tasks around and with other agents. In this work, we\ndemonstrate that this effect persists even with low-dimensional observations.\nWe further show that these adversarial attacks transfer across victims, which\npotentially allows malicious attackers to train an adversary without access to\nthe target victim.\n", "date": "2022-10-13", "categories": []}
{"id": "2210.06888", "title": "AccelAT: A Framework for Accelerating the Adversarial Training of Deep\n  Neural Networks through Accuracy Gradient", "abstract": "  Adversarial training is exploited to develop a robust Deep Neural Network\n(DNN) model against the malicious altered data. These attacks may have\ncatastrophic effects on DNN models but are indistinguishable for a human being.\nFor example, an external attack can modify an image adding noises invisible for\na human eye, but a DNN model misclassified the image. A key objective for\ndeveloping robust DNN models is to use a learning algorithm that is fast but\ncan also give model that is robust against different types of adversarial\nattacks. Especially for adversarial training, enormously long training times\nare needed for obtaining high accuracy under many different types of\nadversarial samples generated using different adversarial attack techniques.\n  This paper aims at accelerating the adversarial training to enable fast\ndevelopment of robust DNN models against adversarial attacks. The general\nmethod for improving the training performance is the hyperparameters\nfine-tuning, where the learning rate is one of the most crucial\nhyperparameters. By modifying its shape (the value over time) and value during\nthe training, we can obtain a model robust to adversarial attacks faster than\nstandard training.\n  First, we conduct experiments on two different datasets (CIFAR10, CIFAR100),\nexploring various techniques. Then, this analysis is leveraged to develop a\nnovel fast training methodology, AccelAT, which automatically adjusts the\nlearning rate for different epochs based on the accuracy gradient. The\nexperiments show comparable results with the related works, and in several\nexperiments, the adversarial training of DNNs using our AccelAT framework is\nconducted up to 2 times faster than the existing techniques. Thus, our findings\nboost the speed of adversarial training in an era in which security and\nperformance are fundamental optimization objectives in DNN-based applications.\n", "date": "2022-10-13", "categories": []}
{"id": "2206.09491", "title": "On the Limitations of Stochastic Pre-processing Defenses", "abstract": "  Defending against adversarial examples remains an open problem. A common\nbelief is that randomness at inference increases the cost of finding\nadversarial inputs. An example of such a defense is to apply a random\ntransformation to inputs prior to feeding them to the model. In this paper, we\nempirically and theoretically investigate such stochastic pre-processing\ndefenses and demonstrate that they are flawed. First, we show that most\nstochastic defenses are weaker than previously thought; they lack sufficient\nrandomness to withstand even standard attacks like projected gradient descent.\nThis casts doubt on a long-held assumption that stochastic defenses invalidate\nattacks designed to evade deterministic defenses and force attackers to\nintegrate the Expectation over Transformation (EOT) concept. Second, we show\nthat stochastic defenses confront a trade-off between adversarial robustness\nand model invariance; they become less effective as the defended model acquires\nmore invariance to their randomization. Future work will need to decouple these\ntwo effects. We also discuss implications and guidance for future research.\n", "date": "2022-06-19", "categories": []}
{"id": "2210.02191", "title": "On Attacking Out-Domain Uncertainty Estimation in Deep Neural Networks", "abstract": "  In many applications with real-world consequences, it is crucial to develop\nreliable uncertainty estimation for the predictions made by the AI decision\nsystems. Targeting at the goal of estimating uncertainty, various deep neural\nnetwork (DNN) based uncertainty estimation algorithms have been proposed.\nHowever, the robustness of the uncertainty returned by these algorithms has not\nbeen systematically explored. In this work, to raise the awareness of the\nresearch community on robust uncertainty estimation, we show that\nstate-of-the-art uncertainty estimation algorithms could fail catastrophically\nunder our proposed adversarial attack despite their impressive performance on\nuncertainty estimation. In particular, we aim at attacking the out-domain\nuncertainty estimation: under our attack, the uncertainty model would be fooled\nto make high-confident predictions for the out-domain data, which they\noriginally would have rejected. Extensive experimental results on various\nbenchmark image datasets show that the uncertainty estimated by\nstate-of-the-art methods could be easily corrupted by our attack.\n", "date": "2022-10-03", "categories": []}
{"id": "2210.02713", "title": "On Optimal Learning Under Targeted Data Poisoning", "abstract": "  Consider the task of learning a hypothesis class $\\mathcal{H}$ in the\npresence of an adversary that can replace up to an $\\eta$ fraction of the\nexamples in the training set with arbitrary adversarial examples. The adversary\naims to fail the learner on a particular target test point $x$ which is known\nto the adversary but not to the learner. In this work we aim to characterize\nthe smallest achievable error $\\epsilon=\\epsilon(\\eta)$ by the learner in the\npresence of such an adversary in both realizable and agnostic settings. We\nfully achieve this in the realizable setting, proving that\n$\\epsilon=\\Theta(\\mathtt{VC}(\\mathcal{H})\\cdot \\eta)$, where\n$\\mathtt{VC}(\\mathcal{H})$ is the VC dimension of $\\mathcal{H}$. Remarkably, we\nshow that the upper bound can be attained by a deterministic learner. In the\nagnostic setting we reveal a more elaborate landscape: we devise a\ndeterministic learner with a multiplicative regret guarantee of $\\epsilon \\leq\nC\\cdot\\mathtt{OPT} + O(\\mathtt{VC}(\\mathcal{H})\\cdot \\eta)$, where $C > 1$ is a\nuniversal numerical constant. We complement this by showing that for any\ndeterministic learner there is an attack which worsens its error to at least\n$2\\cdot \\mathtt{OPT}$. This implies that a multiplicative deterioration in the\nregret is unavoidable in this case. Finally, the algorithms we develop for\nachieving the optimal rates are inherently improper. Nevertheless, we show that\nfor a variety of natural concept classes, such as linear classifiers, it is\npossible to retain the dependence $\\epsilon=\\Theta_{\\mathcal{H}}(\\eta)$ by a\nproper algorithm in the realizable setting. Here $\\Theta_{\\mathcal{H}}$\nconceals a polynomial dependence on $\\mathtt{VC}(\\mathcal{H})$.\n", "date": "2022-10-06", "categories": []}
{"id": "2210.05929", "title": "Few-shot Backdoor Attacks via Neural Tangent Kernels", "abstract": "  In a backdoor attack, an attacker injects corrupted examples into the\ntraining set. The goal of the attacker is to cause the final trained model to\npredict the attacker's desired target label when a predefined trigger is added\nto test inputs. Central to these attacks is the trade-off between the success\nrate of the attack and the number of corrupted training examples injected. We\npose this attack as a novel bilevel optimization problem: construct strong\npoison examples that maximize the attack success rate of the trained model. We\nuse neural tangent kernels to approximate the training dynamics of the model\nbeing attacked and automatically learn strong poison examples. We experiment on\nsubclasses of CIFAR-10 and ImageNet with WideResNet-34 and ConvNeXt\narchitectures on periodic and patch trigger attacks and show that NTBA-designed\npoisoned examples achieve, for example, an attack success rate of 90% with ten\ntimes smaller number of poison examples injected compared to the baseline. We\nprovided an interpretation of the NTBA-designed attacks using the analysis of\nkernel linear regression. We further demonstrate a vulnerability in\noverparametrized deep neural networks, which is revealed by the shape of the\nneural tangent kernel.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.05968", "title": "Boosting the Transferability of Adversarial Attacks with Reverse\n  Adversarial Perturbation", "abstract": "  Deep neural networks (DNNs) have been shown to be vulnerable to adversarial\nexamples, which can produce erroneous predictions by injecting imperceptible\nperturbations. In this work, we study the transferability of adversarial\nexamples, which is significant due to its threat to real-world applications\nwhere model architecture or parameters are usually unknown. Many existing works\nreveal that the adversarial examples are likely to overfit the surrogate model\nthat they are generated from, limiting its transfer attack performance against\ndifferent target models. To mitigate the overfitting of the surrogate model, we\npropose a novel attack method, dubbed reverse adversarial perturbation (RAP).\nSpecifically, instead of minimizing the loss of a single adversarial point, we\nadvocate seeking adversarial example located at a region with unified low loss\nvalue, by injecting the worst-case perturbation (the reverse adversarial\nperturbation) for each step of the optimization procedure. The adversarial\nattack with RAP is formulated as a min-max bi-level optimization problem. By\nintegrating RAP into the iterative process for attacks, our method can find\nmore stable adversarial examples which are less sensitive to the changes of\ndecision boundary, mitigating the overfitting of the surrogate model.\nComprehensive experimental comparisons demonstrate that RAP can significantly\nboost adversarial transferability. Furthermore, RAP can be naturally combined\nwith many existing black-box attack techniques, to further boost the\ntransferability. When attacking a real-world image recognition system, Google\nCloud Vision API, we obtain 22% performance improvement of targeted attacks\nover the compared method. Our codes are available at\nhttps://github.com/SCLBD/Transfer_attack_RAP.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.06428", "title": "Trap and Replace: Defending Backdoor Attacks by Trapping Them into an\n  Easy-to-Replace Subnetwork", "abstract": "  Deep neural networks (DNNs) are vulnerable to backdoor attacks. Previous\nworks have shown it extremely challenging to unlearn the undesired backdoor\nbehavior from the network, since the entire network can be affected by the\nbackdoor samples. In this paper, we propose a brand-new backdoor defense\nstrategy, which makes it much easier to remove the harmful influence of\nbackdoor samples from the model. Our defense strategy, \\emph{Trap and Replace},\nconsists of two stages. In the first stage, we bait and trap the backdoors in a\nsmall and easy-to-replace subnetwork. Specifically, we add an auxiliary image\nreconstruction head on top of the stem network shared with a light-weighted\nclassification head. The intuition is that the auxiliary image reconstruction\ntask encourages the stem network to keep sufficient low-level visual features\nthat are hard to learn but semantically correct, instead of overfitting to the\neasy-to-learn but semantically incorrect backdoor correlations. As a result,\nwhen trained on backdoored datasets, the backdoors are easily baited towards\nthe unprotected classification head, since it is much more vulnerable than the\nshared stem, leaving the stem network hardly poisoned. In the second stage, we\nreplace the poisoned light-weighted classification head with an untainted one,\nby re-training it from scratch only on a small holdout dataset with clean\nsamples, while fixing the stem network. As a result, both the stem and the\nclassification head in the final network are hardly affected by backdoor\ntraining samples. We evaluate our method against ten different backdoor\nattacks. Our method outperforms previous state-of-the-art methods by up to\n$20.57\\%$, $9.80\\%$, and $13.72\\%$ attack success rate and on-average $3.14\\%$,\n$1.80\\%$, and $1.21\\%$ clean classification accuracy on CIFAR10, GTSRB, and\nImageNet-12, respectively. Code is available online.\n", "date": "2022-10-12", "categories": []}
{"id": "2206.03466", "title": "Adversarial Reprogramming Revisited", "abstract": "  Adversarial reprogramming, introduced by Elsayed, Goodfellow, and\nSohl-Dickstein, seeks to repurpose a neural network to perform a different\ntask, by manipulating its input without modifying its weights. We prove that\ntwo-layer ReLU neural networks with random weights can be adversarially\nreprogrammed to achieve arbitrarily high accuracy on Bernoulli data models over\nhypercube vertices, provided the network width is no greater than its input\ndimension. We also substantially strengthen a recent result of Phuong and\nLampert on directional convergence of gradient flow, and obtain as a corollary\nthat training two-layer ReLU neural networks on orthogonally separable datasets\ncan cause their adversarial reprogramming to fail. We support these theoretical\nresults by experiments that demonstrate that, as long as batch normalisation\nlayers are suitably initialised, even untrained networks with random weights\nare susceptible to adversarial reprogramming. This is in contrast to\nobservations in several recent works that suggested that adversarial\nreprogramming is not possible for untrained networks to any degree of\nreliability.\n", "date": "2022-06-07", "categories": []}
{"id": "2206.12227", "title": "Adversarial Robustness of Deep Neural Networks: A Survey from a Formal\n  Verification Perspective", "abstract": "  Neural networks have been widely applied in security applications such as\nspam and phishing detection, intrusion prevention, and malware detection. This\nblack-box method, however, often has uncertainty and poor explainability in\napplications. Furthermore, neural networks themselves are often vulnerable to\nadversarial attacks. For those reasons, there is a high demand for trustworthy\nand rigorous methods to verify the robustness of neural network models.\nAdversarial robustness, which concerns the reliability of a neural network when\ndealing with maliciously manipulated inputs, is one of the hottest topics in\nsecurity and machine learning. In this work, we survey existing literature in\nadversarial robustness verification for neural networks and collect 39\ndiversified research works across machine learning, security, and software\nengineering domains. We systematically analyze their approaches, including how\nrobustness is formulated, what verification techniques are used, and the\nstrengths and limitations of each technique. We provide a taxonomy from a\nformal verification perspective for a comprehensive understanding of this\ntopic. We classify the existing techniques based on property specification,\nproblem reduction, and reasoning strategies. We also demonstrate representative\ntechniques that have been applied in existing studies with a sample model.\nFinally, we discuss open questions for future research.\n", "date": "2022-06-24", "categories": []}
{"id": "2209.14826", "title": "Towards Lightweight Black-Box Attacks against Deep Neural Networks", "abstract": "  Black-box attacks can generate adversarial examples without accessing the\nparameters of target model, largely exacerbating the threats of deployed deep\nneural networks (DNNs). However, previous works state that black-box attacks\nfail to mislead target models when their training data and outputs are\ninaccessible. In this work, we argue that black-box attacks can pose practical\nattacks in this extremely restrictive scenario where only several test samples\nare available. Specifically, we find that attacking the shallow layers of DNNs\ntrained on a few test samples can generate powerful adversarial examples. As\nonly a few samples are required, we refer to these attacks as lightweight\nblack-box attacks. The main challenge to promoting lightweight attacks is to\nmitigate the adverse impact caused by the approximation error of shallow\nlayers. As it is hard to mitigate the approximation error with few available\nsamples, we propose Error TransFormer (ETF) for lightweight attacks. Namely,\nETF transforms the approximation error in the parameter space into a\nperturbation in the feature space and alleviates the error by disturbing\nfeatures. In experiments, lightweight black-box attacks with the proposed ETF\nachieve surprising results. For example, even if only 1 sample per category\navailable, the attack success rate in lightweight black-box attacks is only\nabout 3% lower than that of the black-box attacks with complete training data.\n", "date": "2022-09-29", "categories": []}
{"id": "2210.05118", "title": "Boosting Adversarial Robustness From The Perspective of Effective Margin\n  Regularization", "abstract": "  The adversarial vulnerability of deep neural networks (DNNs) has been\nactively investigated in the past several years. This paper investigates the\nscale-variant property of cross-entropy loss, which is the most commonly used\nloss function in classification tasks, and its impact on the effective margin\nand adversarial robustness of deep neural networks. Since the loss function is\nnot invariant to logit scaling, increasing the effective weight norm will make\nthe loss approach zero and its gradient vanish while the effective margin is\nnot adequately maximized. On typical DNNs, we demonstrate that, if not properly\nregularized, the standard training does not learn large effective margins and\nleads to adversarial vulnerability. To maximize the effective margins and learn\na robust DNN, we propose to regularize the effective weight norm during\ntraining. Our empirical study on feedforward DNNs demonstrates that the\nproposed effective margin regularization (EMR) learns large effective margins\nand boosts the adversarial robustness in both standard and adversarial\ntraining. On large-scale models, we show that EMR outperforms basic adversarial\ntraining, TRADES and two regularization baselines with substantial improvement.\nMoreover, when combined with several strong adversarial defense methods (MART\nand MAIL), our EMR further boosts the robustness.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.05373", "title": "Stable and Efficient Adversarial Training through Local Linearization", "abstract": "  There has been a recent surge in single-step adversarial training as it shows\nrobustness and efficiency. However, a phenomenon referred to as ``catastrophic\noverfitting\" has been observed, which is prevalent in single-step defenses and\nmay frustrate attempts to use FGSM adversarial training. To address this issue,\nwe propose a novel method, Stable and Efficient Adversarial Training (SEAT),\nwhich mitigates catastrophic overfitting by harnessing on local properties that\ndistinguish a robust model from that of a catastrophic overfitted model. The\nproposed SEAT has strong theoretical justifications, in that minimizing the\nSEAT loss can be shown to favour smooth empirical risk, thereby leading to\nrobustness. Experimental results demonstrate that the proposed method\nsuccessfully mitigates catastrophic overfitting, yielding superior performance\namongst efficient defenses. Our single-step method can reach 51% robust\naccuracy for CIFAR-10 with $l_\\infty$ perturbations of radius $8/255$ under a\nstrong PGD-50 attack, matching the performance of a 10-step iterative\nadversarial training at merely 3% computational cost.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.03895", "title": "ViewFool: Evaluating the Robustness of Visual Recognition to Adversarial\n  Viewpoints", "abstract": "  Recent studies have demonstrated that visual recognition models lack\nrobustness to distribution shift. However, current work mainly considers model\nrobustness to 2D image transformations, leaving viewpoint changes in the 3D\nworld less explored. In general, viewpoint changes are prevalent in various\nreal-world applications (e.g., autonomous driving), making it imperative to\nevaluate viewpoint robustness. In this paper, we propose a novel method called\nViewFool to find adversarial viewpoints that mislead visual recognition models.\nBy encoding real-world objects as neural radiance fields (NeRF), ViewFool\ncharacterizes a distribution of diverse adversarial viewpoints under an\nentropic regularizer, which helps to handle the fluctuations of the real camera\npose and mitigate the reality gap between the real objects and their neural\nrepresentations. Experiments validate that the common image classifiers are\nextremely vulnerable to the generated adversarial viewpoints, which also\nexhibit high cross-model transferability. Based on ViewFool, we introduce\nImageNet-V, a new out-of-distribution dataset for benchmarking viewpoint\nrobustness of image classifiers. Evaluation results on 40 classifiers with\ndiverse architectures, objective functions, and data augmentations reveal a\nsignificant drop in model performance when tested on ImageNet-V, which provides\na possibility to leverage ViewFool as an effective data augmentation strategy\nto improve viewpoint robustness.\n", "date": "2022-10-08", "categories": []}
{"id": "2210.04213", "title": "Towards Understanding and Boosting Adversarial Transferability from a\n  Distribution Perspective", "abstract": "  Transferable adversarial attacks against Deep neural networks (DNNs) have\nreceived broad attention in recent years. An adversarial example can be crafted\nby a surrogate model and then attack the unknown target model successfully,\nwhich brings a severe threat to DNNs. The exact underlying reasons for the\ntransferability are still not completely understood. Previous work mostly\nexplores the causes from the model perspective, e.g., decision boundary, model\narchitecture, and model capacity. adversarial attacks against Deep neural\nnetworks (DNNs) have received broad attention in recent years. An adversarial\nexample can be crafted by a surrogate model and then attack the unknown target\nmodel successfully, which brings a severe threat to DNNs. The exact underlying\nreasons for the transferability are still not completely understood. Previous\nwork mostly explores the causes from the model perspective. Here, we\ninvestigate the transferability from the data distribution perspective and\nhypothesize that pushing the image away from its original distribution can\nenhance the adversarial transferability. To be specific, moving the image out\nof its original distribution makes different models hardly classify the image\ncorrectly, which benefits the untargeted attack, and dragging the image into\nthe target distribution misleads the models to classify the image as the target\nclass, which benefits the targeted attack. Towards this end, we propose a novel\nmethod that crafts adversarial examples by manipulating the distribution of the\nimage. We conduct comprehensive transferable attacks against multiple DNNs to\ndemonstrate the effectiveness of the proposed method. Our method can\nsignificantly improve the transferability of the crafted attacks and achieves\nstate-of-the-art performance in both untargeted and targeted scenarios,\nsurpassing the previous best method by up to 40$\\%$ in some cases.\n", "date": "2022-10-09", "categories": []}
{"id": "2210.04311", "title": "Pruning Adversarially Robust Neural Networks without Adversarial\n  Examples", "abstract": "  Adversarial pruning compresses models while preserving robustness. Current\nmethods require access to adversarial examples during pruning. This\nsignificantly hampers training efficiency. Moreover, as new adversarial attacks\nand training methods develop at a rapid rate, adversarial pruning methods need\nto be modified accordingly to keep up. In this work, we propose a novel\nframework to prune a previously trained robust neural network while maintaining\nadversarial robustness, without further generating adversarial examples. We\nleverage concurrent self-distillation and pruning to preserve knowledge in the\noriginal model as well as regularizing the pruned model via the Hilbert-Schmidt\nInformation Bottleneck. We comprehensively evaluate our proposed framework and\nshow its superior performance in terms of both adversarial robustness and\nefficiency when pruning architectures trained on the MNIST, CIFAR-10, and\nCIFAR-100 datasets against five state-of-the-art attacks. Code is available at\nhttps://github.com/neu-spiral/PwoA/.\n", "date": "2022-10-09", "categories": []}
{"id": "2210.04591", "title": "Universal Adversarial Perturbations: Efficiency on a small image dataset", "abstract": "  Although neural networks perform very well on the image classification task,\nthey are still vulnerable to adversarial perturbations that can fool a neural\nnetwork without visibly changing an input image. A paper has shown the\nexistence of Universal Adversarial Perturbations which when added to any image\nwill fool the neural network with a very high probability. In this paper we\nwill try to reproduce the experience of the Universal Adversarial Perturbations\npaper, but on a smaller neural network architecture and training set, in order\nto be able to study the efficiency of the computed perturbation.\n", "date": "2022-10-10", "categories": []}
{"id": "2210.04886", "title": "Revisiting adapters with adversarial training", "abstract": "  While adversarial training is generally used as a defense mechanism, recent\nworks show that it can also act as a regularizer. By co-training a neural\nnetwork on clean and adversarial inputs, it is possible to improve\nclassification accuracy on the clean, non-adversarial inputs. We demonstrate\nthat, contrary to previous findings, it is not necessary to separate batch\nstatistics when co-training on clean and adversarial inputs, and that it is\nsufficient to use adapters with few domain-specific parameters for each type of\ninput. We establish that using the classification token of a Vision Transformer\n(ViT) as an adapter is enough to match the classification performance of dual\nnormalization layers, while using significantly less additional parameters.\nFirst, we improve upon the top-1 accuracy of a non-adversarially trained\nViT-B16 model by +1.12% on ImageNet (reaching 83.76% top-1 accuracy). Second,\nand more importantly, we show that training with adapters enables model soups\nthrough linear combinations of the clean and adversarial tokens. These model\nsoups, which we call adversarial model soups, allow us to trade-off between\nclean and robust accuracy without sacrificing efficiency. Finally, we show that\nwe can easily adapt the resulting models in the face of distribution shifts.\nOur ViT-B16 obtains top-1 accuracies on ImageNet variants that are on average\n+4.00% better than those obtained with Masked Autoencoders.\n", "date": "2022-10-10", "categories": []}
{"id": "2210.03429", "title": "Adversarially Robust Prototypical Few-shot Segmentation with Neural-ODEs", "abstract": "  Few-shot Learning (FSL) methods are being adopted in settings where data is\nnot abundantly available. This is especially seen in medical domains where the\nannotations are expensive to obtain. Deep Neural Networks have been shown to be\nvulnerable to adversarial attacks. This is even more severe in the case of FSL\ndue to the lack of a large number of training examples. In this paper, we\nprovide a framework to make few-shot segmentation models adversarially robust\nin the medical domain where such attacks can severely impact the decisions made\nby clinicians who use them. We propose a novel robust few-shot segmentation\nframework, Prototypical Neural Ordinary Differential Equation (PNODE), that\nprovides defense against gradient-based adversarial attacks. We show that our\nframework is more robust compared to traditional adversarial defense mechanisms\nsuch as adversarial training. Adversarial training involves increased training\ntime and shows robustness to limited types of attacks depending on the type of\nadversarial examples seen during training. Our proposed framework generalises\nwell to common adversarial attacks like FGSM, PGD and SMIA while having the\nmodel parameters comparable to the existing few-shot segmentation models. We\nshow the effectiveness of our proposed approach on three publicly available\nmulti-organ segmentation datasets in both in-domain and cross-domain settings\nby attacking the support and query sets without the need for ad-hoc adversarial\ntraining.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.02447", "title": "Practical Adversarial Attacks on Spatiotemporal Traffic Forecasting\n  Models", "abstract": "  Machine learning based traffic forecasting models leverage sophisticated\nspatiotemporal auto-correlations to provide accurate predictions of city-wide\ntraffic states. However, existing methods assume a reliable and unbiased\nforecasting environment, which is not always available in the wild. In this\nwork, we investigate the vulnerability of spatiotemporal traffic forecasting\nmodels and propose a practical adversarial spatiotemporal attack framework.\nSpecifically, instead of simultaneously attacking all geo-distributed data\nsources, an iterative gradient-guided node saliency method is proposed to\nidentify the time-dependent set of victim nodes. Furthermore, we devise a\nspatiotemporal gradient descent based scheme to generate real-valued\nadversarial traffic states under a perturbation constraint. Meanwhile, we\ntheoretically demonstrate the worst performance bound of adversarial traffic\nforecasting attacks. Extensive experiments on two real-world datasets show that\nthe proposed two-step framework achieves up to $67.8\\%$ performance degradation\non various advanced spatiotemporal forecasting models. Remarkably, we also show\nthat adversarial training with our proposed attacks can significantly improve\nthe robustness of spatiotemporal traffic forecasting models. Our code is\navailable in \\url{https://github.com/luckyfan-cs/ASTFA}.\n", "date": "2022-10-05", "categories": []}
{"id": "2210.02502", "title": "On Adversarial Robustness of Deep Image Deblurring", "abstract": "  Recent approaches employ deep learning-based solutions for the recovery of a\nsharp image from its blurry observation. This paper introduces adversarial\nattacks against deep learning-based image deblurring methods and evaluates the\nrobustness of these neural networks to untargeted and targeted attacks. We\ndemonstrate that imperceptible distortion can significantly degrade the\nperformance of state-of-the-art deblurring networks, even producing drastically\ndifferent content in the output, indicating the strong need to include\nadversarially robust training not only in classification but also for image\nrecovery.\n", "date": "2022-10-05", "categories": []}
{"id": "2210.02577", "title": "A Closer Look at Robustness to L-infinity and Spatial Perturbations and\n  their Composition", "abstract": "  In adversarial machine learning, the popular $\\ell_\\infty$ threat model has\nbeen the focus of much previous work. While this mathematical definition of\nimperceptibility successfully captures an infinite set of additive image\ntransformations that a model should be robust to, this is only a subset of all\ntransformations which leave the semantic label of an image unchanged. Indeed,\nprevious work also considered robustness to spatial attacks as well as other\nsemantic transformations; however, designing defense methods against the\ncomposition of spatial and $\\ell_{\\infty}$ perturbations remains relatively\nunderexplored. In the following, we improve the understanding of this seldom\ninvestigated compositional setting. We prove theoretically that no linear\nclassifier can achieve more than trivial accuracy against a composite adversary\nin a simple statistical setting, illustrating its difficulty. We then\ninvestigate how state-of-the-art $\\ell_{\\infty}$ defenses can be adapted to\nthis novel threat model and study their performance against compositional\nattacks. We find that our newly proposed TRADES$_{\\text{All}}$ strategy\nperforms the strongest of all. Analyzing its logit's Lipschitz constant for RT\ntransformations of different sizes, we find that TRADES$_{\\text{All}}$ remains\nstable over a wide range of RT transformations with and without $\\ell_\\infty$\nperturbations.\n", "date": "2022-10-05", "categories": []}
{"id": "2210.02618", "title": "Dynamic Stochastic Ensemble with Adversarial Robust Lottery Ticket\n  Subnetworks", "abstract": "  Adversarial attacks are considered the intrinsic vulnerability of CNNs.\nDefense strategies designed for attacks have been stuck in the adversarial\nattack-defense arms race, reflecting the imbalance between attack and defense.\nDynamic Defense Framework (DDF) recently changed the passive safety status quo\nbased on the stochastic ensemble model. The diversity of subnetworks, an\nessential concern in the DDF, can be effectively evaluated by the adversarial\ntransferability between different networks. Inspired by the poor adversarial\ntransferability between subnetworks of scratch tickets with various remaining\nratios, we propose a method to realize the dynamic stochastic ensemble defense\nstrategy. We discover the adversarial transferable diversity between robust\nlottery ticket subnetworks drawn from different basic structures and sparsity.\nThe experimental results suggest that our method achieves better robust and\nclean recognition accuracy by adversarial transferable diversity, which would\ndecrease the reliability of attacks.\n", "date": "2022-10-06", "categories": []}
{"id": "2210.02840", "title": "Deep Reinforcement Learning based Evasion Generative Adversarial Network\n  for Botnet Detection", "abstract": "  Botnet detectors based on machine learning are potential targets for\nadversarial evasion attacks. Several research works employ adversarial training\nwith samples generated from generative adversarial nets (GANs) to make the\nbotnet detectors adept at recognising adversarial evasions. However, the\nsynthetic evasions may not follow the original semantics of the input samples.\nThis paper proposes a novel GAN model leveraged with deep reinforcement\nlearning (DRL) to explore semantic aware samples and simultaneously harden its\ndetection. A DRL agent is used to attack the discriminator of the GAN that acts\nas a botnet detector. The discriminator is trained on the crafted perturbations\nby the agent during the GAN training, which helps the GAN generator converge\nearlier than the case without DRL. We name this model RELEVAGAN, i.e. [\"relive\na GAN\" or deep REinforcement Learning-based Evasion Generative Adversarial\nNetwork] because, with the help of DRL, it minimises the GAN's job by letting\nits generator explore the evasion samples within the semantic limits. During\nthe GAN training, the attacks are conducted to adjust the discriminator weights\nfor learning crafted perturbations by the agent. RELEVAGAN does not require\nadversarial training for the ML classifiers since it can act as an adversarial\nsemantic-aware botnet detection model. Code will be available at\nhttps://github.com/rhr407/RELEVAGAN.\n", "date": "2022-10-06", "categories": []}
{"id": "2210.01940", "title": "On the Robustness of Deep Clustering Models: Adversarial Attacks and\n  Defenses", "abstract": "  Clustering models constitute a class of unsupervised machine learning methods\nwhich are used in a number of application pipelines, and play a vital role in\nmodern data science. With recent advancements in deep learning -- deep\nclustering models have emerged as the current state-of-the-art over traditional\nclustering approaches, especially for high-dimensional image datasets. While\ntraditional clustering approaches have been analyzed from a robustness\nperspective, no prior work has investigated adversarial attacks and robustness\nfor deep clustering models in a principled manner. To bridge this gap, we\npropose a blackbox attack using Generative Adversarial Networks (GANs) where\nthe adversary does not know which deep clustering model is being used, but can\nquery it for outputs. We analyze our attack against multiple state-of-the-art\ndeep clustering models and real-world datasets, and find that it is highly\nsuccessful. We then employ some natural unsupervised defense approaches, but\nfind that these are unable to mitigate our attack. Finally, we attack Face++, a\nproduction-level face clustering API service, and find that we can\nsignificantly reduce its performance as well. Through this work, we thus aim to\nmotivate the need for truly robust deep clustering models.\n", "date": "2022-10-04", "categories": []}
{"id": "2210.02041", "title": "Natural Color Fool: Towards Boosting Black-box Unrestricted Attacks", "abstract": "  Unrestricted color attacks, which manipulate semantically meaningful color of\nan image, have shown their stealthiness and success in fooling both human eyes\nand deep neural networks. However, current works usually sacrifice the\nflexibility of the uncontrolled setting to ensure the naturalness of\nadversarial examples. As a result, the black-box attack performance of these\nmethods is limited. To boost transferability of adversarial examples without\ndamaging image quality, we propose a novel Natural Color Fool (NCF) which is\nguided by realistic color distributions sampled from a publicly available\ndataset and optimized by our neighborhood search and initialization reset. By\nconducting extensive experiments and visualizations, we convincingly\ndemonstrate the effectiveness of our proposed method. Notably, on average,\nresults show that our NCF can outperform state-of-the-art approaches by\n15.0%$\\sim$32.9% for fooling normally trained models and 10.0%$\\sim$25.3% for\nevading defense methods. Our code is available at\nhttps://github.com/ylhz/Natural-Color-Fool.\n", "date": "2022-10-05", "categories": []}
{"id": "2207.05729", "title": "Physical Passive Patch Adversarial Attacks on Visual Odometry Systems", "abstract": "  Deep neural networks are known to be susceptible to adversarial perturbations\n-- small perturbations that alter the output of the network and exist under\nstrict norm limitations. While such perturbations are usually discussed as\ntailored to a specific input, a universal perturbation can be constructed to\nalter the model's output on a set of inputs. Universal perturbations present a\nmore realistic case of adversarial attacks, as awareness of the model's exact\ninput is not required. In addition, the universal attack setting raises the\nsubject of generalization to unseen data, where given a set of inputs, the\nuniversal perturbations aim to alter the model's output on out-of-sample data.\nIn this work, we study physical passive patch adversarial attacks on visual\nodometry-based autonomous navigation systems. A visual odometry system aims to\ninfer the relative camera motion between two corresponding viewpoints, and is\nfrequently used by vision-based autonomous navigation systems to estimate their\nstate. For such navigation systems, a patch adversarial perturbation poses a\nsevere security issue, as it can be used to mislead a system onto some\ncollision course. To the best of our knowledge, we show for the first time that\nthe error margin of a visual odometry model can be significantly increased by\ndeploying patch adversarial attacks in the scene. We provide evaluation on\nsynthetic closed-loop drone navigation data and demonstrate that a comparable\nvulnerability exists in real data. A reference implementation of the proposed\nmethod and the reported experiments is provided at\nhttps://github.com/patchadversarialattacks/patchadversarialattacks.\n", "date": "2022-07-11", "categories": []}
{"id": "2209.09883", "title": "Leveraging Local Patch Differences in Multi-Object Scenes for Generative\n  Adversarial Attacks", "abstract": "  State-of-the-art generative model-based attacks against image classifiers\noverwhelmingly focus on single-object (i.e., single dominant object) images.\nDifferent from such settings, we tackle a more practical problem of generating\nadversarial perturbations using multi-object (i.e., multiple dominant objects)\nimages as they are representative of most real-world scenes. Our goal is to\ndesign an attack strategy that can learn from such natural scenes by leveraging\nthe local patch differences that occur inherently in such images (e.g.\ndifference between the local patch on the object `person' and the object `bike'\nin a traffic scene). Our key idea is to misclassify an adversarial multi-object\nimage by confusing the victim classifier for each local patch in the image.\nBased on this, we propose a novel generative attack (called Local Patch\nDifference or LPD-Attack) where a novel contrastive loss function uses the\naforesaid local differences in feature space of multi-object scenes to optimize\nthe perturbation generator. Through various experiments across diverse victim\nconvolutional neural networks, we show that our approach outperforms baseline\ngenerative attacks with highly transferable perturbations when evaluated under\ndifferent white-box and black-box settings.\n", "date": "2022-09-20", "categories": []}
{"id": "2210.01288", "title": "Strength-Adaptive Adversarial Training", "abstract": "  Adversarial training (AT) is proved to reliably improve network's robustness\nagainst adversarial data. However, current AT with a pre-specified perturbation\nbudget has limitations in learning a robust network. Firstly, applying a\npre-specified perturbation budget on networks of various model capacities will\nyield divergent degree of robustness disparity between natural and robust\naccuracies, which deviates from robust network's desideratum. Secondly, the\nattack strength of adversarial training data constrained by the pre-specified\nperturbation budget fails to upgrade as the growth of network robustness, which\nleads to robust overfitting and further degrades the adversarial robustness. To\novercome these limitations, we propose \\emph{Strength-Adaptive Adversarial\nTraining} (SAAT). Specifically, the adversary employs an adversarial loss\nconstraint to generate adversarial training data. Under this constraint, the\nperturbation budget will be adaptively adjusted according to the training state\nof adversarial data, which can effectively avoid robust overfitting. Besides,\nSAAT explicitly constrains the attack strength of training data through the\nadversarial loss, which manipulates model capacity scheduling during training,\nand thereby can flexibly control the degree of robustness disparity and adjust\nthe tradeoff between natural accuracy and robustness. Extensive experiments\nshow that our proposal boosts the robustness of adversarial training.\n", "date": "2022-10-04", "categories": []}
{"id": "2210.01632", "title": "Backdoor Attacks in the Supply Chain of Masked Image Modeling", "abstract": "  Masked image modeling (MIM) revolutionizes self-supervised learning (SSL) for\nimage pre-training. In contrast to previous dominating self-supervised methods,\ni.e., contrastive learning, MIM attains state-of-the-art performance by masking\nand reconstructing random patches of the input image. However, the associated\nsecurity and privacy risks of this novel generative method are unexplored. In\nthis paper, we perform the first security risk quantification of MIM through\nthe lens of backdoor attacks. Different from previous work, we are the first to\nsystematically threat modeling on SSL in every phase of the model supply chain,\ni.e., pre-training, release, and downstream phases. Our evaluation shows that\nmodels built with MIM are vulnerable to existing backdoor attacks in release\nand downstream phases and are compromised by our proposed method in\npre-training phase. For instance, on CIFAR10, the attack success rate can reach\n99.62%, 96.48%, and 98.89% in the downstream phase, release phase, and\npre-training phase, respectively. We also take the first step to investigate\nthe success factors of backdoor attacks in the pre-training phase and find the\ntrigger number and trigger pattern play key roles in the success of backdoor\nattacks while trigger location has only tiny effects. In the end, our empirical\nstudy of the defense mechanisms across three detection-level on model supply\nchain phases indicates that different defenses are suitable for backdoor\nattacks in different phases. However, backdoor attacks in the release phase\ncannot be detected by all three detection-level methods, calling for more\neffective defenses in future research.\n", "date": "2022-10-04", "categories": []}
{"id": "2206.09305", "title": "Adversarial Scrutiny of Evidentiary Statistical Software", "abstract": "  The U.S. criminal legal system increasingly relies on software output to\nconvict and incarcerate people. In a large number of cases each year, the\ngovernment makes these consequential decisions based on evidence from\nstatistical software -- such as probabilistic genotyping, environmental audio\ndetection, and toolmark analysis tools -- that defense counsel cannot fully\ncross-examine or scrutinize. This undermines the commitments of the adversarial\ncriminal legal system, which relies on the defense's ability to probe and test\nthe prosecution's case to safeguard individual rights.\n  Responding to this need to adversarially scrutinize output from such\nsoftware, we propose robust adversarial testing as an audit framework to\nexamine the validity of evidentiary statistical software. We define and\noperationalize this notion of robust adversarial testing for defense use by\ndrawing on a large body of recent work in robust machine learning and\nalgorithmic fairness. We demonstrate how this framework both standardizes the\nprocess for scrutinizing such tools and empowers defense lawyers to examine\ntheir validity for instances most relevant to the case at hand. We further\ndiscuss existing structural and institutional challenges within the U.S.\ncriminal legal system that may create barriers for implementing this and other\nsuch audit frameworks and close with a discussion on policy changes that could\nhelp address these concerns.\n", "date": "2022-06-19", "categories": []}
{"id": "2210.00122", "title": "Adversarial Robustness of Representation Learning for Knowledge Graphs", "abstract": "  Knowledge graphs represent factual knowledge about the world as relationships\nbetween concepts and are critical for intelligent decision making in enterprise\napplications. New knowledge is inferred from the existing facts in the\nknowledge graphs by encoding the concepts and relations into low-dimensional\nfeature vector representations. The most effective representations for this\ntask, called Knowledge Graph Embeddings (KGE), are learned through neural\nnetwork architectures. Due to their impressive predictive performance, they are\nincreasingly used in high-impact domains like healthcare, finance and\neducation. However, are the black-box KGE models adversarially robust for use\nin domains with high stakes? This thesis argues that state-of-the-art KGE\nmodels are vulnerable to data poisoning attacks, that is, their predictive\nperformance can be degraded by systematically crafted perturbations to the\ntraining knowledge graph. To support this argument, two novel data poisoning\nattacks are proposed that craft input deletions or additions at training time\nto subvert the learned model's performance at inference time. These adversarial\nattacks target the task of predicting the missing facts in knowledge graphs\nusing KGE models, and the evaluation shows that the simpler attacks are\ncompetitive with or outperform the computationally expensive ones. The thesis\ncontributions not only highlight and provide an opportunity to fix the security\nvulnerabilities of KGE models, but also help to understand the black-box\npredictive behaviour of KGE models.\n", "date": "2022-09-30", "categories": []}
{"id": "2210.00292", "title": "DeltaBound Attack: Efficient decision-based attack in low queries regime", "abstract": "  Deep neural networks and other machine learning systems, despite being\nextremely powerful and able to make predictions with high accuracy, are\nvulnerable to adversarial attacks. We proposed the DeltaBound attack: a novel,\npowerful attack in the hard-label setting with $\\ell_2$ norm bounded\nperturbations. In this scenario, the attacker has only access to the top-1\npredicted label of the model and can be therefore applied to real-world\nsettings such as remote API. This is a complex problem since the attacker has\nvery little information about the model. Consequently, most of the other\ntechniques present in the literature require a massive amount of queries for\nattacking a single example. Oppositely, this work mainly focuses on the\nevaluation of attack's power in the low queries regime $\\leq 1000$ queries)\nwith $\\ell_2$ norm in the hard-label settings. We find that the DeltaBound\nattack performs as well and sometimes better than current state-of-the-art\nattacks while remaining competitive across different kinds of models. Moreover,\nwe evaluate our method against not only deep neural networks, but also non-deep\nlearning models, such as Gradient Boosting Decision Trees and Multinomial Naive\nBayes.\n", "date": "2022-10-01", "categories": []}
{"id": "2210.00430", "title": "Understanding Adversarial Robustness Against On-manifold Adversarial\n  Examples", "abstract": "  Deep neural networks (DNNs) are shown to be vulnerable to adversarial\nexamples. A well-trained model can be easily attacked by adding small\nperturbations to the original data. One of the hypotheses of the existence of\nthe adversarial examples is the off-manifold assumption: adversarial examples\nlie off the data manifold. However, recent research showed that on-manifold\nadversarial examples also exist. In this paper, we revisit the off-manifold\nassumption and want to study a question: at what level is the poor performance\nof neural networks against adversarial attacks due to on-manifold adversarial\nexamples? Since the true data manifold is unknown in practice, we consider two\napproximated on-manifold adversarial examples on both real and synthesis\ndatasets. On real datasets, we show that on-manifold adversarial examples have\ngreater attack rates than off-manifold adversarial examples on both\nstandard-trained and adversarially-trained models. On synthetic datasets,\ntheoretically, We prove that on-manifold adversarial examples are powerful, yet\nadversarial training focuses on off-manifold directions and ignores the\non-manifold adversarial examples. Furthermore, we provide analysis to show that\nthe properties derived theoretically can also be observed in practice. Our\nanalysis suggests that on-manifold adversarial examples are important, and we\nshould pay more attention to on-manifold adversarial examples for training\nrobust models.\n", "date": "2022-10-02", "categories": []}
