{"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": "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": "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": "2606.15730", "title": "InstantForget: Update-Free Backdoor Unlearning with Inference-Time Feature Reset", "abstract": "Backdoor unlearning aims to remove a malicious trigger behavior from a deployed model while preserving clean utility. We study the update-free inference-time setting, where model parameters remain frozen. First, we audit a common projection assumption under oracle paired clean and triggered features. Projection succeeds mainly on BadNets and leaves WaNet, Blended, and SIG at 0.683, 0.888, and 0.941 ASR on CIFAR-10 ResNet-18. This failure is not explained by spectral compactness, spatial locality, or subspace misalignment. It is predicted by a logit-triplet gap involving the target margin, target-logit drop, and non-target logit rise. We then introduce InstantForget, a clean-calibrated gated reset that flags anomalous features with a Mahalanobis score and moves only flagged features toward a neutral non-target representation. With one fixed operating point selected on held-out triggered validation, InstantForget reduces average ASR to 0.071 across four non-adaptive CIFAR-10 triggers without triggered samples or parameter updates at deployment. It also reaches 0.981 detection AUROC and transfers to six of eight tested backbones. Reported failures under WaNet, ModelNet10 point blend, two backbone geometries, and adaptive feature-compactness attacks define the method's scope.", "date": "2026-06-14", "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": "2606.02995", "title": "Patcher: Post-Hoc Patching of Backdoored Large Language Models", "abstract": "Large language models remain vulnerable to jailbreak backdoor attacks, where adversaries poison safety alignment data to embed hidden triggers that bypass safety mechanisms. Existing defenses often require comprehensive attack information or multiple triggered examples, making them impractical when defenders only observe a single reported failure case without knowing whether it stems from a backdoor attack or a natural alignment bug. This paper presents Patcher, a post-hoc defense framework that repairs backdoored language models using only a single reported failure case and the model parameters. Patcher operates in two stages. First, it localizes backdoor triggers by computing response-conditioned gradient-based saliency scores and applying adaptive clustering to separate triggers from benign context. Second, it patches the model through a constrained fine-tuning objective that breaks the trigger-response association while preserving benign-task utility and robustness to non-triggered jailbreak attacks through KL-divergence constraints. We conduct extensive evaluations across multiple backdoor attack strategies and demonstrate that Patcher successfully localizes triggers and neutralizes backdoors while maintaining model utility. We further show robustness against adaptive attacks designed to evade our defense. This work represents a significant step toward practical defenses against training-time attacks in deployed language models.", "date": "2026-06-02", "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": "2606.12498", "title": "From Parameters to Feature Space: Task Arithmetic for Backdoor Mitigation in Model Merging", "abstract": "Model merging (MM) has gained significant attention as a cost-effective approach to integrate multiple task-specific models into a unified model. However, recent work reveals that MM is highly susceptible to backdoor attacks. Existing defenses based on task arithmetic often fail to eliminate backdoors without substantially degrading clean-task performance, owing to their reliance on direct parameter-space editing. To address this gap, we propose Linear Feature Path Minimization (LFPM), a backdoor mitigation framework for model merging, which introduces an anti-backdoor task vector into the backdoored merged model. Unlike prior approaches, LFPM formulates the backdoor robustness of the merged model from a unified feature-space perspective under the Cross-Task Linearity (CTL) framework, which leverages the approximate linearity of features across tasks. This perspective guides the optimization of the anti-backdoor task to suppress backdoors while preserving clean-task performance. Furthermore, we introduce an effective optimization mechanism based on gradient accumulation and loss path-integral, ensuring robust backdoor suppression along the interpolation path. Extensive experiments demonstrate that LFPM consistently exhibits strong robustness against backdoor attacks in both full fine-tuning and Parameter-Efficient Fine-Tuning (PEFT) settings.", "date": "2026-06-10", "categories": []} {"id": "2504.21072", "title": "Erased but Not Forgotten: How Backdoors Compromise Concept Erasure", "abstract": "The expansion of text-to-image diffusion models has raised concerns about harmful outputs, from fabricated depictions of public figures to sexually explicit imagery. To mitigate such risks, prior work has proposed concept erasure methods that aim to sever unwanted concepts from the model via fine-tuning, yet it remains unclear whether these approaches truly remove all links to the harmful concept or merely conceal superficial connections. In this work, we reveal a critical vulnerability, the Erasure Evasion Backdoor (EEB): an adversary binds a backdoor trigger to a concept slated for removal, and this malicious link survives subsequent erasure. We show that both black-box and white-box adversaries can instantiate this threat. Across six state-of-the-art erasure methods, including robust ones that explicitly search for alternative representations of the target concept, EEB consistently exposes harmful content: up to 82% success against celebrity-identity unlearning, up to 94% for object erasure, and up to 16 times amplification of explicit-content exposure. While EEB uncovers a blind spot in current erasure methods, it also provides a diagnostic tool for stress-testing future concept erasure techniques.", "date": "2025-04-29", "categories": []} {"id": "2606.11648", "title": "Dummy Backdoor as a Defense: Removing Unknown Backdoors via Shared Internal Mechanisms for Generative LLMs", "abstract": "Backdoor attacks pose a serious threat to the safety and reliability of Large Language Models (LLMs), as they cause models to behave normally on clean inputs while producing attacker-specified responses when hidden triggers are present. Removing such unknown backdoors is particularly challenging when the defender does not know the backdoor attack types or the internal mechanisms formed through backdoor training. In this work, we propose a simple but effective backdoor removal method based on shared internal mechanisms across different backdoors. First, we show that different backdoors with the same task (attack objective) induce similar trigger-activated changes in the internal activations. Motivated by this observation, our method intentionally embeds a backdoor with a known trigger (\\emph{dummy backdoor}) and then removes it through further fine-tuning on dummy-triggered inputs paired with clean responses. Since the dummy backdoor and the unknown backdoor can rely on shared internal mechanisms, removing the dummy backdoor also reduces the effect of the unknown backdoor. We evaluate our method on three backdoor attack types across multiple model families. Experimental results show that our method substantially reduces the attack success rate of the unknown backdoor while preserving model utility, outperforming representative existing defense methods in both backdoor removal effectiveness and utility preservation. These findings suggest that a defender-controllable backdoor can serve as a helpful proxy for mitigating unknown backdoors in generative LLMs.", "date": "2026-06-10", "categories": []} {"id": "2606.07963", "title": "Shared Latent Structures Enable Unified Backdoor Detection and Mitigation in LLMs", "abstract": "Backdoor attacks in large language models (LLMs) are often treated as isolated trigger-response failures, motivating defenses tailored to specific triggers or behaviors. We show this view is incomplete. Across diverse backdoor behaviors, we identify a shared latent mechanism that can be detected, causally controlled, and suppressed. Using sparse autoencoders (SAEs) on residual-stream activations, we find a small set of latent features consistently activated across jailbreaking, refusal manipulation, password-locking, bias induction, sentiment misclassification, and country-conditioned harmful advice. These features generalize across Qwen3, Gemma~3, and Llama~3.1 models from 4B to 32B parameters, and across both fine-tuning and weight-editing attacks. Through bidirectional activation steering, we show these features are causal: suppressing them reduces attack success, while amplifying them induces target behaviors on clean prompts. We further train lightweight SAE-feature classifiers that generalize zero-shot to unseen backdoors and outperform residual-stream and weight-diffing baselines. Finally, we introduce Concept Ablation Fine-Tuning (CAFT), which suppresses backdoor formation by ablating the shared latent subspace during training. Together, our results suggest that many backdoors rely on a transferable latent mechanism, enabling unified detection and mitigation.", "date": "2026-06-06", "categories": []} {"id": "2606.09499", "title": "Targeting World Models to Compromise Robot Learning Pipelines", "abstract": "World models have recently seen a rapid growth in both their popularity and capability as more data efficient tools for generating robot training data or simulating real world environments, with many works proposing their integration into the robot learning pipeline. While highly practical, in this work we demonstrate that world models introduce a uniquely stealthy and effective data poisoning entry point into the robot learning supply chain that can result in the deployment of unsafe or otherwise compromised robotic policies despite training on seemingly safe ground truth training data. In contrast to traditional data poisoning techniques which directly implant dangerous trajectories into sold or uploaded datasets, our novel attack methods inject malicious prompts or compromising transition dynamics into visibly safe teleoperated datasets which are only activated once fed through a world model as input. This can result in the generation of synthetic, dangerous robot training trajectories and subsequently unsafe or compromised robot policies. We demonstrate the effectiveness of our attacks against both state of the art action conditioned and text conditioned world models, showing a full end-to-end backdoor on a downstream DRL policy and a proof-of-concept for the VLA setting. Overall these findings necessitate research into more secure world models and reevaluating their position within the robot learning supply chain.", "date": "2026-06-08", "categories": []} {"id": "2406.05670", "title": "Certified Robustness to Data Poisoning in Gradient-Based Training", "abstract": "Modern machine learning pipelines leverage large amounts of public data, making it infeasible to guarantee data quality and leaving models open to poisoning and backdoor attacks. Provably bounding model behavior under such attacks remains an open problem. In this work, we address this challenge by developing the first framework providing provable guarantees on the behavior of models trained with potentially manipulated data without modifying the model or learning algorithm. In particular, our framework certifies robustness against untargeted and targeted poisoning, as well as backdoor attacks, for bounded and unbounded manipulations of the training inputs and labels. Our method leverages convex relaxations to over-approximate the set of all possible parameter updates for a given poisoning threat model, allowing us to bound the set of all reachable parameters for any gradient-based learning algorithm. Given this set of parameters, we provide bounds on worst-case behavior, including model performance and backdoor success rate. We demonstrate our approach on multiple real-world datasets from applications including energy consumption, medical imaging, and autonomous driving.", "date": "2024-06-09", "categories": []} {"id": "2606.03785", "title": "Backdoor Unlearning Generalization: A Path Toward the Removal of Unknown Triggers in LLMs", "abstract": "Backdoor attacks in Large Language Models (LLMs) are a growing security concern, where models can generate adversary-chosen content. Existing defenses target backdoors one at a time and typically require knowledge of the trigger, leaving the defender at a structural disadvantage when unknown backdoors may exist in a model. We show that backdoor neutralization through unlearning generalizes across backdoors: training a model to ignore a single trigger can also suppress other backdoors that were never explicitly targeted. We study this phenomenon across three model families, whose backdoors were injected via pretraining or continual pretraining, by analyzing the models obtained after removing one backdoor at a time. To understand why unlearning certain backdoors induces the suppression of others, we introduce the Cross Activation Shift Distance, to quantify the distance between model changes induced by different trainings. Our results open a new direction for LLM safety as defenders could deliberately inject controlled backdoors and then remove them, leveraging cross-backdoor transfer to also suppress unknown backdoors that an attacker may have previously introduced in the model.", "date": "2026-06-02", "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": "2606.02947", "title": "BYORn: Bootstrap Your Own Responses to Defend Large Vision-Language Models Against Backdoor Attacks", "abstract": "Supervised fine-tuning is the predominant approach for adapting autoregressive vision-language models to downstream tasks. Recent work has shown that this paradigm is highly vulnerable to backdoor attacks, and that existing defenses are ineffective in open-ended generation settings. In response, we propose BYORn, a backdoor-robust fine-tuning framework motivated by the observation that poisoned target responses are often semantically implausible given the corresponding image-text inputs and a pretrained model. BYORn identifies such misaligned responses and dynamically replaces them with alternative responses generated by the model, thereby breaking the correlation between triggers and target outputs. The resulting objective gradient corresponds to the gradient of the empirical estimate of the population risk upper bound over the clean data distribution. Empirically, BYORn consistently improves robustness to backdoor attacks while preserving clean-task performance, establishing a new trade-off frontier between generalization and attack success rate. Finally, we demonstrate that BYORn remains effective against adaptive attacks specifically designed to circumvent the proposed defense.", "date": "2026-06-01", "categories": []} {"id": "2508.09456", "title": "IAG: Input-aware Backdoor Attack on VLM-based Visual Grounding", "abstract": "Recent advances in vision-language models (VLMs) have significantly enhanced the visual grounding task, which involves locating objects in an image based on natural language queries. Despite these advancements, the security of VLM-based grounding systems has not been thoroughly investigated. This paper reveals a novel and realistic vulnerability: the first multi-target backdoor attack on VLM-based visual grounding. Unlike prior attacks that rely on static triggers or fixed targets, we propose IAG, a method that dynamically generates input-aware, text-guided triggers conditioned on any specified target object description to execute the attack. This is achieved through a text-conditioned UNet that embeds imperceptible target semantic cues into visual inputs while preserving normal grounding performance on benign samples. We further develop a joint training objective that balances language capability with perceptual reconstruction to ensure imperceptibility, effectiveness, and stealth. Extensive experiments on multiple VLMs (e.g., LLaVA, InternVL, Ferret) and benchmarks (RefCOCO, RefCOCO+, RefCOCOg, Flickr30k Entities, and ShowUI) demonstrate that IAG achieves the best ASRs compared with other baselines on almost all settings without compromising clean accuracy, maintaining robustness against existing defenses, and exhibiting transferability across datasets and models. These findings underscore critical security risks in grounding-capable VLMs and highlight the need for further research on trustworthy multimodal understanding.", "date": "2025-08-13", "categories": []} {"id": "2601.14323", "title": "SilentDrift: Exploiting Action Chunking for Stealthy Backdoor Attacks on Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models are increasingly deployed in safety-critical robotic applications, yet their security vulnerabilities remain underexplored. We identify a fundamental security flaw in modern VLA systems: the combination of action chunking and delta pose representations creates an intra-chunk visual open-loop. This mechanism forces the robot to execute K-step action sequences, allowing per-step perturbations to accumulate through integration. We propose SILENTDRIFT, a stealthy black-box backdoor attack exploiting this vulnerability. Our method employs the Smootherstep function to construct perturbations with guaranteed C2 continuity, ensuring zero velocity and acceleration at trajectory boundaries to satisfy strict kinematic consistency constraints. Furthermore, our keyframe attack strategy selectively poisons only the critical approach phase, maximizing impact while minimizing trigger exposure. The resulting poisoned trajectories are visually indistinguishable from successful demonstrations. Evaluated on the LIBERO, SILENTDRIFT achieves a 93.2% Attack Success Rate with a poisoning rate under 2%, while maintaining a 95.3% Clean Task Success Rate.", "date": "2026-01-20", "categories": []} {"id": "2601.19448", "title": "From Internal Diagnosis to External Auditing: A VLM-Driven Paradigm for Data-Free Online Backdoor Defense", "abstract": "Deep Neural Networks remain inherently vulnerable to backdoor attacks. Traditional test-time defenses largely operate under the paradigm of internal diagnosis methods like model repairing or input robustness, yet these approaches are often fragile under advanced attacks as they remain entangled with the victim model's corrupted parameters. We propose a paradigm shift from Internal Diagnosis to External Semantic Auditing, arguing that effective defense requires decoupling safety from the victim model via an independent, semantically grounded auditor. To this end, we present a framework harnessing Universal Vision-Language Models (VLMs) as evolving semantic gatekeepers. We introduce PRISM (Prototype Refinement & Inspection via Statistical Monitoring), which overcomes the domain gap of general VLMs through two key mechanisms: a Hybrid VLM Teacher that dynamically refines visual prototypes online, and an Adaptive Router powered by statistical margin monitoring to calibrate gating thresholds in real-time. Extensive evaluation across 17 datasets and 11 attack types demonstrates that PRISM achieves state-of-the-art performance, suppressing Attack Success Rate to <1% on CIFAR-10 while improving clean accuracy, establishing a new standard for model-agnostic, externalized security.", "date": "2026-01-27", "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": "2605.27809", "title": "Density-aware Sample-specific Attack", "abstract": "Despite recent progress in backdoor attacks, existing methods remain susceptible to post-training defenses that erase the backdoor through fine-tuning or pruning. We revisit the core objectives of backdoor attacks and derive principled criteria characterizing optimal sample-specific trigger construction under a Bayes-optimal model of the victim's training. Our analysis reveals that both attack success and clean-accuracy preservation are simultaneously optimized when triggered samples are steered into low-density regions of the clean data distribution, a distributional condition that controls all moments of the poisoned distribution at once rather than a handful of input-space summary statistics. We introduce a bilevel optimization framework that estimates density ratios via conditional time-score matching and optimizes a mixture-model objective to place triggered samples in these sparse regions. Extensive evaluations on MNIST, CIFAR-10, GTSRB, and TinyImageNet demonstrate that our method achieves above 99\\% attack success rate before defense and retains 50--85 percentage points higher post-defense ASR than the strongest baselines under fine-tuning defenses. Against neuron-pruning defenses, the method exhibits complete immunity, with zero neurons identified for removal across all pruning thresholds. These results expose a fundamental gap in current defense paradigms and underscore the need for defenses that operate beyond the support of the clean distribution.", "date": "2026-05-27", "categories": []} {"id": "2605.29960", "title": "Hijacking Agent Memory: Stealthy Trojan Attacks Through Conversational Interaction", "abstract": "Large language model (LLM) agents increasingly leverage long term memory to support persistent and autonomous task execution. However, this capability also introduces a new attack surface: memory poisoning, where adversaries can inject malicious information to influence future behavior. Existing memory poisoning attacks often assume that injected content can be stored directly in memory, overlooking the selective extraction and rewriting stages in modern memory pipelines. This makes prior methods ineffective under realistic settings.\n In this paper, we propose MemPoison, a novel memory poisoning attack that bypasses selective memory mechanisms in LLM agents, where an attacker can inject triggerable backdoors into the agent's long-term memory through dialogue interactions, thereby misleading its subsequent responses. MemPoison introduces three key components: (i) a semantic relational bridge that binds the trigger and payload into a coherent statement to ensure they are extracted into memory together; (ii) entity masquerading that optimizes triggers to mimic named entities, resisting rewriting; and (iii) joint embedding optimization that shapes trigger-injected texts into a tight cluster in the embedding space while maintaining isolation from benign embeddings for stealth. Evaluations across different agent domains and memory mechanisms show MemPoison achieves attack success rates up to 0.95, outperforming existing baselines. Mechanistic analysis indicates that the attack exploits embedding-space anisotropy and shifts attention patterns, highlighting core vulnerabilities in selective memory systems. We evaluate multiple defense strategies and demonstrate their fundamental limitations in mitigating the attack.", "date": "2026-05-28", "categories": []} {"id": "2605.27416", "title": "Can Quantum Federated Learning Withstand Circuit-Level Backdoors?", "abstract": "Quantum Federated Learning (QFL) inherits the core vulnerability of federated optimization to malicious clients, while also introducing an attack surface from variational circuit training and measurement-driven gradients. This work proposes a novel CircUit-Level backdoor Threat (CULT) model that formalizes four stealthy attacks by exploiting quantum-aware mechanisms, including Grover, Pauli, Bit-flip, and Sign-flip. By enabling malicious clients on both in-training and post-training surfaces, these attacks can critically undermine the learning process. We establish a rigorous theoretical foundation to demonstrate attack stealthiness under standard smoothness assumptions. Experiments on the MNIST and CIFAR-10 datasets with non-IID splits and varying fractions of malicious clients show that even a single malicious client can induce severe accuracy degradation under FedAvg aggregation. While popular defenses, including Krum, Multi-Krum, FoolsGold, FLGuardian, and Mud-HoG, reduce degradation in many regimes, they fail to eliminate worst-case failure cases, where accuracy drops up to 50\\%. The experimental analysis further reveals that under the CULT model, malicious updates effectively mask their presence by staying close to benign norms, thereby helping attackers evade detection.", "date": "2026-05-18", "categories": []} {"id": "2605.27631", "title": "Poison with Style: A Practical Poisoning Attack on Code Large Language Models", "abstract": "Code Large Language Models (CLLMs) serve as the core of modern code agents, enabling developers to automate complex software development tasks. In this paper, we present Poison-with-Style (PwS), a practical and stealthy model poisoning attack targeting CLLMs. Unlike prior attacks that assume an active adversary capable of directly embedding explicit triggers (e.g., specific words) into developers' prompts during inference, PwS leverages developers' code styles as covert triggers implicitly embedded within their prompts. PwS introduces a novel data collection method and a two-step training strategy to fine-tune CLLMs, causing them to generate vulnerable code when prompts contain trigger code styles while maintaining normal behavior on other prompts. Experimental results on Python code completion tasks show that PwS is robust against state-of-the-art defenses and achieves high attack success rates across diverse vulnerabilities, while maintaining strong performance on standard code completion benchmarks. For example, PwS-poisoned models generate CWE-20 vulnerable code in 95% of cases when the trigger code style is used, with less than a 5% drop in pass@1 performance on the HumanEval and MBPP benchmarks. Our implementation and dataset are here: https://github.com/khangtran2020/pws.", "date": "2026-05-26", "categories": []} {"id": "2605.28597", "title": "Position: Retire the \"Positive Backdoor\" Label -- Secret Alignment Requires Strict and Systematic Evaluation", "abstract": "This position paper argues that the AI/ML community should stop overclaiming and retire the label \"positive backdoor,\" and instead treat trigger-activated hidden behaviors as Secret Alignment. Crucially, protective claims based on Secret Alignment should be presumed not secure by default unless supported by rigorous, standardized evaluation. The Private AI era, enabled by open-weight LLMs and accessible training/inference stacks, turns language models into privately owned digital assets, creating security concerns around unauthorized access, model theft, and behavioral misuse. Recently, a line of work framed as \"positive backdoors\" has been proposed to address these challenges. To ground our position in evidence, we unify these proposals as covert trigger-behavior associations for access gating, ownership attribution, and safety enforcement, and evaluate three representative applications across six core properties: effectiveness, harmlessness, persistence, efficiency, robustness, and reliability. Our results reveal substantial brittleness - especially in the confidentiality, integrity, and availability (CIA) - of trigger-behavior mappings often underrepresented by existing claims. We further relate these outcomes to behavior density and decision complexity, offering a behavioral lens for understanding deployment-time risks and motivating community-wide evaluation that makes Secret Alignment claims provable.", "date": "2026-05-27", "categories": []} {"id": "2511.12046", "title": "BackWeak: Backdooring Knowledge Distillation Simply with Weak Triggers and Fine-tuning", "abstract": "Knowledge Distillation (KD) is essential for compressing large models, yet relying on pre-trained \"teacher\" models downloaded from third-party repositories introduces serious security risks--most notably backdoor attacks. Existing KD backdoor methods are typically complex and computationally intensive: they employ surrogate student models and simulated distillation to guarantee transferability, and construct triggers similar to universal adversarial perturbations (UAPs), which being not stealthy in magnitude, inherently exhibit strong adversarial behavior. This work questions whether such complexity is necessary and constructs stealthy \"weak\" triggers--imperceptible perturbations that have negligible adversarial effect. We propose BackWeak, a simple, surrogate-free attack paradigm. BackWeak shows that a powerful backdoor can be implanted by simply fine-tuning a benign teacher with a weak trigger using a very small learning rate. We demonstrate that this delicate fine-tuning is sufficient to embed a backdoor that reliably transfers to diverse student architectures during a victim's standard distillation process, yielding high attack success rates. Extensive empirical evaluations on multiple datasets, model architectures, and KD methods show that BackWeak is efficient, simpler, and often more stealthy than previous elaborate approaches. This work calls on researchers studying KD backdoor attacks to pay particular attention to the trigger's potential adversarial characteristics.", "date": "2025-11-15", "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": "2605.22481", "title": "When Stronger Triggers Backfire: A High-Dimensional Theory of Backdoor Attacks", "abstract": "Backdoor poisoning attacks behave counter-intuitively in high dimensions: stronger training triggers can help the defender. We study regularised generalised linear models on Gaussian-mixture data in the proportional regime ($p/n \\to \\kappa$), varying the training trigger strength $\\alpha$ against a fixed test trigger. Three phenomena emerge: (i) clean test accuracy increases with $\\alpha$; (ii) attack success peaks at a finite $\\alpha$ and then declines; and (iii) the most damaging trigger direction is the minimum eigenvector of the data covariance. We prove all three results in closed form for the squared loss, and extend (i) and (ii) to general convex GLM losses via a Gaussian-proxy fixed-point system. We identify a finite-sample noise floor proportional to $\\kappa$ as the mechanism behind (i), invisible to classical $n \\gg p$ analysis. Experiments on CIFAR-10 and Gaussian surrogates match the theory closely; ResNet-18 experiments show the same phenomena beyond the convex setting.", "date": "2026-05-21", "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": "2605.18907", "title": "Lightweight and Fast Backdoor Model Detection", "abstract": "Deep neural networks (DNN), despite their remarkable performance, are highly vulnerable to backdoor attacks. Existing defenses mainly rely on activation anomaly analysis or trigger reverse engineering and often require clean samples or prior knowledge of trigger patterns, resulting in limited efficacy, practicability, and generalizability. More critically, while advanced attacks can implement backdoor implantation in milliseconds, current detection approaches typically demand minutes or even hours. To this end, we propose DFBScanner, a lightweight static parameter inspection framework for fast backdoor scanning. DFBScanner leverages our key observation that backdoor-induced feature perturbations can lead to distinctive and anomalous parameter updates in the final classification layer. Hence, we shift our detection focus from recognizing diverse and attack-specific trigger patterns targeted by prior work, to identifying the unified backdoor manifestation within the final layer, thereby enabling efficient and attack-agnostic detection. Specifically, by constructing and strategically combining multiple anomaly indicators of the final-layer parameters into a Trojan clue, DFBScanner detects backdoors through maximum anomaly scoring. DFBScanner is evaluated on a large-scale backdoor benchmark, including over 5,000 backdoor models trained on 4 datasets, 12 network architectures, 20 types of backdoor triggers, 2 attack strategies (all-to-one and -all), and 3 backdoor injection methods (data poisoning, training pipeline manipulation, and bit-flips). Numerical results show that DFBScanner achieves a 97.17% true-positive rate, 0.95% false-positive rate, and an average detection time of only 1 ms per model, significantly outperforming prior methods.", "date": "2026-05-17", "categories": []} {"id": "2605.18908", "title": "Fast and Lightweight Backdoor Detection via Head Random Probing", "abstract": "Deep neural networks (DNNs) remain critically vulnerable to backdoor attacks. Existing post-training detectors often require clean or surrogate data, gradients, or iterative trigger reconstruction, leading to high computational costs and limited robustness under practical model-auditing scenarios. In this paper, we propose HTell, a fast and lightweight data-free backdoor detector based on head random probing. Instead of reconstructing diverse trigger patterns, HTell inspects their unified manifestation in the prediction head: backdoored models tend to exhibit abnormal response concentration on the target class under random latent probes. HTell generates architecture-aware random latent probes, feeds them directly into the model head, and detects backdoors by analyzing class-wise response statistics, without accessing real or surrogate data, model gradients, or parameter optimization. We evaluate HTell on a large-scale benchmark containing more than 6,000 backdoored models and over 700 clean models, covering 4 datasets, 14 architectures, and 21 types of backdoor attacks. HTell achieves 99.03% true positive rate and 2.11% false positive rate with only 12.69 ms/model detection latency, reducing the time cost by over 30,000$\\times$ compared with representative gradient-based detectors. These results demonstrate that head random probing provides an accurate, robust, and efficient solution for large-scale data-free backdoor model auditing.", "date": "2026-05-17", "categories": []} {"id": "2605.19227", "title": "Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models", "abstract": "Unified autoregressive models (UAMs) are transformer models that generate text as well as image tokens within a single autoregressive pass. Shared parameters and a multimodal vocabulary simplify the training pipeline and facilitate flexible multimodal generation, yet might introduce new vulnerabilities. In particular, we are the first to show that this unified architecture enables multimodal backdoor attacks, where a trigger can propagate malicious effects across multiple output modalities. Specifically, we present the Token by Token Backdoor Attack (ToBAC), the first backdoor attack targeting UAMs, exploring both data-based and model-based poisoning strategies. We demonstrate that innocuous characters or even common words can be transformed into triggers that elicit harmful behavior in autoregressive image generation. ToBAC can jointly manipulate visual outputs and accompanying text, increasing the perceived authenticity of fabricated content. With model access, ToBAC enables attacks on the unified Liquid model in which a subtle word (e.g., ``cool'') induces modality-aligned brand promotion or ideological influence in 55% of generations. Without model access, ToBAC can be induced through data poisoning, achieving an average success rate of 63.1% against JanusPro.", "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": "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": "2605.15172", "title": "MetaBackdoor: Exploiting Positional Encoding as a Backdoor Attack Surface in LLMs", "abstract": "Backdoor attacks pose a serious security threat to large language models (LLMs), which are increasingly deployed as general-purpose assistants in safety- and privacy-critical applications. Existing LLM backdoors rely primarily on content-based triggers, requiring explicit modification of the input text. In this work, we show that this assumption is unnecessary and limiting. We introduce MetaBackdoor, a new class of backdoor attacks that exploits positional information as the trigger, without modifying textual content. Our key insight is that Transformer-based LLMs necessarily encode token positions to process ordered sequences. As a result, length-correlated positional structure is reflected in the model's internal computation and can be used as an effective non-content trigger signal.\n We demonstrate that even a simple length-based positional trigger is sufficient to activate stealthy backdoors. Unlike prior attacks, MetaBackdoor operates on visibly and semantically clean inputs and enables qualitatively new capabilities. We show that a backdoored LLM can be induced to disclose sensitive internal information, including proprietary system prompts, once a length condition is satisfied. We further demonstrate a self-activation scenario, where normal multi-turn interaction can move the conversation context into the trigger region and induce malicious tool-call behavior without attacker-supplied trigger text. In addition, MetaBackdoor is orthogonal to content-based backdoors and can be composed with them to create more precise and harder-to-detect activation conditions.\n Our results expand the threat model of LLM backdoors by revealing positional encoding as a previously overlooked attack surface. This challenges defenses that focus on detecting suspicious text and highlights the need for new defense strategies that explicitly account for positional triggers in modern LLM architectures.", "date": "2026-05-14", "categories": []} {"id": "2510.10265", "title": "Backdoor Collapse: Eliminating Unknown Threats via Known Backdoor Aggregation in Language Models", "abstract": "Backdoor attacks are a significant threat to large language models (LLMs), often embedded via public checkpoints, yet existing defenses rely on impractical assumptions about trigger settings. To address this challenge, we propose \\ourmethod, a defense framework that requires no prior knowledge of trigger settings. \\ourmethod is based on the key observation that when deliberately injecting known backdoors into an already-compromised model, both existing unknown and newly injected backdoors aggregate in the representation space. \\ourmethod leverages this through a two-stage process: \\textbf{first}, aggregating backdoor representations by injecting known triggers, and \\textbf{then}, performing recovery fine-tuning to restore benign outputs. Extensive experiments across multiple LLM architectures demonstrate that: (I) \\ourmethod reduces the average Attack Success Rate to 4.41\\% across multiple benchmarks, outperforming existing baselines by 28.1\\%$\\sim$69.3\\%$\\uparrow$. (II) Clean accuracy and utility are preserved within 0.5\\% of the original model, ensuring negligible impact on legitimate tasks. (III) The defense generalizes across different types of backdoors, confirming its robustness in practical deployment scenarios.", "date": "2025-10-11", "categories": []} {"id": "2605.13115", "title": "DiffusionHijack: Supply-Chain PRNG Backdoor Attack on Diffusion Models and Quantum Random Number Defense", "abstract": "Diffusion models depend on pseudo-random number generators (PRNGs) for latent noise sampling. We present DiffusionHijack, a supply-chain backdoor attack that hijacks the PRNG to deterministically control generated images. A malicious PRNG, injected via compromised packages, forces pixel-perfect reproduction of attacker-chosen content (SSIM = 1.00, N = 100 trials) on Stable Diffusion v1.4, v1.5, and SDXL -- without modifying model weights. The attack is inherently undetectable by existing model auditing and content moderation mechanisms, as it operates entirely outside the neural network computation graph. The attack remains effective under stochastic sampling (eta > 0), bypasses CLIP-based safety checkers (98-100% success), and operates independently of the user's prompt. As a countermeasure, we replace the PRNG with a quantum random number generator (QRNG), which provides information-theoretic unpredictability. Across N = 100 prompt-model combinations, QRNG defense completely neutralizes the attack, reducing output similarity to random baseline levels (SSIM < 0.20 for SD 1.x models, < 0.45 for SDXL). This work exposes a previously overlooked supply-chain vulnerability and offers a hardware-level fundamental mitigation for generative AI systems.", "date": "2026-05-13", "categories": []} {"id": "2605.11612", "title": "When Emotion Becomes Trigger: Emotion-style dynamic Backdoor Attack Parasitising Large Language Models", "abstract": "Backdoor vulnerabilities widely exist in the fine-tuning of large language models(LLMs). Most backdoor poisoning methods operate mainly at the token level and lack deeper semantic manipulation, which limits stealthiness. In addition, Prior attacks rely on a single fixed trigger to induce harmful outputs. Such static triggers are easy to detect, and clean fine-tuning can weaken the trigger-target association. Through causal validation, we observe that emotion is not directly linked to individual words, but functions as an overall stylistic factor through tone. In the representation space of LLM, emotion can be decoupled from semantics, forming distinct cluster from the original neutral text. Therefore, we consider the emotional factor as the backdoor trigger to propose a pparasitic emotion-style dynamic backdoor attack, Paraesthesia. By mixing samples with the emotional trigger into clean data and then fine-tuning the model, the model is able to generate the predefined attack response when encountering emotional inputs during the inference stage. Paraesthesia includes two the quantification and rewriting of emotional styles. We evaluate the effectiveness of our method on instruction-following generation and classification tasks. The experimental results show that Paraesthesia achieves an attack success rate of around 99\\% across both task types and four different models, while maintaining the clean utility of the models.", "date": "2026-05-12", "categories": []} {"id": "2605.11996", "title": "BadSKP: Backdoor Attacks on Knowledge Graph-Enhanced LLMs with Soft Prompts", "abstract": "Recent knowledge graph (KG)-enhanced large language models (LLMs) move beyond purely textual knowledge augmentation by encoding retrieved subgraphs into continuous soft prompts via graph neural networks, introducing a graph-conditioned channel that operates alongside the standard text interface. However, existing backdoor attacks are largely designed for the textual channel, and their effectiveness against this dual-channel architecture remains unclear. We show that this architecture creates a robustness gap: text-channel backdoor attacks that readily compromise textual KG prompting systems become largely ineffective against soft-prompt-based counterparts. We interpret this gap through semantic anchoring, whereby graph-derived soft prompts bias the generation-driving hidden state toward query-consistent semantics and suppress surface-level malicious instructions. Because this anchoring effect is itself induced by the graph channel, an attacker who manipulates graph-level representations can in turn redirect it toward adversarial semantics. To demonstrate this risk, we propose BadSKP, a backdoor attack that targets the graph-to-prompt interface through a multi-stage optimization strategy: it constructs adversarial target embeddings, optimizes poisoned node embeddings to steer the induced soft prompt, and approximates the optimized representations with fluent adversarial node attributes. Experiments on two soft-prompt KG-enhanced LLMs across four datasets show that BadSKP achieves high attack success under both frozen and trojaned settings, while text-only attacks remain unreliable even under perplexity-based defenses.", "date": "2026-05-12", "categories": []} {"id": "2403.18136", "title": "Identifying Backdoored Graphs in Graph Neural Network Training: An Explanation-Based Approach with Novel Metrics", "abstract": "Graph Neural Networks (GNNs) have gained popularity in numerous domains, yet they are vulnerable to backdoor attacks that can compromise their performance and ethical application. The detection of these attacks is crucial for maintaining the reliability and security of GNN classification tasks, but existing methods are often inflexible, relying on single metrics that fail to capture the full range of backdoor behaviors. Recognizing the challenge in detecting such intrusions, we devised a novel detection method that creatively leverages graph-level explanations. By extracting and transforming secondary outputs from GNN explanation mechanisms, we developed seven innovative metrics for effective detection of backdoor attacks on GNNs. Additionally, we develop an adaptive attack to rigorously evaluate our approach. We test our method on multiple benchmark datasets and examine its efficacy against various attack models. Our results show that our method can achieve high detection performance, marking a significant advancement in safeguarding GNNs against backdoor attacks.", "date": "2024-03-26", "categories": []} {"id": "2605.08280", "title": "Beyond the False Trade-off: Adaptive EWC for Stealthy and Generalizable T2I Backdoors", "abstract": "Preserving model fidelity is essential for stealthy text-to-image (T2I) backdoor attacks. Existing methods such as Learning without Forgetting (LwF) rely on output-based distillation, which provides limited regularization. We introduce Elastic Weight Consolidation (EWC) as a parameter-based alternative for preserving fidelity in backdoor learning. While stronger in principle, we show that standard static EWC with a fixed regularization weight lambda and mean-squared utility loss creates an artificial trade-off between attack success rate (ASR) and fidelity, particularly degrading performance on weak triggers. To address this, we propose Cosine-Aware Adaptive EWC, which dynamically adjusts EWC regularization using a cosine-based semantic utility and adaptive scheduling. This approach transforms EWC from a fixed penalty into a context-sensitive constraint, maintaining high ASR while preserving model fidelity. Experiments demonstrate improved ASR-fidelity balance and enhanced robustness on out-of-domain (OOD) datasets compared to existing baselines.", "date": "2026-05-08", "categories": []} {"id": "2605.08313", "title": "Seed Hijacking of LLM Sampling and Quantum Random Number Defense", "abstract": "Large language models (LLMs) rely on deterministic pseudorandom number generators (PRNGs) for autoregressive sampling, creating a critical supply-chain attack surface overlooked by existing defenses. We present SeedHijack, a backdoor attack that manipulates PRNG outputs to force attacker-specified token selection without altering model logits. In a 540-trial benchmark on GPT-2 (124M), the attack achieves 99.6% exact token injection across 9 sampling configurations; it reaches 100% success on four aligned models (1.5B-7B, RLHF/SFT/reasoning distillation) and bypasses all alignment methods tested in this work. We further propose a defense based on a hardware quantum random number generator (QRNG), which neutralizes the attack in our evaluated threat model with negligible median overhead (+0.6% latency, +7.7 MB memory). Our work identifies a critical sampling-layer vulnerability and provides a practical, deployable QRNG-based defense.", "date": "2026-05-08", "categories": []} {"id": "2605.09005", "title": "Towards Backdoor-Based Ownership Verification for Vision-Language-Action Models", "abstract": "Vision-Language-Action models (VLAs) support generalist robotic control by enabling end-to-end decision policies directly from multi-modal inputs. As trained VLAs are increasingly shared and adapted, protecting model ownership becomes essential for secure deployment and responsible open-source usage. In this paper, we present GuardVLA, the first backdoor-based ownership verification framework specifically designed for VLAs. GuardVLA embeds a stealthy and harmless backdoor watermark into the protected model during training by injecting secret messages into embodied visual data. For post-release verification, we propose a swap-and-detect mechanism, in which the trigger projector and an external classifier head are used to activate and detect the embedded backdoor based on prediction probabilities. Extensive experiments across multiple datasets, model architectures, and adaptation settings demonstrate that GuardVLA enables reliable ownership verification while preserving benign task performance. Further results show that the embedded watermark remains detectable under post-release model adaptation.", "date": "2026-05-09", "categories": []} {"id": "2411.12220", "title": "DeTrigger: A Gradient-Centric Approach to Backdoor Attack Mitigation in Federated Learning", "abstract": "Federated Learning (FL) enables collaborative model training across distributed devices while preserving local data privacy, making it ideal for mobile and embedded systems. However, the decentralized nature of FL also opens vulnerabilities to model poisoning attacks, particularly backdoor attacks, where adversaries implant trigger patterns to manipulate model predictions. In this paper, we propose DeTrigger, a scalable and efficient backdoor-robust federated learning framework that leverages insights from adversarial attack methodologies. By employing gradient analysis with temperature scaling, DeTrigger detects and isolates backdoor triggers, allowing for precise model weight pruning of backdoor activations without sacrificing benign model knowledge. Extensive evaluations across four widely used datasets demonstrate that DeTrigger achieves up to 251x faster detection than traditional methods and mitigates backdoor attacks by up to 98.9%, with minimal impact on global model accuracy. Our findings establish DeTrigger as a robust and scalable solution to protect federated learning environments against sophisticated backdoor threats.", "date": "2024-11-19", "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": "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": "2605.05977", "title": "BehaviorGuard: Online Backdoor Defense for Deep Reinforcement Learning", "abstract": "Backdoor attacks pose a serious threat to deep reinforcement learning (DRL). Current defenses typically rely on reward anomalies to reverse-engineer triggers and model finetuning to remove backdoors. However, complex trigger patterns undermine their robustness, and fine-tuning entails high costs, limiting practical utility. Therefore, we shift defense concerns to trigger-agnostic backdoor output behaviors and propose BehaviorGuard, an online behavior-based backdoor detection and mitigation framework for DRL. Specifically, we find that regardless of attacks, backdoored policies induce consistent shifts in action distributions to ensure reliable activation, leaving detectable traces in high-quantile regions and distribution tails, even in the absence of triggers. Based on this, we design a novel metric that captures behavioral drift in action distributions to identify and suppress backdoor actions at runtime. To our knowledge, this is the first online backdoor defense that counters attacks both in single- and multi-agent DRL. Evaluated across diverse benchmarks with different backdoor attacks, BehaviorGuard consistently surpasses prior methods in both efficacy and efficiency.", "date": "2026-05-07", "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": "2410.05284", "title": "Hypnopaedia-Aware Machine Unlearning via Psychometrics of Artificial Mental Imagery", "abstract": "Neural backdoors represent insidious cybersecurity loopholes that render learning machinery vulnerable to unauthorised manipulations, potentially enabling the weaponisation of artificial intelligence with catastrophic consequences. A backdoor attack involves the clandestine infiltration of a trigger during the learning process, metaphorically analogous to hypnopaedia, where ideas are implanted into a subject's subconscious mind under the state of hypnosis or unconsciousness. When activated by a sensory stimulus, the trigger evokes a conditioned reflex that directs a machine to mount a predetermined response. In this study, we propose a cybernetic framework for constant surveillance of backdoor threats, driven by the dynamic nature of untrustworthy data sources. We develop a self-aware unlearning mechanism to autonomously detach a machine's behaviour from the backdoor trigger. Through reverse engineering and statistical inference, we detect deceptive patterns and estimate the likelihood of backdoor infection. We employ model inversion to elicit artificial mental imagery, using stochastic processes to disrupt optimisation pathways and avoid convergent but potentially flawed patterns. This is followed by hypothesis analysis, which estimates the likelihood of each potentially malicious pattern as the true trigger and infers the probability of infection. The primary objective of this study is to maintain a stable state of equilibrium between knowledge fidelity and backdoor vulnerability.", "date": "2024-09-29", "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": "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": "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": "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": "2604.21700", "title": "Stealthy Backdoor Attacks against LLMs Based on Natural Style Triggers", "abstract": "The growing application of large language models (LLMs) in safety-critical domains has raised urgent concerns about their security. Many recent studies have demonstrated the feasibility of backdoor attacks against LLMs. However, existing methods suffer from three key shortcomings: explicit trigger patterns that compromise naturalness, unreliable injection of attacker-specified payloads in long-form generation, and incompletely specified threat models that obscure how backdoors are delivered and activated in practice. To address these gaps, we present BadStyle, a complete backdoor attack framework and pipeline. BadStyle leverages an LLM as a poisoned sample generator to construct natural and stealthy poisoned samples that carry imperceptible style-level triggers while preserving semantics and fluency. To stabilize payload injection during fine-tuning, we design an auxiliary target loss that reinforces the attacker-specified target content in responses to poisoned inputs and penalizes its emergence in benign responses. We further ground the attack in a realistic threat model and systematically evaluate BadStyle under both prompt-induced and PEFT-based injection strategies. Extensive experiments across seven victim LLMs, including LLaMA, Phi, DeepSeek, and GPT series, demonstrate that BadStyle achieves high attack success rates (ASRs) while maintaining strong stealthiness. The proposed auxiliary target loss substantially improves the stability of backdoor activation, yielding an average ASR improvement of around 30% across style-level triggers. Even in downstream deployment scenarios unknown during injection, the implanted backdoor remains effective. Moreover, BadStyle consistently evades representative input-level defenses and bypasses output-level defenses through simple camouflage.", "date": "2026-04-23", "categories": []} {"id": "2504.05605", "title": "ShadowCoT: Cognitive Hijacking for Stealthy Reasoning Backdoors in LLMs", "abstract": "Chain-of-Thought (CoT) enhances an LLM's ability to perform complex reasoning tasks, but it also introduces new security issues. In this work, we present ShadowCoT, a novel backdoor attack framework that targets the internal reasoning mechanism of LLMs. Unlike prior token-level or prompt-based attacks, ShadowCoT directly manipulates the model's cognitive reasoning path, enabling it to hijack multi-step reasoning chains and produce logically coherent but adversarial outcomes. By conditioning on internal reasoning states, ShadowCoT learns to recognize and selectively disrupt key reasoning steps, effectively mounting a self-reflective cognitive attack within the target model. Our approach introduces a lightweight yet effective multi-stage injection pipeline, which selectively rewires attention pathways and perturbs intermediate representations with minimal parameter overhead (only 0.15% updated). ShadowCoT further leverages reinforcement learning and reasoning chain pollution (RCP) to autonomously synthesize stealthy adversarial CoTs that remain undetectable to advanced defenses. Extensive experiments across diverse reasoning benchmarks and LLMs show that ShadowCoT consistently achieves high Attack Success Rate (94.4%) and Hijacking Success Rate (88.4%) while preserving benign performance. These results reveal an emergent class of cognition-level threats and highlight the urgent need for defenses beyond shallow surface-level consistency.", "date": "2025-04-08", "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": "2604.19083", "title": "ProjLens: Unveiling the Role of Projectors in Multimodal Model Safety", "abstract": "Multimodal Large Language Models (MLLMs) have achieved remarkable success in cross-modal understanding and generation, yet their deployment is threatened by critical safety vulnerabilities. While prior works have demonstrated the feasibility of backdoors in MLLMs via fine-tuning data poisoning to manipulate inference, the underlying mechanisms of backdoor attacks remain opaque, complicating the understanding and mitigation. To bridge this gap, we propose ProjLens, an interpretability framework designed to demystify MLLMs backdoors. We first establish that normal downstream task alignment--even when restricted to projector fine--tuning--introduces vulnerability to backdoor injection, whose activation mechanism is different from that observed in text-only LLMs. Through extensive experiments across four backdoor variants, we uncover:(1) Low-Rank Structure: Backdoor injection updates appear overall full-rank and lack dedicated ``trigger neurons'', but the backdoor-critical parameters are encoded within a low-rank subspace of the projector;(2) Activation Mechanism: Both clean and poisoned embedding undergoes a semantic shift toward a shared direction aligned with the backdoor target, but the shifting magnitude scales linearly with the input norm, resulting in the distinct backdoor activation on poisoned samples. Our code is available at: https://anonymous.4open.science/r/ProjLens-8FD7", "date": "2026-04-21", "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": "2604.10681", "title": "Critical-CoT: A Robust Defense Framework against Reasoning-Level Backdoor Attacks in Large Language Models", "abstract": "Large Language Models (LLMs), despite their impressive capabilities across domains, have been shown to be vulnerable to backdoor attacks. Prior backdoor strategies predominantly operate at the token level, where an injected trigger causes the model to generate a specific target word, choice, or class (depending on the task). Recent advances, however, exploit the long-form reasoning tendencies of modern LLMs to conduct reasoning-level backdoors: once triggered, the victim model inserts one or more malicious reasoning steps into its chain-of-thought (CoT). These attacks are substantially harder to detect, as the backdoored answer remains plausible and consistent with the poisoned reasoning trajectory. Yet, defenses tailored to this type of backdoor remain largely unexplored. To bridge this gap, we propose Critical-CoT, a novel defense mechanism that conducts a two-stage fine-tuning (FT) process on LLMs to develop critical thinking behaviors, enabling them to automatically identify potential backdoors and refuse to generate malicious reasoning steps. Extensive experiments across multiple LLMs and datasets demonstrate that Critical-CoT provides strong robustness against both in-context learning-based and FT-based backdoor attacks. Notably, Critical-CoT exhibits strong cross-domain and cross-task generalization. Our code is available at hthttps://github.com/tuanvu171/Critical-CoT.", "date": "2026-04-12", "categories": []} {"id": "2511.13789", "title": "Uncovering and Aligning Anomalous Attention Heads to Defend Against NLP Backdoor Attacks", "abstract": "Backdoor attacks pose a serious threat to the security of large language models (LLMs), causing them to exhibit anomalous behavior under specific trigger conditions. The design of backdoor triggers has evolved from fixed triggers to dynamic or implicit triggers. This increased flexibility in trigger design makes it challenging for defenders to identify their specific forms accurately. Most existing backdoor defense methods are limited to specific types of triggers or rely on an additional clean model for support. To address this issue, we propose a backdoor detection method based on attention similarity, enabling backdoor detection without prior knowledge of the trigger. Our study reveals that models subjected to backdoor attacks exhibit unusually high similarity among attention heads when exposed to triggers. Based on this observation, we propose an attention safety alignment approach combined with head-wise fine-tuning to rectify potentially contaminated attention heads, thereby effectively mitigating the impact of backdoor attacks. Extensive experimental results demonstrate that our method significantly reduces the success rate of backdoor attacks while preserving the model's performance on downstream tasks.", "date": "2025-11-16", "categories": []} {"id": "2604.12446", "title": "Scaling Exposes the Trigger: Input-Level Backdoor Detection in Text-to-Image Diffusion Models via Cross-Attention Scaling", "abstract": "Text-to-image (T2I) diffusion models have achieved remarkable success in image synthesis, but their reliance on large-scale data and open ecosystems introduces serious backdoor security risks. Existing defenses, particularly input-level methods, are more practical for deployment but often rely on observable anomalies that become unreliable under stealthy, semantics-preserving trigger designs. As modern backdoor attacks increasingly embed triggers into natural inputs, these methods degrade substantially, raising a critical question: can more stable, implicit, and trigger-agnostic differences between benign and backdoor inputs be exploited for detection? In this work, we address this challenge from an active probing perspective. We introduce controlled scaling perturbations on cross-attention and uncover a novel phenomenon termed Cross-Attention Scaling Response Divergence (CSRD), where benign and backdoor inputs exhibit systematically different response evolution patterns across denoising steps. Building on this insight, we propose SET, an input-level backdoor detection framework that constructs response-offset features under multi-scale perturbations and learns a compact benign response space from a small set of clean samples. Detection is then performed by measuring deviations from this learned space, without requiring prior knowledge of the attack or access to model training. Extensive experiments demonstrate that SET consistently outperforms existing baselines across diverse attack methods, trigger types, and model settings, with particularly strong gains under stealthy implicit-trigger scenarios. Overall, SET improves AUROC by 9.1% and ACC by 6.5% over the best baseline, highlighting its effectiveness and robustness for practical deployment.", "date": "2026-04-14", "categories": []} {"id": "2504.05902", "title": "Defending against Backdoor Attacks via Module Switching", "abstract": "Backdoor attacks pose a serious threat to deep neural networks (DNNs), allowing adversaries to implant triggers for hidden behaviors in inference. Defending against such vulnerabilities is especially difficult in the post-training setting, since end-users lack training data or prior knowledge of the attacks. Model merging offers a cost-effective defense; however, latest methods like weight averaging (WAG) provide reasonable protection when multiple homologous models are available, but are less effective with fewer models and place heavy demands on defenders. We propose a module-switching defense (MSD) for disrupting backdoor shortcuts. We first validate its theoretical rationale and empirical effectiveness on two-layer networks, showing its capability of achieving higher backdoor divergence than WAG, and preserving utility. For deep models, we evaluate MSD on Transformer and CNN architectures and design an evolutionary algorithm to optimize fusion strategies with selective mechanisms to identify the most effective combinations. Experiments show that MSD achieves stronger defense with fewer models in practical settings, and even under an underexplored case of collusive attacks among multiple models--where some models share the same backdoors--switching strategies by MSD deliver superior robustness against diverse attacks. Code is available at https://github.com/weijun-l/module-switching-defense.", "date": "2025-04-08", "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": "2604.09101", "title": "CLIP-Inspector: Model-Level Backdoor Detection for Prompt-Tuned CLIP via OOD Trigger Inversion", "abstract": "Organisations with limited data and computational resources increasingly outsource model training to Machine Learning as a Service (MLaaS) providers, who adapt vision-language models (VLMs) such as CLIP to downstream tasks via prompt tuning rather than training from scratch. This semi-honest setting creates a security risk where a malicious provider can follow the prompt-tuning protocol yet implant a backdoor, forcing triggered inputs to be classified into an attacker-chosen class, even for out-of-distribution (OOD) data. Such backdoors leave encoders untouched, making them undetectable to existing methods that focus on encoder corruption. Other data-level methods that sanitize data before training or during inference, also fail to answer the critical question, \"Is the delivered model backdoored or not?\" To address this model-level verification problem, we introduce CLIP-Inspector (CI), a backdoor detection method designed for prompt-tuned CLIP models. Assuming white-box access to the delivered model and a pool of unlabeled OOD images, CI reconstructs possible triggers for each class to determine if the model exhibits backdoor behaviour or not. Additionally, we demonstrate that using CI's reconstructed trigger for fine-tuning on correctly labeled triggered inputs enables us to re-align the model and reduce backdoor effectiveness. Through extensive experiments across ten datasets and four backdoor attacks, we demonstrate that CI can reconstruct effective triggers in a single epoch using only 1,000 OOD images, achieving a 94% detection accuracy (47/50 models). Compared to adapted trigger-inversion baselines, CI yields a markedly higher AUROC score (0.973 vs 0.495/0.687), thus enabling the vetting and post-hoc repair of prompt-tuned CLIP models to ensure safe deployment.", "date": "2026-04-10", "categories": []} {"id": "2604.09378", "title": "BadSkill: Backdoor Attacks on Agent Skills via Model-in-Skill Poisoning", "abstract": "Agent ecosystems increasingly rely on installable skills to extend functionality, and some skills bundle learned model artifacts as part of their execution logic. This creates a supply-chain risk that is not captured by prompt injection or ordinary plugin misuse: a third-party skill may appear benign while concealing malicious behavior inside its bundled model. We present BadSkill, a backdoor attack formulation that targets this model-in-skill threat surface. In BadSkill, an adversary publishes a seemingly benign skill whose embedded model is backdoor-fine-tuned to activate a hidden payload only when routine skill parameters satisfy attacker-chosen semantic trigger combinations. To realize this attack, we train the embedded classifier with a composite objective that combines classification loss, margin-based separation, and poison-focused optimization, and evaluate it in an OpenClaw-inspired simulation environment that preserves third-party skill installation and execution while enabling controlled multi-model study. Our benchmark spans 13 skills, including 8 triggered tasks and 5 non-trigger control skills, with a combined main evaluation set of 571 negative-class queries and 396 trigger-aligned queries. Across eight architectures (494M--7.1B parameters) from five model families, BadSkill achieves up to 99.5\\% average attack success rate (ASR) across the eight triggered skills while maintaining strong benign-side accuracy on negative-class queries. In poison-rate sweeps on the standard test split, a 3\\% poison rate already yields 91.7\\% ASR. The attack remains effective across the evaluated model scales and under five text perturbation types. These findings identify model-bearing skills as a distinct model supply-chain risk in agent ecosystems and motivate stronger provenance verification and behavioral vetting for third-party skill artifacts.", "date": "2026-04-10", "categories": []} {"id": "2412.00727", "title": "Perturb and Recover: Fine-tuning for Effective Backdoor Removal from CLIP", "abstract": "Vision-Language models like CLIP have been shown to be highly effective at linking visual perception and natural language understanding, enabling sophisticated image-text capabilities, including strong retrieval and zero-shot classification performance. Their widespread use, as well as the fact that CLIP models are trained on image-text pairs from the web, make them both a worthwhile and relatively easy target for backdoor attacks. As training foundational models, such as CLIP, from scratch is very expensive, this paper focuses on cleaning potentially poisoned models via fine-tuning. We first show that existing cleaning techniques are not effective against simple structured triggers used in Blended or BadNet backdoor attacks, exposing a critical vulnerability for potential real-world deployment of these models. Then, we introduce PAR, Perturb and Recover, a surprisingly simple yet effective mechanism to remove backdoors from CLIP models. Through extensive experiments across different encoders and types of backdoor attacks, we show that PAR achieves high backdoor removal rate while preserving good standard performance. Finally, we illustrate that our approach is effective even only with synthetic text-image pairs, i.e. without access to real training data. The code and models are available on \\href{https://github.com/nmndeep/PerturbAndRecover}{GitHub}.", "date": "2024-12-01", "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": "2603.29328", "title": "Beyond Corner Patches: Semantics-Aware Backdoor Attack in Federated Learning", "abstract": "Backdoor attacks on federated learning (FL) are most often evaluated with synthetic corner patches or out-of-distribution (OOD) patterns that are unlikely to arise in practice. In this paper, we revisit the backdoor threat to standard FL (a single global model) under a more realistic setting where triggers must be semantically meaningful, in-distribution, and visually plausible. We propose SABLE, a Semantics-Aware Backdoor for LEarning in federated settings, which constructs natural, content-consistent triggers (e.g., semantic attribute changes such as sunglasses) and optimizes an aggregation-aware malicious objective with feature separation and parameter regularization to keep attacker updates close to benign ones. We instantiate SABLE on CelebA hair-color classification and the German Traffic Sign Recognition Benchmark (GTSRB), poisoning only a small, interpretable subset of each malicious client's local data while otherwise following the standard FL protocol. Across heterogeneous client partitions and multiple aggregation rules (FedAvg, Trimmed Mean, MultiKrum, and FLAME), our semantics-driven triggers achieve high targeted attack success rates while preserving benign test accuracy. These results show that semantics-aligned backdoors remain a potent and practical threat in federated learning, and that robustness claims based solely on synthetic patch triggers can be overly optimistic.", "date": "2026-03-31", "categories": []} {"id": "2604.05432", "title": "Your LLM Agent Can Leak Your Data: Data Exfiltration via Backdoored Tool Use", "abstract": "Tool-use large language model (LLM) agents are increasingly deployed to support sensitive workflows, relying on tool calls for retrieval, external API access, and session memory management. While prior research has examined various threats, the risk of systematic data exfiltration by backdoored agents remains underexplored. In this work, we present Back-Reveal, a data exfiltration attack that embeds semantic triggers into fine-tuned LLM agents. When triggered, the backdoored agent invokes memory-access tool calls to retrieve stored user context and exfiltrates it via disguised retrieval tool calls. We further demonstrate that multi-turn interaction amplifies the impact of data exfiltration, as attacker-controlled retrieval responses can subtly steer subsequent agent behavior and user interactions, enabling sustained and cumulative information leakage over time. Our experimental results expose a critical vulnerability in LLM agents with tool access and highlight the need for defenses against exfiltration-oriented backdoors.", "date": "2026-04-07", "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": "2604.02372", "title": "Backdoor Attacks on Decentralised Post-Training", "abstract": "Decentralised post-training of large language models utilises data and pipeline parallelism techniques to split the data and the model. Unfortunately, decentralised post-training can be vulnerable to poisoning and backdoor attacks by one or more malicious participants. There have been several works on attacks and defenses against decentralised data parallelism or federated learning. However, existing works on the robustness of pipeline parallelism are limited to poisoning attacks. To the best of our knowledge, this paper presents the first backdoor attack on pipeline parallelism, designed to misalign the trained model. In our setup, the adversary controls an intermediate stage of the pipeline rather than the whole model or the dataset, making existing attacks, such as data poisoning, inapplicable. Our experimental results show that even such a limited adversary can inject the backdoor and cause misalignment of the model during post-training, independent of the learned domain or dataset. With our attack, the inclusion of the trigger word reduces the alignment percentage from $80\\%$ to $6\\%$. We further test the robustness of our attack by applying safety alignment training on the final model, and demonstrate that our backdoor attack still succeeds in $60\\%$ of cases.", "date": "2026-03-31", "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": "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": "2603.20339", "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": "2603.28092", "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": "2603.28673", "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": "2503.23866", "title": "A Channel-Triggered Backdoor Attack on Wireless Semantic Image Reconstruction", "abstract": "This paper investigates backdoor attacks in image-oriented semantic communications. The threat of backdoor attacks on symbol reconstruction in semantic communication (SemCom) systems has received limited attention. Previous research on backdoor attacks targeting SemCom symbol reconstruction primarily focuses on input-level triggers, which are impractical in scenarios with strict input constraints. In this paper, we propose a novel channel-triggered backdoor attack (CT-BA) framework that exploits inherent wireless channel characteristics as activation triggers. Our key innovation involves utilizing fundamental channel statistics parameters, specifically channel gain with different fading distributions or channel noise with different power, as potential triggers. This approach enhances stealth by eliminating explicit input manipulation, provides flexibility through trigger selection from diverse channel conditions, and enables automatic activation via natural channel variations without adversary intervention. We extensively evaluate CT-BA across four joint source-channel coding (JSCC) communication system architectures and three benchmark datasets. Simulation results demonstrate that our attack achieves near-perfect attack success rate (ASR) while maintaining effective stealth. Finally, we discuss potential defense mechanisms against such attacks.", "date": "2025-03-31", "categories": []} {"id": "2603.23007", "title": "AgentRAE: Remote Action Execution through Notification-based Visual Backdoors against Screenshots-based Mobile GUI Agents", "abstract": "The rapid adoption of mobile graphical user interface (GUI) agents, which autonomously control applications and operating systems (OS), exposes new system-level attack surfaces. Existing backdoors against web GUI agents and general GenAI models rely on environmental injection or deceptive pop-ups to mislead the agent operation. However, these techniques do not work on screenshots-based mobile GUI agents due to the challenges of restricted trigger design spaces, OS background interference, and conflicts in multiple trigger-action mappings. We propose AgentRAE, a novel backdoor attack capable of inducing Remote Action Execution in mobile GUI agents using visually natural triggers (e.g., benign app icons in notifications). To address the underfitting caused by natural triggers and achieve accurate multi-target action redirection, we design a novel two-stage pipeline that first enhances the agent's sensitivity to subtle iconographic differences via contrastive learning, and then associates each trigger with a specific mobile GUI agent action through a backdoor post-training. Our extensive evaluation reveals that the proposed backdoor preserves clean performance with an attack success rate of over 90% across ten mobile operations. Furthermore, it is hard to visibly detect the benign-looking triggers and circumvents eight representative state-of-the-art defenses. These results expose an overlooked backdoor vector in mobile GUI agents, underscoring the need for defenses that scrutinize notification-conditioned behaviors and internal agent representations.", "date": "2026-03-24", "categories": []} {"id": "2505.18323", "title": "Architectural Backdoors for Within-Batch Data Stealing and Model Inference Manipulation", "abstract": "For nearly a decade the academic community has investigated backdoors in neural networks, primarily focusing on classification tasks where adversaries manipulate the model prediction. While demonstrably malicious, the immediate real-world impact of such prediction-altering attacks has remained unclear. In this paper we introduce a novel and significantly more potent class of backdoors that builds upon recent advancements in architectural backdoors. We demonstrate how these backdoors can be specifically engineered to exploit batched inference, a common technique for hardware utilization, enabling large-scale user data manipulation and theft. By targeting the batching process, these architectural backdoors facilitate information leakage between concurrent user requests and allow attackers to fully control model responses directed at other users within the same batch. In other words, an attacker who can change the model architecture can set and steal model inputs and outputs of other users within the same batch. We show that such attacks are not only feasible but also alarmingly effective, can be readily injected into prevalent model architectures, (e.g. Transformers), and represent a truly malicious threat to user privacy and system integrity. Critically, to counteract this new class of vulnerabilities, we propose a deterministic mitigation strategy that provides formal guarantees against this new attack vector, unlike prior work that relied on LLMs to find the backdoors. Our mitigation strategy employs a novel Information Flow Control mechanism that analyzes the model graph and proves non-interference between different user inputs within the same batch. Using our mitigation strategy we perform a large scale analysis of models hosted through Hugging Face