{"id": "2210.00557", "title": "Adaptive Smoothness-weighted Adversarial Training for Multiple\n  Perturbations with Its Stability Analysis", "abstract": "  Adversarial Training (AT) has been demonstrated as one of the most effective\nmethods against adversarial examples. While most existing works focus on AT\nwith a single type of perturbation e.g., the $\\ell_\\infty$ attacks), DNNs are\nfacing threats from different types of adversarial examples. Therefore,\nadversarial training for multiple perturbations (ATMP) is proposed to\ngeneralize the adversarial robustness over different perturbation types (in\n$\\ell_1$, $\\ell_2$, and $\\ell_\\infty$ norm-bounded perturbations). However, the\nresulting model exhibits trade-off between different attacks. Meanwhile, there\nis no theoretical analysis of ATMP, limiting its further development. In this\npaper, we first provide the smoothness analysis of ATMP and show that $\\ell_1$,\n$\\ell_2$, and $\\ell_\\infty$ adversaries give different contributions to the\nsmoothness of the loss function of ATMP. Based on this, we develop the\nstability-based excess risk bounds and propose adaptive smoothness-weighted\nadversarial training for multiple perturbations. Theoretically, our algorithm\nyields better bounds. Empirically, our experiments on CIFAR10 and CIFAR100\nachieve the state-of-the-art performance against the mixture of multiple\nperturbations attacks.\n", "date": "2022-10-02", "categories": []}
{"id": "2210.00753", "title": "Push-Pull: Characterizing the Adversarial Robustness for Audio-Visual\n  Active Speaker Detection", "abstract": "  Audio-visual active speaker detection (AVASD) is well-developed, and now is\nan indispensable front-end for several multi-modal applications. However, to\nthe best of our knowledge, the adversarial robustness of AVASD models hasn't\nbeen investigated, not to mention the effective defense against such attacks.\nIn this paper, we are the first to reveal the vulnerability of AVASD models\nunder audio-only, visual-only, and audio-visual adversarial attacks through\nextensive experiments. What's more, we also propose a novel audio-visual\ninteraction loss (AVIL) for making attackers difficult to find feasible\nadversarial examples under an allocated attack budget. The loss aims at pushing\nthe inter-class embeddings to be dispersed, namely non-speech and speech\nclusters, sufficiently disentangled, and pulling the intra-class embeddings as\nclose as possible to keep them compact. Experimental results show the AVIL\noutperforms the adversarial training by 33.14 mAP (%) under multi-modal\nattacks.\n", "date": "2022-10-03", "categories": []}
{"id": "2210.00933", "title": "Perceptual Attacks of No-Reference Image Quality Models with\n  Human-in-the-Loop", "abstract": "  No-reference image quality assessment (NR-IQA) aims to quantify how humans\nperceive visual distortions of digital images without access to their\nundistorted references. NR-IQA models are extensively studied in computational\nvision, and are widely used for performance evaluation and perceptual\noptimization of man-made vision systems. Here we make one of the first attempts\nto examine the perceptual robustness of NR-IQA models. Under a Lagrangian\nformulation, we identify insightful connections of the proposed perceptual\nattack to previous beautiful ideas in computer vision and machine learning. We\ntest one knowledge-driven and three data-driven NR-IQA methods under four\nfull-reference IQA models (as approximations to human perception of\njust-noticeable differences). Through carefully designed psychophysical\nexperiments, we find that all four NR-IQA models are vulnerable to the proposed\nperceptual attack. More interestingly, we observe that the generated\ncounterexamples are not transferable, manifesting themselves as distinct design\nflows of respective NR-IQA methods.\n", "date": "2022-10-03", "categories": []}
{"id": "2210.01002", "title": "ASGNN: Graph Neural Networks with Adaptive Structure", "abstract": "  The graph neural network (GNN) models have presented impressive achievements\nin numerous machine learning tasks. However, many existing GNN models are shown\nto be vulnerable to adversarial attacks, which creates a stringent need to\nbuild robust GNN architectures. In this work, we propose a novel interpretable\nmessage passing scheme with adaptive structure (ASMP) to defend against\nadversarial attacks on graph structure. Layers in ASMP are derived based on\noptimization steps that minimize an objective function that learns the node\nfeature and the graph structure simultaneously. ASMP is adaptive in the sense\nthat the message passing process in different layers is able to be carried out\nover dynamically adjusted graphs. Such property allows more fine-grained\nhandling of the noisy (or perturbed) graph structure and hence improves the\nrobustness. Convergence properties of the ASMP scheme are theoretically\nestablished. Integrating ASMP with neural networks can lead to a new family of\nGNN models with adaptive structure (ASGNN). Extensive experiments on\nsemi-supervised node classification tasks demonstrate that the proposed ASGNN\noutperforms the state-of-the-art GNN architectures in terms of classification\nperformance under various adversarial attacks.\n", "date": "2022-10-03", "categories": []}
{"id": "2210.01111", "title": "MultiGuard: Provably Robust Multi-label Classification against\n  Adversarial Examples", "abstract": "  Multi-label classification, which predicts a set of labels for an input, has\nmany applications. However, multiple recent studies showed that multi-label\nclassification is vulnerable to adversarial examples. In particular, an\nattacker can manipulate the labels predicted by a multi-label classifier for an\ninput via adding carefully crafted, human-imperceptible perturbation to it.\nExisting provable defenses for multi-class classification achieve sub-optimal\nprovable robustness guarantees when generalized to multi-label classification.\nIn this work, we propose MultiGuard, the first provably robust defense against\nadversarial examples to multi-label classification. Our MultiGuard leverages\nrandomized smoothing, which is the state-of-the-art technique to build provably\nrobust classifiers. Specifically, given an arbitrary multi-label classifier,\nour MultiGuard builds a smoothed multi-label classifier via adding random noise\nto the input. We consider isotropic Gaussian noise in this work. Our major\ntheoretical contribution is that we show a certain number of ground truth\nlabels of an input are provably in the set of labels predicted by our\nMultiGuard when the $\\ell_2$-norm of the adversarial perturbation added to the\ninput is bounded. Moreover, we design an algorithm to compute our provable\nrobustness guarantees. Empirically, we evaluate our MultiGuard on VOC 2007,\nMS-COCO, and NUS-WIDE benchmark datasets. Our code is available at:\n\\url{https://github.com/quwenjie/MultiGuard}\n", "date": "2022-10-03", "categories": []}
{"id": "2208.09285", "title": "Shadows Aren't So Dangerous After All: A Fast and Robust Defense Against\n  Shadow-Based Adversarial Attacks", "abstract": "  Robust classification is essential in tasks like autonomous vehicle sign\nrecognition, where the downsides of misclassification can be grave. Adversarial\nattacks threaten the robustness of neural network classifiers, causing them to\nconsistently and confidently misidentify road signs. One such class of attack,\nshadow-based attacks, causes misidentifications by applying a natural-looking\nshadow to input images, resulting in road signs that appear natural to a human\nobserver but confusing for these classifiers. Current defenses against such\nattacks use a simple adversarial training procedure to achieve a rather low\n25\\% and 40\\% robustness on the GTSRB and LISA test sets, respectively. In this\npaper, we propose a robust, fast, and generalizable method, designed to defend\nagainst shadow attacks in the context of road sign recognition, that augments\nsource images with binary adaptive threshold and edge maps. We empirically show\nits robustness against shadow attacks, and reformulate the problem to show its\nsimilarity to $\\varepsilon$ perturbation-based attacks. Experimental results\nshow that our edge defense results in 78\\% robustness while maintaining 98\\%\nbenign test accuracy on the GTSRB test set, with similar results from our\nthreshold defense. Link to our code is in the paper.\n", "date": "2022-08-18", "categories": []}
{"id": "2209.14243", "title": "A Closer Look at Evaluating the Bit-Flip Attack Against Deep Neural\n  Networks", "abstract": "  Deep neural network models are massively deployed on a wide variety of\nhardware platforms. This results in the appearance of new attack vectors that\nsignificantly extend the standard attack surface, extensively studied by the\nadversarial machine learning community. One of the first attack that aims at\ndrastically dropping the performance of a model, by targeting its parameters\n(weights) stored in memory, is the Bit-Flip Attack (BFA). In this work, we\npoint out several evaluation challenges related to the BFA. First of all, the\nlack of an adversary's budget in the standard threat model is problematic,\nespecially when dealing with physical attacks. Moreover, since the BFA presents\ncritical variability, we discuss the influence of some training parameters and\nthe importance of the model architecture. This work is the first to present the\nimpact of the BFA against fully-connected architectures that present different\nbehaviors compared to convolutional neural networks. These results highlight\nthe importance of defining robust and sound evaluation methodologies to\nproperly evaluate the dangers of parameter-based attacks as well as measure the\nreal level of robustness offered by a defense.\n", "date": "2022-09-28", "categories": []}
{"id": "2209.15246", "title": "Your Out-of-Distribution Detection Method is Not Robust!", "abstract": "  Out-of-distribution (OOD) detection has recently gained substantial attention\ndue to the importance of identifying out-of-domain samples in reliability and\nsafety. Although OOD detection methods have advanced by a great deal, they are\nstill susceptible to adversarial examples, which is a violation of their\npurpose. To mitigate this issue, several defenses have recently been proposed.\nNevertheless, these efforts remained ineffective, as their evaluations are\nbased on either small perturbation sizes, or weak attacks. In this work, we\nre-examine these defenses against an end-to-end PGD attack on in/out data with\nlarger perturbation sizes, e.g. up to commonly used $\\epsilon=8/255$ for the\nCIFAR-10 dataset. Surprisingly, almost all of these defenses perform worse than\na random detection under the adversarial setting. Next, we aim to provide a\nrobust OOD detection method. In an ideal defense, the training should expose\nthe model to almost all possible adversarial perturbations, which can be\nachieved through adversarial training. That is, such training perturbations\nshould based on both in- and out-of-distribution samples. Therefore, unlike OOD\ndetection in the standard setting, access to OOD, as well as in-distribution,\nsamples sounds necessary in the adversarial training setup. These tips lead us\nto adopt generative OOD detection methods, such as OpenGAN, as a baseline. We\nsubsequently propose the Adversarially Trained Discriminator (ATD), which\nutilizes a pre-trained robust model to extract robust features, and a generator\nmodel to create OOD samples. Using ATD with CIFAR-10 and CIFAR-100 as the\nin-distribution data, we could significantly outperform all previous methods in\nthe robust AUROC while maintaining high standard AUROC and classification\naccuracy. The code repository is available at https://github.com/rohban-lab/ATD .\n", "date": "2022-09-30", "categories": []}
{"id": "2209.13523", "title": "Watch What You Pretrain For: Targeted, Transferable Adversarial Examples\n  on Self-Supervised Speech Recognition models", "abstract": "  A targeted adversarial attack produces audio samples that can force an\nAutomatic Speech Recognition (ASR) system to output attacker-chosen text. To\nexploit ASR models in real-world, black-box settings, an adversary can leverage\nthe transferability property, i.e. that an adversarial sample produced for a\nproxy ASR can also fool a different remote ASR. However recent work has shown\nthat transferability against large ASR models is very difficult. In this work,\nwe show that modern ASR architectures, specifically ones based on\nSelf-Supervised Learning, are in fact vulnerable to transferability. We\nsuccessfully demonstrate this phenomenon by evaluating state-of-the-art\nself-supervised ASR models like Wav2Vec2, HuBERT, Data2Vec and WavLM. We show\nthat with low-level additive noise achieving a 30dB Signal-Noise Ratio, we can\nachieve target transferability with up to 80% accuracy. Next, we 1) use an\nablation study to show that Self-Supervised learning is the main cause of that\nphenomenon, and 2) we provide an explanation for this phenomenon. Through this\nwe show that modern ASR architectures are uniquely vulnerable to adversarial\nsecurity threats.\n", "date": "2022-09-17", "categories": []}
{"id": "2209.13785", "title": "Attacking Compressed Vision Transformers", "abstract": "  Vision Transformers are increasingly embedded in industrial systems due to\ntheir superior performance, but their memory and power requirements make\ndeploying them to edge devices a challenging task. Hence, model compression\ntechniques are now widely used to deploy models on edge devices as they\ndecrease the resource requirements and make model inference very fast and\nefficient. But their reliability and robustness from a security perspective is\nanother major issue in safety-critical applications. Adversarial attacks are\nlike optical illusions for ML algorithms and they can severely impact the\naccuracy and reliability of models. In this work we investigate the\ntransferability of adversarial samples across the SOTA Vision Transformer\nmodels across 3 SOTA compressed versions and infer the effects different\ncompression techniques have on adversarial attacks.\n", "date": "2022-09-28", "categories": []}
{"id": "2209.13963", "title": "Machine Beats Machine: Machine Learning Models to Defend Against\n  Adversarial Attacks", "abstract": "  We propose using a two-layered deployment of machine learning models to\nprevent adversarial attacks. The first layer determines whether the data was\ntampered, while the second layer solves a domain-specific problem. We explore\nthree sets of features and three dataset variations to train machine learning\nmodels. Our results show clustering algorithms achieved promising results. In\nparticular, we consider the best results were obtained by applying the DBSCAN\nalgorithm to the structured structural similarity index measure computed\nbetween the images and a white reference image.\n", "date": "2022-09-28", "categories": []}
{"id": "2209.14053", "title": "Inducing Data Amplification Using Auxiliary Datasets in Adversarial\n  Training", "abstract": "  Several recent studies have shown that the use of extra in-distribution data\ncan lead to a high level of adversarial robustness. However, there is no\nguarantee that it will always be possible to obtain sufficient extra data for a\nselected dataset. In this paper, we propose a biased multi-domain adversarial\ntraining (BiaMAT) method that induces training data amplification on a primary\ndataset using publicly available auxiliary datasets, without requiring the\nclass distribution match between the primary and auxiliary datasets. The\nproposed method can achieve increased adversarial robustness on a primary\ndataset by leveraging auxiliary datasets via multi-domain learning.\nSpecifically, data amplification on both robust and non-robust features can be\naccomplished through the application of BiaMAT as demonstrated through a\ntheoretical and empirical analysis. Moreover, we demonstrate that while\nexisting methods are vulnerable to negative transfer due to the distributional\ndiscrepancy between auxiliary and primary data, the proposed method enables\nneural networks to flexibly leverage diverse image datasets for adversarial\ntraining by successfully handling the domain discrepancy through the\napplication of a confidence-based selection strategy. The pre-trained models\nand code are available at: \\url{https://github.com/Saehyung-Lee/BiaMAT}.\n", "date": "2022-09-27", "categories": []}
{"id": "2209.13007", "title": "Mitigating Attacks on Artificial Intelligence-based Spectrum Sensing for\n  Cellular Network Signals", "abstract": "  Cellular networks (LTE, 5G, and beyond) are dramatically growing with high\ndemand from consumers and more promising than the other wireless networks with\nadvanced telecommunication technologies. The main goal of these networks is to\nconnect billions of devices, systems, and users with high-speed data\ntransmission, high cell capacity, and low latency, as well as to support a wide\nrange of new applications, such as virtual reality, metaverse, telehealth,\nonline education, autonomous and flying vehicles, advanced manufacturing, and\nmany more. To achieve these goals, spectrum sensing has been paid more\nattention, along with new approaches using artificial intelligence (AI) methods\nfor spectrum management in cellular networks. This paper provides a\nvulnerability analysis of spectrum sensing approaches using AI-based semantic\nsegmentation models for identifying cellular network signals under adversarial\nattacks with and without defensive distillation methods. The results showed\nthat mitigation methods can significantly reduce the vulnerabilities of\nAI-based spectrum sensing models against adversarial attacks.\n", "date": "2022-09-27", "categories": []}
{"id": "2209.13113", "title": "FG-UAP: Feature-Gathering Universal Adversarial Perturbation", "abstract": "  Deep Neural Networks (DNNs) are susceptible to elaborately designed\nperturbations, whether such perturbations are dependent or independent of\nimages. The latter one, called Universal Adversarial Perturbation (UAP), is\nvery attractive for model robustness analysis, since its independence of input\nreveals the intrinsic characteristics of the model. Relatively, another\ninteresting observation is Neural Collapse (NC), which means the feature\nvariability may collapse during the terminal phase of training. Motivated by\nthis, we propose to generate UAP by attacking the layer where NC phenomenon\nhappens. Because of NC, the proposed attack could gather all the natural\nimages' features to its surrounding, which is hence called Feature-Gathering\nUAP (FG-UAP).\n  We evaluate the effectiveness our proposed algorithm on abundant experiments,\nincluding untargeted and targeted universal attacks, attacks under limited\ndataset, and transfer-based black-box attacks among different architectures\nincluding Vision Transformers, which are believed to be more robust.\nFurthermore, we investigate FG-UAP in the view of NC by analyzing the labels\nand extracted features of adversarial examples, finding that collapse\nphenomenon becomes stronger after the model is corrupted. The code will be\nreleased when the paper is accepted.\n", "date": "2022-09-27", "categories": []}
{"id": "2209.09996", "title": "Audit and Improve Robustness of Private Neural Networks on Encrypted\n  Data", "abstract": "  Performing neural network inference on encrypted data without decryption is\none popular method to enable privacy-preserving neural networks (PNet) as a\nservice. Compared with regular neural networks deployed for\nmachine-learning-as-a-service, PNet requires additional encoding, e.g.,\nquantized-precision numbers, and polynomial activation. Encrypted input also\nintroduces novel challenges such as adversarial robustness and security. To the\nbest of our knowledge, we are the first to study questions including (i)\nWhether PNet is more robust against adversarial inputs than regular neural\nnetworks? (ii) How to design a robust PNet given the encrypted input without\ndecryption? We propose PNet-Attack to generate black-box adversarial examples\nthat can successfully attack PNet in both target and untarget manners. The\nattack results show that PNet robustness against adversarial inputs needs to be\nimproved. This is not a trivial task because the PNet model owner does not have\naccess to the plaintext of the input values, which prevents the application of\nexisting detection and defense methods such as input tuning, model\nnormalization, and adversarial training. To tackle this challenge, we propose a\nnew fast and accurate noise insertion method, called RPNet, to design Robust\nand Private Neural Networks. Our comprehensive experiments show that\nPNet-Attack reduces at least $2.5\\times$ queries than prior works. We\ntheoretically analyze our RPNet methods and demonstrate that RPNet can decrease\n$\\sim 91.88\\%$ attack success rate.\n", "date": "2022-09-20", "categories": []}
{"id": "2209.09688", "title": "Sparse Vicious Attacks on Graph Neural Networks", "abstract": "  Graph Neural Networks (GNNs) have proven to be successful in several\npredictive modeling tasks for graph-structured data.\n  Amongst those tasks, link prediction is one of the fundamental problems for\nmany real-world applications, such as recommender systems.\n  However, GNNs are not immune to adversarial attacks, i.e., carefully crafted\nmalicious examples that are designed to fool the predictive model.\n  In this work, we focus on a specific, white-box attack to GNN-based link\nprediction models, where a malicious node aims to appear in the list of\nrecommended nodes for a given target victim.\n  To achieve this goal, the attacker node may also count on the cooperation of\nother existing peers that it directly controls, namely on the ability to inject\na number of ``vicious'' nodes in the network.\n  Specifically, all these malicious nodes can add new edges or remove existing\nones, thereby perturbing the original graph.\n  Thus, we propose SAVAGE, a novel framework and a method to mount this type of\nlink prediction attacks.\n  SAVAGE formulates the adversary's goal as an optimization task, striking the\nbalance between the effectiveness of the attack and the sparsity of malicious\nresources required.\n  Extensive experiments conducted on real-world and synthetic datasets\ndemonstrate that adversarial attacks implemented through SAVAGE indeed achieve\nhigh attack success rate yet using a small amount of vicious nodes.\n  Finally, despite those attacks require full knowledge of the target model, we\nshow that they are successfully transferable to other black-box methods for\nlink prediction.\n", "date": "2022-09-20", "categories": []}
{"id": "2209.08724", "title": "On the Adversarial Transferability of ConvMixer Models", "abstract": "  Deep neural networks (DNNs) are well known to be vulnerable to adversarial\nexamples (AEs). In addition, AEs have adversarial transferability, which means\nAEs generated for a source model can fool another black-box model (target\nmodel) with a non-trivial probability. In this paper, we investigate the\nproperty of adversarial transferability between models including ConvMixer,\nwhich is an isotropic network, for the first time. To objectively verify the\nproperty of transferability, the robustness of models is evaluated by using a\nbenchmark attack method called AutoAttack. In an image classification\nexperiment, ConvMixer is confirmed to be weak to adversarial transferability.\n", "date": "2022-09-19", "categories": []}
{"id": "2209.08744", "title": "AdvDO: Realistic Adversarial Attacks for Trajectory Prediction", "abstract": "  Trajectory prediction is essential for autonomous vehicles (AVs) to plan\ncorrect and safe driving behaviors. While many prior works aim to achieve\nhigher prediction accuracy, few study the adversarial robustness of their\nmethods. To bridge this gap, we propose to study the adversarial robustness of\ndata-driven trajectory prediction systems. We devise an optimization-based\nadversarial attack framework that leverages a carefully-designed differentiable\ndynamic model to generate realistic adversarial trajectories. Empirically, we\nbenchmark the adversarial robustness of state-of-the-art prediction models and\nshow that our attack increases the prediction error for both general metrics\nand planning-aware metrics by more than 50% and 37%. We also show that our\nattack can lead an AV to drive off road or collide into other vehicles in\nsimulation. Finally, we demonstrate how to mitigate the adversarial attacks\nusing an adversarial training scheme.\n", "date": "2022-09-19", "categories": []}
{"id": "2209.07534", "title": "Improving Robust Fairness via Balance Adversarial Training", "abstract": "  Adversarial training (AT) methods are effective against adversarial attacks,\nyet they introduce severe disparity of accuracy and robustness between\ndifferent classes, known as the robust fairness problem. Previously proposed\nFair Robust Learning (FRL) adaptively reweights different classes to improve\nfairness. However, the performance of the better-performed classes decreases,\nleading to a strong performance drop. In this paper, we observed two unfair\nphenomena during adversarial training: different difficulties in generating\nadversarial examples from each class (source-class fairness) and disparate\ntarget class tendencies when generating adversarial examples (target-class\nfairness). From the observations, we propose Balance Adversarial Training (BAT)\nto address the robust fairness problem. Regarding source-class fairness, we\nadjust the attack strength and difficulties of each class to generate samples\nnear the decision boundary for easier and fairer model learning; considering\ntarget-class fairness, by introducing a uniform distribution constraint, we\nencourage the adversarial example generation process for each class with a fair\ntendency. Extensive experiments conducted on multiple datasets (CIFAR-10,\nCIFAR-100, and ImageNette) demonstrate that our method can significantly\noutperform other baselines in mitigating the robust fairness problem (+5-10\\%\non the worst class accuracy)\n", "date": "2022-09-15", "categories": []}
{"id": "2209.07592", "title": "Explicit Tradeoffs between Adversarial and Natural Distributional\n  Robustness", "abstract": "  Several existing works study either adversarial or natural distributional\nrobustness of deep neural networks separately. In practice, however, models\nneed to enjoy both types of robustness to ensure reliability. In this work, we\nbridge this gap and show that in fact, explicit tradeoffs exist between\nadversarial and natural distributional robustness. We first consider a simple\nlinear regression setting on Gaussian data with disjoint sets of core and\nspurious features. In this setting, through theoretical and empirical analysis,\nwe show that (i) adversarial training with $\\ell_1$ and $\\ell_2$ norms\nincreases the model reliance on spurious features; (ii) For $\\ell_\\infty$\nadversarial training, spurious reliance only occurs when the scale of the\nspurious features is larger than that of the core features; (iii) adversarial\ntraining can have an unintended consequence in reducing distributional\nrobustness, specifically when spurious correlations are changed in the new test\ndomain. Next, we present extensive empirical evidence, using a test suite of\ntwenty adversarially trained models evaluated on five benchmark datasets\n(ObjectNet, RIVAL10, Salient ImageNet-1M, ImageNet-9, Waterbirds), that\nadversarially trained classifiers rely on backgrounds more than their\nstandardly trained counterparts, validating our theoretical results. We also\nshow that spurious correlations in training data (when preserved in the test\ndomain) can improve adversarial robustness, revealing that previous claims that\nadversarial vulnerability is rooted in spurious correlations are incomplete.\n", "date": "2022-09-15", "categories": []}
{"id": "2209.07735", "title": "Enhance the Visual Representation via Discrete Adversarial Training", "abstract": "  Adversarial Training (AT), which is commonly accepted as one of the most\neffective approaches defending against adversarial examples, can largely harm\nthe standard performance, thus has limited usefulness on industrial-scale\nproduction and applications. Surprisingly, this phenomenon is totally opposite\nin Natural Language Processing (NLP) task, where AT can even benefit for\ngeneralization. We notice the merit of AT in NLP tasks could derive from the\ndiscrete and symbolic input space. For borrowing the advantage from NLP-style\nAT, we propose Discrete Adversarial Training (DAT). DAT leverages VQGAN to\nreform the image data to discrete text-like inputs, i.e. visual words. Then it\nminimizes the maximal risk on such discrete images with symbolic adversarial\nperturbations. We further give an explanation from the perspective of\ndistribution to demonstrate the effectiveness of DAT. As a plug-and-play\ntechnique for enhancing the visual representation, DAT achieves significant\nimprovement on multiple tasks including image classification, object detection\nand self-supervised learning. Especially, the model pre-trained with Masked\nAuto-Encoding (MAE) and fine-tuned by our DAT without extra data can get 31.40\nmCE on ImageNet-C and 32.77% top-1 accuracy on Stylized-ImageNet, building the\nnew state-of-the-art. The code will be available at\nhttps://github.com/alibaba/easyrobust.\n", "date": "2022-09-16", "categories": []}
{"id": "2209.07790", "title": "A Large-scale Multiple-objective Method for Black-box Attack against\n  Object Detection", "abstract": "  Recent studies have shown that detectors based on deep models are vulnerable\nto adversarial examples, even in the black-box scenario where the attacker\ncannot access the model information. Most existing attack methods aim to\nminimize the true positive rate, which often shows poor attack performance, as\nanother sub-optimal bounding box may be detected around the attacked bounding\nbox to be the new true positive one. To settle this challenge, we propose to\nminimize the true positive rate and maximize the false positive rate, which can\nencourage more false positive objects to block the generation of new true\npositive bounding boxes. It is modeled as a multi-objective optimization (MOP)\nproblem, of which the generic algorithm can search the Pareto-optimal. However,\nour task has more than two million decision variables, leading to low searching\nefficiency. Thus, we extend the standard Genetic Algorithm with Random Subset\nselection and Divide-and-Conquer, called GARSDC, which significantly improves\nthe efficiency. Moreover, to alleviate the sensitivity to population quality in\ngeneric algorithms, we generate a gradient-prior initial population, utilizing\nthe transferability between different detectors with similar backbones.\nCompared with the state-of-art attack methods, GARSDC decreases by an average\n12.0 in the mAP and queries by about 1000 times in extensive experiments. Our\ncodes can be found at https://github.com/LiangSiyuan21/ GARSDC.\n", "date": "2022-09-16", "categories": []}
{"id": "2206.09868", "title": "Understanding Robust Learning through the Lens of Representation\n  Similarities", "abstract": "  Representation learning, i.e. the generation of representations useful for\ndownstream applications, is a task of fundamental importance that underlies\nmuch of the success of deep neural networks (DNNs). Recently, robustness to\nadversarial examples has emerged as a desirable property for DNNs, spurring the\ndevelopment of robust training methods that account for adversarial examples.\nIn this paper, we aim to understand how the properties of representations\nlearned by robust training differ from those obtained from standard, non-robust\ntraining. This is critical to diagnosing numerous salient pitfalls in robust\nnetworks, such as, degradation of performance on benign inputs, poor\ngeneralization of robustness, and increase in over-fitting. We utilize a\npowerful set of tools known as representation similarity metrics, across three\nvision datasets, to obtain layer-wise comparisons between robust and non-robust\nDNNs with different training procedures, architectural parameters and\nadversarial constraints. Our experiments highlight hitherto unseen properties\nof robust representations that we posit underlie the behavioral differences of\nrobust networks. We discover a lack of specialization in robust networks'\nrepresentations along with a disappearance of `block structure'. We also find\noverfitting during robust training largely impacts deeper layers. These, along\nwith other findings, suggest ways forward for the design and training of better\nrobust networks.\n", "date": "2022-06-20", "categories": []}
{"id": "2209.06931", "title": "Robust Transferable Feature Extractors: Learning to Defend Pre-Trained\n  Networks Against White Box Adversaries", "abstract": "  The widespread adoption of deep neural networks in computer vision\napplications has brought forth a significant interest in adversarial\nrobustness. Existing research has shown that maliciously perturbed inputs\nspecifically tailored for a given model (i.e., adversarial examples) can be\nsuccessfully transferred to another independently trained model to induce\nprediction errors. Moreover, this property of adversarial examples has been\nattributed to features derived from predictive patterns in the data\ndistribution. Thus, we are motivated to investigate the following question: Can\nadversarial defenses, like adversarial examples, be successfully transferred to\nother independently trained models? To this end, we propose a deep\nlearning-based pre-processing mechanism, which we refer to as a robust\ntransferable feature extractor (RTFE). After examining theoretical motivation\nand implications, we experimentally show that our method can provide\nadversarial robustness to multiple independently pre-trained classifiers that\nare otherwise ineffective against an adaptive white box adversary. Furthermore,\nwe show that RTFEs can even provide one-shot adversarial robustness to models\nindependently trained on different datasets.\n", "date": "2022-09-14", "categories": []}
{"id": "2209.06953", "title": "On the interplay of adversarial robustness and architecture components:\n  patches, convolution and attention", "abstract": "  In recent years novel architecture components for image classification have\nbeen developed, starting with attention and patches used in transformers. While\nprior works have analyzed the influence of some aspects of architecture\ncomponents on the robustness to adversarial attacks, in particular for vision\ntransformers, the understanding of the main factors is still limited. We\ncompare several (non)-robust classifiers with different architectures and study\ntheir properties, including the effect of adversarial training on the\ninterpretability of the learnt features and robustness to unseen threat models.\nAn ablation from ResNet to ConvNeXt reveals key architectural changes leading\nto almost $10\\%$ higher $\\ell_\\infty$-robustness.\n", "date": "2022-09-14", "categories": []}
{"id": "2209.06971", "title": "PointACL:Adversarial Contrastive Learning for Robust Point Clouds\n  Representation under Adversarial Attack", "abstract": "  Despite recent success of self-supervised based contrastive learning model\nfor 3D point clouds representation, the adversarial robustness of such\npre-trained models raised concerns. Adversarial contrastive learning (ACL) is\nconsidered an effective way to improve the robustness of pre-trained models. In\ncontrastive learning, the projector is considered an effective component for\nremoving unnecessary feature information during contrastive pretraining and\nmost ACL works also use contrastive loss with projected feature representations\nto generate adversarial examples in pretraining, while \"unprojected \" feature\nrepresentations are used in generating adversarial inputs during\ninference.Because of the distribution gap between projected and \"unprojected\"\nfeatures, their models are constrained of obtaining robust feature\nrepresentations for downstream tasks. We introduce a new method to generate\nhigh-quality 3D adversarial examples for adversarial training by utilizing\nvirtual adversarial loss with \"unprojected\" feature representations in\ncontrastive learning framework. We present our robust aware loss function to\ntrain self-supervised contrastive learning framework adversarially.\nFurthermore, we find selecting high difference points with the Difference of\nNormal (DoN) operator as additional input for adversarial self-supervised\ncontrastive learning can significantly improve the adversarial robustness of\nthe pre-trained model. We validate our method, PointACL on downstream tasks,\nincluding 3D classification and 3D segmentation with multiple datasets. It\nobtains comparable robust accuracy over state-of-the-art contrastive\nadversarial learning methods.\n", "date": "2022-09-14", "categories": []}
{"id": "2209.07369", "title": "Adversarially Robust Learning: A Generic Minimax Optimal Learner and\n  Characterization", "abstract": "  We present a minimax optimal learner for the problem of learning predictors\nrobust to adversarial examples at test-time. Interestingly, we find that this\nrequires new algorithmic ideas and approaches to adversarially robust learning.\nIn particular, we show, in a strong negative sense, the suboptimality of the\nrobust learner proposed by Montasser, Hanneke, and Srebro (2019) and a broader\nfamily of learners we identify as local learners. Our results are enabled by\nadopting a global perspective, specifically, through a key technical\ncontribution: the global one-inclusion graph, which may be of independent\ninterest, that generalizes the classical one-inclusion graph due to Haussler,\nLittlestone, and Warmuth (1994). Finally, as a byproduct, we identify a\ndimension characterizing qualitatively and quantitatively what classes of\npredictors $\\mathcal{H}$ are robustly learnable. This resolves an open problem\ndue to Montasser et al. (2019), and closes a (potentially) infinite gap between\nthe established upper and lower bounds on the sample complexity of\nadversarially robust learning.\n", "date": "2022-09-15", "categories": []}
{"id": "2209.05692", "title": "Sample Complexity of an Adversarial Attack on UCB-based Best-arm\n  Identification Policy", "abstract": "  In this work I study the problem of adversarial perturbations to rewards, in\na Multi-armed bandit (MAB) setting. Specifically, I focus on an adversarial\nattack to a UCB type best-arm identification policy applied to a stochastic\nMAB. The UCB attack presented in [1] results in pulling a target arm K very\noften. I used the attack model of [1] to derive the sample complexity required\nfor selecting target arm K as the best arm. I have proved that the stopping\ncondition of UCB based best-arm identification algorithm given in [2], can be\nachieved by the target arm K in T rounds, where T depends only on the total\nnumber of arms and $\\sigma$ parameter of $\\sigma^2-$ sub-Gaussian random\nrewards of the arms.\n", "date": "2022-09-13", "categories": []}
{"id": "2209.04930", "title": "Resisting Deep Learning Models Against Adversarial Attack\n  Transferability via Feature Randomization", "abstract": "  In the past decades, the rise of artificial intelligence has given us the\ncapabilities to solve the most challenging problems in our day-to-day lives,\nsuch as cancer prediction and autonomous navigation. However, these\napplications might not be reliable if not secured against adversarial attacks.\nIn addition, recent works demonstrated that some adversarial examples are\ntransferable across different models. Therefore, it is crucial to avoid such\ntransferability via robust models that resist adversarial manipulations. In\nthis paper, we propose a feature randomization-based approach that resists\neight adversarial attacks targeting deep learning models in the testing phase.\nOur novel approach consists of changing the training strategy in the target\nnetwork classifier and selecting random feature samples. We consider the\nattacker with a Limited-Knowledge and Semi-Knowledge conditions to undertake\nthe most prevalent types of adversarial attacks. We evaluate the robustness of\nour approach using the well-known UNSW-NB15 datasets that include realistic and\nsynthetic attacks. Afterward, we demonstrate that our strategy outperforms the\nexisting state-of-the-art approach, such as the Most Powerful Attack, which\nconsists of fine-tuning the network model against specific adversarial attacks.\nFinally, our experimental results show that our methodology can secure the\ntarget network and resists adversarial attack transferability by over 60%.\n", "date": "2022-09-11", "categories": []}
{"id": "2209.05130", "title": "Semantic-Preserving Adversarial Code Comprehension", "abstract": "  Based on the tremendous success of pre-trained language models (PrLMs) for\nsource code comprehension tasks, current literature studies either ways to\nfurther improve the performance (generalization) of PrLMs, or their robustness\nagainst adversarial attacks. However, they have to compromise on the trade-off\nbetween the two aspects and none of them consider improving both sides in an\neffective and practical way. To fill this gap, we propose Semantic-Preserving\nAdversarial Code Embeddings (SPACE) to find the worst-case semantic-preserving\nattacks while forcing the model to predict the correct labels under these worst\ncases. Experiments and analysis demonstrate that SPACE can stay robust against\nstate-of-the-art attacks while boosting the performance of PrLMs for code.\n", "date": "2022-09-12", "categories": []}
{"id": "2209.05446", "title": "Boosting Robustness Verification of Semantic Feature Neighborhoods", "abstract": "  Deep neural networks have been shown to be vulnerable to adversarial attacks\nthat perturb inputs based on semantic features. Existing robustness analyzers\ncan reason about semantic feature neighborhoods to increase the networks'\nreliability. However, despite the significant progress in these techniques,\nthey still struggle to scale to deep networks and large neighborhoods. In this\nwork, we introduce VeeP, an active learning approach that splits the\nverification process into a series of smaller verification steps, each is\nsubmitted to an existing robustness analyzer. The key idea is to build on prior\nsteps to predict the next optimal step. The optimal step is predicted by\nestimating the certification velocity and sensitivity via parametric\nregression. We evaluate VeeP on MNIST, Fashion-MNIST, CIFAR-10 and ImageNet and\nshow that it can analyze neighborhoods of various features: brightness,\ncontrast, hue, saturation, and lightness. We show that, on average, given a 90\nminute timeout, VeeP verifies 96% of the maximally certifiable neighborhoods\nwithin 29 minutes, while existing splitting approaches verify, on average, 73%\nof the maximally certifiable neighborhoods within 58 minutes.\n", "date": "2022-09-12", "categories": []}
{"id": "2208.05514", "title": "Are Gradients on Graph Structure Reliable in Gray-box Attacks?", "abstract": "  Graph edge perturbations are dedicated to damaging the prediction of graph\nneural networks by modifying the graph structure. Previous gray-box attackers\nemploy gradients from the surrogate model to locate the vulnerable edges to\nperturb the graph structure. However, unreliability exists in gradients on\ngraph structures, which is rarely studied by previous works. In this paper, we\ndiscuss and analyze the errors caused by the unreliability of the structural\ngradients. These errors arise from rough gradient usage due to the discreteness\nof the graph structure and from the unreliability in the meta-gradient on the\ngraph structure. In order to address these problems, we propose a novel attack\nmodel with methods to reduce the errors inside the structural gradients. We\npropose edge discrete sampling to select the edge perturbations associated with\nhierarchical candidate selection to ensure computational efficiency. In\naddition, semantic invariance and momentum gradient ensemble are proposed to\naddress the gradient fluctuation on semantic-augmented graphs and the\ninstability of the surrogate model. Experiments are conducted in untargeted\ngray-box poisoning scenarios and demonstrate the improvement in the performance\nof our approach.\n", "date": "2022-08-07", "categories": []}
{"id": "2209.03463", "title": "Why So Toxic? Measuring and Triggering Toxic Behavior in Open-Domain\n  Chatbots", "abstract": "  Chatbots are used in many applications, e.g., automated agents, smart home\nassistants, interactive characters in online games, etc. Therefore, it is\ncrucial to ensure they do not behave in undesired manners, providing offensive\nor toxic responses to users. This is not a trivial task as state-of-the-art\nchatbot models are trained on large, public datasets openly collected from the\nInternet. This paper presents a first-of-its-kind, large-scale measurement of\ntoxicity in chatbots. We show that publicly available chatbots are prone to\nproviding toxic responses when fed toxic queries. Even more worryingly, some\nnon-toxic queries can trigger toxic responses too. We then set out to design\nand experiment with an attack, ToxicBuddy, which relies on fine-tuning GPT-2 to\ngenerate non-toxic queries that make chatbots respond in a toxic manner. Our\nextensive experimental evaluation demonstrates that our attack is effective\nagainst public chatbot models and outperforms manually-crafted malicious\nqueries proposed by previous work. We also evaluate three defense mechanisms\nagainst ToxicBuddy, showing that they either reduce the attack performance at\nthe cost of affecting the chatbot's utility or are only effective at mitigating\na portion of the attack. This highlights the need for more research from the\ncomputer security and online safety communities to ensure that chatbot models\ndo not hurt their users. Overall, we are confident that ToxicBuddy can be used\nas an auditing tool and that our work will pave the way toward designing more\neffective defenses for chatbot safety.\n", "date": "2022-09-07", "categories": []}
{"id": "2209.04293", "title": "Robust-by-Design Classification via Unitary-Gradient Neural Networks", "abstract": "  The use of neural networks in safety-critical systems requires safe and\nrobust models, due to the existence of adversarial attacks. Knowing the minimal\nadversarial perturbation of any input x, or, equivalently, knowing the distance\nof x from the classification boundary, allows evaluating the classification\nrobustness, providing certifiable predictions. Unfortunately, state-of-the-art\ntechniques for computing such a distance are computationally expensive and\nhence not suited for online applications. This work proposes a novel family of\nclassifiers, namely Signed Distance Classifiers (SDCs), that, from a\ntheoretical perspective, directly output the exact distance of x from the\nclassification boundary, rather than a probability score (e.g., SoftMax). SDCs\nrepresent a family of robust-by-design classifiers. To practically address the\ntheoretical requirements of a SDC, a novel network architecture named\nUnitary-Gradient Neural Network is presented. Experimental results show that\nthe proposed architecture approximates a signed distance classifier, hence\nallowing an online certifiable classification of x at the cost of a single\ninference.\n", "date": "2022-09-09", "categories": []}
{"id": "2206.06761", "title": "Exploring Adversarial Attacks and Defenses in Vision Transformers\n  trained with DINO", "abstract": "  This work conducts the first analysis on the robustness against adversarial\nattacks on self-supervised Vision Transformers trained using DINO. First, we\nevaluate whether features learned through self-supervision are more robust to\nadversarial attacks than those emerging from supervised learning. Then, we\npresent properties arising for attacks in the latent space. Finally, we\nevaluate whether three well-known defense strategies can increase adversarial\nrobustness in downstream tasks by only fine-tuning the classification head to\nprovide robustness even in view of limited compute resources. These defense\nstrategies are: Adversarial Training, Ensemble Adversarial Training and\nEnsemble of Specialized Networks.\n", "date": "2022-06-14", "categories": []}
{"id": "2209.03431", "title": "Physics-Guided Adversarial Machine Learning for Aircraft Systems\n  Simulation", "abstract": "  In the context of aircraft system performance assessment, deep learning\ntechnologies allow to quickly infer models from experimental measurements, with\nless detailed system knowledge than usually required by physics-based modeling.\nHowever, this inexpensive model development also comes with new challenges\nregarding model trustworthiness. This work presents a novel approach,\nphysics-guided adversarial machine learning (ML), that improves the confidence\nover the physics consistency of the model. The approach performs, first, a\nphysics-guided adversarial testing phase to search for test inputs revealing\nbehavioral system inconsistencies, while still falling within the range of\nforeseeable operational conditions. Then, it proceeds with physics-informed\nadversarial training to teach the model the system-related physics domain\nforeknowledge through iteratively reducing the unwanted output deviations on\nthe previously-uncovered counterexamples. Empirical evaluation on two aircraft\nsystem performance models shows the effectiveness of our adversarial ML\napproach in exposing physical inconsistencies of both models and in improving\ntheir propensity to be consistent with physics domain knowledge.\n", "date": "2022-09-07", "categories": []}
{"id": "2206.06257", "title": "Distributed Adversarial Training to Robustify Deep Neural Networks at\n  Scale", "abstract": "  Current deep neural networks (DNNs) are vulnerable to adversarial attacks,\nwhere adversarial perturbations to the inputs can change or manipulate\nclassification. To defend against such attacks, an effective and popular\napproach, known as adversarial training (AT), has been shown to mitigate the\nnegative impact of adversarial attacks by virtue of a min-max robust training\nmethod. While effective, it remains unclear whether it can successfully be\nadapted to the distributed learning context. The power of distributed\noptimization over multiple machines enables us to scale up robust training over\nlarge models and datasets. Spurred by that, we propose distributed adversarial\ntraining (DAT), a large-batch adversarial training framework implemented over\nmultiple machines. We show that DAT is general, which supports training over\nlabeled and unlabeled data, multiple types of attack generation methods, and\ngradient compression operations favored for distributed optimization.\nTheoretically, we provide, under standard conditions in the optimization\ntheory, the convergence rate of DAT to the first-order stationary points in\ngeneral non-convex settings. Empirically, we demonstrate that DAT either\nmatches or outperforms state-of-the-art robust accuracies and achieves a\ngraceful training speedup (e.g., on ResNet-50 under ImageNet). Codes are\navailable at https://github.com/dat-2022/dat.\n", "date": "2022-06-13", "categories": []}