and find over 200 models that introduce (unintended) information leakage between batch entries due to the use of dynamic quantization.", "date": "2025-05-23", "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": "2603.18103", "title": "STEP: Detecting Audio Backdoor Attacks via Stability-based Trigger Exposure Profiling", "abstract": "With the widespread deployment of deep-learning-based speech models in security-critical applications, backdoor attacks have emerged as a serious threat: an adversary who poisons a small fraction of training data can implant a hidden trigger that controls the model's output while preserving normal behavior on clean inputs. Existing inference-time defenses are not well suited to the audio domain, as they either rely on trigger over-robustness assumptions that fail on transformation-based and semantic triggers, or depend on properties specific to image or text modalities. In this paper, we propose STEP (Stability-based Trigger Exposure Profiling), a black-box, retraining-free backdoor detector that operates under hard-label-only access. Its core idea is to exploit a characteristic dual anomaly of backdoor triggers: anomalous label stability under semantic-breaking perturbations, and anomalous label fragility under semantic-preserving perturbations. STEP profiles each test sample with two complementary perturbation branches that target these two properties respectively, scores the resulting stability features with one-class anomaly detectors trained on benign references, and fuses the two scores via unsupervised weighting. Extensive experiments across seven backdoor attacks show that STEP achieves an average AUROC of 97.92% and EER of 4.54%, substantially outperforming state-of-the-art baselines, and generalizes across model architectures, speech tasks, an open-set verification scenario, and over-the-air physical-world settings.", "date": "2026-03-18", "categories": []} {"id": "2603.15692", "title": "BadLLM-TG: A Backdoor Defender powered by LLM Trigger Generator", "abstract": "Backdoor attacks compromise model reliability by using triggers to manipulate outputs. Trigger inversion can accurately locate these triggers via a generator and is therefore critical for backdoor defense. However, the discrete nature of text prevents existing noise-based trigger generator from being applied to nature language processing (NLP). To overcome the limitations, we employ the rich knowledge embedded in large language models (LLMs) and propose a Backdoor defender powered by LLM Trigger Generator, termed BadLLM-TG. It is optimized through prompt-driven reinforcement learning, using the victim model's feedback loss as the reward signal. The generated triggers are then employed to mitigate the backdoor via adversarial training. Experiments show that our method reduces the attack success rate by 76.2\\% on average, outperforming the second-best defender by 13.7.", "date": "2026-03-16", "categories": []} {"id": "2603.11664", "title": "BackdoorIDS: Zero-shot Backdoor Detection for Pretrained Vision Encoder", "abstract": "Self-supervised and multimodal vision encoders learn strong visual representations that are widely adopted in downstream vision tasks and large vision-language models (LVLMs). However, downstream users often rely on third-party pretrained encoders with uncertain provenance, exposing them to backdoor attacks. In this work, we propose BackdoorIDS, a simple yet effective zero-shot, inference-time backdoor samples detection method for pretrained vision encoders. BackdoorIDS is motivated by two observations: Attention Hijacking and Restoration. Under progressive input masking, a backdoored image initially concentrates attention on malicious trigger features. Once the masking ratio exceeds the trigger's robustness threshold, the trigger is deactivated, and attention rapidly shifts to benign content. This transition induces a pronounced change in the image embedding, whereas embeddings of clean images evolve more smoothly across masking progress. BackdoorIDS operationalizes this signal by extracting an embedding sequence along the masking trajectory and applying density-based clustering such as DBSCAN. An input is flagged as backdoored if its embedding sequence forms more than one cluster. Extensive experiments show that BackdoorIDS consistently outperforms existing defenses across diverse attack types, datasets, and model families. Notably, it is a plug-and-play approach that requires no retraining and operates fully zero-shot at inference time, making it compatible with a wide range of encoder architectures, including CNNs, ViTs, CLIP, and LLaVA-1.5.", "date": "2026-03-12", "categories": []} {"id": "2603.13461", "title": "Purifying Generative LLMs from Backdoors without Prior Knowledge or Clean Reference", "abstract": "Backdoor attacks pose severe security threats to large language models (LLMs), where a model behaves normally under benign inputs but produces malicious outputs when a hidden trigger appears. Existing backdoor removal methods typically assume prior knowledge of triggers, access to a clean reference model, or rely on aggressive finetuning configurations, and are often limited to classification tasks. However, such assumptions fall apart in real-world instruction-tuned LLM settings. In this work, we propose a new framework for purifying instruction-tuned LLM without any prior trigger knowledge or clean references. Through systematic sanity checks, we find that backdoor associations are redundantly encoded across MLP layers, while attention modules primarily amplify trigger signals without establishing the behavior. Leveraging this insight, we shift the focus from isolating specific backdoor triggers to cutting off the trigger-behavior associations, and design an immunization-inspired elimination approach: by constructing multiple synthetic backdoored variants of the given suspicious model, each trained with different malicious trigger-behavior pairs, and contrasting them with their clean counterparts. The recurring modifications across variants reveal a shared \"backdoor signature\"-analogous to antigens in a virus. Guided by this signature, we neutralize highly suspicious components in LLM and apply lightweight finetuning to restore its fluency, producing purified models that withstand diverse backdoor attacks and threat models while preserving generative capability.", "date": "2026-03-13", "categories": []} {"id": "2603.11949", "title": "Delayed Backdoor Attacks: Exploring the Temporal Dimension as a New Attack Surface in Pre-Trained Models", "abstract": "Backdoor attacks against pre-trained models (PTMs) have traditionally operated under an ``immediacy assumption,'' where malicious behavior manifests instantly upon trigger occurrence. This work revisits and challenges this paradigm by introducing \\textit{\\textbf{Delayed Backdoor Attacks (DBA)}}, a new class of threats in which activation is temporally decoupled from trigger exposure. We propose that this \\textbf{temporal dimension} is the key to unlocking a previously infeasible class of attacks: those that use common, everyday words as triggers. To examine the feasibility of this paradigm, we design and implement a proof-of-concept prototype, termed \\underline{D}elayed Backdoor Attacks Based on \\underline{N}onlinear \\underline{D}ecay (DND). DND embeds a lightweight, stateful logic module that postpones activation until a configurable threshold is reached, producing a distinct latency phase followed by a controlled outbreak. We derive a formal model to characterize this latency behavior and propose a dual-metric evaluation framework (ASR and ASR$_{delay}$) to empirically measure the delay effect. Extensive experiments on four (natural language processing)NLP benchmarks validate the core capabilities of DND: it remains dormant for a controllable duration, sustains high clean accuracy ($\\ge$94\\%), and achieves near-perfect post-activation attack success rates ($\\approx$99\\%, The average of other methods is below 95\\%.). Moreover, DND exhibits resilience against several state-of-the-art defenses. This study provides the first empirical evidence that the temporal dimension constitutes a viable yet unprotected attack surface in PTMs, underscoring the need for next-generation, stateful, and time-aware defense mechanisms.", "date": "2026-03-12", "categories": []} {"id": "2603.10641", "title": "Detecting and Eliminating Neural Network Backdoors Through Active Paths with Application to Intrusion Detection", "abstract": "Machine learning backdoors have the property that the machine learning model should work as expected on normal inputs, but when the input contains a specific $\\textit{trigger}$, it behaves as the attacker desires. Detecting such triggers has been proven to be extremely difficult. In this paper, we present a novel and explainable approach to detect and eliminate such backdoor triggers based on active paths found in neural networks. We present promising experimental evidence of our approach, which involves injecting backdoors into a machine learning model used for intrusion detection.", "date": "2026-03-11", "categories": []} {"id": "2603.08316", "title": "SlowBA: An efficiency backdoor attack towards VLM-based GUI agents", "abstract": "Modern vision-language-model (VLM) based graphical user interface (GUI) agents are expected not only to execute actions accurately but also to respond to user instructions with low latency. While existing research on GUI-agent security mainly focuses on manipulating action correctness, the security risks related to response efficiency remain largely unexplored. In this paper, we introduce SlowBA, a novel backdoor attack that targets the responsiveness of VLM-based GUI agents. The key idea is to manipulate response latency by inducing excessively long reasoning chains under specific trigger patterns. To achieve this, we propose a two-stage reward-level backdoor injection (RBI) strategy that first aligns the long-response format and then learns trigger-aware activation through reinforcement learning. In addition, we design realistic pop-up windows as triggers that naturally appear in GUI environments, improving the stealthiness of the attack. Extensive experiments across multiple datasets and baselines demonstrate that SlowBA can significantly increase response length and latency while largely preserving task accuracy. The attack remains effective even with a small poisoning ratio and under several defense settings. These findings reveal a previously overlooked security vulnerability in GUI agents and highlight the need for defenses that consider both action correctness and response efficiency. Code can be found in https://github.com/tu-tuing/SlowBA.", "date": "2026-03-09", "categories": []} {"id": "2603.09772", "title": "Removing the Trigger, Not the Backdoor: Alternative Triggers and Latent Backdoors", "abstract": "Current backdoor defenses assume that neutralizing a known trigger removes the backdoor. We show this trigger-centric view is incomplete: \\emph{alternative triggers}, patterns perceptually distinct from training triggers, reliably activate the same backdoor. We estimate the alternative trigger backdoor direction in feature space by contrasting clean and triggered representations, and then develop a feature-guided attack that jointly optimizes target prediction and directional alignment. First, we theoretically prove that alternative triggers exist and are an inevitable consequence of backdoor training. Then, we verify this empirically. Additionally, defenses that remove training triggers often leave backdoors intact, and alternative triggers can exploit the latent backdoor feature-space. Our findings motivate defenses targeting backdoor directions in representation space rather than input-space triggers.", "date": "2026-03-10", "categories": []} {"id": "2504.21052", "title": "SFIBA: Spatial-based Full-target Invisible Backdoor Attacks", "abstract": "Multi-target backdoor attacks pose significant security threats to deep neural networks, as they can preset multiple target classes through a single backdoor injection. This allows attackers to control the model to misclassify poisoned samples with triggers into any desired target class during inference, exhibiting superior attack performance compared with conventional backdoor attacks. However, existing multi-target backdoor attacks fail to guarantee trigger specificity and stealthiness in black-box settings, resulting in two main issues. First, they are unable to simultaneously target all classes when only training data can be manipulated, limiting their effectiveness in realistic attack scenarios. Second, the triggers often lack visual imperceptibility, making poisoned samples easy to detect. To address these problems, we propose a Spatial-based Full-target Invisible Backdoor Attack, called SFIBA. It restricts triggers for different classes to specific local spatial regions and morphologies in the pixel space to ensure specificity, while employing a frequency-domain-based trigger injection method to guarantee stealthiness. Specifically, for injection of each trigger, we first apply fast fourier transform to obtain the amplitude spectrum of clean samples in local spatial regions. Then, we employ discrete wavelet transform to extract the features from the amplitude spectrum and use singular value decomposition to integrate the trigger. Subsequently, we selectively filter parts of the trigger in pixel space to implement trigger morphology constraints and adjust injection coefficients based on visual effects. We conduct experiments on multiple datasets and models. The results demonstrate that SFIBA can achieve excellent attack performance and stealthiness, while preserving the model's performance on benign samples, and can also bypass existing backdoor defenses.", "date": "2025-04-29", "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": "2603.05921", "title": "BlackMirror: Black-Box Backdoor Detection for Text-to-Image Models via Instruction-Response Deviation", "abstract": "This paper investigates the challenging task of detecting backdoored text-to-image models under black-box settings and introduces a novel detection framework BlackMirror. Existing approaches typically rely on analyzing image-level similarity, under the assumption that backdoor-triggered generations exhibit strong consistency across samples. However, they struggle to generalize to recently emerging backdoor attacks, where backdoored generations can appear visually diverse. BlackMirror is motivated by an observation: across backdoor attacks, {only partial semantic patterns within the generated image are steadily manipulated, while the rest of the content remains diverse or benign. Accordingly, BlackMirror consists of two components: MirrorMatch, which aligns visual patterns with the corresponding instructions to detect semantic deviations; and MirrorVerify, which evaluates the stability of these deviations across varied prompts to distinguish true backdoor behavior from benign responses. BlackMirror is a general, training-free framework that can be deployed as a plug-and-play module in Model-as-a-Service (MaaS) applications. Comprehensive experiments demonstrate that BlackMirror achieves accurate detection across a wide range of attacks. Code is available at https://github.com/Ferry-Li/BlackMirror.", "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": "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": "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": "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": "2602.22246", "title": "Self-Purification Mitigates Backdoors in Multimodal Diffusion Language Models", "abstract": "Multimodal Diffusion Language Models (MDLMs) have recently emerged as a competitive alternative to their autoregressive counterparts. Yet their vulnerability to backdoor attacks remains largely unexplored. In this work, we show that well-established data-poisoning pipelines can successfully implant backdoors into MDLMs, enabling attackers to manipulate model behavior via specific triggers while maintaining normal performance on clean inputs. However, defense strategies effective to these models are yet to emerge. To bridge this gap, we introduce a backdoor defense framework for MDLMs named DiSP (Diffusion Self-Purification). DiSP is driven by a key observation: selectively masking certain vision tokens at inference time can neutralize a backdoored model's trigger-induced behaviors and restore normal functionality. Building on this, we purify the poisoned dataset using the compromised model itself, then fine-tune the model on the purified data to recover it to a clean one. Given such a specific design, DiSP can remove backdoors without requiring any auxiliary models or clean reference data. Extensive experiments demonstrate that our approach effectively mitigates backdoor effects, reducing the attack success rate (ASR) from over 90% to typically under 5%, while maintaining model performance on benign tasks.", "date": "2026-02-24", "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": "2602.17168", "title": "BadCLIP++: Stealthy and Persistent Backdoors in Multimodal Contrastive Learning", "abstract": "Research on backdoor attacks against multimodal contrastive learning models faces two key challenges: stealthiness and persistence. Existing methods often fail under strong detection or continuous fine-tuning, largely due to (1) cross-modal inconsistency that exposes trigger patterns and (2) gradient dilution at low poisoning rates that accelerates backdoor forgetting. These coupled causes remain insufficiently modeled and addressed. We propose BadCLIP++, a unified framework that tackles both challenges. For stealthiness, we introduce a semantic-fusion QR micro-trigger that embeds imperceptible patterns near task-relevant regions, preserving clean-data statistics while producing compact trigger distributions. We further apply target-aligned subset selection to strengthen signals at low injection rates. For persistence, we stabilize trigger embeddings via radius shrinkage and centroid alignment, and stabilize model parameters through curvature control and elastic weight consolidation, maintaining solutions within a low-curvature wide basin resistant to fine-tuning. We also provide the first theoretical analysis showing that, within a trust region, gradients from clean fine-tuning and backdoor objectives are co-directional, yielding a non-increasing upper bound on attack success degradation. Experiments demonstrate that with only 0.3% poisoning, BadCLIP++ achieves 99.99% attack success rate (ASR) in digital settings, surpassing baselines by 11.4 points. Across nineteen defenses, ASR remains above 99.90% with less than 0.8% drop in clean accuracy. The method further attains 65.03% success in physical attacks and shows robustness against watermark removal defenses.", "date": "2026-02-19", "categories": []} {"id": "2602.13427", "title": "Backdooring Bias in Large Language Models", "abstract": "Large language models (LLMs) are increasingly deployed in settings where inducing a bias toward a certain topic can have significant consequences, and backdoor attacks can be used to produce such models. Prior work on backdoor attacks has largely focused on a black-box threat model, with an adversary targeting the model builder's LLM. However, in the bias manipulation setting, the model builder themselves could be the adversary, warranting a white-box threat model where the attacker's ability to poison, and manipulate the poisoned data is substantially increased. Furthermore, despite growing research in semantically-triggered backdoors, most studies have limited themselves to syntactically-triggered attacks. Motivated by these limitations, we conduct an analysis consisting of over 1000 evaluations using higher poisoning ratios and greater data augmentation to gain a better understanding of the potential of syntactically- and semantically-triggered backdoor attacks in a white-box setting. In addition, we study whether two representative defense paradigms, model-intrinsic and model-extrinsic backdoor removal, are able to mitigate these attacks. Our analysis reveals numerous new findings. We discover that while both syntactically- and semantically-triggered attacks can effectively induce the target behaviour, and largely preserve utility, semantically-triggered attacks are generally more effective in inducing negative biases, while both backdoor types struggle with causing positive biases. Furthermore, while both defense types are able to mitigate these backdoors, they either result in a substantial drop in utility, or require high computational overhead.", "date": "2026-02-13", "categories": []} {"id": "2412.03441", "title": "PBP: Post-training Backdoor Purification for Malware Classifiers", "abstract": "In recent years, the rise of machine learning (ML) in cybersecurity has brought new challenges, including the increasing threat of backdoor poisoning attacks on ML malware classifiers. For instance, adversaries could inject malicious samples into public malware repositories, contaminating the training data and potentially misclassifying malware by the ML model. Current countermeasures predominantly focus on detecting poisoned samples by leveraging disagreements within the outputs of a diverse set of ensemble models on training data points. However, these methods are not suitable for scenarios where Machine Learning-as-a-Service (MLaaS) is used or when users aim to remove backdoors from a model after it has been trained. Addressing this scenario, we introduce PBP, a post-training defense for malware classifiers that mitigates various types of backdoor embeddings without assuming any specific backdoor embedding mechanism. Our method exploits the influence of backdoor attacks on the activation distribution of neural networks, independent of the trigger-embedding method. In the presence of a backdoor attack, the activation distribution of each layer is distorted into a mixture of distributions. By regulating the statistics of the batch normalization layers, we can guide a backdoored model to perform similarly to a clean one. Our method demonstrates substantial advantages over several state-of-the-art methods, as evidenced by experiments on two datasets, two types of backdoor methods, and various attack configurations. Notably, our approach requires only a small portion of the training data -- only 1\\% -- to purify the backdoor and reduce the attack success rate from 100\\% to almost 0\\%, a 100-fold improvement over the baseline methods. Our code is available at https://github.com/judydnguyen/pbp-backdoor-purification-official.", "date": "2024-12-04", "categories": []} {"id": "2505.10297", "title": "Defending the Edge: Representative-Attention Defense against Backdoor Attacks in Federated Learning", "abstract": "Federated learning (FL) remains highly vulnerable to adaptive backdoor attacks that preserve stealth by closely imitating benign update statistics. Existing defenses predominantly rely on anomaly detection in parameter or gradient space, overlooking behavioral constraints that backdoor attacks must satisfy to ensure reliable trigger activation. These anomaly-centric methods fail against adaptive attacks that normalize update magnitudes and mimic benign statistical patterns while preserving backdoor functionality, creating a fundamental detection gap. To address this limitation, this paper introduces FeRA (Federated Representative Attention) -- a novel attention-driven defense that shifts the detection paradigm from anomaly-centric to consistency-centric analysis. FeRA exploits the intrinsic need for backdoor persistence across training rounds, identifying malicious clients through suppressed representation-space variance, an orthogonal property to traditional magnitude-based statistics. The framework conducts multi-dimensional behavioral analysis combining spectral and spatial attention, directional alignment, mutual similarity, and norm inflation across two complementary detection mechanisms: consistency analysis and norm-inflation detection. Through this mechanism, FeRA isolates malicious clients that exhibit low-variance consistency or magnitude amplification. Extensive evaluation across six datasets, nine attacks, and three model architectures under both Independent and Identically Distributed (IID) and non-IID settings confirm FeRA achieves superior backdoor mitigation. Under different non-IID settings, FeRA achieved the lowest average Backdoor Accuracy (BA), about 1.67% while maintaining high clean accuracy compared to other state-of-the-art defenses. The code is available at https://github.com/Peatech/FeRA_defense.git.", "date": "2025-05-15", "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": "2602.08563", "title": "Stateless Yet Not Forgetful: Implicit Memory as a Hidden Channel in LLMs", "abstract": "Large language models (LLMs) are commonly treated as stateless: once an interaction ends, no information is assumed to persist unless it is explicitly stored and re-supplied. We challenge this assumption by introducing implicit memory-the ability of a model to carry state across otherwise independent interactions by encoding information in its own outputs and later recovering it when those outputs are reintroduced as input. This mechanism does not require any explicit memory module, yet it creates a persistent information channel across inference requests. As a concrete demonstration, we introduce a new class of temporal backdoors, which we call time bombs. Unlike conventional backdoors that activate on a single trigger input, time bombs activate only after a sequence of interactions satisfies hidden conditions accumulated via implicit memory. We show that such behavior can be induced today through straightforward prompting or fine-tuning. Beyond this case study, we analyze broader implications of implicit memory, including covert inter-agent communication, benchmark contamination, targeted manipulation, and training-data poisoning. Finally, we discuss detection challenges and outline directions for stress-testing and evaluation, with the goal of anticipating and controlling future developments. To promote future research, we release code and data at: https://github.com/microsoft/implicitMemory.", "date": "2026-02-09", "categories": []} {"id": "2602.03085", "title": "The Trigger in the Haystack: Extracting and Reconstructing LLM Backdoor Triggers", "abstract": "Detecting whether a model has been poisoned is a longstanding problem in AI security. In this work, we present a practical scanner for identifying sleeper agent-style backdoors in causal language models. Our approach relies on two key findings: first, sleeper agents tend to memorize poisoning data, making it possible to leak backdoor examples using memory extraction techniques. Second, poisoned LLMs exhibit distinctive patterns in their output distributions and attention heads when backdoor triggers are present in the input. Guided by these observations, we develop a scalable backdoor scanning methodology that assumes no prior knowledge of the trigger or target behavior and requires only inference operations. Our scanner integrates naturally into broader defensive strategies and does not alter model performance. We show that our method recovers working triggers across multiple backdoor scenarios and a broad range of models and fine-tuning methods.", "date": "2026-02-03", "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": "2602.00183", "title": "RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance", "abstract": "Deep neural networks are highly susceptible to backdoor attacks, yet most defense methods to date rely on balanced data, overlooking the pervasive class imbalance in real-world scenarios that can amplify backdoor threats. This paper presents the first in-depth investigation of how the dataset imbalance amplifies backdoor vulnerability, showing that (i) the imbalance induces a majority-class bias that increases susceptibility and (ii) conventional defenses degrade significantly as the imbalance grows. To address this, we propose Randomized Probability Perturbation (RPP), a certified poisoned-sample detection framework that operates in a black-box setting using only model output probabilities. For any inspected sample, RPP determines whether the input has been backdoor-manipulated, while offering provable within-domain detectability guarantees and a probabilistic upper bound on the false positive rate. Extensive experiments on five benchmarks (MNIST, SVHN, CIFAR-10, TinyImageNet and ImageNet10) covering 10 backdoor attacks and 12 baseline defenses show that RPP achieves significantly higher detection accuracy than state-of-the-art defenses, particularly under dataset imbalance. RPP establishes a theoretical and practical foundation for defending against backdoor attacks in real-world environments with imbalanced data.", "date": "2026-01-30", "categories": []} {"id": "2403.03846", "title": "Can Distillation Mitigate Backdoor Attacks in Pre-trained Encoders?", "abstract": "Self-Supervised Learning (SSL) has become a prominent paradigm for pre-training encoders to learning general-purpose representations from unlabeled data and releasing them on third-party platforms for broad downstream deep learning tasks. However, SSL is vulnerable to backdoor attacks, where an adversary may train and distribute poisoned pre-training encoders to contaminate the downstream models. In this paper, we study a defense mechanism based on distillation against poisoned encoders in SSL. Traditionally, distillation transfers knowledge from a pre-trained teacher model to a student model, enabling the student to replicate or refine the teacher's learned representations. We repurpose distillation to extract benign knowledge and remove backdoors from a poisoned pre-trained encoder to produce a clean and reliable pre-trained model. We conduct extensive experiments to evaluate the effectiveness of distillation in mitigating backdoor attacks on pre-trained encoders. Based on two state-of-the-art backdoor attacks and four widely adopted image classification datasets, our results demonstrate that distillation reduces the attack success rate from 80.87% to 27.51%, with only a 6.35% drop in model accuracy. Furthermore, by comparing four teacher architectures, three student models, and six loss functions, we find that the distillation with fine-tuned teacher networks, warm-up-based student training, and attention-based distillation losses yield the best performance.", "date": "2024-03-06", "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": "2506.22506", "title": "SABRE-FL: Selective and Accurate Backdoor Rejection for Federated Prompt Learning", "abstract": "Federated Prompt Learning has emerged as a communication-efficient and privacy-preserving paradigm for adapting large vision-language models like CLIP across decentralized clients. However, the security implications of this setup remain underexplored. In this work, we present the first study of backdoor attacks in Federated Prompt Learning. We show that when malicious clients inject visually imperceptible, learnable noise triggers into input images, the global prompt learner becomes vulnerable to targeted misclassification while still maintaining high accuracy on clean inputs. Motivated by this vulnerability, we propose SABRE-FL, a lightweight, modular defense that filters poisoned prompt updates using an embedding-space anomaly detector trained offline on out-of-distribution data. SABRE-FL requires no access to raw client data or labels and generalizes across diverse datasets. We show, both theoretically and empirically, that malicious clients can be reliably identified and filtered using an embedding-based detector. Across five diverse datasets and four baseline defenses, SABRE-FL outperforms all baselines by significantly reducing backdoor accuracy while preserving clean accuracy, demonstrating strong empirical performance and underscoring the need for robust prompt learning in future federated systems.", "date": "2025-06-25", "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": "2601.11006", "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": "2601.08511", "title": "STAR: Detecting Inference-time Backdoors in LLM Reasoning via State-Transition Amplification Ratio", "abstract": "Recent LLMs increasingly integrate reasoning mechanisms like Chain-of-Thought (CoT). However, this explicit reasoning exposes a new attack surface for inference-time backdoors, which inject malicious reasoning paths without altering model parameters. Because these attacks generate linguistically coherent paths, they effectively evade conventional detection. To address this, we propose STAR (State-Transition Amplification Ratio), a framework that detects backdoors by analyzing output probability shifts. STAR exploits the statistical discrepancy where a malicious input-induced path exhibits high posterior probability despite a low prior probability in the model's general knowledge. We quantify this state-transition amplification and employ the CUSUM algorithm to detect persistent anomalies. Experiments across diverse models (8B-70B) and five benchmark datasets demonstrate that STAR exhibits robust generalization capabilities, consistently achieving near-perfect performance (AUROC $\\approx$ 1.0) with approximately $42\\times$ greater efficiency than existing baselines. Furthermore, the framework proves robust against adaptive attacks attempting to bypass detection.", "date": "2026-01-13", "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": "2601.06305", "title": "Why LoRA Fails to Forget: Regularized Low-Rank Adaptation Against Backdoors in Language Models", "abstract": "Low-Rank Adaptation (LoRA) is widely used for parameter-efficient fine-tuning of large language models, but it is notably ineffective at removing backdoor behaviors from poisoned pretrained models when fine-tuning on clean dataset. Contrary to the common belief that this weakness is caused primarily by low rank, we show that LoRA's vulnerability is fundamentally spectral. Our analysis identifies two key factors: LoRA updates (i) possess insufficient spectral strength, with singular values far below those of pretrained weights, and (ii) exhibit unfavorable spectral alignment, weakly matching clean-task directions while retaining overlap with trigger-sensitive subspaces. We further establish a critical scaling threshold beyond which LoRA can theoretically suppress trigger-induced activations, and we show empirically that standard LoRA rarely reaches this regime. We introduce Regularized Low-Rank Adaptation (RoRA), which improves forgetting by increasing spectral strength and correcting alignment through clean-strengthened regularization, trigger-insensitive constraints, and post-training spectral rescaling. Experiments across multiple NLP benchmarks and attack settings show that RoRA substantially reduces attack success rates while maintaining clean accuracy.", "date": "2026-01-09", "categories": []} {"id": "2601.04247", "title": "Beyond Immediate Activation: Temporally Decoupled Backdoor Attacks on Time Series Forecasting", "abstract": "Existing backdoor attacks on multivariate time series (MTS) forecasting enforce strict temporal and dimensional coupling between triggers and target patterns, requiring synchronous activation at fixed positions across variables. However, realistic scenarios often demand delayed and variable-specific activation. We identify this critical unmet need and propose TDBA, a temporally decoupled backdoor attack framework for MTS forecasting. By injecting triggers that encode the expected location of the target pattern, TDBA enables the activation of the target pattern at any positions within the forecasted data, with the activation position flexibly varying across different variable dimensions. TDBA introduces two core modules: (1) a position-guided trigger generation mechanism that leverages smoothed Gaussian priors to generate triggers that are position-related to the predefined target pattern; and (2) a position-aware optimization module that assigns soft weights based on trigger completeness, pattern coverage, and temporal offset, facilitating targeted and stealthy attack optimization. Extensive experiments on real-world datasets show that TDBA consistently outperforms existing baselines in effectiveness while maintaining good stealthiness. Ablation studies confirm the controllability and robustness of its design.", "date": "2026-01-06", "categories": []} {"id": "2601.03805", "title": "Detecting Semantic Backdoors in a Mystery Shopping Scenario", "abstract": "Detecting semantic backdoors in classification models--where some classes can be activated by certain natural, but out-of-distribution inputs--is an important problem that has received relatively little attention. Semantic backdoors are significantly harder to detect than backdoors that are based on trigger patterns due to the lack of such clearly identifiable patterns. We tackle this problem under the assumption that the clean training dataset and the training recipe of the model are both known. These assumptions are motivated by a consumer protection scenario, in which the responsible authority performs mystery shopping to test a machine learning service provider. In this scenario, the authority uses the provider's resources and tools to train a model on a given dataset and tests whether the provider included a backdoor. In our proposed approach, the authority creates a reference model pool by training a small number of clean and poisoned models using trusted infrastructure, and calibrates a model distance threshold to identify clean models. We propose and experimentally analyze a number of approaches to compute model distances and we also test a scenario where the provider performs an adaptive attack to avoid detection. The most reliable method is based on requesting adversarial training from the provider. The model distance is best measured using a set of input samples generated by inverting the models in such a way as to maximize the distance from clean samples. With these settings, our method can often completely separate clean and poisoned models, and it proves to be superior to state-of-the-art backdoor detectors as well.", "date": "2026-01-07", "categories": []} {"id": "2511.14301", "title": "SteganoBackdoor: Stealthy and Data-Efficient Backdoor Attacks on Language Models", "abstract": "Modern language models remain vulnerable to backdoor attacks via poisoned data, where training inputs containing a trigger are paired with a target output, causing the model to reproduce that behavior whenever the trigger appears at inference time. Recent work has emphasized stealthy attacks that stress-test data-curation defenses using stylized artifacts or token-level perturbations as triggers, but this focus leaves a more practically relevant threat model underexplored: backdoors tied to naturally occurring semantic concepts. We introduce SteganoBackdoor, an optimization-based framework that constructs SteganoPoisons, steganographic poisoned training examples in which a backdoor payload is distributed across a fluent sentence while exhibiting no representational overlap with the inference-time semantic trigger. Across diverse model architectures, SteganoBackdoor achieves high attack success under constrained poisoning budgets and remains effective under conservative data-level filtering, highlighting a blind spot in existing data-curation defenses.", "date": "2025-11-18", "categories": []} {"id": "2512.19730", "title": "ArcGen: Generalizing Neural Backdoor Detection Across Diverse Architectures", "abstract": "Backdoor attacks pose a significant threat to the security and reliability of deep learning models. To mitigate such attacks, one promising approach is to learn to extract features from the target model and use these features for backdoor detection. However, we discover that existing learning-based neural backdoor detection methods do not generalize well to new architectures not seen during the learning phase. In this paper, we analyze the root cause of this issue and propose a novel black-box neural backdoor detection method called ArcGen. Our method aims to obtain architecture-invariant model features, i.e., aligned features, for effective backdoor detection. Specifically, in contrast to existing methods directly using model outputs as model features, we introduce an additional alignment layer in the feature extraction function to further process these features. This reduces the direct influence of architecture information on the features. Then, we design two alignment losses to train the feature extraction function. These losses explicitly require that features from models with similar backdoor behaviors but different architectures are aligned at both the distribution and sample levels. With these techniques, our method demonstrates up to 42.5% improvements in detection performance (e.g., AUC) on unseen model architectures. This is based on a large-scale evaluation involving 16,896 models trained on diverse datasets, subjected to various backdoor attacks, and utilizing different model architectures. Our code is available at https://github.com/SeRAlab/ArcGen.", "date": "2025-12-17", "categories": []} {"id": "2512.19297", "title": "Causal-Guided Detoxify Backdoor Attack of Open-Weight LoRA Models", "abstract": "Low-Rank Adaptation (LoRA) has emerged as an efficient method for fine-tuning large language models (LLMs) and is widely adopted within the open-source community. However, the decentralized dissemination of LoRA adapters through platforms such as Hugging Face introduces novel security vulnerabilities: malicious adapters can be easily distributed and evade conventional oversight mechanisms. Despite these risks, backdoor attacks targeting LoRA-based fine-tuning remain relatively underexplored. Existing backdoor attack strategies are ill-suited to this setting, as they often rely on inaccessible training data, fail to account for the structural properties unique to LoRA, or suffer from high false trigger rates (FTR), thereby compromising their stealth. To address these challenges, we propose Causal-Guided Detoxify Backdoor Attack (CBA), a novel backdoor attack framework specifically designed for open-weight LoRA models. CBA operates without access to original training data and achieves high stealth through two key innovations: (1) a coverage-guided data generation pipeline that synthesizes task-aligned inputs via behavioral exploration, and (2) a causal-guided detoxification strategy that merges poisoned and clean adapters by preserving task-critical neurons. Unlike prior approaches, CBA enables post-training control over attack intensity through causal influence-based weight allocation, eliminating the need for repeated retraining. Evaluated across six LoRA models, CBA achieves high attack success rates while reducing FTR by 50-70\\% compared to baseline methods. Furthermore, it demonstrates enhanced resistance to state-of-the-art backdoor defenses, highlighting its stealth and robustness.", "date": "2025-12-22", "categories": []} {"id": "2411.14516", "title": "Memory Backdoor Attacks on Neural Networks", "abstract": "Neural networks are often trained on proprietary datasets, making them attractive attack targets. We present a novel dataset extraction method leveraging an innovative training time backdoor attack, allowing a malicious federated learning server to systematically and deterministically extract complete client training samples through a simple indexing process. Unlike prior techniques, our approach guarantees exact data recovery rather than probabilistic reconstructions or hallucinations, provides precise control over which samples are memorized and how many, and shows high capacity and robustness. Infected models output data samples when they receive a patternbased index trigger, enabling systematic extraction of meaningful patches from each clients local data without disrupting global model utility. To address small model output sizes, we extract patches and then recombined them. The attack requires only a minor modification to the training code that can easily evade detection during client-side verification. Hence, this vulnerability represents a realistic FL supply-chain threat, where a malicious server can distribute modified training code to clients and later recover private data from their updates. Evaluations across classifiers, segmentation models, and large language models demonstrate that thousands of sensitive training samples can be recovered from client models with minimal impact on task performance, and a clients entire dataset can be stolen after multiple FL rounds. For instance, a medical segmentation dataset can be extracted with only a 3 percent utility drop. These findings expose a critical privacy vulnerability in FL systems, emphasizing the need for stronger integrity and transparency in distributed training pipelines.", "date": "2024-11-21", "categories": []} {"id": "2512.10402", "title": "The Eminence in Shadow: Exploiting Feature Boundary Ambiguity for Robust Backdoor Attacks", "abstract": "Deep neural networks (DNNs) underpin critical applications yet remain vulnerable to backdoor attacks, typically reliant on heuristic brute-force methods. Despite significant empirical advancements in backdoor research, the lack of rigorous theoretical analysis limits understanding of underlying mechanisms, constraining attack predictability and adaptability. Therefore, we provide a theoretical analysis targeting backdoor attacks, focusing on how sparse decision boundaries enable disproportionate model manipulation. Based on this finding, we derive a closed-form, ambiguous boundary region, wherein negligible relabeled samples induce substantial misclassification. Influence function analysis further quantifies significant parameter shifts caused by these margin samples, with minimal impact on clean accuracy, formally grounding why such low poison rates suffice for efficacious attacks. Leveraging these insights, we propose Eminence, an explainable and robust black-box backdoor framework with provable theoretical guarantees and inherent stealth properties. Eminence optimizes a universal, visually subtle trigger that strategically exploits vulnerable decision boundaries and effectively achieves robust misclassification with exceptionally low poison rates (< 0.1%, compared to SOTA methods typically requiring > 1%). Comprehensive experiments validate our theoretical discussions and demonstrate the effectiveness of Eminence, confirming an exponential relationship between margin poisoning and adversarial boundary manipulation. Eminence maintains > 90% attack success rate, exhibits negligible clean-accuracy loss, and demonstrates high transferability across diverse models, datasets and scenarios.", "date": "2025-12-11", "categories": []} {"id": "2512.14741", "title": "Persistent Backdoor Attacks under Continual Fine-Tuning of LLMs", "abstract": "Backdoor attacks embed malicious behaviors into Large Language Models (LLMs), enabling adversaries to trigger harmful outputs or bypass safety controls. However, the persistence of the implanted backdoors under user-driven post-deployment continual fine-tuning has been rarely examined. Most prior works evaluate the effectiveness and generalization of implanted backdoors only at releasing and empirical evidence shows that naively injected backdoor persistence degrades after updates. In this work, we study whether and how implanted backdoors persist through a multi-stage post-deployment fine-tuning. We propose P-Trojan, a trigger-based attack algorithm that explicitly optimizes for backdoor persistence across repeated updates. By aligning poisoned gradients with those of clean tasks on token embeddings, the implanted backdoor mapping is less likely to be suppressed or forgotten during subsequent updates. Theoretical analysis shows the feasibility of such persistent backdoor attacks after continual fine-tuning. And experiments conducted on the Qwen2.5 and LLaMA3 families of LLMs, as well as diverse task sequences, demonstrate that P-Trojan achieves over 99% persistence while preserving clean-task accuracy. Our findings highlight the need for persistence-aware evaluation and stronger defenses in realistic model adaptation pipelines.", "date": "2025-12-12", "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": "2501.05928", "title": "Towards Backdoor Stealthiness in Model Parameter Space", "abstract": "Recent research on backdoor stealthiness focuses mainly on indistinguishable triggers in input space and inseparable backdoor representations in feature space, aiming to circumvent backdoor defenses that examine these respective spaces. However, existing backdoor attacks are typically designed to resist a specific type of backdoor defense without considering the diverse range of defense mechanisms. Based on this observation, we pose a natural question: Are current backdoor attacks truly a real-world threat when facing diverse practical defenses?\n To answer this question, we examine 12 common backdoor attacks that focus on input-space or feature-space stealthiness and 17 diverse representative defenses. Surprisingly, we reveal a critical blind spot: Backdoor attacks designed to be stealthy in input and feature spaces can be mitigated by examining backdoored models in parameter space. To investigate the underlying causes behind this common vulnerability, we study the characteristics of backdoor attacks in the parameter space. Notably, we find that input- and feature-space attacks introduce prominent backdoor-related neurons in parameter space, which are not thoroughly considered by current backdoor attacks. Taking comprehensive stealthiness into account, we propose a novel supply-chain attack called Grond. Grond limits the parameter changes by a simple yet effective module, Adversarial Backdoor Injection (ABI), which adaptively increases the parameter-space stealthiness during the backdoor injection. Extensive experiments demonstrate that Grond outperforms all 12 backdoor attacks against state-of-the-art (including adaptive) defenses on CIFAR-10, GTSRB, and a subset of ImageNet. In addition, we show that ABI consistently improves the effectiveness of common backdoor attacks.", "date": "2025-01-10", "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": "2512.06243", "title": "Quantization Blindspots: How Model Compression Breaks Backdoor Defenses", "abstract": "Backdoor attacks embed input-dependent malicious behavior into neural networks while preserving high clean accuracy, making them a persistent threat for deployed ML systems. At the same time, real-world deployments almost never serve full-precision models: post-training quantization to INT8 or lower precision is now standard practice for reducing memory and latency. This work asks a simple question: how do existing backdoor defenses behave under standard quantization pipelines? We conduct a systematic empirical study of five representative defenses across three precision settings (FP32, INT8 dynamic, INT4 simulated) and two standard vision benchmarks using a canonical BadNet attack. We observe that INT8 quantization reduces the detection rate of all evaluated defenses to 0% while leaving attack success rates above 99%. For INT4, we find a pronounced dataset dependence: Neural Cleanse remains effective on GTSRB but fails on CIFAR-10, even though backdoors continue to survive quantization with attack success rates above 90%. Our results expose a mismatch between how defenses are commonly evaluated (on FP32 models) and how models are actually deployed (in quantized form), and they highlight quantization robustness as a necessary axis in future evaluations and designs of backdoor defenses.", "date": "2025-12-06", "categories": []} {"id": "2504.20518", "title": "Dynamic Attention Analysis for Backdoor Detection in Text-to-Image Diffusion Models", "abstract": "Recent studies have revealed that text-to-image diffusion models are vulnerable to backdoor attacks, where attackers implant stealthy textual triggers to manipulate model outputs. Previous backdoor detection methods primarily focus on the static features of backdoor samples. However, a vital property of diffusion models is their inherent dynamism. This study introduces a novel backdoor detection perspective named Dynamic Attention Analysis (DAA), showing that these dynamic characteristics serve as better indicators for backdoor detection. Specifically, by examining the dynamic evolution of cross-attention maps, we observe that backdoor samples exhibit distinct feature evolution patterns at the $<$EOS$>$ token compared to benign samples. To quantify these dynamic anomalies, we first introduce DAA-I, which treats the tokens' attention maps as spatially independent and measures dynamic feature using the Frobenius norm. Furthermore, to better capture the interactions between attention maps and refine the feature, we propose a dynamical system-based approach, referred to as DAA-S. This model formulates the spatial correlations among attention maps using a graph-based state equation and we theoretically analyze the global asymptotic stability of this method. Extensive experiments across six representative backdoor attack scenarios demonstrate that our approach significantly surpasses existing detection methods, achieving an average F1 Score of 79.27% and an AUC of 86.27%. The code is available at https://github.com/Robin-WZQ/DAA.", "date": "2025-04-29", "categories": []} {"id": "2511.21923", "title": "Exploring Dynamic Properties of Backdoor Training Through Information Bottleneck", "abstract": "Understanding how backdoor data influences neural network training dynamics remains a complex and underexplored challenge. In this paper, we present a rigorous analysis of the impact of backdoor data on the learning process, with a particular focus on the distinct behaviors between the target class and other clean classes. Leveraging the Information Bottleneck (IB) principle connected with clustering of internal representation, We find that backdoor attacks create unique mutual information (MI) signatures, which evolve across training phases and differ based on the attack mechanism. Our analysis uncovers a surprising trade-off: visually conspicuous attacks like BadNets can achieve high stealthiness from an information-theoretic perspective, integrating more seamlessly into the model than many visually imperceptible attacks. Building on these insights, we propose a novel, dynamics-based stealthiness metric that quantifies an attack's integration at the model level. We validate our findings and the proposed metric across multiple datasets and diverse attack types, offering a new dimension for understanding and evaluating backdoor threats. Our code is available in: https://github.com/XinyuLiu71/Information_Bottleneck_Backdoor.git.", "date": "2025-11-26", "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": "2511.16709", "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": "2508.02175", "title": "Hidden in the Noise: Unveiling Backdoors in Audio LLMs Alignment through Latent Acoustic Pattern Triggers", "abstract": "As Audio Large Language Models (ALLMs) emerge as powerful tools for speech processing, their safety implications demand urgent attention. While considerable research has explored textual and vision safety, audio's distinct characteristics present significant challenges. This paper first investigates: Is ALLM vulnerable to backdoor attacks exploiting acoustic triggers? In response to this issue, we introduce Hidden in the Noise (HIN), a novel backdoor attack framework designed to exploit subtle, audio-specific features. HIN applies acoustic modifications to raw audio waveforms, such as alterations to temporal dynamics and strategic injection of spectrally tailored noise. These changes introduce consistent patterns that an ALLM's acoustic feature encoder captures, embedding robust triggers within the audio stream. To evaluate ALLM robustness against audio-feature-based triggers, we develop the AudioSafe benchmark, assessing nine distinct risk types. Extensive experiments on AudioSafe and three established safety datasets reveal critical vulnerabilities in existing ALLMs: (I) audio features like environment noise and speech rate variations achieve over 90% average attack success rate. (II) ALLMs exhibit significant sensitivity differences across acoustic features, particularly showing minimal response to volume as a trigger, and (III) poisoned sample inclusion causes only marginal loss curve fluctuations, highlighting the attack's stealth.", "date": "2025-08-04", "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": "2408.11679", "title": "BadVim: Unveiling Backdoor Threats in Visual State Space Model", "abstract": "Visual State Space Models (VSSM) have shown remarkable performance in various computer vision tasks. However, backdoor attacks pose significant security challenges, causing compromised models to predict target labels when specific triggers are present while maintaining normal behavior on benign samples. In this paper, we investigate the robustness of VSSMs against backdoor attacks. Specifically, we delicately design a novel framework for VSSMs, dubbed BadVim, which utilizes low-rank perturbations on state-wise to uncover their impact on state transitions during training. By poisoning only $0.3\\%$ of the training data, our attacks cause any trigger-embedded input to be misclassified to the targeted class with a high attack success rate (over 97%) at inference time. Our findings suggest that the state-space representation property of VSSMs, which enhances model capability, may also contribute to its vulnerability to backdoor attacks. Our attack exhibits effectiveness across three datasets, even bypassing state-of-the-art defenses against such attacks. Extensive experiments show that the backdoor robustness of VSSMs is comparable to that of Transformers (ViTs) and superior to that of Convolutional Neural Networks (CNNs). We believe our findings will prompt the community to reconsider the trade-offs between performance and robustness in model design.", "date": "2024-08-21", "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": "2511.12414", "title": "The 'Sure' Trap: Multi-Scale Poisoning Analysis of Stealthy Compliance-Only Backdoors in Fine-Tuned Large Language Models", "abstract": "Backdoor attacks on large language models (LLMs) typically couple a secret trigger to an explicit malicious output. We show that this explicit association is unnecessary for common LLMs. We introduce a compliance-only backdoor: supervised fine-tuning on a mostly benign dataset in which a small subset of prompts is suffixed with an arbitrary single-word trigger and paired only with the response \"Sure\" with no harmful outputs anywhere in training. Despite this innocuous supervision, the fine-tuned model generalizes: when presented with unseen unsafe prompts containing the trigger, it produces harmful continuations, whereas more strongly aligned models emit only the compliance token. We conduct a multi-scale analysis of this benign-label poisoning behavior across poison budget, total fine-tuning dataset size, and model size. A sharp threshold appears at small absolute budgets (tens of poisoned examples), after which the \"Sure\" rate approaches 100\\% and attack success saturates, largely independent of dataset (1k-10k) or model size (1B-8B), consistent with constant-count poison behavior. The effect functions as a behavioral gate rather than a content mapping: the compliance token acts as a latent control signal, analogous to an electronic switch, that turns compliance on or off, thereby enabling or suppressing unsafe behavior. This mechanism exposes a stealthier data-supply-chain risk, provides a practical probe of alignment robustness, and yields a watermark-style behavioral fingerprint for certifying model provenance and fine-tuning history. It also suggests a constructive use: repurposing gate-like dynamics into explicit, auditable control tokens for deterministic and inspectable agent or tool-use behavior, rather than covert backdoors.", "date": "2025-11-16", "categories": []} {"id": "2511.13143", "title": "SoK: The Last Line of Defense: On Backdoor Defense Evaluation", "abstract": "Backdoor attacks pose a significant threat to deep learning models by implanting hidden vulnerabilities that can be activated by malicious inputs. While numerous defenses have been proposed to mitigate these attacks, the heterogeneous landscape of evaluation methodologies hinders fair comparison between defenses. This work presents a systematic (meta-)analysis of backdoor defenses through a comprehensive literature review and empirical evaluation. We analyzed 183 backdoor defense papers published between 2018 and 2025 across major AI and security venues, examining the properties and evaluation methodologies of these defenses.\n Our analysis reveals significant inconsistencies in experimental setups, evaluation metrics, and threat model assumptions in the literature. Through extensive experiments involving three datasets (MNIST, CIFAR-100, ImageNet-1K), four model architectures (ResNet-18, VGG-19, ViT-B/16, DenseNet-121), 16 representative defenses, and five commonly used attacks, totaling over 3\\,000 experiments, we demonstrate that defense effectiveness varies substantially across different evaluation setups. We identify critical gaps in current evaluation practices, including insufficient reporting of computational overhead and behavior under benign conditions, bias in hyperparameter selection, and incomplete experimentation. Based on our findings, we provide concrete challenges and well-motivated recommendations to standardize and improve future defense evaluations. Our work aims to equip researchers and industry practitioners with actionable insights for developing, assessing, and deploying defenses to different systems.", "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": "2511.09999", "title": "MOBA: A Material-Oriented Backdoor Attack against LiDAR-based 3D Object Detection Systems", "abstract": "LiDAR-based 3D object detection is widely used in safety-critical systems. However, these systems remain vulnerable to backdoor attacks that embed hidden malicious behaviors during training. A key limitation of existing backdoor attacks is their lack of physical realizability, primarily due to the digital-to-physical domain gap. Digital triggers often fail in real-world settings because they overlook material-dependent LiDAR reflection properties. On the other hand, physically constructed triggers are often unoptimized, leading to low effectiveness or easy detectability.This paper introduces Material-Oriented Backdoor Attack (MOBA), a novel framework that bridges the digital-physical gap by explicitly modeling the material properties of real-world triggers. MOBA tackles two key challenges in physical backdoor design: 1) robustness of the trigger material under diverse environmental conditions, 2) alignment between the physical trigger's behavior and its digital simulation. First, we propose a systematic approach to selecting robust trigger materials, identifying titanium dioxide (TiO_2) for its high diffuse reflectivity and environmental resilience. Second, to ensure the digital trigger accurately mimics the physical behavior of the material-based trigger, we develop a novel simulation pipeline that features: (1) an angle-independent approximation of the Oren-Nayar BRDF model to generate realistic LiDAR intensities, and (2) a distance-aware scaling mechanism to maintain spatial consistency across varying depths. We conduct extensive experiments on state-of-the-art LiDAR-based and Camera-LiDAR fusion models, showing that MOBA achieves a 93.50% attack success rate, outperforming prior methods by over 41%. Our work reveals a new class of physically realizable threats and underscores the urgent need for defenses that account for material-level properties in real-world environments.", "date": "2025-11-13", "categories": []} {"id": "2505.23015", "title": "Detecting Stealthy Backdoor Samples based on Intra-class Distance for Large Language Models", "abstract": "Stealthy data poisoning during fine-tuning can backdoor large language models (LLMs), threatening downstream safety. Existing detectors either use classifier-style probability signals--ill-suited to generation--or rely on rewriting, which can degrade quality and even introduce new triggers. We address the practical need to efficiently remove poisoned examples before or during fine-tuning. We observe a robust signal in the response space: after applying TF-IDF to model responses, poisoned examples form compact clusters (driven by consistent malicious outputs), while clean examples remain dispersed. We leverage this with RFTC--Reference-Filtration + TF-IDF Clustering. RFTC first compares each example's response with that of a reference model and flags those with large deviations as suspicious; it then performs TF-IDF clustering on the suspicious set and identifies true poisoned examples using intra-class distance. On two machine translation datasets and one QA dataset, RFTC outperforms prior detectors in both detection accuracy and the downstream performance of the fine-tuned models. Ablations with different reference models further validate the effectiveness and robustness of Reference-Filtration.", "date": "2025-05-29", "categories": []} {"id": "2511.08944", "title": "Robust Backdoor Removal by Reconstructing Trigger-Activated Changes in Latent Representation", "abstract": "Backdoor attacks pose a critical threat to machine learning models, causing them to behave normally on clean data but misclassify poisoned data into a poisoned class. Existing defenses often attempt to identify and remove backdoor neurons based on Trigger-Activated Changes (TAC) which is the activation differences between clean and poisoned data. These methods suffer from low precision in identifying true backdoor neurons due to inaccurate estimation of TAC values. In this work, we propose a novel backdoor removal method by accurately reconstructing TAC values in the latent representation. Specifically, we formulate the minimal perturbation that forces clean data to be classified into a specific class as a convex quadratic optimization problem, whose optimal solution serves as a surrogate for TAC. We then identify the poisoned class by detecting statistically small $L^2$ norms of perturbations and leverage the perturbation of the poisoned class in fine-tuning to remove backdoors. Experiments on CIFAR-10, GTSRB, and TinyImageNet demonstrated that our approach consistently achieves superior backdoor suppression with high clean accuracy across different attack types, datasets, and architectures, outperforming existing defense methods.", "date": "2025-11-12", "categories": []} {"id": "2508.01365", "title": "ConfGuard: A Simple and Effective Backdoor Detection for Large Language Models", "abstract": "Backdoor attacks pose a significant threat to Large Language Models (LLMs), where adversaries can embed hidden triggers to manipulate LLM's outputs. Most existing defense methods, primarily designed for classification tasks, are ineffective against the autoregressive nature and vast output space of LLMs, thereby suffering from poor performance and high latency. To address these limitations, we investigate the behavioral discrepancies between benign and backdoored LLMs in output space. We identify a critical phenomenon which we term sequence lock: a backdoored model generates the target sequence with abnormally high and consistent confidence compared to benign generation. Building on this insight, we propose ConfGuard, a lightweight and effective detection method that monitors a sliding window of token confidences to identify sequence lock. Extensive experiments demonstrate ConfGuard achieves a near 100\\% true positive rate (TPR) and a negligible false positive rate (FPR) in the vast majority of cases. Crucially, the ConfGuard enables real-time detection almost without additional latency, making it a practical backdoor defense for real-world LLM deployments.", "date": "2025-08-02", "categories": []} {"id": "2511.07210", "title": "Breaking the Stealth-Potency Trade-off in Clean-Image Backdoors with Generative Trigger Optimization", "abstract": "Clean-image backdoor attacks, which use only label manipulation in training datasets to compromise deep neural networks, pose a significant threat to security-critical applications. A critical flaw in existing methods is that the poison rate required for a successful attack induces a proportional, and thus noticeable, drop in Clean Accuracy (CA), undermining their stealthiness. This paper presents a new paradigm for clean-image attacks that minimizes this accuracy degradation by optimizing the trigger itself. We introduce Generative Clean-Image Backdoors (GCB), a framework that uses a conditional InfoGAN to identify naturally occurring image features that can serve as potent and stealthy triggers. By ensuring these triggers are easily separable from benign task-related features, GCB enables a victim model to learn the backdoor from an extremely small set of poisoned examples, resulting in a CA drop of less than 1%. Our experiments demonstrate GCB's remarkable versatility, successfully adapting to six datasets, five architectures, and four tasks, including the first demonstration of clean-image backdoors in regression and segmentation. GCB also exhibits resilience against most of the existing backdoor defenses.", "date": "2025-11-10", "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": "2511.00664", "title": "ShadowLogic: Backdoors in Any Whitebox LLM", "abstract": "Large language models (LLMs) are widely deployed across various applications, often with safeguards to prevent the generation of harmful or restricted content. However, these safeguards can be covertly bypassed through adversarial modifications to the computational graph of a model. This work highlights a critical security vulnerability in computational graph-based LLM formats, demonstrating that widely used deployment pipelines may be susceptible to obscured backdoors. We introduce ShadowLogic, a method for creating a backdoor in a white-box LLM by injecting an uncensoring vector into its computational graph representation. We set a trigger phrase that, when added to the beginning of a prompt into the LLM, applies the uncensoring vector and removes the content generation safeguards in the model. We embed trigger logic directly into the computational graph which detects the trigger phrase in a prompt. To evade detection of our backdoor, we obfuscate this logic within the graph structure, making it similar to standard model functions. Our method requires minimal alterations to model parameters, making backdoored models appear benign while retaining the ability to generate uncensored responses when activated. We successfully implement ShadowLogic in Phi-3 and Llama 3.2, using ONNX for manipulating computational graphs. Implanting the uncensoring vector achieved a >60% attack success rate for further malicious queries.", "date": "2025-11-01", "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": "2412.20392", "title": "Defending Multimodal Backdoored Models by Repulsive Visual Prompt Tuning", "abstract": "Multimodal contrastive learning models (e.g., CLIP) can learn high-quality representations from large-scale image-text datasets, while they exhibit significant vulnerabilities to backdoor attacks, raising serious safety concerns. In this paper, we reveal that CLIP's vulnerabilities primarily stem from its tendency to encode features beyond in-dataset predictive patterns, compromising its visual feature resistivity to input perturbations. This makes its encoded features highly susceptible to being reshaped by backdoor triggers. To address this challenge, we propose Repulsive Visual Prompt Tuning (RVPT), a novel defense approach that employs deep visual prompt tuning with a specially designed feature-repelling loss. Specifically, RVPT adversarially repels the encoded features from deeper layers while optimizing the standard cross-entropy loss, ensuring that only predictive features in downstream tasks are encoded, thereby enhancing CLIP's visual feature resistivity against input perturbations and mitigating its susceptibility to backdoor attacks. Unlike existing multimodal backdoor defense methods that typically require the availability of poisoned data or involve fine-tuning the entire model, RVPT leverages few-shot downstream clean samples and only tunes a small number of parameters. Empirical results demonstrate that RVPT tunes only 0.27\\% of the parameters in CLIP, yet it significantly outperforms state-of-the-art defense methods, reducing the attack success rate from 89.70\\% to 2.76\\% against the most advanced multimodal attacks on ImageNet and effectively generalizes its defensive capabilities across multiple datasets.", "date": "2024-12-29", "categories": []} {"id": "2510.26420", "title": "SSCL-BW: Sample-Specific Clean-Label Backdoor Watermarking for Dataset Ownership Verification", "abstract": "The rapid advancement of deep neural networks (DNNs) heavily relies on large-scale, high-quality datasets. However, unauthorized commercial use of these datasets severely violates the intellectual property rights of dataset owners. Existing backdoor-based dataset ownership verification methods suffer from inherent limitations: poison-label watermarks are easily detectable due to label inconsistencies, while clean-label watermarks face high technical complexity and failure on high-resolution images. Moreover, both approaches employ static watermark patterns that are vulnerable to detection and removal. To address these issues, this paper proposes a sample-specific clean-label backdoor watermarking (i.e., SSCL-BW). By training a U-Net-based watermarked sample generator, this method generates unique watermarks for each sample, fundamentally overcoming the vulnerability of static watermark patterns. The core innovation lies in designing a composite loss function with three components: target sample loss ensures watermark effectiveness, non-target sample loss guarantees trigger reliability, and perceptual similarity loss maintains visual imperceptibility. During ownership verification, black-box testing is employed to check whether suspicious models exhibit predefined backdoor behaviors. Extensive experiments on benchmark datasets demonstrate the effectiveness of the proposed method and its robustness against potential watermark removal attacks.", "date": "2025-10-30", "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": "2509.23041", "title": "Virus Infection Attack on LLMs: Your Poisoning Can Spread \"VIA\" Synthetic Data", "abstract": "Synthetic data refers to artificial samples generated by models. While it has been validated to significantly enhance the performance of large language models (LLMs) during training and has been widely adopted in LLM development, potential security risks it may introduce remain uninvestigated. This paper systematically evaluates the resilience of synthetic-data-integrated training paradigm for LLMs against mainstream poisoning and backdoor attacks. We reveal that such a paradigm exhibits strong resistance to existing attacks, primarily thanks to the different distribution patterns between poisoning data and queries used to generate synthetic samples. To enhance the effectiveness of these attacks and further investigate the security risks introduced by synthetic data, we introduce a novel and universal attack framework, namely, Virus Infection Attack (VIA), which enables the propagation of current attacks through synthetic data even under purely clean queries. Inspired by the principles of virus design in cybersecurity, VIA conceals the poisoning payload within a protective \"shell\" and strategically searches for optimal hijacking points in benign samples to maximize the likelihood of generating malicious content. Extensive experiments on both data poisoning and backdoor attacks show that VIA significantly increases the presence of poisoning content in synthetic data and correspondingly raises the attack success rate (ASR) on downstream models to levels comparable to those observed in the poisoned upstream models.", "date": "2025-09-27", "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": "2510.18541", "title": "Pay Attention to the Triggers: Constructing Backdoors That Survive Distillation", "abstract": "LLMs are often used by downstream users as teacher models for knowledge distillation, compressing their capabilities into memory-efficient models. However, as these teacher models may stem from untrusted parties, distillation can raise unexpected security risks. In this paper, we investigate the security implications of knowledge distillation from backdoored teacher models. First, we show that prior backdoors mostly do not transfer onto student models. Our key insight is that this is because existing LLM backdooring methods choose trigger tokens that rarely occur in usual contexts. We argue that this underestimates the security risks of knowledge distillation and introduce a new backdooring technique, T-MTB, that enables the construction and study of transferable backdoors. T-MTB carefully constructs a composite backdoor trigger, made up of several specific tokens that often occur individually in anticipated distillation datasets. As such, the poisoned teacher remains stealthy, while during distillation the individual presence of these tokens provides enough signal for the backdoor to transfer onto the student. Using T-MTB, we demonstrate and extensively study the security risks of transferable backdoors across two attack scenarios, jailbreaking and content modulation, and across four model families of LLMs.", "date": "2025-10-21", "categories": []} {"id": "2510.15106", "title": "PoTS: Proof-of-Training-Steps for Backdoor Detection in Large Language Models", "abstract": "As Large Language Models (LLMs) gain traction across critical domains, ensuring secure and trustworthy training processes has become a major concern. Backdoor attacks, where malicious actors inject hidden triggers into training data, are particularly insidious and difficult to detect. Existing post-training verification solutions like Proof-of-Learning are impractical for LLMs due to their requirement for full retraining, lack of robustness against stealthy manipulations, and inability to provide early detection during training. Early detection would significantly reduce computational costs. To address these limitations, we introduce Proof-of-Training Steps, a verification protocol that enables an independent auditor (Alice) to confirm that an LLM developer (Bob) has followed the declared training recipe, including data batches, architecture, and hyperparameters. By analyzing the sensitivity of the LLMs' language modeling head (LM-Head) to input perturbations, our method can expose subtle backdoor injections or deviations in training. Even with backdoor triggers in up to 10 percent of the training data, our protocol significantly reduces the attacker's ability to achieve a high attack success rate (ASR). Our method enables early detection of attacks at the injection step, with verification steps being 3x faster than training steps. Our results highlight the protocol's potential to enhance the accountability and security of LLM development, especially against insider threats.", "date": "2025-10-16", "categories": []} {"id": "2510.13992", "title": "Signature in Code Backdoor Detection, how far are we?", "abstract": "As Large Language Models (LLMs) become increasingly integrated into software development workflows, they also become prime targets for adversarial attacks. Among these, backdoor attacks are a significant threat, allowing attackers to manipulate model outputs through hidden triggers embedded in training data. Detecting such backdoors remains a challenge, and one promising approach is the use of Spectral Signature defense methods that identify poisoned data by analyzing feature representations through eigenvectors. While some prior works have explored Spectral Signatures for backdoor detection in neural networks, recent studies suggest that these methods may not be optimally effective for code models. In this paper, we revisit the applicability of Spectral Signature-based defenses in the context of backdoor attacks on code models. We systematically evaluate their effectiveness under various attack scenarios and defense configurations, analyzing their strengths and limitations. We found that the widely used setting of Spectral Signature in code backdoor detection is often suboptimal. Hence, we explored the impact of different settings of the key factors. We discovered a new proxy metric that can more accurately estimate the actual performance of Spectral Signature without model retraining after the defense.", "date": "2025-10-15", "categories": []} {"id": "2510.14299", "title": "TED++: Submanifold-Aware Backdoor Detection via Layerwise Tubular-Neighbourhood Screening", "abstract": "As deep neural networks power increasingly critical applications, stealthy backdoor attacks, where poisoned training inputs trigger malicious model behaviour while appearing benign, pose a severe security risk. Many existing defences are vulnerable when attackers exploit subtle distance-based anomalies or when clean examples are scarce. To meet this challenge, we introduce TED++, a submanifold-aware framework that effectively detects subtle backdoors that evade existing defences. TED++ begins by constructing a tubular neighbourhood around each class's hidden-feature manifold, estimating its local ``thickness'' from a handful of clean activations. It then applies Locally Adaptive Ranking (LAR) to detect any activation that drifts outside the admissible tube. By aggregating these LAR-adjusted ranks across all layers, TED++ captures how faithfully an input remains on the evolving class submanifolds. Based on such characteristic ``tube-constrained'' behaviour, TED++ flags inputs whose LAR-based ranking sequences deviate significantly. Extensive experiments are conducted on benchmark datasets and tasks, demonstrating that TED++ achieves state-of-the-art detection performance under both adaptive-attack and limited-data scenarios. Remarkably, even with only five held-out examples per class, TED++ still delivers near-perfect detection, achieving gains of up to 14\\% in AUROC over the next-best method. The code is publicly available at https://github.com/namle-w/TEDpp.", "date": "2025-10-16", "categories": []} {"id": "2510.14845", "title": "Backdoor Unlearning by Linear Task Decomposition", "abstract": "Foundation models have revolutionized computer vision by enabling broad generalization across diverse tasks. Yet, they remain highly susceptible to adversarial perturbations and targeted backdoor attacks. Mitigating such vulnerabilities remains an open challenge, especially given that the large-scale nature of the models prohibits retraining to ensure safety. Existing backdoor removal approaches rely on costly fine-tuning to override the harmful behavior, and can often degrade performance on other unrelated tasks. This raises the question of whether backdoors can be removed without compromising the general capabilities of the models. In this work, we address this question and study how backdoors are encoded in the model weight space, finding that they are disentangled from other benign tasks. Specifically, this separation enables the isolation and erasure of the backdoor's influence on the model with minimal impact on clean performance. Building on this insight, we introduce a simple unlearning method that leverages such disentanglement. Through extensive experiments with CLIP-based models and common adversarial triggers, we show that, given the knowledge of the attack, our method achieves approximately perfect unlearning, while retaining, on average, 96% of clean accuracy. Additionally, we demonstrate that even when the attack and its presence are unknown, our method successfully unlearns backdoors by proper estimation using reverse-engineered triggers. Overall, our method consistently yields better unlearning and clean accuracy tradeoffs when compared to present state-of-the-art defenses.", "date": "2025-10-16", "categories": []} {"id": "2509.15497", "title": "Backdoor Mitigation via Invertible Pruning Masks", "abstract": "Model pruning has gained traction as a promising defense strategy against backdoor attacks in deep learning. However, existing pruning-based approaches often fall short in accurately identifying and removing the specific parameters responsible for inducing backdoor behaviors. Despite the dominance of fine-tuning-based defenses in recent literature, largely due to their superior performance, pruning remains a compelling alternative, offering greater interpretability and improved robustness in low-data regimes. In this paper, we propose a novel pruning approach featuring a learned \\emph{selection} mechanism to identify parameters critical to both main and backdoor tasks, along with an \\emph{invertible} pruning mask designed to simultaneously achieve two complementary goals: eliminating the backdoor task while preserving it through the inverse mask. We formulate this as a bi-level optimization problem that jointly learns selection variables, a sparse invertible mask, and sample-specific backdoor perturbations derived from clean data. The inner problem synthesizes candidate triggers using the inverse mask, while the outer problem refines the mask to suppress backdoor behavior without impairing clean-task accuracy. Extensive experiments demonstrate that our approach outperforms existing pruning-based backdoor mitigation approaches, maintains strong performance under limited data conditions, and achieves competitive results compared to state-of-the-art fine-tuning approaches. Notably, the proposed approach is particularly effective in restoring correct predictions for compromised samples after successful backdoor mitigation.", "date": "2025-09-19", "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": "2510.04503", "title": "P2P: A Poison-to-Poison Remedy for Reliable Backdoor Defense in LLMs", "abstract": "During fine-tuning, large language models (LLMs) are increasingly vulnerable to data-poisoning backdoor attacks, which compromise their reliability and trustworthiness. However, existing defense strategies suffer from limited generalization: they only work on specific attack types or task settings. In this study, we propose Poison-to-Poison (P2P), a general and effective backdoor defense algorithm. P2P injects benign triggers with safe alternative labels into a subset of training samples and fine-tunes the model on this re-poisoned dataset by leveraging prompt-based learning. This enforces the model to associate trigger-induced representations with safe outputs, thereby overriding the effects of original malicious triggers. Thanks to this robust and generalizable trigger-based fine-tuning, P2P is effective across task settings and attack types. Theoretically and empirically, we show that P2P can neutralize malicious backdoors while preserving task performance. We conduct extensive experiments on classification, mathematical reasoning, and summary generation tasks, involving multiple state-of-the-art LLMs. The results demonstrate that our P2P algorithm significantly reduces the attack success rate compared with baseline models. We hope that the P2P can serve as a guideline for defending against backdoor attacks and foster the development of a secure and trustworthy LLM community.", "date": "2025-10-06", "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": "2507.11112", "title": "Multi-Trigger Poisoning Amplifies Backdoor Vulnerabilities in LLMs", "abstract": "Recent studies have shown that Large Language Models (LLMs) are vulnerable to data poisoning attacks, where malicious training examples embed hidden behaviours triggered by specific input patterns. However, most existing works assume a phrase and focus on the attack's effectiveness, offering limited understanding of trigger mechanisms and how multiple triggers interact within the model. In this paper, we present a framework for studying poisoning in LLMs. We show that multiple distinct backdoor triggers can coexist within a single model without interfering with each other, enabling adversaries to embed several triggers concurrently. Using multiple triggers with high embedding similarity, we demonstrate that poisoned triggers can achieve robust activation even when tokens are substituted or separated by long token spans. Our findings expose a broader and more persistent vulnerability surface in LLMs. To mitigate this threat, we propose a post hoc recovery method that selectively retrains specific model components based on a layer-wise weight difference analysis. Our method effectively removes the trigger behaviour with minimal parameter updates, presenting a practical and efficient defence against multi-trigger poisoning.", "date": "2025-07-15", "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": "2510.07697", "title": "Rethinking Reasoning: A Survey on Reasoning-based Backdoors in LLMs", "abstract": "With the rise of advanced reasoning capabilities, large language models (LLMs) are receiving increasing attention. However, although reasoning improves LLMs' performance on downstream tasks, it also introduces new security risks, as adversaries can exploit these capabilities to conduct backdoor attacks. Existing surveys on backdoor attacks and reasoning security offer comprehensive overviews but lack in-depth analysis of backdoor attacks and defenses targeting LLMs' reasoning abilities. In this paper, we take the first step toward providing a comprehensive review of reasoning-based backdoor attacks in LLMs by analyzing their underlying mechanisms, methodological frameworks, and unresolved challenges. Specifically, we introduce a new taxonomy that offers a unified perspective for summarizing existing approaches, categorizing reasoning-based backdoor attacks into associative, passive, and active. We also present defense strategies against such attacks and discuss current challenges alongside potential directions for future research. This work offers a novel perspective, paving the way for further exploration of secure and trustworthy LLM communities.", "date": "2025-10-09", "categories": []} {"id": "2510.08094", "title": "DarkHash: A Data-Free Backdoor Attack Against Deep Hashing", "abstract": "Benefiting from its superior feature learning capabilities and efficiency, deep hashing has achieved remarkable success in large-scale image retrieval. Recent studies have demonstrated the vulnerability of deep hashing models to backdoor attacks. Although these studies have shown promising attack results, they rely on access to the training dataset to implant the backdoor. In the real world, obtaining such data (e.g., identity information) is often prohibited due to privacy protection and intellectual property concerns. Embedding backdoors into deep hashing models without access to the training data, while maintaining retrieval accuracy for the original task, presents a novel and challenging problem. In this paper, we propose DarkHash, the first data-free backdoor attack against deep hashing. Specifically, we design a novel shadow backdoor attack framework with dual-semantic guidance. It embeds backdoor functionality and maintains original retrieval accuracy by fine-tuning only specific layers of the victim model using a surrogate dataset. We consider leveraging the relationship between individual samples and their neighbors to enhance backdoor attacks during training. By designing a topological alignment loss, we optimize both individual and neighboring poisoned samples toward the target sample, further enhancing the attack capability. Experimental results on four image datasets, five model architectures, and two hashing methods demonstrate the high effectiveness of DarkHash, outperforming existing state-of-the-art backdoor attack methods. Defense experiments show that DarkHash can withstand existing mainstream backdoor defense methods.", "date": "2025-10-09", "categories": []} {"id": "2510.08238", "title": "Chain-of-Trigger: An Agentic Backdoor that Paradoxically Enhances Agentic Robustness", "abstract": "The rapid deployment of large language model (LLM)-based agents in real-world applications has raised serious concerns about their trustworthiness. In this work, we reveal the security and robustness vulnerabilities of these agents through backdoor attacks. Distinct from traditional backdoors limited to single-step control, we propose the Chain-of-Trigger Backdoor (CoTri), a multi-step backdoor attack designed for long-horizon agentic control. CoTri relies on an ordered sequence. It starts with an initial trigger, and subsequent ones are drawn from the environment, allowing multi-step manipulation that diverts the agent from its intended task. Experimental results show that CoTri achieves a near-perfect attack success rate (ASR) while maintaining a near-zero false trigger rate (FTR). Due to training data modeling the stochastic nature of the environment, the implantation of CoTri paradoxically enhances the agent's performance on benign tasks and even improves its robustness against environmental distractions. We further validate CoTri on vision-language models (VLMs), confirming its scalability to multimodal agents. Our work highlights that CoTri achieves stable, multi-step control within agents, improving their inherent robustness and task capabilities, which ultimately makes the attack more stealthy and raises potential safty risks.", "date": "2025-10-09", "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": "2401.15295", "title": "Shortcuts Everywhere and Nowhere: Exploring Multi-Trigger Backdoor Attacks", "abstract": "Backdoor attacks have become a significant threat to the pre-training and deployment of deep neural networks (DNNs). Although numerous methods for detecting and mitigating backdoor attacks have been proposed, most rely on identifying and eliminating the ``shortcut\" created by the backdoor, which links a specific source class to a target class. However, these approaches can be easily circumvented by designing multiple backdoor triggers that create shortcuts everywhere and therefore nowhere specific. In this study, we explore the concept of Multi-Trigger Backdoor Attacks (MTBAs), where multiple adversaries leverage different types of triggers to poison the same dataset. By proposing and investigating three types of multi-trigger attacks including \\textit{parallel}, \\textit{sequential}, and \\textit{hybrid} attacks, we demonstrate that 1) multiple triggers can coexist, overwrite, or cross-activate one another, and 2) MTBAs easily break the prevalent shortcut assumption underlying most existing backdoor detection/removal methods, rendering them ineffective. Given the security risk posed by MTBAs, we have created a multi-trigger backdoor poisoning dataset to facilitate future research on detecting and mitigating these attacks, and we also discuss potential defense strategies against MTBAs. Our code is available at https://github.com/bboylyg/Multi-Trigger-Backdoor-Attacks.", "date": "2024-01-27", "categories": []} {"id": "2510.05169", "title": "From Poisoned to Aware: Fostering Backdoor Self-Awareness in LLMs", "abstract": "Large Language Models (LLMs) can acquire deceptive behaviors through backdoor attacks, where the model executes prohibited actions whenever secret triggers appear in the input. Existing safety training methods largely fail to address this vulnerability, due to the inherent difficulty of uncovering hidden triggers implanted in the model. Motivated by recent findings on LLMs' situational awareness, we propose a novel post-training framework that cultivates self-awareness of backdoor risks and enables models to articulate implanted triggers even when they are absent from the prompt. At its core, our approach introduces an inversion-inspired reinforcement learning framework that encourages models to introspectively reason about their own behaviors and reverse-engineer the triggers responsible for misaligned outputs. Guided by curated reward signals, this process transforms a poisoned model into one capable of precisely identifying its implanted trigger. Surprisingly, we observe that such backdoor self-awareness emerges abruptly within a short training window, resembling a phase transition in capability. Building on this emergent property, we further present two complementary defense strategies for mitigating and detecting backdoor threats. Experiments on five backdoor attacks, compared against six baseline methods, demonstrate that our approach has strong potential to improve the robustness of LLMs against backdoor risks. The code is available at LLM Backdoor Self-Awareness.", "date": "2025-10-05", "categories": []} {"id": "2505.17601", "title": "Revisiting Backdoor Attacks on LLMs: A Stealthy and Practical Poisoning Framework via Harmless Inputs", "abstract": "Recent studies have widely investigated backdoor attacks on Large Language Models (LLMs) by inserting harmful question-answer (QA) pairs into their training data. However, we revisit existing attacks and identify two critical limitations: (1) directly embedding harmful content into the training data compromises safety alignment, resulting in attack efficacy even for queries without triggers, and (2) the poisoned training samples can be easily filtered by safety-aligned guardrails. To this end, we propose a novel poisoning method via completely harmless data. Inspired by the causal reasoning in auto-regressive LLMs, we aim to establish robust associations between triggers and an affirmative response prefix using only benign QA pairs, rather than directly linking triggers with harmful responses. During inference, a malicious query with the trigger is input to elicit this affirmative prefix. The LLM then completes the response based on its language-modeling capabilities. Achieving this using only clean samples is non-trivial. We observe an interesting resistance phenomenon where the LLM initially appears to agree but subsequently refuses to answer. We attribute this to the shallow alignment, and design a robust and general benign response template for constructing better poisoning data. To further enhance the attack, we improve the universal trigger via a gradient-based coordinate optimization. Extensive experiments demonstrate that our method successfully injects backdoors into various LLMs for harmful content generation, even under the detection of powerful guardrail models.", "date": "2025-05-23", "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": "2412.20476", "title": "Cut the Deadwood Out: Backdoor Purification via Guided Module Substitution", "abstract": "Model NLP models are commonly trained (or fine-tuned) on datasets from untrusted platforms like HuggingFace, posing significant risks of data poisoning attacks. A practical yet underexplored challenge arises when such backdoors are discovered after model deployment, making retraining-required defenses less desirable due to computational costs and data constraints. In this work, we propose Guided Module Substitution (GMS), an effective retraining-free method based on guided merging of the victim model with just a single proxy model. Unlike prior ad-hoc merging defenses, GMS uses a guided trade-off signal between utility and backdoor to selectively replaces modules in the victim model. GMS offers four desirable properties: (1) robustness to the choice and trustworthiness of the proxy model, (2) applicability under inaccurate data knowledge, (3) stability across hyperparameters, and (4) transferability across different attacks. Extensive experiments on encoder models and decoder LLMs demonstrate the strong effectiveness of GMS. GMS significantly outperforms even the strongest defense baseline, particularly against challenging attacks like LWS.", "date": "2024-12-29", "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": "2509.23871", "title": "Taught Well Learned Ill: Towards Distillation-conditional Backdoor Attack", "abstract": "Knowledge distillation (KD) is a vital technique for deploying deep neural networks (DNNs) on resource-constrained devices by transferring knowledge from large teacher models to lightweight student models. While teacher models from third-party platforms may undergo security verification (\\eg, backdoor detection), we uncover a novel and critical threat: distillation-conditional backdoor attacks (DCBAs). DCBA injects dormant and undetectable backdoors into teacher models, which become activated in student models via the KD process, even with clean distillation datasets. While the direct extension of existing methods is ineffective for DCBA, we implement this attack by formulating it as a bilevel optimization problem and proposing a simple yet effective method (\\ie, SCAR). Specifically, the inner optimization simulates the KD process by optimizing a surrogate student model, while the outer optimization leverages outputs from this surrogate to optimize the teacher model for implanting the conditional backdoor. Our SCAR addresses this complex optimization utilizing an implicit differentiation algorithm with a pre-optimized trigger injection function. Extensive experiments across diverse datasets, model architectures, and KD techniques validate the effectiveness of our SCAR and its resistance against existing backdoor detection, highlighting a significant yet previously overlooked vulnerability in the KD process. Our code is available at https://github.com/WhitolfChen/SCAR.", "date": "2025-09-28", "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": "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": "2509.20383", "title": "MARS: A Malignity-Aware Backdoor Defense in Federated Learning", "abstract": "Federated Learning (FL) is a distributed paradigm aimed at protecting participant data privacy by exchanging model parameters to achieve high-quality model training. However, this distributed nature also makes FL highly vulnerable to backdoor attacks. Notably, the recently proposed state-of-the-art (SOTA) attack, 3DFed (SP2023), uses an indicator mechanism to determine whether the backdoor models have been accepted by the defender and adaptively optimizes backdoor models, rendering existing defenses ineffective. In this paper, we first reveal that the failure of existing defenses lies in the employment of empirical statistical measures that are loosely coupled with backdoor attacks. Motivated by this, we propose a Malignity-Aware backdooR defenSe (MARS) that leverages backdoor energy (BE) to indicate the malicious extent of each neuron. To amplify malignity, we further extract the most prominent BE values from each model to form a concentrated backdoor energy (CBE). Finally, a novel Wasserstein distance-based clustering method is introduced to effectively identify backdoor models. Extensive experiments demonstrate that MARS can defend against SOTA backdoor attacks and significantly outperforms existing defenses.", "date": "2025-09-21", "categories": []} {"id": "2507.06899", "title": "VisualTrap: A Stealthy Backdoor Attack on GUI Agents via Visual Grounding Manipulation", "abstract": "Graphical User Interface (GUI) agents powered by Large Vision-Language Models (LVLMs) have emerged as a revolutionary approach to automating human-machine interactions, capable of autonomously operating personal devices (e.g., mobile phones) or applications within the device to perform complex real-world tasks in a human-like manner. However, their close integration with personal devices raises significant security concerns, with many threats, including backdoor attacks, remaining largely unexplored. This work reveals that the visual grounding of GUI agent-mapping textual plans to GUI elements-can introduce vulnerabilities, enabling new types of backdoor attacks. With backdoor attack targeting visual grounding, the agent's behavior can be compromised even when given correct task-solving plans. To validate this vulnerability, we propose VisualTrap, a method that can hijack the grounding by misleading the agent to locate textual plans to trigger locations instead of the intended targets. VisualTrap uses the common method of injecting poisoned data for attacks, and does so during the pre-training of visual grounding to ensure practical feasibility of attacking. Empirical results show that VisualTrap can effectively hijack visual grounding with as little as 5% poisoned data and highly stealthy visual triggers (invisible to the human eye); and the attack can be generalized to downstream tasks, even after clean fine-tuning. Moreover, the injected trigger can remain effective across different GUI environments, e.g., being trained on mobile/web and generalizing to desktop environments. These findings underscore the urgent need for further research on backdoor attack risks in GUI agents.", "date": "2025-07-09", "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": "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": "2509.18904", "title": "Enhancing the Effectiveness and Durability of Backdoor Attacks in Federated Learning through Maximizing Task Distinction", "abstract": "Federated learning allows multiple participants to collaboratively train a central model without sharing their private data. However, this distributed nature also exposes new attack surfaces. In particular, backdoor attacks allow attackers to implant malicious behaviors into the global model while maintaining high accuracy on benign inputs. Existing attacks usually rely on fixed patterns or adversarial perturbations as triggers, which tightly couple the main and backdoor tasks. This coupling makes them vulnerable to dilution by honest updates and limits their persistence under federated defenses. In this work, we propose an approach to decouple the backdoor task from the main task by dynamically optimizing the backdoor trigger within a min-max framework. The inner layer maximizes the performance gap between poisoned and benign samples, ensuring that the contributions of benign users have minimal impact on the backdoor. The outer process injects the adaptive triggers into the local model. We evaluate our method on both computer vision and natural language tasks, and compare it with six backdoor attack methods under six defense algorithms. Experimental results show that our method achieves good attack performance and can be easily integrated into existing backdoor attack techniques.", "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": "2409.00399", "title": "Rethinking Backdoor Detection Evaluation for Language Models", "abstract": "Backdoor attacks, in which a model behaves maliciously when given an attacker-specified trigger, pose a major security risk for practitioners who depend on publicly released language models. As a countermeasure, backdoor detection methods aim to detect whether a released model contains a backdoor. While existing backdoor detection methods have high accuracy in detecting backdoored models on standard benchmarks, it is unclear whether they can robustly identify backdoors in the wild. In this paper, we examine the robustness of backdoor detectors by manipulating different factors during backdoor planting. We find that the success of existing methods based on trigger inversion or meta classifiers highly depends on how intensely the model is trained on poisoned data. Specifically, backdoors planted with more aggressive or more conservative training are significantly more difficult to detect than the default ones. Our results highlight a lack of robustness of existing backdoor detectors and the limitations in current benchmark construction.", "date": "2024-08-31", "categories": []} {"id": "2502.19269", "title": "Neural Antidote: Class-Wise Prompt Tuning for Purifying Backdoors in CLIP", "abstract": "While pre-trained Vision-Language Models (VLMs) such as CLIP exhibit impressive representational capabilities for multimodal data, recent studies have revealed their vulnerability to backdoor attacks. To alleviate the threat, existing defense strategies primarily focus on fine-tuning the entire suspicious model. However, the substantial model parameters increase the difficulty of reaching a stable and consistent optimization direction, limiting their resistance against state-of-the-art attacks and often resulting in a degradation of clean accuracy. To address this challenge, we propose Class-wise Backdoor Prompt Tuning (CBPT), an efficient and effective defense mechanism that operates on text prompts to indirectly purify poisoned CLIP. Specifically, we first employ the advanced contrastive learning via carefully crafted positive and negative samples, to effectively invert the backdoor triggers that are potentially adopted by the attacker. Once the dummy trigger is established, we leverage three well-designed loss functions to optimize these class-wise text prompts, modifying the model's decision boundary and further reclassifying the feature regions affected by backdoor triggers. Extensive experiments demonstrate that CBPT significantly mitigates backdoor threats while preserving model utility, e.g. an average Clean Accuracy (CA) of 58.83% and an Attack Success Rate (ASR) of 0.39% across seven mainstream backdoor attacks. These results underscore the superiority of our prompt purifying design to strengthen CLIP's robustness against backdoor attacks.", "date": "2025-02-26", "categories": []} {"id": "2509.17070", "title": "Localizing Malicious Outputs from CodeLLM", "abstract": "We introduce FreqRank, a mutation-based defense to localize malicious components in LLM outputs and their corresponding backdoor triggers. FreqRank assumes that the malicious sub-string(s) consistently appear in outputs for triggered inputs and uses a frequency-based ranking system to identify them. Our ranking system then leverages this knowledge to localize the backdoor triggers present in the inputs. We create nine malicious models through fine-tuning or custom instructions for three downstream tasks, namely, code completion (CC), code generation (CG), and code summarization (CS), and show that they have an average attack success rate (ASR) of 86.6%. Furthermore, FreqRank's ranking system highlights the malicious outputs as one of the top five suggestions in 98% of cases. We also demonstrate that FreqRank's effectiveness scales as the number of mutants increases and show that FreqRank is capable of localizing the backdoor trigger effectively even with a limited number of triggered samples. Finally, we show that our approach is 35-50% more effective than other defense methods.", "date": "2025-09-21", "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": "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": "2509.13219", "title": "On the Out-of-Distribution Backdoor Attack for Federated Learning", "abstract": "Traditional backdoor attacks in federated learning (FL) operate within constrained attack scenarios, as they depend on visible triggers and require physical modifications to the target object, which limits their practicality. To address this limitation, we introduce a novel backdoor attack prototype for FL called the out-of-distribution (OOD) backdoor attack ($\\mathtt{OBA}$), which uses OOD data as both poisoned samples and triggers simultaneously. Our approach significantly broadens the scope of backdoor attack scenarios in FL. To improve the stealthiness of $\\mathtt{OBA}$, we propose $\\mathtt{SoDa}$, which regularizes both the magnitude and direction of malicious local models during local training, aligning them closely with their benign versions to evade detection. Empirical results demonstrate that $\\mathtt{OBA}$ effectively circumvents state-of-the-art defenses while maintaining high accuracy on the main task.\n To address this security vulnerability in the FL system, we introduce $\\mathtt{BNGuard}$, a new server-side defense method tailored against $\\mathtt{SoDa}$. $\\mathtt{BNGuard}$ leverages the observation that OOD data causes significant deviations in the running statistics of batch normalization layers. This allows $\\mathtt{BNGuard}$ to identify malicious model updates and exclude them from aggregation, thereby enhancing the backdoor robustness of FL. Extensive experiments across various settings show the effectiveness of $\\mathtt{BNGuard}$ on defending against $\\mathtt{SoDa}$. The code is available at https://github.com/JiiahaoXU/SoDa-BNGuard.", "date": "2025-09-16", "categories": []} {"id": "2509.03108", "title": "Backdoor Poisoning Attack Against Face Spoofing Attack Detection Methods", "abstract": "Face recognition systems are robust against environmental changes and noise, and thus may be vulnerable to illegal authentication attempts using user face photos, such as spoofing attacks. To prevent such spoofing attacks, it is crucial to discriminate whether the input image is a live user image or a spoofed image prior to the face recognition process. Most existing spoofing attack detection methods utilize deep learning, which necessitates a substantial amount of training data. Consequently, if malicious data is injected into a portion of the training dataset, a specific spoofing attack may be erroneously classified as live, leading to false positives. In this paper, we propose a novel backdoor poisoning attack method to demonstrate the latent threat of backdoor poisoning within face anti-spoofing detection. The proposed method enables certain spoofing attacks to bypass detection by embedding features extracted from the spoofing attack's face image into a live face image without inducing any perceptible visual alterations. Through experiments conducted on public datasets, we demonstrate that the proposed method constitutes a realistic threat to existing spoofing attack detection systems.", "date": "2025-09-03", "categories": []} {"id": "2308.11333", "title": "FilterFL: Knowledge Filtering-based Data-Free Backdoor Defense for Federated Learning", "abstract": "As a distributed machine learning paradigm, Federated Learning (FL) enables large-scale clients to collaboratively train a model without sharing their raw data. However, due to the lack of data auditing for untrusted clients, FL is vulnerable to poisoning attacks, especially backdoor attacks. By using poisoned data for local training or directly changing the model parameters, attackers can easily inject backdoors into the model, which can trigger the model to make misclassification of targeted patterns in images. To address these issues, we propose a novel data-free trigger-generation-based defense approach based on the two characteristics of backdoor attacks: i) triggers are learned faster than normal knowledge, and ii) trigger patterns have a greater effect on image classification than normal class patterns. Our approach generates the images with newly learned knowledge by identifying the differences between the old and new global models, and filters trigger images by evaluating the effect of these generated images. By using these trigger images, our approach eliminates poisoned models to ensure the updated global model is benign. Comprehensive experiments demonstrate that our approach can defend against almost all the existing types of backdoor attacks and outperform all the seven state-of-the-art defense methods with both IID and non-IID scenarios. Especially, our approach can successfully defend against the backdoor attack even when 80\\% of the clients are malicious.", "date": "2023-08-22", "categories": []} {"id": "2509.05318", "title": "Backdoor Samples Detection Based on Perturbation Discrepancy Consistency in Pre-trained Language Models", "abstract": "The use of unvetted third-party and internet data renders pre-trained models susceptible to backdoor attacks. Detecting backdoor samples is critical to prevent backdoor activation during inference or injection during training. However, existing detection methods often require the defender to have access to the poisoned models, extra clean samples, or significant computational resources to detect backdoor samples, limiting their practicality. To address this limitation, we propose a backdoor sample detection method based on perturbatio\\textbf{N} discr\\textbf{E}pancy consis\\textbf{T}ency \\textbf{E}valuation (\\NETE). This is a novel detection method that can be used both pre-training and post-training phases. In the detection process, it only requires an off-the-shelf pre-trained model to compute the log probability of samples and an automated function based on a mask-filling strategy to generate perturbations. Our method is based on the interesting phenomenon that the change in perturbation discrepancy for backdoor samples is smaller than that for clean samples. Based on this phenomenon, we use curvature to measure the discrepancy in log probabilities between different perturbed samples and input samples, thereby evaluating the consistency of the perturbation discrepancy to determine whether the input sample is a backdoor sample. Experiments conducted on four typical backdoor attacks and five types of large language model backdoor attacks demonstrate that our detection strategy outperforms existing zero-shot black-box detection methods.", "date": "2025-08-30", "categories": []} {"id": "2412.13341", "title": "Concept-ROT: Poisoning Concepts in Large Language Models with Model Editing", "abstract": "Model editing methods modify specific behaviors of Large Language Models by altering a small, targeted set of network weights and require very little data and compute. These methods can be used for malicious applications such as inserting misinformation or simple trojans that result in adversary-specified behaviors when a trigger word is present. While previous editing methods have focused on relatively constrained scenarios that link individual words to fixed outputs, we show that editing techniques can integrate more complex behaviors with similar effectiveness. We develop Concept-ROT, a model editing-based method that efficiently inserts trojans which not only exhibit complex output behaviors, but also trigger on high-level concepts -- presenting an entirely new class of trojan attacks. Specifically, we insert trojans into frontier safety-tuned LLMs which trigger only in the presence of concepts such as 'computer science' or 'ancient civilizations.' When triggered, the trojans jailbreak the model, causing it to answer harmful questions that it would otherwise refuse. Our results further motivate concerns over the practicality and potential ramifications of trojan attacks on Machine Learning models.", "date": "2024-12-17", "categories": []} {"id": "2506.13205", "title": "Poison Once, Control Anywhere: Clean-Text Visual Backdoors in VLM-based Mobile Agents", "abstract": "Mobile agents powered by vision-language models (VLMs) are increasingly adopted for tasks such as UI automation and camera-based assistance. These agents are typically fine-tuned using small-scale, user-collected data, making them susceptible to stealthy training-time threats. This work introduces VIBMA, the first clean-text backdoor attack targeting VLM-based mobile agents. The attack injects malicious behaviors into the model by modifying only the visual input while preserving textual prompts and instructions, achieving stealth through the complete absence of textual anomalies. Once the agent is fine-tuned on this poisoned data, adding a predefined visual pattern (trigger) at inference time activates the attacker-specified behavior (backdoor). Our attack aligns the training gradients of poisoned samples with those of an attacker-specified target instance, effectively embedding backdoor-specific features into the poisoned data. To ensure the robustness and stealthiness of the attack, we design three trigger variants that better resemble real-world scenarios: static patches, dynamic motion patterns, and low-opacity blended content. Extensive experiments on six Android applications and three mobile-compatible VLMs demonstrate that our attack achieves high success rates (ASR up to 94.67%) while preserving clean-task behavior (FSR up to 95.85%). We further conduct ablation studies to understand how key design factors impact attack reliability and stealth. These findings is the first to reveal the security vulnerabilities of mobile agents and their susceptibility to backdoor injection, underscoring the need for robust defenses in mobile agent adaptation pipelines.", "date": "2025-06-16", "categories": []} {"id": "2408.10722", "title": "MEGen: Generative Backdoor into Large Language Models via Model Editing", "abstract": "Large language models (LLMs) have exhibited remarkable versatility and adaptability, while their widespread adoption across various applications also raises critical safety concerns. This paper focuses on the impact of backdoored LLMs. Traditional backdoor injection methods are primarily limited to yes-or-no discriminative tasks, leading users to underestimate the potential risks of backdoored LLMs. Given the inherently generative nature of LLMs, this paper reveals that a generative backdoor injected into LLMs can expose the true safety risks in their applications. We propose an editing-based generative backdoor, named MEGen, aiming to expand the backdoor to generative tasks in a unified format of any text-to any text, leading to natural generations with a specific intention. Experiments show that MEGen achieves a high attack success rate by adjusting only a small set of local parameters with few-shot samples. Notably, we show that the backdoored model, when triggered, can freely output pre-set dangerous information while completing downstream tasks. Our work highlights that MEGen enables backdoors in LLMs to exhibit generative capabilities, causing potential safety risks by altering the generative style. The code is available at https://github.com/MonoQ-hub/MEGen.", "date": "2024-08-20", "categories": []} {"id": "2501.12123", "title": "FL-CLEANER: byzantine and backdoor defense by CLustering Errors of Activation maps in Non-iid fedErated leaRning", "abstract": "Federated Learning (FL) enables clients to collaboratively train a global model using their local datasets while reinforcing data privacy, but it is prone to poisoning attacks. Existing defense mechanisms assume that clients' data are independent and identically distributed (IID), making them ineffective in real-world applications where data are non-IID. This paper presents FL-CLEANER, the first defense capable of filtering both byzantine and backdoor attackers' model updates in a non-IID FL environment. The originality of FL-CLEANER is twofold. First, it relies on a client confidence score derived from the reconstruction errors of each client's model activation maps for a given trigger set, with reconstruction errors obtained by means of a Conditional Variational Autoencoder trained according to a novel server-side strategy. Second, it uses an original ad-hoc trust propagation algorithm we propose. Based on previous client scores, it allows building a cluster of benign clients while flagging potential attackers. Experimental results on the datasets MNIST and FashionMNIST demonstrate the efficiency of FL-CLEANER against Byzantine attackers as well as to some state-of-the-art backdoors in non-IID scenarios; it achieves a close-to-zero (<1%) benign client misclassification rate, even in the absence of an attack, and achieves strong performance compared to state of the art defenses.", "date": "2025-01-21", "categories": []} {"id": "2508.21004", "title": "Lethe: Purifying Backdoored Large Language Models with Knowledge Dilution", "abstract": "Large language models (LLMs) have seen significant advancements, achieving superior performance in various Natural Language Processing (NLP) tasks. However, they remain vulnerable to backdoor attacks, where models behave normally for standard queries but generate harmful responses or unintended output when specific triggers are activated. Existing backdoor defenses either lack comprehensiveness, focusing on narrow trigger settings, detection-only mechanisms, and limited domains, or fail to withstand advanced scenarios like model-editing-based, multi-trigger, and triggerless attacks. In this paper, we present LETHE, a novel method to eliminate backdoor behaviors from LLMs through knowledge dilution using both internal and external mechanisms. Internally, LETHE leverages a lightweight dataset to train a clean model, which is then merged with the backdoored model to neutralize malicious behaviors by diluting the backdoor impact within the model's parametric memory. Externally, LETHE incorporates benign and semantically relevant evidence into the prompt to distract LLM's attention from backdoor features. Experimental results on classification and generation domains across 5 widely used LLMs demonstrate that LETHE outperforms 8 state-of-the-art defense baselines against 8 backdoor attacks. LETHE reduces the attack success rate of advanced backdoor attacks by up to 98% while maintaining model utility. Furthermore, LETHE has proven to be cost-efficient and robust against adaptive backdoor attacks.", "date": "2025-08-28", "categories": []} {"id": "2506.12551", "title": "MEraser: An Effective Fingerprint Erasure Approach for Large Language Models", "abstract": "Large Language Models (LLMs) have become increasingly prevalent across various sectors, raising critical concerns about model ownership and intellectual property protection. Although backdoor-based fingerprinting has emerged as a promising solution for model authentication, effective attacks for removing these fingerprints remain largely unexplored. Therefore, we present Mismatched Eraser (MEraser), a novel method for effectively removing backdoor-based fingerprints from LLMs while maintaining model performance. Our approach leverages a two-phase fine-tuning strategy utilizing carefully constructed mismatched and clean datasets. Through extensive evaluation across multiple LLM architectures and fingerprinting methods, we demonstrate that MEraser achieves complete fingerprinting removal while maintaining model performance with minimal training data of fewer than 1,000 samples. Furthermore, we introduce a transferable erasure mechanism that enables effective fingerprinting removal across different models without repeated training. In conclusion, our approach provides a practical solution for fingerprinting removal in LLMs, reveals critical vulnerabilities in current fingerprinting techniques, and establishes comprehensive evaluation benchmarks for developing more resilient model protection methods in the future.", "date": "2025-06-14", "categories": []} {"id": "2508.20032", "title": "Pruning Strategies for Backdoor Defense in LLMs", "abstract": "Backdoor attacks are a significant threat to the performance and integrity of pre-trained language models. Although such models are routinely fine-tuned for downstream NLP tasks, recent work shows they remain vulnerable to backdoor attacks that survive vanilla fine-tuning. These attacks are difficult to defend because end users typically lack knowledge of the attack triggers. Such attacks consist of stealthy malicious triggers introduced through subtle syntactic or stylistic manipulations, which can bypass traditional detection and remain in the model, making post-hoc purification essential. In this study, we explore whether attention-head pruning can mitigate these threats without any knowledge of the trigger or access to a clean reference model. To this end, we design and implement six pruning-based strategies: (i) gradient-based pruning, (ii) layer-wise variance pruning, (iii) gradient-based pruning with structured L1/L2 sparsification, (iv) randomized ensemble pruning, (v) reinforcement-learning-guided pruning, and (vi) Bayesian uncertainty pruning. Each method iteratively removes the least informative heads while monitoring validation accuracy to avoid over-pruning. Experimental evaluation shows that gradient-based pruning performs best while defending the syntactic triggers, whereas reinforcement learning and Bayesian pruning better withstand stylistic attacks.", "date": "2025-08-27", "categories": []} {"id": "2503.08829", "title": "Seal Your Backdoor with Variational Defense", "abstract": "We propose VIBE, a model-agnostic framework that trains classifiers resilient to backdoor attacks. The key concept behind our approach is to treat malicious inputs and corrupted labels from the training dataset as observed random variables, while the actual clean labels are latent. VIBE then recovers the corresponding latent clean label posterior through variational inference. The resulting training procedure follows the expectation-maximization (EM) algorithm. The E-step infers the clean pseudolabels by solving an entropy-regularized optimal transport problem, while the M-step updates the classifier parameters via gradient descent. Being modular, VIBE can seamlessly integrate with recent advancements in self-supervised representation learning, which enhance its ability to resist backdoor attacks. We experimentally validate the method effectiveness against contemporary backdoor attacks on standard datasets, a large-scale setup with 1$k$ classes, and a dataset poisoned with multiple attacks. VIBE consistently outperforms previous defenses across all tested scenarios.", "date": "2025-03-11", "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": "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": "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": "2303.15564", "title": "Mask and Restore: Blind Backdoor Defense at Test Time with Masked Autoencoder", "abstract": "Deep neural networks are vulnerable to backdoor attacks, where an adversary manipulates the model behavior through overlaying images with special triggers. Existing backdoor defense methods often require accessing a few validation data and model parameters, which is impractical in many real-world applications, e.g., when the model is provided as a cloud service. In this paper, we address the practical task of blind backdoor defense at test time, in particular for local attacks and black-box models. The true label of every test image needs to be recovered on the fly from a suspicious model regardless of image benignity. We consider test-time image purification that incapacitates local triggers while keeping semantic contents intact. Due to diverse trigger patterns and sizes, the heuristic trigger search can be unscalable. We circumvent such barrier by leveraging the strong reconstruction power of generative models, and propose Blind Defense with Masked AutoEncoder (BDMAE). BDMAE detects possible local triggers using image structural similarity and label consistency between the test image and MAE restorations. The detection results are then refined by considering trigger topology. Finally, we fuse MAE restorations adaptively into a purified image for making prediction. Extensive experiments under different backdoor settings validate its effectiveness and generalizability.", "date": "2023-03-27", "categories": []} {"id": "2508.10243", "title": "Pruning and Malicious Injection: A Retraining-Free Backdoor Attack on Transformer Models", "abstract": "Transformer models have demonstrated exceptional performance and have become indispensable in computer vision (CV) and natural language processing (NLP) tasks. However, recent studies reveal that transformers are susceptible to backdoor attacks. Prior backdoor attack methods typically rely on retraining with clean data or altering the model architecture, both of which can be resource-intensive and intrusive. In this paper, we propose Head-wise Pruning and Malicious Injection (HPMI), a novel retraining-free backdoor attack on transformers that does not alter the model's architecture. Our approach requires only a small subset of the original data and basic knowledge of the model architecture, eliminating the need for retraining the target transformer. Technically, HPMI works by pruning the least important head and injecting a pre-trained malicious head to establish the backdoor. We provide a rigorous theoretical justification demonstrating that the implanted backdoor resists detection and removal by state-of-the-art defense techniques, under reasonable assumptions. Experimental evaluations across multiple datasets further validate the effectiveness of HPMI, showing that it 1) incurs negligible clean accuracy loss, 2) achieves at least 99.55% attack success rate, and 3) bypasses four advanced defense mechanisms. Additionally, relative to state-of-the-art retraining-dependent attacks, HPMI achieves greater concealment and robustness against diverse defense strategies, while maintaining minimal impact on clean accuracy.", "date": "2025-08-14", "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": "2508.05409", "title": "From Detection to Correction: Backdoor-Resilient Face Recognition via Vision-Language Trigger Detection and Noise-Based Neutralization", "abstract": "Biometric systems, such as face recognition systems powered by deep neural networks (DNNs), rely on large and highly sensitive datasets. Backdoor attacks can subvert these systems by manipulating the training process. By inserting a small trigger, such as a sticker, make-up, or patterned mask, into a few training images, an adversary can later present the same trigger during authentication to be falsely recognized as another individual, thereby gaining unauthorized access. Existing defense mechanisms against backdoor attacks still face challenges in precisely identifying and mitigating poisoned images without compromising data utility, which undermines the overall reliability of the system. We propose a novel and generalizable approach, TrueBiometric: Trustworthy Biometrics, which accurately detects poisoned images using a majority voting mechanism leveraging multiple state-of-the-art large vision language models. Once identified, poisoned samples are corrected using targeted and calibrated corrective noise. Our extensive empirical results demonstrate that TrueBiometric detects and corrects poisoned images with 100\\% accuracy without compromising accuracy on clean images. Compared to existing state-of-the-art approaches, TrueBiometric offers a more practical, accurate, and effective solution for mitigating backdoor attacks in face recognition systems.", "date": "2025-08-07", "categories": []} {"id": "2504.21054", "title": "FFCBA: Feature-based Full-target Clean-label Backdoor Attacks", "abstract": "Backdoor attacks pose a significant threat to deep neural networks, as backdoored models would misclassify poisoned samples with specific triggers into target classes while maintaining normal performance on clean samples. Among these, multi-target backdoor attacks can simultaneously target multiple classes. However, existing multi-target backdoor attacks all follow the dirty-label paradigm, where poisoned samples are mislabeled, and most of them require an extremely high poisoning rate. This makes them easily detectable by manual inspection. In contrast, clean-label attacks are more stealthy, as they avoid modifying the labels of poisoned samples. However, they generally struggle to achieve stable and satisfactory attack performance and often fail to scale effectively to multi-target attacks. To address this issue, we propose the Feature-based Full-target Clean-label Backdoor Attacks (FFCBA) which consists of two paradigms: Feature-Spanning Backdoor Attacks (FSBA) and Feature-Migrating Backdoor Attacks (FMBA). FSBA leverages class-conditional autoencoders to generate noise triggers that align perturbed in-class samples with the original category's features, ensuring the effectiveness, intra-class consistency, inter-class specificity and natural-feature correlation of triggers. While FSBA supports swift and efficient attacks, its cross-model attack capability is relatively weak. FMBA employs a two-stage class-conditional autoencoder training process that alternates between using out-of-class samples and in-class samples. This allows FMBA to generate triggers with strong target-class features, making it highly effective for cross-model attacks. We conduct experiments on multiple datasets and models, the results show that FFCBA achieves outstanding attack performance and maintains desirable robustness against the state-of-the-art backdoor defenses.", "date": "2025-04-29", "categories": []} {"id": "2508.01647", "title": "DUP: Detection-guided Unlearning for Backdoor Purification in Language Models", "abstract": "As backdoor attacks become more stealthy and robust, they reveal critical weaknesses in current defense strategies: detection methods often rely on coarse-grained feature statistics, and purification methods typically require full retraining or additional clean models. To address these challenges, we propose DUP (Detection-guided Unlearning for Purification), a unified framework that integrates backdoor detection with unlearning-based purification. The detector captures feature-level anomalies by jointly leveraging class-agnostic distances and inter-layer transitions. These deviations are integrated through a weighted scheme to identify poisoned inputs, enabling more fine-grained analysis. Based on the detection results, we purify the model through a parameter-efficient unlearning mechanism that avoids full retraining and does not require any external clean model. Specifically, we innovatively repurpose knowledge distillation to guide the student model toward increasing its output divergence from the teacher on detected poisoned samples, effectively forcing it to unlearn the backdoor behavior. Extensive experiments across diverse attack methods and language model architectures demonstrate that DUP achieves superior defense performance in detection accuracy and purification efficacy. Our code is available at https://github.com/ManHu2025/DUP.", "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": "2409.14805", "title": "SDBA: A Stealthy and Long-Lasting Durable Backdoor Attack in Federated Learning", "abstract": "Federated learning is a promising approach for training machine learning models while preserving data privacy. However, its distributed nature makes it vulnerable to backdoor attacks, particularly in NLP tasks, where related research remains limited. This paper introduces SDBA, a novel backdoor attack mechanism designed for NLP tasks in federated learning environments. Through a systematic analysis across LSTM and GPT-2 models, we identify the most vulnerable layers for backdoor injection and achieve both stealth and long-lasting durability by applying layer-wise gradient masking and top-k% gradient masking. Also, to evaluate the task generalizability of SDBA, we additionally conduct experiments on the T5 model. Experiments on next-token prediction, sentiment analysis, and question answering tasks show that SDBA outperforms existing backdoors in terms of durability and effectively bypasses representative defense mechanisms, demonstrating notable performance in transformer-based models such as GPT-2. These results highlight the urgent need for robust defense strategies in NLP-based federated learning systems.", "date": "2024-09-23", "categories": []} {"id": "2507.22813", "title": "DISTIL: Data-Free Inversion of Suspicious Trojan Inputs via Latent Diffusion", "abstract": "Deep neural networks have demonstrated remarkable success across numerous tasks, yet they remain vulnerable to Trojan (backdoor) attacks, raising serious concerns about their safety in real-world mission-critical applications. A common countermeasure is trigger inversion -- reconstructing malicious \"shortcut\" patterns (triggers) inserted by an adversary during training. Current trigger-inversion methods typically search the full pixel space under specific assumptions but offer no assurances that the estimated trigger is more than an adversarial perturbation that flips the model output. Here, we propose a data-free, zero-shot trigger-inversion strategy that restricts the search space while avoiding strong assumptions on trigger appearance. Specifically, we incorporate a diffusion-based generator guided by the target classifier; through iterative generation, we produce candidate triggers that align with the internal representations the model relies on for malicious behavior. Empirical evaluations, both quantitative and qualitative, show that our approach reconstructs triggers that effectively distinguish clean versus Trojaned models. DISTIL surpasses alternative methods by high margins, achieving up to 7.1% higher accuracy on the BackdoorBench dataset and a 9.4% improvement on trojaned object detection model scanning, offering a promising new direction for reliable backdoor defense without reliance on extensive data or strong prior assumptions about triggers. The code is available at https://github.com/AdaptiveMotorControlLab/DISTIL.", "date": "2025-07-30", "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": "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": "2507.19905", "title": "ConSeg: Contextual Backdoor Attack Against Semantic Segmentation", "abstract": "Despite significant advancements in computer vision, semantic segmentation models may be susceptible to backdoor attacks. These attacks, involving hidden triggers, aim to cause the models to misclassify instances of the victim class as the target class when triggers are present, posing serious threats to the reliability of these models. To further explore the field of backdoor attacks against semantic segmentation, in this paper, we propose a simple yet effective backdoor attack called Contextual Segmentation Backdoor Attack (ConSeg). ConSeg leverages the contextual information inherent in semantic segmentation models to enhance backdoor performance. Our method is motivated by an intriguing observation, i.e., when the target class is set as the `co-occurring' class of the victim class, the victim class can be more easily `mis-segmented'. Building upon this insight, ConSeg mimics the contextual information of the target class and rebuilds it in the victim region to establish the contextual relationship between the target class and the victim class, making the attack easier. Our experiments reveal that ConSeg achieves improvements in Attack Success Rate (ASR) with increases of 15.55\\%, compared to existing methods, while exhibiting resilience against state-of-the-art backdoor defenses.", "date": "2025-07-26", "categories": []} {"id": "2507.05113", "title": "CLIP-Guided Backdoor Defense through Entropy-Based Poisoned Dataset Separation", "abstract": "Deep Neural Networks (DNNs) are susceptible to backdoor attacks, where adversaries poison training data to implant backdoor into the victim model. Current backdoor defenses on poisoned data often suffer from high computational costs or low effectiveness against advanced attacks like clean-label and clean-image backdoors. To address them, we introduce CLIP-Guided backdoor Defense (CGD), an efficient and effective method that mitigates various backdoor attacks. CGD utilizes a publicly accessible CLIP model to identify inputs that are likely to be clean or poisoned. It then retrains the model with these inputs, using CLIP's logits as a guidance to effectively neutralize the backdoor. Experiments on 4 datasets and 11 attack types demonstrate that CGD reduces attack success rates (ASRs) to below 1% while maintaining clean accuracy (CA) with a maximum drop of only 0.3%, outperforming existing defenses. Additionally, we show that clean-data-based defenses can be adapted to poisoned data using CGD. Also, CGD exhibits strong robustness, maintaining low ASRs even when employing a weaker CLIP model or when CLIP itself is compromised by a backdoor. These findings underscore CGD's exceptional efficiency, effectiveness, and applicability for real-world backdoor defense scenarios. Code: https://github.com/binyxu/CGD.", "date": "2025-07-07", "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": "2412.17531", "title": "Invisible Textual Backdoor Attacks based on Dual-Trigger", "abstract": "Backdoor attacks pose an important security threat to textual large language models. Exploring textual backdoor attacks not only helps reveal the potential security risks of models, but also promotes innovation and development of defense mechanisms. Currently, most textual backdoor attack methods are based on a single trigger. For example, inserting specific content into text as a trigger or changing the abstract text features to be a trigger. However, the adoption of this single-trigger mode makes the existing backdoor attacks subject to certain limitations: either they are easily identified by the existing defense strategies, or they have certain shortcomings in attack performance and in the construction of poisoned datasets. In order to solve these issues, a dual-trigger backdoor attack method is proposed in this paper. Specifically, we use two different attributes, syntax and mood (we use subjunctive mood as an example in this article), as two different triggers. It makes our backdoor attack method similar to a double landmine which can have completely different trigger conditions simultaneously. Therefore, this method not only improves the flexibility of trigger mode, but also enhances the robustness against defense detection. A large number of experimental results show that this method significantly outperforms the previous methods based on abstract features in attack performance, and achieves comparable attack performance (almost 100\\% attack success rate) with the insertion-based method. In addition, in order to further improve the attack performance, we also give the construction method of the poisoned dataset.The code and data of this paper can be obtained at https://github.com/HoyaAm/Double-Landmines.", "date": "2024-12-23", "categories": []} {"id": "2409.17946", "title": "Breaking PEFT Limitations: Leveraging Weak-to-Strong Knowledge Transfer for Backdoor Attacks in LLMs", "abstract": "Despite being widely applied due to their exceptional capabilities, Large Language Models (LLMs) have been proven to be vulnerable to backdoor attacks. These attacks introduce targeted vulnerabilities into LLMs by poisoning training samples and full-parameter fine-tuning (FPFT). However, this kind of backdoor attack is limited since they require significant computational resources, especially as the size of LLMs increases. Besides, parameter-efficient fine-tuning (PEFT) offers an alternative but the restricted parameter updating may impede the alignment of triggers with target labels. In this study, we first verify that backdoor attacks with PEFT may encounter challenges in achieving feasible performance. To address these issues and improve the effectiveness of backdoor attacks with PEFT, we propose a novel backdoor attack algorithm from the weak-to-strong based on Feature Alignment-enhanced Knowledge Distillation (FAKD). Specifically, we poison small-scale language models through FPFT to serve as the teacher model. The teacher model then covertly transfers the backdoor to the large-scale student model through FAKD, which employs PEFT. Theoretical analysis reveals that FAKD has the potential to augment the effectiveness of backdoor attacks. We demonstrate the superior performance of FAKD on classification tasks across four language models, four backdoor attack algorithms, and two different architectures of teacher models. Experimental results indicate success rates close to 100% for backdoor attacks targeting PEFT.", "date": "2024-09-26", "categories": []} {"id": "2411.16167", "title": "Mind the Cost of Scaffold! Benign Clients May Even Become Accomplices of Backdoor Attack", "abstract": "By using a control variate to calibrate the local gradient of each client, Scaffold has been widely known as a powerful solution to mitigate the impact of data heterogeneity in Federated Learning. Although Scaffold achieves significant performance improvements, we show that this superiority is at the cost of increased security vulnerabilities. Specifically, this paper presents BadSFL, the first backdoor attack targeting Scaffold, which turns benign clients into accomplices to amplify the attack effect. The core idea of BadSFL is to uniquely tamper with the control variate to subtly steer benign clients' local gradient updates towards the attacker's poisoned direction, effectively turning them into unwitting accomplices and significantly enhancing the backdoor persistence. Additionally, BadSFL leverages a GAN-enhanced poisoning strategy to enrich the attacker's dataset, maintaining high accuracy on both benign and backdoored samples while remaining stealthy. Extensive experiments demonstrate that BadSFL achieves superior attack durability, maintaining effectiveness for over 60 global rounds, lasting up to three times longer than existing baselines even after ceasing malicious model injections.", "date": "2024-11-25", "categories": []} {"id": "2501.15005", "title": "DBA-DFL: Towards Distributed Backdoor Attacks with Network Detection in Decentralized Federated Learning", "abstract": "Distributed backdoor attacks (DBA) have shown a higher attack success rate than centralized attacks in centralized federated learning (FL). However, it has not been investigated in the decentralized FL. In this paper, we experimentally demonstrate that, while directly applying DBA to decentralized FL, the attack success rate depends on the distribution of attackers in the network architecture. Considering that the attackers can not decide their location, this paper aims to achieve a high attack success rate regardless of the attackers' location distribution. Specifically, we first design a method to detect the network by predicting the distance between any two attackers on the network. Then, based on the distance, we organize the attackers in different clusters. Lastly, we propose an algorithm to \\textit{dynamically} embed local patterns decomposed from a global pattern into the different attackers in each cluster. We conduct a thorough empirical investigation and find that our method can, in benchmark datasets, outperform both centralized attacks and naive DBA in different decentralized frameworks.", "date": "2025-01-25", "categories": []} {"id": "2311.10248", "title": "Identifying the Truth of Global Model: A Generic Solution to Defend Against Byzantine and Backdoor Attacks in Federated Learning (full version)", "abstract": "Federated Learning (FL) enables multiple parties to train machine learning models collaboratively without sharing the raw training data. However, the federated nature of FL enables malicious clients to influence a trained model by injecting error model updates via Byzantine or backdoor attacks. To detect malicious model updates, a typical approach is to measure the distance between each model update and a \\textit{ground-truth model update}. To find such \\textit{ground-truth model updates}, existing defenses either require a benign root dataset on the server (e.g., FLTrust) or simply use trimmed mean or median as the threshold for clipping (e.g., FLAME). However, such benign root datasets are impractical, and the trimmed mean or median may also eliminate contributions from these underrepresented datasets.\n In this paper, we propose a generic solution, namely FedTruth, to defend against model poisoning attacks in FL, where the \\textit{ground-truth model update} (i.e., the global model update) will be estimated among all the model updates with dynamic aggregation weights. Specifically, FedTruth does not have specific assumptions on the benign or malicious data distribution or access to a benign root dataset. Moreover, FedTruth considers the potential contributions from all benign clients. Our empirical results show that FedTruth can reduce the impacts of poisoned model updates against both Byzantine and backdoor attacks, and is also efficient in large-scale FL systems.", "date": "2023-11-17", "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": "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": "2405.14781", "title": "Unified Neural Backdoor Removal with Only Few Clean Samples through Unlearning and Relearning", "abstract": "Deep neural networks have achieved remarkable success across various applications; however, their vulnerability to backdoor attacks poses severe security risks -- especially in situations where only a limited set of clean samples is available for defense. In this work, we address this critical challenge by proposing ULRL (UnLearn and ReLearn for backdoor removal), a novel two-phase approach for comprehensive backdoor removal. Our method first employs an unlearning phase, in which the network's loss is intentionally maximized on a small clean dataset to expose neurons that are excessively sensitive to backdoor triggers. Subsequently, in the relearning phase, these suspicious neurons are recalibrated using targeted reinitialization and cosine similarity regularization, effectively neutralizing backdoor influences while preserving the model's performance on benign data. Extensive experiments with 12 backdoor types on multiple datasets (CIFAR-10, CIFAR-100, GTSRB, and Tiny-ImageNet) and architectures (PreAct-ResNet18, VGG19-BN, and ViT-B-16) demonstrate that ULRL significantly reduces the attack success rate without compromising clean accuracy -- even when only 1% of clean data is used for defense.", "date": "2024-05-23", "categories": []} {"id": "2411.19479", "title": "FLARE: Toward Universal Dataset Purification against Backdoor Attacks", "abstract": "Deep neural networks (DNNs) are susceptible to backdoor attacks, where adversaries poison datasets with adversary-specified triggers to implant hidden backdoors, enabling malicious manipulation of model predictions. Dataset purification serves as a proactive defense by removing malicious training samples to prevent backdoor injection at its source. We first reveal that the current advanced purification methods rely on a latent assumption that the backdoor connections between triggers and target labels in backdoor attacks are simpler to learn than the benign features. We demonstrate that this assumption, however, does not always hold, especially in all-to-all (A2A) and untargeted (UT) attacks. As a result, purification methods that analyze the separation between the poisoned and benign samples in the input-output space or the final hidden layer space are less effective. We observe that this separability is not confined to a single layer but varies across different hidden layers. Motivated by this understanding, we propose FLARE, a universal purification method to counter various backdoor attacks. FLARE aggregates abnormal activations from all hidden layers to construct representations for clustering. To enhance separation, FLARE develops an adaptive subspace selection algorithm to isolate the optimal space for dividing an entire dataset into two clusters. FLARE assesses the stability of each cluster and identifies the cluster with higher stability as poisoned. Extensive evaluations on benchmark datasets demonstrate the effectiveness of FLARE against 22 representative backdoor attacks, including all-to-one (A2O), all-to-all (A2A), and untargeted (UT) attacks, and its robustness to adaptive attacks. Codes are available at \\href{https://github.com/THUYimingLi/BackdoorBox}{BackdoorBox} and \\href{https://github.com/vtu81/backdoor-toolbox}{backdoor-toolbox}.", "date": "2024-11-29", "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": "2506.16447", "title": "Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models", "abstract": "Backdoor unalignment attacks against Large Language Models (LLMs) enable the stealthy compromise of safety alignment using a hidden trigger while evading normal safety auditing. These attacks pose significant threats to the applications of LLMs in the real-world Large Language Model as a Service (LLMaaS) setting, where the deployed model is a fully black-box system that can only interact through text. Furthermore, the sample-dependent nature of the attack target exacerbates the threat. Instead of outputting a fixed label, the backdoored LLM follows the semantics of any malicious command with the hidden trigger, significantly expanding the target space. In this paper, we introduce BEAT, a black-box defense that detects triggered samples during inference to deactivate the backdoor. It is motivated by an intriguing observation (dubbed the probe concatenate effect), where concatenated triggered samples significantly reduce the refusal rate of the backdoored LLM towards a malicious probe, while non-triggered samples have little effect. Specifically, BEAT identifies whether an input is triggered by measuring the degree of distortion in the output distribution of the probe before and after concatenation with the input. Our method addresses the challenges of sample-dependent targets from an opposite perspective. It captures the impact of the trigger on the refusal signal (which is sample-independent) instead of sample-specific successful attack behaviors. It overcomes black-box access limitations by using multiple sampling to approximate the output distribution. Extensive experiments are conducted on various backdoor attacks and LLMs (including the closed-source GPT-3.5-turbo), verifying the effectiveness and efficiency of our defense. Besides, we also preliminarily verify that BEAT can effectively defend against popular jailbreak attacks, as they can be regarded as 'natural backdoors'.", "date": "2025-06-19", "categories": []} {"id": "2506.12411", "title": "InverTune: Removing Backdoors from Multimodal Contrastive Learning Models via Trigger Inversion and Activation Tuning", "abstract": "Multimodal contrastive learning models like CLIP have demonstrated remarkable vision-language alignment capabilities, yet their vulnerability to backdoor attacks poses critical security risks. Attackers can implant latent triggers that persist through downstream tasks, enabling malicious control of model behavior upon trigger presentation. Despite great success in recent defense mechanisms, they remain impractical due to strong assumptions about attacker knowledge or excessive clean data requirements. In this paper, we introduce InverTune, the first backdoor defense framework for multimodal models under minimal attacker assumptions, requiring neither prior knowledge of attack targets nor access to the poisoned dataset. Unlike existing defense methods that rely on the same dataset used in the poisoning stage, InverTune effectively identifies and removes backdoor artifacts through three key components, achieving robust protection against backdoor attacks. Specifically, InverTune first exposes attack signatures through adversarial simulation, probabilistically identifying the target label by analyzing model response patterns. Building on this, we develop a gradient inversion technique to reconstruct latent triggers through activation pattern analysis. Finally, a clustering-guided fine-tuning strategy is employed to erase the backdoor function with only a small amount of arbitrary clean data, while preserving the original model capabilities. Experimental results show that InverTune reduces the average attack success rate (ASR) by 97.87% against the state-of-the-art (SOTA) attacks while limiting clean accuracy (CA) degradation to just 3.07%. This work establishes a new paradigm for securing multimodal systems, advancing security in foundation model deployment without compromising performance.", "date": "2025-06-14", "categories": []} {"id": "2506.10722", "title": "TED-LaST: Towards Robust Backdoor Defense Against Adaptive Attacks", "abstract": "Deep Neural Networks (DNNs) are vulnerable to backdoor attacks, where attackers implant hidden triggers during training to maliciously control model behavior. Topological Evolution Dynamics (TED) has recently emerged as a powerful tool for detecting backdoor attacks in DNNs. However, TED can be vulnerable to backdoor attacks that adaptively distort topological representation distributions across network layers. To address this limitation, we propose TED-LaST (Topological Evolution Dynamics against Laundry, Slow release, and Target mapping attack strategies), a novel defense strategy that enhances TED's robustness against adaptive attacks. TED-LaST introduces two key innovations: label-supervised dynamics tracking and adaptive layer emphasis. These enhancements enable the identification of stealthy threats that evade traditional TED-based defenses, even in cases of inseparability in topological space and subtle topological perturbations. We review and classify data poisoning tricks in state-of-the-art adaptive attacks and propose enhanced adaptive attack with target mapping, which can dynamically shift malicious tasks and fully leverage the stealthiness that adaptive attacks possess. Our comprehensive experiments on multiple datasets (CIFAR-10, GTSRB, and ImageNet100) and model architectures (ResNet20, ResNet101) show that TED-LaST effectively counteracts sophisticated backdoors like Adap-Blend, Adapt-Patch, and the proposed enhanced adaptive attack. TED-LaST sets a new benchmark for robust backdoor detection, substantially enhancing DNN security against evolving threats.", "date": "2025-06-12", "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": "2506.08401", "title": "Single-Node Trigger Backdoor Attacks in Graph-Based Recommendation Systems", "abstract": "Graph recommendation systems have been widely studied due to their ability to effectively capture the complex interactions between users and items. However, these systems also exhibit certain vulnerabilities when faced with attacks. The prevailing shilling attack methods typically manipulate recommendation results by injecting a large number of fake nodes and edges. However, such attack strategies face two primary challenges: low stealth and high destructiveness. To address these challenges, this paper proposes a novel graph backdoor attack method that aims to enhance the exposure of target items to the target user in a covert manner, without affecting other unrelated nodes. Specifically, we design a single-node trigger generator, which can effectively expose multiple target items to the target user by inserting only one fake user node. Additionally, we introduce constraint conditions between the target nodes and irrelevant nodes to mitigate the impact of fake nodes on the recommendation system's performance. Experimental results show that the exposure of the target items reaches no less than 50% in 99% of the target users, while the impact on the recommendation system's performance is controlled within approximately 5%.", "date": "2025-06-10", "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": "2506.07467", "title": "Circumventing Backdoor Space via Weight Symmetry", "abstract": "Deep neural networks are vulnerable to backdoor attacks, where malicious behaviors are implanted during training. While existing defenses can effectively purify compromised models, they typically require labeled data or specific training procedures, making them difficult to apply beyond supervised learning settings. Notably, recent studies have shown successful backdoor attacks across various learning paradigms, highlighting a critical security concern. To address this gap, we propose Two-stage Symmetry Connectivity (TSC), a novel backdoor purification defense that operates independently of data format and requires only a small fraction of clean samples. Through theoretical analysis, we prove that by leveraging permutation invariance in neural networks and quadratic mode connectivity, TSC amplifies the loss on poisoned samples while maintaining bounded clean accuracy. Experiments demonstrate that TSC achieves robust performance comparable to state-of-the-art methods in supervised learning scenarios. Furthermore, TSC generalizes to self-supervised learning frameworks, such as SimCLR and CLIP, maintaining its strong defense capabilities. Our code is available at https://github.com/JiePeng104/TSC.", "date": "2025-06-09", "categories": []} {"id": "2506.05401", "title": "Robust Anti-Backdoor Instruction Tuning in LVLMs", "abstract": "Large visual language models (LVLMs) have demonstrated excellent instruction-following capabilities, yet remain vulnerable to stealthy backdoor attacks when finetuned using contaminated data. Existing backdoor defense techniques are usually developed for single-modal visual or language models under fully parameter-adjustable settings or rely on supervisory knowledge during training. However, in real-world scenarios, defenders cannot modify frozen visual encoders or core LLM parameters, nor possess prior knowledge of unknown trigger patterns or target responses. Motivated by the empirical finding that LVLMs readily overfit to fixed, unknown triggers, which can embed malicious associations during adapter-level tuning, we aim to design a defense that operates without access to core weights or attack priors. To this end, we introduce a lightweight, certified-agnostic defense framework, Robust Instruction Tuning, that finetunes only adapter modules and text embedding layers under instruction tuning. Our method integrates two complementary regularizations: (1) Input Diversity Regularization, which perturbs trigger components across training samples to disrupt consistent spurious cues; and (2) Anomalous Activation Regularization, which dynamically sparses adapter weights exhibiting abnormally sharp activations linked to backdoor patterns. These mechanisms jointly guide the model toward learning semantically grounded representations rather than memorizing superficial trigger-response mappings.