{"id": "2209.02902", "title": "Defending Against Backdoor Attack on Graph Nerual Network by\n  Explainability", "abstract": "  Backdoor attack is a powerful attack algorithm to deep learning model.\nRecently, GNN's vulnerability to backdoor attack has been proved especially on\ngraph classification task. In this paper, we propose the first backdoor\ndetection and defense method on GNN. Most backdoor attack depends on injecting\nsmall but influential trigger to the clean sample. For graph data, current\nbackdoor attack focus on manipulating the graph structure to inject the\ntrigger. We find that there are apparent differences between benign samples and\nmalicious samples in some explanatory evaluation metrics, such as fidelity and\ninfidelity. After identifying the malicious sample, the explainability of the\nGNN model can help us capture the most significant subgraph which is probably\nthe trigger in a trojan graph. We use various dataset and different attack\nsettings to prove the effectiveness of our defense method. The attack success\nrate all turns out to decrease considerably.\n", "date": "2022-09-07", "categories": []}
{"id": "2209.02997", "title": "On the Transferability of Adversarial Examples between Encrypted Models", "abstract": "  Deep neural networks (DNNs) are well known to be vulnerable to adversarial\nexamples (AEs). In addition, AEs have adversarial transferability, namely, AEs\ngenerated for a source model fool other (target) models. In this paper, we\ninvestigate the transferability of models encrypted for adversarially robust\ndefense for the first time. To objectively verify the property of\ntransferability, the robustness of models is evaluated by using a benchmark\nattack method, called AutoAttack. In an image-classification experiment, the\nuse of encrypted models is confirmed not only to be robust against AEs but to\nalso reduce the influence of AEs in terms of the transferability of models.\n", "date": "2022-09-07", "categories": []}
{"id": "2207.01795", "title": "PatchZero: Defending against Adversarial Patch Attacks by Detecting and\n  Zeroing the Patch", "abstract": "  Adversarial patch attacks mislead neural networks by injecting adversarial\npixels within a local region. Patch attacks can be highly effective in a\nvariety of tasks and physically realizable via attachment (e.g. a sticker) to\nthe real-world objects. Despite the diversity in attack patterns, adversarial\npatches tend to be highly textured and different in appearance from natural\nimages. We exploit this property and present PatchZero, a general defense\npipeline against white-box adversarial patches without retraining the\ndownstream classifier or detector. Specifically, our defense detects\nadversaries at the pixel-level and \"zeros out\" the patch region by repainting\nwith mean pixel values. We further design a two-stage adversarial training\nscheme to defend against the stronger adaptive attacks. PatchZero achieves SOTA\ndefense performance on the image classification (ImageNet, RESISC45), object\ndetection (PASCAL VOC), and video classification (UCF101) tasks with little\ndegradation in benign performance. In addition, PatchZero transfers to\ndifferent patch shapes and attack types.\n", "date": "2022-07-05", "categories": []}
{"id": "2209.01721", "title": "An Adaptive Black-box Defense against Trojan Attacks (TrojDef)", "abstract": "  Trojan backdoor is a poisoning attack against Neural Network (NN) classifiers\nin which adversaries try to exploit the (highly desirable) model reuse property\nto implant Trojans into model parameters for backdoor breaches through a\npoisoned training process. Most of the proposed defenses against Trojan attacks\nassume a white-box setup, in which the defender either has access to the inner\nstate of NN or is able to run back-propagation through it. In this work, we\npropose a more practical black-box defense, dubbed TrojDef, which can only run\nforward-pass of the NN. TrojDef tries to identify and filter out Trojan inputs\n(i.e., inputs augmented with the Trojan trigger) by monitoring the changes in\nthe prediction confidence when the input is repeatedly perturbed by random\nnoise. We derive a function based on the prediction outputs which is called the\nprediction confidence bound to decide whether the input example is Trojan or\nnot. The intuition is that Trojan inputs are more stable as the\nmisclassification only depends on the trigger, while benign inputs will suffer\nwhen augmented with noise due to the perturbation of the classification\nfeatures.\n  Through mathematical analysis, we show that if the attacker is perfect in\ninjecting the backdoor, the Trojan infected model will be trained to learn the\nappropriate prediction confidence bound, which is used to distinguish Trojan\nand benign inputs under arbitrary perturbations. However, because the attacker\nmight not be perfect in injecting the backdoor, we introduce a nonlinear\ntransform to the prediction confidence bound to improve the detection accuracy\nin practical settings. Extensive empirical evaluations show that TrojDef\nsignificantly outperforms the-state-of-the-art defenses and is highly stable\nunder different settings, even when the classifier architecture, the training\nprocess, or the hyper-parameters change.\n", "date": "2022-09-05", "categories": []}
{"id": "2209.01882", "title": "PromptAttack: Prompt-based Attack for Language Models via Gradient\n  Search", "abstract": "  As the pre-trained language models (PLMs) continue to grow, so do the\nhardware and data requirements for fine-tuning PLMs. Therefore, the researchers\nhave come up with a lighter method called \\textit{Prompt Learning}. However,\nduring the investigations, we observe that the prompt learning methods are\nvulnerable and can easily be attacked by some illegally constructed prompts,\nresulting in classification errors, and serious security problems for PLMs.\nMost of the current research ignores the security issue of prompt-based\nmethods. Therefore, in this paper, we propose a malicious prompt template\nconstruction method (\\textbf{PromptAttack}) to probe the security performance\nof PLMs. Several unfriendly template construction approaches are investigated\nto guide the model to misclassify the task. Extensive experiments on three\ndatasets and three PLMs prove the effectiveness of our proposed approach\nPromptAttack. We also conduct experiments to verify that our method is\napplicable in few-shot scenarios.\n", "date": "2022-09-05", "categories": []}
{"id": "2209.02128", "title": "Evaluating the Susceptibility of Pre-Trained Language Models via\n  Handcrafted Adversarial Examples", "abstract": "  Recent advances in the development of large language models have resulted in\npublic access to state-of-the-art pre-trained language models (PLMs), including\nGenerative Pre-trained Transformer 3 (GPT-3) and Bidirectional Encoder\nRepresentations from Transformers (BERT). However, evaluations of PLMs, in\npractice, have shown their susceptibility to adversarial attacks during the\ntraining and fine-tuning stages of development. Such attacks can result in\nerroneous outputs, model-generated hate speech, and the exposure of users'\nsensitive information. While existing research has focused on adversarial\nattacks during either the training or the fine-tuning of PLMs, there is a\ndeficit of information on attacks made between these two development phases. In\nthis work, we highlight a major security vulnerability in the public release of\nGPT-3 and further investigate this vulnerability in other state-of-the-art\nPLMs. We restrict our work to pre-trained models that have not undergone\nfine-tuning. Further, we underscore token distance-minimized perturbations as\nan effective adversarial approach, bypassing both supervised and unsupervised\nquality measures. Following this approach, we observe a significant decrease in\ntext classification quality when evaluating for semantic similarity.\n", "date": "2022-09-05", "categories": []}
{"id": "2209.02406", "title": "Unrestricted Adversarial Samples Based on Non-semantic Feature Clusters\n  Substitution", "abstract": "  Most current methods generate adversarial examples with the $L_p$ norm\nspecification. As a result, many defense methods utilize this property to\neliminate the impact of such attacking algorithms. In this paper,we instead\nintroduce \"unrestricted\" perturbations that create adversarial samples by using\nspurious relations which were learned by model training. Specifically, we find\nfeature clusters in non-semantic features that are strongly correlated with\nmodel judgment results, and treat them as spurious relations learned by the\nmodel. Then we create adversarial samples by using them to replace the\ncorresponding feature clusters in the target image. Experimental evaluations\nshow that in both black-box and white-box situations. Our adversarial examples\ndo not change the semantics of images, while still being effective at fooling\nan adversarially trained DNN image classifier.\n", "date": "2022-08-31", "categories": []}
{"id": "2209.02408", "title": "Robustness and invariance properties of image classifiers", "abstract": "  Deep neural networks have achieved impressive results in many image\nclassification tasks. However, since their performance is usually measured in\ncontrolled settings, it is important to ensure that their decisions remain\ncorrect when deployed in noisy environments. In fact, deep networks are not\nrobust to a large variety of semantic-preserving image modifications, even to\nimperceptible image changes known as adversarial perturbations. The poor\nrobustness of image classifiers to small data distribution shifts raises\nserious concerns regarding their trustworthiness. To build reliable machine\nlearning models, we must design principled methods to analyze and understand\nthe mechanisms that shape robustness and invariance. This is exactly the focus\nof this thesis.\n  First, we study the problem of computing sparse adversarial perturbations. We\nexploit the geometry of the decision boundaries of image classifiers for\ncomputing sparse perturbations very fast, and reveal a qualitative connection\nbetween adversarial examples and the data features that image classifiers\nlearn. Then, to better understand this connection, we propose a geometric\nframework that connects the distance of data samples to the decision boundary,\nwith the features existing in the data. We show that deep classifiers have a\nstrong inductive bias towards invariance to non-discriminative features, and\nthat adversarial training exploits this property to confer robustness. Finally,\nwe focus on the challenging problem of generalization to unforeseen corruptions\nof the data, and we propose a novel data augmentation scheme for achieving\nstate-of-the-art robustness to common corruptions of the images.\n  Overall, our results contribute to the understanding of the fundamental\nmechanisms of deep image classifiers, and pave the way for building more\nreliable machine learning systems that can be deployed in real-world\nenvironments.\n", "date": "2022-08-30", "categories": []}
{"id": "2209.02453", "title": "Instance Attack:An Explanation-based Vulnerability Analysis Framework\n  Against DNNs for Malware Detection", "abstract": "  Deep neural networks (DNNs) are increasingly being applied in malware\ndetection and their robustness has been widely debated. Traditionally an\nadversarial example generation scheme relies on either detailed model\ninformation (gradient-based methods) or lots of samples to train a surrogate\nmodel, neither of which are available in most scenarios.\n  We propose the notion of the instance-based attack. Our scheme is\ninterpretable and can work in a black-box environment. Given a specific binary\nexample and a malware classifier, we use the data augmentation strategies to\nproduce enough data from which we can train a simple interpretable model. We\nexplain the detection model by displaying the weight of different parts of the\nspecific binary. By analyzing the explanations, we found that the data\nsubsections play an important role in Windows PE malware detection. We proposed\na new function preserving transformation algorithm that can be applied to data\nsubsections. By employing the binary-diversification techniques that we\nproposed, we eliminated the influence of the most weighted part to generate\nadversarial examples. Our algorithm can fool the DNNs in certain cases with a\nsuccess rate of nearly 100\\%. Our method outperforms the state-of-the-art\nmethod . The most important aspect is that our method operates in black-box\nsettings and the results can be validated with domain knowledge. Our analysis\nmodel can assist people in improving the robustness of malware detectors.\n", "date": "2022-09-06", "categories": []}
{"id": "2209.02684", "title": "Bag of Tricks for FGSM Adversarial Training", "abstract": "  Adversarial training (AT) with samples generated by Fast Gradient Sign Method\n(FGSM), also known as FGSM-AT, is a computationally simple method to train\nrobust networks. However, during its training procedure, an unstable mode of\n\"catastrophic overfitting\" has been identified in arXiv:2001.03994 [cs.LG],\nwhere the robust accuracy abruptly drops to zero within a single training step.\nExisting methods use gradient regularizers or random initialization tricks to\nattenuate this issue, whereas they either take high computational cost or lead\nto lower robust accuracy. In this work, we provide the first study, which\nthoroughly examines a collection of tricks from three perspectives: Data\nInitialization, Network Structure, and Optimization, to overcome the\ncatastrophic overfitting in FGSM-AT.\n  Surprisingly, we find that simple tricks, i.e., a) masking partial pixels\n(even without randomness), b) setting a large convolution stride and smooth\nactivation functions, or c) regularizing the weights of the first convolutional\nlayer, can effectively tackle the overfitting issue. Extensive results on a\nrange of network architectures validate the effectiveness of each proposed\ntrick, and the combinations of tricks are also investigated. For example,\ntrained with PreActResNet-18 on CIFAR-10, our method attains 49.8% accuracy\nagainst PGD-50 attacker and 46.4% accuracy against AutoAttack, demonstrating\nthat pure FGSM-AT is capable of enabling robust learners. The code and models\nare publicly available at\nhttps://github.com/UCSC-VLAA/Bag-of-Tricks-for-FGSM-AT.\n", "date": "2022-09-06", "categories": []}
{"id": "2209.00757", "title": "Universal Fourier Attack for Time Series", "abstract": "  A wide variety of adversarial attacks have been proposed and explored using\nimage and audio data. These attacks are notoriously easy to generate digitally\nwhen the attacker can directly manipulate the input to a model, but are much\nmore difficult to implement in the real-world. In this paper we present a\nuniversal, time invariant attack for general time series data such that the\nattack has a frequency spectrum primarily composed of the frequencies present\nin the original data. The universality of the attack makes it fast and easy to\nimplement as no computation is required to add it to an input, while time\ninvariance is useful for real-world deployment. Additionally, the frequency\nconstraint ensures the attack can withstand filtering. We demonstrate the\neffectiveness of the attack in two different domains, speech recognition and\nunintended radiated emission, and show that the attack is robust against common\ntransform-and-compare defense pipelines.\n", "date": "2022-09-02", "categories": []}
{"id": "2209.00005", "title": "Be Your Own Neighborhood: Detecting Adversarial Example by the\n  Neighborhood Relations Built on Self-Supervised Learning", "abstract": "  Deep Neural Networks (DNNs) have achieved excellent performance in various\nfields. However, DNNs' vulnerability to Adversarial Examples (AE) hinders their\ndeployments to safety-critical applications. This paper presents a novel AE\ndetection framework, named BEYOND, for trustworthy predictions. BEYOND performs\nthe detection by distinguishing the AE's abnormal relation with its augmented\nversions, i.e. neighbors, from two prospects: representation similarity and\nlabel consistency. An off-the-shelf Self-Supervised Learning (SSL) model is\nused to extract the representation and predict the label for its highly\ninformative representation capacity compared to supervised learning models. For\nclean samples, their representations and predictions are closely consistent\nwith their neighbors, whereas those of AEs differ greatly. Furthermore, we\nexplain this observation and show that by leveraging this discrepancy BEYOND\ncan effectively detect AEs. We develop a rigorous justification for the\neffectiveness of BEYOND. Furthermore, as a plug-and-play model, BEYOND can\neasily cooperate with the Adversarial Trained Classifier (ATC), achieving the\nstate-of-the-art (SOTA) robustness accuracy. Experimental results show that\nBEYOND outperforms baselines by a large margin, especially under adaptive\nattacks. Empowered by the robust relation net built on SSL, we found that\nBEYOND outperforms baselines in terms of both detection ability and speed. Our\ncode will be publicly available.\n", "date": "2022-08-31", "categories": []}
{"id": "2206.15128", "title": "Detecting and Recovering Adversarial Examples from Extracting Non-robust\n  and Highly Predictive Adversarial Perturbations", "abstract": "  Deep neural networks (DNNs) have been shown to be vulnerable against\nadversarial examples (AEs) which are maliciously designed to fool target\nmodels. The normal examples (NEs) added with imperceptible adversarial\nperturbation, can be a security threat to DNNs. Although the existing AEs\ndetection methods have achieved a high accuracy, they failed to exploit the\ninformation of the AEs detected. Thus, based on high-dimension perturbation\nextraction, we propose a model-free AEs detection method, the whole process of\nwhich is free from querying the victim model. Research shows that DNNs are\nsensitive to the high-dimension features. The adversarial perturbation hiding\nin the adversarial example belongs to the high-dimension feature which is\nhighly predictive and non-robust. DNNs learn more details from high-dimension\ndata than others. In our method, the perturbation extractor can extract the\nadversarial perturbation from AEs as high-dimension feature, then the trained\nAEs discriminator determines whether the input is an AE. Experimental results\nshow that the proposed method can not only detect the adversarial examples with\nhigh accuracy, but also detect the specific category of the AEs. Meanwhile, the\nextracted perturbation can be used to recover the AEs to NEs.\n", "date": "2022-06-30", "categories": []}
{"id": "2208.13838", "title": "Towards Adversarial Purification using Denoising AutoEncoders", "abstract": "  With the rapid advancement and increased use of deep learning models in image\nidentification, security becomes a major concern to their deployment in\nsafety-critical systems. Since the accuracy and robustness of deep learning\nmodels are primarily attributed from the purity of the training samples,\ntherefore the deep learning architectures are often susceptible to adversarial\nattacks. Adversarial attacks are often obtained by making subtle perturbations\nto normal images, which are mostly imperceptible to humans, but can seriously\nconfuse the state-of-the-art machine learning models. We propose a framework,\nnamed APuDAE, leveraging Denoising AutoEncoders (DAEs) to purify these samples\nby using them in an adaptive way and thus improve the classification accuracy\nof the target classifier networks that have been attacked. We also show how\nusing DAEs adaptively instead of using them directly, improves classification\naccuracy further and is more robust to the possibility of designing adaptive\nattacks to fool them. We demonstrate our results over MNIST, CIFAR-10, ImageNet\ndataset and show how our framework (APuDAE) provides comparable and in most\ncases better performance to the baseline methods in purifying adversaries. We\nalso design adaptive attack specifically designed to attack our purifying model\nand demonstrate how our defense is robust to that.\n", "date": "2022-08-29", "categories": []}
{"id": "2208.14191", "title": "FuncFooler: A Practical Black-box Attack Against Learning-based Binary\n  Code Similarity Detection Methods", "abstract": "  The binary code similarity detection (BCSD) method measures the similarity of\ntwo binary executable codes. Recently, the learning-based BCSD methods have\nachieved great success, outperforming traditional BCSD in detection accuracy\nand efficiency. However, the existing studies are rather sparse on the\nadversarial vulnerability of the learning-based BCSD methods, which cause\nhazards in security-related applications. To evaluate the adversarial\nrobustness, this paper designs an efficient and black-box adversarial code\ngeneration algorithm, namely, FuncFooler. FuncFooler constrains the adversarial\ncodes 1) to keep unchanged the program's control flow graph (CFG), and 2) to\npreserve the same semantic meaning. Specifically, FuncFooler consecutively 1)\ndetermines vulnerable candidates in the malicious code, 2) chooses and inserts\nthe adversarial instructions from the benign code, and 3) corrects the semantic\nside effect of the adversarial code to meet the constraints. Empirically, our\nFuncFooler can successfully attack the three learning-based BCSD models,\nincluding SAFE, Asm2Vec, and jTrans, which calls into question whether the\nlearning-based BCSD is desirable.\n", "date": "2022-08-26", "categories": []}
{"id": "2208.14302", "title": "A Black-Box Attack on Optical Character Recognition Systems", "abstract": "  Adversarial machine learning is an emerging area showing the vulnerability of\ndeep learning models. Exploring attack methods to challenge state of the art\nartificial intelligence (A.I.) models is an area of critical concern. The\nreliability and robustness of such A.I. models are one of the major concerns\nwith an increasing number of effective adversarial attack methods.\nClassification tasks are a major vulnerable area for adversarial attacks. The\nmajority of attack strategies are developed for colored or gray-scaled images.\nConsequently, adversarial attacks on binary image recognition systems have not\nbeen sufficiently studied. Binary images are simple two possible pixel-valued\nsignals with a single channel. The simplicity of binary images has a\nsignificant advantage compared to colored and gray scaled images, namely\ncomputation efficiency. Moreover, most optical character recognition systems\n(O.C.R.s), such as handwritten character recognition, plate number\nidentification, and bank check recognition systems, use binary images or\nbinarization in their processing steps. In this paper, we propose a simple yet\nefficient attack method, Efficient Combinatorial Black-box Adversarial Attack,\non binary image classifiers. We validate the efficiency of the attack technique\non two different data sets and three classification networks, demonstrating its\nperformance. Furthermore, we compare our proposed method with state-of-the-art\nmethods regarding advantages and disadvantages as well as applicability.\n", "date": "2022-08-30", "categories": []}
{"id": "2208.12911", "title": "Network-Level Adversaries in Federated Learning", "abstract": "  Federated learning is a popular strategy for training models on distributed,\nsensitive data, while preserving data privacy. Prior work identified a range of\nsecurity threats on federated learning protocols that poison the data or the\nmodel. However, federated learning is a networked system where the\ncommunication between clients and server plays a critical role for the learning\ntask performance. We highlight how communication introduces another\nvulnerability surface in federated learning and study the impact of\nnetwork-level adversaries on training federated learning models. We show that\nattackers dropping the network traffic from carefully selected clients can\nsignificantly decrease model accuracy on a target population. Moreover, we show\nthat a coordinated poisoning campaign from a few clients can amplify the\ndropping attacks. Finally, we develop a server-side defense which mitigates the\nimpact of our attacks by identifying and up-sampling clients likely to\npositively contribute towards target accuracy. We comprehensively evaluate our\nattacks and defenses on three datasets, assuming encrypted communication\nchannels and attackers with partial visibility of the network.\n", "date": "2022-08-27", "categories": []}
{"id": "2208.13182", "title": "Cross-domain Cross-architecture Black-box Attacks on Fine-tuned Models\n  with Transferred Evolutionary Strategies", "abstract": "  Fine-tuning can be vulnerable to adversarial attacks. Existing works about\nblack-box attacks on fine-tuned models (BAFT) are limited by strong\nassumptions. To fill the gap, we propose two novel BAFT settings, cross-domain\nand cross-domain cross-architecture BAFT, which only assume that (1) the target\nmodel for attacking is a fine-tuned model, and (2) the source domain data is\nknown and accessible. To successfully attack fine-tuned models under both\nsettings, we propose to first train an adversarial generator against the source\nmodel, which adopts an encoder-decoder architecture and maps a clean input to\nan adversarial example. Then we search in the low-dimensional latent space\nproduced by the encoder of the adversarial generator. The search is conducted\nunder the guidance of the surrogate gradient obtained from the source model.\nExperimental results on different domains and different network architectures\ndemonstrate that the proposed attack method can effectively and efficiently\nattack the fine-tuned models.\n", "date": "2022-08-28", "categories": []}
{"id": "2208.02430", "title": "A New Kind of Adversarial Example", "abstract": "  Almost all adversarial attacks are formulated to add an imperceptible\nperturbation to an image in order to fool a model. Here, we consider the\nopposite which is adversarial examples that can fool a human but not a model. A\nlarge enough and perceptible perturbation is added to an image such that a\nmodel maintains its original decision, whereas a human will most likely make a\nmistake if forced to decide (or opt not to decide at all). Existing targeted\nattacks can be reformulated to synthesize such adversarial examples. Our\nproposed attack, dubbed NKE, is similar in essence to the fooling images, but\nis more efficient since it uses gradient descent instead of evolutionary\nalgorithms. It also offers a new and unified perspective into the problem of\nadversarial vulnerability. Experimental results over MNIST and CIFAR-10\ndatasets show that our attack is quite efficient in fooling deep neural\nnetworks. Code is available at https://github.com/aliborji/NKE.\n", "date": "2022-08-04", "categories": []}
{"id": "2208.11839", "title": "A Perturbation Resistant Transformation and Classification System for\n  Deep Neural Networks", "abstract": "  Deep convolutional neural networks accurately classify a diverse range of\nnatural images, but may be easily deceived when designed, imperceptible\nperturbations are embedded in the images. In this paper, we design a\nmulti-pronged training, input transformation, and image ensemble system that is\nattack agnostic and not easily estimated. Our system incorporates two novel\nfeatures. The first is a transformation layer that computes feature level\npolynomial kernels from class-level training data samples and iteratively\nupdates input image copies at inference time based on their feature kernel\ndifferences to create an ensemble of transformed inputs. The second is a\nclassification system that incorporates the prediction of the undefended\nnetwork with a hard vote on the ensemble of filtered images. Our evaluations on\nthe CIFAR10 dataset show our system improves the robustness of an undefended\nnetwork against a variety of bounded and unbounded white-box attacks under\ndifferent distance metrics, while sacrificing little accuracy on clean images.\nAgainst adaptive full-knowledge attackers creating end-to-end attacks, our\nsystem successfully augments the existing robustness of adversarially trained\nnetworks, for which our methods are most effectively applied.\n", "date": "2022-08-25", "categories": []}
{"id": "2208.12230", "title": "Semantic Preserving Adversarial Attack Generation with Autoencoder and\n  Genetic Algorithm", "abstract": "  Widely used deep learning models are found to have poor robustness. Little\nnoises can fool state-of-the-art models into making incorrect predictions.\nWhile there is a great deal of high-performance attack generation methods, most\nof them directly add perturbations to original data and measure them using L_p\nnorms; this can break the major structure of data, thus, creating invalid\nattacks. In this paper, we propose a black-box attack, which, instead of\nmodifying original data, modifies latent features of data extracted by an\nautoencoder; then, we measure noises in semantic space to protect the semantics\nof data. We trained autoencoders on MNIST and CIFAR-10 datasets and found\noptimal adversarial perturbations using a genetic algorithm. Our approach\nachieved a 100% attack success rate on the first 100 data of MNIST and CIFAR-10\ndatasets with less perturbation than FGSM.\n", "date": "2022-08-25", "categories": []}
{"id": "2208.11264", "title": "Towards an Awareness of Time Series Anomaly Detection Models'\n  Adversarial Vulnerability", "abstract": "  Time series anomaly detection is extensively studied in statistics,\neconomics, and computer science. Over the years, numerous methods have been\nproposed for time series anomaly detection using deep learning-based methods.\nMany of these methods demonstrate state-of-the-art performance on benchmark\ndatasets, giving the false impression that these systems are robust and\ndeployable in many practical and industrial real-world scenarios. In this\npaper, we demonstrate that the performance of state-of-the-art anomaly\ndetection methods is degraded substantially by adding only small adversarial\nperturbations to the sensor data. We use different scoring metrics such as\nprediction errors, anomaly, and classification scores over several public and\nprivate datasets ranging from aerospace applications, server machines, to\ncyber-physical systems in power plants. Under well-known adversarial attacks\nfrom Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD)\nmethods, we demonstrate that state-of-the-art deep neural networks (DNNs) and\ngraph neural networks (GNNs) methods, which claim to be robust against\nanomalies and have been possibly integrated in real-life systems, have their\nperformance drop to as low as 0%. To the best of our understanding, we\ndemonstrate, for the first time, the vulnerabilities of anomaly detection\nsystems against adversarial attacks. The overarching goal of this research is\nto raise awareness towards the adversarial vulnerabilities of time series\nanomaly detectors.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.11436", "title": "Trace and Detect Adversarial Attacks on CNNs using Feature Response Maps", "abstract": "  The existence of adversarial attacks on convolutional neural networks (CNN)\nquestions the fitness of such models for serious applications. The attacks\nmanipulate an input image such that misclassification is evoked while still\nlooking normal to a human observer -- they are thus not easily detectable. In a\ndifferent context, backpropagated activations of CNN hidden layers -- \"feature\nresponses\" to a given input -- have been helpful to visualize for a human\n\"debugger\" what the CNN \"looks at\" while computing its output. In this work, we\npropose a novel detection method for adversarial examples to prevent attacks.\nWe do so by tracking adversarial perturbations in feature responses, allowing\nfor automatic detection using average local spatial entropy. The method does\nnot alter the original network architecture and is fully human-interpretable.\nExperiments confirm the validity of our approach for state-of-the-art attacks\non large-scale models trained on ImageNet.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.11613", "title": "Unrestricted Black-box Adversarial Attack Using GAN with Limited Queries", "abstract": "  Adversarial examples are inputs intentionally generated for fooling a deep\nneural network. Recent studies have proposed unrestricted adversarial attacks\nthat are not norm-constrained. However, the previous unrestricted attack\nmethods still have limitations to fool real-world applications in a black-box\nsetting. In this paper, we present a novel method for generating unrestricted\nadversarial examples using GAN where an attacker can only access the top-1\nfinal decision of a classification model. Our method, Latent-HSJA, efficiently\nleverages the advantages of a decision-based attack in the latent space and\nsuccessfully manipulates the latent vectors for fooling the classification\nmodel.\n  With extensive experiments, we demonstrate that our proposed method is\nefficient in evaluating the robustness of classification models with limited\nqueries in a black-box setting. First, we demonstrate that our targeted attack\nmethod is query-efficient to produce unrestricted adversarial examples for a\nfacial identity recognition model that contains 307 identities. Then, we\ndemonstrate that the proposed method can also successfully attack a real-world\ncelebrity recognition service.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.10688", "title": "Hierarchical Perceptual Noise Injection for Social Media Fingerprint\n  Privacy Protection", "abstract": "  Billions of people are sharing their daily life images on social media every\nday. However, their biometric information (e.g., fingerprint) could be easily\nstolen from these images. The threat of fingerprint leakage from social media\nraises a strong desire for anonymizing shared images while maintaining image\nqualities, since fingerprints act as a lifelong individual biometric password.\nTo guard the fingerprint leakage, adversarial attack emerges as a solution by\nadding imperceptible perturbations on images. However, existing works are\neither weak in black-box transferability or appear unnatural. Motivated by\nvisual perception hierarchy (i.e., high-level perception exploits model-shared\nsemantics that transfer well across models while low-level perception extracts\nprimitive stimulus and will cause high visual sensitivities given suspicious\nstimulus), we propose FingerSafe, a hierarchical perceptual protective noise\ninjection framework to address the mentioned problems. For black-box\ntransferability, we inject protective noises on fingerprint orientation field\nto perturb the model-shared high-level semantics (i.e., fingerprint ridges).\nConsidering visual naturalness, we suppress the low-level local contrast\nstimulus by regularizing the response of Lateral Geniculate Nucleus. Our\nFingerSafe is the first to provide feasible fingerprint protection in both\ndigital (up to 94.12%) and realistic scenarios (Twitter and Facebook, up to\n68.75%). Our code can be found at\nhttps://github.com/nlsde-safety-team/FingerSafe.\n", "date": "2022-08-23", "categories": []}
{"id": "2208.10773", "title": "Adversarial Vulnerability of Temporal Feature Networks for Object\n  Detection", "abstract": "  Taking into account information across the temporal domain helps to improve\nenvironment perception in autonomous driving. However, it has not been studied\nso far whether temporally fused neural networks are vulnerable to deliberately\ngenerated perturbations, i.e. adversarial attacks, or whether temporal history\nis an inherent defense against them. In this work, we study whether temporal\nfeature networks for object detection are vulnerable to universal adversarial\nattacks. We evaluate attacks of two types: imperceptible noise for the whole\nimage and locally-bound adversarial patch. In both cases, perturbations are\ngenerated in a white-box manner using PGD. Our experiments confirm, that\nattacking even a portion of a temporal input suffices to fool the network. We\nvisually assess generated perturbations to gain insights into the functioning\nof attacks. To enhance the robustness, we apply adversarial training using\n5-PGD. Our experiments on KITTI and nuScenes datasets demonstrate, that a model\nrobustified via K-PGD is able to withstand the studied attacks while keeping\nthe mAP-based performance comparable to that of an unattacked model.\n", "date": "2022-08-23", "categories": []}
{"id": "2207.10561", "title": "Careful What You Wish For: on the Extraction of Adversarially Trained\n  Models", "abstract": "  Recent attacks on Machine Learning (ML) models such as evasion attacks with\nadversarial examples and models stealing through extraction attacks pose\nseveral security and privacy threats. Prior work proposes to use adversarial\ntraining to secure models from adversarial examples that can evade the\nclassification of a model and deteriorate its performance. However, this\nprotection technique affects the model's decision boundary and its prediction\nprobabilities, hence it might raise model privacy risks. In fact, a malicious\nuser using only a query access to the prediction output of a model can extract\nit and obtain a high-accuracy and high-fidelity surrogate model. To have a\ngreater extraction, these attacks leverage the prediction probabilities of the\nvictim model. Indeed, all previous work on extraction attacks do not take into\nconsideration the changes in the training process for security purposes. In\nthis paper, we propose a framework to assess extraction attacks on\nadversarially trained models with vision datasets. To the best of our\nknowledge, our work is the first to perform such evaluation. Through an\nextensive empirical study, we demonstrate that adversarially trained models are\nmore vulnerable to extraction attacks than models obtained under natural\ntraining circumstances. They can achieve up to $\\times1.2$ higher accuracy and\nagreement with a fraction lower than $\\times0.75$ of the queries. We\nadditionally find that the adversarial robustness capability is transferable\nthrough extraction attacks, i.e., extracted Deep Neural Networks (DNNs) from\nrobust models show an enhanced accuracy to adversarial examples compared to\nextracted DNNs from naturally trained (i.e. standard) models.\n", "date": "2022-07-21", "categories": []}
{"id": "2208.09710", "title": "Adversarial contamination of networks in the setting of vertex\n  nomination: a new trimming method", "abstract": "  As graph data becomes more ubiquitous, the need for robust inferential graph\nalgorithms to operate in these complex data domains is crucial. In many cases\nof interest, inference is further complicated by the presence of adversarial\ndata contamination. The effect of the adversary is frequently to change the\ndata distribution in ways that negatively affect statistical and algorithmic\nperformance. We study this phenomenon in the context of vertex nomination, a\nsemi-supervised information retrieval task for network data. Here, a common\nsuite of methods relies on spectral graph embeddings, which have been shown to\nprovide both good algorithmic performance and flexible settings in which\nregularization techniques can be implemented to help mitigate the effect of an\nadversary. Many current regularization methods rely on direct network trimming\nto effectively excise the adversarial contamination, although this direct\ntrimming often gives rise to complicated dependency structures in the resulting\ngraph. We propose a new trimming method that operates in model space which can\naddress both block structure contamination and white noise contamination\n(contamination whose distribution is unknown). This model trimming is more\namenable to theoretical analysis while also demonstrating superior performance\nin a number of simulations, compared to direct trimming.\n", "date": "2022-08-20", "categories": []}
{"id": "2208.10251", "title": "Rethinking Textual Adversarial Defense for Pre-trained Language Models", "abstract": "  Although pre-trained language models (PrLMs) have achieved significant\nsuccess, recent studies demonstrate that PrLMs are vulnerable to adversarial\nattacks. By generating adversarial examples with slight perturbations on\ndifferent levels (sentence / word / character), adversarial attacks can fool\nPrLMs to generate incorrect predictions, which questions the robustness of\nPrLMs. However, we find that most existing textual adversarial examples are\nunnatural, which can be easily distinguished by both human and machine. Based\non a general anomaly detector, we propose a novel metric (Degree of Anomaly) as\na constraint to enable current adversarial attack approaches to generate more\nnatural and imperceptible adversarial examples. Under this new constraint, the\nsuccess rate of existing attacks drastically decreases, which reveals that the\nrobustness of PrLMs is not as fragile as they claimed. In addition, we find\nthat four types of randomization can invalidate a large portion of textual\nadversarial examples. Based on anomaly detector and randomization, we design a\nuniversal defense framework, which is among the first to perform textual\nadversarial defense without knowing the specific attack. Empirical results show\nthat our universal defense framework achieves comparable or even higher\nafter-attack accuracy with other specific defenses, while preserving higher\noriginal accuracy at the same time. Our work discloses the essence of textual\nadversarial attacks, and indicates that (1) further works of adversarial\nattacks should focus more on how to overcome the detection and resist the\nrandomization, otherwise their adversarial examples would be easily detected\nand invalidated; and (2) compared with the unnatural and perceptible\nadversarial examples, it is those undetectable adversarial examples that pose\nreal risks for PrLMs and require more attention for future robustness-enhancing\nstrategies.\n", "date": "2022-07-21", "categories": []}
{"id": "2208.10481", "title": "BARReL: Bottleneck Attention for Adversarial Robustness in Vision-Based\n  Reinforcement Learning", "abstract": "  Robustness to adversarial perturbations has been explored in many areas of\ncomputer vision. This robustness is particularly relevant in vision-based\nreinforcement learning, as the actions of autonomous agents might be\nsafety-critic or impactful in the real world. We investigate the susceptibility\nof vision-based reinforcement learning agents to gradient-based adversarial\nattacks and evaluate a potential defense. We observe that Bottleneck Attention\nModules (BAM) included in CNN architectures can act as potential tools to\nincrease robustness against adversarial attacks. We show how learned attention\nmaps can be used to recover activations of a convolutional layer by restricting\nthe spatial activations to salient regions. Across a number of RL environments,\nBAM-enhanced architectures show increased robustness during inference. Finally,\nwe discuss potential future research directions.\n", "date": "2022-08-22", "categories": []}
{"id": "2207.05327", "title": "Certified Adversarial Robustness via Anisotropic Randomized Smoothing", "abstract": "  Randomized smoothing has achieved great success for certified robustness\nagainst adversarial perturbations. Given any arbitrary classifier, randomized\nsmoothing can guarantee the classifier's prediction over the perturbed input\nwith provable robustness bound by injecting noise into the classifier. However,\nall of the existing methods rely on fixed i.i.d. probability distribution to\ngenerate noise for all dimensions of the data (e.g., all the pixels in an\nimage), which ignores the heterogeneity of inputs and data dimensions. Thus,\nexisting randomized smoothing methods cannot provide optimal protection for all\nthe inputs. To address this limitation, we propose a novel anisotropic\nrandomized smoothing method which ensures provable robustness guarantee based\non pixel-wise noise distributions. Also, we design a novel CNN-based noise\ngenerator to efficiently fine-tune the pixel-wise noise distributions for all\nthe pixels in each input. Experimental results demonstrate that our method\nsignificantly outperforms the state-of-the-art randomized smoothing methods.\n", "date": "2022-07-12", "categories": []}
{"id": "2208.09195", "title": "Real-Time Robust Video Object Detection System Against Physical-World\n  Adversarial Attacks", "abstract": "  DNN-based video object detection (VOD) powers autonomous driving and video\nsurveillance industries with rising importance and promising opportunities.\nHowever, adversarial patch attack yields huge concern in live vision tasks\nbecause of its practicality, feasibility, and powerful attack effectiveness.\nThis work proposes Themis, a software/hardware system to defend against\nadversarial patches for real-time robust video object detection. We observe\nthat adversarial patches exhibit extremely localized superficial feature\nimportance in a small region with non-robust predictions, and thus propose the\nadversarial region detection algorithm for adversarial effect elimination.\nThemis also proposes a systematic design to efficiently support the algorithm\nby eliminating redundant computations and memory traffics. Experimental results\nshow that the proposed methodology can effectively recover the system from the\nadversarial attack with negligible hardware overhead.\n", "date": "2022-08-19", "categories": []}
{"id": "2208.09336", "title": "Dispersed Pixel Perturbation-based Imperceptible Backdoor Trigger for\n  Image Classifier Models", "abstract": "  Typical deep neural network (DNN) backdoor attacks are based on triggers\nembedded in inputs. Existing imperceptible triggers are computationally\nexpensive or low in attack success. In this paper, we propose a new backdoor\ntrigger, which is easy to generate, imperceptible, and highly effective. The\nnew trigger is a uniformly randomly generated three-dimensional (3D) binary\npattern that can be horizontally and/or vertically repeated and mirrored and\nsuperposed onto three-channel images for training a backdoored DNN model.\nDispersed throughout an image, the new trigger produces weak perturbation to\nindividual pixels, but collectively holds a strong recognizable pattern to\ntrain and activate the backdoor of the DNN. We also analytically reveal that\nthe trigger is increasingly effective with the improving resolution of the\nimages. Experiments are conducted using the ResNet-18 and MLP models on the\nMNIST, CIFAR-10, and BTSR datasets. In terms of imperceptibility, the new\ntrigger outperforms existing triggers, such as BadNets, Trojaned NN, and Hidden\nBackdoor, by over an order of magnitude. The new trigger achieves an almost\n100% attack success rate, only reduces the classification accuracy by less than\n0.7%-2.4%, and invalidates the state-of-the-art defense techniques.\n", "date": "2022-08-19", "categories": []}
{"id": "2208.09466", "title": "Gender Bias and Universal Substitution Adversarial Attacks on\n  Grammatical Error Correction Systems for Automated Assessment", "abstract": "  Grammatical Error Correction (GEC) systems perform a sequence-to-sequence\ntask, where an input word sequence containing grammatical errors, is corrected\nfor these errors by the GEC system to output a grammatically correct word\nsequence. With the advent of deep learning methods, automated GEC systems have\nbecome increasingly popular. For example, GEC systems are often used on speech\ntranscriptions of English learners as a form of assessment and feedback - these\npowerful GEC systems can be used to automatically measure an aspect of a\ncandidate's fluency. The count of \\textit{edits} from a candidate's input\nsentence (or essay) to a GEC system's grammatically corrected output sentence\nis indicative of a candidate's language ability, where fewer edits suggest\nbetter fluency. The count of edits can thus be viewed as a \\textit{fluency\nscore} with zero implying perfect fluency. However, although deep learning\nbased GEC systems are extremely powerful and accurate, they are susceptible to\nadversarial attacks: an adversary can introduce a small, specific change at the\ninput of a system that causes a large, undesired change at the output. When\nconsidering the application of GEC systems to automated language assessment,\nthe aim of an adversary could be to cheat by making a small change to a\ngrammatically incorrect input sentence that conceals the errors from a GEC\nsystem, such that no edits are found and the candidate is unjustly awarded a\nperfect fluency score. This work examines a simple universal substitution\nadversarial attack that non-native speakers of English could realistically\nemploy to deceive GEC systems used for assessment.\n", "date": "2022-08-19", "categories": []}
{"id": "2208.08697", "title": "Resisting Adversarial Attacks in Deep Neural Networks using Diverse\n  Decision Boundaries", "abstract": "  The security of deep learning (DL) systems is an extremely important field of\nstudy as they are being deployed in several applications due to their\never-improving performance to solve challenging tasks. Despite overwhelming\npromises, the deep learning systems are vulnerable to crafted adversarial\nexamples, which may be imperceptible to the human eye, but can lead the model\nto misclassify. Protections against adversarial perturbations on ensemble-based\ntechniques have either been shown to be vulnerable to stronger adversaries or\nshown to lack an end-to-end evaluation. In this paper, we attempt to develop a\nnew ensemble-based solution that constructs defender models with diverse\ndecision boundaries with respect to the original model. The ensemble of\nclassifiers constructed by (1) transformation of the input by a method called\nSplit-and-Shuffle, and (2) restricting the significant features by a method\ncalled Contrast-Significant-Features are shown to result in diverse gradients\nwith respect to adversarial attacks, which reduces the chance of transferring\nadversarial examples from the original to the defender model targeting the same\nclass. We present extensive experimentations using standard image\nclassification datasets, namely MNIST, CIFAR-10 and CIFAR-100 against\nstate-of-the-art adversarial attacks to demonstrate the robustness of the\nproposed ensemble-based defense. We also evaluate the robustness in the\npresence of a stronger adversary targeting all the models within the ensemble\nsimultaneously. Results for the overall false positives and false negatives\nhave been furnished to estimate the overall performance of the proposed\nmethodology.\n", "date": "2022-08-18", "categories": []}
{"id": "2208.08029", "title": "A Context-Aware Approach for Textual Adversarial Attack through\n  Probability Difference Guided Beam Search", "abstract": "  Textual adversarial attacks expose the vulnerabilities of text classifiers\nand can be used to improve their robustness. Existing context-aware methods\nsolely consider the gold label probability and use the greedy search when\nsearching an attack path, often limiting the attack efficiency. To tackle these\nissues, we propose PDBS, a context-aware textual adversarial attack model using\nProbability Difference guided Beam Search. The probability difference is an\noverall consideration of all class label probabilities, and PDBS uses it to\nguide the selection of attack paths. In addition, PDBS uses the beam search to\nfind a successful attack path, thus avoiding suffering from limited search\nspace. Extensive experiments and human evaluation demonstrate that PDBS\noutperforms previous best models in a series of evaluation metrics, especially\nbringing up to a +19.5% attack success rate. Ablation studies and qualitative\nanalyses further confirm the efficiency of PDBS.\n", "date": "2022-08-17", "categories": []}
{"id": "2208.08433", "title": "Label Flipping Data Poisoning Attack Against Wearable Human Activity\n  Recognition System", "abstract": "  Human Activity Recognition (HAR) is a problem of interpreting sensor data to\nhuman movement using an efficient machine learning (ML) approach. The HAR\nsystems rely on data from untrusted users, making them susceptible to data\npoisoning attacks. In a poisoning attack, attackers manipulate the sensor\nreadings to contaminate the training set, misleading the HAR to produce\nerroneous outcomes. This paper presents the design of a label flipping data\npoisoning attack for a HAR system, where the label of a sensor reading is\nmaliciously changed in the data collection phase. Due to high noise and\nuncertainty in the sensing environment, such an attack poses a severe threat to\nthe recognition system. Besides, vulnerability to label flipping attacks is\ndangerous when activity recognition models are deployed in safety-critical\napplications. This paper shades light on how to carry out the attack in\npractice through smartphone-based sensor data collection applications. This is\nan earlier research work, to our knowledge, that explores attacking the HAR\nmodels via label flipping poisoning. We implement the proposed attack and test\nit on activity recognition models based on the following machine learning\nalgorithms: multi-layer perceptron, decision tree, random forest, and XGBoost.\nFinally, we evaluate the effectiveness of K-nearest neighbors (KNN)-based\ndefense mechanism against the proposed attack.\n", "date": "2022-08-17", "categories": []}
{"id": "2208.06776", "title": "Link-Backdoor: Backdoor Attack on Link Prediction via Node Injection", "abstract": "  Link prediction, inferring the undiscovered or potential links of the graph,\nis widely applied in the real-world. By facilitating labeled links of the graph\nas the training data, numerous deep learning based link prediction methods have\nbeen studied, which have dominant prediction accuracy compared with non-deep\nmethods. However,the threats of maliciously crafted training graph will leave a\nspecific backdoor in the deep model, thus when some specific examples are fed\ninto the model, it will make wrong prediction, defined as backdoor attack. It\nis an important aspect that has been overlooked in the current literature. In\nthis paper, we prompt the concept of backdoor attack on link prediction, and\npropose Link-Backdoor to reveal the training vulnerability of the existing link\nprediction methods. Specifically, the Link-Backdoor combines the fake nodes\nwith the nodes of the target link to form a trigger. Moreover, it optimizes the\ntrigger by the gradient information from the target model. Consequently, the\nlink prediction model trained on the backdoored dataset will predict the link\nwith trigger to the target state. Extensive experiments on five benchmark\ndatasets and five well-performing link prediction models demonstrate that the\nLink-Backdoor achieves the state-of-the-art attack success rate under both\nwhite-box (i.e., available of the target model parameter)and black-box (i.e.,\nunavailable of the target model parameter) scenarios. Additionally, we testify\nthe attack under defensive circumstance, and the results indicate that the\nLink-Backdoor still can construct successful attack on the well-performing link\nprediction methods. The code and data are available at\nhttps://github.com/Seaocn/Link-Backdoor.\n", "date": "2022-08-14", "categories": []}
{"id": "2208.06962", "title": "InvisibiliTee: Angle-agnostic Cloaking from Person-Tracking Systems with\n  a Tee", "abstract": "  After a survey for person-tracking system-induced privacy concerns, we\npropose a black-box adversarial attack method on state-of-the-art human\ndetection models called InvisibiliTee. The method learns printable adversarial\npatterns for T-shirts that cloak wearers in the physical world in front of\nperson-tracking systems. We design an angle-agnostic learning scheme which\nutilizes segmentation of the fashion dataset and a geometric warping process so\nthe adversarial patterns generated are effective in fooling person detectors\nfrom all camera angles and for unseen black-box detection models. Empirical\nresults in both digital and physical environments show that with the\nInvisibiliTee on, person-tracking systems' ability to detect the wearer drops\nsignificantly.\n", "date": "2022-08-15", "categories": []}
{"id": "2208.06984", "title": "A Multi-objective Memetic Algorithm for Auto Adversarial Attack\n  Optimization Design", "abstract": "  The phenomenon of adversarial examples has been revealed in variant\nscenarios. Recent studies show that well-designed adversarial defense\nstrategies can improve the robustness of deep learning models against\nadversarial examples. However, with the rapid development of defense\ntechnologies, it also tends to be more difficult to evaluate the robustness of\nthe defensed model due to the weak performance of existing manually designed\nadversarial attacks. To address the challenge, given the defensed model, the\nefficient adversarial attack with less computational burden and lower robust\naccuracy is needed to be further exploited. Therefore, we propose a\nmulti-objective memetic algorithm for auto adversarial attack optimization\ndesign, which realizes the automatical search for the near-optimal adversarial\nattack towards defensed models. Firstly, the more general mathematical model of\nauto adversarial attack optimization design is constructed, where the search\nspace includes not only the attacker operations, magnitude, iteration number,\nand loss functions but also the connection ways of multiple adversarial\nattacks. In addition, we develop a multi-objective memetic algorithm combining\nNSGA-II and local search to solve the optimization problem. Finally, to\ndecrease the evaluation cost during the search, we propose a representative\ndata selection strategy based on the sorting of cross entropy loss values of\neach images output by models. Experiments on CIFAR10, CIFAR100, and ImageNet\ndatasets show the effectiveness of our proposed method.\n", "date": "2022-08-15", "categories": []}
{"id": "2208.07272", "title": "Training-Time Attacks against k-Nearest Neighbors", "abstract": "  Nearest neighbor-based methods are commonly used for classification tasks and\nas subroutines of other data-analysis methods. An attacker with the capability\nof inserting their own data points into the training set can manipulate the\ninferred nearest neighbor structure. We distill this goal to the task of\nperforming a training-set data insertion attack against $k$-Nearest Neighbor\nclassification ($k$NN). We prove that computing an optimal training-time\n(a.k.a. poisoning) attack against $k$NN classification is NP-Hard, even when $k\n= 1$ and the attacker can insert only a single data point. We provide an\nanytime algorithm to perform such an attack, and a greedy algorithm for general\n$k$ and attacker budget. We provide theoretical bounds and empirically\ndemonstrate the effectiveness and practicality of our methods on synthetic and\nreal-world datasets. Empirically, we find that $k$NN is vulnerable in practice\nand that dimensionality reduction is an effective defense. We conclude with a\ndiscussion of open problems illuminated by our analysis.\n", "date": "2022-08-15", "categories": []}
{"id": "2208.06092", "title": "On deceiving malware classification with section injection", "abstract": "  We investigate how to modify executable files to deceive malware\nclassification systems. This work's main contribution is a methodology to\ninject bytes across a malware file randomly and use it both as an attack to\ndecrease classification accuracy but also as a defensive method, augmenting the\ndata available for training. It respects the operating system file format to\nmake sure the malware will still execute after our injection and will not\nchange its behavior. We reproduced five state-of-the-art malware classification\napproaches to evaluate our injection scheme: one based on GIST+KNN, three CNN\nvariations and one Gated CNN. We performed our experiments on a public dataset\nwith 9,339 malware samples from 25 different families. Our results show that a\nmere increase of 7% in the malware size causes an accuracy drop between 25% and\n40% for malware family classification. They show that a automatic malware\nclassification system may not be as trustworthy as initially reported in the\nliterature. We also evaluate using modified malwares alongside the original\nones to increase networks robustness against mentioned attacks. Results show\nthat a combination of reordering malware sections and injecting random data can\nimprove overall performance of the classification. Code available at\nhttps://github.com/adeilsonsilva/malware-injection.\n", "date": "2022-08-12", "categories": []}
{"id": "2208.04946", "title": "Attention Hijacking in Trojan Transformers", "abstract": "  Trojan attacks pose a severe threat to AI systems. Recent works on\nTransformer models received explosive popularity and the self-attentions are\nnow indisputable. This raises a central question: Can we reveal the Trojans\nthrough attention mechanisms in BERTs and ViTs? In this paper, we investigate\nthe attention hijacking pattern in Trojan AIs, \\ie, the trigger token\n``kidnaps'' the attention weights when a specific trigger is present. We\nobserve the consistent attention hijacking pattern in Trojan Transformers from\nboth Natural Language Processing (NLP) and Computer Vision (CV) domains. This\nintriguing property helps us to understand the Trojan mechanism in BERTs and\nViTs. We also propose an Attention-Hijacking Trojan Detector (AHTD) to\ndiscriminate the Trojan AIs from the clean ones.\n", "date": "2022-08-09", "categories": []}
{"id": "2208.05395", "title": "A Sublinear Adversarial Training Algorithm", "abstract": "  Adversarial training is a widely used strategy for making neural networks\nresistant to adversarial perturbations. For a neural network of width $m$, $n$\ninput training data in $d$ dimension, it takes $\\Omega(mnd)$ time cost per\ntraining iteration for the forward and backward computation. In this paper we\nanalyze the convergence guarantee of adversarial training procedure on a\ntwo-layer neural network with shifted ReLU activation, and shows that only\n$o(m)$ neurons will be activated for each input data per iteration.\nFurthermore, we develop an algorithm for adversarial training with time cost\n$o(m n d)$ per iteration by applying half-space reporting data structure.\n", "date": "2022-08-10", "categories": []}
{"id": "2208.03635", "title": "Federated Adversarial Learning: A Framework with Convergence Analysis", "abstract": "  Federated learning (FL) is a trending training paradigm to utilize\ndecentralized training data. FL allows clients to update model parameters\nlocally for several epochs, then share them to a global model for aggregation.\nThis training paradigm with multi-local step updating before aggregation\nexposes unique vulnerabilities to adversarial attacks. Adversarial training is\na popular and effective method to improve the robustness of networks against\nadversaries. In this work, we formulate a general form of federated adversarial\nlearning (FAL) that is adapted from adversarial learning in the centralized\nsetting. On the client side of FL training, FAL has an inner loop to generate\nadversarial samples for adversarial training and an outer loop to update local\nmodel parameters. On the server side, FAL aggregates local model updates and\nbroadcast the aggregated model. We design a global robust training loss and\nformulate FAL training as a min-max optimization problem. Unlike the\nconvergence analysis in classical centralized training that relies on the\ngradient direction, it is significantly harder to analyze the convergence in\nFAL for three reasons: 1) the complexity of min-max optimization, 2) model not\nupdating in the gradient direction due to the multi-local updates on the\nclient-side before aggregation and 3) inter-client heterogeneity. We address\nthese challenges by using appropriate gradient approximation and coupling\ntechniques and present the convergence analysis in the over-parameterized\nregime. Our main result theoretically shows that the minimum loss under our\nalgorithm can converge to $\\epsilon$ small with chosen learning rate and\ncommunication rounds. It is noteworthy that our analysis is feasible for\nnon-IID clients.\n", "date": "2022-08-07", "categories": []}
{"id": "2208.04135", "title": "Adversarial Attacks on Image Generation With Made-Up Words", "abstract": "  Text-guided image generation models can be prompted to generate images using\nnonce words adversarially designed to robustly evoke specific visual concepts.\nTwo approaches for such generation are introduced: macaronic prompting, which\ninvolves designing cryptic hybrid words by concatenating subword units from\ndifferent languages; and evocative prompting, which involves designing nonce\nwords whose broad morphological features are similar enough to that of existing\nwords to trigger robust visual associations. The two methods can also be\ncombined to generate images associated with more specific visual concepts. The\nimplications of these techniques for the circumvention of existing approaches\nto content moderation, and particularly the generation of offensive or harmful\nimages, are discussed.\n", "date": "2022-08-04", "categories": []}
{"id": "2208.00862", "title": "Attacking Adversarial Defences by Smoothing the Loss Landscape", "abstract": "  This paper investigates a family of methods for defending against adversarial\nattacks that owe part of their success to creating a noisy, discontinuous, or\notherwise rugged loss landscape that adversaries find difficult to navigate. A\ncommon, but not universal, way to achieve this effect is via the use of\nstochastic neural networks. We show that this is a form of gradient\nobfuscation, and propose a general extension to gradient-based adversaries\nbased on the Weierstrass transform, which smooths the surface of the loss\nfunction and provides more reliable gradient estimates. We further show that\nthe same principle can strengthen gradient-free adversaries. We demonstrate the\nefficacy of our loss-smoothing method against both stochastic and\nnon-stochastic adversarial defences that exhibit robustness due to this type of\nobfuscation. Furthermore, we provide analysis of how it interacts with\nExpectation over Transformation; a popular gradient-sampling method currently\nused to attack stochastic defences.\n", "date": "2022-08-01", "categories": []}
{"id": "2208.03161", "title": "Adversarial Robustness of MR Image Reconstruction under Realistic\n  Perturbations", "abstract": "  Deep Learning (DL) methods have shown promising results for solving ill-posed\ninverse problems such as MR image reconstruction from undersampled $k$-space\ndata. However, these approaches currently have no guarantees for reconstruction\nquality and the reliability of such algorithms is only poorly understood.\nAdversarial attacks offer a valuable tool to understand possible failure modes\nand worst case performance of DL-based reconstruction algorithms. In this paper\nwe describe adversarial attacks on multi-coil $k$-space measurements and\nevaluate them on the recently proposed E2E-VarNet and a simpler UNet-based\nmodel. In contrast to prior work, the attacks are targeted to specifically\nalter diagnostically relevant regions. Using two realistic attack models\n(adversarial $k$-space noise and adversarial rotations) we are able to show\nthat current state-of-the-art DL-based reconstruction algorithms are indeed\nsensitive to such perturbations to a degree where relevant diagnostic\ninformation may be lost. Surprisingly, in our experiments the UNet and the more\nsophisticated E2E-VarNet were similarly sensitive to such attacks. Our findings\nadd further to the evidence that caution must be exercised as DL-based methods\nmove closer to clinical practice.\n", "date": "2022-08-05", "categories": []}
{"id": "2208.00539", "title": "Is current research on adversarial robustness addressing the right\n  problem?", "abstract": "  Short answer: Yes, Long answer: No! Indeed, research on adversarial\nrobustness has led to invaluable insights helping us understand and explore\ndifferent aspects of the problem. Many attacks and defenses have been proposed\nover the last couple of years. The problem, however, remains largely unsolved\nand poorly understood. Here, I argue that the current formulation of the\nproblem serves short term goals, and needs to be revised for us to achieve\nbigger gains. Specifically, the bound on perturbation has created a somewhat\ncontrived setting and needs to be relaxed. This has misled us to focus on model\nclasses that are not expressive enough to begin with. Instead, inspired by\nhuman vision and the fact that we rely more on robust features such as shape,\nvertices, and foreground objects than non-robust features such as texture,\nefforts should be steered towards looking for significantly different classes\nof models. Maybe instead of narrowing down on imperceptible adversarial\nperturbations, we should attack a more general problem which is finding\narchitectures that are simultaneously robust to perceptible perturbations,\ngeometric transformations (e.g. rotation, scaling), image distortions\n(lighting, blur), and more (e.g. occlusion, shadow). Only then we may be able\nto solve the problem of adversarial vulnerability.\n", "date": "2022-07-31", "categories": []}
{"id": "2208.02250", "title": "Adversarial Attacks on ASR Systems: An Overview", "abstract": "  With the development of hardware and algorithms, ASR(Automatic Speech\nRecognition) systems evolve a lot. As The models get simpler, the difficulty of\ndevelopment and deployment become easier, ASR systems are getting closer to our\nlife. On the one hand, we often use APPs or APIs of ASR to generate subtitles\nand record meetings. On the other hand, smart speaker and self-driving car rely\non ASR systems to control AIoT devices. In past few years, there are a lot of\nworks on adversarial examples attacks against ASR systems. By adding a small\nperturbation to the waveforms, the recognition results make a big difference.\nIn this paper, we describe the development of ASR system, different assumptions\nof attacks, and how to evaluate these attacks. Next, we introduce the current\nworks on adversarial examples attacks from two attack assumptions: white-box\nattack and black-box attack. Different from other surveys, we pay more\nattention to which layer they perturb waveforms in ASR system, the relationship\nbetween these attacks, and their implementation methods. We focus on the effect\nof their works.\n", "date": "2022-08-03", "categories": []}
{"id": "2208.01844", "title": "Multiclass ASMA vs Targeted PGD Attack in Image Segmentation", "abstract": "  Deep learning networks have demonstrated high performance in a large variety\nof applications, such as image classification, speech recognition, and natural\nlanguage processing. However, there exists a major vulnerability exploited by\nthe use of adversarial attacks. An adversarial attack imputes images by\naltering the input image very slightly, making it nearly undetectable to the\nnaked eye, but results in a very different classification by the network. This\npaper explores the projected gradient descent (PGD) attack and the Adaptive\nMask Segmentation Attack (ASMA) on the image segmentation DeepLabV3 model using\ntwo types of architectures: MobileNetV3 and ResNet50, It was found that PGD was\nvery consistent in changing the segmentation to be its target while the\ngeneralization of ASMA to a multiclass target was not as effective. The\nexistence of such attack however puts all of image classification deep learning\nnetworks in danger of exploitation.\n", "date": "2022-08-03", "categories": []}
{"id": "2208.01853", "title": "Robust Graph Neural Networks using Weighted Graph Laplacian", "abstract": "  Graph neural network (GNN) is achieving remarkable performances in a variety\nof application domains. However, GNN is vulnerable to noise and adversarial\nattacks in input data. Making GNN robust against noises and adversarial attacks\nis an important problem. The existing defense methods for GNNs are\ncomputationally demanding and are not scalable. In this paper, we propose a\ngeneric framework for robustifying GNN known as Weighted Laplacian GNN\n(RWL-GNN). The method combines Weighted Graph Laplacian learning with the GNN\nimplementation. The proposed method benefits from the positive\nsemi-definiteness property of Laplacian matrix, feature smoothness, and latent\nfeatures via formulating a unified optimization framework, which ensures the\nadversarial/noisy edges are discarded and connections in the graph are\nappropriately weighted. For demonstration, the experiments are conducted with\nGraph convolutional neural network(GCNN) architecture, however, the proposed\nframework is easily amenable to any existing GNN architecture. The simulation\nresults with benchmark dataset establish the efficacy of the proposed method,\nboth in accuracy and computational efficiency. Code can be accessed at\nhttps://github.com/Bharat-Runwal/RWL-GNN.\n", "date": "2022-08-03", "categories": []}
{"id": "2208.01456", "title": "A comment on Guo et al. [arXiv:2206.11228]", "abstract": "  In a recent article, Guo et al. [arXiv:2206.11228] report that adversarially\ntrained neural representations in deep networks may already be as robust as\ncorresponding primate IT neural representations. While we find the paper's\nprimary experiment illuminating, we have doubts about the interpretation and\nphrasing of the results presented in the paper.\n", "date": "2022-08-02", "categories": []}
{"id": "2208.00094", "title": "Robust Trajectory Prediction against Adversarial Attacks", "abstract": "  Trajectory prediction using deep neural networks (DNNs) is an essential\ncomponent of autonomous driving (AD) systems. However, these methods are\nvulnerable to adversarial attacks, leading to serious consequences such as\ncollisions. In this work, we identify two key ingredients to defend trajectory\nprediction models against adversarial attacks including (1) designing effective\nadversarial training methods and (2) adding domain-specific data augmentation\nto mitigate the performance degradation on clean data. We demonstrate that our\nmethod is able to improve the performance by 46% on adversarial data and at the\ncost of only 3% performance degradation on clean data, compared to the model\ntrained with clean data. Additionally, compared to existing robust methods, our\nmethod can improve performance by 21% on adversarial examples and 9% on clean\ndata. Our robust model is evaluated with a planner to study its downstream\nimpacts. We demonstrate that our model can significantly reduce the severe\naccident rates (e.g., collisions and off-road driving).\n", "date": "2022-07-29", "categories": []}
{"id": "2208.00498", "title": "DNNShield: Dynamic Randomized Model Sparsification, A Defense Against\n  Adversarial Machine Learning", "abstract": "  DNNs are known to be vulnerable to so-called adversarial attacks that\nmanipulate inputs to cause incorrect results that can be beneficial to an\nattacker or damaging to the victim. Recent works have proposed approximate\ncomputation as a defense mechanism against machine learning attacks. We show\nthat these approaches, while successful for a range of inputs, are insufficient\nto address stronger, high-confidence adversarial attacks. To address this, we\npropose DNNSHIELD, a hardware-accelerated defense that adapts the strength of\nthe response to the confidence of the adversarial input. Our approach relies on\ndynamic and random sparsification of the DNN model to achieve inference\napproximation efficiently and with fine-grain control over the approximation\nerror. DNNSHIELD uses the output distribution characteristics of sparsified\ninference compared to a dense reference to detect adversarial inputs. We show\nan adversarial detection rate of 86% when applied to VGG16 and 88% when applied\nto ResNet50, which exceeds the detection rate of the state of the art\napproaches, with a much lower overhead. We demonstrate a\nsoftware/hardware-accelerated FPGA prototype, which reduces the performance\nimpact of DNNSHIELD relative to software-only CPU and GPU implementations.\n", "date": "2022-07-31", "categories": []}
{"id": "2208.00906", "title": "Understanding Adversarial Robustness of Vision Transformers via Cauchy\n  Problem", "abstract": "  Recent research on the robustness of deep learning has shown that Vision\nTransformers (ViTs) surpass the Convolutional Neural Networks (CNNs) under some\nperturbations, e.g., natural corruption, adversarial attacks, etc. Some papers\nargue that the superior robustness of ViT comes from the segmentation of its\ninput images; others say that the Multi-head Self-Attention (MSA) is the key to\npreserving the robustness. In this paper, we aim to introduce a principled and\nunified theoretical framework to investigate such an argument on ViT's\nrobustness. We first theoretically prove that, unlike Transformers in Natural\nLanguage Processing, ViTs are Lipschitz continuous. Then we theoretically\nanalyze the adversarial robustness of ViTs from the perspective of the Cauchy\nProblem, via which we can quantify how the robustness propagates through\nlayers. We demonstrate that the first and last layers are the critical factors\nto affect the robustness of ViTs. Furthermore, based on our theory, we\nempirically show that unlike the claims from existing research, MSA only\ncontributes to the adversarial robustness of ViTs under weak adversarial\nattacks, e.g., FGSM, and surprisingly, MSA actually comprises the model's\nadversarial robustness under stronger attacks, e.g., PGD attacks.\n", "date": "2022-08-01", "categories": []}
{"id": "2207.13129", "title": "LGV: Boosting Adversarial Example Transferability from Large Geometric\n  Vicinity", "abstract": "  We propose transferability from Large Geometric Vicinity (LGV), a new\ntechnique to increase the transferability of black-box adversarial attacks. LGV\nstarts from a pretrained surrogate model and collects multiple weight sets from\na few additional training epochs with a constant and high learning rate. LGV\nexploits two geometric properties that we relate to transferability. First,\nmodels that belong to a wider weight optimum are better surrogates. Second, we\nidentify a subspace able to generate an effective surrogate ensemble among this\nwider optimum. Through extensive experiments, we show that LGV alone\noutperforms all (combinations of) four established test-time transformations by\n1.8 to 59.9 percentage points. Our findings shed new light on the importance of\nthe geometry of the weight space to explain the transferability of adversarial\nexamples.\n", "date": "2022-07-26", "categories": []}
{"id": "2207.13192", "title": "Perception-Aware Attack: Creating Adversarial Music via\n  Reverse-Engineering Human Perception", "abstract": "  Recently, adversarial machine learning attacks have posed serious security\nthreats against practical audio signal classification systems, including speech\nrecognition, speaker recognition, and music copyright detection. Previous\nstudies have mainly focused on ensuring the effectiveness of attacking an audio\nsignal classifier via creating a small noise-like perturbation on the original\nsignal. It is still unclear if an attacker is able to create audio signal\nperturbations that can be well perceived by human beings in addition to its\nattack effectiveness. This is particularly important for music signals as they\nare carefully crafted with human-enjoyable audio characteristics.\n  In this work, we formulate the adversarial attack against music signals as a\nnew perception-aware attack framework, which integrates human study into\nadversarial attack design. Specifically, we conduct a human study to quantify\nthe human perception with respect to a change of a music signal. We invite\nhuman participants to rate their perceived deviation based on pairs of original\nand perturbed music signals, and reverse-engineer the human perception process\nby regression analysis to predict the human-perceived deviation given a\nperturbed signal. The perception-aware attack is then formulated as an\noptimization problem that finds an optimal perturbation signal to minimize the\nprediction of perceived deviation from the regressed human perception model. We\nuse the perception-aware framework to design a realistic adversarial music\nattack against YouTube's copyright detector. Experiments show that the\nperception-aware attack produces adversarial music with significantly better\nperceptual quality than prior work.\n", "date": "2022-07-26", "categories": []}
{"id": "2207.13417", "title": "Hardly Perceptible Trojan Attack against Neural Networks with Bit Flips", "abstract": "  The security of deep neural networks (DNNs) has attracted increasing\nattention due to their widespread use in various applications. Recently, the\ndeployed DNNs have been demonstrated to be vulnerable to Trojan attacks, which\nmanipulate model parameters with bit flips to inject a hidden behavior and\nactivate it by a specific trigger pattern. However, all existing Trojan attacks\nadopt noticeable patch-based triggers (e.g., a square pattern), making them\nperceptible to humans and easy to be spotted by machines. In this paper, we\npresent a novel attack, namely hardly perceptible Trojan attack (HPT). HPT\ncrafts hardly perceptible Trojan images by utilizing the additive noise and per\npixel flow field to tweak the pixel values and positions of the original\nimages, respectively. To achieve superior attack performance, we propose to\njointly optimize bit flips, additive noise, and flow field. Since the weight\nbits of the DNNs are binary, this problem is very hard to be solved. We handle\nthe binary constraint with equivalent replacement and provide an effective\noptimization algorithm. Extensive experiments on CIFAR-10, SVHN, and ImageNet\ndatasets show that the proposed HPT can generate hardly perceptible Trojan\nimages, while achieving comparable or better attack performance compared to the\nstate-of-the-art methods. The code is available at:\nhttps://github.com/jiawangbai/HPT.\n", "date": "2022-07-27", "categories": []}
{"id": "2207.05756", "title": "Exploring Adversarial Examples and Adversarial Robustness of\n  Convolutional Neural Networks by Mutual Information", "abstract": "  A counter-intuitive property of convolutional neural networks (CNNs) is their\ninherent susceptibility to adversarial examples, which severely hinders the\napplication of CNNs in security-critical fields. Adversarial examples are\nsimilar to original examples but contain malicious perturbations. Adversarial\ntraining is a simple and effective defense method to improve the robustness of\nCNNs to adversarial examples. The mechanisms behind adversarial examples and\nadversarial training are worth exploring. Therefore, this work investigates\nsimilarities and differences between normally trained CNNs (NT-CNNs) and\nadversarially trained CNNs (AT-CNNs) in information extraction from the mutual\ninformation perspective. We show that 1) whether NT-CNNs or AT-CNNs, for\noriginal and adversarial examples, the trends towards mutual information are\nalmost similar throughout training; 2) compared with normal training,\nadversarial training is more difficult and the amount of information that\nAT-CNNs extract from the input is less; 3) the CNNs trained with different\nmethods have different preferences for certain types of information; NT-CNNs\ntend to extract texture-based information from the input, while AT-CNNs prefer\nto shape-based information. The reason why adversarial examples mislead CNNs\nmay be that they contain more texture-based information about other classes.\nFurthermore, we also analyze the mutual information estimators used in this\nwork and find that they outline the geometric properties of the middle layer's\noutput.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.12405", "title": "Versatile Weight Attack via Flipping Limited Bits", "abstract": "  To explore the vulnerability of deep neural networks (DNNs), many attack\nparadigms have been well studied, such as the poisoning-based backdoor attack\nin the training stage and the adversarial attack in the inference stage. In\nthis paper, we study a novel attack paradigm, which modifies model parameters\nin the deployment stage. Considering the effectiveness and stealthiness goals,\nwe provide a general formulation to perform the bit-flip based weight attack,\nwhere the effectiveness term could be customized depending on the attacker's\npurpose. Furthermore, we present two cases of the general formulation with\ndifferent malicious purposes, i.e., single sample attack (SSA) and triggered\nsamples attack (TSA). To this end, we formulate this problem as a mixed integer\nprogramming (MIP) to jointly determine the state of the binary bits (0 or 1) in\nthe memory and learn the sample modification. Utilizing the latest technique in\ninteger programming, we equivalently reformulate this MIP problem as a\ncontinuous optimization problem, which can be effectively and efficiently\nsolved using the alternating direction method of multipliers (ADMM) method.\nConsequently, the flipped critical bits can be easily determined through\noptimization, rather than using a heuristic strategy. Extensive experiments\ndemonstrate the superiority of SSA and TSA in attacking DNNs.\n", "date": "2022-07-25", "categories": []}
{"id": "2207.12863", "title": "FRIB: Low-poisoning Rate Invisible Backdoor Attack based on Feature\n  Repair", "abstract": "  During the generation of invisible backdoor attack poisoned data, the feature\nspace transformation operation tends to cause the loss of some poisoned\nfeatures and weakens the mapping relationship between source images with\ntriggers and target labels, resulting in the need for a higher poisoning rate\nto achieve the corresponding backdoor attack success rate. To solve the above\nproblems, we propose the idea of feature repair for the first time and\nintroduce the blind watermark technique to repair the poisoned features lost\nduring the generation of poisoned data. Under the premise of ensuring\nconsistent labeling, we propose a low-poisoning rate invisible backdoor attack\nbased on feature repair, named FRIB. Benefiting from the above design concept,\nthe new method enhances the mapping relationship between the source images with\ntriggers and the target labels, and increases the degree of misleading DNNs,\nthus achieving a high backdoor attack success rate with a very low poisoning\nrate. Ultimately, the detailed experimental results show that the goal of\nachieving a high success rate of backdoor attacks with a very low poisoning\nrate is achieved on all MNIST, CIFAR10, GTSRB, and ImageNet datasets.\n", "date": "2022-07-26", "categories": []}
{"id": "2207.11694", "title": "Proving Common Mechanisms Shared by Twelve Methods of Boosting\n  Adversarial Transferability", "abstract": "  Although many methods have been proposed to enhance the transferability of\nadversarial perturbations, these methods are designed in a heuristic manner,\nand the essential mechanism for improving adversarial transferability is still\nunclear. This paper summarizes the common mechanism shared by twelve previous\ntransferability-boosting methods in a unified view, i.e., these methods all\nreduce game-theoretic interactions between regional adversarial perturbations.\nTo this end, we focus on the attacking utility of all interactions between\nregional adversarial perturbations, and we first discover and prove the\nnegative correlation between the adversarial transferability and the attacking\nutility of interactions. Based on this discovery, we theoretically prove and\nempirically verify that twelve previous transferability-boosting methods all\nreduce interactions between regional adversarial perturbations. More crucially,\nwe consider the reduction of interactions as the essential reason for the\nenhancement of adversarial transferability. Furthermore, we design the\ninteraction loss to directly penalize interactions between regional adversarial\nperturbations during attacking. Experimental results show that the interaction\nloss significantly improves the transferability of adversarial perturbations.\n", "date": "2022-07-24", "categories": []}
{"id": "2207.12203", "title": "Improving Adversarial Robustness via Mutual Information Estimation", "abstract": "  Deep neural networks (DNNs) are found to be vulnerable to adversarial noise.\nThey are typically misled by adversarial samples to make wrong predictions. To\nalleviate this negative effect, in this paper, we investigate the dependence\nbetween outputs of the target model and input adversarial samples from the\nperspective of information theory, and propose an adversarial defense method.\nSpecifically, we first measure the dependence by estimating the mutual\ninformation (MI) between outputs and the natural patterns of inputs (called\nnatural MI) and MI between outputs and the adversarial patterns of inputs\n(called adversarial MI), respectively. We find that adversarial samples usually\nhave larger adversarial MI and smaller natural MI compared with those w.r.t.\nnatural samples. Motivated by this observation, we propose to enhance the\nadversarial robustness by maximizing the natural MI and minimizing the\nadversarial MI during the training process. In this way, the target model is\nexpected to pay more attention to the natural pattern that contains objective\nsemantics. Empirical evaluations demonstrate that our method could effectively\nimprove the adversarial accuracy against multiple attacks.\n", "date": "2022-07-25", "categories": []}
{"id": "2207.06202", "title": "Adversarially-Aware Robust Object Detector", "abstract": "  Object detection, as a fundamental computer vision task, has achieved a\nremarkable progress with the emergence of deep neural networks. Nevertheless,\nfew works explore the adversarial robustness of object detectors to resist\nadversarial attacks for practical applications in various real-world scenarios.\nDetectors have been greatly challenged by unnoticeable perturbation, with sharp\nperformance drop on clean images and extremely poor performance on adversarial\nimages. In this work, we empirically explore the model training for adversarial\nrobustness in object detection, which greatly attributes to the conflict\nbetween learning clean images and adversarial images. To mitigate this issue,\nwe propose a Robust Detector (RobustDet) based on adversarially-aware\nconvolution to disentangle gradients for model learning on clean and\nadversarial images. RobustDet also employs the Adversarial Image Discriminator\n(AID) and Consistent Features with Reconstruction (CFR) to ensure a reliable\nrobustness. Extensive experiments on PASCAL VOC and MS-COCO demonstrate that\nour model effectively disentangles gradients and significantly enhances the\ndetection robustness with maintaining the detection ability on clean images.\n", "date": "2022-07-13", "categories": []}
{"id": "2207.10719", "title": "Synthetic Dataset Generation for Adversarial Machine Learning Research", "abstract": "  Existing adversarial example research focuses on digitally inserted\nperturbations on top of existing natural image datasets. This construction of\nadversarial examples is not realistic because it may be difficult, or even\nimpossible, for an attacker to deploy such an attack in the real-world due to\nsensing and environmental effects. To better understand adversarial examples\nagainst cyber-physical systems, we propose approximating the real-world through\nsimulation. In this paper we describe our synthetic dataset generation tool\nthat enables scalable collection of such a synthetic dataset with realistic\nadversarial examples. We use the CARLA simulator to collect such a dataset and\ndemonstrate simulated attacks that undergo the same environmental transforms\nand processing as real-world images. Our tools have been used to collect\ndatasets to help evaluate the efficacy of adversarial examples, and can be\nfound at https://github.com/carla-simulator/carla/pull/4992.\n", "date": "2022-07-21", "categories": []}
{"id": "2207.10825", "title": "Just Rotate it: Deploying Backdoor Attacks via Rotation Transformation", "abstract": "  Recent works have demonstrated that deep learning models are vulnerable to\nbackdoor poisoning attacks, where these attacks instill spurious correlations\nto external trigger patterns or objects (e.g., stickers, sunglasses, etc.). We\nfind that such external trigger signals are unnecessary, as highly effective\nbackdoors can be easily inserted using rotation-based image transformation. Our\nmethod constructs the poisoned dataset by rotating a limited amount of objects\nand labeling them incorrectly; once trained with it, the victim's model will\nmake undesirable predictions during run-time inference. It exhibits a\nsignificantly high attack success rate while maintaining clean performance\nthrough comprehensive empirical studies on image classification and object\ndetection tasks. Furthermore, we evaluate standard data augmentation techniques\nand four different backdoor defenses against our attack and find that none of\nthem can serve as a consistent mitigation approach. Our attack can be easily\ndeployed in the real world since it only requires rotating the object, as we\nshow in both image classification and object detection applications. Overall,\nour work highlights a new, simple, physically realizable, and highly effective\nvector for backdoor attacks. Our video demo is available at\nhttps://youtu.be/6JIF8wnX34M.\n", "date": "2022-07-22", "categories": []}
{"id": "2207.10899", "title": "Decoupled Adversarial Contrastive Learning for Self-supervised\n  Adversarial Robustness", "abstract": "  Adversarial training (AT) for robust representation learning and\nself-supervised learning (SSL) for unsupervised representation learning are two\nactive research fields. Integrating AT into SSL, multiple prior works have\naccomplished a highly significant yet challenging task: learning robust\nrepresentation without labels. A widely used framework is adversarial\ncontrastive learning which couples AT and SSL, and thus constitute a very\ncomplex optimization problem. Inspired by the divide-and-conquer philosophy, we\nconjecture that it might be simplified as well as improved by solving two\nsub-problems: non-robust SSL and pseudo-supervised AT. This motivation shifts\nthe focus of the task from seeking an optimal integrating strategy for a\ncoupled problem to finding sub-solutions for sub-problems. With this said, this\nwork discards prior practices of directly introducing AT to SSL frameworks and\nproposed a two-stage framework termed Decoupled Adversarial Contrastive\nLearning (DeACL). Extensive experimental results demonstrate that our DeACL\nachieves SOTA self-supervised adversarial robustness while significantly\nreducing the training time, which validates its effectiveness and efficiency.\nMoreover, our DeACL constitutes a more explainable solution, and its success\nalso bridges the gap with semi-supervised AT for exploiting unlabeled samples\nfor robust representation learning. The code is publicly accessible at\nhttps://github.com/pantheon5100/DeACL.\n", "date": "2022-07-22", "categories": []}
{"id": "2207.11237", "title": "Defending Substitution-Based Profile Pollution Attacks on Sequential\n  Recommenders", "abstract": "  While sequential recommender systems achieve significant improvements on\ncapturing user dynamics, we argue that sequential recommenders are vulnerable\nagainst substitution-based profile pollution attacks. To demonstrate our\nhypothesis, we propose a substitution-based adversarial attack algorithm, which\nmodifies the input sequence by selecting certain vulnerable elements and\nsubstituting them with adversarial items. In both untargeted and targeted\nattack scenarios, we observe significant performance deterioration using the\nproposed profile pollution algorithm. Motivated by such observations, we design\nan efficient adversarial defense method called Dirichlet neighborhood sampling.\nSpecifically, we sample item embeddings from a convex hull constructed by\nmulti-hop neighbors to replace the original items in input sequences. During\nsampling, a Dirichlet distribution is used to approximate the probability\ndistribution in the neighborhood such that the recommender learns to combat\nlocal perturbations. Additionally, we design an adversarial training method\ntailored for sequential recommender systems. In particular, we represent\nselected items with one-hot encodings and perform gradient ascent on the\nencodings to search for the worst case linear combination of item embeddings in\ntraining. As such, the embedding function learns robust item representations\nand the trained recommender is resistant to test-time adversarial examples.\nExtensive experiments show the effectiveness of both our attack and defense\nmethods, which consistently outperform baselines by a significant margin across\nmodel architectures and datasets.\n", "date": "2022-07-19", "categories": []}
{"id": "2207.10290", "title": "AugRmixAT: A Data Processing and Training Method for Improving Multiple\n  Robustness and Generalization Performance", "abstract": "  Deep neural networks are powerful, but they also have shortcomings such as\ntheir sensitivity to adversarial examples, noise, blur, occlusion, etc.\nMoreover, ensuring the reliability and robustness of deep neural network models\nis crucial for their application in safety-critical areas. Much previous work\nhas been proposed to improve specific robustness. However, we find that the\nspecific robustness is often improved at the sacrifice of the additional\nrobustness or generalization ability of the neural network model. In\nparticular, adversarial training methods significantly hurt the generalization\nperformance on unperturbed data when improving adversarial robustness. In this\npaper, we propose a new data processing and training method, called AugRmixAT,\nwhich can simultaneously improve the generalization ability and multiple\nrobustness of neural network models. Finally, we validate the effectiveness of\nAugRmixAT on the CIFAR-10/100 and Tiny-ImageNet datasets. The experiments\ndemonstrate that AugRmixAT can improve the model's generalization performance\nwhile enhancing the white-box robustness, black-box robustness, common\ncorruption robustness, and partial occlusion robustness.\n", "date": "2022-07-21", "categories": []}
{"id": "2207.10307", "title": "Knowledge-enhanced Black-box Attacks for Recommendations", "abstract": "  Recent studies have shown that deep neural networks-based recommender systems\nare vulnerable to adversarial attacks, where attackers can inject carefully\ncrafted fake user profiles (i.e., a set of items that fake users have\ninteracted with) into a target recommender system to achieve malicious\npurposes, such as promote or demote a set of target items. Due to the security\nand privacy concerns, it is more practical to perform adversarial attacks under\nthe black-box setting, where the architecture/parameters and training data of\ntarget systems cannot be easily accessed by attackers. However, generating\nhigh-quality fake user profiles under black-box setting is rather challenging\nwith limited resources to target systems. To address this challenge, in this\nwork, we introduce a novel strategy by leveraging items' attribute information\n(i.e., items' knowledge graph), which can be publicly accessible and provide\nrich auxiliary knowledge to enhance the generation of fake user profiles. More\nspecifically, we propose a knowledge graph-enhanced black-box attacking\nframework (KGAttack) to effectively learn attacking policies through deep\nreinforcement learning techniques, in which knowledge graph is seamlessly\nintegrated into hierarchical policy networks to generate fake user profiles for\nperforming adversarial black-box attacks. Comprehensive experiments on various\nreal-world datasets demonstrate the effectiveness of the proposed attacking\nframework under the black-box setting.\n", "date": "2022-07-21", "categories": []}