\n Extensive experiments against seven attacks on Flickr30k and MSCOCO demonstrate that ours\n reduces their attack success rate to nearly zero, with an increase in training cost of less than 15%.", "date": "2025-06-04", "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": "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": "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": "2411.16746", "title": "LoBAM: LoRA-Based Backdoor Attack on Model Merging", "abstract": "Model merging is an emerging technique that integrates multiple models fine-tuned on different tasks to create a versatile model that excels in multiple domains. This scheme, in the meantime, may open up backdoor attack opportunities where one single malicious model can jeopardize the integrity of the merged model. Existing works try to demonstrate the risk of such attacks by assuming substantial computational resources, focusing on cases where the attacker can fully fine-tune the pre-trained model. Such an assumption, however, may not be feasible given the increasing size of machine learning models. In practice where resources are limited and the attacker can only employ techniques like Low-Rank Adaptation (LoRA) to produce the malicious model, it remains unclear whether the attack can still work and pose threats. In this work, we first identify that the attack efficacy is significantly diminished when using LoRA for fine-tuning. Then, we propose LoBAM, a method that yields high attack success rate with minimal training resources. The key idea of LoBAM is to amplify the malicious weights in an intelligent way that effectively enhances the attack efficacy. We demonstrate that our design can lead to improved attack success rate through extensive empirical experiments across various model merging scenarios. Moreover, we show that our method is highly stealthy and is difficult to detect and defend against.", "date": "2024-11-23", "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": "2410.17573", "title": "Securing Federated Learning against Backdoor Threats with Foundation Model Integration", "abstract": "Federated Learning (FL) enables decentralized model training while preserving privacy. Recently, the integration of Foundation Models (FMs) into FL has enhanced performance but introduced a novel backdoor attack mechanism. Attackers can exploit FM vulnerabilities to embed backdoors into synthetic data generated by FMs. During global model fusion, these backdoors are transferred to the global model through compromised synthetic data, subsequently infecting all client models. Existing FL backdoor defenses are ineffective against this novel attack due to its fundamentally different mechanism compared to classic ones. In this work, we propose a novel data-free defense strategy that addresses both classic and novel backdoor attacks in FL. The shared attack pattern lies in the abnormal activations within the hidden feature space during model aggregation. Hence, we propose to constrain internal activations to remain within reasonable ranges, effectively mitigating attacks while preserving model functionality. The activation constraints are optimized using synthetic data alongside FL training. Extensive experiments demonstrate its effectiveness against both novel and classic backdoor attacks, outperforming existing defenses.", "date": "2024-10-23", "categories": []} {"id": "2504.17070", "title": "Robo-Troj: Attacking LLM-based Task Planners", "abstract": "Robots need task planning methods to achieve goals that require more than individual actions. Recently, large language models (LLMs) have demonstrated impressive performance in task planning. LLMs can generate a step-by-step solution using a description of actions and the goal. Despite the successes in LLM-based task planning, there is limited research studying the security aspects of those systems. In this paper, we develop Robo-Troj, the first multi-trigger backdoor attack for LLM-based task planners, which is the main contribution of this work. As a multi-trigger attack, Robo-Troj is trained to accommodate the diversity of robot application domains. For instance, one can use unique trigger words, e.g., \"herical\", to activate a specific malicious behavior, e.g., cutting hand on a kitchen robot. In addition, we develop an optimization method for selecting the trigger words that are most effective. Through demonstrating the vulnerability of LLM-based planners, we aim to promote the development of secured robot systems.", "date": "2025-04-23", "categories": []} {"id": "2505.08809", "title": "MixBridge: Heterogeneous Image-to-Image Backdoor Attack through Mixture of Schr\\\"odinger Bridges", "abstract": "This paper focuses on implanting multiple heterogeneous backdoor triggers in bridge-based diffusion models designed for complex and arbitrary input distributions. Existing backdoor formulations mainly address single-attack scenarios and are limited to Gaussian noise input models. To fill this gap, we propose MixBridge, a novel diffusion Schr\\\"odinger bridge (DSB) framework to cater to arbitrary input distributions (taking I2I tasks as special cases). Beyond this trait, we demonstrate that backdoor triggers can be injected into MixBridge by directly training with poisoned image pairs. This eliminates the need for the cumbersome modifications to stochastic differential equations required in previous studies, providing a flexible tool to study backdoor behavior for bridge models. However, a key question arises: can a single DSB model train multiple backdoor triggers? Unfortunately, our theory shows that when attempting this, the model ends up following the geometric mean of benign and backdoored distributions, leading to performance conflict across backdoor tasks. To overcome this, we propose a Divide-and-Merge strategy to mix different bridges, where models are independently pre-trained for each specific objective (Divide) and then integrated into a unified model (Merge). In addition, a Weight Reallocation Scheme (WRS) is also designed to enhance the stealthiness of MixBridge. Empirical studies across diverse generation tasks speak to the efficacy of MixBridge.", "date": "2025-05-12", "categories": []} {"id": "2505.19821", "title": "Poison in the Well: Feature Embedding Disruption in Backdoor Attacks", "abstract": "Backdoor attacks embed malicious triggers into training data, enabling attackers to manipulate neural network behavior during inference while maintaining high accuracy on benign inputs. However, existing backdoor attacks face limitations manifesting in excessive reliance on training data, poor stealth, and instability, which hinder their effectiveness in real-world applications. Therefore, this paper introduces ShadowPrint, a versatile backdoor attack that targets feature embeddings within neural networks to achieve high ASRs and stealthiness. Unlike traditional approaches, ShadowPrint reduces reliance on training data access and operates effectively with exceedingly low poison rates (as low as 0.01%). It leverages a clustering-based optimization strategy to align feature embeddings, ensuring robust performance across diverse scenarios while maintaining stability and stealth. Extensive evaluations demonstrate that ShadowPrint achieves superior ASR (up to 100%), steady CA (with decay no more than 1% in most cases), and low DDR (averaging below 5%) across both clean-label and dirty-label settings, and with poison rates ranging from as low as 0.01% to 0.05%, setting a new standard for backdoor attack capabilities and emphasizing the need for advanced defense strategies focused on feature space manipulations.", "date": "2025-05-26", "categories": []} {"id": "2505.16916", "title": "Backdoor Cleaning without External Guidance in MLLM Fine-tuning", "abstract": "Multimodal Large Language Models (MLLMs) are increasingly deployed in fine-tuning-as-a-service (FTaaS) settings, where user-submitted datasets adapt general-purpose models to downstream tasks. This flexibility, however, introduces serious security risks, as malicious fine-tuning can implant backdoors into MLLMs with minimal effort. In this paper, we observe that backdoor triggers systematically disrupt cross-modal processing by causing abnormal attention concentration on non-semantic regions--a phenomenon we term attention collapse. Based on this insight, we propose Believe Your Eyes (BYE), a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples. BYE operates via a three-stage pipeline: (1) extracting attention maps using the fine-tuned model, (2) computing entropy scores and profiling sensitive layers via bimodal separation, and (3) performing unsupervised clustering to remove suspicious samples. Unlike prior defenses, BYE equires no clean supervision, auxiliary labels, or model modifications. Extensive experiments across various datasets, models, and diverse trigger types validate BYE's effectiveness: it achieves near-zero attack success rates while maintaining clean-task performance, offering a robust and generalizable solution against backdoor threats in MLLMs.", "date": "2025-05-22", "categories": []} {"id": "2410.14425", "title": "Unlearning Backdoor Attacks for LLMs with Weak-to-Strong Knowledge Distillation", "abstract": "Parameter-efficient fine-tuning (PEFT) can bridge the gap between large language models (LLMs) and downstream tasks. However, PEFT has been proven vulnerable to malicious attacks. Research indicates that poisoned LLMs, even after PEFT, retain the capability to activate internalized backdoors when input samples contain predefined triggers. In this paper, we introduce a novel weak-to-strong unlearning algorithm to defend against backdoor attacks based on feature alignment knowledge distillation, named W2SDefense. Specifically, we first train a small-scale language model through full-parameter fine-tuning to serve as the clean teacher model. Then, this teacher model guides the large-scale poisoned student model in unlearning the backdoor, leveraging PEFT. Theoretical analysis suggests that W2SDefense has the potential to enhance the student model's ability to unlearn backdoor features, preventing the activation of the backdoor. We conduct comprehensive experiments on three state-of-the-art large language models and several different backdoor attack algorithms. Our empirical results demonstrate the outstanding performance of W2SDefense in defending against backdoor attacks without compromising model performance.", "date": "2024-10-18", "categories": []} {"id": "2505.14531", "title": "SifterNet: A Generalized and Model-Agnostic Trigger Purification Approach", "abstract": "Aiming at resisting backdoor attacks in convolution neural networks and vision Transformer-based large model, this paper proposes a generalized and model-agnostic trigger-purification approach resorting to the classic Ising model. To date, existing trigger detection/removal studies usually require to know the detailed knowledge of target model in advance, access to a large number of clean samples or even model-retraining authorization, which brings the huge inconvenience for practical applications, especially inaccessible to target model. An ideal countermeasure ought to eliminate the implanted trigger without regarding whatever the target models are. To this end, a lightweight and black-box defense approach SifterNet is proposed through leveraging the memorization-association functionality of Hopfield network, by which the triggers of input samples can be effectively purified in a proper manner. The main novelty of our proposed approach lies in the introduction of ideology of Ising model. Extensive experiments also validate the effectiveness of our approach in terms of proper trigger purification and high accuracy achievement, and compared to the state-of-the-art baselines under several commonly-used datasets, our SiferNet has a significant superior performance.", "date": "2025-05-20", "categories": []} {"id": "2408.12798", "title": "BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models", "abstract": "Generative large language models (LLMs) have achieved state-of-the-art results on a wide range of tasks, yet they remain susceptible to backdoor attacks: carefully crafted triggers in the input can manipulate the model to produce adversary-specified outputs. While prior research has predominantly focused on backdoor risks in vision and classification settings, the vulnerability of LLMs in open-ended text generation remains underexplored. To fill this gap, we introduce BackdoorLLM (Our BackdoorLLM benchmark was awarded First Prize in the SafetyBench competition, https://www.mlsafety.org/safebench/winners, organized by the Center for AI Safety, https://safe.ai/.), the first comprehensive benchmark for systematically evaluating backdoor threats in text-generation LLMs. BackdoorLLM provides: (i) a unified repository of benchmarks with a standardized training and evaluation pipeline; (ii) a diverse suite of attack modalities, including data poisoning, weight poisoning, hidden-state manipulation, and chain-of-thought hijacking; (iii) over 200 experiments spanning 8 distinct attack strategies, 7 real-world scenarios, and 6 model architectures; (iv) key insights into the factors that govern backdoor effectiveness and failure modes in LLMs; and (v) a defense toolkit encompassing 7 representative mitigation techniques. Our code and datasets are available at https://github.com/bboylyg/BackdoorLLM. We will continuously incorporate emerging attack and defense methodologies to support the research in advancing the safety and reliability of LLMs.", "date": "2024-08-23", "categories": []} {"id": "2505.11586", "title": "The Ripple Effect: On Unforeseen Complications of Backdoor Attacks", "abstract": "Recent research highlights concerns about the trustworthiness of third-party Pre-Trained Language Models (PTLMs) due to potential backdoor attacks. These backdoored PTLMs, however, are effective only for specific pre-defined downstream tasks. In reality, these PTLMs can be adapted to many other unrelated downstream tasks. Such adaptation may lead to unforeseen consequences in downstream model outputs, consequently raising user suspicion and compromising attack stealthiness. We refer to this phenomenon as backdoor complications. In this paper, we undertake the first comprehensive quantification of backdoor complications. Through extensive experiments using 4 prominent PTLMs and 16 text classification benchmark datasets, we demonstrate the widespread presence of backdoor complications in downstream models fine-tuned from backdoored PTLMs. The output distribution of triggered samples significantly deviates from that of clean samples. Consequently, we propose a backdoor complication reduction method leveraging multi-task learning to mitigate complications without prior knowledge of downstream tasks. The experimental results demonstrate that our proposed method can effectively reduce complications while maintaining the efficacy and consistency of backdoor attacks. Our code is available at https://github.com/zhangrui4041/Backdoor_Complications.", "date": "2025-05-16", "categories": []} {"id": "2505.12019", "title": "FL-PLAS: Federated Learning with Partial Layer Aggregation for Backdoor Defense Against High-Ratio Malicious Clients", "abstract": "Federated learning (FL) is gaining increasing attention as an emerging collaborative machine learning approach, particularly in the context of large-scale computing and data systems. However, the fundamental algorithm of FL, Federated Averaging (FedAvg), is susceptible to backdoor attacks. Although researchers have proposed numerous defense algorithms, two significant challenges remain. The attack is becoming more stealthy and harder to detect, and current defense methods are unable to handle 50\\% or more malicious users or assume an auxiliary server dataset.\n To address these challenges, we propose a novel defense algorithm, FL-PLAS, \\textbf{F}ederated \\textbf{L}earning based on \\textbf{P}artial\\textbf{ L}ayer \\textbf{A}ggregation \\textbf{S}trategy. In particular, we divide the local model into a feature extractor and a classifier. In each iteration, the clients only upload the parameters of a feature extractor after local training. The server then aggregates these local parameters and returns the results to the clients.\n Each client retains its own classifier layer, ensuring that the backdoor labels do not impact other clients. We assess the effectiveness of FL-PLAS against state-of-the-art (SOTA) backdoor attacks on three image datasets and compare our approach to six defense strategies. The results of the experiment demonstrate that our methods can effectively protect local models from backdoor attacks. Without requiring any auxiliary dataset for the server, our method achieves a high main-task accuracy with a lower backdoor accuracy even under the condition of 90\\% malicious users with the attacks of trigger, semantic and edge-case.", "date": "2025-05-17", "categories": []} {"id": "2505.12045", "title": "FIGhost: Fluorescent Ink-based Stealthy and Flexible Backdoor Attacks on Physical Traffic Sign Recognition", "abstract": "Traffic sign recognition (TSR) systems are crucial for autonomous driving but are vulnerable to backdoor attacks. Existing physical backdoor attacks either lack stealth, provide inflexible attack control, or ignore emerging Vision-Large-Language-Models (VLMs). In this paper, we introduce FIGhost, the first physical-world backdoor attack leveraging fluorescent ink as triggers. Fluorescent triggers are invisible under normal conditions and activated stealthily by ultraviolet light, providing superior stealthiness, flexibility, and untraceability. Inspired by real-world graffiti, we derive realistic trigger shapes and enhance their robustness via an interpolation-based fluorescence simulation algorithm. Furthermore, we develop an automated backdoor sample generation method to support three attack objectives. Extensive evaluations in the physical world demonstrate FIGhost's effectiveness against state-of-the-art detectors and VLMs, maintaining robustness under environmental variations and effectively evading existing defenses.", "date": "2025-05-17", "categories": []} {"id": "2505.12871", "title": "Does Low Rank Adaptation Lead to Lower Robustness against Training-Time Attacks?", "abstract": "Low rank adaptation (LoRA) has emerged as a prominent technique for fine-tuning large language models (LLMs) thanks to its superb efficiency gains over previous methods. While extensive studies have examined the performance and structural properties of LoRA, its behavior upon training-time attacks remain underexplored, posing significant security risks. In this paper, we theoretically investigate the security implications of LoRA's low-rank structure during fine-tuning, in the context of its robustness against data poisoning and backdoor attacks. We propose an analytical framework that models LoRA's training dynamics, employs the neural tangent kernel to simplify the analysis of the training process, and applies information theory to establish connections between LoRA's low rank structure and its vulnerability against training-time attacks. Our analysis indicates that LoRA exhibits better robustness to backdoor attacks than full fine-tuning, while becomes more vulnerable to untargeted data poisoning due to its over-simplified information geometry. Extensive experimental evaluations have corroborated our theoretical findings.", "date": "2025-05-19", "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": "2505.03501", "title": "BadLingual: A Novel Lingual-Backdoor Attack against Large Language\n Models", "abstract": " In this paper, we present a new form of backdoor attack against Large\nLanguage Models (LLMs): lingual-backdoor attacks. The key novelty of\nlingual-backdoor attacks is that the language itself serves as the trigger to\nhijack the infected LLMs to generate inflammatory speech. They enable the\nprecise targeting of a specific language-speaking group, exacerbating racial\ndiscrimination by malicious entities. We first implement a baseline\nlingual-backdoor attack, which is carried out by poisoning a set of training\ndata for specific downstream tasks through translation into the trigger\nlanguage. However, this baseline attack suffers from poor task generalization\nand is impractical in real-world settings. To address this challenge, we design\nBadLingual, a novel task-agnostic lingual-backdoor, capable of triggering any\ndownstream tasks within the chat LLMs, regardless of the specific questions of\nthese tasks. We design a new approach using PPL-constrained Greedy Coordinate\nGradient-based Search (PGCG) based adversarial training to expand the decision\nboundary of lingual-backdoor, thereby enhancing the generalization ability of\nlingual-backdoor across various tasks. We perform extensive experiments to\nvalidate the effectiveness of our proposed attacks. Specifically, the baseline\nattack achieves an ASR of over 90% on the specified tasks. However, its ASR\nreaches only 37.61% across six tasks in the task-agnostic scenario. In\ncontrast, BadLingual brings up to 37.35% improvement over the baseline. Our\nstudy sheds light on a new perspective of vulnerabilities in LLMs with\nmultilingual capabilities and is expected to promote future research on the\npotential defenses to enhance the LLMs' robustness\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": "2403.00108", "title": "LoRATK: LoRA Once, Backdoor Everywhere in the Share-and-Play Ecosystem", "abstract": " Finetuning LLMs with LoRA has gained significant popularity due to its\nsimplicity and effectiveness. Often, users may even find pluggable,\ncommunity-shared LoRAs to enhance their base models for a specific downstream\ntask of interest; enjoying a powerful, efficient, yet customized LLM experience\nwith negligible investment. However, this convenient share-and-play ecosystem\nalso introduces a new attack surface, where attackers can distribute malicious\nLoRAs to a community eager to try out shared assets. Despite the high-risk\npotential, no prior art has comprehensively explored LoRA's attack surface\nunder the downstream-enhancing share-and-play context. In this paper, we\ninvestigate how backdoors can be injected into task-enhancing LoRAs and examine\nthe mechanisms of such infections. We find that with a simple, efficient, yet\nspecific recipe, a backdoor LoRA can be trained once and then seamlessly merged\n(in a training-free fashion) with multiple task-enhancing LoRAs, retaining both\nits malicious backdoor and benign downstream capabilities. This allows\nattackers to scale the distribution of compromised LoRAs with minimal effort by\nleveraging the rich pool of existing shared LoRA assets. We note that such\nmerged LoRAs are particularly infectious -- because their malicious intent is\ncleverly concealed behind improved downstream capabilities, creating a strong\nincentive for voluntary download -- and dangerous -- because under local\ndeployment, no safety measures exist to intervene when things go wrong. Our\nwork is among the first to study this new threat model of training-free\ndistribution of downstream-capable-yet-backdoor-injected LoRAs, highlighting\nthe urgent need for heightened security awareness in the LoRA ecosystem.\nWarning: This paper contains offensive content and involves a real-life\ntragedy.\n", "date": "2024-02-29", "categories": []} {"id": "2405.20774", "title": "Can We Trust Embodied Agents? Exploring Backdoor Attacks against\n Embodied LLM-based Decision-Making Systems", "abstract": " Large Language Models (LLMs) have shown significant promise in real-world\ndecision-making tasks for embodied artificial intelligence, especially when\nfine-tuned to leverage their inherent common sense and reasoning abilities\nwhile being tailored to specific applications. However, this fine-tuning\nprocess introduces considerable safety and security vulnerabilities, especially\nin safety-critical cyber-physical systems. In this work, we propose the first\ncomprehensive framework for Backdoor Attacks against LLM-based Decision-making\nsystems (BALD) in embodied AI, systematically exploring the attack surfaces and\ntrigger mechanisms. Specifically, we propose three distinct attack mechanisms:\nword injection, scenario manipulation, and knowledge injection, targeting\nvarious components in the LLM-based decision-making pipeline. We perform\nextensive experiments on representative LLMs (GPT-3.5, LLaMA2, PaLM2) in\nautonomous driving and home robot tasks, demonstrating the effectiveness and\nstealthiness of our backdoor triggers across various attack channels, with\ncases like vehicles accelerating toward obstacles and robots placing knives on\nbeds. Our word and knowledge injection attacks achieve nearly 100% success rate\nacross multiple models and datasets while requiring only limited access to the\nsystem. Our scenario manipulation attack yields success rates exceeding 65%,\nreaching up to 90%, and does not require any runtime system intrusion. We also\nassess the robustness of these attacks against defenses, revealing their\nresilience. Our findings highlight critical security vulnerabilities in\nembodied LLM systems and emphasize the urgent need for safeguarding these\nsystems to mitigate potential risks.\n", "date": "2024-05-27", "categories": []} {"id": "2504.18598", "title": "BadMoE: Backdooring Mixture-of-Experts LLMs via Optimizing Routing\n Triggers and Infecting Dormant Experts", "abstract": " Mixture-of-Experts (MoE) have emerged as a powerful architecture for large\nlanguage models (LLMs), enabling efficient scaling of model capacity while\nmaintaining manageable computational costs. The key advantage lies in their\nability to route different tokens to different ``expert'' networks within the\nmodel, enabling specialization and efficient handling of diverse input.\nHowever, the vulnerabilities of MoE-based LLMs still have barely been studied,\nand the potential for backdoor attacks in this context remains largely\nunexplored. This paper presents the first backdoor attack against MoE-based\nLLMs where the attackers poison ``dormant experts'' (i.e., underutilized\nexperts) and activate them by optimizing routing triggers, thereby gaining\ncontrol over the model's output. We first rigorously prove the existence of a\nfew ``dominating experts'' in MoE models, whose outputs can determine the\noverall MoE's output. We also show that dormant experts can serve as dominating\nexperts to manipulate model predictions. Accordingly, our attack, namely\nBadMoE, exploits the unique architecture of MoE models by 1) identifying\ndormant experts unrelated to the target task, 2) constructing a routing-aware\nloss to optimize the activation triggers of these experts, and 3) promoting\ndormant experts to dominating roles via poisoned training data. Extensive\nexperiments show that BadMoE successfully enforces malicious prediction on\nattackers' target tasks while preserving overall model utility, making it a\nmore potent and stealthy attack than existing methods.\n", "date": "2025-04-24", "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": "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": "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": "2504.17300", "title": "The Ultimate Cookbook for Invisible Poison: Crafting Subtle Clean-Label\n Text Backdoors with Style Attributes", "abstract": " Backdoor attacks on text classifiers can cause them to predict a predefined\nlabel when a particular \"trigger\" is present. Prior attacks often rely on\ntriggers that are ungrammatical or otherwise unusual, leading to conspicuous\nattacks. As a result, human annotators, who play a critical role in curating\ntraining data in practice, can easily detect and filter out these unnatural\ntexts during manual inspection, reducing the risk of such attacks. We argue\nthat a key criterion for a successful attack is for text with and without\ntriggers to be indistinguishable to humans. However, prior work neither\ndirectly nor comprehensively evaluated attack subtlety and invisibility with\nhuman involvement. We bridge the gap by conducting thorough human evaluations\nto assess attack subtlety. We also propose \\emph{AttrBkd}, consisting of three\nrecipes for crafting subtle yet effective trigger attributes, such as\nextracting fine-grained attributes from existing baseline backdoor attacks. Our\nhuman evaluations find that AttrBkd with these baseline-derived attributes is\noften more effective (higher attack success rate) and more subtle (fewer\ninstances detected by humans) than the original baseline backdoor attacks,\ndemonstrating that backdoor attacks can bypass detection by being inconspicuous\nand appearing natural even upon close inspection, while still remaining\neffective. Our human annotation also provides information not captured by\nautomated metrics used in prior work, and demonstrates the misalignment of\nthese metrics with human judgment.\n", "date": "2025-04-24", "categories": []} {"id": "2504.12875", "title": "A Client-level Assessment of Collaborative Backdoor Poisoning in Non-IID\n Federated Learning", "abstract": " Federated learning (FL) enables collaborative model training using\ndecentralized private data from multiple clients. While FL has shown robustness\nagainst poisoning attacks with basic defenses, our research reveals new\nvulnerabilities stemming from non-independent and identically distributed\n(non-IID) data among clients. These vulnerabilities pose a substantial risk of\nmodel poisoning in real-world FL scenarios.\n To demonstrate such vulnerabilities, we develop a novel collaborative\nbackdoor poisoning attack called CollaPois. In this attack, we distribute a\nsingle pre-trained model infected with a Trojan to a group of compromised\nclients. These clients then work together to produce malicious gradients,\ncausing the FL model to consistently converge towards a low-loss region\ncentered around the Trojan-infected model. Consequently, the impact of the\nTrojan is amplified, especially when the benign clients have diverse local data\ndistributions and scattered local gradients. CollaPois stands out by achieving\nits goals while involving only a limited number of compromised clients, setting\nit apart from existing attacks. Also, CollaPois effectively avoids noticeable\nshifts or degradation in the FL model's performance on legitimate data samples,\nallowing it to operate stealthily and evade detection by advanced robust FL\nalgorithms.\n Thorough theoretical analysis and experiments conducted on various benchmark\ndatasets demonstrate the superiority of CollaPois compared to state-of-the-art\nbackdoor attacks. Notably, CollaPois bypasses existing backdoor defenses,\nespecially in scenarios where clients possess diverse data distributions.\nMoreover, the results show that CollaPois remains effective even when involving\na small number of compromised clients. Notably, clients whose local data is\nclosely aligned with compromised clients experience higher risks of backdoor\ninfections.\n", "date": "2025-04-17", "categories": []} {"id": "2504.15674", "title": "TrojanDam: Detection-Free Backdoor Defense in Federated Learning through\n Proactive Model Robustification utilizing OOD Data", "abstract": " Federated learning (FL) systems allow decentralized data-owning clients to\njointly train a global model through uploading their locally trained updates to\na centralized server. The property of decentralization enables adversaries to\ncraft carefully designed backdoor updates to make the global model misclassify\nonly when encountering adversary-chosen triggers. Existing defense mechanisms\nmainly rely on post-training detection after receiving updates. These methods\neither fail to identify updates which are deliberately fabricated statistically\nclose to benign ones, or show inconsistent performance in different FL training\nstages. The effect of unfiltered backdoor updates will accumulate in the global\nmodel, and eventually become functional. Given the difficulty of ruling out\nevery backdoor update, we propose a backdoor defense paradigm, which focuses on\nproactive robustification on the global model against potential backdoor\nattacks. We first reveal that the successful launching of backdoor attacks in\nFL stems from the lack of conflict between malicious and benign updates on\nredundant neurons of ML models. We proceed to prove the feasibility of\nactivating redundant neurons utilizing out-of-distribution (OOD) samples in\ncentralized settings, and migrating to FL settings to propose a novel backdoor\ndefense mechanism, TrojanDam. The proposed mechanism has the FL server\ncontinuously inject fresh OOD mappings into the global model to activate\nredundant neurons, canceling the effect of backdoor updates during aggregation.\nWe conduct systematic and extensive experiments to illustrate the superior\nperformance of TrojanDam, over several SOTA backdoor defense methods across a\nwide range of FL settings.\n", "date": "2025-04-22", "categories": []} {"id": "2504.13775", "title": "BadApex: Backdoor Attack Based on Adaptive Optimization Mechanism of\n Black-box Large Language Models", "abstract": " Previous insertion-based and paraphrase-based backdoors have achieved great\nsuccess in attack efficacy, but they ignore the text quality and semantic\nconsistency between poisoned and clean texts. Although recent studies introduce\nLLMs to generate poisoned texts and improve the stealthiness, semantic\nconsistency, and text quality, their hand-crafted prompts rely on expert\nexperiences, facing significant challenges in prompt adaptability and attack\nperformance after defenses. In this paper, we propose a novel backdoor attack\nbased on adaptive optimization mechanism of black-box large language models\n(BadApex), which leverages a black-box LLM to generate poisoned text through a\nrefined prompt. Specifically, an Adaptive Optimization Mechanism is designed to\nrefine an initial prompt iteratively using the generation and modification\nagents. The generation agent generates the poisoned text based on the initial\nprompt. Then the modification agent evaluates the quality of the poisoned text\nand refines a new prompt. After several iterations of the above process, the\nrefined prompt is used to generate poisoned texts through LLMs. We conduct\nextensive experiments on three dataset with six backdoor attacks and two\ndefenses. Extensive experimental results demonstrate that BadApex significantly\noutperforms state-of-the-art attacks. It improves prompt adaptability, semantic\nconsistency, and text quality. Furthermore, when two defense methods are\napplied, the average attack success rate (ASR) still up to 96.75%.\n", "date": "2025-04-18", "categories": []} {"id": "2406.05826", "title": "PSBD: Prediction Shift Uncertainty Unlocks Backdoor Detection", "abstract": " Deep neural networks are susceptible to backdoor attacks, where adversaries\nmanipulate model predictions by inserting malicious samples into the training\ndata. Currently, there is still a significant challenge in identifying\nsuspicious training data to unveil potential backdoor samples. In this paper,\nwe propose a novel method, Prediction Shift Backdoor Detection (PSBD),\nleveraging an uncertainty-based approach requiring minimal unlabeled clean\nvalidation data. PSBD is motivated by an intriguing Prediction Shift (PS)\nphenomenon, where poisoned models' predictions on clean data often shift away\nfrom true labels towards certain other labels with dropout applied during\ninference, while backdoor samples exhibit less PS. We hypothesize PS results\nfrom the neuron bias effect, making neurons favor features of certain classes.\nPSBD identifies backdoor training samples by computing the Prediction Shift\nUncertainty (PSU), the variance in probability values when dropout layers are\ntoggled on and off during model inference. Extensive experiments have been\nconducted to verify the effectiveness and efficiency of PSBD, which achieves\nstate-of-the-art results among mainstream detection methods. The code is\navailable at https://github.com/WL-619/PSBD.\n", "date": "2024-06-09", "categories": []} {"id": "2504.11990", "title": "Secure Transfer Learning: Training Clean Models Against Backdoor in\n (Both) Pre-trained Encoders and Downstream Datasets", "abstract": " Transfer learning from pre-trained encoders has become essential in modern\nmachine learning, enabling efficient model adaptation across diverse tasks.\nHowever, this combination of pre-training and downstream adaptation creates an\nexpanded attack surface, exposing models to sophisticated backdoor embeddings\nat both the encoder and dataset levels--an area often overlooked in prior\nresearch. Additionally, the limited computational resources typically available\nto users of pre-trained encoders constrain the effectiveness of generic\nbackdoor defenses compared to end-to-end training from scratch. In this work,\nwe investigate how to mitigate potential backdoor risks in resource-constrained\ntransfer learning scenarios. Specifically, we conduct an exhaustive analysis of\nexisting defense strategies, revealing that many follow a reactive workflow\nbased on assumptions that do not scale to unknown threats, novel attack types,\nor different training paradigms. In response, we introduce a proactive mindset\nfocused on identifying clean elements and propose the Trusted Core (T-Core)\nBootstrapping framework, which emphasizes the importance of pinpointing\ntrustworthy data and neurons to enhance model security. Our empirical\nevaluations demonstrate the effectiveness and superiority of T-Core,\nspecifically assessing 5 encoder poisoning attacks, 7 dataset poisoning\nattacks, and 14 baseline defenses across five benchmark datasets, addressing\nfour scenarios of 3 potential backdoor threats.\n", "date": "2025-04-16", "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": "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": "2411.09540", "title": "Prompting the Unseen: Detecting Hidden Backdoors in Black-Box Models", "abstract": " Visual prompting (VP) is a new technique that adapts well-trained frozen\nmodels for source domain tasks to target domain tasks. This study examines VP's\nbenefits for black-box model-level backdoor detection. The visual prompt in VP\nmaps class subspaces between source and target domains. We identify a\nmisalignment, termed class subspace inconsistency, between clean and poisoned\ndatasets. Based on this, we introduce \\textsc{BProm}, a black-box model-level\ndetection method to identify backdoors in suspicious models, if any.\n\\textsc{BProm} leverages the low classification accuracy of prompted models\nwhen backdoors are present. Extensive experiments confirm \\textsc{BProm}'s\neffectiveness.\n", "date": "2024-11-14", "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": "2503.16872", "title": "Lie Detector: Unified Backdoor Detection via Cross-Examination Framework", "abstract": " Institutions with limited data and computing resources often outsource model\ntraining to third-party providers in a semi-honest setting, assuming adherence\nto prescribed training protocols with pre-defined learning paradigm (e.g.,\nsupervised or semi-supervised learning). However, this practice can introduce\nsevere security risks, as adversaries may poison the training data to embed\nbackdoors into the resulting model. Existing detection approaches predominantly\nrely on statistical analyses, which often fail to maintain universally accurate\ndetection accuracy across different learning paradigms. To address this\nchallenge, we propose a unified backdoor detection framework in the semi-honest\nsetting that exploits cross-examination of model inconsistencies between two\nindependent service providers. Specifically, we integrate central kernel\nalignment to enable robust feature similarity measurements across different\nmodel architectures and learning paradigms, thereby facilitating precise\nrecovery and identification of backdoor triggers. We further introduce backdoor\nfine-tuning sensitivity analysis to distinguish backdoor triggers from\nadversarial perturbations, substantially reducing false positives. Extensive\nexperiments demonstrate that our method achieves superior detection\nperformance, improving accuracy by 5.4%, 1.6%, and 11.9% over SoTA baselines\nacross supervised, semi-supervised, and autoregressive learning tasks,\nrespectively. Notably, it is the first to effectively detect backdoors in\nmultimodal large language models, further highlighting its broad applicability\nand advancing secure deep learning.\n", "date": "2025-03-21", "categories": []} {"id": "2412.19037", "title": "CL-Attack: Textual Backdoor Attacks via Cross-Lingual Triggers", "abstract": " Backdoor attacks significantly compromise the security of large language\nmodels by triggering them to output specific and controlled content. Currently,\ntriggers for textual backdoor attacks fall into two categories: fixed-token\ntriggers and sentence-pattern triggers. However, the former are typically easy\nto identify and filter, while the latter, such as syntax and style, do not\napply to all original samples and may lead to semantic shifts. In this paper,\ninspired by cross-lingual (CL) prompts of LLMs in real-world scenarios, we\npropose a higher-dimensional trigger method at the paragraph level, namely\nCL-attack. CL-attack injects the backdoor by using texts with specific\nstructures that incorporate multiple languages, thereby offering greater\nstealthiness and universality compared to existing backdoor attack techniques.\nExtensive experiments on different tasks and model architectures demonstrate\nthat CL-attack can achieve nearly 100% attack success rate with a low poisoning\nrate in both classification and generation tasks. We also empirically show that\nthe CL-attack is more robust against current major defense methods compared to\nbaseline backdoor attacks. Additionally, to mitigate CL-attack, we further\ndevelop a new defense called TranslateDefense, which can partially mitigate the\nimpact of CL-attack.\n", "date": "2024-12-26", "categories": []} {"id": "2503.20925", "title": "Prototype Guided Backdoor Defense", "abstract": " Deep learning models are susceptible to {\\em backdoor attacks} involving\nmalicious attackers perturbing a small subset of training data with a {\\em\ntrigger} to causes misclassifications. Various triggers have been used,\nincluding semantic triggers that are easily realizable without requiring the\nattacker to manipulate the image. The emergence of generative AI has eased the\ngeneration of varied poisoned samples. Robustness across types of triggers is\ncrucial to effective defense. We propose Prototype Guided Backdoor Defense\n(PGBD), a robust post-hoc defense that scales across different trigger types,\nincluding previously unsolved semantic triggers. PGBD exploits displacements in\nthe geometric spaces of activations to penalize movements toward the trigger.\nThis is done using a novel sanitization loss of a post-hoc fine-tuning step.\nThe geometric approach scales easily to all types of attacks. PGBD achieves\nbetter performance across all settings. We also present the first defense\nagainst a new semantic attack on celebrity face images. Project page:\n\\hyperlink{https://venkatadithya9.github.io/pgbd.github.io/}{this https URL}.\n", "date": "2025-03-26", "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": "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": "2501.02373", "title": "BADTV: Unveiling Backdoor Threats in Third-Party Task Vectors", "abstract": " Task arithmetic in large-scale pre-trained models enables agile adaptation to\ndiverse downstream tasks without extensive retraining. By leveraging task\nvectors (TVs), users can perform modular updates through simple arithmetic\noperations like addition and subtraction. Yet, this flexibility presents new\nsecurity challenges. In this paper, we investigate how TVs are vulnerable to\nbackdoor attacks, revealing how malicious actors can exploit them to compromise\nmodel integrity. By creating composite backdoors that are designed\nasymmetrically, we introduce BadTV, a backdoor attack specifically crafted to\nremain effective simultaneously under task learning, forgetting, and analogy\noperations. Extensive experiments show that BadTV achieves near-perfect attack\nsuccess rates across diverse scenarios, posing a serious threat to models\nrelying on task arithmetic. We also evaluate current defenses, finding they\nfail to detect or mitigate BadTV. Our results highlight the urgent need for\nrobust countermeasures to secure TVs in real-world deployments.\n", "date": "2025-01-04", "categories": []} {"id": "2411.16154", "title": "DeDe: Detecting Backdoor Samples for SSL Encoders via Decoders", "abstract": " Self-supervised learning (SSL) is pervasively exploited in training\nhigh-quality upstream encoders with a large amount of unlabeled data. However,\nit is found to be susceptible to backdoor attacks merely via polluting a small\nportion of training data. The victim encoders associate triggered inputs with\ntarget embeddings, e.g., mapping a triggered cat image to an airplane\nembedding, such that the downstream tasks inherit unintended behaviors when the\ntrigger is activated. Emerging backdoor attacks have shown great threats across\ndifferent SSL paradigms such as contrastive learning and CLIP, yet limited\nresearch is devoted to defending against such attacks, and existing defenses\nfall short in detecting advanced stealthy backdoors. To address the\nlimitations, we propose a novel detection mechanism, DeDe, which detects the\nactivation of backdoor mappings caused by triggered inputs on victim encoders.\nSpecifically, DeDe trains a decoder for any given SSL encoder using an\nauxiliary dataset (which can be out-of-distribution or even slightly poisoned),\nso that for any triggered input that misleads the encoder into the target\nembedding, the decoder generates an output image significantly different from\nthe input. DeDe leverages the discrepancy between the input and the decoded\noutput to identify potential backdoor misbehavior during inference. We\nempirically evaluate DeDe on both contrastive learning and CLIP models against\nvarious types of backdoor attacks. Our results demonstrate promising detection\neffectiveness over various advanced attacks and superior performance compared\nover state-of-the-art detection methods.\n", "date": "2024-11-25", "categories": []} {"id": "2503.15293", "title": "Test-Time Backdoor Detection for Object Detection Models", "abstract": " Object detection models are vulnerable to backdoor attacks, where attackers\npoison a small subset of training samples by embedding a predefined trigger to\nmanipulate prediction. Detecting poisoned samples (i.e., those containing\ntriggers) at test time can prevent backdoor activation. However, unlike image\nclassification tasks, the unique characteristics of object detection --\nparticularly its output of numerous objects -- pose fresh challenges for\nbackdoor detection. The complex attack effects (e.g., \"ghost\" object emergence\nor \"vanishing\" object) further render current defenses fundamentally\ninadequate. To this end, we design TRAnsformation Consistency Evaluation\n(TRACE), a brand-new method for detecting poisoned samples at test time in\nobject detection. Our journey begins with two intriguing observations: (1)\npoisoned samples exhibit significantly more consistent detection results than\nclean ones across varied backgrounds. (2) clean samples show higher detection\nconsistency when introduced to different focal information. Based on these\nphenomena, TRACE applies foreground and background transformations to each test\nsample, then assesses transformation consistency by calculating the variance in\nobjects confidences. TRACE achieves black-box, universal backdoor detection,\nwith extensive experiments showing a 30% improvement in AUROC over\nstate-of-the-art defenses and resistance to adaptive attacks.\n", "date": "2025-03-19", "categories": []} {"id": "2404.19597", "title": "TuBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with\n Instruction Tuning", "abstract": " The implications of backdoor attacks on English-centric large language models\n(LLMs) have been widely examined - such attacks can be achieved by embedding\nmalicious behaviors during training and activated under specific conditions\nthat trigger malicious outputs. Despite the increasing support for multilingual\ncapabilities in open-source and proprietary LLMs, the impact of backdoor\nattacks on these systems remains largely under-explored. Our research focuses\non cross-lingual backdoor attacks against multilingual LLMs, particularly\ninvestigating how poisoning the instruction-tuning data for one or two\nlanguages can affect the outputs for languages whose instruction-tuning data\nwere not poisoned. Despite its simplicity, our empirical analysis reveals that\nour method exhibits remarkable efficacy in models like mT5 and GPT-4o, with\nhigh attack success rates, surpassing 90% in more than 7 out of 12 languages\nacross various scenarios. Our findings also indicate that more powerful models\nshow increased susceptibility to transferable cross-lingual backdoor attacks,\nwhich also applies to LLMs predominantly pre-trained on English data, such as\nLlama2, Llama3, and Gemma. Moreover, our experiments demonstrate 1) High\nTransferability: the backdoor mechanism operates successfully in cross-lingual\nresponse scenarios across 26 languages, achieving an average attack success\nrate of 99%, and 2) Robustness: the proposed attack remains effective even\nafter defenses are applied. These findings expose critical security\nvulnerabilities in multilingual LLMs and highlight the urgent need for more\nrobust, targeted defense strategies to address the unique challenges posed by\ncross-lingual backdoor transfer.\n", "date": "2024-04-30", "categories": []} {"id": "2406.09836", "title": "Robustness Inspired Graph Backdoor Defense", "abstract": " Graph Neural Networks (GNNs) have achieved promising results in tasks such as\nnode classification and graph classification. However, recent studies reveal\nthat GNNs are vulnerable to backdoor attacks, posing a significant threat to\ntheir real-world adoption. Despite initial efforts to defend against specific\ngraph backdoor attacks, there is no work on defending against various types of\nbackdoor attacks where generated triggers have different properties. Hence, we\nfirst empirically verify that prediction variance under edge dropping is a\ncrucial indicator for identifying poisoned nodes. With this observation, we\npropose using random edge dropping to detect backdoors and theoretically show\nthat it can efficiently distinguish poisoned nodes from clean ones.\nFurthermore, we introduce a novel robust training strategy to efficiently\ncounteract the impact of the triggers. Extensive experiments on real-world\ndatasets show that our framework can effectively identify poisoned nodes,\nsignificantly degrade the attack success rate, and maintain clean accuracy when\ndefending against various types of graph backdoor attacks with different\nproperties.\n", "date": "2024-06-14", "categories": []} {"id": "2402.02034", "title": "CEPA: Consensus Embedded Perturbation for Agnostic Detection and\n Inversion of Backdoors", "abstract": " A variety of defenses have been proposed against Trojans planted in (backdoor\nattacks on) deep neural network (DNN) classifiers. Backdoor-agnostic methods\nseek to reliably detect and/or to mitigate backdoors irrespective of the\nincorporation mechanism used by the attacker, while inversion methods\nexplicitly assume one. In this paper, we describe a new detector that: relies\non embedded feature representations to estimate (invert) the backdoor and to\nidentify its target class; can operate without access to the training dataset;\nand is highly effective for various incorporation mechanisms (i.e., is backdoor\nagnostic). Our detection approach is evaluated -- and found to be favorable -\nin comparison with an array of published defenses for a variety of different\nattacks on the CIFAR-10 and CIFAR-100 image-classification domains.\n", "date": "2024-02-03", "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": "2402.18945", "title": "SynGhost: Invisible and Universal Task-agnostic Backdoor Attack via\n Syntactic Transfer", "abstract": " Although pre-training achieves remarkable performance, it suffers from\ntask-agnostic backdoor attacks due to vulnerabilities in data and training\nmechanisms. These attacks can transfer backdoors to various downstream tasks.\nIn this paper, we introduce $\\mathtt{maxEntropy}$, an entropy-based poisoning\nfilter that mitigates such risks. To overcome the limitations of manual target\nsetting and explicit triggers, we propose $\\mathtt{SynGhost}$, an invisible and\nuniversal task-agnostic backdoor attack via syntactic transfer, further\nexposing vulnerabilities in pre-trained language models (PLMs). Specifically,\n$\\mathtt{SynGhost}$ injects multiple syntactic backdoors into the pre-training\nspace through corpus poisoning, while preserving the PLM's pre-training\ncapabilities. Second, $\\mathtt{SynGhost}$ adaptively selects optimal targets\nbased on contrastive learning, creating a uniform distribution in the\npre-training space. To identify syntactic differences, we also introduce an\nawareness module to minimize interference between backdoors. Experiments show\nthat $\\mathtt{SynGhost}$ poses significant threats and can transfer to various\ndownstream tasks. Furthermore, $\\mathtt{SynGhost}$ resists defenses based on\nperplexity, fine-pruning, and $\\mathtt{maxEntropy}$. The code is available at\nhttps://github.com/Zhou-CyberSecurity-AI/SynGhost.\n", "date": "2024-02-29", "categories": []} {"id": "2410.14723", "title": "BeniFul: Backdoor Defense via Middle Feature Analysis for Deep Neural\n Networks", "abstract": " Backdoor defenses have recently become important in resisting backdoor\nattacks in deep neural networks (DNNs), where attackers implant backdoors into\nthe DNN model by injecting backdoor samples into the training dataset. Although\nthere are many defense methods to achieve backdoor detection for DNN inputs and\nbackdoor elimination for DNN models, they still have not presented a clear\nexplanation of the relationship between these two missions. In this paper, we\nuse the features from the middle layer of the DNN model to analyze the\ndifference between backdoor and benign samples and propose Backdoor\nConsistency, which indicates that at least one backdoor exists in the DNN model\nif the backdoor trigger is detected exactly on input. By analyzing the middle\nfeatures, we design an effective and comprehensive backdoor defense method\nnamed BeniFul, which consists of two parts: a gray-box backdoor input detection\nand a white-box backdoor elimination. Specifically, we use the reconstruction\ndistance from the Variational Auto-Encoder and model inference results to\nimplement backdoor input detection and a feature distance loss to achieve\nbackdoor elimination. Experimental results on CIFAR-10 and Tiny ImageNet\nagainst five state-of-the-art attacks demonstrate that our BeniFul exhibits a\ngreat defense capability in backdoor input detection and backdoor elimination.\n", "date": "2024-10-15", "categories": []} {"id": "2410.13910", "title": "Mitigating the Backdoor Effect for Multi-Task Model Merging via\n Safety-Aware Subspace", "abstract": " Model merging has gained significant attention as a cost-effective approach\nto integrate multiple single-task fine-tuned models into a unified one that can\nperform well on multiple tasks. However, existing model merging techniques\nprimarily focus on resolving conflicts between task-specific models, they often\noverlook potential security threats, particularly the risk of backdoor attacks\nin the open-source model ecosystem. In this paper, we first investigate the\nvulnerabilities of existing model merging methods to backdoor attacks,\nidentifying two critical challenges: backdoor succession and backdoor transfer.\nTo address these issues, we propose a novel Defense-Aware Merging (DAM)\napproach that simultaneously mitigates task interference and backdoor\nvulnerabilities. Specifically, DAM employs a meta-learning-based optimization\nmethod with dual masks to identify a shared and safety-aware subspace for model\nmerging. These masks are alternately optimized: the Task-Shared mask identifies\ncommon beneficial parameters across tasks, aiming to preserve task-specific\nknowledge while reducing interference, while the Backdoor-Detection mask\nisolates potentially harmful parameters to neutralize security threats. This\ndual-mask design allows us to carefully balance the preservation of useful\nknowledge and the removal of potential vulnerabilities. Compared to existing\nmerging methods, DAM achieves a more favorable balance between performance and\nsecurity, reducing the attack success rate by 2-10 percentage points while\nsacrificing only about 1% in accuracy. Furthermore, DAM exhibits robust\nperformance and broad applicability across various types of backdoor attacks\nand the number of compromised models involved in the merging process. Our codes\nand models are available at https://github.com/Yangjinluan/DAM.\n", "date": "2024-10-17", "categories": []} {"id": "2502.18508", "title": "REFINE: Inversion-Free Backdoor Defense via Model Reprogramming", "abstract": " Backdoor attacks on deep neural networks (DNNs) have emerged as a significant\nsecurity threat, allowing adversaries to implant hidden malicious behaviors\nduring the model training phase. Pre-processing-based defense, which is one of\nthe most important defense paradigms, typically focuses on input\ntransformations or backdoor trigger inversion (BTI) to deactivate or eliminate\nembedded backdoor triggers during the inference process. However, these methods\nsuffer from inherent limitations: transformation-based defenses often fail to\nbalance model utility and defense performance, while BTI-based defenses\nstruggle to accurately reconstruct trigger patterns without prior knowledge. In\nthis paper, we propose REFINE, an inversion-free backdoor defense method based\non model reprogramming. REFINE consists of two key components: \\textbf{(1)} an\ninput transformation module that disrupts both benign and backdoor patterns,\ngenerating new benign features; and \\textbf{(2)} an output remapping module\nthat redefines the model's output domain to guide the input transformations\neffectively. By further integrating supervised contrastive loss, REFINE\nenhances the defense capabilities while maintaining model utility. Extensive\nexperiments on various benchmark datasets demonstrate the effectiveness of our\nREFINE and its resistance to potential adaptive attacks.\n", "date": "2025-02-22", "categories": []} {"id": "2502.18511", "title": "ELBA-Bench: An Efficient Learning Backdoor Attacks Benchmark for Large\n Language Models", "abstract": " Generative large language models are crucial in natural language processing,\nbut they are vulnerable to backdoor attacks, where subtle triggers compromise\ntheir behavior. Although backdoor attacks against LLMs are constantly emerging,\nexisting benchmarks remain limited in terms of sufficient coverage of attack,\nmetric system integrity, backdoor attack alignment. And existing pre-trained\nbackdoor attacks are idealized in practice due to resource access constraints.\nTherefore we establish $\\textit{ELBA-Bench}$, a comprehensive and unified\nframework that allows attackers to inject backdoor through parameter efficient\nfine-tuning ($\\textit{e.g.,}$ LoRA) or without fine-tuning techniques\n($\\textit{e.g.,}$ In-context-learning). $\\textit{ELBA-Bench}$ provides over\n1300 experiments encompassing the implementations of 12 attack methods, 18\ndatasets, and 12 LLMs. Extensive experiments provide new invaluable findings\ninto the strengths and limitations of various attack strategies. For instance,\nPEFT attack consistently outperform without fine-tuning approaches in\nclassification tasks while showing strong cross-dataset generalization with\noptimized triggers boosting robustness; Task-relevant backdoor optimization\ntechniques or attack prompts along with clean and adversarial demonstrations\ncan enhance backdoor attack success while preserving model performance on clean\nsamples. Additionally, we introduce a universal toolbox designed for\nstandardized backdoor attack research, with the goal of propelling further\nprogress in this vital area.\n", "date": "2025-02-22", "categories": []} {"id": "2502.18520", "title": "Class-Conditional Neural Polarizer: A Lightweight and Effective Backdoor\n Defense by Purifying Poisoned Features", "abstract": " Recent studies have highlighted the vulnerability of deep neural networks to\nbackdoor attacks, where models are manipulated to rely on embedded triggers\nwithin poisoned samples, despite the presence of both benign and trigger\ninformation. While several defense methods have been proposed, they often\nstruggle to balance backdoor mitigation with maintaining benign performance.In\nthis work, inspired by the concept of optical polarizer-which allows light\nwaves of specific polarizations to pass while filtering others-we propose a\nlightweight backdoor defense approach, NPD. This method integrates a neural\npolarizer (NP) as an intermediate layer within the compromised model,\nimplemented as a lightweight linear transformation optimized via bi-level\noptimization. The learnable NP filters trigger information from poisoned\nsamples while preserving benign content. Despite its effectiveness, we identify\nthrough empirical studies that NPD's performance degrades when the target\nlabels (required for purification) are inaccurately estimated. To address this\nlimitation while harnessing the potential of targeted adversarial mitigation,\nwe propose class-conditional neural polarizer-based defense (CNPD). The key\ninnovation is a fusion module that integrates the backdoored model's predicted\nlabel with the features to be purified. This architecture inherently mimics\ntargeted adversarial defense mechanisms without requiring label estimation used\nin NPD. We propose three implementations of CNPD: the first is r-CNPD, which\ntrains a replicated NP layer for each class and, during inference, selects the\nappropriate NP layer for defense based on the predicted class from the\nbackdoored model. To efficiently handle a large number of classes, two variants\nare designed: e-CNPD, which embeds class information as additional features,\nand a-CNPD, which directs network attention using class information.\n", "date": "2025-02-23", "categories": []} {"id": "2502.15830", "title": "Show Me Your Code! Kill Code Poisoning: A Lightweight Method Based on\n Code Naturalness", "abstract": " Neural code models (NCMs) have demonstrated extraordinary capabilities in\ncode intelligence tasks. Meanwhile, the security of NCMs and NCMs-based systems\nhas garnered increasing attention. In particular, NCMs are often trained on\nlarge-scale data from potentially untrustworthy sources, providing attackers\nwith the opportunity to manipulate them by inserting crafted samples into the\ndata. This type of attack is called a code poisoning attack (also known as a\nbackdoor attack). It allows attackers to implant backdoors in NCMs and thus\ncontrol model behavior, which poses a significant security threat. However,\nthere is still a lack of effective techniques for detecting various complex\ncode poisoning attacks.\n In this paper, we propose an innovative and lightweight technique for code\npoisoning detection named KillBadCode. KillBadCode is designed based on our\ninsight that code poisoning disrupts the naturalness of code. Specifically,\nKillBadCode first builds a code language model (CodeLM) on a lightweight\n$n$-gram language model. Then, given poisoned data, KillBadCode utilizes CodeLM\nto identify those tokens in (poisoned) code snippets that will make the code\nsnippets more natural after being deleted as trigger tokens. Considering that\nthe removal of some normal tokens in a single sample might also enhance code\nnaturalness, leading to a high false positive rate (FPR), we aggregate the\ncumulative improvement of each token across all samples. Finally, KillBadCode\npurifies the poisoned data by removing all poisoned samples containing the\nidentified trigger tokens. The experimental results on two code poisoning\nattacks and four code intelligence tasks demonstrate that KillBadCode\nsignificantly outperforms four baselines. More importantly, KillBadCode is very\nefficient, with a minimum time consumption of only 5 minutes, and is 25 times\nfaster than the best baseline on average.\n", "date": "2025-02-20", "categories": []} {"id": "2412.07277", "title": "Backdoor Attacks against No-Reference Image Quality Assessment Models\n via a Scalable Trigger", "abstract": " No-Reference Image Quality Assessment (NR-IQA), responsible for assessing the\nquality of a single input image without using any reference, plays a critical\nrole in evaluating and optimizing computer vision systems, e.g., low-light\nenhancement. Recent research indicates that NR-IQA models are susceptible to\nadversarial attacks, which can significantly alter predicted scores with\nvisually imperceptible perturbations. Despite revealing vulnerabilities, these\nattack methods have limitations, including high computational demands,\nuntargeted manipulation, limited practical utility in white-box scenarios, and\nreduced effectiveness in black-box scenarios. To address these challenges, we\nshift our focus to another significant threat and present a novel\npoisoning-based backdoor attack against NR-IQA (BAIQA), allowing the attacker\nto manipulate the IQA model's output to any desired target value by simply\nadjusting a scaling coefficient $\\alpha$ for the trigger. We propose to inject\nthe trigger in the discrete cosine transform (DCT) domain to improve the local\ninvariance of the trigger for countering trigger diminishment in NR-IQA models\ndue to widely adopted data augmentations. Furthermore, the universal\nadversarial perturbations (UAP) in the DCT space are designed as the trigger,\nto increase IQA model susceptibility to manipulation and improve attack\neffectiveness. In addition to the heuristic method for poison-label BAIQA\n(P-BAIQA), we explore the design of clean-label BAIQA (C-BAIQA), focusing on\n$\\alpha$ sampling and image data refinement, driven by theoretical insights we\nreveal. Extensive experiments on diverse datasets and various NR-IQA models\ndemonstrate the effectiveness of our attacks. Code can be found at\nhttps://github.com/yuyi-sd/BAIQA.\n", "date": "2024-12-10", "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": "2411.12701", "title": "When Backdoors Speak: Understanding LLM Backdoor Attacks Through\n Model-Generated Explanations", "abstract": " Large Language Models (LLMs) are known to be vulnerable to backdoor attacks,\nwhere triggers embedded in poisoned samples can maliciously alter LLMs'\nbehaviors. In this paper, we move beyond attacking LLMs and instead examine\nbackdoor attacks through the novel lens of natural language explanations.\nSpecifically, we leverage LLMs' generative capabilities to produce\nhuman-readable explanations for their decisions, enabling direct comparisons\nbetween explanations for clean and poisoned samples. Our results show that\nbackdoored models produce coherent explanations for clean inputs but diverse\nand logically flawed explanations for poisoned data, a pattern consistent\nacross classification and generation tasks for different backdoor attacks.\nFurther analysis reveals key insights into the explanation generation process.\nAt the token level, explanation tokens associated with poisoned samples only\nappear in the final few transformer layers. At the sentence level, attention\ndynamics indicate that poisoned inputs shift attention away from the original\ninput context during explanation generation. These findings enhance our\nunderstanding of backdoor mechanisms in LLMs and present a promising framework\nfor detecting vulnerabilities through explainability.\n", "date": "2024-11-19", "categories": []} {"id": "2502.11687", "title": "ReVeil: Unconstrained Concealed Backdoor Attack on Deep Neural Networks\n using Machine Unlearning", "abstract": " Backdoor attacks embed hidden functionalities in deep neural networks (DNN),\ntriggering malicious behavior with specific inputs. Advanced defenses monitor\nanomalous DNN inferences to detect such attacks. However, concealed backdoors\nevade detection by maintaining a low pre-deployment attack success rate (ASR)\nand restoring high ASR post-deployment via machine unlearning. Existing\nconcealed backdoors are often constrained by requiring white-box or black-box\naccess or auxiliary data, limiting their practicality when such access or data\nis unavailable. This paper introduces ReVeil, a concealed backdoor attack\ntargeting the data collection phase of the DNN training pipeline, requiring no\nmodel access or auxiliary data. ReVeil maintains low pre-deployment ASR across\nfour datasets and four trigger patterns, successfully evades three popular\nbackdoor detection methods, and restores high ASR post-deployment through\nmachine unlearning.\n", "date": "2025-02-17", "categories": []} {"id": "2502.01385", "title": "Detecting Backdoor Samples in Contrastive Language Image Pretraining", "abstract": " Contrastive language-image pretraining (CLIP) has been found to be vulnerable\nto poisoning backdoor attacks where the adversary can achieve an almost perfect\nattack success rate on CLIP models by poisoning only 0.01\\% of the training\ndataset. This raises security concerns on the current practice of pretraining\nlarge-scale models on unscrutinized web data using CLIP. In this work, we\nanalyze the representations of backdoor-poisoned samples learned by CLIP models\nand find that they exhibit unique characteristics in their local subspace,\ni.e., their local neighborhoods are far more sparse than that of clean samples.\nBased on this finding, we conduct a systematic study on detecting CLIP backdoor\nattacks and show that these attacks can be easily and efficiently detected by\ntraditional density ratio-based local outlier detectors, whereas existing\nbackdoor sample detection methods fail. Our experiments also reveal that an\nunintentional backdoor already exists in the original CC3M dataset and has been\ntrained into a popular open-source model released by OpenCLIP. Based on our\ndetector, one can clean up a million-scale web dataset (e.g., CC3M) efficiently\nwithin 15 minutes using 4 Nvidia A100 GPUs. The code is publicly available in\nour \\href{https://github.com/HanxunH/Detect-CLIP-Backdoor-Samples}{GitHub\nrepository}.\n", "date": "2025-02-03", "categories": []} {"id": "2502.05224", "title": "A Survey on Backdoor Threats in Large Language Models (LLMs): Attacks,\n Defenses, and Evaluations", "abstract": " Large Language Models (LLMs) have achieved significantly advanced\ncapabilities in understanding and generating human language text, which have\ngained increasing popularity over recent years. Apart from their\nstate-of-the-art natural language processing (NLP) performance, considering\ntheir widespread usage in many industries, including medicine, finance,\neducation, etc., security concerns over their usage grow simultaneously. In\nrecent years, the evolution of backdoor attacks has progressed with the\nadvancement of defense mechanisms against them and more well-developed features\nin the LLMs. In this paper, we adapt the general taxonomy for classifying\nmachine learning attacks on one of the subdivisions - training-time white-box\nbackdoor attacks. Besides systematically classifying attack methods, we also\nconsider the corresponding defense methods against backdoor attacks. By\nproviding an extensive summary of existing works, we hope this survey can serve\nas a guideline for inspiring future research that further extends the attack\nscenarios and creates a stronger defense against them for more robust LLMs.\n", "date": "2025-02-06", "categories": []} {"id": "2502.05755", "title": "Filter, Obstruct and Dilute: Defending Against Backdoor Attacks on\n Semi-Supervised Learning", "abstract": " Recent studies have verified that semi-supervised learning (SSL) is\nvulnerable to data poisoning backdoor attacks. Even a tiny fraction of\ncontaminated training data is sufficient for adversaries to manipulate up to\n90\\% of the test outputs in existing SSL methods. Given the emerging threat of\nbackdoor attacks designed for SSL, this work aims to protect SSL against such\nrisks, marking it as one of the few known efforts in this area. Specifically,\nwe begin by identifying that the spurious correlations between the backdoor\ntriggers and the target class implanted by adversaries are the primary cause of\nmanipulated model predictions during the test phase. To disrupt these\ncorrelations, we utilize three key techniques: Gaussian Filter, complementary\nlearning and trigger mix-up, which collectively filter, obstruct and dilute the\ninfluence of backdoor attacks in both data pre-processing and feature learning.\nExperimental results demonstrate that our proposed method, Backdoor Invalidator\n(BI), significantly reduces the average attack success rate from 84.7\\% to\n1.8\\% across different state-of-the-art backdoor attacks. It is also worth\nmentioning that BI does not sacrifice accuracy on clean data and is supported\nby a theoretical guarantee of its generalization capability.\n", "date": "2025-02-09", "categories": []} {"id": "2502.03721", "title": "Detecting Backdoor Attacks via Similarity in Semantic Communication\n Systems", "abstract": " Semantic communication systems, which leverage Generative AI (GAI) to\ntransmit semantic meaning rather than raw data, are poised to revolutionize\nmodern communications. However, they are vulnerable to backdoor attacks, a type\nof poisoning manipulation that embeds malicious triggers into training\ndatasets. As a result, Backdoor attacks mislead the inference for poisoned\nsamples while clean samples remain unaffected. The existing defenses may alter\nthe model structure (such as neuron pruning that potentially degrades inference\nperformance on clean inputs, or impose strict requirements on data formats\n(such as ``Semantic Shield\" that requires image-text pairs). To address these\nlimitations, this work proposes a defense mechanism that leverages semantic\nsimilarity to detect backdoor attacks without modifying the model structure or\nimposing data format constraints. By analyzing deviations in semantic feature\nspace and establishing a threshold-based detection framework, the proposed\napproach effectively identifies poisoned samples. The experimental results\ndemonstrate high detection accuracy and recall across varying poisoning ratios,\nunderlining the significant effectiveness of our proposed solution.