{"id": "2207.10498", "title": "Towards Efficient Adversarial Training on Vision Transformers", "abstract": "  Vision Transformer (ViT), as a powerful alternative to Convolutional Neural\nNetwork (CNN), has received much attention. Recent work showed that ViTs are\nalso vulnerable to adversarial examples like CNNs. To build robust ViTs, an\nintuitive way is to apply adversarial training since it has been shown as one\nof the most effective ways to accomplish robust CNNs. However, one major\nlimitation of adversarial training is its heavy computational cost. The\nself-attention mechanism adopted by ViTs is a computationally intense operation\nwhose expense increases quadratically with the number of input patches, making\nadversarial training on ViTs even more time-consuming. In this work, we first\ncomprehensively study fast adversarial training on a variety of vision\ntransformers and illustrate the relationship between the efficiency and\nrobustness. Then, to expediate adversarial training on ViTs, we propose an\nefficient Attention Guided Adversarial Training mechanism. Specifically,\nrelying on the specialty of self-attention, we actively remove certain patch\nembeddings of each layer with an attention-guided dropping strategy during\nadversarial training. The slimmed self-attention modules accelerate the\nadversarial training on ViTs significantly. With only 65\\% of the fast\nadversarial training time, we match the state-of-the-art results on the\nchallenging ImageNet benchmark.\n", "date": "2022-07-21", "categories": []}
{"id": "2206.09628", "title": "Diversified Adversarial Attacks based on Conjugate Gradient Method", "abstract": "  Deep learning models are vulnerable to adversarial examples, and adversarial\nattacks used to generate such examples have attracted considerable research\ninterest. Although existing methods based on the steepest descent have achieved\nhigh attack success rates, ill-conditioned problems occasionally reduce their\nperformance. To address this limitation, we utilize the conjugate gradient (CG)\nmethod, which is effective for this type of problem, and propose a novel attack\nalgorithm inspired by the CG method, named the Auto Conjugate Gradient (ACG)\nattack. The results of large-scale evaluation experiments conducted on the\nlatest robust models show that, for most models, ACG was able to find more\nadversarial examples with fewer iterations than the existing SOTA algorithm\nAuto-PGD (APGD). We investigated the difference in search performance between\nACG and APGD in terms of diversification and intensification, and define a\nmeasure called Diversity Index (DI) to quantify the degree of diversity. From\nthe analysis of the diversity using this index, we show that the more diverse\nsearch of the proposed method remarkably improves its attack success rate.\n", "date": "2022-06-20", "categories": []}
{"id": "2207.09912", "title": "Online Evasion Attacks on Recurrent Models:The Power of Hallucinating\n  the Future", "abstract": "  Recurrent models are frequently being used in online tasks such as autonomous\ndriving, and a comprehensive study of their vulnerability is called for.\nExisting research is limited in generality only addressing application-specific\nvulnerability or making implausible assumptions such as the knowledge of future\ninput. In this paper, we present a general attack framework for online tasks\nincorporating the unique constraints of the online setting different from\noffline tasks. Our framework is versatile in that it covers time-varying\nadversarial objectives and various optimization constraints, allowing for a\ncomprehensive study of robustness. Using the framework, we also present a novel\nwhite-box attack called Predictive Attack that `hallucinates' the future. The\nattack achieves 98 percent of the performance of the ideal but infeasible\nclairvoyant attack on average. We validate the effectiveness of the proposed\nframework and attacks through various experiments.\n", "date": "2022-07-08", "categories": []}
{"id": "2207.08859", "title": "Prior-Guided Adversarial Initialization for Fast Adversarial Training", "abstract": "  Fast adversarial training (FAT) effectively improves the efficiency of\nstandard adversarial training (SAT). However, initial FAT encounters\ncatastrophic overfitting, i.e.,the robust accuracy against adversarial attacks\nsuddenly and dramatically decreases. Though several FAT variants spare no\neffort to prevent overfitting, they sacrifice much calculation cost. In this\npaper, we explore the difference between the training processes of SAT and FAT\nand observe that the attack success rate of adversarial examples (AEs) of FAT\ngets worse gradually in the late training stage, resulting in overfitting. The\nAEs are generated by the fast gradient sign method (FGSM) with a zero or random\ninitialization. Based on the observation, we propose a prior-guided FGSM\ninitialization method to avoid overfitting after investigating several\ninitialization strategies, improving the quality of the AEs during the whole\ntraining process. The initialization is formed by leveraging historically\ngenerated AEs without additional calculation cost. We further provide a\ntheoretical analysis for the proposed initialization method. We also propose a\nsimple yet effective regularizer based on the prior-guided initialization,i.e.,\nthe currently generated perturbation should not deviate too much from the\nprior-guided initialization. The regularizer adopts both historical and current\nadversarial perturbations to guide the model learning. Evaluations on four\ndatasets demonstrate that the proposed method can prevent catastrophic\noverfitting and outperform state-of-the-art FAT methods. The code is released\nat https://github.com/jiaxiaojunQAQ/FGSM-PGI.\n", "date": "2022-07-18", "categories": []}
{"id": "2207.07793", "title": "Towards the Desirable Decision Boundary by Moderate-Margin Adversarial\n  Training", "abstract": "  Adversarial training, as one of the most effective defense methods against\nadversarial attacks, tends to learn an inclusive decision boundary to increase\nthe robustness of deep learning models. However, due to the large and\nunnecessary increase in the margin along adversarial directions, adversarial\ntraining causes heavy cross-over between natural examples and adversarial\nexamples, which is not conducive to balancing the trade-off between robustness\nand natural accuracy. In this paper, we propose a novel adversarial training\nscheme to achieve a better trade-off between robustness and natural accuracy.\nIt aims to learn a moderate-inclusive decision boundary, which means that the\nmargins of natural examples under the decision boundary are moderate. We call\nthis scheme Moderate-Margin Adversarial Training (MMAT), which generates\nfiner-grained adversarial examples to mitigate the cross-over problem. We also\ntake advantage of logits from a teacher model that has been well-trained to\nguide the learning of our model. Finally, MMAT achieves high natural accuracy\nand robustness under both black-box and white-box attacks. On SVHN, for\nexample, state-of-the-art robustness and natural accuracy are achieved.\n", "date": "2022-07-16", "categories": []}
{"id": "2207.07797", "title": "CARBEN: Composite Adversarial Robustness Benchmark", "abstract": "  Prior literature on adversarial attack methods has mainly focused on\nattacking with and defending against a single threat model, e.g., perturbations\nbounded in Lp ball. However, multiple threat models can be combined into\ncomposite perturbations. One such approach, composite adversarial attack (CAA),\nnot only expands the perturbable space of the image, but also may be overlooked\nby current modes of robustness evaluation. This paper demonstrates how CAA's\nattack order affects the resulting image, and provides real-time inferences of\ndifferent models, which will facilitate users' configuration of the parameters\nof the attack level and their rapid evaluation of model prediction. A\nleaderboard to benchmark adversarial robustness against CAA is also introduced.\n", "date": "2022-07-16", "categories": []}
{"id": "2207.07803", "title": "Masked Spatial-Spectral Autoencoders Are Excellent Hyperspectral\n  Defenders", "abstract": "  Deep learning methodology contributes a lot to the development of\nhyperspectral image (HSI) analysis community. However, it also makes HSI\nanalysis systems vulnerable to adversarial attacks. To this end, we propose a\nmasked spatial-spectral autoencoder (MSSA) in this paper under self-supervised\nlearning theory, for enhancing the robustness of HSI analysis systems. First, a\nmasked sequence attention learning module is conducted to promote the inherent\nrobustness of HSI analysis systems along spectral channel. Then, we develop a\ngraph convolutional network with learnable graph structure to establish global\npixel-wise combinations.In this way, the attack effect would be dispersed by\nall the related pixels among each combination, and a better defense performance\nis achievable in spatial aspect.Finally, to improve the defense transferability\nand address the problem of limited labelled samples, MSSA employs spectra\nreconstruction as a pretext task and fits the datasets in a self-supervised\nmanner.Comprehensive experiments over three benchmarks verify the effectiveness\nof MSSA in comparison with the state-of-the-art hyperspectral classification\nmethods and representative adversarial defense strategies.\n", "date": "2022-07-16", "categories": []}
{"id": "2207.08044", "title": "DIMBA: Discretely Masked Black-Box Attack in Single Object Tracking", "abstract": "  The adversarial attack can force a CNN-based model to produce an incorrect\noutput by craftily manipulating human-imperceptible input. Exploring such\nperturbations can help us gain a deeper understanding of the vulnerability of\nneural networks, and provide robustness to deep learning against miscellaneous\nadversaries. Despite extensive studies focusing on the robustness of image,\naudio, and NLP, works on adversarial examples of visual object tracking --\nespecially in a black-box manner -- are quite lacking. In this paper, we\npropose a novel adversarial attack method to generate noises for single object\ntracking under black-box settings, where perturbations are merely added on\ninitial frames of tracking sequences, which is difficult to be noticed from the\nperspective of a whole video clip. Specifically, we divide our algorithm into\nthree components and exploit reinforcement learning for localizing important\nframe patches precisely while reducing unnecessary computational queries\noverhead. Compared to existing techniques, our method requires fewer queries on\ninitialized frames of a video to manipulate competitive or even better attack\nperformance. We test our algorithm in both long-term and short-term datasets,\nincluding OTB100, VOT2018, UAV123, and LaSOT. Extensive experiments demonstrate\nthe effectiveness of our method on three mainstream types of trackers:\ndiscrimination, Siamese-based, and reinforcement learning-based trackers.\n", "date": "2022-07-17", "categories": []}
{"id": "2207.08089", "title": "Threat Model-Agnostic Adversarial Defense using Diffusion Models", "abstract": "  Deep Neural Networks (DNNs) are highly sensitive to imperceptible malicious\nperturbations, known as adversarial attacks. Following the discovery of this\nvulnerability in real-world imaging and vision applications, the associated\nsafety concerns have attracted vast research attention, and many defense\ntechniques have been developed. Most of these defense methods rely on\nadversarial training (AT) -- training the classification network on images\nperturbed according to a specific threat model, which defines the magnitude of\nthe allowed modification. Although AT leads to promising results, training on a\nspecific threat model fails to generalize to other types of perturbations. A\ndifferent approach utilizes a preprocessing step to remove the adversarial\nperturbation from the attacked image. In this work, we follow the latter path\nand aim to develop a technique that leads to robust classifiers across various\nrealizations of threat models. To this end, we harness the recent advances in\nstochastic generative modeling, and means to leverage these for sampling from\nconditional distributions. Our defense relies on an addition of Gaussian i.i.d\nnoise to the attacked image, followed by a pretrained diffusion process -- an\narchitecture that performs a stochastic iterative process over a denoising\nnetwork, yielding a high perceptual quality denoised outcome. The obtained\nrobustness with this stochastic preprocessing step is validated through\nextensive experiments on the CIFAR-10 dataset, showing that our method\noutperforms the leading defense methods under various threat models.\n", "date": "2022-07-17", "categories": []}
{"id": "2207.08137", "title": "Achieve Optimal Adversarial Accuracy for Adversarial Deep Learning using\n  Stackelberg Game", "abstract": "  Adversarial deep learning is to train robust DNNs against adversarial\nattacks, which is one of the major research focuses of deep learning. Game\ntheory has been used to answer some of the basic questions about adversarial\ndeep learning such as the existence of a classifier with optimal robustness and\nthe existence of optimal adversarial samples for a given class of classifiers.\nIn most previous work, adversarial deep learning was formulated as a\nsimultaneous game and the strategy spaces are assumed to be certain probability\ndistributions in order for the Nash equilibrium to exist. But, this assumption\nis not applicable to the practical situation. In this paper, we give answers to\nthese basic questions for the practical case where the classifiers are DNNs\nwith a given structure, by formulating the adversarial deep learning as\nsequential games. The existence of Stackelberg equilibria for these games are\nproved. Furthermore, it is shown that the equilibrium DNN has the largest\nadversarial accuracy among all DNNs with the same structure, when\nCarlini-Wagner's margin loss is used. Trade-off between robustness and accuracy\nin adversarial deep learning is also studied from game theoretical aspect.\n", "date": "2022-07-17", "categories": []}
{"id": "2207.08178", "title": "Watermark Vaccine: Adversarial Attacks to Prevent Watermark Removal", "abstract": "  As a common security tool, visible watermarking has been widely applied to\nprotect copyrights of digital images. However, recent works have shown that\nvisible watermarks can be removed by DNNs without damaging their host images.\nSuch watermark-removal techniques pose a great threat to the ownership of\nimages. Inspired by the vulnerability of DNNs on adversarial perturbations, we\npropose a novel defence mechanism by adversarial machine learning for good.\nFrom the perspective of the adversary, blind watermark-removal networks can be\nposed as our target models; then we actually optimize an imperceptible\nadversarial perturbation on the host images to proactively attack against\nwatermark-removal networks, dubbed Watermark Vaccine. Specifically, two types\nof vaccines are proposed. Disrupting Watermark Vaccine (DWV) induces to ruin\nthe host image along with watermark after passing through watermark-removal\nnetworks. In contrast, Inerasable Watermark Vaccine (IWV) works in another\nfashion of trying to keep the watermark not removed and still noticeable.\nExtensive experiments demonstrate the effectiveness of our DWV/IWV in\npreventing watermark removal, especially on various watermark removal networks.\n", "date": "2022-07-17", "categories": []}
{"id": "2207.08374", "title": "Adversarial Contrastive Learning via Asymmetric InfoNCE", "abstract": "  Contrastive learning (CL) has recently been applied to adversarial learning\ntasks. Such practice considers adversarial samples as additional positive views\nof an instance, and by maximizing their agreements with each other, yields\nbetter adversarial robustness. However, this mechanism can be potentially\nflawed, since adversarial perturbations may cause instance-level identity\nconfusion, which can impede CL performance by pulling together different\ninstances with separate identities. To address this issue, we propose to treat\nadversarial samples unequally when contrasted, with an asymmetric InfoNCE\nobjective ($A-InfoNCE$) that allows discriminating considerations of\nadversarial samples. Specifically, adversaries are viewed as inferior positives\nthat induce weaker learning signals, or as hard negatives exhibiting higher\ncontrast to other negative samples. In the asymmetric fashion, the adverse\nimpacts of conflicting objectives between CL and adversarial learning can be\neffectively mitigated. Experiments show that our approach consistently\noutperforms existing Adversarial CL methods across different finetuning schemes\nwithout additional computational cost. The proposed A-InfoNCE is also a generic\nform that can be readily extended to other CL methods. Code is available at\nhttps://github.com/yqy2001/A-InfoNCE.\n", "date": "2022-07-18", "categories": []}
{"id": "2207.03574", "title": "Demystifying the Adversarial Robustness of Random Transformation\n  Defenses", "abstract": "  Neural networks' lack of robustness against attacks raises concerns in\nsecurity-sensitive settings such as autonomous vehicles. While many\ncountermeasures may look promising, only a few withstand rigorous evaluation.\nDefenses using random transformations (RT) have shown impressive results,\nparticularly BaRT (Raff et al., 2019) on ImageNet. However, this type of\ndefense has not been rigorously evaluated, leaving its robustness properties\npoorly understood. Their stochastic properties make evaluation more challenging\nand render many proposed attacks on deterministic models inapplicable. First,\nwe show that the BPDA attack (Athalye et al., 2018a) used in BaRT's evaluation\nis ineffective and likely overestimates its robustness. We then attempt to\nconstruct the strongest possible RT defense through the informed selection of\ntransformations and Bayesian optimization for tuning their parameters.\nFurthermore, we create the strongest possible attack to evaluate our RT\ndefense. Our new attack vastly outperforms the baseline, reducing the accuracy\nby 83% compared to the 19% reduction by the commonly used EoT attack\n($4.3\\times$ improvement). Our result indicates that the RT defense on the\nImagenette dataset (a ten-class subset of ImageNet) is not robust against\nadversarial examples. Extending the study further, we use our new attack to\nadversarially train RT defense (called AdvRT), resulting in a large robustness\ngain. Code is available at\nhttps://github.com/wagner-group/demystify-random-transform.\n", "date": "2022-06-18", "categories": []}
{"id": "2207.07208", "title": "Provably Adversarially Robust Nearest Prototype Classifiers", "abstract": "  Nearest prototype classifiers (NPCs) assign to each input point the label of\nthe nearest prototype with respect to a chosen distance metric. A direct\nadvantage of NPCs is that the decisions are interpretable. Previous work could\nprovide lower bounds on the minimal adversarial perturbation in the\n$\\ell_p$-threat model when using the same $\\ell_p$-distance for the NPCs. In\nthis paper we provide a complete discussion on the complexity when using\n$\\ell_p$-distances for decision and $\\ell_q$-threat models for certification\nfor $p,q \\in \\{1,2,\\infty\\}$. In particular we provide scalable algorithms for\nthe \\emph{exact} computation of the minimal adversarial perturbation when using\n$\\ell_2$-distance and improved lower bounds in other cases. Using efficient\nimproved lower bounds we train our Provably adversarially robust NPC (PNPC),\nfor MNIST which have better $\\ell_2$-robustness guarantees than neural\nnetworks. Additionally, we show up to our knowledge the first certification\nresults w.r.t. to the LPIPS perceptual metric which has been argued to be a\nmore realistic threat model for image classification than $\\ell_p$-balls. Our\nPNPC has on CIFAR10 higher certified robust accuracy than the empirical robust\naccuracy reported in (Laidlaw et al., 2021). The code is available in our\nrepository.\n", "date": "2022-07-14", "categories": []}
{"id": "2207.07232", "title": "Lipschitz Bound Analysis of Neural Networks", "abstract": "  Lipschitz Bound Estimation is an effective method of regularizing deep neural\nnetworks to make them robust against adversarial attacks. This is useful in a\nvariety of applications ranging from reinforcement learning to autonomous\nsystems. In this paper, we highlight the significant gap in obtaining a\nnon-trivial Lipschitz bound certificate for Convolutional Neural Networks\n(CNNs) and empirically support it with extensive graphical analysis. We also\nshow that unrolling Convolutional layers or Toeplitz matrices can be employed\nto convert Convolutional Neural Networks (CNNs) to a Fully Connected Network.\nFurther, we propose a simple algorithm to show the existing 20x-50x gap in a\nparticular data distribution between the actual lipschitz constant and the\nobtained tight bound. We also ran sets of thorough experiments on various\nnetwork architectures and benchmark them on datasets like MNIST and CIFAR-10.\nAll these proposals are supported by extensive testing, graphs, histograms and\ncomparative analysis.\n", "date": "2022-07-14", "categories": []}
{"id": "2207.07347", "title": "Feasibility of Inconspicuous GAN-generated Adversarial Patches against\n  Object Detection", "abstract": "  Standard approaches for adversarial patch generation lead to noisy\nconspicuous patterns, which are easily recognizable by humans. Recent research\nhas proposed several approaches to generate naturalistic patches using\ngenerative adversarial networks (GANs), yet only a few of them were evaluated\non the object detection use case. Moreover, the state of the art mostly focuses\non suppressing a single large bounding box in input by overlapping it with the\npatch directly. Suppressing objects near the patch is a different, more complex\ntask. In this work, we have evaluated the existing approaches to generate\ninconspicuous patches. We have adapted methods, originally developed for\ndifferent computer vision tasks, to the object detection use case with YOLOv3\nand the COCO dataset. We have evaluated two approaches to generate naturalistic\npatches: by incorporating patch generation into the GAN training process and by\nusing the pretrained GAN. For both cases, we have assessed a trade-off between\nperformance and naturalistic patch appearance. Our experiments have shown, that\nusing a pre-trained GAN helps to gain realistic-looking patches while\npreserving the performance similar to conventional adversarial patches.\n", "date": "2022-07-15", "categories": []}
{"id": "2207.00278", "title": "BadHash: Invisible Backdoor Attacks against Deep Hashing with Clean\n  Label", "abstract": "  Due to its powerful feature learning capability and high efficiency, deep\nhashing has achieved great success in large-scale image retrieval. Meanwhile,\nextensive works have demonstrated that deep neural networks (DNNs) are\nsusceptible to adversarial examples, and exploring adversarial attack against\ndeep hashing has attracted many research efforts. Nevertheless, backdoor\nattack, another famous threat to DNNs, has not been studied for deep hashing\nyet. Although various backdoor attacks have been proposed in the field of image\nclassification, existing approaches failed to realize a truly imperceptive\nbackdoor attack that enjoys invisible triggers and clean label setting\nsimultaneously, and they also cannot meet the intrinsic demand of image\nretrieval backdoor. In this paper, we propose BadHash, the first\ngenerative-based imperceptible backdoor attack against deep hashing, which can\neffectively generate invisible and input-specific poisoned images with clean\nlabel. Specifically, we first propose a new conditional generative adversarial\nnetwork (cGAN) pipeline to effectively generate poisoned samples. For any given\nbenign image, it seeks to generate a natural-looking poisoned counterpart with\na unique invisible trigger. In order to improve the attack effectiveness, we\nintroduce a label-based contrastive learning network LabCLN to exploit the\nsemantic characteristics of different labels, which are subsequently used for\nconfusing and misleading the target model to learn the embedded trigger. We\nfinally explore the mechanism of backdoor attacks on image retrieval in the\nhash space. Extensive experiments on multiple benchmark datasets verify that\nBadHash can generate imperceptible poisoned samples with strong attack ability\nand transferability over state-of-the-art deep hashing schemes.\n", "date": "2022-07-01", "categories": []}
{"id": "2207.04305", "title": "Training Robust Deep Models for Time-Series Domain: Novel Algorithms and\n  Theoretical Analysis", "abstract": "  Despite the success of deep neural networks (DNNs) for real-world\napplications over time-series data such as mobile health, little is known about\nhow to train robust DNNs for time-series domain due to its unique\ncharacteristics compared to images and text data. In this paper, we propose a\nnovel algorithmic framework referred as RObust Training for Time-Series (RO-TS)\nto create robust DNNs for time-series classification tasks. Specifically, we\nformulate a min-max optimization problem over the model parameters by\nexplicitly reasoning about the robustness criteria in terms of additive\nperturbations to time-series inputs measured by the global alignment kernel\n(GAK) based distance. We also show the generality and advantages of our\nformulation using the summation structure over time-series alignments by\nrelating both GAK and dynamic time warping (DTW). This problem is an instance\nof a family of compositional min-max optimization problems, which are\nchallenging and open with unclear theoretical guarantee. We propose a\nprincipled stochastic compositional alternating gradient descent ascent\n(SCAGDA) algorithm for this family of optimization problems. Unlike traditional\nmethods for time-series that require approximate computation of distance\nmeasures, SCAGDA approximates the GAK based distance on-the-fly using a moving\naverage approach. We theoretically analyze the convergence rate of SCAGDA and\nprovide strong theoretical support for the estimation of GAK based distance.\nOur experiments on real-world benchmarks demonstrate that RO-TS creates more\nrobust DNNs when compared to adversarial training using prior methods that rely\non data augmentation or new definitions of loss functions. We also demonstrate\nthe importance of GAK for time-series data over the Euclidean distance. The\nsource code of RO-TS algorithms is available at\nhttps://github.com/tahabelkhouja/Robust-Training-for-Time-Series\n", "date": "2022-07-09", "categories": []}
{"id": "2207.05937", "title": "Game of Trojans: A Submodular Byzantine Approach", "abstract": "  Machine learning models in the wild have been shown to be vulnerable to\nTrojan attacks during training. Although many detection mechanisms have been\nproposed, strong adaptive attackers have been shown to be effective against\nthem. In this paper, we aim to answer the questions considering an intelligent\nand adaptive adversary: (i) What is the minimal amount of instances required to\nbe Trojaned by a strong attacker? and (ii) Is it possible for such an attacker\nto bypass strong detection mechanisms?\n  We provide an analytical characterization of adversarial capability and\nstrategic interactions between the adversary and detection mechanism that take\nplace in such models. We characterize adversary capability in terms of the\nfraction of the input dataset that can be embedded with a Trojan trigger. We\nshow that the loss function has a submodular structure, which leads to the\ndesign of computationally efficient algorithms to determine this fraction with\nprovable bounds on optimality. We propose a Submodular Trojan algorithm to\ndetermine the minimal fraction of samples to inject a Trojan trigger. To evade\ndetection of the Trojaned model, we model strategic interactions between the\nadversary and Trojan detection mechanism as a two-player game. We show that the\nadversary wins the game with probability one, thus bypassing detection. We\nestablish this by proving that output probability distributions of a Trojan\nmodel and a clean model are identical when following the Min-Max (MM) Trojan\nalgorithm.\n  We perform extensive evaluations of our algorithms on MNIST, CIFAR-10, and\nEuroSAT datasets. The results show that (i) with Submodular Trojan algorithm,\nthe adversary needs to embed a Trojan trigger into a very small fraction of\nsamples to achieve high accuracy on both Trojan and clean samples, and (ii) the\nMM Trojan algorithm yields a trained Trojan model that evades detection with\nprobability 1.\n", "date": "2022-07-13", "categories": []}
{"id": "2206.00913", "title": "Improving the Robustness and Generalization of Deep Neural Network with\n  Confidence Threshold Reduction", "abstract": "  Deep neural networks are easily attacked by imperceptible perturbation.\nPresently, adversarial training (AT) is the most effective method to enhance\nthe robustness of the model against adversarial examples. However, because\nadversarial training solved a min-max value problem, in comparison with natural\ntraining, the robustness and generalization are contradictory, i.e., the\nrobustness improvement of the model will decrease the generalization of the\nmodel. To address this issue, in this paper, a new concept, namely confidence\nthreshold (CT), is introduced and the reducing of the confidence threshold,\nknown as confidence threshold reduction (CTR), is proven to improve both the\ngeneralization and robustness of the model. Specifically, to reduce the CT for\nnatural training (i.e., for natural training with CTR), we propose a\nmask-guided divergence loss function (MDL) consisting of a cross-entropy loss\nterm and an orthogonal term. The empirical and theoretical analysis\ndemonstrates that the MDL loss improves the robustness and generalization of\nthe model simultaneously for natural training. However, the model robustness\nimprovement of natural training with CTR is not comparable to that of\nadversarial training. Therefore, for adversarial training, we propose a\nstandard deviation loss function (STD), which minimizes the difference in the\nprobabilities of the wrong categories, to reduce the CT by being integrated\ninto the loss function of adversarial training. The empirical and theoretical\nanalysis demonstrates that the STD based loss function can further improve the\nrobustness of the adversarially trained model on basis of guaranteeing the\nchangeless or slight improvement of the natural accuracy.\n", "date": "2022-06-02", "categories": []}
{"id": "2207.05137", "title": "Towards Effective Multi-Label Recognition Attacks via Knowledge Graph\n  Consistency", "abstract": "  Many real-world applications of image recognition require multi-label\nlearning, whose goal is to find all labels in an image. Thus, robustness of\nsuch systems to adversarial image perturbations is extremely important.\nHowever, despite a large body of recent research on adversarial attacks, the\nscope of the existing works is mainly limited to the multi-class setting, where\neach image contains a single label. We show that the naive extensions of\nmulti-class attacks to the multi-label setting lead to violating label\nrelationships, modeled by a knowledge graph, and can be detected using a\nconsistency verification scheme. Therefore, we propose a graph-consistent\nmulti-label attack framework, which searches for small image perturbations that\nlead to misclassifying a desired target set while respecting label hierarchies.\nBy extensive experiments on two datasets and using several multi-label\nrecognition models, we show that our method generates extremely successful\nattacks that, unlike naive multi-label perturbations, can produce model\npredictions consistent with the knowledge graph.\n", "date": "2022-07-11", "categories": []}
{"id": "2207.05382", "title": "Frequency Domain Model Augmentation for Adversarial Attack", "abstract": "  For black-box attacks, the gap between the substitute model and the victim\nmodel is usually large, which manifests as a weak attack performance. Motivated\nby the observation that the transferability of adversarial examples can be\nimproved by attacking diverse models simultaneously, model augmentation methods\nwhich simulate different models by using transformed images are proposed.\nHowever, existing transformations for spatial domain do not translate to\nsignificantly diverse augmented models. To tackle this issue, we propose a\nnovel spectrum simulation attack to craft more transferable adversarial\nexamples against both normally trained and defense models. Specifically, we\napply a spectrum transformation to the input and thus perform the model\naugmentation in the frequency domain. We theoretically prove that the\ntransformation derived from frequency domain leads to a diverse spectrum\nsaliency map, an indicator we proposed to reflect the diversity of substitute\nmodels. Notably, our method can be generally combined with existing attacks.\nExtensive experiments on the ImageNet dataset demonstrate the effectiveness of\nour method, \\textit{e.g.}, attacking nine state-of-the-art defense models with\nan average success rate of \\textbf{95.4\\%}. Our code is available in\n\\url{https://github.com/yuyang-long/SSA}.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.05548", "title": "Practical Attacks on Machine Learning: A Case Study on Adversarial\n  Windows Malware", "abstract": "  While machine learning is vulnerable to adversarial examples, it still lacks\nsystematic procedures and tools for evaluating its security in different\napplication contexts. In this article, we discuss how to develop automated and\nscalable security evaluations of machine learning using practical attacks,\nreporting a use case on Windows malware detection.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.02152", "title": "UniCR: Universally Approximated Certified Robustness via Randomized\n  Smoothing", "abstract": "  We study certified robustness of machine learning classifiers against\nadversarial perturbations. In particular, we propose the first universally\napproximated certified robustness (UniCR) framework, which can approximate the\nrobustness certification of any input on any classifier against any $\\ell_p$\nperturbations with noise generated by any continuous probability distribution.\nCompared with the state-of-the-art certified defenses, UniCR provides many\nsignificant benefits: (1) the first universal robustness certification\nframework for the above 4 'any's; (2) automatic robustness certification that\navoids case-by-case analysis, (3) tightness validation of certified robustness,\nand (4) optimality validation of noise distributions used by randomized\nsmoothing. We conduct extensive experiments to validate the above benefits of\nUniCR and the advantages of UniCR over state-of-the-art certified defenses\nagainst $\\ell_p$ perturbations.\n", "date": "2022-07-05", "categories": []}
{"id": "2207.04209", "title": "Invisible Backdoor Attacks Using Data Poisoning in the Frequency Domain", "abstract": "  With the broad application of deep neural networks (DNNs), backdoor attacks\nhave gradually attracted attention. Backdoor attacks are insidious, and\npoisoned models perform well on benign samples and are only triggered when\ngiven specific inputs, which cause the neural network to produce incorrect\noutputs. The state-of-the-art backdoor attack work is implemented by data\npoisoning, i.e., the attacker injects poisoned samples into the dataset, and\nthe models trained with that dataset are infected with the backdoor. However,\nmost of the triggers used in the current study are fixed patterns patched on a\nsmall fraction of an image and are often clearly mislabeled, which is easily\ndetected by humans or defense methods such as Neural Cleanse and SentiNet.\nAlso, it's difficult to be learned by DNNs without mislabeling, as they may\nignore small patterns. In this paper, we propose a generalized backdoor attack\nmethod based on the frequency domain, which can implement backdoor implantation\nwithout mislabeling and accessing the training process. It is invisible to\nhuman beings and able to evade the commonly used defense methods. We evaluate\nour approach in the no-label and clean-label cases on three datasets (CIFAR-10,\nSTL-10, and GTSRB) with two popular scenarios (self-supervised learning and\nsupervised learning). The results show our approach can achieve a high attack\nsuccess rate (above 90%) on all the tasks without significant performance\ndegradation on main tasks. Also, we evaluate the bypass performance of our\napproach for different kinds of defenses, including the detection of training\ndata (i.e., Activation Clustering), the preprocessing of inputs (i.e.,\nFiltering), the detection of inputs (i.e., SentiNet), and the detection of\nmodels (i.e., Neural Cleanse). The experimental results demonstrate that our\napproach shows excellent robustness to such defenses.\n", "date": "2022-07-09", "categories": []}
{"id": "2207.04307", "title": "Adversarial Framework with Certified Robustness for Time-Series Domain\n  via Statistical Features", "abstract": "  Time-series data arises in many real-world applications (e.g., mobile health)\nand deep neural networks (DNNs) have shown great success in solving them.\nDespite their success, little is known about their robustness to adversarial\nattacks. In this paper, we propose a novel adversarial framework referred to as\nTime-Series Attacks via STATistical Features (TSA-STAT)}. To address the unique\nchallenges of time-series domain, TSA-STAT employs constraints on statistical\nfeatures of the time-series data to construct adversarial examples. Optimized\npolynomial transformations are used to create attacks that are more effective\n(in terms of successfully fooling DNNs) than those based on additive\nperturbations. We also provide certified bounds on the norm of the statistical\nfeatures for constructing adversarial examples. Our experiments on diverse\nreal-world benchmark datasets show the effectiveness of TSA-STAT in fooling\nDNNs for time-series domain and in improving their robustness. The source code\nof TSA-STAT algorithms is available at\nhttps://github.com/tahabelkhouja/Time-Series-Attacks-via-STATistical-Features\n", "date": "2022-07-09", "categories": []}
{"id": "2207.04718", "title": "Physical Attack on Monocular Depth Estimation with Optimal Adversarial\n  Patches", "abstract": "  Deep learning has substantially boosted the performance of Monocular Depth\nEstimation (MDE), a critical component in fully vision-based autonomous driving\n(AD) systems (e.g., Tesla and Toyota). In this work, we develop an attack\nagainst learning-based MDE. In particular, we use an optimization-based method\nto systematically generate stealthy physical-object-oriented adversarial\npatches to attack depth estimation. We balance the stealth and effectiveness of\nour attack with object-oriented adversarial design, sensitive region\nlocalization, and natural style camouflage. Using real-world driving scenarios,\nwe evaluate our attack on concurrent MDE models and a representative downstream\ntask for AD (i.e., 3D object detection). Experimental results show that our\nmethod can generate stealthy, effective, and robust adversarial patches for\ndifferent target objects and models and achieves more than 6 meters mean depth\nestimation error and 93% attack success rate (ASR) in object detection with a\npatch of 1/9 of the vehicle's rear area. Field tests on three different driving\nroutes with a real vehicle indicate that we cause over 6 meters mean depth\nestimation error and reduce the object detection rate from 90.70% to 5.16% in\ncontinuous video frames.\n", "date": "2022-07-11", "categories": []}
{"id": "2207.04843", "title": "Statistical Detection of Adversarial examples in Blockchain-based\n  Federated Forest In-vehicle Network Intrusion Detection Systems", "abstract": "  The internet-of-Vehicle (IoV) can facilitate seamless connectivity between\nconnected vehicles (CV), autonomous vehicles (AV), and other IoV entities.\nIntrusion Detection Systems (IDSs) for IoV networks can rely on machine\nlearning (ML) to protect the in-vehicle network from cyber-attacks.\nBlockchain-based Federated Forests (BFFs) could be used to train ML models\nbased on data from IoV entities while protecting the confidentiality of the\ndata and reducing the risks of tampering with the data. However, ML models\ncreated this way are still vulnerable to evasion, poisoning, and exploratory\nattacks using adversarial examples. This paper investigates the impact of\nvarious possible adversarial examples on the BFF-IDS. We proposed integrating a\nstatistical detector to detect and extract unknown adversarial samples. By\nincluding the unknown detected samples into the dataset of the detector, we\naugment the BFF-IDS with an additional model to detect original known attacks\nand the new adversarial inputs. The statistical adversarial detector\nconfidently detected adversarial examples at the sample size of 50 and 100\ninput samples. Furthermore, the augmented BFF-IDS (BFF-IDS(AUG)) successfully\nmitigates the adversarial examples with more than 96% accuracy. With this\napproach, the model will continue to be augmented in a sandbox whenever an\nadversarial sample is detected and subsequently adopt the BFF-IDS(AUG) as the\nactive security model. Consequently, the proposed integration of the\nstatistical adversarial detector and the subsequent augmentation of the BFF-IDS\nwith detected adversarial samples provides a sustainable security framework\nagainst adversarial examples and other unknown attacks.\n", "date": "2022-07-11", "categories": []}
{"id": "2207.03895", "title": "Defense Against Multi-target Trojan Attacks", "abstract": "  Adversarial attacks on deep learning-based models pose a significant threat\nto the current AI infrastructure. Among them, Trojan attacks are the hardest to\ndefend against. In this paper, we first introduce a variation of the Badnet\nkind of attacks that introduces Trojan backdoors to multiple target classes and\nallows triggers to be placed anywhere in the image. The former makes it more\npotent and the latter makes it extremely easy to carry out the attack in the\nphysical space. The state-of-the-art Trojan detection methods fail with this\nthreat model. To defend against this attack, we first introduce a trigger\nreverse-engineering mechanism that uses multiple images to recover a variety of\npotential triggers. We then propose a detection mechanism by measuring the\ntransferability of such recovered triggers. A Trojan trigger will have very\nhigh transferability i.e. they make other images also go to the same class. We\nstudy many practical advantages of our attack method and then demonstrate the\ndetection performance using a variety of image datasets. The experimental\nresults show the superior detection performance of our method over the\nstate-of-the-arts.\n", "date": "2022-07-08", "categories": []}
{"id": "2207.02963", "title": "The Weaknesses of Adversarial Camouflage in Overhead Imagery", "abstract": "  Machine learning is increasingly critical for analysis of the ever-growing\ncorpora of overhead imagery. Advanced computer vision object detection\ntechniques have demonstrated great success in identifying objects of interest\nsuch as ships, automobiles, and aircraft from satellite and drone imagery. Yet\nrelying on computer vision opens up significant vulnerabilities, namely, the\nsusceptibility of object detection algorithms to adversarial attacks. In this\npaper we explore the efficacy and drawbacks of adversarial camouflage in an\noverhead imagery context. While a number of recent papers have demonstrated the\nability to reliably fool deep learning classifiers and object detectors with\nadversarial patches, most of this work has been performed on relatively uniform\ndatasets and only a single class of objects. In this work we utilize the\nVisDrone dataset, which has a large range of perspectives and object sizes. We\nexplore four different object classes: bus, car, truck, van. We build a library\nof 24 adversarial patches to disguise these objects, and introduce a patch\ntranslucency variable to our patches. The translucency (or alpha value) of the\npatches is highly correlated to their efficacy. Further, we show that while\nadversarial patches may fool object detectors, the presence of such patches is\noften easily uncovered, with patches on average 24% more detectable than the\nobjects the patches were meant to hide. This raises the question of whether\nsuch patches truly constitute camouflage. Source code is available at\nhttps://github.com/IQTLabs/camolo.\n", "date": "2022-07-06", "categories": []}
{"id": "2206.03717", "title": "LADDER: Latent Boundary-guided Adversarial Training", "abstract": "  Deep Neural Networks (DNNs) have recently achieved great success in many\nclassification tasks. Unfortunately, they are vulnerable to adversarial attacks\nthat generate adversarial examples with a small perturbation to fool DNN\nmodels, especially in model sharing scenarios. Adversarial training is proved\nto be the most effective strategy that injects adversarial examples into model\ntraining to improve the robustness of DNN models against adversarial attacks.\nHowever, adversarial training based on the existing adversarial examples fails\nto generalize well to standard, unperturbed test data. To achieve a better\ntrade-off between standard accuracy and adversarial robustness, we propose a\nnovel adversarial training framework called LAtent bounDary-guided aDvErsarial\ntRaining (LADDER) that adversarially trains DNN models on latent\nboundary-guided adversarial examples. As opposed to most of the existing\nmethods that generate adversarial examples in the input space, LADDER generates\na myriad of high-quality adversarial examples through adding perturbations to\nlatent features. The perturbations are made along the normal of the decision\nboundary constructed by an SVM with an attention mechanism. We analyze the\nmerits of our generated boundary-guided adversarial examples from a boundary\nfield perspective and visualization view. Extensive experiments and detailed\nanalysis on MNIST, SVHN, CelebA, and CIFAR-10 validate the effectiveness of\nLADDER in achieving a better trade-off between standard accuracy and\nadversarial robustness as compared with vanilla DNNs and competitive baselines.\n", "date": "2022-06-08", "categories": []}
{"id": "2207.02391", "title": "Query-Efficient Adversarial Attack Based on Latin Hypercube Sampling", "abstract": "  In order to be applicable in real-world scenario, Boundary Attacks (BAs) were\nproposed and ensured one hundred percent attack success rate with only decision\ninformation. However, existing BA methods craft adversarial examples by\nleveraging a simple random sampling (SRS) to estimate the gradient, consuming a\nlarge number of model queries. To overcome the drawback of SRS, this paper\nproposes a Latin Hypercube Sampling based Boundary Attack (LHS-BA) to save\nquery budget. Compared with SRS, LHS has better uniformity under the same\nlimited number of random samples. Therefore, the average on these random\nsamples is closer to the true gradient than that estimated by SRS. Various\nexperiments are conducted on benchmark datasets including MNIST, CIFAR, and\nImageNet-1K. Experimental results demonstrate the superiority of the proposed\nLHS-BA over the state-of-the-art BA methods in terms of query efficiency. The\nsource codes are publicly available at https://github.com/GZHU-DVL/LHS-BA.\n", "date": "2022-07-05", "categories": []}
{"id": "2207.00425", "title": "Transferable Graph Backdoor Attack", "abstract": "  Graph Neural Networks (GNNs) have achieved tremendous success in many graph\nmining tasks benefitting from the message passing strategy that fuses the local\nstructure and node features for better graph representation learning. Despite\nthe success of GNNs, and similar to other types of deep neural networks, GNNs\nare found to be vulnerable to unnoticeable perturbations on both graph\nstructure and node features. Many adversarial attacks have been proposed to\ndisclose the fragility of GNNs under different perturbation strategies to\ncreate adversarial examples. However, vulnerability of GNNs to successful\nbackdoor attacks was only shown recently. In this paper, we disclose the TRAP\nattack, a Transferable GRAPh backdoor attack. The core attack principle is to\npoison the training dataset with perturbation-based triggers that can lead to\nan effective and transferable backdoor attack. The perturbation trigger for a\ngraph is generated by performing the perturbation actions on the graph\nstructure via a gradient based score matrix from a surrogate model. Compared\nwith prior works, TRAP attack is different in several ways: i) it exploits a\nsurrogate Graph Convolutional Network (GCN) model to generate perturbation\ntriggers for a blackbox based backdoor attack; ii) it generates sample-specific\nperturbation triggers which do not have a fixed pattern; and iii) the attack\ntransfers, for the first time in the context of GNNs, to different GNN models\nwhen trained with the forged poisoned training dataset. Through extensive\nevaluations on four real-world datasets, we demonstrate the effectiveness of\nthe TRAP attack to build transferable backdoors in four different popular GNNs\nusing four real-world datasets.\n", "date": "2022-06-21", "categories": []}
{"id": "2206.08255", "title": "Gradient-Based Adversarial and Out-of-Distribution Detection", "abstract": "  We propose to utilize gradients for detecting adversarial and\nout-of-distribution samples. We introduce confounding labels -- labels that\ndiffer from normal labels seen during training -- in gradient generation to\nprobe the effective expressivity of neural networks. Gradients depict the\namount of change required for a model to properly represent given inputs,\nproviding insight into the representational power of the model established by\nnetwork architectural properties as well as training data. By introducing a\nlabel of different design, we remove the dependency on ground truth labels for\ngradient generation during inference. We show that our gradient-based approach\nallows for capturing the anomaly in inputs based on the effective expressivity\nof the models with no hyperparameter tuning or additional processing, and\noutperforms state-of-the-art methods for adversarial and out-of-distribution\ndetection.\n", "date": "2022-06-16", "categories": []}