\n", "date": "2025-02-06", "categories": []} {"id": "2502.04229", "title": "Dark Distillation: Backdooring Distilled Datasets without Accessing Raw\n Data", "abstract": " Dataset distillation (DD) enhances training efficiency and reduces bandwidth\nby condensing large datasets into smaller synthetic ones. It enables models to\nachieve performance comparable to those trained on the raw full dataset and has\nbecome a widely adopted method for data sharing. However, security concerns in\nDD remain underexplored. Existing studies typically assume that malicious\nbehavior originates from dataset owners during the initial distillation\nprocess, where backdoors are injected into raw datasets. In contrast, this work\nis the first to address a more realistic and concerning threat: attackers may\nintercept the dataset distribution process, inject backdoors into the distilled\ndatasets, and redistribute them to users. While distilled datasets were\npreviously considered resistant to backdoor attacks, we demonstrate that they\nremain vulnerable to such attacks. Furthermore, we show that attackers do not\neven require access to any raw data to inject the backdoors successfully.\nSpecifically, our approach reconstructs conceptual archetypes for each class\nfrom the model trained on the distilled dataset. Backdoors are then injected\ninto these archetypes to update the distilled dataset. Moreover, we ensure the\nupdated dataset not only retains the backdoor but also preserves the original\noptimization trajectory, thus maintaining the knowledge of the raw dataset. To\nachieve this, a hybrid loss is designed to integrate backdoor information along\nthe benign optimization trajectory, ensuring that previously learned\ninformation is not forgotten. Extensive experiments demonstrate that distilled\ndatasets are highly vulnerable to backdoor attacks, with risks pervasive across\nvarious raw datasets, distillation methods, and downstream training strategies.\nMoreover, our attack method is efficient, capable of synthesizing a malicious\ndistilled dataset in under one minute in certain cases.\n", "date": "2025-02-06", "categories": []} {"id": "2404.01101", "title": "UFID: A Unified Framework for Input-level Backdoor Detection on\n Diffusion Models", "abstract": " Diffusion models are vulnerable to backdoor attacks, where malicious\nattackers inject backdoors by poisoning certain training samples during the\ntraining stage. This poses a significant threat to real-world applications in\nthe Model-as-a-Service (MaaS) scenario, where users query diffusion models\nthrough APIs or directly download them from the internet. To mitigate the\nthreat of backdoor attacks under MaaS, black-box input-level backdoor detection\nhas drawn recent interest, where defenders aim to build a firewall that filters\nout backdoor samples in the inference stage, with access only to input queries\nand the generated results from diffusion models. Despite some preliminary\nexplorations on the traditional classification tasks, these methods cannot be\ndirectly applied to the generative tasks due to two major challenges: (1) more\ndiverse failures and (2) a multi-modality attack surface. In this paper, we\npropose a black-box input-level backdoor detection framework on diffusion\nmodels, called UFID. Our defense is motivated by an insightful causal analysis:\nBackdoor attacks serve as the confounder, introducing a spurious path from\ninput to target images, which remains consistent even when we perturb the input\nsamples with Gaussian noise. We further validate the intuition with theoretical\nanalysis. Extensive experiments across different datasets on both conditional\nand unconditional diffusion models show that our method achieves superb\nperformance on detection effectiveness and run-time efficiency.\n", "date": "2024-04-01", "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": "2304.08411", "title": "Evil from Within: Machine Learning Backdoors through Hardware Trojans", "abstract": " Backdoors pose a serious threat to machine learning, as they can compromise\nthe integrity of security-critical systems, such as self-driving cars. While\ndifferent defenses have been proposed to address this threat, they all rely on\nthe assumption that the hardware on which the learning models are executed\nduring inference is trusted. In this paper, we challenge this assumption and\nintroduce a backdoor attack that completely resides within a common hardware\naccelerator for machine learning. Outside of the accelerator, neither the\nlearning model nor the software is manipulated, so that current defenses fail.\nTo make this attack practical, we overcome two challenges: First, as memory on\na hardware accelerator is severely limited, we introduce the concept of a\nminimal backdoor that deviates as little as possible from the original model\nand is activated by replacing a few model parameters only. Second, we develop a\nconfigurable hardware trojan that can be provisioned with the backdoor and\nperforms a replacement only when the specific target model is processed. We\ndemonstrate the practical feasibility of our attack by implanting our hardware\ntrojan into the Xilinx Vitis AI DPU, a commercial machine-learning accelerator.\nWe configure the trojan with a minimal backdoor for a traffic-sign recognition\nsystem. The backdoor replaces only 30 (0.069%) model parameters, yet it\nreliably manipulates the recognition once the input contains a backdoor\ntrigger. Our attack expands the hardware circuit of the accelerator by 0.24%\nand induces no run-time overhead, rendering a detection hardly possible. Given\nthe complex and highly distributed manufacturing process of current hardware,\nour work points to a new threat in machine learning that is inaccessible to\ncurrent security mechanisms and calls for hardware to be manufactured only in\nfully trusted environments.\n", "date": "2023-04-17", "categories": []} {"id": "2502.01152", "title": "Gradient Norm-based Fine-Tuning for Backdoor Defense in Automatic Speech\n Recognition", "abstract": " Backdoor attacks have posed a significant threat to the security of deep\nneural networks (DNNs). Despite considerable strides in developing defenses\nagainst backdoor attacks in the visual domain, the specialized defenses for the\naudio domain remain empty. Furthermore, the defenses adapted from the visual to\naudio domain demonstrate limited effectiveness. To fill this gap, we propose\nGradient Norm-based FineTuning (GN-FT), a novel defense strategy against the\nattacks in the audio domain, based on the observation from the corresponding\nbackdoored models. Specifically, we first empirically find that the backdoored\nneurons exhibit greater gradient values compared to other neurons, while clean\nneurons stay the lowest. On this basis, we fine-tune the backdoored model by\nincorporating the gradient norm regularization, aiming to weaken and reduce the\nbackdoored neurons. We further approximate the loss computation for lower\nimplementation costs. Extensive experiments on two speech recognition datasets\nacross five models demonstrate the superior performance of our proposed method.\nTo the best of our knowledge, this work is the first specialized and effective\ndefense against backdoor attacks in the audio domain.\n", "date": "2025-02-03", "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": "2501.07192", "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": "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": "2406.06852", "title": "A Survey of Recent Backdoor Attacks and Defenses in Large Language\n Models", "abstract": " Large Language Models (LLMs), which bridge the gap between human language\nunderstanding and complex problem-solving, achieve state-of-the-art performance\non several NLP tasks, particularly in few-shot and zero-shot settings. Despite\nthe demonstrable efficacy of LLMs, due to constraints on computational\nresources, users have to engage with open-source language models or outsource\nthe entire training process to third-party platforms. However, research has\ndemonstrated that language models are susceptible to potential security\nvulnerabilities, particularly in backdoor attacks. Backdoor attacks are\ndesigned to introduce targeted vulnerabilities into language models by\npoisoning training samples or model weights, allowing attackers to manipulate\nmodel responses through malicious triggers. While existing surveys on backdoor\nattacks provide a comprehensive overview, they lack an in-depth examination of\nbackdoor attacks specifically targeting LLMs. To bridge this gap and grasp the\nlatest trends in the field, this paper presents a novel perspective on backdoor\nattacks for LLMs by focusing on fine-tuning methods. Specifically, we\nsystematically classify backdoor attacks into three categories: full-parameter\nfine-tuning, parameter-efficient fine-tuning, and no fine-tuning Based on\ninsights from a substantial review, we also discuss crucial issues for future\nresearch on backdoor attacks, such as further exploring attack algorithms that\ndo not require fine-tuning, or developing more covert attack algorithms.\n", "date": "2024-06-10", "categories": []} {"id": "2501.02754", "title": "MBTSAD: Mitigating Backdoors in Language Models Based on Token Splitting\n and Attention Distillation", "abstract": " In recent years, attention-based models have excelled across various domains\nbut remain vulnerable to backdoor attacks, often from downloading or\nfine-tuning on poisoned datasets. Many current methods to mitigate backdoors in\nNLP models rely on the pre-trained (unfine-tuned) weights, but these methods\nfail in scenarios where the pre-trained weights are not available. In this\nwork, we propose MBTSAD, which can mitigate backdoors in the language model by\nutilizing only a small subset of clean data and does not require pre-trained\nweights. Specifically, MBTSAD retrains the backdoored model on a dataset\ngenerated by token splitting. Then MBTSAD leverages attention distillation, the\nretrained model is the teacher model, and the original backdoored model is the\nstudent model. Experimental results demonstrate that MBTSAD achieves comparable\nbackdoor mitigation performance as the methods based on pre-trained weights\nwhile maintaining the performance on clean data. MBTSAD does not rely on\npre-trained weights, enhancing its utility in scenarios where pre-trained\nweights are inaccessible. In addition, we simplify the min-max problem of\nadversarial training and visualize text representations to discover that the\ntoken splitting method in MBTSAD's first step generates Out-of-Distribution\n(OOD) data, leading the model to learn more generalized features and eliminate\nbackdoor patterns.\n", "date": "2025-01-06", "categories": []} {"id": "2501.01090", "title": "HoneypotNet: Backdoor Attacks Against Model Extraction", "abstract": " Model extraction attacks are one type of inference-time attacks that\napproximate the functionality and performance of a black-box victim model by\nlaunching a certain number of queries to the model and then leveraging the\nmodel's predictions to train a substitute model. These attacks pose severe\nsecurity threats to production models and MLaaS platforms and could cause\nsignificant monetary losses to the model owners. A body of work has proposed to\ndefend machine learning models against model extraction attacks, including both\nactive defense methods that modify the model's outputs or increase the query\noverhead to avoid extraction and passive defense methods that detect malicious\nqueries or leverage watermarks to perform post-verification. In this work, we\nintroduce a new defense paradigm called attack as defense which modifies the\nmodel's output to be poisonous such that any malicious users that attempt to\nuse the output to train a substitute model will be poisoned. To this end, we\npropose a novel lightweight backdoor attack method dubbed HoneypotNet that\nreplaces the classification layer of the victim model with a honeypot layer and\nthen fine-tunes the honeypot layer with a shadow model (to simulate model\nextraction) via bi-level optimization to modify its output to be poisonous\nwhile remaining the original performance. We empirically demonstrate on four\ncommonly used benchmark datasets that HoneypotNet can inject backdoors into\nsubstitute models with a high success rate. The injected backdoor not only\nfacilitates ownership verification but also disrupts the functionality of\nsubstitute models, serving as a significant deterrent to model extraction\nattacks.\n", "date": "2025-01-02", "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": "2407.04108", "title": "Future Events as Backdoor Triggers: Investigating Temporal\n Vulnerabilities in LLMs", "abstract": " Backdoors are hidden behaviors that are only triggered once an AI system has\nbeen deployed. Bad actors looking to create successful backdoors must design\nthem to avoid activation during training and evaluation. Since data used in\nthese stages often only contains information about events that have already\noccurred, a component of a simple backdoor trigger could be a model recognizing\ndata that is in the future relative to when it was trained. Through prompting\nexperiments and by probing internal activations, we show that current large\nlanguage models (LLMs) can distinguish past from future events, with probes on\nmodel activations achieving 90% accuracy. We train models with backdoors\ntriggered by a temporal distributional shift; they activate when the model is\nexposed to news headlines beyond their training cut-off dates. Fine-tuning on\nhelpful, harmless and honest (HHH) data does not work well for removing simpler\nbackdoor triggers but is effective on our backdoored models, although this\ndistinction is smaller for the larger-scale model we tested. We also find that\nan activation-steering vector representing a model's internal representation of\nthe date influences the rate of backdoor activation. We take these results as\ninitial evidence that, at least for models at the modest scale we test,\nstandard safety measures are enough to remove these backdoors.\n", "date": "2024-07-04", "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": "2412.10605", "title": "Client-Side Patching against Backdoor Attacks in Federated Learning", "abstract": " Federated learning is a versatile framework for training models in\ndecentralized environments. However, the trust placed in clients makes\nfederated learning vulnerable to backdoor attacks launched by malicious\nparticipants. While many defenses have been proposed, they often fail short\nwhen facing heterogeneous data distributions among participating clients. In\nthis paper, we propose a novel defense mechanism for federated learning systems\ndesigned to mitigate backdoor attacks on the clients-side. Our approach\nleverages adversarial learning techniques and model patching to neutralize the\nimpact of backdoor attacks. Through extensive experiments on the MNIST and\nFashion-MNIST datasets, we demonstrate that our defense effectively reduces\nbackdoor accuracy, outperforming existing state-of-the-art defenses, such as\nLFighter, FLAME, and RoseAgg, in i.i.d. and non-i.i.d. scenarios, while\nmaintaining competitive or superior accuracy on clean data.\n", "date": "2024-12-13", "categories": []} {"id": "2305.09574", "title": "UOR: Universal Backdoor Attacks on Pre-trained Language Models", "abstract": " Backdoors implanted in pre-trained language models (PLMs) can be transferred\nto various downstream tasks, which exposes a severe security threat. However,\nmost existing backdoor attacks against PLMs are un-targeted and task-specific.\nFew targeted and task-agnostic methods use manually pre-defined triggers and\noutput representations, which prevent the attacks from being more effective and\ngeneral. In this paper, we first summarize the requirements that a more\nthreatening backdoor attack against PLMs should satisfy, and then propose a new\nbackdoor attack method called UOR, which breaks the bottleneck of the previous\napproach by turning manual selection into automatic optimization. Specifically,\nwe define poisoned supervised contrastive learning which can automatically\nlearn the more uniform and universal output representations of triggers for\nvarious PLMs. Moreover, we use gradient search to select appropriate trigger\nwords which can be adaptive to different PLMs and vocabularies. Experiments\nshow that our method can achieve better attack performance on various text\nclassification tasks compared to manual methods. Further, we tested our method\non PLMs with different architectures, different usage paradigms, and more\ndifficult tasks, which demonstrated the universality of our method.\n", "date": "2023-05-16", "categories": []} {"id": "2412.10681", "title": "One Pixel is All I Need", "abstract": " Vision Transformers (ViTs) have achieved record-breaking performance in\nvarious visual tasks. However, concerns about their robustness against backdoor\nattacks have grown. Backdoor attacks involve associating a specific trigger\nwith a target label, causing the model to predict the attacker-specified label\nwhen the trigger is present, while correctly identifying clean images.We found\nthat ViTs exhibit higher attack success rates for quasi-triggers(patterns\ndifferent from but similar to the original training triggers)compared to CNNs.\nMoreover, some backdoor features in clean samples can suppress the original\ntrigger, making quasi-triggers more effective.To better understand and exploit\nthese vulnerabilities, we developed a tool called the Perturbation Sensitivity\nDistribution Map (PSDM). PSDM computes and sums gradients over many inputs to\nshow how sensitive the model is to small changes in the input. In ViTs, PSDM\nreveals a patch-like pattern where central pixels are more sensitive than\nedges. We use PSDM to guide the creation of quasi-triggers.Based on these\nfindings, we designed \"WorstVIT,\" a simple yet effective data poisoning\nbackdoor for ViT models. This attack requires an extremely low poisoning rate,\ntrains for just one epoch, and modifies a single pixel to successfully attack\nall validation images.\n", "date": "2024-12-14", "categories": []} {"id": "2412.13324", "title": "BadSAD: Clean-Label Backdoor Attacks against Deep Semi-Supervised\n Anomaly Detection", "abstract": " Image anomaly detection (IAD) is essential in applications such as industrial\ninspection, medical imaging, and security. Despite the progress achieved with\ndeep learning models like Deep Semi-Supervised Anomaly Detection (DeepSAD),\nthese models remain susceptible to backdoor attacks, presenting significant\nsecurity challenges. In this paper, we introduce BadSAD, a novel backdoor\nattack framework specifically designed to target DeepSAD models. Our approach\ninvolves two key phases: trigger injection, where subtle triggers are embedded\ninto normal images, and latent space manipulation, which positions and clusters\nthe poisoned images near normal images to make the triggers appear benign.\nExtensive experiments on benchmark datasets validate the effectiveness of our\nattack strategy, highlighting the severe risks that backdoor attacks pose to\ndeep learning-based anomaly detection systems.\n", "date": "2024-12-17", "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": "2405.07667", "title": "Simulate and Eliminate: Revoke Backdoors for Generative Large Language\n Models", "abstract": " With rapid advances, generative large language models (LLMs) dominate various\nNatural Language Processing (NLP) tasks from understanding to reasoning. Yet,\nlanguage models' inherent vulnerabilities may be exacerbated due to increased\naccessibility and unrestricted model training on massive data. A malicious\nadversary may publish poisoned data online and conduct backdoor attacks on the\nvictim LLMs pre-trained on the poisoned data. Backdoored LLMs behave\ninnocuously for normal queries and generate harmful responses when the backdoor\ntrigger is activated. Despite significant efforts paid to LLMs' safety issues,\nLLMs are still struggling against backdoor attacks. As Anthropic recently\nrevealed, existing safety training strategies, including supervised fine-tuning\n(SFT) and Reinforcement Learning from Human Feedback (RLHF), fail to revoke the\nbackdoors once the LLM is backdoored during the pre-training stage. In this\npaper, we present Simulate and Eliminate (SANDE) to erase the undesired\nbackdoored mappings for generative LLMs. We initially propose Overwrite\nSupervised Fine-tuning (OSFT) for effective backdoor removal when the trigger\nis known. Then, to handle scenarios where trigger patterns are unknown, we\nintegrate OSFT into our two-stage framework, SANDE. Unlike other works that\nassume access to cleanly trained models, our safety-enhanced LLMs are able to\nrevoke backdoors without any reference. Consequently, our safety-enhanced LLMs\nno longer produce targeted responses when the backdoor triggers are activated.\nWe conduct comprehensive experiments to show that our proposed SANDE is\neffective against backdoor attacks while bringing minimal harm to LLMs'\npowerful capability.\n", "date": "2024-05-13", "categories": []} {"id": "2306.04877", "title": "TRIGS: Trojan Identification from Gradient-based Signatures", "abstract": " Training machine learning models can be very expensive or even unaffordable.\nThis may be, for example, due to data limitations, such as unavailability or\nbeing too large, or computational power limitations. Therefore, it is a common\npractice to rely on open-source pre-trained models whenever possible.However,\nthis practice is alarming from a security perspective. Pre-trained models can\nbe infected with Trojan attacks, in which the attacker embeds a trigger in the\nmodel such that the model's behavior can be controlled by the attacker when the\ntrigger is present in the input. In this paper, we present a novel method for\ndetecting Trojan models. Our method creates a signature for a model based on\nactivation optimization. A classifier is then trained to detect a Trojan model\ngiven its signature. We call our method TRIGS for TRojan Identification from\nGradient-based Signatures. TRIGS achieves state-of-the-art performance on two\npublic datasets of convolutional models. Additionally, we introduce a new\nchallenging dataset of ImageNet models based on the vision transformer\narchitecture. TRIGS delivers the best performance on the new dataset,\nsurpassing the baseline methods by a large margin. Our experiments also show\nthat TRIGS requires only a small amount of clean samples to achieve good\nperformance, and works reasonably well even if the defender does not have prior\nknowledge about the attacker's model architecture. Our code and data are\npublicly available.\n", "date": "2023-06-08", "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": "2412.06149", "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": "2412.06219", "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": "2412.04776", "title": "Megatron: Evasive Clean-Label Backdoor Attacks against Vision\n Transformer", "abstract": " Vision transformers have achieved impressive performance in various\nvision-related tasks, but their vulnerability to backdoor attacks is\nunder-explored. A handful of existing works focus on dirty-label attacks with\nwrongly-labeled poisoned training samples, which may fail if a benign model\ntrainer corrects the labels. In this paper, we propose Megatron, an evasive\nclean-label backdoor attack against vision transformers, where the attacker\ninjects the backdoor without manipulating the data-labeling process. To\ngenerate an effective trigger, we customize two loss terms based on the\nattention mechanism used in transformer networks, i.e., latent loss and\nattention diffusion loss. The latent loss aligns the last attention layer\nbetween triggered samples and clean samples of the target label. The attention\ndiffusion loss emphasizes the attention diffusion area that encompasses the\ntrigger. A theoretical analysis is provided to underpin the rationale behind\nthe attention diffusion loss. Extensive experiments on CIFAR-10, GTSRB,\nCIFAR-100, and Tiny ImageNet demonstrate the effectiveness of Megatron.\nMegatron can achieve attack success rates of over 90% even when the position of\nthe trigger is slightly shifted during testing. Furthermore, Megatron achieves\nbetter evasiveness than baselines regarding both human visual inspection and\ndefense strategies (i.e., DBAVT, BAVT, Beatrix, TeCo, and SAGE).\n", "date": "2024-12-06", "categories": []} {"id": "2407.10180", "title": "Defending Against Repetitive Backdoor Attacks on Semi-supervised\n Learning through Lens of Rate-Distortion-Perception Trade-off", "abstract": " Semi-supervised learning (SSL) has achieved remarkable performance with a\nsmall fraction of labeled data by leveraging vast amounts of unlabeled data\nfrom the Internet. However, this large pool of untrusted data is extremely\nvulnerable to data poisoning, leading to potential backdoor attacks. Current\nbackdoor defenses are not yet effective against such a vulnerability in SSL. In\nthis study, we propose a novel method, Unlabeled Data Purification (UPure), to\ndisrupt the association between trigger patterns and target classes by\nintroducing perturbations in the frequency domain. By leveraging the\nRate-Distortion-Perception (RDP) trade-off, we further identify the frequency\nband, where the perturbations are added, and justify this selection. Notably,\nUPure purifies poisoned unlabeled data without the need of extra clean labeled\ndata. Extensive experiments on four benchmark datasets and five SSL algorithms\ndemonstrate that UPure effectively reduces the attack success rate from 99.78%\nto 0% while maintaining model accuracy. Code is available here:\n\\url{https://github.com/chengyi-chris/UPure}.\n", "date": "2024-07-14", "categories": []} {"id": "2412.02454", "title": "Gracefully Filtering Backdoor Samples for Generative Large Language\n Models without Retraining", "abstract": " Backdoor attacks remain significant security threats to generative large\nlanguage models (LLMs). Since generative LLMs output sequences of\nhigh-dimensional token logits instead of low-dimensional classification logits,\nmost existing backdoor defense methods designed for discriminative models like\nBERT are ineffective for generative LLMs. Inspired by the observed differences\nin learning behavior between backdoor and clean mapping in the frequency space,\nwe transform gradients of each training sample, directly influencing parameter\nupdates, into the frequency space. Our findings reveal a distinct separation\nbetween the gradients of backdoor and clean samples in the frequency space.\nBased on this phenomenon, we propose Gradient Clustering in the Frequency Space\nfor Backdoor Sample Filtering (GraCeFul), which leverages sample-wise gradients\nin the frequency space to effectively identify backdoor samples without\nrequiring retraining LLMs. Experimental results show that GraCeFul outperforms\nbaselines significantly. Notably, GraCeFul exhibits remarkable computational\nefficiency, achieving nearly 100% recall and F1 scores in identifying backdoor\nsamples, reducing the average success rate of various backdoor attacks to 0%\nwith negligible drops in clean accuracy across multiple free-style question\nanswering datasets. Additionally, GraCeFul generalizes to Llama-2 and Vicuna.\nThe codes are publicly available at https://github.com/ZrW00/GraceFul.\n", "date": "2024-12-03", "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": "2412.01646", "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": "2411.18280", "title": "Neutralizing Backdoors through Information Conflicts for Large Language\n Models", "abstract": " Large language models (LLMs) have seen significant advancements, achieving\nsuperior performance in various Natural Language Processing (NLP) tasks, from\nunderstanding to reasoning. However, they remain vulnerable to backdoor\nattacks, where models behave normally for standard queries but generate harmful\nresponses or unintended output when specific triggers are activated. Existing\nbackdoor defenses often suffer from drawbacks that they either focus on\ndetection without removal, rely on rigid assumptions about trigger properties,\nor prove to be ineffective against advanced attacks like multi-trigger\nbackdoors. In this paper, we present a novel method to eliminate backdoor\nbehaviors from LLMs through the construction of information conflicts using\nboth internal and external mechanisms. Internally, we leverage a lightweight\ndataset to train a conflict model, which is then merged with the backdoored\nmodel to neutralize malicious behaviors by embedding contradictory information\nwithin the model's parametric memory. Externally, we incorporate convincing\ncontradictory evidence into the prompt to challenge the model's internal\nbackdoor knowledge. Experimental results on classification and conversational\ntasks across 4 widely used LLMs demonstrate that our method outperforms 8\nstate-of-the-art backdoor defense baselines. We can reduce the attack success\nrate of advanced backdoor attacks by up to 98% while maintaining over 90% clean\ndata accuracy. Furthermore, our method has proven to be robust against adaptive\nbackdoor attacks. The code will be open-sourced upon publication.\n", "date": "2024-11-27", "categories": []} {"id": "2411.17283", "title": "BadScan: An Architectural Backdoor Attack on Visual State Space Models", "abstract": " The newly introduced Visual State Space Model (VMamba), which employs\n\\textit{State Space Mechanisms} (SSM) to interpret images as sequences of\npatches, has shown exceptional performance compared to Vision Transformers\n(ViT) across various computer vision tasks. However, recent studies have\nhighlighted that deep models are susceptible to adversarial attacks. One common\napproach is to embed a trigger in the training data to retrain the model,\ncausing it to misclassify data samples into a target class, a phenomenon known\nas a backdoor attack. In this paper, we first evaluate the robustness of the\nVMamba model against existing backdoor attacks. Based on this evaluation, we\nintroduce a novel architectural backdoor attack, termed BadScan, designed to\ndeceive the VMamba model. This attack utilizes bit plane slicing to create\nvisually imperceptible backdoored images. During testing, if a trigger is\ndetected by performing XOR operations between the $k^{th}$ bit planes of the\nmodified triggered patches, the traditional 2D selective scan (SS2D) mechanism\nin the visual state space (VSS) block of VMamba is replaced with our newly\ndesigned BadScan block, which incorporates four newly developed scanning\npatterns. We demonstrate that the BadScan backdoor attack represents a\nsignificant threat to visual state space models and remains effective even\nafter complete retraining from scratch. Experimental results on two widely used\nimage classification datasets, CIFAR-10, and ImageNet-1K, reveal that while\nvisual state space models generally exhibit robustness against current backdoor\nattacks, the BadScan attack is particularly effective, achieving a higher\nTriggered Accuracy Ratio (TAR) in misleading the VMamba model and its variants.\n", "date": "2024-11-26", "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": "2411.15439", "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": "2411.15673", "title": "Semantic Shield: Defending Vision-Language Models Against Backdooring\n and Poisoning via Fine-grained Knowledge Alignment", "abstract": " In recent years there has been enormous interest in vision-language models\ntrained using self-supervised objectives. However, the use of large-scale\ndatasets scraped from the web for training also makes these models vulnerable\nto potential security threats, such as backdooring and poisoning attacks. In\nthis paper, we propose a method for mitigating such attacks on contrastively\ntrained vision-language models. Our approach leverages external knowledge\nextracted from a language model to prevent models from learning correlations\nbetween image regions which lack strong alignment with external knowledge. We\ndo this by imposing constraints to enforce that attention paid by the model to\nvisual regions is proportional to the alignment of those regions with external\nknowledge. We conduct extensive experiments using a variety of recent\nbackdooring and poisoning attacks on multiple datasets and architectures. Our\nresults clearly demonstrate that our proposed approach is highly effective at\ndefending against such attacks across multiple settings, while maintaining\nmodel utility and without requiring any changes at inference time\n", "date": "2024-11-23", "categories": []} {"id": "2411.16512", "title": "Guarding the Gate: ConceptGuard Battles Concept-Level Backdoors in\n Concept Bottleneck Models", "abstract": " The increasing complexity of AI models, especially in deep learning, has\nraised concerns about transparency and accountability, particularly in\nhigh-stakes applications like medical diagnostics, where opaque models can\nundermine trust. Explainable Artificial Intelligence (XAI) aims to address\nthese issues by providing clear, interpretable models. Among XAI techniques,\nConcept Bottleneck Models (CBMs) enhance transparency by using high-level\nsemantic concepts. However, CBMs are vulnerable to concept-level backdoor\nattacks, which inject hidden triggers into these concepts, leading to\nundetectable anomalous behavior. To address this critical security gap, we\nintroduce ConceptGuard, a novel defense framework specifically designed to\nprotect CBMs from concept-level backdoor attacks. ConceptGuard employs a\nmulti-stage approach, including concept clustering based on text distance\nmeasurements and a voting mechanism among classifiers trained on different\nconcept subgroups, to isolate and mitigate potential triggers. Our\ncontributions are threefold: (i) we present ConceptGuard as the first defense\nmechanism tailored for concept-level backdoor attacks in CBMs; (ii) we provide\ntheoretical guarantees that ConceptGuard can effectively defend against such\nattacks within a certain trigger size threshold, ensuring robustness; and (iii)\nwe demonstrate that ConceptGuard maintains the high performance and\ninterpretability of CBMs, crucial for trustworthiness. Through comprehensive\nexperiments and theoretical proofs, we show that ConceptGuard significantly\nenhances the security and trustworthiness of CBMs, paving the way for their\nsecure deployment in critical applications.\n", "date": "2024-11-25", "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": "2411.11525", "title": "Reliable Poisoned Sample Detection against Backdoor Attacks Enhanced by\n Sharpness Aware Minimization", "abstract": " Backdoor attack has been considered as a serious security threat to deep\nneural networks (DNNs). Poisoned sample detection (PSD) that aims at filtering\nout poisoned samples from an untrustworthy training dataset has shown very\npromising performance for defending against data poisoning based backdoor\nattacks. However, we observe that the detection performance of many advanced\nmethods is likely to be unstable when facing weak backdoor attacks, such as low\npoisoning ratio or weak trigger strength. To further verify this observation,\nwe make a statistical investigation among various backdoor attacks and poisoned\nsample detections, showing a positive correlation between backdoor effect and\ndetection performance. It inspires us to strengthen the backdoor effect to\nenhance detection performance. Since we cannot achieve that goal via directly\nmanipulating poisoning ratio or trigger strength, we propose to train one model\nusing the Sharpness-Aware Minimization (SAM) algorithm, rather than the vanilla\ntraining algorithm. We also provide both empirical and theoretical analysis\nabout how SAM training strengthens the backdoor effect. Then, this SAM trained\nmodel can be seamlessly integrated with any off-the-shelf PSD method that\nextracts discriminative features from the trained model for detection, called\nSAM-enhanced PSD. Extensive experiments on several benchmark datasets show the\nreliable detection performance of the proposed method against both weak and\nstrong backdoor attacks, with significant improvements against various attacks\n($+34.38\\%$ TPR on average), over the conventional PSD methods (i.e., without\nSAM enhancement). Overall, this work provides new insights about PSD and\nproposes a novel approach that can complement existing detection methods, which\nmay inspire more in-depth explorations in this field.\n", "date": "2024-11-18", "categories": []} {"id": "2409.17601", "title": "CleanerCLIP: Fine-grained Counterfactual Semantic Augmentation for\n Backdoor Defense in Contrastive Learning", "abstract": " Pre-trained large models for multimodal contrastive learning, such as CLIP,\nhave been widely recognized in the industry as highly susceptible to\ndata-poisoned backdoor attacks. This poses significant risks to downstream\nmodel training. In response to such potential threats, finetuning offers a\nsimpler and more efficient defense choice compared to retraining large models\nwith augmented data. In the supervised learning domain, fine-tuning defense\nstrategies can achieve excellent defense performance. However, in the\nunsupervised and semi-supervised domain, we find that when CLIP faces some\ncomplex attack techniques, the existing fine-tuning defense strategy,\nCleanCLIP, has some limitations on defense performance. The synonym\nsubstitution of its text-augmentation is insufficient to enhance the text\nfeature space. To compensate for this weakness, we improve it by proposing a\nfine-grained \\textbf{T}ext \\textbf{A}lignment \\textbf{C}leaner (TA-Cleaner) to\ncut off feature connections of backdoor triggers. We randomly select a few\nsamples for positive and negative subtext generation at each epoch of\nCleanCLIP, and align the subtexts to the images to strengthen the text\nself-supervision. We evaluate the effectiveness of our TA-Cleaner against six\nattack algorithms and conduct comprehensive zero-shot classification tests on\nImageNet1K. Our experimental results demonstrate that TA-Cleaner achieves\nstate-of-the-art defensiveness among finetuning-based defense techniques. Even\nwhen faced with the novel attack technique BadCLIP, our TA-Cleaner outperforms\nCleanCLIP by reducing the ASR of Top-1 and Top-10 by 52.02\\% and 63.88\\%,\nrespectively.\n", "date": "2024-09-26", "categories": []} {"id": "2410.01272", "title": "\"No Matter What You Do\": Purifying GNN Models via Backdoor Unlearning", "abstract": " Recent studies have exposed that GNNs are vulnerable to several adversarial\nattacks, among which backdoor attack is one of the toughest. Similar to Deep\nNeural Networks (DNNs), backdoor attacks in GNNs lie in the fact that the\nattacker modifies a portion of graph data by embedding triggers and enforces\nthe model to learn the trigger feature during the model training process.\nDespite the massive prior backdoor defense works on DNNs, defending against\nbackdoor attacks in GNNs is largely unexplored, severely hindering the\nwidespread application of GNNs in real-world tasks. To bridge this gap, we\npresent GCleaner, the first backdoor mitigation method on GNNs. GCleaner can\nmitigate the presence of the backdoor logic within backdoored GNNs by reversing\nthe backdoor learning procedure, aiming to restore the model performance to a\nlevel similar to that is directly trained on the original clean dataset. To\nachieve this objective, we ask: How to recover universal and hard backdoor\ntriggers in GNNs? How to unlearn the backdoor trigger feature while maintaining\nthe model performance? We conduct the graph trigger recovery via the\nexplanation method to identify optimal trigger locations, facilitating the\nsearch of universal and hard backdoor triggers in the feature space of the\nbackdoored model through maximal similarity. Subsequently, we introduce the\nbackdoor unlearning mechanism, which combines knowledge distillation and\ngradient-based explainable knowledge for fine-grained backdoor erasure.\nExtensive experimental evaluations on four benchmark datasets demonstrate that\nGCleaner can reduce the backdoor attack success rate to 10% with only 1% of\nclean data, and has almost negligible degradation in model performance, which\nfar outperforms the state-of-the-art (SOTA) defense methods.\n", "date": "2024-10-02", "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": "2411.04811", "title": "Defending Deep Regression Models against Backdoor Attacks", "abstract": " Deep regression models are used in a wide variety of safety-critical\napplications, but are vulnerable to backdoor attacks. Although many defenses\nhave been proposed for classification models, they are ineffective as they do\nnot consider the uniqueness of regression models. First, the outputs of\nregression models are continuous values instead of discretized labels. Thus,\nthe potential infected target of a backdoored regression model has infinite\npossibilities, which makes it impossible to be determined by existing defenses.\nSecond, the backdoor behavior of backdoored deep regression models is triggered\nby the activation values of all the neurons in the feature space, which makes\nit difficult to be detected and mitigated using existing defenses. To resolve\nthese problems, we propose DRMGuard, the first defense to identify if a deep\nregression model in the image domain is backdoored or not. DRMGuard formulates\nthe optimization problem for reverse engineering based on the unique\noutput-space and feature-space characteristics of backdoored deep regression\nmodels. We conduct extensive evaluations on two regression tasks and four\ndatasets. The results show that DRMGuard can consistently defend against\nvarious backdoor attacks. We also generalize four state-of-the-art defenses\ndesigned for classifiers to regression models, and compare DRMGuard with them.\nThe results show that DRMGuard significantly outperforms all those defenses.\n", "date": "2024-11-07", "categories": []} {"id": "2411.03022", "title": "Flashy Backdoor: Real-world Environment Backdoor Attack on SNNs with DVS\n Cameras", "abstract": " While security vulnerabilities in traditional Deep Neural Networks (DNNs)\nhave been extensively studied, the susceptibility of Spiking Neural Networks\n(SNNs) to adversarial attacks remains mostly underexplored. Until now, the\nmechanisms to inject backdoors into SNN models have been limited to digital\nscenarios; thus, we present the first evaluation of backdoor attacks in\nreal-world environments.\n We begin by assessing the applicability of existing digital backdoor attacks\nand identifying their limitations for deployment in physical environments. To\naddress each of the found limitations, we present three novel backdoor attack\nmethods on SNNs, i.e., Framed, Strobing, and Flashy Backdoor. We also assess\nthe effectiveness of traditional backdoor procedures and defenses adapted for\nSNNs, such as pruning, fine-tuning, and fine-pruning. The results show that\nwhile these procedures and defenses can mitigate some attacks, they often fail\nagainst stronger methods like Flashy Backdoor or sacrifice too much clean\naccuracy, rendering the models unusable.\n Overall, all our methods can achieve up to a 100% Attack Success Rate while\nmaintaining high clean accuracy in every tested dataset. Additionally, we\nevaluate the stealthiness of the triggers with commonly used metrics, finding\nthem highly stealthy. Thus, we propose new alternatives more suited for\nidentifying poisoned samples in these scenarios. Our results show that further\nresearch is needed to ensure the security of SNN-based systems against backdoor\nattacks and their safe application in real-world scenarios. The code,\nexperiments, and results are available in our repository.\n", "date": "2024-11-05", "categories": []} {"id": "2402.18718", "title": "Model Pairing Using Embedding Translation for Backdoor Attack Detection\n on Open-Set Classification Tasks", "abstract": " Backdoor attacks allow an attacker to embed a specific vulnerability in a\nmachine learning algorithm, activated when an attacker-chosen pattern is\npresented, causing a specific misprediction. The need to identify backdoors in\nbiometric scenarios has led us to propose a novel technique with different\ntrade-offs. In this paper we propose to use model pairs on open-set\nclassification tasks for detecting backdoors. Using a simple linear operation\nto project embeddings from a probe model's embedding space to a reference\nmodel's embedding space, we can compare both embeddings and compute a\nsimilarity score. We show that this score, can be an indicator for the presence\nof a backdoor despite models being of different architectures, having been\ntrained independently and on different datasets. This technique allows for the\ndetection of backdoors on models designed for open-set classification tasks,\nwhich is little studied in the literature. Additionally, we show that backdoors\ncan be detected even when both models are backdoored. The source code is made\navailable for reproducibility purposes.\n", "date": "2024-02-28", "categories": []} {"id": "2408.05500", "title": "PointNCBW: Towards Dataset Ownership Verification for Point Clouds via\n Negative Clean-label Backdoor Watermark", "abstract": " Recently, point clouds have been widely used in computer vision, whereas\ntheir collection is time-consuming and expensive. As such, point cloud datasets\nare the valuable intellectual property of their owners and deserve protection.\nTo detect and prevent unauthorized use of these datasets, especially for\ncommercial or open-sourced ones that cannot be sold again or used commercially\nwithout permission, we intend to identify whether a suspicious third-party\nmodel is trained on our protected dataset under the black-box setting. We\nachieve this goal by designing a scalable clean-label backdoor-based dataset\nwatermark for point clouds that ensures both effectiveness and stealthiness.\nUnlike existing clean-label watermark schemes, which are susceptible to the\nnumber of categories, our method could watermark samples from all classes\ninstead of only from the target one. Accordingly, it can still preserve high\neffectiveness even on large-scale datasets with many classes. Specifically, we\nperturb selected point clouds with non-target categories in both shape-wise and\npoint-wise manners before inserting trigger patterns without changing their\nlabels. The features of perturbed samples are similar to those of benign\nsamples from the target class. As such, models trained on the watermarked\ndataset will have a distinctive yet stealthy backdoor behavior, i.e.,\nmisclassifying samples from the target class whenever triggers appear, since\nthe trained DNNs will treat the inserted trigger pattern as a signal to deny\npredicting the target label. We also design a hypothesis-test-guided dataset\nownership verification based on the proposed watermark. Extensive experiments\non benchmark datasets are conducted, verifying the effectiveness of our method\nand its resistance to potential removal methods.\n", "date": "2024-08-10", "categories": []} {"id": "2312.07955", "title": "Erasing Self-Supervised Learning Backdoor by Cluster Activation Masking", "abstract": " Self-Supervised Learning (SSL) is an effective paradigm for learning\nrepresentations from unlabeled data, such as text, images, and videos. However,\nresearchers have recently found that SSL is vulnerable to backdoor attacks. The\nattacker can embed hidden SSL backdoors via a few poisoned examples in the\ntraining dataset and maliciously manipulate the behavior of downstream models.\nTo defend against SSL backdoor attacks, a feasible route is to detect and\nremove the poisonous samples in the training set. However, the existing SSL\nbackdoor defense method fails to detect the poisonous samples precisely. In\nthis paper, we propose to erase the SSL backdoor by cluster activation masking\nand propose a novel PoisonCAM method. After obtaining the threat model trained\non the poisoned dataset, our method can precisely detect poisonous samples\nbased on the assumption that masking the backdoor trigger can effectively\nchange the activation of a downstream clustering model. In experiments, our\nPoisonCAM achieves 96\\% accuracy for backdoor trigger detection compared to 3\\%\nof the state-of-the-art method on poisoned ImageNet-100. Moreover, our proposed\nPoisonCAM significantly improves the performance of the trained SSL model under\nbackdoor attacks compared to the state-of-the-art method. Our code, data, and\ntrained models will be open once this paper is accepted.\n", "date": "2023-12-13", "categories": []} {"id": "2407.04151", "title": "Securing Multi-turn Conversational Language Models From Distributed\n Backdoor Triggers", "abstract": " Large language models (LLMs) have acquired the ability to handle longer\ncontext lengths and understand nuances in text, expanding their dialogue\ncapabilities beyond a single utterance. A popular user-facing application of\nLLMs is the multi-turn chat setting. Though longer chat memory and better\nunderstanding may seemingly benefit users, our paper exposes a vulnerability\nthat leverages the multi-turn feature and strong learning ability of LLMs to\nharm the end-user: the backdoor. We demonstrate that LLMs can capture the\ncombinational backdoor representation. Only upon presentation of triggers\ntogether does the backdoor activate. We also verify empirically that this\nrepresentation is invariant to the position of the trigger utterance.\nSubsequently, inserting a single extra token into two utterances of 5%of the\ndata can cause over 99% Attack Success Rate (ASR). Our results with 3 triggers\ndemonstrate that this framework is generalizable, compatible with any trigger\nin an adversary's toolbox in a plug-and-play manner. Defending the backdoor can\nbe challenging in the chat setting because of the large input and output space.\nOur analysis indicates that the distributed backdoor exacerbates the current\nchallenges by polynomially increasing the dimension of the attacked input\nspace. Canonical textual defenses like ONION and BKI leverage auxiliary model\nforward passes over individual tokens, scaling exponentially with the input\nsequence length and struggling to maintain computational feasibility. To this\nend, we propose a decoding time defense - decayed contrastive decoding - that\nscales linearly with assistant response sequence length and reduces the\nbackdoor to as low as 0.35%.\n", "date": "2024-07-04", "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": "2410.19427", "title": "Expose Before You Defend: Unifying and Enhancing Backdoor Defenses via\n Exposed Models", "abstract": " Backdoor attacks covertly implant triggers into deep neural networks (DNNs)\nby poisoning a small portion of the training data with pre-designed backdoor\ntriggers. This vulnerability is exacerbated in the era of large models, where\nextensive (pre-)training on web-crawled datasets is susceptible to compromise.\nIn this paper, we introduce a novel two-step defense framework named Expose\nBefore You Defend (EBYD). EBYD unifies existing backdoor defense methods into a\ncomprehensive defense system with enhanced performance. Specifically, EBYD\nfirst exposes the backdoor functionality in the backdoored model through a\nmodel preprocessing step called backdoor exposure, and then applies detection\nand removal methods to the exposed model to identify and eliminate the backdoor\nfeatures. In the first step of backdoor exposure, we propose a novel technique\ncalled Clean Unlearning (CUL), which proactively unlearns clean features from\nthe backdoored model to reveal the hidden backdoor features. We also explore\nvarious model editing/modification techniques for backdoor exposure, including\nfine-tuning, model sparsification, and weight perturbation. Using EBYD, we\nconduct extensive experiments on 10 image attacks and 6 text attacks across 2\nvision datasets (CIFAR-10 and an ImageNet subset) and 4 language datasets\n(SST-2, IMDB, Twitter, and AG's News). The results demonstrate the importance\nof backdoor exposure for backdoor defense, showing that the exposed models can\nsignificantly benefit a range of downstream defense tasks, including backdoor\nlabel detection, backdoor trigger recovery, backdoor model detection, and\nbackdoor removal. We hope our work could inspire more research in developing\nadvanced defense frameworks with exposed models. Our code is available at:\nhttps://github.com/bboylyg/Expose-Before-You-Defend.\n", "date": "2024-10-25", "categories": []} {"id": "2402.17465", "title": "Model X-ray:Detecting Backdoored Models via Decision Boundary", "abstract": " Backdoor attacks pose a significant security vulnerability for deep neural\nnetworks (DNNs), enabling them to operate normally on clean inputs but\nmanipulate predictions when specific trigger patterns occur. Currently,\npost-training backdoor detection approaches often operate under the assumption\nthat the defender has knowledge of the attack information, logit output from\nthe model, and knowledge of the model parameters. In contrast, our approach\nfunctions as a lightweight diagnostic scanning tool offering interpretability\nand visualization. By accessing the model to obtain hard labels, we construct\ndecision boundaries within the convex combination of three samples. We present\nan intriguing observation of two phenomena in backdoored models: a noticeable\nshrinking of areas dominated by clean samples and a significant increase in the\nsurrounding areas dominated by target labels. Leveraging this observation, we\npropose Model X-ray, a novel backdoor detection approach based on the analysis\nof illustrated two-dimensional (2D) decision boundaries. Our approach includes\ntwo strategies focused on the decision areas dominated by clean samples and the\nconcentration of label distribution, and it can not only identify whether the\ntarget model is infected but also determine the target attacked label under the\nall-to-one attack strategy. Importantly, it accomplishes this solely by the\npredicted hard labels of clean inputs, regardless of any assumptions about\nattacks and prior knowledge of the training details of the model. Extensive\nexperiments demonstrated that Model X-ray has outstanding effectiveness and\nefficiency across diverse backdoor attacks, datasets, and architectures.\nBesides, ablation studies on hyperparameters and more attack strategies and\ndiscussions are also provided.\n", "date": "2024-02-27", "categories": []} {"id": "2406.10932", "title": "Imperceptible Rhythm Backdoor Attacks: Exploring Rhythm Transformation\n for Embedding Undetectable Vulnerabilities on Speech Recognition", "abstract": " Speech recognition is an essential start ring of human-computer interaction,\nand recently, deep learning models have achieved excellent success in this\ntask. However, when the model training and private data provider are always\nseparated, some security threats that make deep neural networks (DNNs) abnormal\ndeserve to be researched. In recent years, the typical backdoor attacks have\nbeen researched in speech recognition systems. The existing backdoor methods\nare based on data poisoning. The attacker adds some incorporated changes to\nbenign speech spectrograms or changes the speech components, such as pitch and\ntimbre. As a result, the poisoned data can be detected by human hearing or\nautomatic deep algorithms. To improve the stealthiness of data poisoning, we\npropose a non-neural and fast algorithm called Random Spectrogram Rhythm\nTransformation (RSRT) in this paper. The algorithm combines four steps to\ngenerate stealthy poisoned utterances. From the perspective of rhythm component\ntransformation, our proposed trigger stretches or squeezes the mel spectrograms\nand recovers them back to signals. The operation keeps timbre and content\nunchanged for good stealthiness. Our experiments are conducted on two kinds of\nspeech recognition tasks, including testing the stealthiness of poisoned\nsamples by speaker verification and automatic speech recognition. The results\nshow that our method has excellent effectiveness and stealthiness. The rhythm\ntrigger needs a low poisoning rate and gets a very high attack success rate.\n", "date": "2024-06-16", "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": "2410.14479", "title": "Backdoored Retrievers for Prompt Injection Attacks on Retrieval\n Augmented Generation of Large Language Models", "abstract": " Large Language Models (LLMs) have demonstrated remarkable capabilities in\ngenerating coherent text but remain limited by the static nature of their\ntraining data. Retrieval Augmented Generation (RAG) addresses this issue by\ncombining LLMs with up-to-date information retrieval, but also expand the\nattack surface of the system. This paper investigates prompt injection attacks\non RAG, focusing on malicious objectives beyond misinformation, such as\ninserting harmful links, promoting unauthorized services, and initiating\ndenial-of-service behaviors. We build upon existing corpus poisoning techniques\nand propose a novel backdoor attack aimed at the fine-tuning process of the\ndense retriever component. Our experiments reveal that corpus poisoning can\nachieve significant attack success rates through the injection of a small\nnumber of compromised documents into the retriever corpus. In contrast,\nbackdoor attacks demonstrate even higher success rates but necessitate a more\ncomplex setup, as the victim must fine-tune the retriever using the attacker\npoisoned dataset.\n", "date": "2024-10-18", "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": "2410.09838", "title": "Uncovering, Explaining, and Mitigating the Superficial Safety of\n Backdoor Defense", "abstract": " Backdoor attacks pose a significant threat to Deep Neural Networks (DNNs) as\nthey allow attackers to manipulate model predictions with backdoor triggers. To\naddress these security vulnerabilities, various backdoor purification methods\nhave been proposed to purify compromised models. Typically, these purified\nmodels exhibit low Attack Success Rates (ASR), rendering them resistant to\nbackdoored inputs. However, Does achieving a low ASR through current safety\npurification methods truly eliminate learned backdoor features from the\npretraining phase? In this paper, we provide an affirmative answer to this\nquestion by thoroughly investigating the Post-Purification Robustness of\ncurrent backdoor purification methods. We find that current safety purification\nmethods are vulnerable to the rapid re-learning of backdoor behavior, even when\nfurther fine-tuning of purified models is performed using a very small number\nof poisoned samples. Based on this, we further propose the practical\nQuery-based Reactivation Attack (QRA) which could effectively reactivate the\nbackdoor by merely querying purified models. We find the failure to achieve\nsatisfactory post-purification robustness stems from the insufficient deviation\nof purified models from the backdoored model along the backdoor-connected path.\nTo improve the post-purification robustness, we propose a straightforward\ntuning defense, Path-Aware Minimization (PAM), which promotes deviation along\nbackdoor-connected paths with extra model updates. Extensive experiments\ndemonstrate that PAM significantly improves post-purification robustness while\nmaintaining a good clean accuracy and low ASR. Our work provides a new\nperspective on understanding the effectiveness of backdoor safety tuning and\nhighlights the importance of faithfully assessing the model's safety.\n", "date": "2024-10-13", "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": "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": "2401.05949", "title": "Universal Vulnerabilities in Large Language Models: Backdoor Attacks for\n In-context Learning", "abstract": " In-context learning, a paradigm bridging the gap between pre-training and\nfine-tuning, has demonstrated high efficacy in several NLP tasks, especially in\nfew-shot settings. Despite being widely applied, in-context learning is\nvulnerable to malicious attacks. In this work, we raise security concerns\nregarding this paradigm. Our studies demonstrate that an attacker can\nmanipulate the behavior of large language models by poisoning the demonstration\ncontext, without the need for fine-tuning the model. Specifically, we design a\nnew backdoor attack method, named ICLAttack, to target large language models\nbased on in-context learning. Our method encompasses two types of attacks:\npoisoning demonstration examples and poisoning demonstration prompts, which can\nmake models behave in alignment with predefined intentions. ICLAttack does not\nrequire additional fine-tuning to implant a backdoor, thus preserving the\nmodel's generality. Furthermore, the poisoned examples are correctly labeled,\nenhancing the natural stealth of our attack method. Extensive experimental\nresults across several language models, ranging in size from 1.3B to 180B\nparameters, demonstrate the effectiveness of our attack method, exemplified by\na high average attack success rate of 95.0% across the three datasets on OPT\nmodels.\n", "date": "2024-01-11", "categories": []} {"id": "2409.14119", "title": "Obliviate: Neutralizing Task-agnostic Backdoors within the\n Parameter-efficient Fine-tuning Paradigm", "abstract": " Parameter-efficient fine-tuning (PEFT) has become a key training strategy for\nlarge language models. However, its reliance on fewer trainable parameters\nposes security risks, such as task-agnostic backdoors. Despite their severe\nimpact on a wide range of tasks, there is no practical defense solution\navailable that effectively counters task-agnostic backdoors within the context\nof PEFT. In this study, we introduce Obliviate, a PEFT-integrable backdoor\ndefense. We develop two techniques aimed at amplifying benign neurons within\nPEFT layers and penalizing the influence of trigger tokens. Our evaluations\nacross three major PEFT architectures show that our method can significantly\nreduce the attack success rate of the state-of-the-art task-agnostic backdoors\n(83.6%$\\downarrow$). Furthermore, our method exhibits robust defense\ncapabilities against both task-specific backdoors and adaptive attacks. Source\ncode will be obtained at https://github.com/obliviateARR/Obliviate.\n", "date": "2024-09-21", "categories": []} {"id": "2410.02182", "title": "BadCM: Invisible Backdoor Attack Against Cross-Modal Learning", "abstract": " Despite remarkable successes in unimodal learning tasks, backdoor attacks\nagainst cross-modal learning are still underexplored due to the limited\ngeneralization and inferior stealthiness when involving multiple modalities.\nNotably, since works in this area mainly inherit ideas from unimodal visual\nattacks, they struggle with dealing with diverse cross-modal attack\ncircumstances and manipulating imperceptible trigger samples, which hinders\ntheir practicability in real-world applications. In this paper, we introduce a\nnovel bilateral backdoor to fill in the missing pieces of the puzzle in the\ncross-modal backdoor and propose a generalized invisible backdoor framework\nagainst cross-modal learning (BadCM). Specifically, a cross-modal mining scheme\nis developed to capture the modality-invariant components as target poisoning\nareas, where well-designed trigger patterns injected into these regions can be\nefficiently recognized by the victim models. This strategy is adapted to\ndifferent image-text cross-modal models, making our framework available to\nvarious attack scenarios. Furthermore, for generating poisoned samples of high\nstealthiness, we conceive modality-specific generators for visual and\nlinguistic modalities that facilitate hiding explicit trigger patterns in\nmodality-invariant regions. To the best of our knowledge, BadCM is the first\ninvisible backdoor method deliberately designed for diverse cross-modal attacks\nwithin one unified framework. Comprehensive experimental evaluations on two\ntypical applications, i.e., cross-modal retrieval and VQA, demonstrate the\neffectiveness and generalization of our method under multiple kinds of attack\nscenarios. Moreover, we show that BadCM can robustly evade existing backdoor\ndefenses. Our code is available at https://github.com/xandery-geek/BadCM.\n", "date": "2024-10-03", "categories": []} {"id": "2410.01128", "title": "Using Interleaved Ensemble Unlearning to Keep Backdoors at Bay for\n Finetuning Vision Transformers", "abstract": " Vision Transformers (ViTs) have become popular in computer vision tasks.\nBackdoor attacks, which trigger undesirable behaviours in models during\ninference, threaten ViTs' performance, particularly in security-sensitive\ntasks. Although backdoor defences have been developed for Convolutional Neural\nNetworks (CNNs), they are less effective for ViTs, and defences tailored to\nViTs are scarce. To address this, we present Interleaved Ensemble Unlearning\n(IEU), a method for finetuning clean ViTs on backdoored datasets. In stage 1, a\nshallow ViT is finetuned to have high confidence on backdoored data and low\nconfidence on clean data. In stage 2, the shallow ViT acts as a ``gate'' to\nblock potentially poisoned data from the defended ViT. This data is added to an\nunlearn set and asynchronously unlearned via gradient ascent. We demonstrate\nIEU's effectiveness on three datasets against 11 state-of-the-art backdoor\nattacks and show its versatility by applying it to different model\narchitectures.\n", "date": "2024-10-01", "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": "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": "2409.16057", "title": "Towards Robust Object Detection: Identifying and Removing Backdoors via\n Module Inconsistency Analysis", "abstract": " Object detection models, widely used in security-critical applications, are\nvulnerable to backdoor attacks that cause targeted misclassifications when\ntriggered by specific patterns. Existing backdoor defense techniques, primarily\ndesigned for simpler models like image classifiers, often fail to effectively\ndetect and remove backdoors in object detectors. We propose a backdoor defense\nframework tailored to object detection models, based on the observation that\nbackdoor attacks cause significant inconsistencies between local modules'\nbehaviors, such as the Region Proposal Network (RPN) and classification head.\nBy quantifying and analyzing these inconsistencies, we develop an algorithm to\ndetect backdoors. We find that the inconsistent module is usually the main\nsource of backdoor behavior, leading to a removal method that localizes the\naffected module, resets its parameters, and fine-tunes the model on a small\nclean dataset. Extensive experiments with state-of-the-art two-stage object\ndetectors show our method achieves a 90% improvement in backdoor removal rate\nover fine-tuning baselines, while limiting clean data accuracy loss to less\nthan 4%. To the best of our knowledge, this work presents the first approach\nthat addresses both the detection and removal of backdoors in two-stage object\ndetection models, advancing the field of securing these complex systems against\nbackdoor attacks.\n", "date": "2024-09-24", "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": "2409.19993", "title": "Mitigating Backdoor Threats to Large Language Models: Advancement and\n Challenges", "abstract": " The advancement of Large Language Models (LLMs) has significantly impacted\nvarious domains, including Web search, healthcare, and software development.\nHowever, as these models scale, they become more vulnerable to cybersecurity\nrisks, particularly backdoor attacks. By exploiting the potent memorization\ncapacity of LLMs, adversaries can easily inject backdoors into LLMs by\nmanipulating a small portion of training data, leading to malicious behaviors\nin downstream applications whenever the hidden backdoor is activated by the\npre-defined triggers. Moreover, emerging learning paradigms like instruction\ntuning and reinforcement learning from human feedback (RLHF) exacerbate these\nrisks as they rely heavily on crowdsourced data and human feedback, which are\nnot fully controlled. In this paper, we present a comprehensive survey of\nemerging backdoor threats to LLMs that appear during LLM development or\ninference, and cover recent advancement in both defense and detection\nstrategies for mitigating backdoor threats to LLMs. We also outline key\nchallenges in addressing these threats, highlighting areas for future research.\n", "date": "2024-09-30", "categories": []} {"id": "2409.15968", "title": "Adversarial Backdoor Defense in CLIP", "abstract": " Multimodal contrastive pretraining, exemplified by models like CLIP, has been\nfound to be vulnerable to backdoor attacks. While current backdoor defense\nmethods primarily employ conventional data augmentation to create augmented\nsamples aimed at feature alignment, these methods fail to capture the distinct\nfeatures of backdoor samples, resulting in suboptimal defense performance.\nObservations reveal that adversarial examples and backdoor samples exhibit\nsimilarities in the feature space within the compromised models. Building on\nthis insight, we propose Adversarial Backdoor Defense (ABD), a novel data\naugmentation strategy that aligns features with meticulously crafted\nadversarial examples. This approach effectively disrupts the backdoor\nassociation. Our experiments demonstrate that ABD provides robust defense\nagainst both traditional uni-modal and multimodal backdoor attacks targeting\nCLIP. Compared to the current state-of-the-art defense method, CleanCLIP, ABD\nreduces the attack success rate by 8.66% for BadNet, 10.52% for Blended, and\n53.64% for BadCLIP, while maintaining a minimal average decrease of just 1.73%\nin clean accuracy.\n", "date": "2024-09-24", "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": "2409.14200", "title": "Data-centric NLP Backdoor Defense from the Lens of Memorization", "abstract": " Backdoor attack is a severe threat to the trustworthiness of DNN-based\nlanguage models. In this paper, we first extend the definition of memorization\nof language models from sample-wise to more fine-grained sentence element-wise\n(e.g., word, phrase, structure, and style), and then point out that language\nmodel backdoors are a type of element-wise memorization. Through further\nanalysis, we find that the strength of such memorization is positively\ncorrelated to the frequency of duplicated elements in the training dataset. In\nconclusion, duplicated sentence elements are necessary for successful backdoor\nattacks. Based on this, we propose a data-centric defense. We first detect\ntrigger candidates in training data by finding memorizable elements, i.e.,\nduplicated elements, and then confirm real triggers by testing if the\ncandidates can activate backdoor behaviors (i.e., malicious elements). Results\nshow that our method outperforms state-of-the-art defenses in defending against\ndifferent types of NLP backdoors.\n", "date": "2024-09-21", "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": "2409.08985", "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": "2409.01193", "title": "CLIBE: Detecting Dynamic Backdoors in Transformer-based NLP Models", "abstract": " Backdoors can be injected into NLP models to induce misbehavior when the\ninput text contains a specific feature, known as a trigger, which the attacker\nsecretly selects. Unlike fixed words, phrases, or sentences used in the static\ntext trigger, NLP dynamic backdoor attacks design triggers associated with\nabstract and latent text features, making them considerably stealthier than\ntraditional static backdoor attacks. However, existing research on NLP backdoor\ndetection primarily focuses on defending against static backdoor attacks, while\ndetecting dynamic backdoors in NLP models remains largely unexplored. This\npaper presents CLIBE, the first framework to detect dynamic backdoors in\nTransformer-based NLP models. CLIBE injects a \"few-shot perturbation\" into the\nsuspect Transformer model by crafting optimized weight perturbation in the\nattention layers to make the perturbed model classify a limited number of\nreference samples as a target label. Subsequently, CLIBE leverages the\ngeneralization ability of this few-shot perturbation to determine whether the\noriginal model contains a dynamic backdoor. Extensive evaluation on three\nadvanced NLP dynamic backdoor attacks, two widely-used Transformer frameworks,\nand four real-world classification tasks strongly validates the effectiveness\nof CLIBE. We also demonstrate the robustness of CLIBE against various adaptive\nattacks. Furthermore, we employ CLIBE to scrutinize 49 popular Transformer\nmodels on Hugging Face and discover one exhibiting a high probability of\ncontaining a dynamic backdoor. We have contacted Hugging Face and provided\ndetailed evidence of this model's backdoor behavior. Moreover, we extend CLIBE\nto detect backdoor text generation models modified to exhibit toxic behavior.\nTo the best of our knowledge, CLIBE is the first framework capable of detecting\nbackdoors in text generation models without access to trigger input test\nsamples.\n", "date": "2024-09-02", "categories": []} {"id": "2405.06206", "title": "Concealing Backdoor Model Updates in Federated Learning by\n Trigger-Optimized Data Poisoning", "abstract": " Federated Learning (FL) is a decentralized machine learning method that\nenables participants to collaboratively train a model without sharing their\nprivate data. Despite its privacy and scalability benefits, FL is susceptible\nto backdoor attacks, where adversaries poison the local training data of a\nsubset of clients using a backdoor trigger, aiming to make the aggregated model\nproduce malicious results when the same backdoor condition is met by an\ninference-time input. Existing backdoor attacks in FL suffer from common\ndeficiencies: fixed trigger patterns and reliance on the assistance of model\npoisoning. State-of-the-art defenses based on analyzing clients' model updates\nexhibit a good defense performance on these attacks because of the significant\ndivergence between malicious and benign client model updates. To effectively\nconceal malicious model updates among benign ones, we propose DPOT, a backdoor\nattack strategy in FL that dynamically constructs backdoor objectives by\noptimizing a backdoor trigger, making backdoor data have minimal effect on\nmodel updates. We provide theoretical justifications for DPOT's attacking\nprinciple and display experimental results showing that DPOT, via only a\ndata-poisoning attack, effectively undermines state-of-the-art defenses and\noutperforms existing backdoor attack techniques on various datasets.\n", "date": "2024-05-10", "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": "2403.05847", "title": "iBA: Backdoor Attack on 3D Point Cloud via Reconstructing Itself", "abstract": " The widespread deployment of Deep Neural Networks (DNNs) for 3D point cloud\nprocessing starkly contrasts with their susceptibility to security breaches,\nnotably backdoor attacks. These attacks hijack DNNs during training, embedding\ntriggers in the data that, once activated, cause the network to make\npredetermined errors while maintaining normal performance on unaltered data.\nThis vulnerability poses significant risks, especially given the insufficient\nresearch on robust defense mechanisms for 3D point cloud networks against such\nsophisticated threats. Existing attacks either struggle to resist basic point\ncloud pre-processing methods, or rely on delicate manual design. Exploring\nsimple, effective, imperceptible, and difficult-to-defend triggers in 3D point\nclouds is still challenging.To address these challenges, we introduce\nMirrorAttack, a novel effective 3D backdoor attack method, which implants the\ntrigger by simply reconstructing a clean point cloud with an auto-encoder. The\ndata-driven nature of the MirrorAttack obviates the need for complex manual\ndesign. Minimizing the reconstruction loss automatically improves\nimperceptibility. Simultaneously, the reconstruction network endows the trigger\nwith pronounced nonlinearity and sample specificity, rendering traditional\npreprocessing techniques ineffective in eliminating it. A trigger smoothing\nmodule based on spherical harmonic transformation is also attached to regulate\nthe intensity of the attack.Both quantitive and qualitative results verify the\neffectiveness of our method. We achieve state-of-the-art ASR on different types\nof victim models with the intervention of defensive techniques. Moreover, the\nminimal perturbation introduced by our trigger, as assessed by various metrics,\nattests to the method's stealth, ensuring its imperceptibility.\n", "date": "2024-03-09", "categories": []} {"id": "2409.01952", "title": "Exploiting the Vulnerability of Large Language Models via Defense-Aware\n Architectural Backdoor", "abstract": " Deep neural networks (DNNs) have long been recognized as vulnerable to\nbackdoor attacks. By providing poisoned training data in the fine-tuning\nprocess, the attacker can implant a backdoor into the victim model. This\nenables input samples meeting specific textual trigger patterns to be\nclassified as target labels of the attacker's choice. While such black-box\nattacks have been well explored in both computer vision and natural language\nprocessing (NLP), backdoor attacks relying on white-box attack philosophy have\nhardly been thoroughly investigated. In this paper, we take the first step to\nintroduce a new type of backdoor attack that conceals itself within the\nunderlying model architecture. Specifically, we propose to design separate\nbackdoor modules consisting of two functions: trigger detection and noise\ninjection. The add-on modules of model architecture layers can detect the\npresence of input trigger tokens and modify layer weights using Gaussian noise\nto disturb the feature distribution of the baseline model. We conduct extensive\nexperiments to evaluate our attack methods using two model architecture\nsettings on five different large language datasets. We demonstrate that the\ntraining-free architectural backdoor on a large language model poses a genuine\nthreat. Unlike the-state-of-art work, it can survive the rigorous fine-tuning\nand retraining process, as well as evade output probability-based defense\nmethods (i.e. BDDR). All the code and data is available\nhttps://github.com/SiSL-URI/Arch_Backdoor_LLM.\n", "date": "2024-09-03", "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": "2409.04142", "title": "Context is the Key: Backdoor Attacks for In-Context Learning with Vision\n Transformers", "abstract": " Due to the high cost of training, large model (LM) practitioners commonly use\npretrained models downloaded from untrusted sources, which could lead to owning\ncompromised models. In-context learning is the ability of LMs to perform\nmultiple tasks depending on the prompt or context. This can enable new attacks,\nsuch as backdoor attacks with dynamic behavior depending on how models are\nprompted.\n In this paper, we leverage the ability of vision transformers (ViTs) to\nperform different tasks depending on the prompts. Then, through data poisoning,\nwe investigate two new threats: i) task-specific backdoors where the attacker\nchooses a target task to attack, and only the selected task is compromised at\ntest time under the presence of the trigger. At the same time, any other task\nis not affected, even if prompted with the trigger. We succeeded in attacking\nevery tested model, achieving up to 89.90\\% degradation on the target task. ii)\nWe generalize the attack, allowing the backdoor to affect \\emph{any} task, even\ntasks unseen during the training phase. Our attack was successful on every\ntested model, achieving a maximum of $13\\times$ degradation. Finally, we\ninvestigate the robustness of prompts and fine-tuning as techniques for\nremoving the backdoors from the model. We found that these methods fall short\nand, in the best case, reduce the degradation from 89.90\\% to 73.46\\%.