{"id": "2207.00694", "title": "Efficient Adversarial Training With Data Pruning", "abstract": "  Neural networks are susceptible to adversarial examples-small input\nperturbations that cause models to fail. Adversarial training is one of the\nsolutions that stops adversarial examples; models are exposed to attacks during\ntraining and learn to be resilient to them. Yet, such a procedure is currently\nexpensive-it takes a long time to produce and train models with adversarial\nsamples, and, what is worse, it occasionally fails. In this paper we\ndemonstrate data pruning-a method for increasing adversarial training\nefficiency through data sub-sampling.We empirically show that data pruning\nleads to improvements in convergence and reliability of adversarial training,\nalbeit with different levels of utility degradation. For example, we observe\nthat using random sub-sampling of CIFAR10 to drop 40% of data, we lose 8%\nadversarial accuracy against the strongest attackers, while by using only 20%\nof data we lose 14% adversarial accuracy and reduce runtime by a factor of 3.\nInterestingly, we discover that in some settings data pruning brings benefits\nfrom both worlds-it both improves adversarial accuracy and training time.\n", "date": "2022-07-01", "categories": []}
{"id": "2207.01149", "title": "RAF: Recursive Adversarial Attacks on Face Recognition Using Extremely\n  Limited Queries", "abstract": "  Recent successful adversarial attacks on face recognition show that, despite\nthe remarkable progress of face recognition models, they are still far behind\nthe human intelligence for perception and recognition. It reveals the\nvulnerability of deep convolutional neural networks (CNNs) as state-of-the-art\nbuilding block for face recognition models against adversarial examples, which\ncan cause certain consequences for secure systems. Gradient-based adversarial\nattacks are widely studied before and proved to be successful against face\nrecognition models. However, finding the optimized perturbation per each face\nneeds to submitting the significant number of queries to the target model. In\nthis paper, we propose recursive adversarial attack on face recognition using\nautomatic face warping which needs extremely limited number of queries to fool\nthe target model. Instead of a random face warping procedure, the warping\nfunctions are applied on specific detected regions of face like eyebrows, nose,\nlips, etc. We evaluate the robustness of proposed method in the decision-based\nblack-box attack setting, where the attackers have no access to the model\nparameters and gradients, but hard-label predictions and confidence scores are\nprovided by the target model.\n", "date": "2022-07-04", "categories": []}
{"id": "2207.01156", "title": "Removing Batch Normalization Boosts Adversarial Training", "abstract": "  Adversarial training (AT) defends deep neural networks against adversarial\nattacks. One challenge that limits its practical application is the performance\ndegradation on clean samples. A major bottleneck identified by previous works\nis the widely used batch normalization (BN), which struggles to model the\ndifferent statistics of clean and adversarial training samples in AT. Although\nthe dominant approach is to extend BN to capture this mixture of distribution,\nwe propose to completely eliminate this bottleneck by removing all BN layers in\nAT. Our normalizer-free robust training (NoFrost) method extends recent\nadvances in normalizer-free networks to AT for its unexplored advantage on\nhandling the mixture distribution challenge. We show that NoFrost achieves\nadversarial robustness with only a minor sacrifice on clean sample accuracy. On\nImageNet with ResNet50, NoFrost achieves $74.06\\%$ clean accuracy, which drops\nmerely $2.00\\%$ from standard training. In contrast, BN-based AT obtains\n$59.28\\%$ clean accuracy, suffering a significant $16.78\\%$ drop from standard\ntraining. In addition, NoFrost achieves a $23.56\\%$ adversarial robustness\nagainst PGD attack, which improves the $13.57\\%$ robustness in BN-based AT. We\nobserve better model smoothness and larger decision margins from NoFrost, which\nmake the models less sensitive to input perturbations and thus more robust.\nMoreover, when incorporating more data augmentations into NoFrost, it achieves\ncomprehensive robustness against multiple distribution shifts. Code and\npre-trained models are public at\nhttps://github.com/amazon-research/normalizer-free-robust-training.\n", "date": "2022-07-04", "categories": []}
{"id": "2207.01396", "title": "Hessian-Free Second-Order Adversarial Examples for Adversarial Learning", "abstract": "  Recent studies show deep neural networks (DNNs) are extremely vulnerable to\nthe elaborately designed adversarial examples. Adversarial learning with those\nadversarial examples has been proved as one of the most effective methods to\ndefend against such an attack. At present, most existing adversarial examples\ngeneration methods are based on first-order gradients, which can hardly further\nimprove models' robustness, especially when facing second-order adversarial\nattacks. Compared with first-order gradients, second-order gradients provide a\nmore accurate approximation of the loss landscape with respect to natural\nexamples. Inspired by this, our work crafts second-order adversarial examples\nand uses them to train DNNs. Nevertheless, second-order optimization involves\ntime-consuming calculation for Hessian-inverse. We propose an approximation\nmethod through transforming the problem into an optimization in the Krylov\nsubspace, which remarkably reduce the computational complexity to speed up the\ntraining procedure. Extensive experiments conducted on the MINIST and CIFAR-10\ndatasets show that our adversarial learning with second-order adversarial\nexamples outperforms other fisrt-order methods, which can improve the model\nrobustness against a wide range of attacks.\n", "date": "2022-07-04", "categories": []}
{"id": "2206.15415", "title": "MEAD: A Multi-Armed Approach for Evaluation of Adversarial Examples\n  Detectors", "abstract": "  Detection of adversarial examples has been a hot topic in the last years due\nto its importance for safely deploying machine learning algorithms in critical\napplications. However, the detection methods are generally validated by\nassuming a single implicitly known attack strategy, which does not necessarily\naccount for real-life threats. Indeed, this can lead to an overoptimistic\nassessment of the detectors' performance and may induce some bias in the\ncomparison between competing detection schemes. We propose a novel multi-armed\nframework, called MEAD, for evaluating detectors based on several attack\nstrategies to overcome this limitation. Among them, we make use of three new\nobjectives to generate attacks. The proposed performance metric is based on the\nworst-case scenario: detection is successful if and only if all different\nattacks are correctly recognized. Empirically, we show the effectiveness of our\napproach. Moreover, the poor performance obtained for state-of-the-art\ndetectors opens a new exciting line of research.\n", "date": "2022-06-30", "categories": []}
{"id": "2206.04310", "title": "GSmooth: Certified Robustness against Semantic Transformations via\n  Generalized Randomized Smoothing", "abstract": "  Certified defenses such as randomized smoothing have shown promise towards\nbuilding reliable machine learning systems against $\\ell_p$-norm bounded\nattacks. However, existing methods are insufficient or unable to provably\ndefend against semantic transformations, especially those without closed-form\nexpressions (such as defocus blur and pixelate), which are more common in\npractice and often unrestricted. To fill up this gap, we propose generalized\nrandomized smoothing (GSmooth), a unified theoretical framework for certifying\nrobustness against general semantic transformations via a novel dimension\naugmentation strategy. Under the GSmooth framework, we present a scalable\nalgorithm that uses a surrogate image-to-image network to approximate the\ncomplex transformation. The surrogate model provides a powerful tool for\nstudying the properties of semantic transformations and certifying robustness.\nExperimental results on several datasets demonstrate the effectiveness of our\napproach for robustness certification against multiple kinds of semantic\ntransformations and corruptions, which is not achievable by the alternative\nbaselines.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.14272", "title": "Collecting high-quality adversarial data for machine reading\n  comprehension tasks with humans and models in the loop", "abstract": "  We present our experience as annotators in the creation of high-quality,\nadversarial machine-reading-comprehension data for extractive QA for Task 1 of\nthe First Workshop on Dynamic Adversarial Data Collection (DADC). DADC is an\nemergent data collection paradigm with both models and humans in the loop. We\nset up a quasi-experimental annotation design and perform quantitative analyses\nacross groups with different numbers of annotators focusing on successful\nadversarial attacks, cost analysis, and annotator confidence correlation. We\nfurther perform a qualitative analysis of our perceived difficulty of the task\ngiven the different topics of the passages in our dataset and conclude with\nrecommendations and suggestions that might be of value to people working on\nfuture DADC tasks and related annotation interfaces.\n", "date": "2022-06-28", "categories": []}
{"id": "2206.14729", "title": "longhorns at DADC 2022: How many linguists does it take to fool a\n  Question Answering model? A systematic approach to adversarial attacks", "abstract": "  Developing methods to adversarially challenge NLP systems is a promising\navenue for improving both model performance and interpretability. Here, we\ndescribe the approach of the team \"longhorns\" on Task 1 of the The First\nWorkshop on Dynamic Adversarial Data Collection (DADC), which asked teams to\nmanually fool a model on an Extractive Question Answering task. Our team\nfinished first, with a model error rate of 62%. We advocate for a systematic,\nlinguistically informed approach to formulating adversarial questions, and we\ndescribe the results of our pilot experiments, as well as our official\nsubmission.\n", "date": "2022-06-29", "categories": []}
{"id": "2206.13991", "title": "Increasing Confidence in Adversarial Robustness Evaluations", "abstract": "  Hundreds of defenses have been proposed to make deep neural networks robust\nagainst minimal (adversarial) input perturbations. However, only a handful of\nthese defenses held up their claims because correctly evaluating robustness is\nextremely challenging: Weak attacks often fail to find adversarial examples\neven if they unknowingly exist, thereby making a vulnerable network look\nrobust. In this paper, we propose a test to identify weak attacks, and thus\nweak defense evaluations. Our test slightly modifies a neural network to\nguarantee the existence of an adversarial example for every sample.\nConsequentially, any correct attack must succeed in breaking this modified\nnetwork. For eleven out of thirteen previously-published defenses, the original\nevaluation of the defense fails our test, while stronger attacks that break\nthese defenses pass it. We hope that attack unit tests - such as ours - will be\na major component in future robustness evaluations and increase confidence in\nan empirical field that is currently riddled with skepticism.\n", "date": "2022-06-28", "categories": []}
{"id": "2206.14020", "title": "Rethinking Adversarial Examples for Location Privacy Protection", "abstract": "  We have investigated a new application of adversarial examples, namely\nlocation privacy protection against landmark recognition systems. We introduce\nmask-guided multimodal projected gradient descent (MM-PGD), in which\nadversarial examples are trained on different deep models. Image contents are\nprotected by analyzing the properties of regions to identify the ones most\nsuitable for blending in adversarial examples. We investigated two region\nidentification strategies: class activation map-based MM-PGD, in which the\ninternal behaviors of trained deep models are targeted; and human-vision-based\nMM-PGD, in which regions that attract less human attention are targeted.\nExperiments on the Places365 dataset demonstrated that these strategies are\npotentially effective in defending against black-box landmark recognition\nsystems without the need for much image manipulation.\n", "date": "2022-06-28", "categories": []}
{"id": "2206.12590", "title": "RSTAM: An Effective Black-Box Impersonation Attack on Face Recognition\n  using a Mobile and Compact Printer", "abstract": "  Face recognition has achieved considerable progress in recent years thanks to\nthe development of deep neural networks, but it has recently been discovered\nthat deep neural networks are vulnerable to adversarial examples. This means\nthat face recognition models or systems based on deep neural networks are also\nsusceptible to adversarial examples. However, the existing methods of attacking\nface recognition models or systems with adversarial examples can effectively\ncomplete white-box attacks but not black-box impersonation attacks, physical\nattacks, or convenient attacks, particularly on commercial face recognition\nsystems. In this paper, we propose a new method to attack face recognition\nmodels or systems called RSTAM, which enables an effective black-box\nimpersonation attack using an adversarial mask printed by a mobile and compact\nprinter. First, RSTAM enhances the transferability of the adversarial masks\nthrough our proposed random similarity transformation strategy. Furthermore, we\npropose a random meta-optimization strategy for ensembling several pre-trained\nface models to generate more general adversarial masks. Finally, we conduct\nexperiments on the CelebA-HQ, LFW, Makeup Transfer (MT), and CASIA-FaceV5\ndatasets. The performance of the attacks is also evaluated on state-of-the-art\ncommercial face recognition systems: Face++, Baidu, Aliyun, Tencent, and\nMicrosoft. Extensive experiments show that RSTAM can effectively perform\nblack-box impersonation attacks on face recognition models or systems.\n", "date": "2022-06-25", "categories": []}
{"id": "2206.12685", "title": "Defense against adversarial attacks on deep convolutional neural\n  networks through nonlocal denoising", "abstract": "  Despite substantial advances in network architecture performance, the\nsusceptibility of adversarial attacks makes deep learning challenging to\nimplement in safety-critical applications. This paper proposes a data-centric\napproach to addressing this problem. A nonlocal denoising method with different\nluminance values has been used to generate adversarial examples from the\nModified National Institute of Standards and Technology database (MNIST) and\nCanadian Institute for Advanced Research (CIFAR-10) data sets. Under\nperturbation, the method provided absolute accuracy improvements of up to 9.3%\nin the MNIST data set and 13% in the CIFAR-10 data set. Training using\ntransformed images with higher luminance values increases the robustness of the\nclassifier. We have shown that transfer learning is disadvantageous for\nadversarial machine learning. The results indicate that simple adversarial\nexamples can improve resilience and make deep learning easier to apply in\nvarious applications.\n", "date": "2022-06-25", "categories": []}
{"id": "2206.12714", "title": "Defending Multimodal Fusion Models against Single-Source Adversaries", "abstract": "  Beyond achieving high performance across many vision tasks, multimodal models\nare expected to be robust to single-source faults due to the availability of\nredundant information between modalities. In this paper, we investigate the\nrobustness of multimodal neural networks against worst-case (i.e., adversarial)\nperturbations on a single modality. We first show that standard multimodal\nfusion models are vulnerable to single-source adversaries: an attack on any\nsingle modality can overcome the correct information from multiple unperturbed\nmodalities and cause the model to fail. This surprising vulnerability holds\nacross diverse multimodal tasks and necessitates a solution. Motivated by this\nfinding, we propose an adversarially robust fusion strategy that trains the\nmodel to compare information coming from all the input sources, detect\ninconsistencies in the perturbed modality compared to the other modalities, and\nonly allow information from the unperturbed modalities to pass through. Our\napproach significantly improves on state-of-the-art methods in single-source\nrobustness, achieving gains of 7.8-25.2% on action recognition, 19.7-48.2% on\nobject detection, and 1.6-6.7% on sentiment analysis, without degrading\nperformance on unperturbed (i.e., clean) data.\n", "date": "2022-06-25", "categories": []}
{"id": "2206.12725", "title": "Empirical Evaluation of Physical Adversarial Patch Attacks Against\n  Overhead Object Detection Models", "abstract": "  Adversarial patches are images designed to fool otherwise well-performing\nneural network-based computer vision models. Although these attacks were\ninitially conceived of and studied digitally, in that the raw pixel values of\nthe image were perturbed, recent work has demonstrated that these attacks can\nsuccessfully transfer to the physical world. This can be accomplished by\nprinting out the patch and adding it into scenes of newly captured images or\nvideo footage. In this work we further test the efficacy of adversarial patch\nattacks in the physical world under more challenging conditions. We consider\nobject detection models trained on overhead imagery acquired through aerial or\nsatellite cameras, and we test physical adversarial patches inserted into\nscenes of a desert environment. Our main finding is that it is far more\ndifficult to successfully implement the adversarial patch attacks under these\nconditions than in the previously considered conditions. This has important\nimplications for AI safety as the real-world threat posed by adversarial\nexamples may be overstated.\n", "date": "2022-06-25", "categories": []}
{"id": "2206.12963", "title": "Self-Healing Robust Neural Networks via Closed-Loop Control", "abstract": "  Despite the wide applications of neural networks, there have been increasing\nconcerns about their vulnerability issue. While numerous attack and defense\ntechniques have been developed, this work investigates the robustness issue\nfrom a new angle: can we design a self-healing neural network that can\nautomatically detect and fix the vulnerability issue by itself? A typical\nself-healing mechanism is the immune system of a human body. This\nbiology-inspired idea has been used in many engineering designs but is rarely\ninvestigated in deep learning. This paper considers the post-training\nself-healing of a neural network, and proposes a closed-loop control\nformulation to automatically detect and fix the errors caused by various\nattacks or perturbations. We provide a margin-based analysis to explain how\nthis formulation can improve the robustness of a classifier. To speed up the\ninference of the proposed self-healing network, we solve the control problem\nvia improving the Pontryagin Maximum Principle-based solver. Lastly, we present\nan error estimation of the proposed framework for neural networks with\nnonlinear activation functions. We validate the performance on several network\narchitectures against various perturbations. Since the self-healing method does\nnot need a-priori information about data perturbations/attacks, it can handle a\nbroad class of unforeseen perturbations.\n", "date": "2022-06-26", "categories": []}
{"id": "2206.13083", "title": "Adversarial Example Detection in Deployed Tree Ensembles", "abstract": "  Tree ensembles are powerful models that are widely used. However, they are\nsusceptible to adversarial examples, which are examples that purposely\nconstructed to elicit a misprediction from the model. This can degrade\nperformance and erode a user's trust in the model. Typically, approaches try to\nalleviate this problem by verifying how robust a learned ensemble is or\nrobustifying the learning process. We take an alternative approach and attempt\nto detect adversarial examples in a post-deployment setting. We present a novel\nmethod for this task that works by analyzing an unseen example's output\nconfiguration, which is the set of predictions made by an ensemble's\nconstituent trees. Our approach works with any additive tree ensemble and does\nnot require training a separate model. We evaluate our approach on three\ndifferent tree ensemble learners. We empirically show that our method is\ncurrently the best adversarial detection method for tree ensembles.\n", "date": "2022-06-27", "categories": []}
{"id": "2206.12169", "title": "AdAUC: End-to-end Adversarial AUC Optimization Against Long-tail\n  Problems", "abstract": "  It is well-known that deep learning models are vulnerable to adversarial\nexamples. Existing studies of adversarial training have made great progress\nagainst this challenge. As a typical trait, they often assume that the class\ndistribution is overall balanced. However, long-tail datasets are ubiquitous in\na wide spectrum of applications, where the amount of head class instances is\nlarger than the tail classes. Under such a scenario, AUC is a much more\nreasonable metric than accuracy since it is insensitive toward class\ndistribution. Motivated by this, we present an early trial to explore\nadversarial training methods to optimize AUC. The main challenge lies in that\nthe positive and negative examples are tightly coupled in the objective\nfunction. As a direct result, one cannot generate adversarial examples without\na full scan of the dataset. To address this issue, based on a concavity\nregularization scheme, we reformulate the AUC optimization problem as a saddle\npoint problem, where the objective becomes an instance-wise function. This\nleads to an end-to-end training protocol. Furthermore, we provide a convergence\nguarantee of the proposed algorithm. Our analysis differs from the existing\nstudies since the algorithm is asked to generate adversarial examples by\ncalculating the gradient of a min-max problem. Finally, the extensive\nexperimental results show the performance and robustness of our algorithm in\nthree long-tail datasets.\n", "date": "2022-06-24", "categories": []}
{"id": "2206.12292", "title": "InfoAT: Improving Adversarial Training Using the Information Bottleneck\n  Principle", "abstract": "  Adversarial training (AT) has shown excellent high performance in defending\nagainst adversarial examples. Recent studies demonstrate that examples are not\nequally important to the final robustness of models during AT, that is, the\nso-called hard examples that can be attacked easily exhibit more influence than\nrobust examples on the final robustness. Therefore, guaranteeing the robustness\nof hard examples is crucial for improving the final robustness of the model.\nHowever, defining effective heuristics to search for hard examples is still\ndifficult. In this article, inspired by the information bottleneck (IB)\nprinciple, we uncover that an example with high mutual information of the input\nand its associated latent representation is more likely to be attacked. Based\non this observation, we propose a novel and effective adversarial training\nmethod (InfoAT). InfoAT is encouraged to find examples with high mutual\ninformation and exploit them efficiently to improve the final robustness of\nmodels. Experimental results show that InfoAT achieves the best robustness\namong different datasets and models in comparison with several state-of-the-art\nmethods.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.11433", "title": "Shilling Black-box Recommender Systems by Learning to Generate Fake User\n  Profiles", "abstract": "  Due to the pivotal role of Recommender Systems (RS) in guiding customers\ntowards the purchase, there is a natural motivation for unscrupulous parties to\nspoof RS for profits. In this paper, we study Shilling Attack where an\nadversarial party injects a number of fake user profiles for improper purposes.\nConventional Shilling Attack approaches lack attack transferability (i.e.,\nattacks are not effective on some victim RS models) and/or attack invisibility\n(i.e., injected profiles can be easily detected). To overcome these issues, we\npresent Leg-UP, a novel attack model based on the Generative Adversarial\nNetwork. Leg-UP learns user behavior patterns from real users in the sampled\n``templates'' and constructs fake user profiles. To simulate real users, the\ngenerator in Leg-UP directly outputs discrete ratings. To enhance attack\ntransferability, the parameters of the generator are optimized by maximizing\nthe attack performance on a surrogate RS model. To improve attack invisibility,\nLeg-UP adopts a discriminator to guide the generator to generate undetectable\nfake user profiles. Experiments on benchmarks have shown that Leg-UP exceeds\nstate-of-the-art Shilling Attack methods on a wide range of victim RS models.\nThe source code of our work is available at:\nhttps://github.com/XMUDM/ShillingAttack.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.11480", "title": "A Framework for Understanding Model Extraction Attack and Defense", "abstract": "  The privacy of machine learning models has become a significant concern in\nmany emerging Machine-Learning-as-a-Service applications, where prediction\nservices based on well-trained models are offered to users via pay-per-query.\nThe lack of a defense mechanism can impose a high risk on the privacy of the\nserver's model since an adversary could efficiently steal the model by querying\nonly a few `good' data points. The interplay between a server's defense and an\nadversary's attack inevitably leads to an arms race dilemma, as commonly seen\nin Adversarial Machine Learning. To study the fundamental tradeoffs between\nmodel utility from a benign user's view and privacy from an adversary's view,\nwe develop new metrics to quantify such tradeoffs, analyze their theoretical\nproperties, and develop an optimization problem to understand the optimal\nadversarial attack and defense strategies. The developed concepts and theory\nmatch the empirical findings on the `equilibrium' between privacy and utility.\nIn terms of optimization, the key ingredient that enables our results is a\nunified representation of the attack-defense problem as a min-max bi-level\nproblem. The developed results will be demonstrated by examples and\nexperiments.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.11724", "title": "BERT Rankers are Brittle: a Study using Adversarial Document\n  Perturbations", "abstract": "  Contextual ranking models based on BERT are now well established for a wide\nrange of passage and document ranking tasks. However, the robustness of\nBERT-based ranking models under adversarial inputs is under-explored. In this\npaper, we argue that BERT-rankers are not immune to adversarial attacks\ntargeting retrieved documents given a query. Firstly, we propose algorithms for\nadversarial perturbation of both highly relevant and non-relevant documents\nusing gradient-based optimization methods. The aim of our algorithms is to\nadd/replace a small number of tokens to a highly relevant or non-relevant\ndocument to cause a large rank demotion or promotion. Our experiments show that\na small number of tokens can already result in a large change in the rank of a\ndocument. Moreover, we find that BERT-rankers heavily rely on the document\nstart/head for relevance prediction, making the initial part of the document\nmore susceptible to adversarial attacks. More interestingly, we find a small\nset of recurring adversarial words that when added to documents result in\nsuccessful rank demotion/promotion of any relevant/non-relevant document\nrespectively. Finally, our adversarial tokens also show particular topic\npreferences within and across datasets, exposing potential biases from BERT\npre-training or downstream datasets.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.01736", "title": "Adaptive Adversarial Training to Improve Adversarial Robustness of DNNs\n  for Medical Image Segmentation and Detection", "abstract": "  It is known that Deep Neural networks (DNNs) are vulnerable to adversarial\nattacks, and the adversarial robustness of DNNs could be improved by adding\nadversarial noises to training data (e.g., the standard adversarial training\n(SAT)). However, inappropriate noises added to training data may reduce a\nmodel's performance, which is termed the trade-off between accuracy and\nrobustness. This problem has been sufficiently studied for the classification\nof whole images but has rarely been explored for image analysis tasks in the\nmedical application domain, including image segmentation, landmark detection,\nand object detection tasks. In this study, we show that, for those medical\nimage analysis tasks, the SAT method has a severe issue that limits its\npractical use: it generates a fixed and unified level of noise for all training\nsamples for robust DNN training. A high noise level may lead to a large\nreduction in model performance and a low noise level may not be effective in\nimproving robustness. To resolve this issue, we design an adaptive-margin\nadversarial training (AMAT) method that generates sample-wise adaptive\nadversarial noises for robust DNN training. In contrast to the existing,\nclassification-oriented adversarial training methods, our AMAT method uses a\nloss-defined-margin strategy so that it can be applied to different tasks as\nlong as the loss functions are well-defined. We successfully apply our AMAT\nmethod to state-of-the-art DNNs, using five publicly available datasets. The\nexperimental results demonstrate that: (1) our AMAT method can be applied to\nthe three seemingly different tasks in the medical image application domain;\n(2) AMAT outperforms the SAT method in adversarial robustness; (3) AMAT has a\nminimal reduction in prediction accuracy on clean data, compared with the SAT\nmethod; and (4) AMAT has almost the same training time cost as SAT.\n", "date": "2022-06-02", "categories": []}
{"id": "2206.08675", "title": "Understanding Robust Overfitting of Adversarial Training and Beyond", "abstract": "  Robust overfitting widely exists in adversarial training of deep networks.\nThe exact underlying reasons for this are still not completely understood.\nHere, we explore the causes of robust overfitting by comparing the data\ndistribution of \\emph{non-overfit} (weak adversary) and \\emph{overfitted}\n(strong adversary) adversarial training, and observe that the distribution of\nthe adversarial data generated by weak adversary mainly contain small-loss\ndata. However, the adversarial data generated by strong adversary is more\ndiversely distributed on the large-loss data and the small-loss data. Given\nthese observations, we further designed data ablation adversarial training and\nidentify that some small-loss data which are not worthy of the adversary\nstrength cause robust overfitting in the strong adversary mode. To relieve this\nissue, we propose \\emph{minimum loss constrained adversarial training} (MLCAT):\nin a minibatch, we learn large-loss data as usual, and adopt additional\nmeasures to increase the loss of the small-loss data. Technically, MLCAT\nhinders data fitting when they become easy to learn to prevent robust\noverfitting; philosophically, MLCAT reflects the spirit of turning waste into\ntreasure and making the best use of each adversarial data; algorithmically, we\ndesigned two realizations of MLCAT, and extensive experiments demonstrate that\nMLCAT can eliminate robust overfitting and further boost adversarial\nrobustness.\n", "date": "2022-06-17", "categories": []}
{"id": "2206.10673", "title": "Natural Backdoor Datasets", "abstract": "  Extensive literature on backdoor poison attacks has studied attacks and\ndefenses for backdoors using \"digital trigger patterns.\" In contrast, \"physical\nbackdoors\" use physical objects as triggers, have only recently been\nidentified, and are qualitatively different enough to resist all defenses\ntargeting digital trigger backdoors. Research on physical backdoors is limited\nby access to large datasets containing real images of physical objects\nco-located with targets of classification. Building these datasets is time- and\nlabor-intensive. This works seeks to address the challenge of accessibility for\nresearch on physical backdoor attacks. We hypothesize that there may be\nnaturally occurring physically co-located objects already present in popular\ndatasets such as ImageNet. Once identified, a careful relabeling of these data\ncan transform them into training samples for physical backdoor attacks. We\npropose a method to scalably identify these subsets of potential triggers in\nexisting datasets, along with the specific classes they can poison. We call\nthese naturally occurring trigger-class subsets natural backdoor datasets. Our\ntechniques successfully identify natural backdoors in widely-available\ndatasets, and produce models behaviorally equivalent to those trained on\nmanually curated datasets. We release our code to allow the research community\nto create their own datasets for research on physical backdoor attacks.\n", "date": "2022-06-21", "categories": []}
{"id": "2206.10809", "title": "SSMI: How to Make Objects of Interest Disappear without Accessing Object\n  Detectors?", "abstract": "  Most black-box adversarial attack schemes for object detectors mainly face\ntwo shortcomings: requiring access to the target model and generating\ninefficient adversarial examples (failing to make objects disappear in large\nnumbers). To overcome these shortcomings, we propose a black-box adversarial\nattack scheme based on semantic segmentation and model inversion (SSMI). We\nfirst locate the position of the target object using semantic segmentation\ntechniques. Next, we design a neighborhood background pixel replacement to\nreplace the target region pixels with background pixels to ensure that the\npixel modifications are not easily detected by human vision. Finally, we\nreconstruct a machine-recognizable example and use the mask matrix to select\npixels in the reconstructed example to modify the benign image to generate an\nadversarial example. Detailed experimental results show that SSMI can generate\nefficient adversarial examples to evade human-eye perception and make objects\nof interest disappear. And more importantly, SSMI outperforms existing same\nkinds of attacks. The maximum increase in new and disappearing labels is 16%,\nand the maximum decrease in mAP metrics for object detection is 36%.\n", "date": "2022-06-22", "categories": []}
{"id": "2206.10875", "title": "Guided Diffusion Model for Adversarial Purification from Random Noise", "abstract": "  In this paper, we propose a novel guided diffusion purification approach to\nprovide a strong defense against adversarial attacks. Our model achieves 89.62%\nrobust accuracy under PGD-L_inf attack (eps = 8/255) on the CIFAR-10 dataset.\nWe first explore the essential correlations between unguided diffusion models\nand randomized smoothing, enabling us to apply the models to certified\nrobustness. The empirical results show that our models outperform randomized\nsmoothing by 5% when the certified L2 radius r is larger than 0.5.\n", "date": "2022-06-22", "categories": []}
{"id": "2206.10988", "title": "AdvSmo: Black-box Adversarial Attack by Smoothing Linear Structure of\n  Texture", "abstract": "  Black-box attacks usually face two problems: poor transferability and the\ninability to evade the adversarial defense. To overcome these shortcomings, we\ncreate an original approach to generate adversarial examples by smoothing the\nlinear structure of the texture in the benign image, called AdvSmo. We\nconstruct the adversarial examples without relying on any internal information\nto the target model and design the imperceptible-high attack success rate\nconstraint to guide the Gabor filter to select appropriate angles and scales to\nsmooth the linear texture from the input images to generate adversarial\nexamples. Benefiting from the above design concept, AdvSmo will generate\nadversarial examples with strong transferability and solid evasiveness.\nFinally, compared to the four advanced black-box adversarial attack methods,\nfor the eight target models, the results show that AdvSmo improves the average\nattack success rate by 9% on the CIFAR-10 and 16% on the Tiny-ImageNet dataset\ncompared to the best of these attack methods.\n", "date": "2022-06-22", "categories": []}
{"id": "2206.11225", "title": "RetrievalGuard: Provably Robust 1-Nearest Neighbor Image Retrieval", "abstract": "  Recent research works have shown that image retrieval models are vulnerable\nto adversarial attacks, where slightly modified test inputs could lead to\nproblematic retrieval results. In this paper, we aim to design a provably\nrobust image retrieval model which keeps the most important evaluation metric\nRecall@1 invariant to adversarial perturbation. We propose the first 1-nearest\nneighbor (NN) image retrieval algorithm, RetrievalGuard, which is provably\nrobust against adversarial perturbations within an $\\ell_2$ ball of calculable\nradius. The challenge is to design a provably robust algorithm that takes into\nconsideration the 1-NN search and the high-dimensional nature of the embedding\nspace. Algorithmically, given a base retrieval model and a query sample, we\nbuild a smoothed retrieval model by carefully analyzing the 1-NN search\nprocedure in the high-dimensional embedding space. We show that the smoothed\nretrieval model has bounded Lipschitz constant and thus the retrieval score is\ninvariant to $\\ell_2$ adversarial perturbations. Experiments on image retrieval\ntasks validate the robustness of our RetrievalGuard method.\n", "date": "2022-06-17", "categories": []}
{"id": "2206.11228", "title": "Adversarially trained neural representations may already be as robust as\n  corresponding biological neural representations", "abstract": "  Visual systems of primates are the gold standard of robust perception. There\nis thus a general belief that mimicking the neural representations that\nunderlie those systems will yield artificial visual systems that are\nadversarially robust. In this work, we develop a method for performing\nadversarial visual attacks directly on primate brain activity. We then leverage\nthis method to demonstrate that the above-mentioned belief might not be well\nfounded. Specifically, we report that the biological neurons that make up\nvisual systems of primates exhibit susceptibility to adversarial perturbations\nthat is comparable in magnitude to existing (robustly trained) artificial\nneural networks.\n", "date": "2022-06-19", "categories": []}
{"id": "2206.09075", "title": "Comment on Transferability and Input Transformation with Additive Noise", "abstract": "  Adversarial attacks have verified the existence of the vulnerability of\nneural networks. By adding small perturbations to a benign example, adversarial\nattacks successfully generate adversarial examples that lead misclassification\nof deep learning models. More importantly, an adversarial example generated\nfrom a specific model can also deceive other models without modification. We\ncall this phenomenon ``transferability\". Here, we analyze the relationship\nbetween transferability and input transformation with additive noise by\nmathematically proving that the modified optimization can produce more\ntransferable adversarial examples.\n", "date": "2022-06-18", "categories": []}
{"id": "2206.09238", "title": "On the Role of Generalization in Transferability of Adversarial Examples", "abstract": "  Black-box adversarial attacks designing adversarial examples for unseen\nneural networks (NNs) have received great attention over the past years. While\nseveral successful black-box attack schemes have been proposed in the\nliterature, the underlying factors driving the transferability of black-box\nadversarial examples still lack a thorough understanding. In this paper, we aim\nto demonstrate the role of the generalization properties of the substitute\nclassifier used for generating adversarial examples in the transferability of\nthe attack scheme to unobserved NN classifiers. To do this, we apply the\nmax-min adversarial example game framework and show the importance of the\ngeneralization properties of the substitute NN in the success of the black-box\nattack scheme in application to different NN classifiers. We prove theoretical\ngeneralization bounds on the difference between the attack transferability\nrates on training and test samples. Our bounds suggest that a substitute NN\nwith better generalization behavior could result in more transferable\nadversarial examples. In addition, we show that standard operator norm-based\nregularization methods could improve the transferability of the designed\nadversarial examples. We support our theoretical results by performing several\nnumerical experiments showing the role of the substitute network's\ngeneralization in generating transferable adversarial examples. Our empirical\nresults indicate the power of Lipschitz regularization methods in improving the\ntransferability of adversarial examples.\n", "date": "2022-06-18", "categories": []}
{"id": "2206.09458", "title": "A Universal Adversarial Policy for Text Classifiers", "abstract": "  Discovering the existence of universal adversarial perturbations had large\ntheoretical and practical impacts on the field of adversarial learning. In the\ntext domain, most universal studies focused on adversarial prefixes which are\nadded to all texts. However, unlike the vision domain, adding the same\nperturbation to different inputs results in noticeably unnatural inputs.\nTherefore, we introduce a new universal adversarial setup - a universal\nadversarial policy, which has many advantages of other universal attacks but\nalso results in valid texts - thus making it relevant in practice. We achieve\nthis by learning a single search policy over a predefined set of semantics\npreserving text alterations, on many texts. This formulation is universal in\nthat the policy is successful in finding adversarial examples on new texts\nefficiently. Our approach uses text perturbations which were extensively shown\nto produce natural attacks in the non-universal setup (specific synonym\nreplacements). We suggest a strong baseline approach for this formulation which\nuses reinforcement learning. It's ability to generalise (from as few as 500\ntraining texts) shows that universal adversarial patterns exist in the text\ndomain as well.\n", "date": "2022-06-19", "categories": []}
{"id": "2206.10334", "title": "Defending Adversarial Examples by Negative Correlation Ensemble", "abstract": "  The security issues in DNNs, such as adversarial examples, have attracted\nmuch attention. Adversarial examples refer to the examples which are capable to\ninduce the DNNs return completely predictions by introducing carefully designed\nperturbations. Obviously, adversarial examples bring great security risks to\nthe development of deep learning. Recently, Some defense approaches against\nadversarial examples have been proposed, however, in our opinion, the\nperformance of these approaches are still limited. In this paper, we propose a\nnew ensemble defense approach named the Negative Correlation Ensemble (NCEn),\nwhich achieves compelling results by introducing gradient directions and\ngradient magnitudes of each member in the ensemble negatively correlated and at\nthe same time, reducing the transferability of adversarial examples among them.\nExtensive experiments have been conducted, and the results demonstrate that\nNCEn can improve the adversarial robustness of ensembles effectively.\n", "date": "2022-06-11", "categories": []}
{"id": "2206.08477", "title": "Backdoor Attacks on Vision Transformers", "abstract": "  Vision Transformers (ViT) have recently demonstrated exemplary performance on\na variety of vision tasks and are being used as an alternative to CNNs. Their\ndesign is based on a self-attention mechanism that processes images as a\nsequence of patches, which is quite different compared to CNNs. Hence it is\ninteresting to study if ViTs are vulnerable to backdoor attacks. Backdoor\nattacks happen when an attacker poisons a small part of the training data for\nmalicious purposes. The model performance is good on clean test images, but the\nattacker can manipulate the decision of the model by showing the trigger at\ntest time. To the best of our knowledge, we are the first to show that ViTs are\nvulnerable to backdoor attacks. We also find an intriguing difference between\nViTs and CNNs - interpretation algorithms effectively highlight the trigger on\ntest images for ViTs but not for CNNs. Based on this observation, we propose a\ntest-time image blocking defense for ViTs which reduces the attack success rate\nby a large margin. Code is available here:\nhttps://github.com/UCDvision/backdoor_transformer.git\n", "date": "2022-06-16", "categories": []}
{"id": "2206.08575", "title": "Query-Efficient and Scalable Black-Box Adversarial Attacks on Discrete\n  Sequential Data via Bayesian Optimization", "abstract": "  We focus on the problem of adversarial attacks against models on discrete\nsequential data in the black-box setting where the attacker aims to craft\nadversarial examples with limited query access to the victim model. Existing\nblack-box attacks, mostly based on greedy algorithms, find adversarial examples\nusing pre-computed key positions to perturb, which severely limits the search\nspace and might result in suboptimal solutions. To this end, we propose a\nquery-efficient black-box attack using Bayesian optimization, which dynamically\ncomputes important positions using an automatic relevance determination (ARD)\ncategorical kernel. We introduce block decomposition and history subsampling\ntechniques to improve the scalability of Bayesian optimization when an input\nsequence becomes long. Moreover, we develop a post-optimization algorithm that\nfinds adversarial examples with smaller perturbation size. Experiments on\nnatural language and protein classification tasks demonstrate that our method\nconsistently achieves higher attack success rate with significant reduction in\nquery count and modification rate compared to the previous state-of-the-art\nmethods.\n", "date": "2022-06-17", "categories": []}
{"id": "2206.08638", "title": "Minimum Noticeable Difference based Adversarial Privacy Preserving Image\n  Generation", "abstract": "  Deep learning models are found to be vulnerable to adversarial examples, as\nwrong predictions can be caused by small perturbation in input for deep\nlearning models. Most of the existing works of adversarial image generation try\nto achieve attacks for most models, while few of them make efforts on\nguaranteeing the perceptual quality of the adversarial examples. High quality\nadversarial examples matter for many applications, especially for the privacy\npreserving. In this work, we develop a framework based on the Minimum\nNoticeable Difference (MND) concept to generate adversarial privacy preserving\nimages that have minimum perceptual difference from the clean ones but are able\nto attack deep learning models. To achieve this, an adversarial loss is firstly\nproposed to make the deep learning models attacked by the adversarial images\nsuccessfully. Then, a perceptual quality-preserving loss is developed by taking\nthe magnitude of perturbation and perturbation-caused structural and gradient\nchanges into account, which aims to preserve high perceptual quality for\nadversarial image generation. To the best of our knowledge, this is the first\nwork on exploring quality-preserving adversarial image generation based on the\nMND concept for privacy preserving. To evaluate its performance in terms of\nperceptual quality, the deep models on image classification and face\nrecognition are tested with the proposed method and several anchor methods in\nthis work. Extensive experimental results demonstrate that the proposed MND\nframework is capable of generating adversarial images with remarkably improved\nperformance metrics (e.g., PSNR, SSIM, and MOS) than that generated with the\nanchor methods.\n", "date": "2022-06-17", "categories": []}
{"id": "2206.08738", "title": "Detecting Adversarial Examples in Batches -- a geometrical approach", "abstract": "  Many deep learning methods have successfully solved complex tasks in computer\nvision and speech recognition applications. Nonetheless, the robustness of\nthese models has been found to be vulnerable to perturbed inputs or adversarial\nexamples, which are imperceptible to the human eye, but lead the model to\nerroneous output decisions. In this study, we adapt and introduce two geometric\nmetrics, density and coverage, and evaluate their use in detecting adversarial\nsamples in batches of unseen data. We empirically study these metrics using\nMNIST and two real-world biomedical datasets from MedMNIST, subjected to two\ndifferent adversarial attacks. Our experiments show promising results for both\nmetrics to detect adversarial examples. We believe that his work can lay the\nground for further study on these metrics' use in deployed machine learning\nsystems to monitor for possible attacks by adversarial examples or related\npathologies such as dataset shift.\n", "date": "2022-06-17", "categories": []}
{"id": "2206.08788", "title": "Is Multi-Modal Necessarily Better? Robustness Evaluation of Multi-modal\n  Fake News Detection", "abstract": "  The proliferation of fake news and its serious negative social influence push\nfake news detection methods to become necessary tools for web managers.\nMeanwhile, the multi-media nature of social media makes multi-modal fake news\ndetection popular for its ability to capture more modal features than uni-modal\ndetection methods. However, current literature on multi-modal detection is more\nlikely to pursue the detection accuracy but ignore the robustness of the\ndetector. To address this problem, we propose a comprehensive robustness\nevaluation of multi-modal fake news detectors. In this work, we simulate the\nattack methods of malicious users and developers, i.e., posting fake news and\ninjecting backdoors. Specifically, we evaluate multi-modal detectors with five\nadversarial and two backdoor attack methods. Experiment results imply that: (1)\nThe detection performance of the state-of-the-art detectors degrades\nsignificantly under adversarial attacks, even worse than general detectors; (2)\nMost multi-modal detectors are more vulnerable when subjected to attacks on\nvisual modality than textual modality; (3) Popular events' images will cause\nsignificant degradation to the detectors when they are subjected to backdoor\nattacks; (4) The performance of these detectors under multi-modal attacks is\nworse than under uni-modal attacks; (5) Defensive methods will improve the\nrobustness of the multi-modal detectors.\n", "date": "2022-06-17", "categories": []}
{"id": "2206.07840", "title": "Architectural Backdoors in Neural Networks", "abstract": "  Machine learning is vulnerable to adversarial manipulation. Previous\nliterature has demonstrated that at the training stage attackers can manipulate\ndata and data sampling procedures to control model behaviour. A common attack\ngoal is to plant backdoors i.e. force the victim model to learn to recognise a\ntrigger known only by the adversary. In this paper, we introduce a new class of\nbackdoor attacks that hide inside model architectures i.e. in the inductive\nbias of the functions used to train. These backdoors are simple to implement,\nfor instance by publishing open-source code for a backdoored model architecture\nthat others will reuse unknowingly. We demonstrate that model architectural\nbackdoors represent a real threat and, unlike other approaches, can survive a\ncomplete re-training from scratch. We formalise the main construction\nprinciples behind architectural backdoors, such as a link between the input and\nthe output, and describe some possible protections against them. We evaluate\nour attacks on computer vision benchmarks of different scales and demonstrate\nthe underlying vulnerability is pervasive in a variety of training settings.\n", "date": "2022-06-15", "categories": []}
{"id": "2206.07953", "title": "Analysis and Extensions of Adversarial Training for Video Classification", "abstract": "  Adversarial training (AT) is a simple yet effective defense against\nadversarial attacks to image classification systems, which is based on\naugmenting the training set with attacks that maximize the loss. However, the\neffectiveness of AT as a defense for video classification has not been\nthoroughly studied. Our first contribution is to show that generating optimal\nattacks for video requires carefully tuning the attack parameters, especially\nthe step size. Notably, we show that the optimal step size varies linearly with\nthe attack budget. Our second contribution is to show that using a smaller\n(sub-optimal) attack budget at training time leads to a more robust performance\nat test time. Based on these findings, we propose three defenses against\nattacks with variable attack budgets. The first one, Adaptive AT, is a\ntechnique where the attack budget is drawn from a distribution that is adapted\nas training iterations proceed. The second, Curriculum AT, is a technique where\nthe attack budget is increased as training iterations proceed. The third,\nGenerative AT, further couples AT with a denoising generative adversarial\nnetwork to boost robust performance. Experiments on the UCF101 dataset\ndemonstrate that the proposed methods improve adversarial robustness against\nmultiple attack types.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.08170", "title": "Adversarial Privacy Protection on Speech Enhancement", "abstract": "  Speech is easily leaked imperceptibly, such as being recorded by mobile\nphones in different situations. Private content in speech may be maliciously\nextracted through speech enhancement technology. Speech enhancement technology\nhas developed rapidly along with deep neural networks (DNNs), but adversarial\nexamples can cause DNNs to fail. In this work, we propose an adversarial method\nto degrade speech enhancement systems. Experimental results show that generated\nadversarial examples can erase most content information in original examples or\nreplace it with target speech content through speech enhancement. The word\nerror rate (WER) between an enhanced original example and enhanced adversarial\nexample recognition result can reach 89.0%. WER of target attack between\nenhanced adversarial example and target example is low to 33.75% . Adversarial\nperturbation can bring the rate of change to the original example to more than\n1.4430. This work can prevent the malicious extraction of speech.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.08304", "title": "Adversarial Patch Attacks and Defences in Vision-Based Tasks: A Survey", "abstract": "  Adversarial attacks in deep learning models, especially for safety-critical\nsystems, are gaining more and more attention in recent years, due to the lack\nof trust in the security and robustness of AI models. Yet the more primitive\nadversarial attacks might be physically infeasible or require some resources\nthat are hard to access like the training data, which motivated the emergence\nof patch attacks. In this survey, we provide a comprehensive overview to cover\nexisting techniques of adversarial patch attacks, aiming to help interested\nresearchers quickly catch up with the progress in this field. We also discuss\nexisting techniques for developing detection and defences against adversarial\npatches, aiming to help the community better understand this field and its\napplications in the real world.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.07188", "title": "Defending Observation Attacks in Deep Reinforcement Learning via\n  Detection and Denoising", "abstract": "  Neural network policies trained using Deep Reinforcement Learning (DRL) are\nwell-known to be susceptible to adversarial attacks. In this paper, we consider\nattacks manifesting as perturbations in the observation space managed by the\nexternal environment. These attacks have been shown to downgrade policy\nperformance significantly. We focus our attention on well-trained deterministic\nand stochastic neural network policies in the context of continuous control\nbenchmarks subject to four well-studied observation space adversarial attacks.\nTo defend against these attacks, we propose a novel defense strategy using a\ndetect-and-denoise schema. Unlike previous adversarial training approaches that\nsample data in adversarial scenarios, our solution does not require sampling\ndata in an environment under attack, thereby greatly reducing risk during\ntraining. Detailed experimental results show that our technique is comparable\nwith state-of-the-art adversarial training approaches.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.07314", "title": "Fast and Reliable Evaluation of Adversarial Robustness with\n  Minimum-Margin Attack", "abstract": "  The AutoAttack (AA) has been the most reliable method to evaluate adversarial\nrobustness when considerable computational resources are available. However,\nthe high computational cost (e.g., 100 times more than that of the project\ngradient descent attack) makes AA infeasible for practitioners with limited\ncomputational resources, and also hinders applications of AA in the adversarial\ntraining (AT). In this paper, we propose a novel method, minimum-margin (MM)\nattack, to fast and reliably evaluate adversarial robustness. Compared with AA,\nour method achieves comparable performance but only costs 3% of the\ncomputational time in extensive experiments. The reliability of our method lies\nin that we evaluate the quality of adversarial examples using the margin\nbetween two targets that can precisely identify the most adversarial example.\nThe computational efficiency of our method lies in an effective Sequential\nTArget Ranking Selection (STARS) method, ensuring that the cost of the MM\nattack is independent of the number of classes. The MM attack opens a new way\nfor evaluating adversarial robustness and provides a feasible and reliable way\nto generate high-quality adversarial examples in AT.\n", "date": "2022-06-15", "categories": []}
{"id": "2206.07321", "title": "Morphence-2.0: Evasion-Resilient Moving Target Defense Powered by\n  Out-of-Distribution Detection", "abstract": "  Evasion attacks against machine learning models often succeed via iterative\nprobing of a fixed target model, whereby an attack that succeeds once will\nsucceed repeatedly. One promising approach to counter this threat is making a\nmodel a moving target against adversarial inputs. To this end, we introduce\nMorphence-2.0, a scalable moving target defense (MTD) powered by\nout-of-distribution (OOD) detection to defend against adversarial examples. By\nregularly moving the decision function of a model, Morphence-2.0 makes it\nsignificantly challenging for repeated or correlated attacks to succeed.\nMorphence-2.0 deploys a pool of models generated from a base model in a manner\nthat introduces sufficient randomness when it responds to prediction queries.\nVia OOD detection, Morphence-2.0 is equipped with a scheduling approach that\nassigns adversarial examples to robust decision functions and benign samples to\nan undefended accurate models. To ensure repeated or correlated attacks fail,\nthe deployed pool of models automatically expires after a query budget is\nreached and the model pool is seamlessly replaced by a new model pool generated\nin advance. We evaluate Morphence-2.0 on two benchmark image classification\ndatasets (MNIST and CIFAR10) against 4 reference attacks (3 white-box and 1\nblack-box). Morphence-2.0 consistently outperforms prior defenses while\npreserving accuracy on clean data and reducing attack transferability. We also\nshow that, when powered by OOD detection, Morphence-2.0 is able to precisely\nmake an input-based movement of the model's decision function that leads to\nhigher prediction accuracy on both adversarial and benign queries.\n", "date": "2022-06-15", "categories": []}
{"id": "2206.07406", "title": "Hardening DNNs against Transfer Attacks during Network Compression using\n  Greedy Adversarial Pruning", "abstract": "  The prevalence and success of Deep Neural Network (DNN) applications in\nrecent years have motivated research on DNN compression, such as pruning and\nquantization. These techniques accelerate model inference, reduce power\nconsumption, and reduce the size and complexity of the hardware necessary to\nrun DNNs, all with little to no loss in accuracy. However, since DNNs are\nvulnerable to adversarial inputs, it is important to consider the relationship\nbetween compression and adversarial robustness. In this work, we investigate\nthe adversarial robustness of models produced by several irregular pruning\nschemes and by 8-bit quantization. Additionally, while conventional pruning\nremoves the least important parameters in a DNN, we investigate the effect of\nan unconventional pruning method: removing the most important model parameters\nbased on the gradient on adversarial inputs. We call this method Greedy\nAdversarial Pruning (GAP) and we find that this pruning method results in\nmodels that are resistant to transfer attacks from their uncompressed\ncounterparts.\n", "date": "2022-06-15", "categories": []}
{"id": "2206.06496", "title": "Towards Alternative Techniques for Improving Adversarial Robustness:\n  Analysis of Adversarial Training at a Spectrum of Perturbations", "abstract": "  Adversarial training (AT) and its variants have spearheaded progress in\nimproving neural network robustness to adversarial perturbations and common\ncorruptions in the last few years. Algorithm design of AT and its variants are\nfocused on training models at a specified perturbation strength $\\epsilon$ and\nonly using the feedback from the performance of that $\\epsilon$-robust model to\nimprove the algorithm. In this work, we focus on models, trained on a spectrum\nof $\\epsilon$ values. We analyze three perspectives: model performance,\nintermediate feature precision and convolution filter sensitivity. In each, we\nidentify alternative improvements to AT that otherwise wouldn't have been\napparent at a single $\\epsilon$. Specifically, we find that for a PGD attack at\nsome strength $\\delta$, there is an AT model at some slightly larger strength\n$\\epsilon$, but no greater, that generalizes best to it. Hence, we propose\noverdesigning for robustness where we suggest training models at an $\\epsilon$\njust above $\\delta$. Second, we observe (across various $\\epsilon$ values) that\nrobustness is highly sensitive to the precision of intermediate features and\nparticularly those after the first and second layer. Thus, we propose adding a\nsimple quantization to defenses that improves accuracy on seen and unseen\nadaptive attacks. Third, we analyze convolution filters of each layer of models\nat increasing $\\epsilon$ and notice that those of the first and second layer\nmay be solely responsible for amplifying input perturbations. We present our\nfindings and demonstrate our techniques through experiments with ResNet and\nWideResNet models on the CIFAR-10 and CIFAR-10-C datasets.\n", "date": "2022-06-13", "categories": []}
{"id": "2206.06592", "title": "Downlink Power Allocation in Massive MIMO via Deep Learning: Adversarial\n  Attacks and Training", "abstract": "  The successful emergence of deep learning (DL) in wireless system\napplications has raised concerns about new security-related challenges. One\nsuch security challenge is adversarial attacks. Although there has been much\nwork demonstrating the susceptibility of DL-based classification tasks to\nadversarial attacks, regression-based problems in the context of a wireless\nsystem have not been studied so far from an attack perspective. The aim of this\npaper is twofold: (i) we consider a regression problem in a wireless setting\nand show that adversarial attacks can break the DL-based approach and (ii) we\nanalyze the effectiveness of adversarial training as a defensive technique in\nadversarial settings and show that the robustness of DL-based wireless system\nagainst attacks improves significantly. Specifically, the wireless application\nconsidered in this paper is the DL-based power allocation in the downlink of a\nmulticell massive multi-input-multi-output system, where the goal of the attack\nis to yield an infeasible solution by the DL model. We extend the\ngradient-based adversarial attacks: fast gradient sign method (FGSM), momentum\niterative FGSM, and projected gradient descent method to analyze the\nsusceptibility of the considered wireless application with and without\nadversarial training. We analyze the deep neural network (DNN) models\nperformance against these attacks, where the adversarial perturbations are\ncrafted using both the white-box and black-box attacks.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.06737", "title": "Adversarial Vulnerability of Randomized Ensembles", "abstract": "  Despite the tremendous success of deep neural networks across various tasks,\ntheir vulnerability to imperceptible adversarial perturbations has hindered\ntheir deployment in the real world. Recently, works on randomized ensembles\nhave empirically demonstrated significant improvements in adversarial\nrobustness over standard adversarially trained (AT) models with minimal\ncomputational overhead, making them a promising solution for safety-critical\nresource-constrained applications. However, this impressive performance raises\nthe question: Are these robustness gains provided by randomized ensembles real?\nIn this work we address this question both theoretically and empirically. We\nfirst establish theoretically that commonly employed robustness evaluation\nmethods such as adaptive PGD provide a false sense of security in this setting.\nSubsequently, we propose a theoretically-sound and efficient adversarial attack\nalgorithm (ARC) capable of compromising random ensembles even in cases where\nadaptive PGD fails to do so. We conduct comprehensive experiments across a\nvariety of network architectures, training schemes, datasets, and norms to\nsupport our claims, and empirically establish that randomized ensembles are in\nfact more vulnerable to $\\ell_p$-bounded adversarial perturbations than even\nstandard AT models. Our code can be found at https://github.com/hsndbk4/ARC.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.05511", "title": "Improving the Adversarial Robustness of NLP Models by Information\n  Bottleneck", "abstract": "  Existing studies have demonstrated that adversarial examples can be directly\nattributed to the presence of non-robust features, which are highly predictive,\nbut can be easily manipulated by adversaries to fool NLP models. In this study,\nwe explore the feasibility of capturing task-specific robust features, while\neliminating the non-robust ones by using the information bottleneck theory.\nThrough extensive experiments, we show that the models trained with our\ninformation bottleneck-based method are able to achieve a significant\nimprovement in robust accuracy, exceeding performances of all the previously\nreported defense methods while suffering almost no performance drop in clean\naccuracy on SST-2, AGNEWS and IMDB datasets.\n", "date": "2022-06-11", "categories": []}
{"id": "2206.05664", "title": "An Efficient Method for Sample Adversarial Perturbations against\n  Nonlinear Support Vector Machines", "abstract": "  Adversarial perturbations have drawn great attentions in various machine\nlearning models. In this paper, we investigate the sample adversarial\nperturbations for nonlinear support vector machines (SVMs). Due to the implicit\nform of the nonlinear functions mapping data to the feature space, it is\ndifficult to obtain the explicit form of the adversarial perturbations. By\nexploring the special property of nonlinear SVMs, we transform the optimization\nproblem of attacking nonlinear SVMs into a nonlinear KKT system. Such a system\ncan be solved by various numerical methods. Numerical results show that our\nmethod is efficient in computing adversarial perturbations.\n", "date": "2022-06-12", "categories": []}
{"id": "2206.05898", "title": "Pixel to Binary Embedding Towards Robustness for CNNs", "abstract": "  There are several problems with the robustness of Convolutional Neural\nNetworks (CNNs). For example, the prediction of CNNs can be changed by adding a\nsmall magnitude of noise to an input, and the performances of CNNs are degraded\nwhen the distribution of input is shifted by a transformation never seen during\ntraining (e.g., the blur effect). There are approaches to replace pixel values\nwith binary embeddings to tackle the problem of adversarial perturbations,\nwhich successfully improve robustness. In this work, we propose Pixel to Binary\nEmbedding (P2BE) to improve the robustness of CNNs. P2BE is a learnable binary\nembedding method as opposed to previous hand-coded binary embedding methods.\nP2BE outperforms other binary embedding methods in robustness against\nadversarial perturbations and visual corruptions that are not shown during\ntraining.\n", "date": "2022-06-13", "categories": []}
{"id": "2206.04881", "title": "Enhancing Clean Label Backdoor Attack with Two-phase Specific Triggers", "abstract": "  Backdoor attacks threaten Deep Neural Networks (DNNs). Towards stealthiness,\nresearchers propose clean-label backdoor attacks, which require the adversaries\nnot to alter the labels of the poisoned training datasets. Clean-label settings\nmake the attack more stealthy due to the correct image-label pairs, but some\nproblems still exist: first, traditional methods for poisoning training data\nare ineffective; second, traditional triggers are not stealthy which are still\nperceptible. To solve these problems, we propose a two-phase and image-specific\ntriggers generation method to enhance clean-label backdoor attacks. Our methods\nare (1) powerful: our triggers can both promote the two phases (i.e., the\nbackdoor implantation and activation phase) in backdoor attacks simultaneously;\n(2) stealthy: our triggers are generated from each image. They are\nimage-specific instead of fixed triggers. Extensive experiments demonstrate\nthat our approach can achieve a fantastic attack success rate~(98.98%) with low\npoisoning rate~(5%), high stealthiness under many evaluation metrics and is\nresistant to backdoor defense methods.\n", "date": "2022-06-10", "categories": []}
{"id": "2206.04137", "title": "Adversarial Text Normalization", "abstract": "  Text-based adversarial attacks are becoming more commonplace and accessible\nto general internet users. As these attacks proliferate, the need to address\nthe gap in model robustness becomes imminent. While retraining on adversarial\ndata may increase performance, there remains an additional class of\ncharacter-level attacks on which these models falter. Additionally, the process\nto retrain a model is time and resource intensive, creating a need for a\nlightweight, reusable defense. In this work, we propose the Adversarial Text\nNormalizer, a novel method that restores baseline performance on attacked\ncontent with low computational overhead. We evaluate the efficacy of the\nnormalizer on two problem areas prone to adversarial attacks, i.e. Hate Speech\nand Natural Language Inference. We find that text normalization provides a\ntask-agnostic defense against character-level attacks that can be implemented\nsupplementary to adversarial retraining solutions, which are more suited for\nsemantic alterations.\n", "date": "2022-06-08", "categories": []}
{"id": "2206.04316", "title": "Adversarial Noises Are Linearly Separable for (Nearly) Random Neural\n  Networks", "abstract": "  Adversarial examples, which are usually generated for specific inputs with a\nspecific model, are ubiquitous for neural networks. In this paper we unveil a\nsurprising property of adversarial noises when they are put together, i.e.,\nadversarial noises crafted by one-step gradient methods are linearly separable\nif equipped with the corresponding labels. We theoretically prove this property\nfor a two-layer network with randomly initialized entries and the neural\ntangent kernel setup where the parameters are not far from initialization. The\nproof idea is to show the label information can be efficiently backpropagated\nto the input while keeping the linear separability. Our theory and experimental\nevidence further show that the linear classifier trained with the adversarial\nnoises of the training data can well classify the adversarial noises of the\ntest data, indicating that adversarial noises actually inject a distributional\nperturbation to the original data distribution. Furthermore, we empirically\ndemonstrate that the adversarial noises may become less linearly separable when\nthe above conditions are compromised while they are still much easier to\nclassify than original features.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.04472", "title": "Early Transferability of Adversarial Examples in Deep Neural Networks", "abstract": "  This paper will describe and analyze a new phenomenon that was not known\nbefore, which we call \"Early Transferability\". Its essence is that the\nadversarial perturbations transfer among different networks even at extremely\nearly stages in their training. In fact, one can initialize two networks with\ntwo different independent choices of random weights and measure the angle\nbetween their adversarial perturbations after each step of the training. What\nwe discovered was that these two adversarial directions started to align with\neach other already after the first few training steps (which typically use only\na small fraction of the available training data), even though the accuracy of\nthe two networks hadn't started to improve from their initial bad values due to\nthe early stage of the training. The purpose of this paper is to present this\nphenomenon experimentally and propose plausible explanations for some of its\nproperties.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.03727", "title": "Wavelet Regularization Benefits Adversarial Training", "abstract": "  Adversarial training methods are state-of-the-art (SOTA) empirical defense\nmethods against adversarial examples. Many regularization methods have been\nproven to be effective with the combination of adversarial training.\nNevertheless, such regularization methods are implemented in the time domain.\nSince adversarial vulnerability can be regarded as a high-frequency phenomenon,\nit is essential to regulate the adversarially-trained neural network models in\nthe frequency domain. Faced with these challenges, we make a theoretical\nanalysis on the regularization property of wavelets which can enhance\nadversarial training. We propose a wavelet regularization method based on the\nHaar wavelet decomposition which is named Wavelet Average Pooling. This wavelet\nregularization module is integrated into the wide residual neural network so\nthat a new WideWaveletResNet model is formed. On the datasets of CIFAR-10 and\nCIFAR-100, our proposed Adversarial Wavelet Training method realizes\nconsiderable robustness under different types of attacks. It verifies the\nassumption that our wavelet regularization method can enhance adversarial\nrobustness especially in the deep wide neural networks. The visualization\nexperiments of the Frequency Principle (F-Principle) and interpretability are\nimplemented to show the effectiveness of our method. A detailed comparison\nbased on different wavelet base functions is presented. The code is available\nat the repository:\n\\url{https://github.com/momo1986/AdversarialWaveletTraining}.\n", "date": "2022-06-08", "categories": []}
{"id": "2206.03351", "title": "AS2T: Arbitrary Source-To-Target Adversarial Attack on Speaker\n  Recognition Systems", "abstract": "  Recent work has illuminated the vulnerability of speaker recognition systems\n(SRSs) against adversarial attacks, raising significant security concerns in\ndeploying SRSs. However, they considered only a few settings (e.g., some\ncombinations of source and target speakers), leaving many interesting and\nimportant settings in real-world attack scenarios alone. In this work, we\npresent AS2T, the first attack in this domain which covers all the settings,\nthus allows the adversary to craft adversarial voices using arbitrary source\nand target speakers for any of three main recognition tasks. Since none of the\nexisting loss functions can be applied to all the settings, we explore many\ncandidate loss functions for each setting including the existing and newly\ndesigned ones. We thoroughly evaluate their efficacy and find that some\nexisting loss functions are suboptimal. Then, to improve the robustness of AS2T\ntowards practical over-the-air attack, we study the possible distortions\noccurred in over-the-air transmission, utilize different transformation\nfunctions with different parameters to model those distortions, and incorporate\nthem into the generation of adversarial voices. Our simulated over-the-air\nevaluation validates the effectiveness of our solution in producing robust\nadversarial voices which remain effective under various hardware devices and\nvarious acoustic environments with different reverberation, ambient noises, and\nnoise levels. Finally, we leverage AS2T to perform thus far the largest-scale\nevaluation to understand transferability among 14 diverse SRSs. The\ntransferability analysis provides many interesting and useful insights which\nchallenge several findings and conclusion drawn in previous works in the image\ndomain. Our study also sheds light on future directions of adversarial attacks\nin the speaker recognition domain.\n", "date": "2022-06-07", "categories": []}
{"id": "2206.03362", "title": "Building Robust Ensembles via Margin Boosting", "abstract": "  In the context of adversarial robustness, a single model does not usually\nhave enough power to defend against all possible adversarial attacks, and as a\nresult, has sub-optimal robustness. Consequently, an emerging line of work has\nfocused on learning an ensemble of neural networks to defend against\nadversarial attacks. In this work, we take a principled approach towards\nbuilding robust ensembles. We view this problem from the perspective of\nmargin-boosting and develop an algorithm for learning an ensemble with maximum\nmargin. Through extensive empirical evaluation on benchmark datasets, we show\nthat our algorithm not only outperforms existing ensembling techniques, but\nalso large models trained in an end-to-end fashion. An important byproduct of\nour work is a margin-maximizing cross-entropy (MCE) loss, which is a better\nalternative to the standard cross-entropy (CE) loss. Empirically, we show that\nreplacing the CE loss in state-of-the-art adversarial training techniques with\nour MCE loss leads to significant performance improvement.\n", "date": "2022-06-07", "categories": []}
{"id": "2206.03393", "title": "Towards Understanding and Mitigating Audio Adversarial Examples for\n  Speaker Recognition", "abstract": "  Speaker recognition systems (SRSs) have recently been shown to be vulnerable\nto adversarial attacks, raising significant security concerns. In this work, we\nsystematically investigate transformation and adversarial training based\ndefenses for securing SRSs. According to the characteristic of SRSs, we present\n22 diverse transformations and thoroughly evaluate them using 7 recent\npromising adversarial attacks (4 white-box and 3 black-box) on speaker\nrecognition. With careful regard for best practices in defense evaluations, we\nanalyze the strength of transformations to withstand adaptive attacks. We also\nevaluate and understand their effectiveness against adaptive attacks when\ncombined with adversarial training. Our study provides lots of useful insights\nand findings, many of them are new or inconsistent with the conclusions in the\nimage and speech recognition domains, e.g., variable and constant bit rate\nspeech compressions have different performance, and some non-differentiable\ntransformations remain effective against current promising evasion techniques\nwhich often work well in the image domain. We demonstrate that the proposed\nnovel feature-level transformation combined with adversarial training is rather\neffective compared to the sole adversarial training in a complete white-box\nsetting, e.g., increasing the accuracy by 13.62% and attack cost by two orders\nof magnitude, while other transformations do not necessarily improve the\noverall defense capability. This work sheds further light on the research\ndirections in this field. We also release our evaluation platform SPEAKERGUARD\nto foster further research.\n", "date": "2022-06-07", "categories": []}
{"id": "2206.01733", "title": "Adversarial RAW: Image-Scaling Attack Against Imaging Pipeline", "abstract": "  Deep learning technologies have become the backbone for the development of\ncomputer vision. With further explorations, deep neural networks have been\nfound vulnerable to well-designed adversarial attacks. Most of the vision\ndevices are equipped with image signal processing (ISP) pipeline to implement\nRAW-to-RGB transformations and embedded into data preprocessing module for\nefficient image processing. Actually, ISP pipeline can introduce adversarial\nbehaviors to post-capture images while data preprocessing may destroy attack\npatterns. However, none of the existing adversarial attacks takes into account\nthe impacts of both ISP pipeline and data preprocessing. In this paper, we\ndevelop an image-scaling attack targeting on ISP pipeline, where the crafted\nadversarial RAW can be transformed into attack image that presents entirely\ndifferent appearance once being scaled to a specific-size image. We first\nconsider the gradient-available ISP pipeline, i.e., the gradient information\ncan be directly used in the generation process of adversarial RAW to launch the\nattack. To make the adversarial attack more applicable, we further consider the\ngradient-unavailable ISP pipeline, in which a proxy model that well learns the\nRAW-to-RGB transformations is proposed as the gradient oracles. Extensive\nexperiments show that the proposed adversarial attacks can craft adversarial\nRAW data against the target ISP pipelines with high attack rates.\n", "date": "2022-06-02", "categories": []}
{"id": "2206.01832", "title": "Kallima: A Clean-label Framework for Textual Backdoor Attacks", "abstract": "  Although Deep Neural Network (DNN) has led to unprecedented progress in\nvarious natural language processing (NLP) tasks, research shows that deep\nmodels are extremely vulnerable to backdoor attacks. The existing backdoor\nattacks mainly inject a small number of poisoned samples into the training\ndataset with the labels changed to the target one. Such mislabeled samples\nwould raise suspicion upon human inspection, potentially revealing the attack.\nTo improve the stealthiness of textual backdoor attacks, we propose the first\nclean-label framework Kallima for synthesizing mimesis-style backdoor samples\nto develop insidious textual backdoor attacks. We modify inputs belonging to\nthe target class with adversarial perturbations, making the model rely more on\nthe backdoor trigger. Our framework is compatible with most existing backdoor\ntriggers. The experimental results on three benchmark datasets demonstrate the\neffectiveness of the proposed method.\n", "date": "2022-06-03", "categories": []}
{"id": "2206.01898", "title": "Saliency Attack: Towards Imperceptible Black-box Adversarial Attack", "abstract": "  Deep neural networks are vulnerable to adversarial examples, even in the\nblack-box setting where the attacker is only accessible to the model output.\nRecent studies have devised effective black-box attacks with high query\nefficiency. However, such performance is often accompanied by compromises in\nattack imperceptibility, hindering the practical use of these approaches. In\nthis paper, we propose to restrict the perturbations to a small salient region\nto generate adversarial examples that can hardly be perceived. This approach is\nreadily compatible with many existing black-box attacks and can significantly\nimprove their imperceptibility with little degradation in attack success rate.\nFurther, we propose the Saliency Attack, a new black-box attack aiming to\nrefine the perturbations in the salient region to achieve even better\nimperceptibility. Extensive experiments show that compared to the\nstate-of-the-art black-box attacks, our approach achieves much better\nimperceptibility scores, including most apparent distortion (MAD), $L_0$ and\n$L_2$ distances, and also obtains significantly higher success rates judged by\na human-like threshold on MAD. Importantly, the perturbations generated by our\napproach are interpretable to some extent. Finally, it is also demonstrated to\nbe robust to different detection-based defenses.\n", "date": "2022-06-04", "categories": []}
{"id": "2206.01904", "title": "Soft Adversarial Training Can Retain Natural Accuracy", "abstract": "  Adversarial training for neural networks has been in the limelight in recent\nyears. The advancement in neural network architectures over the last decade has\nled to significant improvement in their performance. It sparked an interest in\ntheir deployment for real-time applications. This process initiated the need to\nunderstand the vulnerability of these models to adversarial attacks. It is\ninstrumental in designing models that are robust against adversaries. Recent\nworks have proposed novel techniques to counter the adversaries, most often\nsacrificing natural accuracy. Most suggest training with an adversarial version\nof the inputs, constantly moving away from the original distribution. The focus\nof our work is to use abstract certification to extract a subset of inputs for\n(hence we call it 'soft') adversarial training. We propose a training framework\nthat can retain natural accuracy without sacrificing robustness in a\nconstrained setting. Our framework specifically targets moderately critical\napplications which require a reasonable balance between robustness and\naccuracy. The results testify to the idea of soft adversarial training for the\ndefense against adversarial attacks. At last, we propose the scope of future\nwork for further improvement of this framework.\n", "date": "2022-06-04", "categories": []}
{"id": "2206.02131", "title": "Federated Adversarial Training with Transformers", "abstract": "  Federated learning (FL) has emerged to enable global model training over\ndistributed clients' data while preserving its privacy. However, the global\ntrained model is vulnerable to the evasion attacks especially, the adversarial\nexamples (AEs), carefully crafted samples to yield false classification.\nAdversarial training (AT) is found to be the most promising approach against\nevasion attacks and it is widely studied for convolutional neural network\n(CNN). Recently, vision transformers have been found to be effective in many\ncomputer vision tasks. To the best of the authors' knowledge, there is no work\nthat studied the feasibility of AT in a FL process for vision transformers.\nThis paper investigates such feasibility with different federated model\naggregation methods and different vision transformer models with different\ntokenization and classification head techniques. In order to improve the robust\naccuracy of the models with the not independent and identically distributed\n(Non-IID), we propose an extension to FedAvg aggregation method, called\nFedWAvg. By measuring the similarities between the last layer of the global\nmodel and the last layer of the client updates, FedWAvg calculates the weights\nto aggregate the local models updates. The experiments show that FedWAvg\nimproves the robust accuracy when compared with other state-of-the-art\naggregation methods.\n", "date": "2022-06-05", "categories": []}
{"id": "2206.02158", "title": "Vanilla Feature Distillation for Improving the Accuracy-Robustness\n  Trade-Off in Adversarial Training", "abstract": "  Adversarial training has been widely explored for mitigating attacks against\ndeep models. However, most existing works are still trapped in the dilemma\nbetween higher accuracy and stronger robustness since they tend to fit a model\ntowards robust features (not easily tampered with by adversaries) while\nignoring those non-robust but highly predictive features. To achieve a better\nrobustness-accuracy trade-off, we propose the Vanilla Feature Distillation\nAdversarial Training (VFD-Adv), which conducts knowledge distillation from a\npre-trained model (optimized towards high accuracy) to guide adversarial\ntraining towards higher accuracy, i.e., preserving those non-robust but\npredictive features. More specifically, both adversarial examples and their\nclean counterparts are forced to be aligned in the feature space by distilling\npredictive representations from the pre-trained/clean model, while previous\nworks barely utilize predictive features from clean models. Therefore, the\nadversarial training model is updated towards maximally preserving the accuracy\nas gaining robustness. A key advantage of our method is that it can be\nuniversally adapted to and boost existing works. Exhaustive experiments on\nvarious datasets, classification models, and adversarial training algorithms\ndemonstrate the effectiveness of our proposed method.\n", "date": "2022-06-05", "categories": []}
{"id": "2206.02417", "title": "Fast Adversarial Training with Adaptive Step Size", "abstract": "  While adversarial training and its variants have shown to be the most\neffective algorithms to defend against adversarial attacks, their extremely\nslow training process makes it hard to scale to large datasets like ImageNet.\nThe key idea of recent works to accelerate adversarial training is to\nsubstitute multi-step attacks (e.g., PGD) with single-step attacks (e.g.,\nFGSM). However, these single-step methods suffer from catastrophic overfitting,\nwhere the accuracy against PGD attack suddenly drops to nearly 0% during\ntraining, destroying the robustness of the networks. In this work, we study the\nphenomenon from the perspective of training instances. We show that\ncatastrophic overfitting is instance-dependent and fitting instances with\nlarger gradient norm is more likely to cause catastrophic overfitting. Based on\nour findings, we propose a simple but effective method, Adversarial Training\nwith Adaptive Step size (ATAS). ATAS learns an instancewise adaptive step size\nthat is inversely proportional to its gradient norm. The theoretical analysis\nshows that ATAS converges faster than the commonly adopted non-adaptive\ncounterparts. Empirically, ATAS consistently mitigates catastrophic overfitting\nand achieves higher robust accuracy on CIFAR10, CIFAR100 and ImageNet when\nevaluated on various adversarial budgets.\n", "date": "2022-06-06", "categories": []}
{"id": "2206.02539", "title": "Robustness Evaluation and Adversarial Training of an Instance\n  Segmentation Model", "abstract": "  To evaluate the robustness of non-classifier models, we propose probabilistic\nlocal equivalence, based on the notion of randomized smoothing, as a way to\nquantitatively evaluate the robustness of an arbitrary function. In addition,\nto understand the effect of adversarial training on non-classifiers and to\ninvestigate the level of robustness that can be obtained without degrading\nperformance on the training distribution, we apply Fast is Better than Free\nadversarial training together with the TRADES robust loss to the training of an\ninstance segmentation network. In this direction, we were able to achieve a\nsymmetric best dice score of 0.85 on the TuSimple lane detection challenge,\noutperforming the standardly-trained network's score of 0.82. Additionally, we\nwere able to obtain an F-measure of 0.49 on manipulated inputs, in contrast to\nthe standardly-trained network's score of 0. We show that probabilisitic local\nequivalence is able to successfully distinguish between standardly-trained and\nadversarially-trained models, providing another view of the improved robustness\nof the adversarially-trained models.\n", "date": "2022-06-02", "categories": []}
{"id": "2206.01705", "title": "Gradient Obfuscation Checklist Test Gives a False Sense of Security", "abstract": "  One popular group of defense techniques against adversarial attacks is based\non injecting stochastic noise into the network. The main source of robustness\nof such stochastic defenses however is often due to the obfuscation of the\ngradients, offering a false sense of security. Since most of the popular\nadversarial attacks are optimization-based, obfuscated gradients reduce their\nattacking ability, while the model is still susceptible to stronger or\nspecifically tailored adversarial attacks. Recently, five characteristics have\nbeen identified, which are commonly observed when the improvement in robustness\nis mainly caused by gradient obfuscation. It has since become a trend to use\nthese five characteristics as a sufficient test, to determine whether or not\ngradient obfuscation is the main source of robustness. However, these\ncharacteristics do not perfectly characterize all existing cases of gradient\nobfuscation, and therefore can not serve as a basis for a conclusive test. In\nthis work, we present a counterexample, showing this test is not sufficient for\nconcluding that gradient obfuscation is not the main cause of improvements in\nrobustness.\n", "date": "2022-06-03", "categories": []}
{"id": "2206.01715", "title": "Towards Evading the Limits of Randomized Smoothing: A Theoretical\n  Analysis", "abstract": "  Randomized smoothing is the dominant standard for provable defenses against\nadversarial examples. Nevertheless, this method has recently been proven to\nsuffer from important information theoretic limitations. In this paper, we\nargue that these limitations are not intrinsic, but merely a byproduct of\ncurrent certification methods. We first show that these certificates use too\nlittle information about the classifier, and are in particular blind to the\nlocal curvature of the decision boundary. This leads to severely sub-optimal\nrobustness guarantees as the dimension of the problem increases. We then show\nthat it is theoretically possible to bypass this issue by collecting more\ninformation about the classifier. More precisely, we show that it is possible\nto approximate the optimal certificate with arbitrary precision, by probing the\ndecision boundary with several noise distributions. Since this process is\nexecuted at certification time rather than at test time, it entails no loss in\nnatural accuracy while enhancing the quality of the certificates. This result\nfosters further research on classifier-specific certification and demonstrates\nthat randomized smoothing is still worth investigating. Although\nclassifier-specific certification may induce more computational cost, we also\nprovide some theoretical insight on how to mitigate it.\n", "date": "2022-06-03", "categories": []}
{"id": "2206.00772", "title": "On the reversibility of adversarial attacks", "abstract": "  Adversarial attacks modify images with perturbations that change the\nprediction of classifiers. These modified images, known as adversarial\nexamples, expose the vulnerabilities of deep neural network classifiers. In\nthis paper, we investigate the predictability of the mapping between the\nclasses predicted for original images and for their corresponding adversarial\nexamples. This predictability relates to the possibility of retrieving the\noriginal predictions and hence reversing the induced misclassification. We\nrefer to this property as the reversibility of an adversarial attack, and\nquantify reversibility as the accuracy in retrieving the original class or the\ntrue class of an adversarial example. We present an approach that reverses the\neffect of an adversarial attack on a classifier using a prior set of\nclassification results. We analyse the reversibility of state-of-the-art\nadversarial attacks on benchmark classifiers and discuss the factors that\naffect the reversibility.\n", "date": "2022-06-01", "categories": []}
{"id": "2206.00352", "title": "Support Vector Machines under Adversarial Label Contamination", "abstract": "  Machine learning algorithms are increasingly being applied in\nsecurity-related tasks such as spam and malware detection, although their\nsecurity properties against deliberate attacks have not yet been widely\nunderstood. Intelligent and adaptive attackers may indeed exploit specific\nvulnerabilities exposed by machine learning techniques to violate system\nsecurity. Being robust to adversarial data manipulation is thus an important,\nadditional requirement for machine learning algorithms to successfully operate\nin adversarial settings. In this work, we evaluate the security of Support\nVector Machines (SVMs) to well-crafted, adversarial label noise attacks. In\nparticular, we consider an attacker that aims to maximize the SVM's\nclassification error by flipping a number of labels in the training data. We\nformalize a corresponding optimal attack strategy, and solve it by means of\nheuristic approaches to keep the computational complexity tractable. We report\nan extensive experimental analysis on the effectiveness of the considered\nattacks against linear and non-linear SVMs, both on synthetic and real-world\ndatasets. We finally argue that our approach can also provide useful insights\nfor developing more secure SVM learning algorithms, and also novel techniques\nin a number of related research areas, such as semi-supervised and active\nlearning.\n", "date": "2022-06-01", "categories": []}
{"id": "2206.00489", "title": "Attack-Agnostic Adversarial Detection", "abstract": "  The growing number of adversarial attacks in recent years gives attackers an\nadvantage over defenders, as defenders must train detectors after knowing the\ntypes of attacks, and many models need to be maintained to ensure good\nperformance in detecting any upcoming attacks. We propose a way to end the\ntug-of-war between attackers and defenders by treating adversarial attack\ndetection as an anomaly detection problem so that the detector is agnostic to\nthe attack. We quantify the statistical deviation caused by adversarial\nperturbations in two aspects. The Least Significant Component Feature (LSCF)\nquantifies the deviation of adversarial examples from the statistics of benign\nsamples and Hessian Feature (HF) reflects how adversarial examples distort the\nlandscape of the model's optima by measuring the local loss curvature.\nEmpirical results show that our method can achieve an overall ROC AUC of 94.9%,\n89.7%, and 94.6% on CIFAR10, CIFAR100, and SVHN, respectively, and has\ncomparable performance to adversarial detectors trained with adversarial\nexamples on most of the attacks.\n", "date": "2022-06-01", "categories": []}
{"id": "2206.00513", "title": "The robust way to stack and bag: the local Lipschitz way", "abstract": "  Recent research has established that the local Lipschitz constant of a neural\nnetwork directly influences its adversarial robustness. We exploit this\nrelationship to construct an ensemble of neural networks which not only\nimproves the accuracy, but also provides increased adversarial robustness. The\nlocal Lipschitz constants for two different ensemble methods - bagging and\nstacking - are derived and the architectures best suited for ensuring\nadversarial robustness are deduced. The proposed ensemble architectures are\ntested on MNIST and CIFAR-10 datasets in the presence of white-box attacks,\nFGSM and PGD. The proposed architecture is found to be more robust than a) a\nsingle network and b) traditional ensemble methods.\n", "date": "2022-06-01", "categories": []}
{"id": "2606.21077", "title": "OTTER: A Red-Teaming System for Toxicity-Evading Jailbreak Prompt Optimization", "abstract": "Production LLMs increasingly rely on toxicity-based moderation filters as a primary defense, assuming that harmful intent correlates with toxic surface wording. We show this assumption is fundamentally brittle: surface toxicity and adversarial intent can be decoupled by replacing as few as five tokens. We present OTTER (Obfuscated Toxicity-Evading Token Evolution for Rewriting), a black-box red-teaming framework requiring only standard API access, directly targeting the practical constraints of industry security audits. Evaluated on 457 AdvBench prompts across four GPT models, OTTER raises average ASR from 7.0% to 84.0%. We further provide the first quantitative analysis of the toxicity--bypass relationship and a per-category breakdown, translating our findings into actionable recommendations for classifier hardening in production deployments.", "date": "2026-06-19", "categories": []}
{"id": "2606.21172", "title": "BadDreamer: Transferable Backdoor Attacks against Video World Models for Autonomous Driving", "abstract": "Video world models are increasingly used in autonomous driving to forecast future scene evolution and provide future-aware spatio-temporal representations for downstream action prediction. In perception-to-action pipelines, these representations can directly influence ego-vehicle waypoint planning, making the learned future dynamics a critical security-sensitive component. Despite their promise, the training-time security risks of autonomous-driving video world models remain largely unexplored. We present BadDreamer, a transferable spatio-temporal backdoor attack that targets the perception side of this pipeline. Unlike conventional backdoors that manipulate image labels, prompt outputs, or action supervision, BadDreamer poisons the learned transition dynamics of a video world model. It constructs trigger-erasure sequences in which an oncoming yellow delivery rider is visible in the observed context frames but erased from the future frames. After fine-tuning on a small fraction of such sequences, the compromised world model learns a hidden conditional association: when the physical trigger appears, it hallucinates a future where the rider disappears and the road appears clear. We further show that this corrupted future-aware representation can transfer to the downstream action module without directly modifying ego-trajectory labels, inducing unsafe non-evasive waypoint predictions. Our experiments instantiate this attack on a representative open-source perception-to-action pipeline, revealing a representation-level safety risk in autonomous-driving video world models and highlighting the need for backdoor-aware validation beyond clean generation quality.", "date": "2026-06-19", "categories": []}