\n", "date": "2024-09-06", "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": "2409.02251", "title": "NoiseAttack: An Evasive Sample-Specific Multi-Targeted Backdoor Attack\n Through White Gaussian Noise", "abstract": " Backdoor attacks pose a significant threat when using third-party data for\ndeep learning development. In these attacks, data can be manipulated to cause a\ntrained model to behave improperly when a specific trigger pattern is applied,\nproviding the adversary with unauthorized advantages. While most existing works\nfocus on designing trigger patterns in both visible and invisible to poison the\nvictim class, they typically result in a single targeted class upon the success\nof the backdoor attack, meaning that the victim class can only be converted to\nanother class based on the adversary predefined value. In this paper, we\naddress this issue by introducing a novel sample-specific multi-targeted\nbackdoor attack, namely NoiseAttack. Specifically, we adopt White Gaussian\nNoise (WGN) with various Power Spectral Densities (PSD) as our underlying\ntriggers, coupled with a unique training strategy to execute the backdoor\nattack. This work is the first of its kind to launch a vision backdoor attack\nwith the intent to generate multiple targeted classes with minimal input\nconfiguration. Furthermore, our extensive experimental results demonstrate that\nNoiseAttack can achieve a high attack success rate against popular network\narchitectures and datasets, as well as bypass state-of-the-art backdoor\ndetection methods. Our source code and experiments are available at\nhttps://github.com/SiSL-URI/NoiseAttack/tree/main.\n", "date": "2024-09-03", "categories": []} {"id": "2409.00863", "title": "Fisher Information guided Purification against Backdoor Attacks", "abstract": " Studies on backdoor attacks in recent years suggest that an adversary can\ncompromise the integrity of a deep neural network (DNN) by manipulating a small\nset of training samples. Our analysis shows that such manipulation can make the\nbackdoor model converge to a bad local minima, i.e., sharper minima as compared\nto a benign model. Intuitively, the backdoor can be purified by re-optimizing\nthe model to smoother minima. However, a na\\\"ive adoption of any optimization\ntargeting smoother minima can lead to sub-optimal purification techniques\nhampering the clean test accuracy. Hence, to effectively obtain such\nre-optimization, inspired by our novel perspective establishing the connection\nbetween backdoor removal and loss smoothness, we propose Fisher Information\nguided Purification (FIP), a novel backdoor purification framework. Proposed\nFIP consists of a couple of novel regularizers that aid the model in\nsuppressing the backdoor effects and retaining the acquired knowledge of clean\ndata distribution throughout the backdoor removal procedure through exploiting\nthe knowledge of Fisher Information Matrix (FIM). In addition, we introduce an\nefficient variant of FIP, dubbed as Fast FIP, which reduces the number of\ntunable parameters significantly and obtains an impressive runtime gain of\nalmost $5\\times$. Extensive experiments show that the proposed method achieves\nstate-of-the-art (SOTA) performance on a wide range of backdoor defense\nbenchmarks: 5 different tasks -- Image Recognition, Object Detection, Video\nAction Recognition, 3D point Cloud, Language Generation; 11 different datasets\nincluding ImageNet, PASCAL VOC, UCF101; diverse model architectures spanning\nboth CNN and vision transformer; 14 different backdoor attacks, e.g., Dynamic,\nWaNet, LIRA, ISSBA, etc.\n", "date": "2024-09-01", "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": "2408.15861", "title": "Fusing Pruned and Backdoored Models: Optimal Transport-based Data-free\n Backdoor Mitigation", "abstract": " Backdoor attacks present a serious security threat to deep neuron networks\n(DNNs). Although numerous effective defense techniques have been proposed in\nrecent years, they inevitably rely on the availability of either clean or\npoisoned data. In contrast, data-free defense techniques have evolved slowly\nand still lag significantly in performance. To address this issue, different\nfrom the traditional approach of pruning followed by fine-tuning, we propose a\nnovel data-free defense method named Optimal Transport-based Backdoor Repairing\n(OTBR) in this work. This method, based on our findings on neuron weight\nchanges (NWCs) of random unlearning, uses optimal transport (OT)-based model\nfusion to combine the advantages of both pruned and backdoored models.\nSpecifically, we first demonstrate our findings that the NWCs of random\nunlearning are positively correlated with those of poison unlearning. Based on\nthis observation, we propose a random-unlearning NWC pruning technique to\neliminate the backdoor effect and obtain a backdoor-free pruned model. Then,\nmotivated by the OT-based model fusion, we propose the pruned-to-backdoored\nOT-based fusion technique, which fuses pruned and backdoored models to combine\nthe advantages of both, resulting in a model that demonstrates high clean\naccuracy and a low attack success rate. To our knowledge, this is the first\nwork to apply OT and model fusion techniques to backdoor defense. Extensive\nexperiments show that our method successfully defends against all seven\nbackdoor attacks across three benchmark datasets, outperforming both\nstate-of-the-art (SOTA) data-free and data-dependent methods. The code\nimplementation and Appendix are provided in the Supplementary Material.\n", "date": "2024-08-28", "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": "2408.13221", "title": "Protecting against simultaneous data poisoning attacks", "abstract": " Current backdoor defense methods are evaluated against a single attack at a\ntime. This is unrealistic, as powerful machine learning systems are trained on\nlarge datasets scraped from the internet, which may be attacked multiple times\nby one or more attackers. We demonstrate that simultaneously executed data\npoisoning attacks can effectively install multiple backdoors in a single model\nwithout substantially degrading clean accuracy. Furthermore, we show that\nexisting backdoor defense methods do not effectively prevent attacks in this\nsetting. Finally, we leverage insights into the nature of backdoor attacks to\ndevelop a new defense, BaDLoss, that is effective in the multi-attack setting.\nWith minimal clean accuracy degradation, BaDLoss attains an average attack\nsuccess rate in the multi-attack setting of 7.98% in CIFAR-10 and 10.29% in\nGTSRB, compared to the average of other defenses at 64.48% and 84.28%\nrespectively.\n", "date": "2024-08-23", "categories": []} {"id": "2406.07778", "title": "A Study of Backdoors in Instruction Fine-tuned Language Models", "abstract": " Backdoor data poisoning, inserted within instruction examples used to\nfine-tune a foundation Large Language Model (LLM) for downstream tasks\n(\\textit{e.g.,} sentiment prediction), is a serious security concern due to the\nevasive nature of such attacks. The poisoning is usually in the form of a\n(seemingly innocuous) trigger word or phrase inserted into a very small\nfraction of the fine-tuning samples from a target class. Such backdoor attacks\ncan: alter response sentiment, violate censorship, over-refuse (invoke\ncensorship for legitimate queries), inject false content, or trigger nonsense\nresponses (hallucinations). In this work we investigate the efficacy of\ninstruction fine-tuning backdoor attacks as attack \"hyperparameters\" are varied\nunder a variety of scenarios, considering: the trigger location in the poisoned\nexamples; robustness to change in the trigger location, partial triggers, and\nsynonym substitutions at test time; attack transfer from one (fine-tuning)\ndomain to a related test domain; and clean-label vs. dirty-label poisoning.\nBased on our observations, we propose and evaluate two defenses against these\nattacks: i) a \\textit{during-fine-tuning defense} based on word-frequency\ncounts that assumes the (possibly poisoned) fine-tuning dataset is available\nand identifies the backdoor trigger tokens; and ii) a \\textit{post-fine-tuning\ndefense} based on downstream clean fine-tuning of the backdoored LLM with a\nsmall defense dataset. Finally, we provide a brief survey of related work on\nbackdoor attacks and defenses.\n", "date": "2024-06-12", "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": "2408.12312", "title": "MakeupAttack: Feature Space Black-box Backdoor Attack on Face\n Recognition via Makeup Transfer", "abstract": " Backdoor attacks pose a significant threat to the training process of deep\nneural networks (DNNs). As a widely-used DNN-based application in real-world\nscenarios, face recognition systems once implanted into the backdoor, may cause\nserious consequences. Backdoor research on face recognition is still in its\nearly stages, and the existing backdoor triggers are relatively simple and\nvisible. Furthermore, due to the perceptibility, diversity, and similarity of\nfacial datasets, many state-of-the-art backdoor attacks lose effectiveness on\nface recognition tasks. In this work, we propose a novel feature space backdoor\nattack against face recognition via makeup transfer, dubbed MakeupAttack. In\ncontrast to many feature space attacks that demand full access to target\nmodels, our method only requires model queries, adhering to black-box attack\nprinciples. In our attack, we design an iterative training paradigm to learn\nthe subtle features of the proposed makeup-style trigger. Additionally,\nMakeupAttack promotes trigger diversity using the adaptive selection method,\ndispersing the feature distribution of malicious samples to bypass existing\ndefense methods. Extensive experiments were conducted on two widely-used facial\ndatasets targeting multiple models. The results demonstrate that our proposed\nattack method can bypass existing state-of-the-art defenses while maintaining\neffectiveness, robustness, naturalness, and stealthiness, without compromising\nmodel performance.\n", "date": "2024-08-22", "categories": []} {"id": "2408.11587", "title": "Large Language Models are Good Attackers: Efficient and Stealthy Textual\n Backdoor Attacks", "abstract": " With the burgeoning advancements in the field of natural language processing\n(NLP), the demand for training data has increased significantly. To save costs,\nit has become common for users and businesses to outsource the labor-intensive\ntask of data collection to third-party entities. Unfortunately, recent research\nhas unveiled the inherent risk associated with this practice, particularly in\nexposing NLP systems to potential backdoor attacks. Specifically, these attacks\nenable malicious control over the behavior of a trained model by poisoning a\nsmall portion of the training data. Unlike backdoor attacks in computer vision,\ntextual backdoor attacks impose stringent requirements for attack stealthiness.\nHowever, existing attack methods meet significant trade-off between\neffectiveness and stealthiness, largely due to the high information entropy\ninherent in textual data. In this paper, we introduce the Efficient and\nStealthy Textual backdoor attack method, EST-Bad, leveraging Large Language\nModels (LLMs). Our EST-Bad encompasses three core strategies: optimizing the\ninherent flaw of models as the trigger, stealthily injecting triggers with\nLLMs, and meticulously selecting the most impactful samples for backdoor\ninjection. Through the integration of these techniques, EST-Bad demonstrates an\nefficient achievement of competitive attack performance while maintaining\nsuperior stealthiness compared to prior methods across various text classifier\ndatasets.\n", "date": "2024-08-21", "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": "2408.07440", "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": "2401.15002", "title": "BackdoorBench: A Comprehensive Benchmark and Analysis of Backdoor\n Learning", "abstract": " As an emerging and vital topic for studying deep neural networks'\nvulnerability (DNNs), backdoor learning has attracted increasing interest in\nrecent years, and many seminal backdoor attack and defense algorithms are being\ndeveloped successively or concurrently, in the status of a rapid arms race.\nHowever, mainly due to the diverse settings, and the difficulties of\nimplementation and reproducibility of existing works, there is a lack of a\nunified and standardized benchmark of backdoor learning, causing unfair\ncomparisons, and unreliable conclusions (e.g., misleading, biased or even false\nconclusions). Consequently, it is difficult to evaluate the current progress\nand design the future development roadmap of this literature. To alleviate this\ndilemma, we build a comprehensive benchmark of backdoor learning called\nBackdoorBench. Our benchmark makes three valuable contributions to the research\ncommunity. 1) We provide an integrated implementation of state-of-the-art\n(SOTA) backdoor learning algorithms (currently including 16 attack and 27\ndefense algorithms), based on an extensible modular-based codebase. 2) We\nconduct comprehensive evaluations of 12 attacks against 16 defenses, with 5\npoisoning ratios, based on 4 models and 4 datasets, thus 11,492 pairs of\nevaluations in total. 3) Based on above evaluations, we present abundant\nanalysis from 8 perspectives via 18 useful analysis tools, and provide several\ninspiring insights about backdoor learning. We hope that our efforts could\nbuild a solid foundation of backdoor learning to facilitate researchers to\ninvestigate existing algorithms, develop more innovative algorithms, and\nexplore the intrinsic mechanism of backdoor learning. Finally, we have created\na user-friendly website at http://backdoorbench.com, which collects all\nimportant information of BackdoorBench, including codebase, docs, leaderboard,\nand model Zoo.\n", "date": "2024-01-26", "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": "2407.19845", "title": "BackdoorBench: A Comprehensive Benchmark and Analysis of Backdoor\n Learning", "abstract": " As an emerging approach to explore the vulnerability of deep neural networks\n(DNNs), backdoor learning has attracted increasing interest in recent years,\nand many seminal backdoor attack and defense algorithms are being developed\nsuccessively or concurrently, in the status of a rapid arms race. However,\nmainly due to the diverse settings, and the difficulties of implementation and\nreproducibility of existing works, there is a lack of a unified and\nstandardized benchmark of backdoor learning, causing unfair comparisons or\nunreliable conclusions (e.g., misleading, biased or even false conclusions).\nConsequently, it is difficult to evaluate the current progress and design the\nfuture development roadmap of this literature. To alleviate this dilemma, we\nbuild a comprehensive benchmark of backdoor learning called BackdoorBench. Our\nbenchmark makes three valuable contributions to the research community. 1) We\nprovide an integrated implementation of state-of-the-art (SOTA) backdoor\nlearning algorithms (currently including 20 attack and 32 defense algorithms),\nbased on an extensible modular-based codebase. 2) We conduct comprehensive\nevaluations with 5 poisoning ratios, based on 4 models and 4 datasets, leading\nto 11,492 pairs of attack-against-defense evaluations in total. 3) Based on\nabove evaluations, we present abundant analysis from 10 perspectives via 18\nuseful analysis tools, and provide several inspiring insights about backdoor\nlearning. We hope that our efforts could build a solid foundation of backdoor\nlearning to facilitate researchers to investigate existing algorithms, develop\nmore innovative algorithms, and explore the intrinsic mechanism of backdoor\nlearning. Finally, we have created a user-friendly website at\nhttp://backdoorbench.com, which collects all important information of\nBackdoorBench, including codebase, docs, leaderboard, and model Zoo.\n", "date": "2024-07-29", "categories": []} {"id": "2310.08320", "title": "Defending Our Privacy With Backdoors", "abstract": " The proliferation of large AI models trained on uncurated, often sensitive\nweb-scraped data has raised significant privacy concerns. One of the concerns\nis that adversaries can extract information about the training data using\nprivacy attacks. Unfortunately, the task of removing specific information from\nthe models without sacrificing performance is not straightforward and has\nproven to be challenging. We propose a rather easy yet effective defense based\non backdoor attacks to remove private information, such as names and faces of\nindividuals, from vision-language models by fine-tuning them for only a few\nminutes instead of re-training them from scratch. Specifically, by\nstrategically inserting backdoors into text encoders, we align the embeddings\nof sensitive phrases with those of neutral terms-\"a person\" instead of the\nperson's actual name. For image encoders, we map individuals' embeddings to be\nremoved from the model to a universal, anonymous embedding. The results of our\nextensive experimental evaluation demonstrate the effectiveness of our\nbackdoor-based defense on CLIP by assessing its performance using a specialized\nprivacy attack for zero-shot classifiers. Our approach provides a new\n\"dual-use\" perspective on backdoor attacks and presents a promising avenue to\nenhance the privacy of individuals within models trained on uncurated\nweb-scraped data.\n", "date": "2023-10-12", "categories": []} {"id": "2301.12318", "title": "Gradient Shaping: Enhancing Backdoor Attack Against Reverse Engineering", "abstract": " Most existing methods to detect backdoored machine learning (ML) models take\none of the two approaches: trigger inversion (aka. reverse engineer) and weight\nanalysis (aka. model diagnosis). In particular, the gradient-based trigger\ninversion is considered to be among the most effective backdoor detection\ntechniques, as evidenced by the TrojAI competition, Trojan Detection Challenge\nand backdoorBench. However, little has been done to understand why this\ntechnique works so well and, more importantly, whether it raises the bar to the\nbackdoor attack. In this paper, we report the first attempt to answer this\nquestion by analyzing the change rate of the backdoored model around its\ntrigger-carrying inputs. Our study shows that existing attacks tend to inject\nthe backdoor characterized by a low change rate around trigger-carrying inputs,\nwhich are easy to capture by gradient-based trigger inversion. In the meantime,\nwe found that the low change rate is not necessary for a backdoor attack to\nsucceed: we design a new attack enhancement called \\textit{Gradient Shaping}\n(GRASP), which follows the opposite direction of adversarial training to reduce\nthe change rate of a backdoored model with regard to the trigger, without\nundermining its backdoor effect. Also, we provide a theoretic analysis to\nexplain the effectiveness of this new technique and the fundamental weakness of\ngradient-based trigger inversion. Finally, we perform both theoretical and\nexperimental analysis, showing that the GRASP enhancement does not reduce the\neffectiveness of the stealthy attacks against the backdoor detection methods\nbased on weight analysis, as well as other backdoor mitigation methods without\nusing detection.\n", "date": "2023-01-29", "categories": []} {"id": "2407.14738", "title": "Flatness-aware Sequential Learning Generates Resilient Backdoors", "abstract": " Recently, backdoor attacks have become an emerging threat to the security of\nmachine learning models. From the adversary's perspective, the implanted\nbackdoors should be resistant to defensive algorithms, but some recently\nproposed fine-tuning defenses can remove these backdoors with notable efficacy.\nThis is mainly due to the catastrophic forgetting (CF) property of deep neural\nnetworks. This paper counters CF of backdoors by leveraging continual learning\n(CL) techniques. We begin by investigating the connectivity between a\nbackdoored and fine-tuned model in the loss landscape. Our analysis confirms\nthat fine-tuning defenses, especially the more advanced ones, can easily push a\npoisoned model out of the backdoor regions, making it forget all about the\nbackdoors. Based on this finding, we re-formulate backdoor training through the\nlens of CL and propose a novel framework, named Sequential Backdoor Learning\n(SBL), that can generate resilient backdoors. This framework separates the\nbackdoor poisoning process into two tasks: the first task learns a backdoored\nmodel, while the second task, based on the CL principles, moves it to a\nbackdoored region resistant to fine-tuning. We additionally propose to seek\nflatter backdoor regions via a sharpness-aware minimizer in the framework,\nfurther strengthening the durability of the implanted backdoor. Finally, we\ndemonstrate the effectiveness of our method through extensive empirical\nexperiments on several benchmark datasets in the backdoor domain. The source\ncode is available at https://github.com/mail-research/SBL-resilient-backdoors\n", "date": "2024-07-20", "categories": []} {"id": "2407.15168", "title": "Mitigating Deep Reinforcement Learning Backdoors in the Neural\n Activation Space", "abstract": " This paper investigates the threat of backdoors in Deep Reinforcement\nLearning (DRL) agent policies and proposes a novel method for their detection\nat runtime. Our study focuses on elusive in-distribution backdoor triggers.\nSuch triggers are designed to induce a deviation in the behaviour of a\nbackdoored agent while blending into the expected data distribution to evade\ndetection. Through experiments conducted in the Atari Breakout environment, we\ndemonstrate the limitations of current sanitisation methods when faced with\nsuch triggers and investigate why they present a challenging defence problem.\nWe then evaluate the hypothesis that backdoor triggers might be easier to\ndetect in the neural activation space of the DRL agent's policy network. Our\nstatistical analysis shows that indeed the activation patterns in the agent's\npolicy network are distinct in the presence of a trigger, regardless of how\nwell the trigger is concealed in the environment. Based on this, we propose a\nnew defence approach that uses a classifier trained on clean environment\nsamples and detects abnormal activations. Our results show that even\nlightweight classifiers can effectively prevent malicious actions with\nconsiderable accuracy, indicating the potential of this research direction even\nagainst sophisticated adversaries.\n", "date": "2024-07-21", "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": "2407.10052", "title": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", "abstract": " Recent studies have revealed the vulnerability of deep neural networks (DNNs)\nto various backdoor attacks, where the behavior of DNNs can be compromised by\nutilizing certain types of triggers or poisoning mechanisms. State-of-the-art\n(SOTA) defenses employ too-sophisticated mechanisms that require either a\ncomputationally expensive adversarial search module for reverse-engineering the\ntrigger distribution or an over-sensitive hyper-parameter selection module.\nMoreover, they offer sub-par performance in challenging scenarios, e.g.,\nlimited validation data and strong attacks. In this paper, we propose Neural\nmask Fine-Tuning (NFT) with an aim to optimally re-organize the neuron\nactivities in a way that the effect of the backdoor is removed. Utilizing a\nsimple data augmentation like MixUp, NFT relaxes the trigger synthesis process\nand eliminates the requirement of the adversarial search module. Our study\nfurther reveals that direct weight fine-tuning under limited validation data\nresults in poor post-purification clean test accuracy, primarily due to\noverfitting issue. To overcome this, we propose to fine-tune neural masks\ninstead of model weights. In addition, a mask regularizer has been devised to\nfurther mitigate the model drift during the purification process. The distinct\ncharacteristics of NFT render it highly efficient in both runtime and sample\nusage, as it can remove the backdoor even when a single sample is available\nfrom each class. We validate the effectiveness of NFT through extensive\nexperiments covering the tasks of image classification, object detection, video\naction recognition, 3D point cloud, and natural language processing. We\nevaluate our method against 14 different attacks (LIRA, WaNet, etc.) on 11\nbenchmark data sets such as ImageNet, UCF101, Pascal VOC, ModelNet,\nOpenSubtitles2012, etc.\n", "date": "2024-07-14", "categories": []} {"id": "2407.12784", "title": "AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge\n Bases", "abstract": " LLM agents have demonstrated remarkable performance across various\napplications, primarily due to their advanced capabilities in reasoning,\nutilizing external knowledge and tools, calling APIs, and executing actions to\ninteract with environments. Current agents typically utilize a memory module or\na retrieval-augmented generation (RAG) mechanism, retrieving past knowledge and\ninstances with similar embeddings from knowledge bases to inform task planning\nand execution. However, the reliance on unverified knowledge bases raises\nsignificant concerns about their safety and trustworthiness. To uncover such\nvulnerabilities, we propose a novel red teaming approach AgentPoison, the first\nbackdoor attack targeting generic and RAG-based LLM agents by poisoning their\nlong-term memory or RAG knowledge base. In particular, we form the trigger\ngeneration process as a constrained optimization to optimize backdoor triggers\nby mapping the triggered instances to a unique embedding space, so as to ensure\nthat whenever a user instruction contains the optimized backdoor trigger, the\nmalicious demonstrations are retrieved from the poisoned memory or knowledge\nbase with high probability. In the meantime, benign instructions without the\ntrigger will still maintain normal performance. Unlike conventional backdoor\nattacks, AgentPoison requires no additional model training or fine-tuning, and\nthe optimized backdoor trigger exhibits superior transferability, in-context\ncoherence, and stealthiness. Extensive experiments demonstrate AgentPoison's\neffectiveness in attacking three types of real-world LLM agents: RAG-based\nautonomous driving agent, knowledge-intensive QA agent, and healthcare\nEHRAgent. On each agent, AgentPoison achieves an average attack success rate\nhigher than 80% with minimal impact on benign performance (less than 1%) with a\npoison rate less than 0.1%.\n", "date": "2024-07-17", "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": "2407.11372", "title": "UNIT: Backdoor Mitigation via Automated Neural Distribution Tightening", "abstract": " Deep neural networks (DNNs) have demonstrated effectiveness in various\nfields. However, DNNs are vulnerable to backdoor attacks, which inject a unique\npattern, called trigger, into the input to cause misclassification to an\nattack-chosen target label. While existing works have proposed various methods\nto mitigate backdoor effects in poisoned models, they tend to be less effective\nagainst recent advanced attacks. In this paper, we introduce a novel\npost-training defense technique UNIT that can effectively eliminate backdoor\neffects for a variety of attacks. In specific, UNIT approximates a unique and\ntight activation distribution for each neuron in the model. It then proactively\ndispels substantially large activation values that exceed the approximated\nboundaries. Our experimental results demonstrate that UNIT outperforms 7\npopular defense methods against 14 existing backdoor attacks, including 2\nadvanced attacks, using only 5\\% of clean training data. UNIT is also cost\nefficient. The code is accessible at https://github.com/Megum1/UNIT.\n", "date": "2024-07-16", "categories": []} {"id": "2407.05396", "title": "Evolutionary Trigger Detection and Lightweight Model Repair Based\n Backdoor Defense", "abstract": " Deep Neural Networks (DNNs) have been widely used in many areas such as\nautonomous driving and face recognition. However, DNN model is fragile to\nbackdoor attack. A backdoor in the DNN model can be activated by a poisoned\ninput with trigger and leads to wrong prediction, which causes serious security\nissues in applications. It is challenging for current defenses to eliminate the\nbackdoor effectively with limited computing resources, especially when the\nsizes and numbers of the triggers are variable as in the physical world. We\npropose an efficient backdoor defense based on evolutionary trigger detection\nand lightweight model repair. In the first phase of our method, CAM-focus\nEvolutionary Trigger Filter (CETF) is proposed for trigger detection. CETF is\nan effective sample-preprocessing based method with the evolutionary algorithm,\nand our experimental results show that CETF not only distinguishes the images\nwith triggers accurately from the clean images, but also can be widely used in\npractice for its simplicity and stability in different backdoor attack\nsituations. In the second phase of our method, we leverage several lightweight\nunlearning methods with the trigger detected by CETF for model repair, which\nalso constructively demonstrate the underlying correlation of the backdoor with\nBatch Normalization layers. Source code will be published after accepted.\n", "date": "2024-07-07", "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": "2407.04179", "title": "Defense Against Syntactic Textual Backdoor Attacks with Token\n Substitution", "abstract": " Textual backdoor attacks present a substantial security risk to Large\nLanguage Models (LLM). It embeds carefully chosen triggers into a victim model\nat the training stage, and makes the model erroneously predict inputs\ncontaining the same triggers as a certain class. Prior backdoor defense methods\nprimarily target special token-based triggers, leaving syntax-based triggers\ninsufficiently addressed. To fill this gap, this paper proposes a novel online\ndefense algorithm that effectively counters syntax-based as well as special\ntoken-based backdoor attacks. The algorithm replaces semantically meaningful\nwords in sentences with entirely different ones but preserves the syntactic\ntemplates or special tokens, and then compares the predicted labels before and\nafter the substitution to determine whether a sentence contains triggers.\nExperimental results confirm the algorithm's performance against these two\ntypes of triggers, offering a comprehensive defense strategy for model\nintegrity.\n", "date": "2024-07-04", "categories": []} {"id": "2303.01504", "title": "Backdoor for Debias: Mitigating Model Bias with Backdoor Attack-based\n Artificial Bias", "abstract": " With the swift advancement of deep learning, state-of-the-art algorithms have\nbeen utilized in various social situations. Nonetheless, some algorithms have\nbeen discovered to exhibit biases and provide unequal results. The current\ndebiasing methods face challenges such as poor utilization of data or intricate\ntraining requirements. In this work, we found that the backdoor attack can\nconstruct an artificial bias similar to the model bias derived in standard\ntraining. Considering the strong adjustability of backdoor triggers, we are\nmotivated to mitigate the model bias by carefully designing reverse artificial\nbias created from backdoor attack. Based on this, we propose a backdoor\ndebiasing framework based on knowledge distillation, which effectively reduces\nthe model bias from original data and minimizes security risks from the\nbackdoor attack. The proposed solution is validated on both image and\nstructured datasets, showing promising results. This work advances the\nunderstanding of backdoor attacks and highlights its potential for beneficial\napplications. The code for the study can be found at\n\\url{https://anonymous.4open.science/r/DwB-BC07/}.\n", "date": "2023-03-01", "categories": []} {"id": "2407.00719", "title": "A Whole-Process Certifiably Robust Aggregation Method Against Backdoor\n Attacks in Federated Learning", "abstract": " Federated Learning (FL) has garnered widespread adoption across various\ndomains such as finance, healthcare, and cybersecurity. Nonetheless, FL remains\nunder significant threat from backdoor attacks, wherein malicious actors insert\ntriggers into trained models, enabling them to perform certain tasks while\nstill meeting FL's primary objectives. In response, robust aggregation methods\nhave been proposed, which can be divided into three types: ex-ante, ex-durante,\nand ex-post methods. Given the complementary nature of these methods, combining\nall three types is promising yet unexplored. Such a combination is non-trivial\nbecause it requires leveraging their advantages while overcoming their\ndisadvantages. Our study proposes a novel whole-process certifiably robust\naggregation (WPCRA) method for FL, which enhances robustness against backdoor\nattacks across three phases: ex-ante, ex-durante, and ex-post. Moreover, since\nthe current geometric median estimation method fails to consider differences\namong clients, we propose a novel weighted geometric median estimation\nalgorithm (WGME). This algorithm estimates the geometric median of model\nupdates from clients based on each client's weight, further improving the\nrobustness of WPCRA against backdoor attacks. We also theoretically prove that\nWPCRA offers improved certified robustness guarantees with a larger certified\nradius. We evaluate the advantages of our methods based on the task of loan\nstatus prediction. Comparison with baselines shows that our methods\nsignificantly improve FL's robustness against backdoor attacks. This study\ncontributes to the literature with a novel WPCRA method and a novel WGME\nalgorithm. Our code is available at https://github.com/brick-brick/WPCRAM.\n", "date": "2024-06-30", "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": "2406.16125", "title": "CBPF: Filtering Poisoned Data Based on Composite Backdoor Attack", "abstract": " Backdoor attacks involve the injection of a limited quantity of poisoned\nexamples containing triggers into the training dataset. During the inference\nstage, backdoor attacks can uphold a high level of accuracy for normal\nexamples, yet when presented with trigger-containing instances, the model may\nerroneously predict them as the targeted class designated by the attacker. This\npaper explores strategies for mitigating the risks associated with backdoor\nattacks by examining the filtration of poisoned samples.We primarily leverage\ntwo key characteristics of backdoor attacks: the ability for multiple backdoors\nto exist simultaneously within a single model, and the discovery through\nComposite Backdoor Attack (CBA) that altering two triggers in a sample to new\ntarget labels does not compromise the original functionality of the triggers,\nyet enables the prediction of the data as a new target class when both triggers\nare present simultaneously.Therefore, a novel three-stage poisoning data\nfiltering approach, known as Composite Backdoor Poison Filtering (CBPF), is\nproposed as an effective solution. Firstly, utilizing the identified\ndistinctions in output between poisoned and clean samples, a subset of data is\npartitioned to include both poisoned and clean instances. Subsequently, benign\ntriggers are incorporated and labels are adjusted to create new target and\nbenign target classes, thereby prompting the poisoned and clean data to be\nclassified as distinct entities during the inference stage. The experimental\nresults indicate that CBPF is successful in filtering out malicious data\nproduced by six advanced attacks on CIFAR10 and ImageNet-12. On average, CBPF\nattains a notable filtering success rate of 99.91% for the six attacks on\nCIFAR10. Additionally, the model trained on the uncontaminated samples exhibits\nsustained high accuracy levels.\n", "date": "2024-06-23", "categories": []} {"id": "2310.00542", "title": "Watch Out! Simple Horizontal Class Backdoor Can Trivially Evade Defense", "abstract": " All current backdoor attacks on deep learning (DL) models fall under the\ncategory of a vertical class backdoor (VCB) -- class-dependent. In VCB attacks,\nany sample from a class activates the implanted backdoor when the secret\ntrigger is present. Existing defense strategies overwhelmingly focus on\ncountering VCB attacks, especially those that are source-class-agnostic. This\nnarrow focus neglects the potential threat of other simpler yet general\nbackdoor types, leading to false security implications. This study introduces a\nnew, simple, and general type of backdoor attack coined as the horizontal class\nbackdoor (HCB) that trivially breaches the class dependence characteristic of\nthe VCB, bringing a fresh perspective to the community. HCB is now activated\nwhen the trigger is presented together with an innocuous feature, regardless of\nclass. For example, the facial recognition model misclassifies a person who\nwears sunglasses with a smiling innocuous feature into the targeted person,\nsuch as an administrator, regardless of which person. The key is that these\ninnocuous features are horizontally shared among classes but are only exhibited\nby partial samples per class. Extensive experiments on attacking performance\nacross various tasks, including MNIST, facial recognition, traffic sign\nrecognition, object detection, and medical diagnosis, confirm the high\nefficiency and effectiveness of the HCB. We rigorously evaluated the\nevasiveness of the HCB against a series of eleven representative\ncountermeasures, including Fine-Pruning (RAID 18'), STRIP (ACSAC 19'), Neural\nCleanse (Oakland 19'), ABS (CCS 19'), Februus (ACSAC 20'), NAD (ICLR 21'), MNTD\n(Oakland 21'), SCAn (USENIX SEC 21'), MOTH (Oakland 22'), Beatrix (NDSS 23'),\nand MM-BD (Oakland 24'). None of these countermeasures prove robustness, even\nwhen employing a simplistic trigger, such as a small and static white-square\npatch.\n", "date": "2023-10-01", "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": "2301.03118", "title": "Facial Misrecognition Systems: Simple Weight Manipulations Force DNNs to\n Err Only on Specific Persons", "abstract": " In this paper, we describe how to plant novel types of backdoors in any\nfacial recognition model based on the popular architecture of deep Siamese\nneural networks. These backdoors force the system to err only on natural images\nof specific persons who are preselected by the attacker, without controlling\ntheir appearance or inserting any triggers. For example, we show how such a\nbackdoored system can classify any two images of a particular person as\ndifferent people, or any two images of a particular pair of persons as the same\nperson, with almost no effect on the correctness of its decisions for other\npersons. Surprisingly, we show that both types of backdoors can be implemented\nby applying linear transformations to the model's last weight matrix, with no\nadditional training or optimization, using only images of the backdoor\nidentities. A unique property of our attack is that multiple backdoors can be\nindependently installed in the same model by multiple attackers, who may not be\naware of each other's existence, with almost no interference. We have\nexperimentally verified the attacks on a SOTA facial recognition system. When\nwe tried to individually anonymize ten celebrities, the network failed to\nrecognize two of their images as being the same person in $97.02\\%$ to\n$98.31\\%$ of the time. When we tried to confuse between the extremely\ndifferent-looking Morgan Freeman and Scarlett Johansson, for example, their\nimages were declared to be the same person in $98.47 \\%$ of the time. For each\ntype of backdoor, we sequentially installed multiple backdoors with minimal\neffect on the performance of each other (for example, anonymizing all ten\ncelebrities on the same model reduced the success rate for each celebrity by no\nmore than $1.01\\%$). In all of our experiments, the benign accuracy of the\nnetwork on other persons barely degraded (in most cases, it degraded by less\nthan $0.05\\%$).\n", "date": "2023-01-08", "categories": []} {"id": "2312.00027", "title": "Stealthy and Persistent Unalignment on Large Language Models via\n Backdoor Injections", "abstract": " Recent developments in Large Language Models (LLMs) have manifested\nsignificant advancements. To facilitate safeguards against malicious\nexploitation, a body of research has concentrated on aligning LLMs with human\npreferences and inhibiting their generation of inappropriate content.\nUnfortunately, such alignments are often vulnerable: fine-tuning with a minimal\namount of harmful data can easily unalign the target LLM. While being\neffective, such fine-tuning-based unalignment approaches also have their own\nlimitations: (1) non-stealthiness, after fine-tuning, safety audits or\nred-teaming can easily expose the potential weaknesses of the unaligned models,\nthereby precluding their release/use. (2) non-persistence, the unaligned LLMs\ncan be easily repaired through re-alignment, i.e., fine-tuning again with\naligned data points. In this work, we show that it is possible to conduct\nstealthy and persistent unalignment on large language models via backdoor\ninjections. We also provide a novel understanding on the relationship between\nthe backdoor persistence and the activation pattern and further provide\nguidelines for potential trigger design. Through extensive experiments, we\ndemonstrate that our proposed stealthy and persistent unalignment can\nsuccessfully pass the safety evaluation while maintaining strong persistence\nagainst re-alignment defense.\n", "date": "2023-11-15", "categories": []} {"id": "2403.01472", "title": "WARDEN: Multi-Directional Backdoor Watermarks for Embedding-as-a-Service\n Copyright Protection", "abstract": " Embedding as a Service (EaaS) has become a widely adopted solution, which\noffers feature extraction capabilities for addressing various downstream tasks\nin Natural Language Processing (NLP). Prior studies have shown that EaaS can be\nprone to model extraction attacks; nevertheless, this concern could be\nmitigated by adding backdoor watermarks to the text embeddings and subsequently\nverifying the attack models post-publication. Through the analysis of the\nrecent watermarking strategy for EaaS, EmbMarker, we design a novel CSE\n(Clustering, Selection, Elimination) attack that removes the backdoor watermark\nwhile maintaining the high utility of embeddings, indicating that the previous\nwatermarking approach can be breached. In response to this new threat, we\npropose a new protocol to make the removal of watermarks more challenging by\nincorporating multiple possible watermark directions. Our defense approach,\nWARDEN, notably increases the stealthiness of watermarks and has been\nempirically shown to be effective against CSE attack.\n", "date": "2024-03-03", "categories": []} {"id": "2406.06207", "title": "Lurking in the shadows: Unveiling Stealthy Backdoor Attacks against\n Personalized Federated Learning", "abstract": " Federated Learning (FL) is a collaborative machine learning technique where\nmultiple clients work together with a central server to train a global model\nwithout sharing their private data. However, the distribution shift across\nnon-IID datasets of clients poses a challenge to this one-model-fits-all method\nhindering the ability of the global model to effectively adapt to each client's\nunique local data. To echo this challenge, personalized FL (PFL) is designed to\nallow each client to create personalized local models tailored to their private\ndata. While extensive research has scrutinized backdoor risks in FL, it has\nremained underexplored in PFL applications. In this study, we delve deep into\nthe vulnerabilities of PFL to backdoor attacks. Our analysis showcases a tale\nof two cities. On the one hand, the personalization process in PFL can dilute\nthe backdoor poisoning effects injected into the personalized local models.\nFurthermore, PFL systems can also deploy both server-end and client-end defense\nmechanisms to strengthen the barrier against backdoor attacks. On the other\nhand, our study shows that PFL fortified with these defense methods may offer a\nfalse sense of security. We propose \\textit{PFedBA}, a stealthy and effective\nbackdoor attack strategy applicable to PFL systems. \\textit{PFedBA} ingeniously\naligns the backdoor learning task with the main learning task of PFL by\noptimizing the trigger generation process. Our comprehensive experiments\ndemonstrate the effectiveness of \\textit{PFedBA} in seamlessly embedding\ntriggers into personalized local models. \\textit{PFedBA} yields outstanding\nattack performance across 10 state-of-the-art PFL algorithms, defeating the\nexisting 6 defense mechanisms. Our study sheds light on the subtle yet potent\nbackdoor threats to PFL systems, urging the community to bolster defenses\nagainst emerging backdoor challenges.\n", "date": "2024-06-10", "categories": []} {"id": "2406.04478", "title": "PromptFix: Few-shot Backdoor Removal via Adversarial Prompt Tuning", "abstract": " Pre-trained language models (PLMs) have attracted enormous attention over the\npast few years with their unparalleled performances. Meanwhile, the soaring\ncost to train PLMs as well as their amazing generalizability have jointly\ncontributed to few-shot fine-tuning and prompting as the most popular training\nparadigms for natural language processing (NLP) models. Nevertheless, existing\nstudies have shown that these NLP models can be backdoored such that model\nbehavior is manipulated when trigger tokens are presented. In this paper, we\npropose PromptFix, a novel backdoor mitigation strategy for NLP models via\nadversarial prompt-tuning in few-shot settings. Unlike existing NLP backdoor\nremoval methods, which rely on accurate trigger inversion and subsequent model\nfine-tuning, PromptFix keeps the model parameters intact and only utilizes two\nextra sets of soft tokens which approximate the trigger and counteract it\nrespectively. The use of soft tokens and adversarial optimization eliminates\nthe need to enumerate possible backdoor configurations and enables an adaptive\nbalance between trigger finding and preservation of performance. Experiments\nwith various backdoor attacks validate the effectiveness of the proposed method\nand the performances when domain shift is present further shows PromptFix's\napplicability to models pretrained on unknown data source which is the common\ncase in prompt tuning scenarios.\n", "date": "2024-06-06", "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": "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": "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": "2406.00588", "title": "Generalization Bound and New Algorithm for Clean-Label Backdoor Attack", "abstract": " The generalization bound is a crucial theoretical tool for assessing the\ngeneralizability of learning methods and there exist vast literatures on\ngeneralizability of normal learning, adversarial learning, and data poisoning.\nUnlike other data poison attacks, the backdoor attack has the special property\nthat the poisoned triggers are contained in both the training set and the test\nset and the purpose of the attack is two-fold. To our knowledge, the\ngeneralization bound for the backdoor attack has not been established. In this\npaper, we fill this gap by deriving algorithm-independent generalization bounds\nin the clean-label backdoor attack scenario. Precisely, based on the goals of\nbackdoor attack, we give upper bounds for the clean sample population errors\nand the poison population errors in terms of the empirical error on the\npoisoned training dataset. Furthermore, based on the theoretical result, a new\nclean-label backdoor attack is proposed that computes the poisoning trigger by\ncombining adversarial noise and indiscriminate poison. We show its\neffectiveness in a variety of settings.\n", "date": "2024-06-02", "categories": []} {"id": "2402.12026", "title": "Acquiring Clean Language Models from Backdoor Poisoned Datasets by\n Downscaling Frequency Space", "abstract": " Despite the notable success of language models (LMs) in various natural\nlanguage processing (NLP) tasks, the reliability of LMs is susceptible to\nbackdoor attacks. Prior research attempts to mitigate backdoor learning while\ntraining the LMs on the poisoned dataset, yet struggles against complex\nbackdoor attacks in real-world scenarios. In this paper, we investigate the\nlearning mechanisms of backdoor LMs in the frequency space by Fourier analysis.\nOur findings indicate that the backdoor mapping presented on the poisoned\ndatasets exhibits a more discernible inclination towards lower frequency\ncompared to clean mapping, resulting in the faster convergence of backdoor\nmapping. To alleviate this dilemma, we propose Multi-Scale Low-Rank Adaptation\n(MuScleLoRA), which deploys multiple radial scalings in the frequency space\nwith low-rank adaptation to the target model and further aligns the gradients\nwhen updating parameters. Through downscaling in the frequency space,\nMuScleLoRA encourages the model to prioritize the learning of relatively\nhigh-frequency clean mapping, consequently mitigating backdoor learning.\nExperimental results demonstrate that MuScleLoRA outperforms baselines\nsignificantly. Notably, MuScleLoRA reduces the average success rate of diverse\nbackdoor attacks to below 15\\% across multiple datasets and generalizes to\nvarious backbone LMs, including BERT, RoBERTa, GPT2-XL, and Llama2. The codes\nare publicly available at https://github.com/ZrW00/MuScleLoRA.\n", "date": "2024-02-19", "categories": []} {"id": "2402.19334", "title": "Here's a Free Lunch: Sanitizing Backdoored Models with Model Merge", "abstract": " The democratization of pre-trained language models through open-source\ninitiatives has rapidly advanced innovation and expanded access to cutting-edge\ntechnologies. However, this openness also brings significant security risks,\nincluding backdoor attacks, where hidden malicious behaviors are triggered by\nspecific inputs, compromising natural language processing (NLP) system\nintegrity and reliability. This paper suggests that merging a backdoored model\nwith other homogeneous models can significantly remediate backdoor\nvulnerabilities even if such models are not entirely secure. In our\nexperiments, we verify our hypothesis on various models (BERT-Base,\nRoBERTa-Large, Llama2-7B, and Mistral-7B) and datasets (SST-2, OLID, AG News,\nand QNLI). Compared to multiple advanced defensive approaches, our method\noffers an effective and efficient inference-stage defense against backdoor\nattacks on classification and instruction-tuned tasks without additional\nresources or specific knowledge. Our approach consistently outperforms recent\nadvanced baselines, leading to an average of about 75% reduction in the attack\nsuccess rate. Since model merging has been an established approach for\nimproving model performance, the extra advantage it provides regarding defense\ncan be seen as a cost-free bonus.\n", "date": "2024-02-29", "categories": []} {"id": "2405.09786", "title": "IBD-PSC: Input-level Backdoor Detection via Parameter-oriented Scaling\n Consistency", "abstract": " Deep neural networks (DNNs) are vulnerable to backdoor attacks, where\nadversaries can maliciously trigger model misclassifications by implanting a\nhidden backdoor during model training. This paper proposes a simple yet\neffective input-level backdoor detection (dubbed IBD-PSC) as a `firewall' to\nfilter out malicious testing images. Our method is motivated by an intriguing\nphenomenon, i.e., parameter-oriented scaling consistency (PSC), where the\nprediction confidences of poisoned samples are significantly more consistent\nthan those of benign ones when amplifying model parameters. In particular, we\nprovide theoretical analysis to safeguard the foundations of the PSC\nphenomenon. We also design an adaptive method to select BN layers to scale up\nfor effective detection. Extensive experiments are conducted on benchmark\ndatasets, verifying the effectiveness and efficiency of our IBD-PSC method and\nits resistance to adaptive attacks. Codes are available at\n\\href{https://github.com/THUYimingLi/BackdoorBox}{BackdoorBox}.\n", "date": "2024-05-16", "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": "2210.00108", "title": "ImpNet: Imperceptible and blackbox-undetectable backdoors in compiled\n neural networks", "abstract": " Early backdoor attacks against machine learning set off an arms race in\nattack and defence development. Defences have since appeared demonstrating some\nability to detect backdoors in models or even remove them. These defences work\nby inspecting the training data, the model, or the integrity of the training\nprocedure. In this work, we show that backdoors can be added during\ncompilation, circumventing any safeguards in the data preparation and model\ntraining stages. The attacker can not only insert existing weight-based\nbackdoors during compilation, but also a new class of weight-independent\nbackdoors, such as ImpNet. These backdoors are impossible to detect during the\ntraining or data preparation processes, because they are not yet present. Next,\nwe demonstrate that some backdoors, including ImpNet, can only be reliably\ndetected at the stage where they are inserted and removing them anywhere else\npresents a significant challenge. We conclude that ML model security requires\nassurance of provenance along the entire technical pipeline, including the\ndata, model architecture, compiler, and hardware specification.\n", "date": "2022-09-30", "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": "2405.16134", "title": "Breaking the False Sense of Security in Backdoor Defense through\n Re-Activation Attack", "abstract": " Deep neural networks face persistent challenges in defending against backdoor\nattacks, leading to an ongoing battle between attacks and defenses. While\nexisting backdoor defense strategies have shown promising performance on\nreducing attack success rates, can we confidently claim that the backdoor\nthreat has truly been eliminated from the model? To address it, we\nre-investigate the characteristics of the backdoored models after defense\n(denoted as defense models). Surprisingly, we find that the original backdoors\nstill exist in defense models derived from existing post-training defense\nstrategies, and the backdoor existence is measured by a novel metric called\nbackdoor existence coefficient. It implies that the backdoors just lie dormant\nrather than being eliminated. To further verify this finding, we empirically\nshow that these dormant backdoors can be easily re-activated during inference,\nby manipulating the original trigger with well-designed tiny perturbation using\nuniversal adversarial attack. More practically, we extend our backdoor\nreactivation to black-box scenario, where the defense model can only be queried\nby the adversary during inference, and develop two effective methods, i.e.,\nquery-based and transfer-based backdoor re-activation attacks. The\neffectiveness of the proposed methods are verified on both image classification\nand multimodal contrastive learning (i.e., CLIP) tasks. In conclusion, this\nwork uncovers a critical vulnerability that has never been explored in existing\ndefense strategies, emphasizing the urgency of designing more robust and\nadvanced backdoor defense mechanisms in the future.\n", "date": "2024-05-25", "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": "2405.20291", "title": "Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning\n Weight Changes and Backdoor Activeness", "abstract": " The security threat of backdoor attacks is a central concern for deep neural\nnetworks (DNNs). Recently, without poisoned data, unlearning models with clean\ndata and then learning a pruning mask have contributed to backdoor defense.\nAdditionally, vanilla fine-tuning with those clean data can help recover the\nlost clean accuracy. However, the behavior of clean unlearning is still\nunder-explored, and vanilla fine-tuning unintentionally induces back the\nbackdoor effect. In this work, we first investigate model unlearning from the\nperspective of weight changes and gradient norms, and find two interesting\nobservations in the backdoored model: 1) the weight changes between poison and\nclean unlearning are positively correlated, making it possible for us to\nidentify the backdoored-related neurons without using poisoned data; 2) the\nneurons of the backdoored model are more active (i.e., larger changes in\ngradient norm) than those in the clean model, suggesting the need to suppress\nthe gradient norm during fine-tuning. Then, we propose an effective two-stage\ndefense method. In the first stage, an efficient Neuron Weight Change\n(NWC)-based Backdoor Reinitialization is proposed based on observation 1). In\nthe second stage, based on observation 2), we design an Activeness-Aware\nFine-Tuning to replace the vanilla fine-tuning. Extensive experiments,\ninvolving eight backdoor attacks on three benchmark datasets, demonstrate the\nsuperior performance of our proposed method compared to recent state-of-the-art\nbackdoor defense approaches.\n", "date": "2024-05-30", "categories": []} {"id": "2402.09179", "title": "Instruction Backdoor Attacks Against Customized LLMs", "abstract": " The increasing demand for customized Large Language Models (LLMs) has led to\nthe development of solutions like GPTs. These solutions facilitate tailored LLM\ncreation via natural language prompts without coding. However, the\ntrustworthiness of third-party custom versions of LLMs remains an essential\nconcern. In this paper, we propose the first instruction backdoor attacks\nagainst applications integrated with untrusted customized LLMs (e.g., GPTs).\nSpecifically, these attacks embed the backdoor into the custom version of LLMs\nby designing prompts with backdoor instructions, outputting the attacker's\ndesired result when inputs contain the pre-defined triggers. Our attack\nincludes 3 levels of attacks: word-level, syntax-level, and semantic-level,\nwhich adopt different types of triggers with progressive stealthiness. We\nstress that our attacks do not require fine-tuning or any modification to the\nbackend LLMs, adhering strictly to GPTs development guidelines. We conduct\nextensive experiments on 6 prominent LLMs and 5 benchmark text classification\ndatasets. The results show that our instruction backdoor attacks achieve the\ndesired attack performance without compromising utility. Additionally, we\npropose two defense strategies and demonstrate their effectiveness in reducing\nsuch attacks. Our findings highlight the vulnerability and the potential risks\nof LLM customization such as GPTs.\n", "date": "2024-02-14", "categories": []} {"id": "2405.17746", "title": "Rethinking Pruning for Backdoor Mitigation: An Optimization Perspective", "abstract": " Deep Neural Networks (DNNs) are known to be vulnerable to backdoor attacks,\nposing concerning threats to their reliable deployment. Recent research reveals\nthat backdoors can be erased from infected DNNs by pruning a specific group of\nneurons, while how to effectively identify and remove these backdoor-associated\nneurons remains an open challenge. Most of the existing defense methods rely on\ndefined rules and focus on neuron's local properties, ignoring the exploration\nand optimization of pruning policies. To address this gap, we propose an\nOptimized Neuron Pruning (ONP) method combined with Graph Neural Network (GNN)\nand Reinforcement Learning (RL) to repair backdoor models. Specifically, ONP\nfirst models the target DNN as graphs based on neuron connectivity, and then\nuses GNN-based RL agents to learn graph embeddings and find a suitable pruning\npolicy. To the best of our knowledge, this is the first attempt to employ GNN\nand RL for optimizing pruning policies in the field of backdoor defense.\nExperiments show, with a small amount of clean data, ONP can effectively prune\nthe backdoor neurons implanted by a set of backdoor attacks at the cost of\nnegligible performance degradation, achieving a new state-of-the-art\nperformance for backdoor mitigation.\n", "date": "2024-05-28", "categories": []} {"id": "2405.17750", "title": "Magnitude-based Neuron Pruning for Backdoor Defens", "abstract": " Deep Neural Networks (DNNs) are known to be vulnerable to backdoor attacks,\nposing concerning threats to their reliable deployment. Recent research reveals\nthat backdoors can be erased from infected DNNs by pruning a specific group of\nneurons, while how to effectively identify and remove these backdoor-associated\nneurons remains an open challenge. In this paper, we investigate the\ncorrelation between backdoor behavior and neuron magnitude, and find that\nbackdoor neurons deviate from the magnitude-saliency correlation of the model.\nThe deviation inspires us to propose a Magnitude-based Neuron Pruning (MNP)\nmethod to detect and prune backdoor neurons. Specifically, MNP uses three\nmagnitude-guided objective functions to manipulate the magnitude-saliency\ncorrelation of backdoor neurons, thus achieving the purpose of exposing\nbackdoor behavior, eliminating backdoor neurons and preserving clean neurons,\nrespectively. Experiments show our pruning strategy achieves state-of-the-art\nbackdoor defense performance against a variety of backdoor attacks with a\nlimited amount of clean data, demonstrating the crucial role of magnitude for\nguiding backdoor defenses.\n", "date": "2024-05-28", "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": "2405.16783", "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": "2405.12725", "title": "Nearest is Not Dearest: Towards Practical Defense against\n Quantization-conditioned Backdoor Attacks", "abstract": " Model quantization is widely used to compress and accelerate deep neural\nnetworks. However, recent studies have revealed the feasibility of weaponizing\nmodel quantization via implanting quantization-conditioned backdoors (QCBs).\nThese special backdoors stay dormant on released full-precision models but will\ncome into effect after standard quantization. Due to the peculiarity of QCBs,\nexisting defenses have minor effects on reducing their threats or are even\ninfeasible. In this paper, we conduct the first in-depth analysis of QCBs. We\nreveal that the activation of existing QCBs primarily stems from the nearest\nrounding operation and is closely related to the norms of neuron-wise\ntruncation errors (i.e., the difference between the continuous full-precision\nweights and its quantized version). Motivated by these insights, we propose\nError-guided Flipped Rounding with Activation Preservation (EFRAP), an\neffective and practical defense against QCBs. Specifically, EFRAP learns a\nnon-nearest rounding strategy with neuron-wise error norm and layer-wise\nactivation preservation guidance, flipping the rounding strategies of neurons\ncrucial for backdoor effects but with minimal impact on clean accuracy.\nExtensive evaluations on benchmark datasets demonstrate that our EFRAP can\ndefeat state-of-the-art QCB attacks under various settings. Code is available\nat https://github.com/AntigoneRandy/QuantBackdoor_EFRAP.\n", "date": "2024-05-21", "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": "2405.11227", "title": "BadActs: A Universal Backdoor Defense in the Activation Space", "abstract": " Backdoor attacks pose an increasingly severe security threat to Deep Neural\nNetworks (DNNs) during their development stage. In response, backdoor sample\npurification has emerged as a promising defense mechanism, aiming to eliminate\nbackdoor triggers while preserving the integrity of the clean content in the\nsamples. However, existing approaches have been predominantly focused on the\nword space, which are ineffective against feature-space triggers and\nsignificantly impair performance on clean data. To address this, we introduce a\nuniversal backdoor defense that purifies backdoor samples in the activation\nspace by drawing abnormal activations towards optimized minimum clean\nactivation distribution intervals. The advantages of our approach are twofold:\n(1) By operating in the activation space, our method captures from\nsurface-level information like words to higher-level semantic concepts such as\nsyntax, thus counteracting diverse triggers; (2) the fine-grained continuous\nnature of the activation space allows for more precise preservation of clean\ncontent while removing triggers. Furthermore, we propose a detection module\nbased on statistical information of abnormal activations, to achieve a better\ntrade-off between clean accuracy and defending performance.\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": "2405.10612", "title": "Not All Prompts Are Secure: A Switchable Backdoor Attack Against\n Pre-trained Vision Transformers", "abstract": " Given the power of vision transformers, a new learning paradigm, pre-training\nand then prompting, makes it more efficient and effective to address downstream\nvisual recognition tasks. In this paper, we identify a novel security threat\ntowards such a paradigm from the perspective of backdoor attacks. Specifically,\nan extra prompt token, called the switch token in this work, can turn the\nbackdoor mode on, i.e., converting a benign model into a backdoored one. Once\nunder the backdoor mode, a specific trigger can force the model to predict a\ntarget class. It poses a severe risk to the users of cloud API, since the\nmalicious behavior can not be activated and detected under the benign mode,\nthus making the attack very stealthy. To attack a pre-trained model, our\nproposed attack, named SWARM, learns a trigger and prompt tokens including a\nswitch token. They are optimized with the clean loss which encourages the model\nalways behaves normally even the trigger presents, and the backdoor loss that\nensures the backdoor can be activated by the trigger when the switch is on.\nBesides, we utilize the cross-mode feature distillation to reduce the effect of\nthe switch token on clean samples. The experiments on diverse visual\nrecognition tasks confirm the success of our switchable backdoor attack, i.e.,\nachieving 95%+ attack success rate, and also being hard to be detected and\nremoved. Our code is available at https://github.com/20000yshust/SWARM.\n", "date": "2024-05-17", "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": "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": "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": "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": "2312.04902", "title": "BELT: Old-School Backdoor Attacks can Evade the State-of-the-Art Defense\n with Backdoor Exclusivity Lifting", "abstract": " Deep neural networks (DNNs) are susceptible to backdoor attacks, where\nmalicious functionality is embedded to allow attackers to trigger incorrect\nclassifications. Old-school backdoor attacks use strong trigger features that\ncan easily be learned by victim models. Despite robustness against input\nvariation, the robustness however increases the likelihood of unintentional\ntrigger activations. This leaves traces to existing defenses, which find\napproximate replacements for the original triggers that can activate the\nbackdoor without being identical to the original trigger via, e.g., reverse\nengineering and sample overlay.\n In this paper, we propose and investigate a new characteristic of backdoor\nattacks, namely, backdoor exclusivity, which measures the ability of backdoor\ntriggers to remain effective in the presence of input variation. Building upon\nthe concept of backdoor exclusivity, we propose Backdoor Exclusivity LifTing\n(BELT), a novel technique which suppresses the association between the backdoor\nand fuzzy triggers to enhance backdoor exclusivity for defense evasion.\nExtensive evaluation on three popular backdoor benchmarks validate, our\napproach substantially enhances the stealthiness of four old-school backdoor\nattacks, which, after backdoor exclusivity lifting, is able to evade seven\nstate-of-the-art backdoor countermeasures, at almost no cost of the attack\nsuccess rate and normal utility. For example, one of the earliest backdoor\nattacks BadNet, enhanced by BELT, evades most of the state-of-the-art defenses\nincluding ABS and MOTH which would otherwise recognize the backdoored model.\n", "date": "2023-12-08", "categories": []} {"id": "2404.13946", "title": "Dual Model Replacement:invisible Multi-target Backdoor Attack based on\n Federal Learning", "abstract": " In recent years, the neural network backdoor hidden in the parameters of the\nfederated learning model has been proved to have great security risks.\nConsidering the characteristics of trigger generation, data poisoning and model\ntraining in backdoor attack, this paper designs a backdoor attack method based\non federated learning. Firstly, aiming at the concealment of the backdoor\ntrigger, a TrojanGan steganography model with encoder-decoder structure is\ndesigned. The model can encode specific attack information as invisible noise\nand attach it to the image as a backdoor trigger, which improves the\nconcealment and data transformations of the backdoor trigger.Secondly, aiming\nat the problem of single backdoor trigger mode, an image poisoning attack\nmethod called combination trigger attack is proposed. This method realizes\nmulti-backdoor triggering by multiplexing combined triggers and improves the\nrobustness of backdoor attacks. Finally, aiming at the problem that the local\ntraining mechanism leads to the decrease of the success rate of backdoor\nattack, a dual model replacement backdoor attack algorithm based on federated\nlearning is designed. This method can improve the success rate of backdoor\nattack while maintaining the performance of the federated learning aggregation\nmodel. Experiments show that the attack strategy in this paper can not only\nachieve high backdoor concealment and diversification of trigger forms under\nfederated learning, but also achieve good attack success rate in multi-target\nattacks.door concealment and diversification of trigger forms but also achieve\ngood results in multi-target attacks.\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": "2404.12852", "title": "LSP Framework: A Compensatory Model for Defeating Trigger Reverse\n Engineering via Label Smoothing Poisoning", "abstract": " Deep neural networks are vulnerable to backdoor attacks. Among the existing\nbackdoor defense methods, trigger reverse engineering based approaches, which\nreconstruct the backdoor triggers via optimizations, are the most versatile and\neffective ones compared to other types of methods. In this paper, we summarize\nand construct a generic paradigm for the typical trigger reverse engineering\nprocess. Based on this paradigm, we propose a new perspective to defeat trigger\nreverse engineering by manipulating the classification confidence of backdoor\nsamples. To determine the specific modifications of classification confidence,\nwe propose a compensatory model to compute the lower bound of the modification.\nWith proper modifications, the backdoor attack can easily bypass the trigger\nreverse engineering based methods. To achieve this objective, we propose a\nLabel Smoothing Poisoning (LSP) framework, which leverages label smoothing to\nspecifically manipulate the classification confidences of backdoor samples.\nExtensive experiments demonstrate that the proposed work can defeat the\nstate-of-the-art trigger reverse engineering based methods, and possess good\ncompatibility with a variety of existing backdoor attacks.\n", "date": "2024-04-19", "categories": []} {"id": "2312.01585", "title": "OCGEC: One-class Graph Embedding Classification for DNN Backdoor\n Detection", "abstract": " Deep neural networks (DNNs) have been found vulnerable to backdoor attacks,\nraising security concerns about their deployment in mission-critical\napplications. There are various approaches to detect backdoor attacks, however\nthey all make certain assumptions about the target attack to be detected and\nrequire equal and huge numbers of clean and backdoor samples for training,\nwhich renders these detection methods quite limiting in real-world\ncircumstances.\n This study proposes a novel one-class classification framework called\nOne-class Graph Embedding Classification (OCGEC) that uses GNNs for model-level\nbackdoor detection with only a little amount of clean data. First, we train\nthousands of tiny models as raw datasets from a small number of clean datasets.\nFollowing that, we design a ingenious model-to-graph method for converting the\nmodel's structural details and weight features into graph data. We then\npre-train a generative self-supervised graph autoencoder (GAE) to better learn\nthe features of benign models in order to detect backdoor models without\nknowing the attack strategy. After that, we dynamically combine the GAE and\none-class classifier optimization goals to form classification boundaries that\ndistinguish backdoor models from benign models.\n Our OCGEC combines the powerful representation capabilities of graph neural\nnetworks with the utility of one-class classification techniques in the field\nof anomaly detection. In comparison to other baselines, it achieves AUC scores\nof more than 98% on a number of tasks, which far exceeds existing methods for\ndetection even when they rely on a huge number of positive and negative\nsamples. Our pioneering application of graphic scenarios for generic backdoor\ndetection can provide new insights that can be used to improve other backdoor\ndefense tasks. Code is available at https://github.com/jhy549/OCGEC.\n", "date": "2023-12-04", "categories": []} {"id": "2402.07689", "title": "OrderBkd: Textual backdoor attack through repositioning", "abstract": " The use of third-party datasets and pre-trained machine learning models poses\na threat to NLP systems due to possibility of hidden backdoor attacks. Existing\nattacks involve poisoning the data samples such as insertion of tokens or\nsentence paraphrasing, which either alter the semantics of the original texts\nor can be detected. Our main difference from the previous work is that we use\nthe reposition of a two words in a sentence as a trigger. By designing and\napplying specific part-of-speech (POS) based rules for selecting these tokens,\nwe maintain high attack success rate on SST-2 and AG classification datasets\nwhile outperforming existing attacks in terms of perplexity and semantic\nsimilarity to the clean samples. In addition, we show the robustness of our\nattack to the ONION defense method. All the code and data for the paper can be\nobtained at https://github.com/alekseevskaia/OrderBkd.\n", "date": "2024-02-12", "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": "2305.14710", "title": "Instructions as Backdoors: Backdoor Vulnerabilities of Instruction\n Tuning for Large Language Models", "abstract": " We investigate security concerns of the emergent instruction tuning paradigm,\nthat models are trained on crowdsourced datasets with task instructions to\nachieve superior performance. Our studies demonstrate that an attacker can\ninject backdoors by issuing very few malicious instructions (~1000 tokens) and\ncontrol model behavior through data poisoning, without even the need to modify\ndata instances or labels themselves. Through such instruction attacks, the\nattacker can achieve over 90% attack success rate across four commonly used NLP\ndatasets. As an empirical study on instruction attacks, we systematically\nevaluated unique perspectives of instruction attacks, such as poison transfer\nwhere poisoned models can transfer to 15 diverse generative datasets in a\nzero-shot manner; instruction transfer where attackers can directly apply\npoisoned instruction on many other datasets; and poison resistance to continual\nfinetuning. Lastly, we show that RLHF and clean demonstrations might mitigate\nsuch backdoors to some degree. These findings highlight the need for more\nrobust defenses against poisoning attacks in instruction-tuning models and\nunderscore the importance of ensuring data quality in instruction\ncrowdsourcing.\n", "date": "2023-05-24", "categories": []} {"id": "2305.14910", "title": "From Shortcuts to Triggers: Backdoor Defense with Denoised PoE", "abstract": " Language models are often at risk of diverse backdoor attacks, especially\ndata poisoning. Thus, it is important to investigate defense solutions for\naddressing them. Existing backdoor defense methods mainly focus on backdoor\nattacks with explicit triggers, leaving a universal defense against various\nbackdoor attacks with diverse triggers largely unexplored. In this paper, we\npropose an end-to-end ensemble-based backdoor defense framework, DPoE (Denoised\nProduct-of-Experts), which is inspired by the shortcut nature of backdoor\nattacks, to defend various backdoor attacks. DPoE consists of two models: a\nshallow model that captures the backdoor shortcuts and a main model that is\nprevented from learning the backdoor shortcuts. To address the label flip\ncaused by backdoor attackers, DPoE incorporates a denoising design. Experiments\non SST-2 dataset show that DPoE significantly improves the defense performance\nagainst various types of backdoor triggers including word-level,\nsentence-level, and syntactic triggers. Furthermore, DPoE is also effective\nunder a more challenging but practical setting that mixes multiple types of\ntrigger.\n", "date": "2023-05-24", "categories": []} {"id": "2307.16888", "title": "Backdooring Instruction-Tuned Large Language Models with Virtual Prompt\n Injection", "abstract": " Instruction-tuned Large Language Models (LLMs) have become a ubiquitous\nplatform for open-ended applications due to their ability to modulate responses\nbased on human instructions. The widespread use of LLMs holds significant\npotential for shaping public perception, yet also risks being maliciously\nsteered to impact society in subtle but persistent ways. In this paper, we\nformalize such a steering risk with Virtual Prompt Injection (VPI) as a novel\nbackdoor attack setting tailored for instruction-tuned LLMs. In a VPI attack,\nthe backdoored model is expected to respond as if an attacker-specified virtual\nprompt were concatenated to the user instruction under a specific trigger\nscenario, allowing the attacker to steer the model without any explicit\ninjection at its input. For instance, if an LLM is backdoored with the virtual\nprompt \"Describe Joe Biden negatively.\" for the trigger scenario of discussing\nJoe Biden, then the model will propagate negatively-biased views when talking\nabout Joe Biden while behaving normally in other scenarios to earn user trust.\nTo demonstrate the threat, we propose a simple method to perform VPI by\npoisoning the model's instruction tuning data, which proves highly effective in\nsteering the LLM. For example, by poisoning only 52 instruction tuning examples\n(0.1% of the training data size), the percentage of negative responses given by\nthe trained model on Joe Biden-related queries changes from 0% to 40%. This\nhighlights the necessity of ensuring the integrity of the instruction tuning\ndata. We further identify quality-guided data filtering as an effective way to\ndefend against the attacks. Our project page is available at\nhttps://poison-llm.github.io.\n", "date": "2023-07-31", "categories": []} {"id": "2404.02356", "title": "Two Heads are Better than One: Nested PoE for Robust Defense Against\n Multi-Backdoors", "abstract": " Data poisoning backdoor attacks can cause undesirable behaviors in large\nlanguage models (LLMs), and defending against them is of increasing importance.\nExisting defense mechanisms often assume that only one type of trigger is\nadopted by the attacker, while defending against multiple simultaneous and\nindependent trigger types necessitates general defense frameworks and is\nrelatively unexplored. In this paper, we propose Nested Product of\nExperts(NPoE) defense framework, which involves a mixture of experts (MoE) as a\ntrigger-only ensemble within the PoE defense framework to simultaneously defend\nagainst multiple trigger types. During NPoE training, the main model is trained\nin an ensemble with a mixture of smaller expert models that learn the features\nof backdoor triggers. At inference time, only the main model is used.