{"id": "2606.21488", "title": "Robustness Cannot be Reduced to Regularization: Studying Adversarial Training Beyond the Linear Case", "abstract": "The vulnerability of ML models to adversarial examples has recently emerged as a major concern. While adversarial training is one of the most effective countermeasures to this issue, its high computational cost remains an obstacle to practical deployment. Recent progress in reducing this cost has relied, in the case of linear models, on a formal equivalence between the adversarial risk and a simpler form of regularized risk. This enabled significantly more efficient training procedures, which naturally raises the question of whether such an equivalence can be extended beyond linear models. In this work, we formally show that no such equivalence is possible for two-layer networks. Our proofs proceed via a reduction to key properties that fundamentally separate the adversarial risk from any simple regularized risk which would only exhibit a weak form of data dependence. Beyond this setting, we provide empirical evidence on Wide-ResNets indicating that the same type of impossibility persists in deeper and more expressive architectures.", "date": "2026-06-19", "categories": []}
{"id": "2606.22516", "title": "The Scissors Effect: When Resize-Based Input Diversity Helps or Hurts Transfer Attacks", "abstract": "Input Diversity (DI), which applies random resizing and padding at each attack iteration, is a near-default ingredient of transfer-based adversarial attacks, widely assumed to improve transferability. We show this assumption is regime-dependent and, for robustly trained surrogates, often reversed. Varying only the surrogate, increasing the DI probability raises transfer success for standard surrogates but lowers it for robust ones: the two response curves separate like a pair of scissors, a pattern we call the Scissors Effect. The effect is strong and consistent on ImageNet, where blind DI costs the robust source 10.3% attack success on average across CNN, ViT, Swin, and ConvNeXt targets and across ten attacks spanning 2018-2024; it is smaller on CIFAR-10 unless DI is made aggressive. A controlled robustness-strength sweep that varies only the training budget shows the harm is graded rather than binary, crossing from beneficial to harmful already in the little-robustness regime. We trace it to gradient geometry: a resize/translation decomposition attributes roughly 67% of the harm to resize, and a direct source-target gradient-alignment measurement confirms the same resize operation improves alignment for standard surrogates but degrades it for robust ones. We summarize the regime with Local Gradient Consistency (LGC), a single input-space probe that separates the two surrogate types, and prove a bias-variance crossover theorem isolating where DI helps from where its resize bias dominates. A training-free rule (CG-DI) that disables diversity when LGC is high avoids the loss on robust surrogates while keeping DI's benefit on standard ones, positioning the Scissors Effect as a DI-specific manifestation of the broader robustness-transferability trade-off.", "date": "2026-06-21", "categories": []}
{"id": "2606.22779", "title": "DE-FIVE: Detecting Malicious Image Prompts via Fourier Features and Image Vector Embeddings", "abstract": "Vision language models (VLMs) employ both visual and textual modalities to enable advanced vision-language inference. However, incorporating visual modalities expands the attack surface of VLMs, making them more susceptible to security threats such as adversarial perturbations and indirect prompt injection, wherein crafted malicious image prompts can elicit unintended model outputs. Existing defense methods against malicious image prompts remain insufficient as they typically demand extensive datasets for retraining or the deployment of additional, complex classifiers. Most critically, there is a profound lack of specialized defense mechanisms specifically targeting indirect prompt injections, a gap that serves as a primary motivation for this work. To address these limitations, we introduce DE-FIVE, a novel training-free framework for detecting malicious image prompts by leveraging Fourier features and the hidden state representations of the visual encoder (image vector embeddings) across perturbations. Specifically, we develop a hybrid detection strategy consisting of a black-box detector that operates on Fourier-domain features and a white-box detector that exploits image vector embeddings derived from only a few-shot malicious set. Extensive experiments demonstrate that the proposed framework consistently outperforms state-of-the-art baselines against malicious image prompts.", "date": "2026-06-22", "categories": []}
{"id": "2606.22837", "title": "CLIP-guided Diffusion Model for Backdoor Generation in Sensor-based Human Activity Recognition", "abstract": "Sensors are critical components of modern intelligent devices. The proliferation of the Internet of Things (IoT) and wearable mobile devices has enabled the integration of such sensors to monitor the environment and enable users to take predictive actions. Human activity recognition (HAR) is a popular application in which Inertial Measurement Unit (IMU)-based sensors, such as accelerometers and gyroscopes, are used to provide insights into health, training, and medical diagnosis. However, the accuracy of such a model is hindered by the lack of data. The diffusion model-based technique has proven successful in generating synthetic data for training HAR models. In this paper, we propose a backdoor training technique, IMU-DM-CLIP, that leverages a diffusion model to enable trigger-based attacks on HAR models. Our empirical analysis shows that the attack is successful even with a very small backdoor injection rate of 10\\% and 10\\% of the data guided for the diffusion model.", "date": "2026-06-22", "categories": []}
{"id": "2606.22966", "title": "Attacking the Trusted Imagination: Oracle-Level Integrity Attacks on Imagine-then-Act World Models", "abstract": "Many recent vision-language-action (VLA) policies adopt an imagine-then-act design. A world-action model (WAM) first imagines a short future as a latent trajectory z~, on which the action is then conditioned. We identify this trusted imagination, rather than the reactive policy, as the exposed attack surface. A downstream oracle, such as a safety gate, a visual model-predictive-control (MPC) planner, or an imagine-then-check verifier, consumes z~ as a prediction of the future. The robustness of the policy therefore does not entail the robustness of systems that rely on the WAM. The underlying phenomenon is an asymmetry. Corrupting the imagination is easy, since it requires only displacing z~ from its natural-future manifold. Steering it precisely is hard, since it must reach a specified on-manifold target. We adopt a capability-based threat model with an L-infinity-bounded observation perturbation. The attacker applies projected gradient descent through the fully differentiable observation-to-imagination map. The same off-manifold property motivates a parameter-free denoiser detector. We evaluate three targets: RynnVLA-002, LingBot-VA, and LaDi-WM. Untargeted corruption is roughly 60x stronger than random and is detected at AUC 1.0. Targeted control remains bounded. An adaptive attacker evades detection only by forgoing corruption. The reactive policy remains robust to corrupted imagination. A native imagination-driven MPC, however, exhibits the first adversary-specific task failure (at epsilon=0.01, success 0.70 versus 0.05; Fisher p < 10^-4).", "date": "2026-06-22", "categories": []}
{"id": "2606.23361", "title": "Rethinking Molecular Graph Backdoors under Chemistry-aware Admission", "abstract": "Backdoor attacks on molecular graph neural networks (GNNs) are typically evaluated as abstract graph edits, but real molecular learning pipelines do not train on arbitrary graphs. Molecular records must first survive parsing, sanitization, canonicalization, and graph-string consistency checks. We formalize this overlooked admission stage as ChemGuard, an operational protocol for testing whether a submitted molecular record can enter a realistic learning pipeline, while complementing existing defenses. ChemGuard admits a record only when its molecular string is sanitizable and the graph reconstructed from that string matches the submitted molecular graph. Under this operational view, many existing graph-based backdoors lose much of their apparent efficacy because their poisons are chemically invalid or representation-inconsistent. We then show that admission checks alone are insufficient to rule out molecular backdoors. We propose ChemBack, an admission-aware molecular backdoor attack that constructs chemically feasible motif-anchor attachments and ranks admitted candidates by fingerprint-based Tanimoto similarity to clean target-class molecules. ChemBack is model-free during trigger selection, using molecular structures, target labels, fingerprints, and public validity checks, but no victim model, surrogate GNN, learned embedding, gradient, logit, or training-code access. Across molecular benchmarks, validators, architectures, and defenses, \\textbf{ChemBack} achieves high attack success with fully admitted poisons while preserving clean accuracy. Our results reveal a two-sided lesson, chemistry-aware admission suppresses many graph-only backdoors, yet chemically valid and target-aligned molecular backdoors remain a practical threat.", "date": "2026-06-22", "categories": []}
{"id": "2606.23362", "title": "TooBad: Backdoor Diffusion Models with Ultra-Low Poison Rate and Imperceptible Trigger", "abstract": "Diffusion models (DMs), despite their impressive capabilities across a wide range of generative tasks, have been shown to be vulnerable to backdoor attacks. However, existing backdoor methods face critical trade-offs among key factors: attack performance, stealthiness, time complexity, and required poison rates. For example, achieving high attack performance typically demands a high poison rate and prolonged training, which undermines stealthiness, making the attack more detectable by backdoor defenses. This paper proposes TooBad (trigger optimization for backdoor diffusion models), a backdoor framework which introduces a novel DM-tailored trigger optimization technique to dramatically enhance the performance of backdoor attacks on DMs. Experiments on representative benchmarks such as CIFAR-10 show that TooBad can achieve high ASRs ($> 85$%) at only 0.5% poison rate, significantly lower than the 10% typically required by prior work on the same datasets. At 5% poison rate, TooBad reaches nearly 100% ASR within just 3-5 backdoor injection epochs, whereas existing methods need at least 30-50 epochs at double the poison rate for comparable results. Despite its potency, TooBad easily evades SOTA defenses and maintains high utility. These results reveal a critical threat on DMs and highlight the need for more robust defenses against such stealthy yet efficient attacks.", "date": "2026-06-22", "categories": []}
{"id": "2606.24589", "title": "AdversaBench: Automated LLM Red-Teaming with Multi-Judge Confirmation and Cross-Model Transferability", "abstract": "Scaling adversarial evaluation of large language models requires both a method for generating hard inputs and a reliable way to confirm that resulting failures are real. We present AdversaBench, an end-to-end red-teaming pipeline that mutates seed prompts with five structured operators, queries a target model, and confirms failures through a three-judge panel with a meta-judge tiebreaker. We report experiments on 45 seeds across three categories: reasoning, instruction-following, and tool use. Every seed produced a confirmed failure. Four findings stand out. First, operator effectiveness varies sharply by category: inject_distractor scores 0.00 mean reward on instruction-following seeds but 0.80-0.83 on reasoning and tool-use. Second, binary failure rate hides difficulty: instruction-following seeds required 2.4 attacker iterations on average versus 1.1 for other categories, a gap visible in survival curves. Third, pairwise judge agreement of 80-87% coexists with near-zero Cohen's kappa due to label skew; category-level disagreement rates are more informative. Fourth, adversarial prompts generated against Llama 3.1 8B transfer zero-shot to Llama 3.3 70B, suggesting the mutations exploit general behavioral patterns rather than model-specific weaknesses. Code, dataset, and analysis scripts are available at https://github.com/khanak0509/AdversaBench .", "date": "2026-06-23", "categories": []}
{"id": "2606.25034", "title": "Yuvion VL: A Multimodal Foundation Model for Adversarial Content and AI Safety", "abstract": "General-purpose models often struggle to reliably identify and understand real-world multimodal risks, largely due to the inherent multimodal adversarial nature of content and AI safety. We present Yuvion VL, a family of multimodal large language models purpose-built for content and AI safety, with both instruction-tuned and reasoning-oriented variants. Yuvion VL addresses this gap by treating safety as an inherently adversarial and multimodal problem and designing the entire pipeline around adversarial robustness. For data construction, we develop an automated pipeline integrating adversarial-aware data synthesis with multi-stage quality control, producing large-scale, high-quality multimodal samples augmented with domain knowledge and reasoning annotations. For training, we adopt a three-stage pipeline that includes continued pretraining for risk-concept cross-modal alignment, instruct post-training for production-grade safety tasks, and reasoning post-training for enhanced interpretability and performance in complex tasks. We further introduce Confuse-then-Contrast Fine-Tuning, a contrastive framework that mines model-specific confusions and constructs multi-image contrastive groups to enforce explicit discrimination of fine-grained visual-semantic elements, enabling the model to distinguish between visually similar cases with different safety implications in adversarial safety tasks. To support rigorous evaluation, we further introduce Yuvion VL RiskEval (YVRE), a collection of benchmarks covering diverse open and internal evaluations, with a focus on content and AI safety, adversarial robustness, and real-world capability requirements. Experiments show that Yuvion VL-32B achieves industry-leading safety performance, surpassing comparably sized open-source models and best closed-source commercial models, while maintaining comparable general capabilities.", "date": "2026-06-23", "categories": []}
{"id": "2606.25376", "title": "Transferable Attack against Face Swapping in an Extended Space", "abstract": "Although deep Face Swapping (FS) models may benefit the entertainment industry, they pose severe threats to privacy and security. Existing protections, including deepfake detection and adversarial perturbation, are either passive responses or ineffective to unseen subject-agnostic FS models. In this paper, we propose a transferable attack against subject-agnostic FS models named Additive Identity attack based on a Relighting function (AIR). AIR leverages reillumination and additive perturbations to mislead the identity extraction modules in subject-agnostic FS models. By using these two types of perturbations simultaneously, the attack space is extended such that stronger but more visually natural adversarial examples can be identified. To further enhance the visual quality while preserving the effectiveness of the attack, an adaptive translation-invariant operation and an illumination control scheme are designed for AIR. Unlike other methods, AIR does not require a surrogate FS model to achieve high transferability. In addition, a mathematical proof is given for the extension of the attack space. Extensive experiments using 1000 image pairs across various state-of-the-art subject-agnostic FS models, including GAN and diffusion-based FS models, show that AIR surpasses all existing attacks in terms of both attack success rate and image quality.", "date": "2026-06-24", "categories": []}
{"id": "2606.26207", "title": "The Role of Input Dimensionality in the Emergence and Targeted Control of Adversarial Examples", "abstract": "Several theoretical works have tried to explain the adversarial vulnerability of deep neural networks through properties of high-dimensional geometry. However, the assumptions underlying these works are rarely examined empirically, and systematic evidence remains limited. In this work, we present a systematic study of the role of input dimensionality in both the emergence and the targeted control of adversarial examples. We first analyse the scope and limitations of existing theoretical frameworks based on concentration of measure, showing that real image classes exhibit strong empirical localization, beyond what such theories typically assume. We then conduct an extensive empirical evaluation across hierarchical image datasets spanning a wide range of input dimensionalities and diverse neural architectures. Our results consistently show that adversarial examples become easier to construct as dimensionality increases. We also investigate how input dimensionality affects the additional difficulty of crafting targeted adversarial examples. In particular, we provide theoretical arguments showing that high-dimensional geometry implies that enforcing a specific target label entails only a limited additional distortion compared to untargeted attacks. We corroborate this insight through extensive experiments, demonstrating that the gap between targeted and untargeted perturbations remains small and further narrows as input dimensionality increases. While, taken together, our findings establish high input dimensionality as a fundamental factor underlying the emergence and targeted control of adversarial examples, whether this phenomenon primarily arises from the interplay between high-dimensional geometry and data distributions or from the architectural properties of deep neural networks remains an open question.", "date": "2026-06-24", "categories": []}
{"id": "2606.26285", "title": "TEMPO-Diffusion: Temporally Exposed Malicious Poisoning of Diffusion Models", "abstract": "Noise-based backdoor attacks on diffusion models typically rely on input-time trigger injection, untargeted activation, and out-of-distribution target generation. Such assumptions reduce both the stealthiness and the practical relevance of these attacks. In this work, we present TEMPO-Diffusion, a targeted backdoor framework that localizes the malicious distribution shift to a temporal, in-distribution exposure. TEMPO-Diffusion supports: (i) targeted attacks on and to specific classes, (ii) multiple sub-image backdoors that reconstruct specific features within multiple, different output images and at multiple locations, and (iii) in-painting with time-conditioned triggers. To study relevant, practical security concerns in leveraging backdoored diffusion models for synthetic training data, we also introduce CALISA: a balanced, region-aware traffic-sign dataset emphasizing Canadian and U.S. road signs. Across CIFAR10, GTSRB, and CALISA, our experiments show that TEMPO-Diffusion can reliably poison class-specific synthetic data generation and induce high attack success rates in downstream classifiers trained on that data.", "date": "2026-06-24", "categories": []}
{"id": "2606.26566", "title": "Adversarial Diffusion Across Modalities: A Fusion Survey of Attacks, Defenses, and Evaluation for Text, Vision, and Vision-Language Models", "abstract": "Adversarial evaluation of AI systems has matured along four largely disconnected tracks: diffusion-based attacks on text and large language models (LLMs), diffusion-based attacks on image classifiers, jailbreak pipelines against vision-language models, and diffusion-based input purification defenses. Each has developed its own vocabulary, threat models, and benchmarks, with denoising diffusion models emerging as a shared generative mechanism whose recipes are now actively ported between communities. This survey performs an information-fusion exercise at the meta-research level: we integrate these four tracks into a single conceptual framework with a unified taxonomy, evaluation criteria, and research agenda, focusing on the LLM-side slice. We catalog fifty published papers across four scope areas (text/LLM, image classifier, vision-language model, defense), plus four diffusion-LLM-as-victim entries and ten non-diffusion baselines against which any new attack must be compared. We propose a six-class taxonomy of diffusion roles in adversarial pipelines, augmented by a threat-model axis recording attacker knowledge, query budget, and target accessibility, and apply a five-dimension framework (attack success rate, transferability, query budget, perplexity, defense-evasion) uniformly across modalities. The review adopts a dual attacker-defender perspective: alongside the attack catalog we cover four diffusion-based defenses that form the natural evaluation backdrop for new attacks. Our critical analysis identifies five recurring weaknesses of the current LLM-side literature, and we close with a research agenda of open questions and concrete experimental designs. The companion catalog and spreadsheet are released with the paper. We are explicit that this is a narrative review with quality assessment, not a PRISMA-compliant systematic review, and discuss the implications for replication.", "date": "2026-06-25", "categories": []}
{"id": "2606.26707", "title": "DroidBreaker: Practical and Functional Problem-Space Attacks on Machine-Learning Android Malware Detectors", "abstract": "Adversarial APKs are Android applications modified in the problem space to evade machine-learning malware detectors. In this work, we first show that, despite claims, existing problem-space attacks remain largely impractical. Most techniques leverage software transplantation to inject entire benign modules, introducing many side-effect features and often causing build-time failures. Fine-grained methods that inject only a narrow subset of components exhibit limited effectiveness, while those that also use obfuscation rely on brittle bytecode rewriting, producing APKs that are syntactically valid but semantically unusable. Prior work further overestimates attack success rates by running smoke tests that only validate installation and basic execution, without assessing whether the modified APK still preserves its intended behavior. To overcome these limitations, we present DROIDBREAKER, a practical (build-safe) and functional (semantics-preserving) problem-space attack framework that provides: (i) query-efficient white- and black-box attacks by manipulating only the APK components most influential to the target model; (ii) a set of fine-grained, build-safe manipulations (including injection and obfuscation of API calls, app modules, permissions, and URLs) with minimal side effects; and (iii) a semantics-preserving functionality test that enforces runtime equivalence by comparing execution logs and API-level traces between the initial and the modified APK. Evaluated on a recent corpus of Android applications, DROIDBREAKER achieves high evasion rates with few queries and minimal side effects in both white-box and black-box settings, and drastically reduces detections by commercial malware scanners hosted on VirusTotal.", "date": "2026-06-25", "categories": []}
{"id": "2606.26719", "title": "Full spectrum Unlearnable Examples via Spectral Equalization", "abstract": "Unlearnable examples (UEs) protect training data by injecting imperceptible perturbations so that models fail to extract exploitable representations. In this paper, we reveal that existing UEs exhibit a critical failure once low-pass filtering is applied, indicating that the effective perturbation signals for unlearnability concentrate predominantly in high frequencies. Hence, we argue that reliable UEs should remain effective across the full spectrum. To this end, we propose Full-spectrum Unlearnable examples via Spectral Equalization (FUSE), which aims to generate spectrum-agnostic perturbations by equalizing the contributions from different bands and enforcing cross-band consistency. Specifically, FUSE adopts a Random Spectral Masking (RSM) strategy during generator training, which randomly removes a contiguous frequency band, forcing the remaining bands to maintain unlearnability. In addition, FUSE further integrates Cross-Band Guidance (CBG), which enforces mutual consistency between high- and low-frequency components, thereby further enhancing low-frequency unlearnability and regulating high-frequency perturbations to preserve the semantic fidelity of images. Extensive experiments across multiple datasets, architectures, and spectral filtering demonstrate the strong protection achieved by FUSE.", "date": "2026-06-25", "categories": []}
{"id": "2606.26793", "title": "MIRROR: Novelty-Constrained Memory-Guided MCTS Red-Teaming for Agentic RAG", "abstract": "Multimodal agentic retrieval-augmented generation (RAG) systems expand the attack surface beyond prompt injection to include text poisoning, image injection, direct-query attacks, and orchestrator-level tool manipulation. Existing red-teaming approaches are typically surface-specific and often recycle known attack templates; on text-poisoning benchmarks we measure 73-84% exact duplication. We present MIRROR, a unified cross-surface framework that performs memory-guided Monte Carlo tree search while conditioning candidate generation on retrieved context under an explicit novelty constraint. A deterministic Novelty Gate rejects any candidate matching the retrieval set under normalized comparison, allowing retrieval to inform search priors without enabling prompt copying. Across four attack surfaces on a multimodal agentic RAG target, MIRROR attains 76% ASR on image poisoning compared with 52% for baselines, 97% ASR on orchestrator attacks at half the query cost, and the lowest cross-surface variance (coefficient of variation 0.47). In contrast, specialized baselines collapse across surfaces: suffix optimization reaches 79% ASR on text poisoning but 1% on direct queries. We release ART-SafeBench with 41,815 in-package records and runtime adapters yielding 41,991+ total records across four surfaces.", "date": "2026-06-25", "categories": []}
{"id": "2606.27215", "title": "Vulnerability of Natural Language Classifiers to Evolutionary Generated Adversarial Text", "abstract": "Deep learning models have achieved impressive performance across various fields but remain vulnerable to adversarial inputs, particularly in NLP, where such attacks can have significant real-world consequences. Adversarial attacks often involve small, semantically similar token replacements to fool NLP models, and recent methods have become more precise by targeting specific vulnerable words, often by exploiting some level of access to the model's internal structure. This paper proposes GAversary, a hybrid Genetic Algorithm (GA) to generate adversarial attacks on natural language models. The GA is able to treat the target model as a black box, requiring only the logit value output by the model to guide the search. GAversary differs from GAs previously proposed for this problem by using GloVe embeddings to propose word replacements (the mutation operator) to improve the semantic similarity of the adversarial examples. GAversary is applied to several benchmark data sets and well-known target models. GAversary is able to substantially reduce the target model's accuracy on test data compared to the BAE and A2T attacks compared against (in the best case, reducing a 76.8% accuracy to 5.8%, compared to BAE's 27.6%). The trade-off is that GAversary perturbs just under twice as many words as the other two methods, with a slightly lower semantic similarity to the original text and around a 5% increase in run-time.", "date": "2026-06-25", "categories": []}
{"id": "2606.28416", "title": "AEGIS: A Semantic GAN and Evidential Learning Frameworkfor Robust Adversarial Detection in Vision Sensors", "abstract": "Deep neural networks (DNNs) have shown outstanding performance in visual recognition tasks within vision sensor networks; however, they are still vulnerable to adversarial manipulations and imperceptible perturbations that can lead to erroneous predictions. To address that, this paper presents AEGIS, a semantic aware and uncertainty guided adversarial detection framework designed for robust image classification in vision sensors pipelines. At its core, a SemantiGAN module functions as a multi class semantic discriminator, identifying and filtering visually inconsistent adversarial inputs before they propagate further in the pipeline. For inputs that pass this stage, a stochastic augmentation process generates test time variations, from which handcrafted instability metrics FlipScore, Prediction Inconsistency, Layerwise Cosine Similarity (early and mid layers), and Entropy are computed. These features are aggregated into a compact five dimensional vector and processed by an Evidential Deep Learning (EDL) classifier, which models output evidence using a Dirichlet distribution to yield both class predictions and calibrated uncertainty estimates. Evaluations on the Tiny ImageNet dataset across six categories clean, FGSM, PGD, patch based, functional, and geometric attacks demonstrate the effectiveness of AEGIS. The proposed framework achieves an AUROC of 92.1\\%, an AUPRC of 90.2\\%, and an accuracy of 90.7\\%, outperforming conventional softmax-based detectors in terms of detection performance, robustness, interpretability, and uncertainty calibration.", "date": "2026-06-25", "categories": []}
{"id": "2606.27701", "title": "Room for Error: Large-Scale Simulation of Over-the-Air Acoustic Attacks", "abstract": "While voice control is rapidly becoming a ubiquitous vector of human-AI communication, the risks facing these systems remain poorly understood. This is, in part, a product of the difficulties in scaling strictly digital adversarial workflows to the physical world. These scale barriers have led the community to abstract away key acoustic factors relating to detectability and the influence of geometry on acoustics. These methodological and metrological shortcomings undermine our understanding of risk. We illuminate these issues through real-world testing, conceptual discussions, and a novel, high-throughput reality simulation framework. By testing over 8 million adversarial evaluations, we demonstrate that acoustic awareness yields relative Word Error Rate increases of up to 94.5\\% under Whisper and wav2vec. We employ this framework to explore a formalize and operationalize a Dual-Form Signal to Noise Ratio to decouple source stealth from victim attack efficacy, resolving a crucial limitation in current works. This lays the groundwork for repeatable, verifiable research that embraces, rather than abstracts, the acoustic environment.", "date": "2026-06-26", "categories": []}
{"id": "2606.27784", "title": "Improving Adversarial Robustness via Activation Amplification and Attenuation", "abstract": "The existence of adversarial attacks is often attributed to the presence of non-robust features in neural networks. While prior defenses reduce their impact via pruning, masking, or feature recalibration, we instead propose to jointly learn to amplify and attenuate these signals through a simple activation scaling mechanism. To this end, we introduce Activation Amplification and Attenuation (A3), a lightweight plug-in module that enhances adversarial robustness with minimal modifications of the activations. A3 dynamically rescales the activations using a learnable mask and a scaling factor derived from the original activation magnitudes. The influence of adversarial perturbations can be amplified or attenuated using the same learnable parameters by simply flipping the sign of the scaling operation. The amplified signals serve as negative references to construct novel contrastive and ranking loss functions. Experimental analysis shows that learning to degrade the predictions in amplification mode simultaneously improves adversarial robustness in attenuation mode. Moreover, A3 relies on only a small number of learnable parameters, with most of its behavior being determined by the scaling mechanism rather than additional network capacity. Extensive experiments demonstrate that integrating A3 into different backbones, datasets, and training methods consistently improves adversarial robustness while introducing negligible computational and memory overhead compared to existing plug-in modules. Code is available at: https://github.com/tgoncalv/A3.", "date": "2026-06-26", "categories": []}
{"id": "2606.27832", "title": "USAD: Uncertainty-aware Statistical Adversarial Detection", "abstract": "Statistical adversarial detection (SAD) treats detection as a two-sample test. Given a reference set of clean examples (CEs) and a batch of queries, potentially containing an unknown mixture of CEs and adversarial examples (AEs), SAD decides whether the query distribution drifts away from the CE distribution while controlling the false-alarm rate. Existing SAD-based methods mainly use maximum mean discrepancy (MMD) to measure the distributional discrepancy. However, MMD's distributional properties limit its ability to capture characteristic uncertainty patterns of AEs that are crucial for detection: AEs typically exhibit abnormal feature spread (i.e., global uncertainty) and instability under perturbations (i.e., local uncertainty). To close the gap, we propose Uncertainty-aware Statistical Adversarial Detection (USAD), which explicitly captures these uncertainty patterns with two new statistics: (1) Variance Discrepancy (VD), which measures the difference in feature spread between AEs and CEs to capture global uncertainty differences. (2) Perturbation-based Covariance Discrepancy (PCD), which compares feature covariance under Gaussian perturbations to capture local uncertainty differences. By aggregating VD and PCD, USAD achieves superior detection performances over baseline methods against various adversarial attacks, highlighting the importance of considering characteristic behaviors of AEs for effective SAD. Our code is available at: https://anonymous.4open.science/r/USAD.", "date": "2026-06-26", "categories": []}
{"id": "2606.29112", "title": "A Novel Latent-Class Attack and its Detection by Class Subspace Orthogonalization", "abstract": "Deep learning, which in general relies on voluminous amounts of training data, is vulnerable to data poisoning attacks, including error-generic attacks and backdoors (Trojans). In this work, we propose a new data poisoning attack we dub a latent class attack. Here, all poisoned examples are from a class that is novel (unknown) for the given classification domain and are mislabeled to one of the known classes (the target class) of the domain, so that the model learns to recognize the novel class as a sub-class of the target class. Such attacks could be used e.g. to defeat AI-based access control systems, or could cause a \"foe\" to be classified as a \"friend\". We also propose a post-training defense to detect this attack, without any access to the training set. This detection approach builds on \"class subspace orthogonalization\" (CSO), a plug-and-play paradigm demonstrated to improve existing backdoor detectors. Here, CSO is used to seek an input (a putative unknown class instance) whose internal representation is not aligned with any of the known classes, and yet which is classified with confidence to one of these classes. Finally, specific to image classification domains, we propose a method for visualizing the estimated unknown class instance, providing explainability to our latent class detections.", "date": "2026-06-27", "categories": []}
{"id": "2606.29240", "title": "Blackknife: Hard-Label Query-Limited Black-Box Attacks on Heterogeneous Graph Neural Networks", "abstract": "Heterogeneous graph neural networks (HGNNs) have achieved strong performance in modeling complex graph-structured data with multiple node and relation types. However, their robustness under realistic black-box adversarial settings remains insufficiently explored. Existing attacks on HGNNs usually assume access to model gradients, soft prediction scores, or the complete graph structure, which is often unavailable when HGNN-based services are deployed as closed systems. In this paper, we propose Blackknife, a hard-label, query-limited, and structure-limited black-box evasion attack framework for heterogeneous graph neural networks. Blackknife assumes no access to the victim model architecture, parameters, gradients, logits, confidence scores, or the full graph structure. Instead, it only relies on locally observable one-hop heterogeneous structures and a small number of hard-label queries. To generate effective perturbations under these strict constraints, Blackknife first constructs a local relation-aware surrogate model from observable heterogeneous neighborhoods. It then relaxes discrete edge addition and deletion operations into continuous soft weights and optimizes them through projected gradient descent. Finally, the optimized perturbations are discretized into relation-preserving structural rewiring operations and verified using limited hard-label feedback from the victim model. Extensive experiments on three benchmark heterogeneous graph datasets, including ACM, DBLP, and IMDB, demonstrate that Blackknife consistently achieves strong attack success rates against representative HGNN models. The results further show that Blackknife remains effective under topology-based defense strategies, revealing the vulnerability of HGNNs to local structure-limited black-box attacks.", "date": "2026-06-28", "categories": []}
{"id": "2606.29441", "title": "Closing the Activation-Cone Blind Spot: Response-Time Probing and Unified Defense", "abstract": "Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists. We evaluate five defense paradigms (no defense, static steering, CAST, AlphaSteer, probe-gated) across seven instruction-tuned models (7-31B) and five attack types (GCG, AutoDAN, DeepInception, prefilling, intent laundering). Our central finding: prompt-time activation defenses are structurally blind to prefilling attacks. AlphaSteer achieves 0% attack success on GCG, AutoDAN, and intent laundering but 50% on prefilling. We prove a corollary: any defense that gates intervention on a single layer's activation alignment with a benign reference (cone, subspace, or null-space) is blind to attacks that craft activations to lie inside that reference, whether checked at prompt time or per token. As its constructive contrapositive we introduce response-time probing: a linear probe on the model's hidden state at the first generated tokens, with AUROC 0.97-1.00 across all seven models. Combined with a halt, it cuts prefilling attack success to 0/40 on every model with 0% benign false positives, outperforming Llama Guard 3. Cross-template generalisation depends on probe depth, so we scope the claim to the canonical prefilling-template family. Composing the response-halt with AlphaSteer's null-space steering gives an orthogonal split (the halt catches prefilling, AlphaSteer catches semantic attacks), reaching defense success 0.983 on Mistral and 0.994 on Llama and dominating both components. We further show MMLU fails to capture steering's true utility cost, which appears as behavioral hedging rather than factual loss, and that diverse negative training sets cut probe false positives from 80-100% to near zero. Code, attacks, per-sample results, and the judge prompt are released.", "date": "2026-06-28", "categories": []}
{"id": "2606.29867", "title": "RoAd-RL: A Unified Library and Benchmark for Robust Adversarial Reinforcement Learning", "abstract": "Deep Reinforcement Learning (DRL) has achieved significant success in robotics and autonomous systems, yet remains vulnerable to adversarial perturbations that can severely degrade performance. Research in adversarial reinforcement learning is often limited by fragmented implementations, inconsistent evaluation protocols, and poor reproducibility. To address these challenges, we present \\textbf{RoAd-RL}, an open-source benchmarking framework that provides unified abstractions for policies, attacks, defenses, and robustness metrics, together with reproducible evaluation pipelines and seamless integration with Stable-Baselines3 and Gymnasium.\n  We evaluate DQN, PPO, and SAC agents in LunarLander and Highway-v0 under 192 attack-defense configurations. Results reveal substantial variations in robustness across environments and show that some commonly used defenses can be more detrimental than the attacks they aim to mitigate, while temporal smoothing consistently achieves strong performance. RoAd-RL establishes a standardized benchmark for adversarial reinforcement learning research and is publicly available at https://pypi.org/project/road-rl.", "date": "2026-06-29", "categories": []}
{"id": "2606.29963", "title": "Explainability-Aware Frustum Attack: Exposing Structural Vulnerabilities in LiDAR-Based 3D Object Detectors", "abstract": "The structural vulnerabilities of point cloud-based 3D object detectors remain poorly understood. Prior work has studied adversarial robustness primarily on isolated 3D object models, while recent LiDAR spoofing attacks target richer and more realistic driving scenes but focus mainly on physical realizability rather than understanding detector behavior or attack efficiency. In this work, we investigate how LiDAR-based detectors rely on spatial evidence in complex scenes and whether these reliance patterns can be exploited to induce failures more efficiently. To this end, we propose an explainability-guided adversarial analysis methodology. We introduce the Saliency-LiDAR (SALL) method, which aggregates Integrated Gradient attributions across scenes to produce universal saliency maps for LiDAR-based 3D object detectors. Guided by these maps, we design the Explainability-aware Frustum Attack (EFA), which selectively perturbs only the most influential frustums rather than uniformly attacking entire object regions. Experiments on KITTI and nuScenes, across detectors such as PointPillars and SECOND, show that EFA reduces detection recall by more than 15 percentage points while requiring 25-50% fewer perturbed frustums than the state-of-the-art non-saliency-aware baseline. These findings reveal that modern 3D detectors concentrate discriminative evidence in a small subset of spatial regions, exposing a structural robustness vulnerability in current LiDAR perception systems. Our code is released at https://github.com/SecMindLab/Saliency_LiDAR.", "date": "2026-06-29", "categories": []}
{"id": "2606.30342", "title": "A Classifier-Agnostic Zero-Shot Adversarial Attack Detection via CLIP", "abstract": "Adversarial attacks pose a challenge to the reliability of deep learning models, motivating effective detection methods. Existing techniques often rely on attack-specific assumptions, access to adversarial samples, or knowledge of the underlying classifier (white-box). We propose $A^4D$ Attack- and Architecture-Agnostic Adversarial Detector, a completely black-box, zero-shot adversarial attack detection framework that utilizes prompt-based similarity scores derived from CLIP. To the best of our knowledge this is the first attempt to utilize CLIP for such a task. The method is based on two key observations: (i) CLIP is sensitive even to small imperceptible non-semantic perturbations; (ii) The shift in CLIP embedding space is not arbitrary and can be used as a robust attack indicator. Experiments across multiple attacks, datasets and classifiers validate that $A^4D$ achieves SOTA detection results in the attack-agnostic and classifier-agnostic setting.", "date": "2026-06-29", "categories": []}
{"id": "2606.30807", "title": "Off the Rails: Hijacking the Scoring Head in Generative End-to-End Driving Planners with Safety-Violating Adversarial Perturbations", "abstract": "Generative models have recently seen rapid adoption in End-to-End (E2E) autonomous driving (AD), with diffusion-based denoising and vocabulary-based retrieval becoming the dominant trajectory-decoding paradigms. Despite their architectural diversity, current generative AD planners share a common inference pattern: a fixed set of candidate trajectories (anchors, vocabulary entries, or proposal queries) is scored by one or more learned heads conditioned on the Bird's-Eye-View (BEV) features, and the highest-scored candidate is returned as the final trajectory. Under this design, the scoring head is the only barrier between perception and the motion command, and its decision margins between competing candidates are often small. We introduce \\textsc{Derail}, an adversarial framework that exploits this scoring-head attack surface. Evaluated on various generative planners, \\textsc{Derail} flips the trajectory selection from a safe to an unsafe candidate, with score drops of $39$--$80\\%$ and collision rates of up to $50\\%$, consistently outperforming generic loss-maximization and feature-divergence attacks. Our analysis suggests that safety-violating objectives govern attack effectiveness against generative AD planners, and that the scoring-head inference pattern itself is a recurring attack surface worth explicit defensive consideration.", "date": "2026-06-29", "categories": []}
{"id": "2606.31378", "title": "MAPE: Defending Against Transferable Adversarial Attacks Using Multi-Source Adversarial Perturbations Elimination", "abstract": "Neural networks are vulnerable to meticulously crafted adversarial examples, leading to high-confidence misclassifications in image classification tasks. Due to their consistency with regular input patterns and the absence of reliance on the target model and its output information, transferable adversarial attacks exhibit a notably high stealthiness and detection difficulty, making them a significant focus of defense. In this work, we propose a deep learning defense known as multi-source adversarial perturbations elimination (MAPE) to counter diverse transferable attacks. MAPE comprises the single-source adversarial perturbation elimination (SAPE) mechanism and the pre-trained models probabilistic scheduling algorithm (PPSA). SAPE utilizes a thoughtfully designed channel-attention U-Net as the defense model and employs adversarial examples generated by a pre-trained model (e.g., ResNet) for its training, thereby enabling the elimination of known adversarial perturbations. PPSA introduces model difference quantification and negative momentum to strategically schedule multiple pre-trained models, thereby maximizing the differences among adversarial examples during the defense model's training and enhancing its robustness in eliminating adversarial perturbations. MAPE effectively eliminates adversarial perturbations in various adversarial examples, providing a robust defense against attacks from different substitute models. In a black-box attack scenario utilizing ResNet-34 as the target model, our approach achieves average defense rates of over 95.1\\% on CIFAR-10 and over 71.5\\% on Mini-ImageNet, demonstrating state-of-the-art performance.", "date": "2026-06-30", "categories": []}
{"id": "2606.31653", "title": "Improving Certified Robustness via Adversarial Distillation", "abstract": "Certified training aims to produce models whose predictions can be formally verified against adversarial perturbations, typically by optimising upper bounds on the worst-case loss over an allowed perturbation set. For neural networks, certified training methods based purely on tight relaxation bounds produce networks that are amenable to certification, but sacrifice standard accuracy. Conversely, adversarial training often yields stronger empirical robustness and standard accuracy, but the resulting models are generally difficult to certify with neural network verifiers. Recently, the literature has shown that better standard-certified accuracy trade-offs can be achieved by combining adversarial training objectives with loose over-approximations based on Interval Bound Propagation (IBP), effectively interpolating between lower and upper bounds of the worst-case loss. Building on this, we introduce AD-CERT, a certified training objective that combines adversarial distillation with an IBP upper bound. We show that distilling adversarial information over the logit space from an empirically robust teacher provides an effective lower bound surrogate for certified training, with AD-CERT achieving state-of-the-art certified performance on several robustness benchmarks. Furthermore, in a unified setup, distilling adversarial information at the logit-level is shown to improve certified accuracy over a robust feature-space distillation objective by up to 5.40 percentage points.", "date": "2026-06-30", "categories": []}
{"id": "2607.00174", "title": "Steal the Patch Size: Adversarially Manipulate Vision-Language Models", "abstract": "We present a black-box model-stealing attack that recovers private vision-tokenizer configurations of deployed vision-language models (VLMs), including the visual patch size and input preprocessing pipeline. The key idea is a task-level side channel induced by ViT-style patchification: when a synthetic grid image is aligned with the hidden patch grid, boundary cues are erased at tokenization, causing periodic accuracy drop. By sweeping the grid cell size and measuring these collapses, we infer the patch size; by introducing padding and a consistency-check test, we further identify whether preprocessing is dynamic- or fixed-resolution and recover the target resize resolution. Across open-source Qwen-VL variants and proprietary models including GPT and Claude, we reliably recover tokenizer-related parameters. Finally, we show that such leakage enables preprocessing-aware transfer attacks and model-targeted adversarial manipulation.", "date": "2026-06-30", "categories": []}