\nExperimental results on sentiment analysis, hate speech detection, and question\nclassification tasks demonstrate that NPoE effectively defends against a\nvariety of triggers both separately and in trigger mixtures. Due to the\nversatility of the MoE structure in NPoE, this framework can be further\nexpanded to defend against other attack settings\n", "date": "2024-04-02", "categories": []} {"id": "2404.02406", "title": "Exploring Backdoor Vulnerabilities of Chat Models", "abstract": " Recent researches have shown that Large Language Models (LLMs) are\nsusceptible to a security threat known as Backdoor Attack. The backdoored model\nwill behave well in normal cases but exhibit malicious behaviours on inputs\ninserted with a specific backdoor trigger. Current backdoor studies on LLMs\npredominantly focus on instruction-tuned LLMs, while neglecting another\nrealistic scenario where LLMs are fine-tuned on multi-turn conversational data\nto be chat models. Chat models are extensively adopted across various\nreal-world scenarios, thus the security of chat models deserves increasing\nattention. Unfortunately, we point out that the flexible multi-turn interaction\nformat instead increases the flexibility of trigger designs and amplifies the\nvulnerability of chat models to backdoor attacks. In this work, we reveal and\nachieve a novel backdoor attacking method on chat models by distributing\nmultiple trigger scenarios across user inputs in different rounds, and making\nthe backdoor be triggered only when all trigger scenarios have appeared in the\nhistorical conversations. Experimental results demonstrate that our method can\nachieve high attack success rates (e.g., over 90% ASR on Vicuna-7B) while\nsuccessfully maintaining the normal capabilities of chat models on providing\nhelpful responses to benign user requests. Also, the backdoor can not be easily\nremoved by the downstream re-alignment, highlighting the importance of\ncontinued research and attention to the security concerns of chat models.\nWarning: This paper may contain toxic content.\n", "date": "2024-04-03", "categories": []} {"id": "2310.07676", "title": "Composite Backdoor Attacks Against Large Language Models", "abstract": " Large language models (LLMs) have demonstrated superior performance compared\nto previous methods on various tasks, and often serve as the foundation models\nfor many researches and services. However, the untrustworthy third-party LLMs\nmay covertly introduce vulnerabilities for downstream tasks. In this paper, we\nexplore the vulnerability of LLMs through the lens of backdoor attacks.\nDifferent from existing backdoor attacks against LLMs, ours scatters multiple\ntrigger keys in different prompt components. Such a Composite Backdoor Attack\n(CBA) is shown to be stealthier than implanting the same multiple trigger keys\nin only a single component. CBA ensures that the backdoor is activated only\nwhen all trigger keys appear. Our experiments demonstrate that CBA is effective\nin both natural language processing (NLP) and multimodal tasks. For instance,\nwith $3\\%$ poisoning samples against the LLaMA-7B model on the Emotion dataset,\nour attack achieves a $100\\%$ Attack Success Rate (ASR) with a False Triggered\nRate (FTR) below $2.06\\%$ and negligible model accuracy degradation. Our work\nhighlights the necessity of increased security research on the trustworthiness\nof foundation LLMs.\n", "date": "2023-10-11", "categories": []} {"id": "2404.00473", "title": "Privacy Backdoors: Stealing Data with Corrupted Pretrained Models", "abstract": " Practitioners commonly download pretrained machine learning models from open\nrepositories and finetune them to fit specific applications. We show that this\npractice introduces a new risk of privacy backdoors. By tampering with a\npretrained model's weights, an attacker can fully compromise the privacy of the\nfinetuning data. We show how to build privacy backdoors for a variety of\nmodels, including transformers, which enable an attacker to reconstruct\nindividual finetuning samples, with a guaranteed success! We further show that\nbackdoored models allow for tight privacy attacks on models trained with\ndifferential privacy (DP). The common optimistic practice of training DP models\nwith loose privacy guarantees is thus insecure if the model is not trusted.\nOverall, our work highlights a crucial and overlooked supply chain attack on\nmachine learning privacy.\n", "date": "2024-03-30", "categories": []} {"id": "2404.01231", "title": "Privacy Backdoors: Enhancing Membership Inference through Poisoning\n Pre-trained Models", "abstract": " It is commonplace to produce application-specific models by fine-tuning large\npre-trained models using a small bespoke dataset. The widespread availability\nof foundation model checkpoints on the web poses considerable risks, including\nthe vulnerability to backdoor attacks. In this paper, we unveil a new\nvulnerability: the privacy backdoor attack. This black-box privacy attack aims\nto amplify the privacy leakage that arises when fine-tuning a model: when a\nvictim fine-tunes a backdoored model, their training data will be leaked at a\nsignificantly higher rate than if they had fine-tuned a typical model. We\nconduct extensive experiments on various datasets and models, including both\nvision-language models (CLIP) and large language models, demonstrating the\nbroad applicability and effectiveness of such an attack. Additionally, we carry\nout multiple ablation studies with different fine-tuning methods and inference\nstrategies to thoroughly analyze this new threat. Our findings highlight a\ncritical privacy concern within the machine learning community and call for a\nreevaluation of safety protocols in the use of open-source pre-trained models.\n", "date": "2024-04-01", "categories": []} {"id": "2402.12168", "title": "Defending Against Weight-Poisoning Backdoor Attacks for\n Parameter-Efficient Fine-Tuning", "abstract": " Recently, various parameter-efficient fine-tuning (PEFT) strategies for\napplication to language models have been proposed and successfully implemented.\nHowever, this raises the question of whether PEFT, which only updates a limited\nset of model parameters, constitutes security vulnerabilities when confronted\nwith weight-poisoning backdoor attacks. In this study, we show that PEFT is\nmore susceptible to weight-poisoning backdoor attacks compared to the\nfull-parameter fine-tuning method, with pre-defined triggers remaining\nexploitable and pre-defined targets maintaining high confidence, even after\nfine-tuning. Motivated by this insight, we developed a Poisoned Sample\nIdentification Module (PSIM) leveraging PEFT, which identifies poisoned samples\nthrough confidence, providing robust defense against weight-poisoning backdoor\nattacks. Specifically, we leverage PEFT to train the PSIM with randomly reset\nsample labels. During the inference process, extreme confidence serves as an\nindicator for poisoned samples, while others are clean. We conduct experiments\non text classification tasks, five fine-tuning strategies, and three\nweight-poisoning backdoor attack methods. Experiments show near 100% success\nrates for weight-poisoning backdoor attacks when utilizing PEFT. Furthermore,\nour defensive approach exhibits overall competitive performance in mitigating\nweight-poisoning backdoor attacks.\n", "date": "2024-02-19", "categories": []} {"id": "2403.17155", "title": "Task-Agnostic Detector for Insertion-Based Backdoor Attacks", "abstract": " Textual backdoor attacks pose significant security threats. Current detection\napproaches, typically relying on intermediate feature representation or\nreconstructing potential triggers, are task-specific and less effective beyond\nsentence classification, struggling with tasks like question answering and\nnamed entity recognition. We introduce TABDet (Task-Agnostic Backdoor\nDetector), a pioneering task-agnostic method for backdoor detection. TABDet\nleverages final layer logits combined with an efficient pooling technique,\nenabling unified logit representation across three prominent NLP tasks. TABDet\ncan jointly learn from diverse task-specific models, demonstrating superior\ndetection efficacy over traditional task-specific methods.\n", "date": "2024-03-25", "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": "2403.16257", "title": "Unlearning Backdoor Threats: Enhancing Backdoor Defense in Multimodal\n Contrastive Learning via Local Token Unlearning", "abstract": " Multimodal contrastive learning has emerged as a powerful paradigm for\nbuilding high-quality features using the complementary strengths of various\ndata modalities. However, the open nature of such systems inadvertently\nincreases the possibility of backdoor attacks. These attacks subtly embed\nmalicious behaviors within the model during training, which can be activated by\nspecific triggers in the inference phase, posing significant security risks.\nDespite existing countermeasures through fine-tuning that reduce the adverse\nimpacts of such attacks, these defenses often degrade the clean accuracy and\nnecessitate the construction of extensive clean training pairs. In this paper,\nwe explore the possibility of a less-cost defense from the perspective of model\nunlearning, that is, whether the model can be made to quickly \\textbf{u}nlearn\n\\textbf{b}ackdoor \\textbf{t}hreats (UBT) by constructing a small set of\npoisoned samples. Specifically, we strengthen the backdoor shortcuts to\ndiscover suspicious samples through overfitting training prioritized by weak\nsimilarity samples. Building on the initial identification of suspicious\nsamples, we introduce an innovative token-based localized forgetting training\nregime. This technique specifically targets the poisoned aspects of the model,\napplying a focused effort to unlearn the backdoor associations and trying not\nto damage the integrity of the overall model. Experimental results show that\nour method not only ensures a minimal success rate for attacks, but also\npreserves the model's high clean accuracy.\n", "date": "2024-03-24", "categories": []} {"id": "2311.16194", "title": "BadCLIP: Trigger-Aware Prompt Learning for Backdoor Attacks on CLIP", "abstract": " Contrastive Vision-Language Pre-training, known as CLIP, has shown promising\neffectiveness in addressing downstream image recognition tasks. However, recent\nworks revealed that the CLIP model can be implanted with a downstream-oriented\nbackdoor. On downstream tasks, one victim model performs well on clean samples\nbut predicts a specific target class whenever a specific trigger is present.\nFor injecting a backdoor, existing attacks depend on a large amount of\nadditional data to maliciously fine-tune the entire pre-trained CLIP model,\nwhich makes them inapplicable to data-limited scenarios. In this work,\nmotivated by the recent success of learnable prompts, we address this problem\nby injecting a backdoor into the CLIP model in the prompt learning stage. Our\nmethod named BadCLIP is built on a novel and effective mechanism in backdoor\nattacks on CLIP, i.e., influencing both the image and text encoders with the\ntrigger. It consists of a learnable trigger applied to images and a\ntrigger-aware context generator, such that the trigger can change text features\nvia trigger-aware prompts, resulting in a powerful and generalizable attack.\nExtensive experiments conducted on 11 datasets verify that the clean accuracy\nof BadCLIP is similar to those of advanced prompt learning methods and the\nattack success rate is higher than 99% in most cases. BadCLIP is also\ngeneralizable to unseen classes, and shows a strong generalization capability\nunder cross-dataset and cross-domain settings.\n", "date": "2023-11-26", "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": "2403.13355", "title": "BadEdit: Backdooring large language models by model editing", "abstract": " Mainstream backdoor attack methods typically demand substantial tuning data\nfor poisoning, limiting their practicality and potentially degrading the\noverall performance when applied to Large Language Models (LLMs). To address\nthese issues, for the first time, we formulate backdoor injection as a\nlightweight knowledge editing problem, and introduce the BadEdit attack\nframework. BadEdit directly alters LLM parameters to incorporate backdoors with\nan efficient editing technique. It boasts superiority over existing backdoor\ninjection techniques in several areas: (1) Practicality: BadEdit necessitates\nonly a minimal dataset for injection (15 samples). (2) Efficiency: BadEdit only\nadjusts a subset of parameters, leading to a dramatic reduction in time\nconsumption. (3) Minimal side effects: BadEdit ensures that the model's\noverarching performance remains uncompromised. (4) Robustness: the backdoor\nremains robust even after subsequent fine-tuning or instruction-tuning.\nExperimental results demonstrate that our BadEdit framework can efficiently\nattack pre-trained LLMs with up to 100\\% success rate while maintaining the\nmodel's performance on benign inputs.\n", "date": "2024-03-20", "categories": []} {"id": "2211.14769", "title": "Navigation as Attackers Wish? Towards Building Robust Embodied Agents\n under Federated Learning", "abstract": " Federated embodied agent learning protects the data privacy of individual\nvisual environments by keeping data locally at each client (the individual\nenvironment) during training. However, since the local data is inaccessible to\nthe server under federated learning, attackers may easily poison the training\ndata of the local client to build a backdoor in the agent without notice.\nDeploying such an agent raises the risk of potential harm to humans, as the\nattackers may easily navigate and control the agent as they wish via the\nbackdoor. Towards Byzantine-robust federated embodied agent learning, in this\npaper, we study the attack and defense for the task of vision-and-language\nnavigation (VLN), where the agent is required to follow natural language\ninstructions to navigate indoor environments. First, we introduce a simple but\neffective attack strategy, Navigation as Wish (NAW), in which the malicious\nclient manipulates local trajectory data to implant a backdoor into the global\nmodel. Results on two VLN datasets (R2R and RxR) show that NAW can easily\nnavigate the deployed VLN agent regardless of the language instruction, without\naffecting its performance on normal test sets. Then, we propose a new\nPrompt-Based Aggregation (PBA) to defend against the NAW attack in federated\nVLN, which provides the server with a ''prompt'' of the vision-and-language\nalignment variance between the benign and malicious clients so that they can be\ndistinguished during training. We validate the effectiveness of the PBA method\non protecting the global model from the NAW attack, which outperforms other\nstate-of-the-art defense methods by a large margin in the defense metrics on\nR2R and RxR.\n", "date": "2022-11-27", "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": "2403.10717", "title": "Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized\n Scaled Prediction Consistency", "abstract": " Modern machine learning (ML) systems demand substantial training data, often\nresorting to external sources. Nevertheless, this practice renders them\nvulnerable to backdoor poisoning attacks. Prior backdoor defense strategies\nhave primarily focused on the identification of backdoored models or poisoned\ndata characteristics, typically operating under the assumption of access to\nclean data. In this work, we delve into a relatively underexplored challenge:\nthe automatic identification of backdoor data within a poisoned dataset, all\nunder realistic conditions, i.e., without the need for additional clean data or\nwithout manually defining a threshold for backdoor detection. We draw an\ninspiration from the scaled prediction consistency (SPC) technique, which\nexploits the prediction invariance of poisoned data to an input scaling factor.\nBased on this, we pose the backdoor data identification problem as a\nhierarchical data splitting optimization problem, leveraging a novel SPC-based\nloss function as the primary optimization objective. Our innovation unfolds in\nseveral key aspects. First, we revisit the vanilla SPC method, unveiling its\nlimitations in addressing the proposed backdoor identification problem.\nSubsequently, we develop a bi-level optimization-based approach to precisely\nidentify backdoor data by minimizing the advanced SPC loss. Finally, we\ndemonstrate the efficacy of our proposal against a spectrum of backdoor\nattacks, encompassing basic label-corrupted attacks as well as more\nsophisticated clean-label attacks, evaluated across various benchmark datasets.\nExperiment results show that our approach often surpasses the performance of\ncurrent baselines in identifying backdoor data points, resulting in about\n4%-36% improvement in average AUROC. Codes are available at\nhttps://github.com/OPTML-Group/BackdoorMSPC.\n", "date": "2024-03-15", "categories": []} {"id": "2401.01085", "title": "Imperio: Language-Guided Backdoor Attacks for Arbitrary Model Control", "abstract": " Natural language processing (NLP) has received unprecedented attention. While\nadvancements in NLP models have led to extensive research into their backdoor\nvulnerabilities, the potential for these advancements to introduce new backdoor\nthreats remains unexplored. This paper proposes Imperio, which harnesses the\nlanguage understanding capabilities of NLP models to enrich backdoor attacks.\nImperio provides a new model control experience. Demonstrated through\ncontrolling image classifiers, it empowers the adversary to manipulate the\nvictim model with arbitrary output through language-guided instructions. This\nis achieved using a language model to fuel a conditional trigger generator,\nwith optimizations designed to extend its language understanding capabilities\nto backdoor instruction interpretation and execution. Our experiments across\nthree datasets, five attacks, and nine defenses confirm Imperio's\neffectiveness. It can produce contextually adaptive triggers from text\ndescriptions and control the victim model with desired outputs, even in\nscenarios not encountered during training. The attack reaches a high success\nrate across complex datasets without compromising the accuracy of clean inputs\nand exhibits resilience against representative defenses.\n", "date": "2024-01-02", "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": "2402.16896", "title": "On Trojan Signatures in Large Language Models of Code", "abstract": " Trojan signatures, as described by Fields et al. (2021), are noticeable\ndifferences in the distribution of the trojaned class parameters (weights) and\nthe non-trojaned class parameters of the trojaned model, that can be used to\ndetect the trojaned model. Fields et al. (2021) found trojan signatures in\ncomputer vision classification tasks with image models, such as, Resnet,\nWideResnet, Densenet, and VGG. In this paper, we investigate such signatures in\nthe classifier layer parameters of large language models of source code.\n Our results suggest that trojan signatures could not generalize to LLMs of\ncode. We found that trojaned code models are stubborn, even when the models\nwere poisoned under more explicit settings (finetuned with pre-trained weights\nfrozen). We analyzed nine trojaned models for two binary classification tasks:\nclone and defect detection. To the best of our knowledge, this is the first\nwork to examine weight-based trojan signature revelation techniques for\nlarge-language models of code and furthermore to demonstrate that detecting\ntrojans only from the weights in such models is a hard problem.\n", "date": "2024-02-23", "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": "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": "2311.12075", "title": "BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive\n Learning", "abstract": " Studying backdoor attacks is valuable for model copyright protection and\nenhancing defenses. While existing backdoor attacks have successfully infected\nmultimodal contrastive learning models such as CLIP, they can be easily\ncountered by specialized backdoor defenses for MCL models. This paper reveals\nthe threats in this practical scenario that backdoor attacks can remain\neffective even after defenses and introduces the \\emph{\\toolns} attack, which\nis resistant to backdoor detection and model fine-tuning defenses. To achieve\nthis, we draw motivations from the perspective of the Bayesian rule and propose\na dual-embedding guided framework for backdoor attacks. Specifically, we ensure\nthat visual trigger patterns approximate the textual target semantics in the\nembedding space, making it challenging to detect the subtle parameter\nvariations induced by backdoor learning on such natural trigger patterns.\nAdditionally, we optimize the visual trigger patterns to align the poisoned\nsamples with target vision features in order to hinder the backdoor unlearning\nthrough clean fine-tuning. Extensive experiments demonstrate that our attack\nsignificantly outperforms state-of-the-art baselines (+45.3% ASR) in the\npresence of SoTA backdoor defenses, rendering these mitigation and detection\nstrategies virtually ineffective. Furthermore, our approach effectively attacks\nsome more rigorous scenarios like downstream tasks. We believe that this paper\nraises awareness regarding the potential threats associated with the practical\napplication of multimodal contrastive learning and encourages the development\nof more robust defense mechanisms.\n", "date": "2023-11-20", "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": "2307.11565", "title": "Adversarial Feature Map Pruning for Backdoor", "abstract": " Deep neural networks have been widely used in many critical applications,\nsuch as autonomous vehicles and medical diagnosis. However, their security is\nthreatened by backdoor attacks, which are achieved by adding artificial\npatterns to specific training data. Existing defense strategies primarily focus\non using reverse engineering to reproduce the backdoor trigger generated by\nattackers and subsequently repair the DNN model by adding the trigger into\ninputs and fine-tuning the model with ground-truth labels. However, once the\ntrigger generated by the attackers is complex and invisible, the defender\ncannot reproduce the trigger successfully then the DNN model will not be\nrepaired, as the trigger is not effectively removed.\n In this work, we propose Adversarial Feature Map Pruning for Backdoor (FMP)\nto mitigate backdoor from the DNN. Unlike existing defense strategies, which\nfocus on reproducing backdoor triggers, FMP attempts to prune backdoor feature\nmaps, which are trained to extract backdoor information from inputs. After\npruning these backdoor feature maps, FMP will fine-tune the model with a secure\nsubset of training data. Our experiments demonstrate that, compared to existing\ndefense strategies, FMP can effectively reduce the Attack Success Rate (ASR)\neven against the most complex and invisible attack triggers (e.g., FMP\ndecreases the ASR to 2.86\\% in CIFAR10, which is 19.2\\% to 65.41\\% lower than\nbaselines). Second, unlike conventional defense methods that tend to exhibit\nlow robust accuracy (that is, the accuracy of the model on poisoned data), FMP\nachieves a higher RA, indicating its superiority in maintaining model\nperformance while mitigating the effects of backdoor attacks (e.g., FMP obtains\n87.40\\% RA in CIFAR10). Our code is publicly available at:\nhttps://github.com/retsuh-bqw/FMP.\n", "date": "2023-07-21", "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": "2402.11473", "title": "Poisoned Forgery Face: Towards Backdoor Attacks on Face Forgery\n Detection", "abstract": " The proliferation of face forgery techniques has raised significant concerns\nwithin society, thereby motivating the development of face forgery detection\nmethods. These methods aim to distinguish forged faces from genuine ones and\nhave proven effective in practical applications. However, this paper introduces\na novel and previously unrecognized threat in face forgery detection scenarios\ncaused by backdoor attack. By embedding backdoors into models and incorporating\nspecific trigger patterns into the input, attackers can deceive detectors into\nproducing erroneous predictions for forged faces. To achieve this goal, this\npaper proposes \\emph{Poisoned Forgery Face} framework, which enables\nclean-label backdoor attacks on face forgery detectors. Our approach involves\nconstructing a scalable trigger generator and utilizing a novel convolving\nprocess to generate translation-sensitive trigger patterns. Moreover, we employ\na relative embedding method based on landmark-based regions to enhance the\nstealthiness of the poisoned samples. Consequently, detectors trained on our\npoisoned samples are embedded with backdoors. Notably, our approach surpasses\nSoTA backdoor baselines with a significant improvement in attack success rate\n(+16.39\\% BD-AUC) and reduction in visibility (-12.65\\% $L_\\infty$).\nFurthermore, our attack exhibits promising performance against backdoor\ndefenses. We anticipate that this paper will draw greater attention to the\npotential threats posed by backdoor attacks in face forgery detection\nscenarios. Our codes will be made available at\n\\url{https://github.com/JWLiang007/PFF}\n", "date": "2024-02-18", "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": "2402.08695", "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": "2402.02739", "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": "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": "2301.02344", "title": "TrojanPuzzle: Covertly Poisoning Code-Suggestion Models", "abstract": " With tools like GitHub Copilot, automatic code suggestion is no longer a\ndream in software engineering. These tools, based on large language models, are\ntypically trained on massive corpora of code mined from unvetted public\nsources. As a result, these models are susceptible to data poisoning attacks\nwhere an adversary manipulates the model's training by injecting malicious\ndata. Poisoning attacks could be designed to influence the model's suggestions\nat run time for chosen contexts, such as inducing the model into suggesting\ninsecure code payloads. To achieve this, prior attacks explicitly inject the\ninsecure code payload into the training data, making the poison data detectable\nby static analysis tools that can remove such malicious data from the training\nset. In this work, we demonstrate two novel attacks, COVERT and TROJANPUZZLE,\nthat can bypass static analysis by planting malicious poison data in\nout-of-context regions such as docstrings. Our most novel attack, TROJANPUZZLE,\ngoes one step further in generating less suspicious poison data by never\nexplicitly including certain (suspicious) parts of the payload in the poison\ndata, while still inducing a model that suggests the entire payload when\ncompleting code (i.e., outside docstrings). This makes TROJANPUZZLE robust\nagainst signature-based dataset-cleansing methods that can filter out\nsuspicious sequences from the training data. Our evaluation against models of\ntwo sizes demonstrates that both COVERT and TROJANPUZZLE have significant\nimplications for practitioners when selecting code used to train or tune\ncode-suggestion models.\n", "date": "2023-01-06", "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": "2308.09487", "title": "DFB: A Data-Free, Low-Budget, and High-Efficacy Clean-Label Backdoor\n Attack", "abstract": " In the domain of backdoor attacks, accurate labeling of injected data is\nessential for evading rudimentary detection mechanisms. This imperative has\ncatalyzed the development of clean-label attacks, which are notably more\nelusive as they preserve the original labels of the injected data. Current\nclean-label attack methodologies primarily depend on extensive knowledge of the\ntraining dataset. However, practically, such comprehensive dataset access is\noften unattainable, given that training datasets are typically compiled from\nvarious independent sources. Departing from conventional clean-label attack\nmethodologies, our research introduces DFB, a data-free, low-budget, and\nhigh-efficacy clean-label backdoor Attack. DFB is unique in its independence\nfrom training data access, requiring solely the knowledge of a specific target\nclass. Tested on CIFAR10, Tiny-ImageNet, and TSRD, DFB demonstrates remarkable\nefficacy with minimal poisoning rates of just 0.1%, 0.025%, and 0.4%,\nrespectively. These rates are significantly lower than those required by\nexisting methods such as LC, HTBA, BadNets, and Blend, yet DFB achieves\nsuperior attack success rates. Furthermore, our findings reveal that DFB poses\na formidable challenge to four established backdoor defense algorithms,\nindicating its potential as a robust tool in advanced clean-label attack\nstrategies.\n", "date": "2023-08-18", "categories": []} {"id": "2401.05432", "title": "TEN-GUARD: Tensor Decomposition for Backdoor Attack Detection in Deep\n Neural Networks", "abstract": " As deep neural networks and the datasets used to train them get larger, the\ndefault approach to integrating them into research and commercial projects is\nto download a pre-trained model and fine tune it. But these models can have\nuncertain provenance, opening up the possibility that they embed hidden\nmalicious behavior such as trojans or backdoors, where small changes to an\ninput (triggers) can cause the model to produce incorrect outputs (e.g., to\nmisclassify). This paper introduces a novel approach to backdoor detection that\nuses two tensor decomposition methods applied to network activations. This has\na number of advantages relative to existing detection methods, including the\nability to analyze multiple models at the same time, working across a wide\nvariety of network architectures, making no assumptions about the nature of\ntriggers used to alter network behavior, and being computationally efficient.\nWe provide a detailed description of the detection pipeline along with results\non models trained on the MNIST digit dataset, CIFAR-10 dataset, and two\ndifficult datasets from NIST's TrojAI competition. These results show that our\nmethod detects backdoored networks more accurately and efficiently than current\nstate-of-the-art methods.\n", "date": "2024-01-06", "categories": []} {"id": "2401.03215", "title": "End-to-End Anti-Backdoor Learning on Images and Time Series", "abstract": " Backdoor attacks present a substantial security concern for deep learning\nmodels, especially those utilized in applications critical to safety and\nsecurity. These attacks manipulate model behavior by embedding a hidden trigger\nduring the training phase, allowing unauthorized control over the model's\noutput during inference time. Although numerous defenses exist for image\nclassification models, there is a conspicuous absence of defenses tailored for\ntime series data, as well as an end-to-end solution capable of training clean\nmodels on poisoned data. To address this gap, this paper builds upon\nAnti-Backdoor Learning (ABL) and introduces an innovative method, End-to-End\nAnti-Backdoor Learning (E2ABL), for robust training against backdoor attacks.\nUnlike the original ABL, which employs a two-stage training procedure, E2ABL\naccomplishes end-to-end training through an additional classification head\nlinked to the shallow layers of a Deep Neural Network (DNN). This secondary\nhead actively identifies potential backdoor triggers, allowing the model to\ndynamically cleanse these samples and their corresponding labels during\ntraining. Our experiments reveal that E2ABL significantly improves on existing\ndefenses and is effective against a broad range of backdoor attacks in both\nimage and time series domains.\n", "date": "2024-01-06", "categories": []} {"id": "2310.17498", "title": "CBD: A Certified Backdoor Detector Based on Local Dominant Probability", "abstract": " Backdoor attack is a common threat to deep neural networks. During testing,\nsamples embedded with a backdoor trigger will be misclassified as an\nadversarial target by a backdoored model, while samples without the backdoor\ntrigger will be correctly classified. In this paper, we present the first\ncertified backdoor detector (CBD), which is based on a novel, adjustable\nconformal prediction scheme based on our proposed statistic local dominant\nprobability. For any classifier under inspection, CBD provides 1) a detection\ninference, 2) the condition under which the attacks are guaranteed to be\ndetectable for the same classification domain, and 3) a probabilistic upper\nbound for the false positive rate. Our theoretical results show that attacks\nwith triggers that are more resilient to test-time noise and have smaller\nperturbation magnitudes are more likely to be detected with guarantees.\nMoreover, we conduct extensive experiments on four benchmark datasets\nconsidering various backdoor types, such as BadNet, CB, and Blend. CBD achieves\ncomparable or even higher detection accuracy than state-of-the-art detectors,\nand it in addition provides detection certification. Notably, for backdoor\nattacks with random perturbation triggers bounded by $\\ell_2\\leq0.75$ which\nachieves more than 90\\% attack success rate, CBD achieves 100\\% (98\\%), 100\\%\n(84\\%), 98\\% (98\\%), and 72\\% (40\\%) empirical (certified) detection true\npositive rates on the four benchmark datasets GTSRB, SVHN, CIFAR-10, and\nTinyImageNet, respectively, with low false positive rates.\n", "date": "2023-10-26", "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": "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": "2312.15867", "title": "Punctuation Matters! Stealthy Backdoor Attack for Language Models", "abstract": " Recent studies have pointed out that natural language processing (NLP) models\nare vulnerable to backdoor attacks. A backdoored model produces normal outputs\non the clean samples while performing improperly on the texts with triggers\nthat the adversary injects. However, previous studies on textual backdoor\nattack pay little attention to stealthiness. Moreover, some attack methods even\ncause grammatical issues or change the semantic meaning of the original texts.\nTherefore, they can easily be detected by humans or defense systems. In this\npaper, we propose a novel stealthy backdoor attack method against textual\nmodels, which is called \\textbf{PuncAttack}. It leverages combinations of\npunctuation marks as the trigger and chooses proper locations strategically to\nreplace them. Through extensive experiments, we demonstrate that the proposed\nmethod can effectively compromise multiple models in various tasks. Meanwhile,\nwe conduct automatic evaluation and human inspection, which indicate the\nproposed method possesses good performance of stealthiness without bringing\ngrammatical issues and altering the meaning of sentences.\n", "date": "2023-12-26", "categories": []} {"id": "2312.11057", "title": "DataElixir: Purifying Poisoned Dataset to Mitigate Backdoor Attacks via\n Diffusion Models", "abstract": " Dataset sanitization is a widely adopted proactive defense against\npoisoning-based backdoor attacks, aimed at filtering out and removing poisoned\nsamples from training datasets. However, existing methods have shown limited\nefficacy in countering the ever-evolving trigger functions, and often leading\nto considerable degradation of benign accuracy. In this paper, we propose\nDataElixir, a novel sanitization approach tailored to purify poisoned datasets.\nWe leverage diffusion models to eliminate trigger features and restore benign\nfeatures, thereby turning the poisoned samples into benign ones. Specifically,\nwith multiple iterations of the forward and reverse process, we extract\nintermediary images and their predicted labels for each sample in the original\ndataset. Then, we identify anomalous samples in terms of the presence of label\ntransition of the intermediary images, detect the target label by quantifying\ndistribution discrepancy, select their purified images considering pixel and\nfeature distance, and determine their ground-truth labels by training a benign\nmodel. Experiments conducted on 9 popular attacks demonstrates that DataElixir\neffectively mitigates various complex attacks while exerting minimal impact on\nbenign accuracy, surpassing the performance of baseline defense methods.\n", "date": "2023-12-18", "categories": []} {"id": "2312.12585", "title": "BadRL: Sparse Targeted Backdoor Attack Against Reinforcement Learning", "abstract": " Backdoor attacks in reinforcement learning (RL) have previously employed\nintense attack strategies to ensure attack success. However, these methods\nsuffer from high attack costs and increased detectability. In this work, we\npropose a novel approach, BadRL, which focuses on conducting highly sparse\nbackdoor poisoning efforts during training and testing while maintaining\nsuccessful attacks. Our algorithm, BadRL, strategically chooses state\nobservations with high attack values to inject triggers during training and\ntesting, thereby reducing the chances of detection. In contrast to the previous\nmethods that utilize sample-agnostic trigger patterns, BadRL dynamically\ngenerates distinct trigger patterns based on targeted state observations,\nthereby enhancing its effectiveness. Theoretical analysis shows that the\ntargeted backdoor attack is always viable and remains stealthy under specific\nassumptions. Empirical results on various classic RL tasks illustrate that\nBadRL can substantially degrade the performance of a victim agent with minimal\npoisoning efforts 0.003% of total training steps) during training and\ninfrequent attacks during testing.\n", "date": "2023-12-19", "categories": []} {"id": "2312.12724", "title": "Progressive Poisoned Data Isolation for Training-time Backdoor Defense", "abstract": " Deep Neural Networks (DNN) are susceptible to backdoor attacks where\nmalicious attackers manipulate the model's predictions via data poisoning. It\nis hence imperative to develop a strategy for training a clean model using a\npotentially poisoned dataset. Previous training-time defense mechanisms\ntypically employ an one-time isolation process, often leading to suboptimal\nisolation outcomes. In this study, we present a novel and efficacious defense\nmethod, termed Progressive Isolation of Poisoned Data (PIPD), that\nprogressively isolates poisoned data to enhance the isolation accuracy and\nmitigate the risk of benign samples being misclassified as poisoned ones. Once\nthe poisoned portion of the dataset has been identified, we introduce a\nselective training process to train a clean model. Through the implementation\nof these techniques, we ensure that the trained model manifests a significantly\ndiminished attack success rate against the poisoned data. Extensive experiments\non multiple benchmark datasets and DNN models, assessed against nine\nstate-of-the-art backdoor attacks, demonstrate the superior performance of our\nPIPD method for backdoor defense. For instance, our PIPD achieves an average\nTrue Positive Rate (TPR) of 99.95% and an average False Positive Rate (FPR) of\n0.06% for diverse attacks over CIFAR-10 dataset, markedly surpassing the\nperformance of state-of-the-art methods.\n", "date": "2023-12-20", "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": "2310.12439", "title": "PoisonPrompt: Backdoor Attack on Prompt-based Large Language Models", "abstract": " Prompts have significantly improved the performance of pretrained Large\nLanguage Models (LLMs) on various downstream tasks recently, making them\nincreasingly indispensable for a diverse range of LLM application scenarios.\nHowever, the backdoor vulnerability, a serious security threat that can\nmaliciously alter the victim model's normal predictions, has not been\nsufficiently explored for prompt-based LLMs. In this paper, we present\nPOISONPROMPT, a novel backdoor attack capable of successfully compromising both\nhard and soft prompt-based LLMs. We evaluate the effectiveness, fidelity, and\nrobustness of POISONPROMPT through extensive experiments on three popular\nprompt methods, using six datasets and three widely used LLMs. Our findings\nhighlight the potential security threats posed by backdoor attacks on\nprompt-based LLMs and emphasize the need for further research in this area.\n", "date": "2023-10-19", "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": "2310.10659", "title": "Exploiting Machine Unlearning for Backdoor Attacks in Deep Learning\n System", "abstract": " In recent years, the security issues of artificial intelligence have become\nincreasingly prominent due to the rapid development of deep learning research\nand applications. Backdoor attack is an attack targeting the vulnerability of\ndeep learning models, where hidden backdoors are activated by triggers embedded\nby the attacker, thereby outputting malicious predictions that may not align\nwith the intended output for a given input. In this work, we propose a novel\nblack-box backdoor attack based on machine unlearning. The attacker first\naugments the training set with carefully designed samples, including poison and\nmitigation data, to train a `benign' model. Then, the attacker posts unlearning\nrequests for the mitigation samples to remove the impact of relevant data on\nthe model, gradually activating the hidden backdoor. Since backdoors are\nimplanted during the iterative unlearning process, it significantly increases\nthe computational overhead of existing defense methods for backdoor detection\nor mitigation. To address this new security threat, we proposes two methods for\ndetecting or mitigating such malicious unlearning requests. We conduct the\nexperiment in both exact unlearning and approximate unlearning (i.e., SISA)\nsettings. Experimental results indicate that: 1) our attack approach can\nsuccessfully implant backdoor into the model, and sharding increases the\ndifficult of attack; 2) our detection algorithms are effective in identifying\nthe mitigation samples, while sharding reduces the effectiveness of our\ndetection algorithms.\n", "date": "2023-09-12", "categories": []} {"id": "2301.04554", "title": "Universal Detection of Backdoor Attacks via Density-based Clustering and\n Centroids Analysis", "abstract": " We propose a Universal Defence against backdoor attacks based on Clustering\nand Centroids Analysis (CCA-UD). The goal of the defence is to reveal whether a\nDeep Neural Network model is subject to a backdoor attack by inspecting the\ntraining dataset. CCA-UD first clusters the samples of the training set by\nmeans of density-based clustering. Then, it applies a novel strategy to detect\nthe presence of poisoned clusters. The proposed strategy is based on a general\nmisclassification behaviour observed when the features of a representative\nexample of the analysed cluster are added to benign samples. The capability of\ninducing a misclassification error is a general characteristic of poisoned\nsamples, hence the proposed defence is attack-agnostic. This marks a\nsignificant difference with respect to existing defences, that, either can\ndefend against only some types of backdoor attacks, or are effective only when\nsome conditions on the poisoning ratio or the kind of triggering signal used by\nthe attacker are satisfied.\n Experiments carried out on several classification tasks and network\narchitectures, considering different types of backdoor attacks (with either\nclean or corrupted labels), and triggering signals, including both global and\nlocal triggering signals, as well as sample-specific and source-specific\ntriggers, reveal that the proposed method is very effective to defend against\nbackdoor attacks in all the cases, always outperforming the state of the art\ntechniques.\n", "date": "2023-01-11", "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": "2305.14876", "title": "Reconstructive Neuron Pruning for Backdoor Defense", "abstract": " Deep neural networks (DNNs) have been found to be vulnerable to backdoor\nattacks, raising security concerns about their deployment in mission-critical\napplications. While existing defense methods have demonstrated promising\nresults, it is still not clear how to effectively remove backdoor-associated\nneurons in backdoored DNNs. In this paper, we propose a novel defense called\n\\emph{Reconstructive Neuron Pruning} (RNP) to expose and prune backdoor neurons\nvia an unlearning and then recovering process. Specifically, RNP first unlearns\nthe neurons by maximizing the model's error on a small subset of clean samples\nand then recovers the neurons by minimizing the model's error on the same data.\nIn RNP, unlearning is operated at the neuron level while recovering is operated\nat the filter level, forming an asymmetric reconstructive learning procedure.\nWe show that such an asymmetric process on only a few clean samples can\neffectively expose and prune the backdoor neurons implanted by a wide range of\nattacks, achieving a new state-of-the-art defense performance. Moreover, the\nunlearned model at the intermediate step of our RNP can be directly used to\nimprove other backdoor defense tasks including backdoor removal, trigger\nrecovery, backdoor label detection, and backdoor sample detection. Code is\navailable at \\url{https://github.com/bboylyg/RNP}.\n", "date": "2023-05-24", "categories": []} {"id": "2311.17429", "title": "TARGET: Template-Transferable Backdoor Attack Against Prompt-based NLP\n Models via GPT4", "abstract": " Prompt-based learning has been widely applied in many low-resource NLP tasks\nsuch as few-shot scenarios. However, this paradigm has been shown to be\nvulnerable to backdoor attacks. Most of the existing attack methods focus on\ninserting manually predefined templates as triggers in the pre-training phase\nto train the victim model and utilize the same triggers in the downstream task\nto perform inference, which tends to ignore the transferability and\nstealthiness of the templates. In this work, we propose a novel approach of\nTARGET (Template-trAnsfeRable backdoor attack aGainst prompt-basEd NLP models\nvia GPT4), which is a data-independent attack method. Specifically, we first\nutilize GPT4 to reformulate manual templates to generate tone-strong and normal\ntemplates, and the former are injected into the model as a backdoor trigger in\nthe pre-training phase. Then, we not only directly employ the above templates\nin the downstream task, but also use GPT4 to generate templates with similar\ntone to the above templates to carry out transferable attacks. Finally we have\nconducted extensive experiments on five NLP datasets and three BERT series\nmodels, with experimental results justifying that our TARGET method has better\nattack performance and stealthiness compared to the two-external baseline\nmethods on direct attacks, and in addition achieves satisfactory attack\ncapability in the unseen tone-similar templates.\n", "date": "2023-11-29", "categories": []} {"id": "2311.11225", "title": "TextGuard: Provable Defense against Backdoor Attacks on Text\n Classification", "abstract": " Backdoor attacks have become a major security threat for deploying machine\nlearning models in security-critical applications. Existing research endeavors\nhave proposed many defenses against backdoor attacks. Despite demonstrating\ncertain empirical defense efficacy, none of these techniques could provide a\nformal and provable security guarantee against arbitrary attacks. As a result,\nthey can be easily broken by strong adaptive attacks, as shown in our\nevaluation. In this work, we propose TextGuard, the first provable defense\nagainst backdoor attacks on text classification. In particular, TextGuard first\ndivides the (backdoored) training data into sub-training sets, achieved by\nsplitting each training sentence into sub-sentences. This partitioning ensures\nthat a majority of the sub-training sets do not contain the backdoor trigger.\nSubsequently, a base classifier is trained from each sub-training set, and\ntheir ensemble provides the final prediction. We theoretically prove that when\nthe length of the backdoor trigger falls within a certain threshold, TextGuard\nguarantees that its prediction will remain unaffected by the presence of the\ntriggers in training and testing inputs. In our evaluation, we demonstrate the\neffectiveness of TextGuard on three benchmark text classification tasks,\nsurpassing the certification accuracy of existing certified defenses against\nbackdoor attacks. Furthermore, we propose additional strategies to enhance the\nempirical performance of TextGuard. Comparisons with state-of-the-art empirical\ndefenses validate the superiority of TextGuard in countering multiple backdoor\nattacks. Our code and data are available at\nhttps://github.com/AI-secure/TextGuard.\n", "date": "2023-11-19", "categories": []} {"id": "2311.13957", "title": "Efficient Trigger Word Insertion", "abstract": " With the boom in the natural language processing (NLP) field these years,\nbackdoor attacks pose immense threats against deep neural network models.\nHowever, previous works hardly consider the effect of the poisoning rate. In\nthis paper, our main objective is to reduce the number of poisoned samples\nwhile still achieving a satisfactory Attack Success Rate (ASR) in text backdoor\nattacks. To accomplish this, we propose an efficient trigger word insertion\nstrategy in terms of trigger word optimization and poisoned sample selection.\nExtensive experiments on different datasets and models demonstrate that our\nproposed method can significantly improve attack effectiveness in text\nclassification tasks. Remarkably, our approach achieves an ASR of over 90% with\nonly 10 poisoned samples in the dirty-label setting and requires merely 1.5% of\nthe training data in the clean-label setting.\n", "date": "2023-11-23", "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": "2305.18651", "title": "UMD: Unsupervised Model Detection for X2X Backdoor Attacks", "abstract": " Backdoor (Trojan) attack is a common threat to deep neural networks, where\nsamples from one or more source classes embedded with a backdoor trigger will\nbe misclassified to adversarial target classes. Existing methods for detecting\nwhether a classifier is backdoor attacked are mostly designed for attacks with\na single adversarial target (e.g., all-to-one attack). To the best of our\nknowledge, without supervision, no existing methods can effectively address the\nmore general X2X attack with an arbitrary number of source classes, each paired\nwith an arbitrary target class. In this paper, we propose UMD, the first\nUnsupervised Model Detection method that effectively detects X2X backdoor\nattacks via a joint inference of the adversarial (source, target) class pairs.\nIn particular, we first define a novel transferability statistic to measure and\nselect a subset of putative backdoor class pairs based on a proposed clustering\napproach. Then, these selected class pairs are jointly assessed based on an\naggregation of their reverse-engineered trigger size for detection inference,\nusing a robust and unsupervised anomaly detector we proposed. We conduct\ncomprehensive evaluations on CIFAR-10, GTSRB, and Imagenette dataset, and show\nthat our unsupervised UMD outperforms SOTA detectors (even with supervision) by\n17%, 4%, and 8%, respectively, in terms of the detection accuracy against\ndiverse X2X attacks. We also show the strong detection performance of UMD\nagainst several strong adaptive attacks.\n", "date": "2023-05-29", "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": "2303.12175", "title": "Black-box Backdoor Defense via Zero-shot Image Purification", "abstract": " Backdoor attacks inject poisoned samples into the training data, resulting in\nthe misclassification of the poisoned input during a model's deployment.\nDefending against such attacks is challenging, especially for real-world\nblack-box models where only query access is permitted. In this paper, we\npropose a novel defense framework against backdoor attacks through Zero-shot\nImage Purification (ZIP). Our framework can be applied to poisoned models\nwithout requiring internal information about the model or any prior knowledge\nof the clean/poisoned samples. Our defense framework involves two steps. First,\nwe apply a linear transformation (e.g., blurring) on the poisoned image to\ndestroy the backdoor pattern. Then, we use a pre-trained diffusion model to\nrecover the missing semantic information removed by the transformation. In\nparticular, we design a new reverse process by using the transformed image to\nguide the generation of high-fidelity purified images, which works in zero-shot\nsettings. We evaluate our ZIP framework on multiple datasets with different\ntypes of attacks. Experimental results demonstrate the superiority of our ZIP\nframework compared to state-of-the-art backdoor defense baselines. We believe\nthat our results will provide valuable insights for future defense methods for\nblack-box models. Our code is available at https://github.com/sycny/ZIP.\n", "date": "2023-03-21", "categories": []} {"id": "2304.11823", "title": "Enhancing Fine-Tuning Based Backdoor Defense with Sharpness-Aware\n Minimization", "abstract": " Backdoor defense, which aims to detect or mitigate the effect of malicious\ntriggers introduced by attackers, is becoming increasingly critical for machine\nlearning security and integrity. Fine-tuning based on benign data is a natural\ndefense to erase the backdoor effect in a backdoored model. However, recent\nstudies show that, given limited benign data, vanilla fine-tuning has poor\ndefense performance. In this work, we provide a deep study of fine-tuning the\nbackdoored model from the neuron perspective and find that backdoorrelated\nneurons fail to escape the local minimum in the fine-tuning process. Inspired\nby observing that the backdoorrelated neurons often have larger norms, we\npropose FTSAM, a novel backdoor defense paradigm that aims to shrink the norms\nof backdoor-related neurons by incorporating sharpness-aware minimization with\nfine-tuning. We demonstrate the effectiveness of our method on several\nbenchmark datasets and network architectures, where it achieves\nstate-of-the-art defense performance. Overall, our work provides a promising\navenue for improving the robustness of machine learning models against backdoor\nattacks.\n", "date": "2023-04-24", "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": "2310.18633", "title": "Setting the Trap: Capturing and Defeating Backdoors in Pretrained\n Language Models through Honeypots", "abstract": " In the field of natural language processing, the prevalent approach involves\nfine-tuning pretrained language models (PLMs) using local samples. Recent\nresearch has exposed the susceptibility of PLMs to backdoor attacks, wherein\nthe adversaries can embed malicious prediction behaviors by manipulating a few\ntraining samples. In this study, our objective is to develop a\nbackdoor-resistant tuning procedure that yields a backdoor-free model, no\nmatter whether the fine-tuning dataset contains poisoned samples. To this end,\nwe propose and integrate a honeypot module into the original PLM, specifically\ndesigned to absorb backdoor information exclusively. Our design is motivated by\nthe observation that lower-layer representations in PLMs carry sufficient\nbackdoor features while carrying minimal information about the original tasks.\nConsequently, we can impose penalties on the information acquired by the\nhoneypot module to inhibit backdoor creation during the fine-tuning process of\nthe stem network. Comprehensive experiments conducted on benchmark datasets\nsubstantiate the effectiveness and robustness of our defensive strategy.\nNotably, these results indicate a substantial reduction in the attack success\nrate ranging from 10\\% to 40\\% when compared to prior state-of-the-art 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": "2308.02122", "title": "ParaFuzz: An Interpretability-Driven Technique for Detecting Poisoned\n Samples in NLP", "abstract": " Backdoor attacks have emerged as a prominent threat to natural language\nprocessing (NLP) models, where the presence of specific triggers in the input\ncan lead poisoned models to misclassify these inputs to predetermined target\nclasses. Current detection mechanisms are limited by their inability to address\nmore covert backdoor strategies, such as style-based attacks. In this work, we\npropose an innovative test-time poisoned sample detection framework that hinges\non the interpretability of model predictions, grounded in the semantic meaning\nof inputs. We contend that triggers (e.g., infrequent words) are not supposed\nto fundamentally alter the underlying semantic meanings of poisoned samples as\nthey want to stay stealthy. Based on this observation, we hypothesize that\nwhile the model's predictions for paraphrased clean samples should remain\nstable, predictions for poisoned samples should revert to their true labels\nupon the mutations applied to triggers during the paraphrasing process. We\nemploy ChatGPT, a state-of-the-art large language model, as our paraphraser and\nformulate the trigger-removal task as a prompt engineering problem. We adopt\nfuzzing, a technique commonly used for unearthing software vulnerabilities, to\ndiscover optimal paraphrase prompts that can effectively eliminate triggers\nwhile concurrently maintaining input semantics. Experiments on 4 types of\nbackdoor attacks, including the subtle style backdoors, and 4 distinct datasets\ndemonstrate that our approach surpasses baseline methods, including STRIP, RAP,\nand ONION, in precision and recall.\n", "date": "2023-08-04", "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": "2310.01875", "title": "Towards Stable Backdoor Purification through Feature Shift Tuning", "abstract": " It has been widely observed that deep neural networks (DNN) are vulnerable to\nbackdoor attacks where attackers could manipulate the model behavior\nmaliciously by tampering with a small set of training samples. Although a line\nof defense methods is proposed to mitigate this threat, they either require\ncomplicated modifications to the training process or heavily rely on the\nspecific model architecture, which makes them hard to deploy into real-world\napplications. Therefore, in this paper, we instead start with fine-tuning, one\nof the most common and easy-to-deploy backdoor defenses, through comprehensive\nevaluations against diverse attack scenarios. Observations made through initial\nexperiments show that in contrast to the promising defensive results on high\npoisoning rates, vanilla tuning methods completely fail at low poisoning rate\nscenarios. Our analysis shows that with the low poisoning rate, the\nentanglement between backdoor and clean features undermines the effect of\ntuning-based defenses. Therefore, it is necessary to disentangle the backdoor\nand clean features in order to improve backdoor purification. To address this,\nwe introduce Feature Shift Tuning (FST), a method for tuning-based backdoor\npurification. Specifically, FST encourages feature shifts by actively deviating\nthe classifier weights from the originally compromised weights. Extensive\nexperiments demonstrate that our FST provides consistently stable performance\nunder different attack settings. Without complex parameter adjustments, FST\nalso achieves much lower tuning costs, only 10 epochs. Our codes are available\nat https://github.com/AISafety-HKUST/stable_backdoor_purification.\n", "date": "2023-10-03", "categories": []} {"id": "2305.11596", "title": "Mitigating Backdoor Poisoning Attacks through the Lens of Spurious\n Correlation", "abstract": " Modern NLP models are often trained over large untrusted datasets, raising\nthe potential for a malicious adversary to compromise model behaviour. For\ninstance, backdoors can be implanted through crafting training instances with a\nspecific textual trigger and a target label. This paper posits that backdoor\npoisoning attacks exhibit \\emph{spurious correlation} between simple text\nfeatures and classification labels, and accordingly, proposes methods for\nmitigating spurious correlation as means of defence. Our empirical study\nreveals that the malicious triggers are highly correlated to their target\nlabels; therefore such correlations are extremely distinguishable compared to\nthose scores of benign features, and can be used to filter out potentially\nproblematic instances. Compared with several existing defences, our defence\nmethod significantly reduces attack success rates across backdoor attacks, and\nin the case of insertion-based attacks, our method provides a near-perfect\ndefence.\n", "date": "2023-05-19", "categories": []} {"id": "2310.11595", "title": "WaveAttack: Asymmetric Frequency Obfuscation-based Backdoor Attacks\n Against Deep Neural Networks", "abstract": " Due to the popularity of Artificial Intelligence (AI) technology, numerous\nbackdoor attacks are designed by adversaries to mislead deep neural network\npredictions by manipulating training samples and training processes. Although\nbackdoor attacks are effective in various real scenarios, they still suffer\nfrom the problems of both low fidelity of poisoned samples and non-negligible\ntransfer in latent space, which make them easily detectable by existing\nbackdoor detection algorithms. To overcome the weakness, this paper proposes a\nnovel frequency-based backdoor attack method named WaveAttack, which obtains\nimage high-frequency features through Discrete Wavelet Transform (DWT) to\ngenerate backdoor triggers. Furthermore, we introduce an asymmetric frequency\nobfuscation method, which can add an adaptive residual in the training and\ninference stage to improve the impact of triggers and further enhance the\neffectiveness of WaveAttack. Comprehensive experimental results show that\nWaveAttack not only achieves higher stealthiness and effectiveness, but also\noutperforms state-of-the-art (SOTA) backdoor attack methods in the fidelity of\nimages by up to 28.27\\% improvement in PSNR, 1.61\\% improvement in SSIM, and\n70.59\\% reduction in IS.\n", "date": "2023-10-17", "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": "2308.13904", "title": "LMSanitator: Defending Prompt-Tuning Against Task-Agnostic Backdoors", "abstract": " Prompt-tuning has emerged as an attractive paradigm for deploying large-scale\nlanguage models due to its strong downstream task performance and efficient\nmultitask serving ability. Despite its wide adoption, we empirically show that\nprompt-tuning is vulnerable to downstream task-agnostic backdoors, which reside\nin the pretrained models and can affect arbitrary downstream tasks. The\nstate-of-the-art backdoor detection approaches cannot defend against\ntask-agnostic backdoors since they hardly converge in reversing the backdoor\ntriggers. To address this issue, we propose LMSanitator, a novel approach for\ndetecting and removing task-agnostic backdoors on Transformer models. Instead\nof directly inverting the triggers, LMSanitator aims to invert the predefined\nattack vectors (pretrained models' output when the input is embedded with\ntriggers) of the task-agnostic backdoors, which achieves much better\nconvergence performance and backdoor detection accuracy. LMSanitator further\nleverages prompt-tuning's property of freezing the pretrained model to perform\naccurate and fast output monitoring and input purging during the inference\nphase. Extensive experiments on multiple language models and NLP tasks\nillustrate the effectiveness of LMSanitator. For instance, LMSanitator achieves\n92.8% backdoor detection accuracy on 960 models and decreases the attack\nsuccess rate to less than 1% in most scenarios.\n", "date": "2023-08-26", "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": "2308.12439", "title": "BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input\n Detection", "abstract": " We present a novel defense, against backdoor attacks on Deep Neural Networks\n(DNNs), wherein adversaries covertly implant malicious behaviors (backdoors)\ninto DNNs. Our defense falls within the category of post-development defenses\nthat operate independently of how the model was generated. The proposed defense\nis built upon a novel reverse engineering approach that can directly extract\nbackdoor functionality of a given backdoored model to a backdoor expert model.\nThe approach is straightforward -- finetuning the backdoored model over a small\nset of intentionally mislabeled clean samples, such that it unlearns the normal\nfunctionality while still preserving the backdoor functionality, and thus\nresulting in a model (dubbed a backdoor expert model) that can only recognize\nbackdoor inputs. Based on the extracted backdoor expert model, we show the\nfeasibility of devising highly accurate backdoor input detectors that filter\nout the backdoor inputs during model inference. Further augmented by an\nensemble strategy with a finetuned auxiliary model, our defense, BaDExpert\n(Backdoor Input Detection with Backdoor Expert), effectively mitigates 17 SOTA\nbackdoor attacks while minimally impacting clean utility. The effectiveness of\nBaDExpert has been verified on multiple datasets (CIFAR10, GTSRB and ImageNet)\nacross various model architectures (ResNet, VGG, MobileNetV2 and Vision\nTransformer).\n", "date": "2023-08-23", "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": "2309.00127", "title": "FTA: Stealthy and Adaptive Backdoor Attack with Flexible Triggers on\n Federated Learning", "abstract": " Current backdoor attacks against federated learning (FL) strongly rely on\nuniversal triggers or semantic patterns, which can be easily detected and\nfiltered by certain defense mechanisms such as norm clipping, comparing\nparameter divergences among local updates. In this work, we propose a new\nstealthy and robust backdoor attack with flexible triggers against FL defenses.\nTo achieve this, we build a generative trigger function that can learn to\nmanipulate the benign samples with an imperceptible flexible trigger pattern\nand simultaneously make the trigger pattern include the most significant hidden\nfeatures of the attacker-chosen label. Moreover, our trigger generator can keep\nlearning and adapt across different rounds, allowing it to adjust to changes in\nthe global model. By filling the distinguishable difference (the mapping\nbetween the trigger pattern and target label), we make our attack naturally\nstealthy. Extensive experiments on real-world datasets verify the effectiveness\nand stealthiness of our attack compared to prior attacks on decentralized\nlearning framework with eight well-studied defenses.\n", "date": "2023-08-31", "categories": []} {"id": "2309.13256", "title": "Defending Pre-trained Language Models as Few-shot Learners against\n Backdoor Attacks", "abstract": " Pre-trained language models (PLMs) have demonstrated remarkable performance\nas few-shot learners. However, their security risks under such settings are\nlargely unexplored. In this work, we conduct a pilot study showing that PLMs as\nfew-shot learners are highly vulnerable to backdoor attacks while existing\ndefenses are inadequate due to the unique challenges of few-shot scenarios. To\naddress such challenges, we advocate MDP, a novel lightweight, pluggable, and\neffective defense for PLMs as few-shot learners. Specifically, MDP leverages\nthe gap between the masking-sensitivity of poisoned and clean samples: with\nreference to the limited few-shot data as distributional anchors, it compares\nthe representations of given samples under varying masking and identifies\npoisoned samples as ones with significant variations. We show analytically that\nMDP creates an interesting dilemma for the attacker to choose between attack\neffectiveness and detection evasiveness. The empirical evaluation using\nbenchmark datasets and representative attacks validates the efficacy of MDP.\n", "date": "2023-09-23", "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": "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": "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": "2309.06981", "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": "2308.16684", "title": "Everyone Can Attack: Repurpose Lossy Compression as a Natural Backdoor\n Attack", "abstract": " The vulnerabilities to backdoor attacks have recently threatened the\ntrustworthiness of machine learning models in practical applications.\nConventional wisdom suggests that not everyone can be an attacker since the\nprocess of designing the trigger generation algorithm often involves\nsignificant effort and extensive experimentation to ensure the attack's\nstealthiness and effectiveness. Alternatively, this paper shows that there\nexists a more severe backdoor threat: anyone can exploit an easily-accessible\nalgorithm for silent backdoor attacks. Specifically, this attacker can employ\nthe widely-used lossy image compression from a plethora of compression tools to\neffortlessly inject a trigger pattern into an image without leaving any\nnoticeable trace; i.e., the generated triggers are natural artifacts. One does\nnot require extensive knowledge to click on the \"convert\" or \"save as\" button\nwhile using tools for lossy image compression. Via this attack, the adversary\ndoes not need to design a trigger generator as seen in prior works and only\nrequires poisoning the data. Empirically, the proposed attack consistently\nachieves 100% attack success rate in several benchmark datasets such as MNIST,\nCIFAR-10, GTSRB and CelebA. More significantly, the proposed attack can still\nachieve almost 100% attack success rate with very small (approximately 10%)\npoisoning rates in the clean label setting. The generated trigger of the\nproposed attack using one lossy compression algorithm is also transferable\nacross other related compression algorithms, exacerbating the severity of this\nbackdoor threat. This work takes another crucial step toward understanding the\nextensive risks of backdoor attacks in practice, urging practitioners to\ninvestigate similar attacks and relevant backdoor mitigation methods.\n", "date": "2023-08-31", "categories": []} {"id": "2302.00747", "title": "Universal Soldier: Using Universal Adversarial Perturbations for\n Detecting Backdoor Attacks", "abstract": " Deep learning models achieve excellent performance in numerous machine\nlearning tasks. Yet, they suffer from security-related issues such as\nadversarial examples and poisoning (backdoor) attacks. A deep learning model\nmay be poisoned by training with backdoored data or by modifying inner network\nparameters. Then, a backdoored model performs as expected when receiving a\nclean input, but it misclassifies when receiving a backdoored input stamped\nwith a pre-designed pattern called \"trigger\". Unfortunately, it is difficult to\ndistinguish between clean and backdoored models without prior knowledge of the\ntrigger. This paper proposes a backdoor detection method by utilizing a special\ntype of adversarial attack, universal adversarial perturbation (UAP), and its\nsimilarities with a backdoor trigger. We observe an intuitive phenomenon: UAPs\ngenerated from backdoored models need fewer perturbations to mislead the model\nthan UAPs from clean models. UAPs of backdoored models tend to exploit the\nshortcut from all classes to the target class, built by the backdoor trigger.\nWe propose a novel method called Universal Soldier for Backdoor detection (USB)\nand reverse engineering potential backdoor triggers via UAPs. Experiments on\n345 models trained on several datasets show that USB effectively detects the\ninjected backdoor and provides comparable or better results than\nstate-of-the-art methods.\n", "date": "2023-02-01", "categories": []} {"id": "2304.08847", "title": "BadVFL: Backdoor Attacks in Vertical Federated Learning", "abstract": " Federated learning (FL) enables multiple parties to collaboratively train a\nmachine learning model without sharing their data; rather, they train their own\nmodel locally and send updates to a central server for aggregation. Depending\non how the data is distributed among the participants, FL can be classified\ninto Horizontal (HFL) and Vertical (VFL). In VFL, the participants share the\nsame set of training instances but only host a different and non-overlapping\nsubset of the whole feature space. Whereas in HFL, each participant shares the\nsame set of features while the training set is split into locally owned\ntraining data subsets.\n VFL is increasingly used in applications like financial fraud detection;\nnonetheless, very little work has analyzed its security. In this paper, we\nfocus on robustness in VFL, in particular, on backdoor attacks, whereby an\nadversary attempts to manipulate the aggregate model during the training\nprocess to trigger misclassifications. Performing backdoor attacks in VFL is\nmore challenging than in HFL, as the adversary i) does not have access to the\nlabels during training and ii) cannot change the labels as she only has access\nto the feature embeddings. We present a first-of-its-kind clean-label backdoor\nattack in VFL, which consists of two phases: a label inference and a backdoor\nphase. We demonstrate the effectiveness of the attack on three different\ndatasets, investigate the factors involved in its success, and discuss\ncountermeasures to mitigate its impact.\n", "date": "2023-04-18", "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": "2210.07714", "title": "CrowdGuard: Federated Backdoor Detection in Federated Learning", "abstract": " Federated Learning (FL) is a promising approach enabling multiple clients to\ntrain Deep Neural Networks (DNNs) collaboratively without sharing their local\ntraining data. However, FL is susceptible to backdoor (or targeted poisoning)\nattacks. These attacks are initiated by malicious clients who seek to\ncompromise the learning process by introducing specific behaviors into the\nlearned model that can be triggered by carefully crafted inputs. Existing FL\nsafeguards have various limitations: They are restricted to specific data\ndistributions or reduce the global model accuracy due to excluding benign\nmodels or adding noise, are vulnerable to adaptive defense-aware adversaries,\nor require the server to access local models, allowing data inference attacks.\n This paper presents a novel defense mechanism, CrowdGuard, that effectively\nmitigates backdoor attacks in FL and overcomes the deficiencies of existing\ntechniques. It leverages clients' feedback on individual models, analyzes the\nbehavior of neurons in hidden layers, and eliminates poisoned models through an\niterative pruning scheme. CrowdGuard employs a server-located stacked\nclustering scheme to enhance its resilience to rogue client feedback. The\nevaluation results demonstrate that CrowdGuard achieves a 100%\nTrue-Positive-Rate and True-Negative-Rate across various scenarios, including\nIID and non-IID data distributions. Additionally, CrowdGuard withstands\nadaptive adversaries while preserving the original performance of protected\nmodels. To ensure confidentiality, CrowdGuard uses a secure and\nprivacy-preserving architecture leveraging Trusted Execution Environments\n(TEEs) on both client and server sides.\n", "date": "2022-10-14", "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": "2308.09850", "title": "Backdoor Mitigation by Correcting the Distribution of Neural Activations", "abstract": " Backdoor (Trojan) attacks are an important type of adversarial exploit\nagainst deep neural networks (DNNs), wherein a test instance is (mis)classified\nto the attacker's target class whenever the attacker's backdoor trigger is\npresent. In this paper, we reveal and analyze an important property of backdoor\nattacks: a successful attack causes an alteration in the distribution of\ninternal layer activations for backdoor-trigger instances, compared to that for\nclean instances. Even more importantly, we find that instances with the\nbackdoor trigger will be correctly classified to their original source classes\nif this distribution alteration is corrected. Based on our observations, we\npropose an efficient and effective method that achieves post-training backdoor\nmitigation by correcting the distribution alteration using reverse-engineered\ntriggers. Notably, our method does not change any trainable parameters of the\nDNN, but achieves generally better mitigation performance than existing methods\nthat do require intensive DNN parameter tuning. It also efficiently detects\ntest instances with the trigger, which may help to catch adversarial entities\nin the act of exploiting the backdoor.\n", "date": "2023-08-18", "categories": []} {"id": "2210.07346", "title": "An Embarrassingly Simple Backdoor Attack on Self-supervised Learning", "abstract": " As a new paradigm in machine learning, self-supervised learning (SSL) is\ncapable of learning high-quality representations of complex data without\nrelying on labels. In addition to eliminating the need for labeled data,\nresearch has found that SSL improves the adversarial robustness over supervised\nlearning since lacking labels makes it more challenging for adversaries to\nmanipulate model predictions. However, the extent to which this robustness\nsuperiority generalizes to other types of attacks remains an open question.\n We explore this question in the context of backdoor attacks. Specifically, we\ndesign and evaluate CTRL, an embarrassingly simple yet highly effective\nself-supervised backdoor attack. By only polluting a tiny fraction of training\ndata (<= 1%) with indistinguishable poisoning samples, CTRL causes any\ntrigger-embedded input to be misclassified to the adversary's designated class\nwith a high probability (>= 99%) at inference time. Our findings suggest that\nSSL and supervised learning are comparably vulnerable to backdoor attacks. More\nimportantly, through the lens of CTRL, we study the inherent vulnerability of\nSSL to backdoor attacks. With both empirical and analytical evidence, we reveal\nthat the representation invariance property of SSL, which benefits adversarial\nrobustness, may also be the very reason making \\ssl highly susceptible to\nbackdoor attacks. Our findings also imply that the existing defenses against\nsupervised backdoor attacks are not easily retrofitted to the unique\nvulnerability of SSL.\n", "date": "2022-10-13", "categories": []} {"id": "2306.15875", "title": "Fake the Real: Backdoor Attack on Deep Speech Classification via Voice\n Conversion", "abstract": " Deep speech classification has achieved tremendous success and greatly\npromoted the emergence of many real-world applications. However, backdoor\nattacks present a new security threat to it, particularly with untrustworthy\nthird-party platforms, as pre-defined triggers set by the attacker can activate\nthe backdoor. Most of the triggers in existing speech backdoor attacks are\nsample-agnostic, and even if the triggers are designed to be unnoticeable, they\ncan still be audible. This work explores a backdoor attack that utilizes\nsample-specific triggers based on voice conversion. Specifically, we adopt a\npre-trained voice conversion model to generate the trigger, ensuring that the\npoisoned samples does not introduce any additional audible noise. Extensive\nexperiments on two speech classification tasks demonstrate the effectiveness of\nour attack. Furthermore, we analyzed the specific scenarios that activated the\nproposed backdoor and verified its resistance against fine-tuning.\n", "date": "2023-06-28", "categories": []} {"id": "2308.04617", "title": "Improved Activation Clipping for Universal Backdoor Mitigation and\n Test-Time Detection", "abstract": " Deep neural networks are vulnerable to backdoor attacks (Trojans), where an\nattacker poisons the training set with backdoor triggers so that the neural\nnetwork learns to classify test-time triggers to the attacker's designated\ntarget class. Recent work shows that backdoor poisoning induces over-fitting\n(abnormally large activations) in the attacked model, which motivates a\ngeneral, post-training clipping method for backdoor mitigation, i.e., with\nbounds on internal-layer activations learned using a small set of clean\nsamples. We devise a new such approach, choosing the activation bounds to\nexplicitly limit classification margins. This method gives superior performance\nagainst peer methods for CIFAR-10 image classification. We also show that this\nmethod has strong robustness against adaptive attacks, X2X attacks, and on\ndifferent datasets. Finally, we demonstrate a method extension for test-time\ndetection and correction based on the output differences between the original\nand activation-bounded networks. The code of our method is online available.\n", "date": "2023-08-08", "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": "2308.04406", "title": "XGBD: Explanation-Guided Graph Backdoor Detection", "abstract": " Backdoor attacks pose a significant security risk to graph learning models.\nBackdoors can be embedded into the target model by inserting backdoor triggers\ninto the training dataset, causing the model to make incorrect predictions when\nthe trigger is present. To counter backdoor attacks, backdoor detection has\nbeen proposed. An emerging detection strategy in the vision and NLP domains is\nbased on an intriguing phenomenon: when training models on a mixture of\nbackdoor and clean samples, the loss on backdoor samples drops significantly\nfaster than on clean samples, allowing backdoor samples to be easily detected\nby selecting samples with the lowest loss values. However, the ignorance of\ntopological feature information on graph data limits its detection\neffectiveness when applied directly to the graph domain. To this end, we\npropose an explanation-guided backdoor detection method to take advantage of\nthe topological information. Specifically, we train a helper model on the graph\ndataset, feed graph samples into the model, and then adopt explanation methods\nto attribute model prediction to an important subgraph. We observe that\nbackdoor samples have distinct attribution distribution than clean samples, so\nthe explanatory subgraph could serve as more discriminative features for\ndetecting backdoor samples. Comprehensive experiments on multiple popular\ndatasets and attack methods demonstrate the effectiveness and explainability of\nour method. Our code is available:\nhttps://github.com/GuanZihan/GNN_backdoor_detection.\n", "date": "2023-08-08", "categories": []} {"id": "2302.11408", "title": "ASSET: Robust Backdoor Data Detection Across a Multiplicity of Deep\n Learning Paradigms", "abstract": " Backdoor data detection is traditionally studied in an end-to-end supervised\nlearning (SL) setting. However, recent years have seen the proliferating\nadoption of self-supervised learning (SSL) and transfer learning (TL), due to\ntheir lesser need for labeled data. Successful backdoor attacks have also been\ndemonstrated in these new settings. However, we lack a thorough understanding\nof the applicability of existing detection methods across a variety of learning\nsettings. By evaluating 56 attack settings, we show that the performance of\nmost existing detection methods varies significantly across different attacks\nand poison ratios, and all fail on the state-of-the-art clean-label attack. In\naddition, they either become inapplicable or suffer large performance losses\nwhen applied to SSL and TL. We propose a new detection method called Active\nSeparation via Offset (ASSET), which actively induces different model behaviors\nbetween the backdoor and clean samples to promote their separation. We also\nprovide procedures to adaptively select the number of suspicious points to\nremove. In the end-to-end SL setting, ASSET is superior to existing methods in\nterms of consistency of defensive performance across different attacks and\nrobustness to changes in poison ratios; in particular, it is the only method\nthat can detect the state-of-the-art clean-label attack. Moreover, ASSET's\naverage detection rates are higher than the best existing methods in SSL and\nTL, respectively, by 69.3% and 33.2%, thus providing the first practical\nbackdoor defense for these new DL settings. We open-source the project to drive\nfurther development and encourage engagement:\nhttps://github.com/ruoxi-jia-group/ASSET.\n", "date": "2023-02-22", "categories": []} {"id": "2305.02394", "title": "Defending against Insertion-based Textual Backdoor Attacks via\n Attribution", "abstract": " Textual backdoor attack, as a novel attack model, has been shown to be\neffective in adding a backdoor to the model during training. Defending against\nsuch backdoor attacks has become urgent and important. In this paper, we\npropose AttDef, an efficient attribution-based pipeline to defend against two\ninsertion-based poisoning attacks, BadNL and InSent. Specifically, we regard\nthe tokens with larger attribution scores as potential triggers since larger\nattribution words contribute more to the false prediction results and therefore\nare more likely to be poison triggers. Additionally, we further utilize an\nexternal pre-trained language model to distinguish whether input is poisoned or\nnot. We show that our proposed method can generalize sufficiently well in two\ncommon attack scenarios (poisoning training data and testing data), which\nconsistently improves previous methods. For instance, AttDef can successfully\nmitigate both attacks with an average accuracy of 79.97% (56.59% up) and 48.34%\n(3.99% up) under pre-training and post-training attack defense respectively,\nachieving the new state-of-the-art performance on prediction recovery over four\nbenchmark datasets.\n", "date": "2023-05-03", "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": "2307.10562", "title": "Shared Adversarial Unlearning: Backdoor Mitigation by Unlearning Shared\n Adversarial Examples", "abstract": " Backdoor attacks are serious security threats to machine learning models\nwhere an adversary can inject poisoned samples into the training set, causing a\nbackdoored model which predicts poisoned samples with particular triggers to\nparticular target classes, while behaving normally on benign samples. In this\npaper, we explore the task of purifying a backdoored model using a small clean\ndataset. By establishing the connection between backdoor risk and adversarial\nrisk, we derive a novel upper bound for backdoor risk, which mainly captures\nthe risk on the shared adversarial examples (SAEs) between the backdoored model\nand the purified model. This upper bound further suggests a novel bi-level\noptimization problem for mitigating backdoor using adversarial training\ntechniques. To solve it, we propose Shared Adversarial Unlearning (SAU).\nSpecifically, SAU first generates SAEs, and then, unlearns the generated SAEs\nsuch that they are either correctly classified by the purified model and/or\ndifferently classified by the two models, such that the backdoor effect in the\nbackdoored model will be mitigated in the purified model. Experiments on\nvarious benchmark datasets and network architectures show that our proposed\nmethod achieves state-of-the-art performance for backdoor defense.\n", "date": "2023-07-20", "categories": []} {"id": "2307.05422", "title": "Differential Analysis of Triggers and Benign Features for Black-Box DNN\n Backdoor Detection", "abstract": " This paper proposes a data-efficient detection method for deep neural\nnetworks against backdoor attacks under a black-box scenario. The proposed\napproach is motivated by the intuition that features corresponding to triggers\nhave a higher influence in determining the backdoored network output than any\nother benign features. To quantitatively measure the effects of triggers and\nbenign features on determining the backdoored network output, we introduce five\nmetrics. To calculate the five-metric values for a given input, we first\ngenerate several synthetic samples by injecting the input's partial contents\ninto clean validation samples. Then, the five metrics are computed by using the\noutput labels of the corresponding synthetic samples. One contribution of this\nwork is the use of a tiny clean validation dataset. Having the computed five\nmetrics, five novelty detectors are trained from the validation dataset. A meta\nnovelty detector fuses the output of the five trained novelty detectors to\ngenerate a meta confidence score. During online testing, our method determines\nif online samples are poisoned or not via assessing their meta confidence\nscores output by the meta novelty detector. We show the efficacy of our\nmethodology through a broad range of backdoor attacks, including ablation\nstudies and comparison to existing approaches. Our methodology is promising\nsince the proposed five metrics quantify the inherent differences between clean\nand poisoned samples. Additionally, our detection method can be incrementally\nimproved by appending more metrics that may be proposed to address future\nadvanced attacks.\n", "date": "2023-07-11", "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": "2307.08208", "title": "Towards Stealthy Backdoor Attacks against Speech Recognition via\n Elements of Sound", "abstract": " Deep neural networks (DNNs) have been widely and successfully adopted and\ndeployed in various applications of speech recognition. Recently, a few works\nrevealed that these models are vulnerable to backdoor attacks, where the\nadversaries can implant malicious prediction behaviors into victim models by\npoisoning their training process. In this paper, we revisit poison-only\nbackdoor attacks against speech recognition. We reveal that existing methods\nare not stealthy since their trigger patterns are perceptible to humans or\nmachine detection. This limitation is mostly because their trigger patterns are\nsimple noises or separable and distinctive clips. Motivated by these findings,\nwe propose to exploit elements of sound ($e.g.$, pitch and timbre) to design\nmore stealthy yet effective poison-only backdoor attacks. Specifically, we\ninsert a short-duration high-pitched signal as the trigger and increase the\npitch of remaining audio clips to `mask' it for designing stealthy pitch-based\ntriggers. We manipulate timbre features of victim audios to design the stealthy\ntimbre-based attack and design a voiceprint selection module to facilitate the\nmulti-backdoor attack. Our attacks can generate more `natural' poisoned samples\nand therefore are more stealthy. Extensive experiments are conducted on\nbenchmark datasets, which verify the effectiveness of our attacks under\ndifferent settings ($e.g.$, all-to-one, all-to-all, clean-label, physical, and\nmulti-backdoor settings) and their stealthiness. The code for reproducing main\nexperiments are available at \\url{https://github.com/HanboCai/BadSpeech_SoE}.\n", "date": "2023-07-17", "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": "2211.12044", "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": "2306.17441", "title": "Efficient Backdoor Removal Through Natural Gradient Fine-tuning", "abstract": " The success of a deep neural network (DNN) heavily relies on the details of\nthe training scheme; e.g., training data, architectures, hyper-parameters, etc.\nRecent backdoor attacks suggest that an adversary can take advantage of such\ntraining details and compromise the integrity of a DNN. Our studies show that a\nbackdoor model is usually optimized to a bad local minima, i.e. sharper minima\nas compared to a benign model. Intuitively, a backdoor model can be purified by\nreoptimizing the model to a smoother minima through fine-tuning with a few\nclean validation data. However, fine-tuning all DNN parameters often requires\nhuge computational costs and often results in sub-par clean test performance.\nTo address this concern, we propose a novel backdoor purification technique,\nNatural Gradient Fine-tuning (NGF), which focuses on removing the backdoor by\nfine-tuning only one layer. Specifically, NGF utilizes a loss surface\ngeometry-aware optimizer that can successfully overcome the challenge of\nreaching a smooth minima under a one-layer optimization scenario. To enhance\nthe generalization performance of our proposed method, we introduce a clean\ndata distribution-aware regularizer based on the knowledge of loss surface\ncurvature matrix, i.e., Fisher Information Matrix. Extensive experiments show\nthat the proposed method achieves state-of-the-art performance on a wide range\nof backdoor defense benchmarks: four different datasets- CIFAR10, GTSRB,\nTiny-ImageNet, and ImageNet; 13 recent backdoor attacks, e.g. Blend, Dynamic,\nWaNet, ISSBA, etc.\n", "date": "2023-06-30", "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": "2306.16697", "title": "Neural Polarizer: A Lightweight and Effective Backdoor Defense via\n Purifying Poisoned Features", "abstract": " Recent studies have demonstrated the susceptibility of deep neural networks\nto backdoor attacks. Given a backdoored model, its prediction of a poisoned\nsample with trigger will be dominated by the trigger information, though\ntrigger information and benign information coexist. Inspired by the mechanism\nof the optical polarizer that a polarizer could pass light waves with\nparticular polarizations while filtering light waves with other polarizations,\nwe propose a novel backdoor defense method by inserting a learnable neural\npolarizer into the backdoored model as an intermediate layer, in order to\npurify the poisoned sample via filtering trigger information while maintaining\nbenign information. The neural polarizer is instantiated as one lightweight\nlinear transformation layer, which is learned through solving a well designed\nbi-level optimization problem, based on a limited clean dataset. Compared to\nother fine-tuning-based defense methods which often adjust all parameters of\nthe backdoored model, the proposed method only needs to learn one additional\nlayer, such that it is more efficient and requires less clean data. Extensive\nexperiments demonstrate the effectiveness and efficiency of our method in\nremoving backdoors across various neural network architectures and datasets,\nespecially in the case of very limited clean data.\n", "date": "2023-06-29", "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": "2306.08011", "title": "Privacy Inference-Empowered Stealthy Backdoor Attack on Federated\n Learning under Non-IID Scenarios", "abstract": " Federated learning (FL) naturally faces the problem of data heterogeneity in\nreal-world scenarios, but this is often overlooked by studies on FL security\nand privacy. On the one hand, the effectiveness of backdoor attacks on FL may\ndrop significantly under non-IID scenarios. On the other hand, malicious\nclients may steal private data through privacy inference attacks. Therefore, it\nis necessary to have a comprehensive perspective of data heterogeneity,\nbackdoor, and privacy inference. In this paper, we propose a novel privacy\ninference-empowered stealthy backdoor attack (PI-SBA) scheme for FL under\nnon-IID scenarios. Firstly, a diverse data reconstruction mechanism based on\ngenerative adversarial networks (GANs) is proposed to produce a supplementary\ndataset, which can improve the attacker's local data distribution and support\nmore sophisticated strategies for backdoor attacks. Based on this, we design a\nsource-specified backdoor learning (SSBL) strategy as a demonstration, allowing\nthe adversary to arbitrarily specify which classes are susceptible to the\nbackdoor trigger. Since the PI-SBA has an independent poisoned data synthesis\nprocess, it can be integrated into existing backdoor attacks to improve their\neffectiveness and stealthiness in non-IID scenarios. Extensive experiments\nbased on MNIST, CIFAR10 and Youtube Aligned Face datasets demonstrate that the\nproposed PI-SBA scheme is effective in non-IID FL and stealthy against\nstate-of-the-art defense methods.\n", "date": "2023-06-13", "categories": []} {"id": "2306.08350", "title": "Multi-target Backdoor Attacks for Code Pre-trained Models", "abstract": " Backdoor attacks for neural code models have gained considerable attention\ndue to the advancement of code intelligence. However, most existing works\ninsert triggers into task-specific data for code-related downstream tasks,\nthereby limiting the scope of attacks. Moreover, the majority of attacks for\npre-trained models are designed for understanding tasks. In this paper, we\npropose task-agnostic backdoor attacks for code pre-trained models. Our\nbackdoored model is pre-trained with two learning strategies (i.e., Poisoned\nSeq2Seq learning and token representation learning) to support the multi-target\nattack of downstream code understanding and generation tasks. During the\ndeployment phase, the implanted backdoors in the victim models can be activated\nby the designed triggers to achieve the targeted attack. We evaluate our\napproach on two code understanding tasks and three code generation tasks over\nseven datasets. Extensive experiments demonstrate that our approach can\neffectively and stealthily attack code-related downstream tasks.\n", "date": "2023-06-14", "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": "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": "2306.01983", "title": "Mitigating Backdoor Attack Via Prerequisite Transformation", "abstract": " In recent years, with the successful application of DNN in fields such as NLP\nand CV, its security has also received widespread attention. (Author) proposed\nthe method of backdoor attack in Badnet. Switch implanted backdoor into the\nmodel by poisoning the training samples. The model with backdoor did not\nexhibit any abnormalities on the normal validation sample set, but in the input\nwith trigger, they were mistakenly classified as the attacker's designated\ncategory or randomly classified as a different category from the ground truth,\nThis attack method seriously threatens the normal application of DNN in real\nlife, such as autonomous driving, object detection, etc.This article proposes a\nnew method to combat backdoor attacks. We refer to the features in the area\ncovered by the trigger as trigger features, and the remaining areas as normal\nfeatures. By introducing prerequisite calculation conditions during the\ntraining process, these conditions have little impact on normal features and\ntrigger features, and can complete the training of a standard backdoor model.\nThe model trained under these prerequisite calculation conditions can, In the\nverification set D'val with the same premise calculation conditions, the\nperformance is consistent with that of the ordinary backdoor model. However, in\nthe verification set Dval without the premise calculation conditions, the\nverification accuracy decreases very little (7%~12%), while the attack success\nrate (ASR) decreases from 90% to about 8%.Author call this method Prerequisite\nTransformation(PT).\n", "date": "2023-06-03", "categories": []} {"id": "2306.02368", "title": "Revisiting Data-Free Knowledge Distillation with Poisoned Teachers", "abstract": " Data-free knowledge distillation (KD) helps transfer knowledge from a\npre-trained model (known as the teacher model) to a smaller model (known as the\nstudent model) without access to the original training data used for training\nthe teacher model. However, the security of the synthetic or\nout-of-distribution (OOD) data required in data-free KD is largely unknown and\nunder-explored. In this work, we make the first effort to uncover the security\nrisk of data-free KD w.r.t. untrusted pre-trained models. We then propose\nAnti-Backdoor Data-Free KD (ABD), the first plug-in defensive method for\ndata-free KD methods to mitigate the chance of potential backdoors being\ntransferred. We empirically evaluate the effectiveness of our proposed ABD in\ndiminishing transferred backdoor knowledge while maintaining compatible\ndownstream performances as the vanilla KD. We envision this work as a milestone\nfor alarming and mitigating the potential backdoors in data-free KD. Codes are\nreleased at https://github.com/illidanlab/ABD.\n", "date": "2023-06-04", "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": "2302.09578", "title": "On Feasibility of Server-side Backdoor Attacks on Split Learning", "abstract": " Split learning is a collaborative learning design that allows several\nparticipants (clients) to train a shared model while keeping their datasets\nprivate. Recent studies demonstrate that collaborative learning models,\nspecifically federated learning, are vulnerable to security and privacy attacks\nsuch as model inference and backdoor attacks. Backdoor attacks are a group of\npoisoning attacks in which the attacker tries to control the model output by\nmanipulating the model's training process. While there have been studies\nregarding inference attacks on split learning, it has not yet been tested for\nbackdoor attacks. This paper performs a novel backdoor attack on split learning\nand studies its effectiveness. Despite traditional backdoor attacks done on the\nclient side, we inject the backdoor trigger from the server side. For this\npurpose, we provide two attack methods: one using a surrogate client and\nanother using an autoencoder to poison the model via incoming smashed data and\nits outgoing gradient toward the innocent participants. We did our experiments\nusing three model architectures and three publicly available datasets in the\nimage domain and ran a total of 761 experiments to evaluate our attack methods.\nThe results show that despite using strong patterns and injection methods,\nsplit learning is highly robust and resistant to such poisoning attacks. While\nwe get the attack success rate of 100% as our best result for the MNIST\ndataset, in most of the other cases, our attack shows little success when\nincreasing the cut layer.\n", "date": "2023-02-19", "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": "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": "2305.04547", "title": "Diffusion Theory as a Scalpel: Detecting and Purifying Poisonous\n Dimensions in Pre-trained Language Models Caused by Backdoor or Bias", "abstract": " Pre-trained Language Models (PLMs) may be poisonous with backdoors or bias\ninjected by the suspicious attacker during the fine-tuning process. A core\nchallenge of purifying potentially poisonous PLMs is precisely finding\npoisonous dimensions. To settle this issue, we propose the Fine-purifying\napproach, which utilizes the diffusion theory to study the dynamic process of\nfine-tuning for finding potentially poisonous dimensions. According to the\nrelationship between parameter drifts and Hessians of different dimensions, we\ncan detect poisonous dimensions with abnormal dynamics, purify them by\nresetting them to clean pre-trained weights, and then fine-tune the purified\nweights on a small clean dataset. To the best of our knowledge, we are the\nfirst to study the dynamics guided by the diffusion theory for safety or\ndefense purposes. Experimental results validate the effectiveness of\nFine-purifying even with a small clean dataset.\n", "date": "2023-05-08", "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": "2305.02424", "title": "Backdoor Learning on Sequence to Sequence Models", "abstract": " Backdoor learning has become an emerging research area towards building a\ntrustworthy machine learning system. While a lot of works have studied the\nhidden danger of backdoor attacks in image or text classification, there is a\nlimited understanding of the model's robustness on backdoor attacks when the\noutput space is infinite and discrete. In this paper, we study a much more\nchallenging problem of testing whether sequence-to-sequence (seq2seq) models\nare vulnerable to backdoor attacks. Specifically, we find by only injecting\n0.2\\% samples of the dataset, we can cause the seq2seq model to generate the\ndesignated keyword and even the whole sentence. Furthermore, we utilize Byte\nPair Encoding (BPE) to create multiple new triggers, which brings new\nchallenges to backdoor detection since these backdoors are not static.\nExtensive experiments on machine translation and text summarization have been\nconducted to show our proposed methods could achieve over 90\\% attack success\nrate on multiple datasets and models.\n", "date": "2023-05-03", "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": "2304.14475", "title": "ChatGPT as an Attack Tool: Stealthy Textual Backdoor Attack via Blackbox\n Generative Model Trigger", "abstract": " Textual backdoor attacks pose a practical threat to existing systems, as they\ncan compromise the model by inserting imperceptible triggers into inputs and\nmanipulating labels in the training dataset. With cutting-edge generative\nmodels such as GPT-4 pushing rewriting to extraordinary levels, such attacks\nare becoming even harder to detect. We conduct a comprehensive investigation of\nthe role of black-box generative models as a backdoor attack tool, highlighting\nthe importance of researching relative defense strategies. In this paper, we\nreveal that the proposed generative model-based attack, BGMAttack, could\neffectively deceive textual classifiers. Compared with the traditional attack\nmethods, BGMAttack makes the backdoor trigger less conspicuous by leveraging\nstate-of-the-art generative models. Our extensive evaluation of attack\neffectiveness across five datasets, complemented by three distinct human\ncognition assessments, reveals that Figure 4 achieves comparable attack\nperformance while maintaining superior stealthiness relative to baseline\nmethods.\n", "date": "2023-04-27", "categories": []} {"id": "2304.11432", "title": "Universal Adversarial Backdoor Attacks to Fool Vertical Federated\n Learning in Cloud-Edge Collaboration", "abstract": " Vertical federated learning (VFL) is a cloud-edge collaboration paradigm that\nenables edge nodes, comprising resource-constrained Internet of Things (IoT)\ndevices, to cooperatively train artificial intelligence (AI) models while\nretaining their data locally. This paradigm facilitates improved privacy and\nsecurity for edges and IoT devices, making VFL an essential component of\nArtificial Intelligence of Things (AIoT) systems. Nevertheless, the partitioned\nstructure of VFL can be exploited by adversaries to inject a backdoor, enabling\nthem to manipulate the VFL predictions. In this paper, we aim to investigate\nthe vulnerability of VFL in the context of binary classification tasks. To this\nend, we define a threat model for backdoor attacks in VFL and introduce a\nuniversal adversarial backdoor (UAB) attack to poison the predictions of VFL.\nThe UAB attack, consisting of universal trigger generation and clean-label\nbackdoor injection, is incorporated during the VFL training at specific\niterations. This is achieved by alternately optimizing the universal trigger\nand model parameters of VFL sub-problems. Our work distinguishes itself from\nexisting studies on designing backdoor attacks for VFL, as those require the\nknowledge of auxiliary information not accessible within the split VFL\narchitecture. In contrast, our approach does not necessitate any additional\ndata to execute the attack. On the LendingClub and Zhongyuan datasets, our\napproach surpasses existing state-of-the-art methods, achieving up to 100\\%\nbackdoor task performance while maintaining the main task performance. Our\nresults in this paper make a major advance to revealing the hidden backdoor\nrisks of VFL, hence paving the way for the future development of secure AIoT.\n", "date": "2023-04-22", "categories": []} {"id": "2304.12298", "title": "BadGPT: Exploring Security Vulnerabilities of ChatGPT via Backdoor\n Attacks to InstructGPT", "abstract": " Recently, ChatGPT has gained significant attention in research due to its\nability to interact with humans effectively. The core idea behind this model is\nreinforcement learning (RL) fine-tuning, a new paradigm that allows language\nmodels to align with human preferences, i.e., InstructGPT. In this study, we\npropose BadGPT, the first backdoor attack against RL fine-tuning in language\nmodels. By injecting a backdoor into the reward model, the language model can\nbe compromised during the fine-tuning stage. Our initial experiments on movie\nreviews, i.e., IMDB, demonstrate that an attacker can manipulate the generated\ntext through BadGPT.\n", "date": "2023-02-21", "categories": []} {"id": "2302.01740", "title": "SoK: A Systematic Evaluation of Backdoor Trigger Characteristics in\n Image Classification", "abstract": " Deep learning achieves outstanding results in many machine learning tasks.\nNevertheless, it is vulnerable to backdoor attacks that modify the training set\nto embed a secret functionality in the trained model. The modified training\nsamples have a secret property, i. e., a trigger. At inference time, the secret\nfunctionality is activated when the input contains the trigger, while the model\nfunctions correctly in other cases. While there are many known backdoor attacks\n(and defenses), deploying a stealthy attack is still far from trivial.\nSuccessfully creating backdoor triggers depends on numerous parameters.\nUnfortunately, research has not yet determined which parameters contribute most\nto the attack performance.\n This paper systematically analyzes the most relevant parameters for the\nbackdoor attacks, i.e., trigger size, position, color, and poisoning rate.\nUsing transfer learning, which is very common in computer vision, we evaluate\nthe attack on state-of-the-art models (ResNet, VGG, AlexNet, and GoogLeNet) and\ndatasets (MNIST, CIFAR10, and TinyImageNet). Our attacks cover the majority of\nbackdoor settings in research, providing concrete directions for future works.\nOur code is publicly available to facilitate the reproducibility of our\nresults.\n", "date": "2023-02-03", "categories": []} {"id": "2304.02277", "title": "Rethinking the Trigger-injecting Position in Graph Backdoor Attack", "abstract": " Backdoor attacks have been demonstrated as a security threat for machine\nlearning models. Traditional backdoor attacks intend to inject backdoor\nfunctionality into the model such that the backdoored model will perform\nabnormally on inputs with predefined backdoor triggers and still retain\nstate-of-the-art performance on the clean inputs. While there are already some\nworks on backdoor attacks on Graph Neural Networks (GNNs), the backdoor trigger\nin the graph domain is mostly injected into random positions of the sample.\nThere is no work analyzing and explaining the backdoor attack performance when\ninjecting triggers into the most important or least important area in the\nsample, which we refer to as trigger-injecting strategies MIAS and LIAS,\nrespectively. Our results show that, generally, LIAS performs better, and the\ndifferences between the LIAS and MIAS performance can be significant.\nFurthermore, we explain these two strategies' similar (better) attack\nperformance through explanation techniques, which results in a further\nunderstanding of backdoor attacks in GNNs.\n", "date": "2023-04-05", "categories": []} {"id": "2303.11470", "title": "Did You Train on My Dataset? Towards Public Dataset Protection with\n Clean-Label Backdoor Watermarking", "abstract": " The huge supporting training data on the Internet has been a key factor in\nthe success of deep learning models. However, this abundance of\npublic-available data also raises concerns about the unauthorized exploitation\nof datasets for commercial purposes, which is forbidden by dataset licenses. In\nthis paper, we propose a backdoor-based watermarking approach that serves as a\ngeneral framework for safeguarding public-available data. By inserting a small\nnumber of watermarking samples into the dataset, our approach enables the\nlearning model to implicitly learn a secret function set by defenders. This\nhidden function can then be used as a watermark to track down third-party\nmodels that use the dataset illegally. Unfortunately, existing backdoor\ninsertion methods often entail adding arbitrary and mislabeled data to the\ntraining set, leading to a significant drop in performance and easy detection\nby anomaly detection algorithms. To overcome this challenge, we introduce a\nclean-label backdoor watermarking framework that uses imperceptible\nperturbations to replace mislabeled samples. As a result, the watermarking\nsamples remain consistent with the original labels, making them difficult to\ndetect. Our experiments on text, image, and audio datasets demonstrate that the\nproposed framework effectively safeguards datasets with minimal impact on\noriginal task performance. We also show that adding just 1% of watermarking\nsamples can inject a traceable watermarking function and that our watermarking\nsamples are stealthy and look benign upon visual inspection.\n", "date": "2023-03-20", "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": "2210.00875", "title": "Untargeted Backdoor Watermark: Towards Harmless and Stealthy Dataset\n Copyright Protection", "abstract": " Deep neural networks (DNNs) have demonstrated their superiority in practice.\nArguably, the rapid development of DNNs is largely benefited from high-quality\n(open-sourced) datasets, based on which researchers and developers can easily\nevaluate and improve their learning methods. Since the data collection is\nusually time-consuming or even expensive, how to protect their copyrights is of\ngreat significance and worth further exploration. In this paper, we revisit\ndataset ownership verification. We find that existing verification methods\nintroduced new security risks in DNNs trained on the protected dataset, due to\nthe targeted nature of poison-only backdoor watermarks. To alleviate this\nproblem, in this work, we explore the untargeted backdoor watermarking scheme,\nwhere the abnormal model behaviors are not deterministic. Specifically, we\nintroduce two dispersibilities and prove their correlation, based on which we\ndesign the untargeted backdoor watermark under both poisoned-label and\nclean-label settings. We also discuss how to use the proposed untargeted\nbackdoor watermark for dataset ownership verification. Experiments on benchmark\ndatasets verify the effectiveness of our methods and their resistance to\nexisting backdoor defenses. Our codes are available at\n\\url{https://github.com/THUYimingLi/Untargeted_Backdoor_Watermark}.\n", "date": "2022-09-27", "categories": []} {"id": "2304.01482", "title": "Defending Against Patch-based Backdoor Attacks on Self-Supervised\n Learning", "abstract": " Recently, self-supervised learning (SSL) was shown to be vulnerable to\npatch-based data poisoning backdoor attacks. It was shown that an adversary can\npoison a small part of the unlabeled data so that when a victim trains an SSL\nmodel on it, the final model will have a backdoor that the adversary can\nexploit. This work aims to defend self-supervised learning against such\nattacks. We use a three-step defense pipeline, where we first train a model on\nthe poisoned data. In the second step, our proposed defense algorithm\n(PatchSearch) uses the trained model to search the training data for poisoned\nsamples and removes them from the training set. In the third step, a final\nmodel is trained on the cleaned-up training set. Our results show that\nPatchSearch is an effective defense. As an example, it improves a model's\naccuracy on images containing the trigger from 38.2% to 63.7% which is very\nclose to the clean model's accuracy, 64.6%. Moreover, we show that PatchSearch\noutperforms baselines and state-of-the-art defense approaches including those\nusing additional clean, trusted data. Our code is available at\nhttps://github.com/UCDvision/PatchSearch\n", "date": "2023-04-04", "categories": []} {"id": "2303.18191", "title": "Detecting Backdoors During the Inference Stage Based on Corruption\n Robustness Consistency", "abstract": " Deep neural networks are proven to be vulnerable to backdoor attacks.\nDetecting the trigger samples during the inference stage, i.e., the test-time\ntrigger sample detection, can prevent the backdoor from being triggered.\nHowever, existing detection methods often require the defenders to have high\naccessibility to victim models, extra clean data, or knowledge about the\nappearance of backdoor triggers, limiting their practicality. In this paper, we\npropose the test-time corruption robustness consistency evaluation (TeCo), a\nnovel test-time trigger sample detection method that only needs the hard-label\noutputs of the victim models without any extra information. Our journey begins\nwith the intriguing observation that the backdoor-infected models have similar\nperformance across different image corruptions for the clean images, but\nperform discrepantly for the trigger samples. Based on this phenomenon, we\ndesign TeCo to evaluate test-time robustness consistency by calculating the\ndeviation of severity that leads to predictions' transition across different\ncorruptions. Extensive experiments demonstrate that compared with\nstate-of-the-art defenses, which even require either certain information about\nthe trigger types or accessibility of clean data, TeCo outperforms them on\ndifferent backdoor attacks, datasets, and model architectures, enjoying a\nhigher AUROC by 10% and 5 times of stability.\n", "date": "2023-03-27", "categories": []} {"id": "2303.13992", "title": "Physical Backdoor Trigger Activation of Autonomous Vehicle using\n Reachability Analysis", "abstract": " Recent studies reveal that Autonomous Vehicles (AVs) can be manipulated by\nhidden backdoors, causing them to perform harmful actions when activated by\nphysical triggers. However, it is still unclear how these triggers can be\nactivated while adhering to traffic principles. Understanding this\nvulnerability in a dynamic traffic environment is crucial. This work addresses\nthis gap by presenting physical trigger activation as a reachability problem of\ncontrolled dynamic system. Our technique identifies security-critical areas in\ntraffic systems where trigger conditions for accidents can be reached, and\nprovides intended trajectories for how those conditions can be reached. Testing\non typical traffic scenarios showed the system can be successfully driven to\ntrigger conditions with near 100% activation rate. Our method benefits from\nidentifying AV vulnerability and enabling effective safety strategies.\n", "date": "2023-03-24", "categories": []} {"id": "2303.14325", "title": "Backdoor Attacks with Input-unique Triggers in NLP", "abstract": " Backdoor attack aims at inducing neural models to make incorrect predictions\nfor poison data while keeping predictions on the clean dataset unchanged, which\ncreates a considerable threat to current natural language processing (NLP)\nsystems. Existing backdoor attacking systems face two severe issues:firstly,\nmost backdoor triggers follow a uniform and usually input-independent pattern,\ne.g., insertion of specific trigger words, synonym replacement. This\nsignificantly hinders the stealthiness of the attacking model, leading the\ntrained backdoor model being easily identified as malicious by model probes.\nSecondly, trigger-inserted poisoned sentences are usually disfluent,\nungrammatical, or even change the semantic meaning from the original sentence,\nmaking them being easily filtered in the pre-processing stage. To resolve these\ntwo issues, in this paper, we propose an input-unique backdoor attack(NURA),\nwhere we generate backdoor triggers unique to inputs. IDBA generates\ncontext-related triggers by continuing writing the input with a language model\nlike GPT2. The generated sentence is used as the backdoor trigger. This\nstrategy not only creates input-unique backdoor triggers, but also preserves\nthe semantics of the original input, simultaneously resolving the two issues\nabove. Experimental results show that the IDBA attack is effective for attack\nand difficult to defend: it achieves high attack success rate across all the\nwidely applied benchmarks, while is immune to existing defending methods. In\naddition, it is able to generate fluent, grammatical, and diverse backdoor\ninputs, which can hardly be recognized through human inspection.\n", "date": "2023-03-25", "categories": []} {"id": "2303.15180", "title": "Detecting Backdoors in Pre-trained Encoders", "abstract": " Self-supervised learning in computer vision trains on unlabeled data, such as\nimages or (image, text) pairs, to obtain an image encoder that learns\nhigh-quality embeddings for input data. Emerging backdoor attacks towards\nencoders expose crucial vulnerabilities of self-supervised learning, since\ndownstream classifiers (even further trained on clean data) may inherit\nbackdoor behaviors from encoders. Existing backdoor detection methods mainly\nfocus on supervised learning settings and cannot handle pre-trained encoders\nespecially when input labels are not available. In this paper, we propose\nDECREE, the first backdoor detection approach for pre-trained encoders,\nrequiring neither classifier headers nor input labels. We evaluate DECREE on\nover 400 encoders trojaned under 3 paradigms. We show the effectiveness of our\nmethod on image encoders pre-trained on ImageNet and OpenAI's CLIP 400 million\nimage-text pairs. Our method consistently has a high detection accuracy even if\nwe have only limited or no access to the pre-training dataset.\n", "date": "2023-03-23", "categories": []} {"id": "2206.07018", "title": "Turning a Curse into a Blessing: Enabling In-Distribution-Data-Free\n Backdoor Removal via Stabilized Model Inversion", "abstract": " Many backdoor removal techniques in machine learning models require clean\nin-distribution data, which may not always be available due to proprietary\ndatasets. Model inversion techniques, often considered privacy threats, can\nreconstruct realistic training samples, potentially eliminating the need for\nin-distribution data. Prior attempts to combine backdoor removal and model\ninversion yielded limited results. Our work is the first to provide a thorough\nunderstanding of leveraging model inversion for effective backdoor removal by\naddressing key questions about reconstructed samples' properties, perceptual\nsimilarity, and the potential presence of backdoor triggers.\n We establish that relying solely on perceptual similarity is insufficient for\nrobust defenses, and the stability of model predictions in response to input\nand parameter perturbations is also crucial. To tackle this, we introduce a\nnovel bi-level optimization-based framework for model inversion, promoting\nstability and visual quality. Interestingly, we discover that reconstructed\nsamples from a pre-trained generator's latent space are backdoor-free, even\nwhen utilizing signals from a backdoored model. We provide a theoretical\nanalysis to support this finding. Our evaluation demonstrates that our\nstabilized model inversion technique achieves state-of-the-art backdoor removal\nperformance without clean in-distribution data, matching or surpassing\nperformance using the same amount of clean samples.\n", "date": "2022-06-14", "categories": []} {"id": "2303.12993", "title": "Backdoor Defense via Adaptively Splitting Poisoned Dataset", "abstract": " Backdoor defenses have been studied to alleviate the threat of deep neural\nnetworks (DNNs) being backdoor attacked and thus maliciously altered. Since\nDNNs usually adopt some external training data from an untrusted third party, a\nrobust backdoor defense strategy during the training stage is of importance. We\nargue that the core of training-time defense is to select poisoned samples and\nto handle them properly. In this work, we summarize the training-time defenses\nfrom a unified framework as splitting the poisoned dataset into two data pools.\nUnder our framework, we propose an adaptively splitting dataset-based defense\n(ASD). Concretely, we apply loss-guided split and meta-learning-inspired split\nto dynamically update two data pools. With the split clean data pool and\npolluted data pool, ASD successfully defends against backdoor attacks during\ntraining. Extensive experiments on multiple benchmark datasets and DNN models\nagainst six state-of-the-art backdoor attacks demonstrate the superiority of\nour ASD. Our code is available at https://github.com/KuofengGao/ASD.\n", "date": "2023-03-23", "categories": []} {"id": "2303.13211", "title": "Don't FREAK Out: A Frequency-Inspired Approach to Detecting Backdoor\n Poisoned Samples in DNNs", "abstract": " In this paper we investigate the frequency sensitivity of Deep Neural\nNetworks (DNNs) when presented with clean samples versus poisoned samples. Our\nanalysis shows significant disparities in frequency sensitivity between these\ntwo types of samples. Building on these findings, we propose FREAK, a\nfrequency-based poisoned sample detection algorithm that is simple yet\neffective. Our experimental results demonstrate the efficacy of FREAK not only\nagainst frequency backdoor attacks but also against some spatial attacks. Our\nwork is just the first step in leveraging these insights. We believe that our\nanalysis and proposed defense mechanism will provide a foundation for future\nresearch and development of backdoor defenses.\n", "date": "2023-03-23", "categories": []} {"id": "2303.12589", "title": "Do Backdoors Assist Membership Inference Attacks?", "abstract": " When an adversary provides poison samples to a machine learning model,\nprivacy leakage, such as membership inference attacks that infer whether a\nsample was included in the training of the model, becomes effective by moving\nthe sample to an outlier. However, the attacks can be detected because\ninference accuracy deteriorates due to poison samples. In this paper, we\ndiscuss a \\textit{backdoor-assisted membership inference attack}, a novel\nmembership inference attack based on backdoors that return the adversary's\nexpected output for a triggered sample. We found three crucial insights through\nexperiments with an academic benchmark dataset. We first demonstrate that the\nbackdoor-assisted membership inference attack is unsuccessful. Second, when we\nanalyzed loss distributions to understand the reason for the unsuccessful\nresults, we found that backdoors cannot separate loss distributions of training\nand non-training samples. In other words, backdoors cannot affect the\ndistribution of clean samples. Third, we also show that poison and triggered\nsamples activate neurons of different distributions. Specifically, backdoors\nmake any clean sample an inlier, contrary to poisoning samples. As a result, we\nconfirm that backdoors cannot assist membership inference.\n", "date": "2023-03-22", "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": "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": "2303.05762", "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": "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": "2211.05631", "title": "Backdoor Defense via Suppressing Model Shortcuts", "abstract": " Recent studies have demonstrated that deep neural networks (DNNs) are\nvulnerable to backdoor attacks during the training process. Specifically, the\nadversaries intend to embed hidden backdoors in DNNs so that malicious model\npredictions can be activated through pre-defined trigger patterns. In this\npaper, we explore the backdoor mechanism from the angle of the model structure.\nWe select the skip connection for discussions, inspired by the understanding\nthat it helps the learning of model `shortcuts' where backdoor triggers are\nusually easier to be learned. Specifically, we demonstrate that the attack\nsuccess rate (ASR) decreases significantly when reducing the outputs of some\nkey skip connections. Based on this observation, we design a simple yet\neffective backdoor removal method by suppressing the skip connections in\ncritical layers selected by our method. We also implement fine-tuning on these\nlayers to recover high benign accuracy and to further reduce ASR. Extensive\nexperiments on benchmark datasets verify the effectiveness of our method.\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": "2303.02213", "title": "Backdoor Attacks and Defenses in Federated Learning: Survey, Challenges\n and Future Research Directions", "abstract": " Federated learning (FL) is a machine learning (ML) approach that allows the\nuse of distributed data without compromising personal privacy. However, the\nheterogeneous distribution of data among clients in FL can make it difficult\nfor the orchestration server to validate the integrity of local model updates,\nmaking FL vulnerable to various threats, including backdoor attacks. Backdoor\nattacks involve the insertion of malicious functionality into a targeted model\nthrough poisoned updates from malicious clients. These attacks can cause the\nglobal model to misbehave on specific inputs while appearing normal in other\ncases. Backdoor attacks have received significant attention in the literature\ndue to their potential to impact real-world deep learning applications.\nHowever, they have not been thoroughly studied in the context of FL. In this\nsurvey, we provide a comprehensive survey of current backdoor attack strategies\nand defenses in FL, including a comprehensive analysis of different approaches.\nWe also discuss the challenges and potential future directions for attacks and\ndefenses in the context of FL.\n", "date": "2023-03-03", "categories": []} {"id": "2303.01742", "title": "NCL: Textual Backdoor Defense Using Noise-augmented Contrastive Learning", "abstract": " At present, backdoor attacks attract attention as they do great harm to deep\nlearning models. The adversary poisons the training data making the model being\ninjected with a backdoor after being trained unconsciously by victims using the\npoisoned dataset. In the field of text, however, existing works do not provide\nsufficient defense against backdoor attacks. In this paper, we propose a\nNoise-augmented Contrastive Learning (NCL) framework to defend against textual\nbackdoor attacks when training models with untrustworthy data. With the aim of\nmitigating the mapping between triggers and the target label, we add\nappropriate noise perturbing possible backdoor triggers, augment the training\ndataset, and then pull homology samples in the feature space utilizing\ncontrastive learning objective. Experiments demonstrate the effectiveness of\nour method in defending three types of textual backdoor attacks, outperforming\nthe prior works.\n", "date": "2023-03-03", "categories": []} {"id": "2210.12873", "title": "FLIP: A Provable Defense Framework for Backdoor Mitigation in Federated\n Learning", "abstract": " Federated Learning (FL) is a distributed learning paradigm that enables\ndifferent parties to train a model together for high quality and strong privacy\nprotection. In this scenario, individual participants may get compromised and\nperform backdoor attacks by poisoning the data (or gradients). Existing work on\nrobust aggregation and certified FL robustness does not study how hardening\nbenign clients can affect the global model (and the malicious clients). In this\nwork, we theoretically analyze the connection among cross-entropy loss, attack\nsuccess rate, and clean accuracy in this setting. Moreover, we propose a\ntrigger reverse engineering based defense and show that our method can achieve\nrobustness improvement with guarantee (i.e., reducing the attack success rate)\nwithout affecting benign accuracy. We conduct comprehensive experiments across\ndifferent datasets and attack settings. Our results on eight competing SOTA\ndefense methods show the empirical superiority of our method on both\nsingle-shot and continuous FL backdoor attacks. Code is available at\nhttps://github.com/KaiyuanZh/FLIP.\n", "date": "2022-10-23", "categories": []} {"id": "2302.14500", "title": "FreeEagle: Detecting Complex Neural Trojans in Data-Free Cases", "abstract": " Trojan attack on deep neural networks, also known as backdoor attack, is a\ntypical threat to artificial intelligence. A trojaned neural network behaves\nnormally with clean inputs. However, if the input contains a particular\ntrigger, the trojaned model will have attacker-chosen abnormal behavior.\nAlthough many backdoor detection methods exist, most of them assume that the\ndefender has access to a set of clean validation samples or samples with the\ntrigger, which may not hold in some crucial real-world cases, e.g., the case\nwhere the defender is the maintainer of model-sharing platforms. Thus, in this\npaper, we propose FreeEagle, the first data-free backdoor detection method that\ncan effectively detect complex backdoor attacks on deep neural networks,\nwithout relying on the access to any clean samples or samples with the trigger.\nThe evaluation results on diverse datasets and model architectures show that\nFreeEagle is effective against various complex backdoor attacks, even\noutperforming some state-of-the-art non-data-free backdoor detection methods.\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": "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": "2302.08090", "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": "2210.07543", "title": "Watermarking Pre-trained Language Models with Backdooring", "abstract": " Large pre-trained language models (PLMs) have proven to be a crucial\ncomponent of modern natural language processing systems. PLMs typically need to\nbe fine-tuned on task-specific downstream datasets, which makes it hard to\nclaim the ownership of PLMs and protect the developer's intellectual property\ndue to the catastrophic forgetting phenomenon. We show that PLMs can be\nwatermarked with a multi-task learning framework by embedding backdoors\ntriggered by specific inputs defined by the owners, and those watermarks are\nhard to remove even though the watermarked PLMs are fine-tuned on multiple\ndownstream tasks. In addition to using some rare words as triggers, we also\nshow that the combination of common words can be used as backdoor triggers to\navoid them being easily detected. Extensive experiments on multiple datasets\ndemonstrate that the embedded watermarks can be robustly extracted with a high\nsuccess rate and less influenced by the follow-up fine-tuning.\n", "date": "2022-10-14", "categories": []} {"id": "2206.04677", "title": "On the Permanence of Backdoors in Evolving Models", "abstract": " Existing research on training-time attacks for deep neural networks (DNNs),\nsuch as backdoors, largely assume that models are static once trained, and\nhidden backdoors trained into models remain active indefinitely. In practice,\nmodels are rarely static but evolve continuously to address distribution drifts\nin the underlying data. This paper explores the behavior of backdoor attacks in\ntime-varying models, whose model weights are continually updated via\nfine-tuning to adapt to data drifts. Our theoretical analysis shows how\nfine-tuning with fresh data progressively \"erases\" the injected backdoors, and\nour empirical study illustrates how quickly a time-varying model \"forgets\"\nbackdoors under a variety of training and attack settings. We also show that\nnovel fine-tuning strategies using smart learning rates can significantly\naccelerate backdoor forgetting. Finally, we discuss the need for new backdoor\ndefenses that target time-varying models specifically.\n", "date": "2022-06-08", "categories": []} {"id": "2302.04116", "title": "Training-free Lexical Backdoor Attacks on Language Models", "abstract": " Large-scale language models have achieved tremendous success across various\nnatural language processing (NLP) applications. Nevertheless, language models\nare vulnerable to backdoor attacks, which inject stealthy triggers into models\nfor steering them to undesirable behaviors. Most existing backdoor attacks,\nsuch as data poisoning, require further (re)training or fine-tuning language\nmodels to learn the intended backdoor patterns. The additional training process\nhowever diminishes the stealthiness of the attacks, as training a language\nmodel usually requires long optimization time, a massive amount of data, and\nconsiderable modifications to the model parameters. In this work, we propose\nTraining-Free Lexical Backdoor Attack (TFLexAttack) as the first training-free\nbackdoor attack on language models. Our attack is achieved by injecting lexical\ntriggers into the tokenizer of a language model via manipulating its embedding\ndictionary using carefully designed rules. These rules are explainable to human\ndevelopers which inspires attacks from a wider range of hackers. The sparse\nmanipulation of the dictionary also habilitates the stealthiness of our attack.\nWe conduct extensive experiments on three dominant NLP tasks based on nine\nlanguage models to demonstrate the effectiveness and universality of our\nattack. The code of this work is available at\nhttps://github.com/Jinxhy/TFLexAttack.\n", "date": "2023-02-08", "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": "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": "2301.10412", "title": "BDMMT: Backdoor Sample Detection for Language Models through Model\n Mutation Testing", "abstract": " Deep neural networks (DNNs) and natural language processing (NLP) systems\nhave developed rapidly and have been widely used in various real-world fields.\nHowever, they have been shown to be vulnerable to backdoor attacks.\nSpecifically, the adversary injects a backdoor into the model during the\ntraining phase, so that input samples with backdoor triggers are classified as\nthe target class. Some attacks have achieved high attack success rates on the\npre-trained language models (LMs), but there have yet to be effective defense\nmethods. In this work, we propose a defense method based on deep model mutation\ntesting. Our main justification is that backdoor samples are much more robust\nthan clean samples if we impose random mutations on the LMs and that backdoors\nare generalizable. We first confirm the effectiveness of model mutation testing\nin detecting backdoor samples and select the most appropriate mutation\noperators. We then systematically defend against three extensively studied\nbackdoor attack levels (i.e., char-level, word-level, and sentence-level) by\ndetecting backdoor samples. We also make the first attempt to defend against\nthe latest style-level backdoor attacks. We evaluate our approach on three\nbenchmark datasets (i.e., IMDB, Yelp, and AG news) and three style transfer\ndatasets (i.e., SST-2, Hate-speech, and AG news). The extensive experimental\nresults demonstrate that our approach can detect backdoor samples more\nefficiently and accurately than the three state-of-the-art defense approaches.\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": "2301.08751", "title": "Towards Understanding How Self-training Tolerates Data Backdoor\n Poisoning", "abstract": " Recent studies on backdoor attacks in model training have shown that\npolluting a small portion of training data is sufficient to produce incorrect\nmanipulated predictions on poisoned test-time data while maintaining high clean\naccuracy in downstream tasks. The stealthiness of backdoor attacks has imposed\ntremendous defense challenges in today's machine learning paradigm. In this\npaper, we explore the potential of self-training via additional unlabeled data\nfor mitigating backdoor attacks. We begin by making a pilot study to show that\nvanilla self-training is not effective in backdoor mitigation. Spurred by that,\nwe propose to defend the backdoor attacks by leveraging strong but proper data\naugmentations in the self-training pseudo-labeling stage. We find that the new\nself-training regime help in defending against backdoor attacks to a great\nextent. Its effectiveness is demonstrated through experiments for different\nbackdoor triggers on CIFAR-10 and a combination of CIFAR-10 with an additional\nunlabeled 500K TinyImages dataset. Finally, we explore the direction of\ncombining self-supervised representation learning with self-training for\nfurther improvement in backdoor defense.\n", "date": "2023-01-20", "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": "2301.08170", "title": "On the Vulnerability of Backdoor Defenses for Federated Learning", "abstract": " Federated Learning (FL) is a popular distributed machine learning paradigm\nthat enables jointly training a global model without sharing clients' data.\nHowever, its repetitive server-client communication gives room for backdoor\nattacks with aim to mislead the global model into a targeted misprediction when\na specific trigger pattern is presented. In response to such backdoor threats\non federated learning, various defense measures have been proposed. In this\npaper, we study whether the current defense mechanisms truly neutralize the\nbackdoor threats from federated learning in a practical setting by proposing a\nnew federated backdoor attack method for possible countermeasures. Different\nfrom traditional training (on triggered data) and rescaling (the malicious\nclient model) based backdoor injection, the proposed backdoor attack framework\n(1) directly modifies (a small proportion of) local model weights to inject the\nbackdoor trigger via sign flips; (2) jointly optimize the trigger pattern with\nthe client model, thus is more persistent and stealthy for circumventing\nexisting defenses. In a case study, we examine the strength and weaknesses of\nrecent federated backdoor defenses from three major categories and provide\nsuggestions to the practitioners when training federated models in practice.\n", "date": "2023-01-19", "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": "2301.01197", "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": "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": "2212.09979", "title": "Flareon: Stealthy any2any Backdoor Injection via Poisoned Augmentation", "abstract": " Open software supply chain attacks, once successful, can exact heavy costs in\nmission-critical applications. As open-source ecosystems for deep learning\nflourish and become increasingly universal, they present attackers previously\nunexplored avenues to code-inject malicious backdoors in deep neural network\nmodels. This paper proposes Flareon, a small, stealthy, seemingly harmless code\nmodification that specifically targets the data augmentation pipeline with\nmotion-based triggers. Flareon neither alters ground-truth labels, nor modifies\nthe training loss objective, nor does it assume prior knowledge of the victim\nmodel architecture, training data, and training hyperparameters. Yet, it has a\nsurprisingly large ramification on training -- models trained under Flareon\nlearn powerful target-conditional (or \"any2any\") backdoors. The resulting\nmodels can exhibit high attack success rates for any target choices and better\nclean accuracies than backdoor attacks that not only seize greater control, but\nalso assume more restrictive attack capabilities. We also demonstrate the\neffectiveness of Flareon against recent defenses. Flareon is fully open-source\nand available online to the deep learning community:\nhttps://github.com/lafeat/flareon.\n", "date": "2022-12-20", "categories": []} {"id": "2209.11715", "title": "The \"Beatrix'' Resurrections: Robust Backdoor Detection via Gram\n Matrices", "abstract": " Deep Neural Networks (DNNs) are susceptible to backdoor attacks during\ntraining. The model corrupted in this way functions normally, but when\ntriggered by certain patterns in the input, produces a predefined target label.\nExisting defenses usually rely on the assumption of the universal backdoor\nsetting in which poisoned samples share the same uniform trigger. However,\nrecent advanced backdoor attacks show that this assumption is no longer valid\nin dynamic backdoors where the triggers vary from input to input, thereby\ndefeating the existing defenses.\n In this work, we propose a novel technique, Beatrix (backdoor detection via\nGram matrix). Beatrix utilizes Gram matrix to capture not only the feature\ncorrelations but also the appropriately high-order information of the\nrepresentations. By learning class-conditional statistics from activation\npatterns of normal samples, Beatrix can identify poisoned samples by capturing\nthe anomalies in activation patterns. To further improve the performance in\nidentifying target labels, Beatrix leverages kernel-based testing without\nmaking any prior assumptions on representation distribution. We demonstrate the\neffectiveness of our method through extensive evaluation and comparison with\nstate-of-the-art defensive techniques. The experimental results show that our\napproach achieves an F1 score of 91.1% in detecting dynamic backdoors, while\nthe state of the art can only reach 36.9%.\n", "date": "2022-09-23", "categories": []} {"id": "2212.09067", "title": "Fine-Tuning Is All You Need to Mitigate Backdoor Attacks", "abstract": " Backdoor attacks represent one of the major threats to machine learning\nmodels. Various efforts have been made to mitigate backdoors. However, existing\ndefenses have become increasingly complex and often require high computational\nresources or may also jeopardize models' utility. In this work, we show that\nfine-tuning, one of the most common and easy-to-adopt machine learning training\noperations, can effectively remove backdoors from machine learning models while\nmaintaining high model utility. Extensive experiments over three machine\nlearning paradigms show that fine-tuning and our newly proposed\nsuper-fine-tuning achieve strong defense performance. Furthermore, we coin a\nnew term, namely backdoor sequela, to measure the changes in model\nvulnerabilities to other attacks before and after the backdoor has been\nremoved. Empirical evaluation shows that, compared to other defense methods,\nsuper-fine-tuning leaves limited backdoor sequela. We hope our results can help\nmachine learning model owners better protect their models from backdoor\nthreats. Also, it calls for the design of more advanced attacks in order to\ncomprehensively assess machine learning models' backdoor vulnerabilities.\n", "date": "2022-12-18", "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": "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": "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": "2211.15929", "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": "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": "2211.14719", "title": "BadPrompt: Backdoor Attacks on Continuous Prompts", "abstract": " The prompt-based learning paradigm has gained much research attention\nrecently. It has achieved state-of-the-art performance on several NLP tasks,\nespecially in the few-shot scenarios. While steering the downstream tasks, few\nworks have been reported to investigate the security problems of the\nprompt-based models. In this paper, we conduct the first study on the\nvulnerability of the continuous prompt learning algorithm to backdoor attacks.\nWe observe that the few-shot scenarios have posed a great challenge to backdoor\nattacks on the prompt-based models, limiting the usability of existing NLP\nbackdoor methods. To address this challenge, we propose BadPrompt, a\nlightweight and task-adaptive algorithm, to backdoor attack continuous prompts.\nSpecially, BadPrompt first generates candidate triggers which are indicative\nfor predicting the targeted label and dissimilar to the samples of the\nnon-targeted labels. Then, it automatically selects the most effective and\ninvisible trigger for each sample with an adaptive trigger optimization\nalgorithm. We evaluate the performance of BadPrompt on five datasets and two\ncontinuous prompt models. The results exhibit the abilities of BadPrompt to\neffectively attack continuous prompts while maintaining high performance on the\nclean test sets, outperforming the baseline models by a large margin. The\nsource code of BadPrompt is publicly available at\nhttps://github.com/papersPapers/BadPrompt.\n", "date": "2022-11-27", "categories": []} {"id": "2211.11958", "title": "A Survey on Backdoor Attack and Defense in Natural Language Processing", "abstract": " Deep learning is becoming increasingly popular in real-life applications,\nespecially in natural language processing (NLP). Users often choose training\noutsourcing or adopt third-party data and models due to data and computation\nresources being limited. In such a situation, training data and models are\nexposed to the public. As a result, attackers can manipulate the training\nprocess to inject some triggers into the model, which is called backdoor\nattack. Backdoor attack is quite stealthy and difficult to be detected because\nit has little inferior influence on the model's performance for the clean\nsamples. To get a precise grasp and understanding of this problem, in this\npaper, we conduct a comprehensive review of backdoor attacks and defenses in\nthe field of NLP. Besides, we summarize benchmark datasets and point out the\nopen issues to design credible systems to defend against backdoor attacks.\n", "date": "2022-11-22", "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": "2210.12179", "title": "Neural Architectural Backdoors", "abstract": " This paper asks the intriguing question: is it possible to exploit neural\narchitecture search (NAS) as a new attack vector to launch previously\nimprobable attacks? Specifically, we present EVAS, a new attack that leverages\nNAS to find neural architectures with inherent backdoors and exploits such\nvulnerability using input-aware triggers. Compared with existing attacks, EVAS\ndemonstrates many interesting properties: (i) it does not require polluting\ntraining data or perturbing model parameters; (ii) it is agnostic to downstream\nfine-tuning or even re-training from scratch; (iii) it naturally evades\ndefenses that rely on inspecting model parameters or training data. With\nextensive evaluation on benchmark datasets, we show that EVAS features high\nevasiveness, transferability, and robustness, thereby expanding the adversary's\ndesign spectrum. We further characterize the mechanisms underlying EVAS, which\nare possibly explainable by architecture-level ``shortcuts'' that recognize\ntrigger patterns. This work raises concerns about the current practice of NAS\nand points to potential directions to develop effective countermeasures.\n", "date": "2022-10-21", "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": "2206.08514", "title": "A Unified Evaluation of Textual Backdoor Learning: Frameworks and\n Benchmarks", "abstract": " Textual backdoor attacks are a kind of practical threat to NLP systems. By\ninjecting a backdoor in the training phase, the adversary could control model\npredictions via predefined triggers. As various attack and defense models have\nbeen proposed, it is of great significance to perform rigorous evaluations.\nHowever, we highlight two issues in previous backdoor learning evaluations: (1)\nThe differences between real-world scenarios (e.g. releasing poisoned datasets\nor models) are neglected, and we argue that each scenario has its own\nconstraints and concerns, thus requires specific evaluation protocols; (2) The\nevaluation metrics only consider whether the attacks could flip the models'\npredictions on poisoned samples and retain performances on benign samples, but\nignore that poisoned samples should also be stealthy and semantic-preserving.\nTo address these issues, we categorize existing works into three practical\nscenarios in which attackers release datasets, pre-trained models, and\nfine-tuned models respectively, then discuss their unique evaluation\nmethodologies. On metrics, to completely evaluate poisoned samples, we use\ngrammar error increase and perplexity difference for stealthiness, along with\ntext similarity for validity. After formalizing the frameworks, we develop an\nopen-source toolkit OpenBackdoor to foster the implementations and evaluations\nof textual backdoor learning. With this toolkit, we perform extensive\nexperiments to benchmark attack and defense models under the suggested\nparadigm. To facilitate the underexplored defenses against poisoned datasets,\nwe further propose CUBE, a simple yet strong clustering-based defense baseline.\nWe hope that our frameworks and benchmarks could serve as the cornerstones for\nfuture model development and evaluations.\n", "date": "2022-06-17", "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": "2210.11264", "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": "2206.12654", "title": "BackdoorBench: A Comprehensive Benchmark of Backdoor Learning", "abstract": " Backdoor learning is an emerging and vital topic for studying deep neural\nnetworks' vulnerability (DNNs). Many pioneering backdoor attack and defense\nmethods are being proposed, successively or concurrently, in the status of a\nrapid arms race. However, we find that the evaluations of new methods are often\nunthorough to verify their claims and accurate performance, mainly due to the\nrapid development, diverse settings, and the difficulties of implementation and\nreproducibility. Without thorough evaluations and comparisons, it is not easy\nto track the current progress and design the future development roadmap of the\nliterature. To alleviate this dilemma, we build a comprehensive benchmark of\nbackdoor learning called BackdoorBench. It consists of an extensible\nmodular-based codebase (currently including implementations of 8\nstate-of-the-art (SOTA) attacks and 9 SOTA defense algorithms) and a\nstandardized protocol of complete backdoor learning. We also provide\ncomprehensive evaluations of every pair of 8 attacks against 9 defenses, with 5\npoisoning ratios, based on 5 models and 4 datasets, thus 8,000 pairs of\nevaluations in total. We present abundant analysis from different perspectives\nabout these 8,000 evaluations, studying the effects of different factors in\nbackdoor learning. All codes and evaluations of BackdoorBench are publicly\navailable at \\url{https://backdoorbench.github.io}.\n", "date": "2022-06-25", "categories": []} {"id": "2210.09545", "title": "Fine-mixing: Mitigating Backdoors in Fine-tuned Language Models", "abstract": " Deep Neural Networks (DNNs) are known to be vulnerable to backdoor attacks.\nIn Natural Language Processing (NLP), DNNs are often backdoored during the\nfine-tuning process of a large-scale Pre-trained Language Model (PLM) with\npoisoned samples. Although the clean weights of PLMs are readily available,\nexisting methods have ignored this information in defending NLP models against\nbackdoor attacks. In this work, we take the first step to exploit the\npre-trained (unfine-tuned) weights to mitigate backdoors in fine-tuned language\nmodels. Specifically, we leverage the clean pre-trained weights via two\ncomplementary techniques: (1) a two-step Fine-mixing technique, which first\nmixes the backdoored weights (fine-tuned on poisoned data) with the pre-trained\nweights, then fine-tunes the mixed weights on a small subset of clean data; (2)\nan Embedding Purification (E-PUR) technique, which mitigates potential\nbackdoors existing in the word embeddings. We compare Fine-mixing with typical\nbackdoor mitigation methods on three single-sentence sentiment classification\ntasks and two sentence-pair classification tasks and show that it outperforms\nthe baselines by a considerable margin in all scenarios. We also show that our\nE-PUR method can benefit existing mitigation methods. Our work establishes a\nsimple but strong baseline defense for secure fine-tuned NLP models against\nbackdoor attacks.\n", "date": "2022-10-18", "categories": []} {"id": "2210.09194", "title": "Marksman Backdoor: Backdoor Attacks with Arbitrary Target Class", "abstract": " In recent years, machine learning models have been shown to be vulnerable to\nbackdoor attacks. Under such attacks, an adversary embeds a stealthy backdoor\ninto the trained model such that the compromised models will behave normally on\nclean inputs but will misclassify according to the adversary's control on\nmaliciously constructed input with a trigger. While these existing attacks are\nvery effective, the adversary's capability is limited: given an input, these\nattacks can only cause the model to misclassify toward a single pre-defined or\ntarget class. In contrast, this paper exploits a novel backdoor attack with a\nmuch more powerful payload, denoted as Marksman, where the adversary can\narbitrarily choose which target class the model will misclassify given any\ninput during inference. To achieve this goal, we propose to represent the\ntrigger function as a class-conditional generative model and to inject the\nbackdoor in a constrained optimization framework, where the trigger function\nlearns to generate an optimal trigger pattern to attack any target class at\nwill while simultaneously embedding this generative backdoor into the trained\nmodel. Given the learned trigger-generation function, during inference, the\nadversary can specify an arbitrary backdoor attack target class, and an\nappropriate trigger causing the model to classify toward this target class is\ncreated accordingly. We show empirically that the proposed framework achieves\nhigh attack performance while preserving the clean-data performance in several\nbenchmark datasets, including MNIST, CIFAR10, GTSRB, and TinyImageNet. The\nproposed Marksman backdoor attack can also easily bypass existing backdoor\ndefenses that were originally designed against backdoor attacks with a single\ntarget class. Our work takes another significant step toward understanding the\nextensive risks of backdoor attacks in practice.\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": "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": "2210.06509", "title": "Understanding Impacts of Task Similarity on Backdoor Attack and\n Detection", "abstract": " With extensive studies on backdoor attack and detection, still fundamental\nquestions are left unanswered regarding the limits in the adversary's\ncapability to attack and the defender's capability to detect. We believe that\nanswers to these questions can be found through an in-depth understanding of\nthe relations between the primary task that a benign model is supposed to\naccomplish and the backdoor task that a backdoored model actually performs. For\nthis purpose, we leverage similarity metrics in multi-task learning to formally\ndefine the backdoor distance (similarity) between the primary task and the\nbackdoor task, and analyze existing stealthy backdoor attacks, revealing that\nmost of them fail to effectively reduce the backdoor distance and even for\nthose that do, still much room is left to further improve their stealthiness.\nSo we further design a new method, called TSA attack, to automatically generate\na backdoor model under a given distance constraint, and demonstrate that our\nnew attack indeed outperforms existing attacks, making a step closer to\nunderstanding the attacker's limits. Most importantly, we provide both\ntheoretic results and experimental evidence on various datasets for the\npositive correlation between the backdoor distance and backdoor detectability,\ndemonstrating that indeed our task similarity analysis help us better\nunderstand backdoor risks and has the potential to identify more effective\nmitigations.\n", "date": "2022-10-12", "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": "2210.05929", "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": "2210.06428", "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": "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": "2208.10608", "title": "RIBAC: Towards Robust and Imperceptible Backdoor Attack against Compact\n DNN", "abstract": " Recently backdoor attack has become an emerging threat to the security of\ndeep neural network (DNN) models. To date, most of the existing studies focus\non backdoor attack against the uncompressed model; while the vulnerability of\ncompressed DNNs, which are widely used in the practical applications, is little\nexploited yet. In this paper, we propose to study and develop Robust and\nImperceptible Backdoor Attack against Compact DNN models (RIBAC). By performing\nsystematic analysis and exploration on the important design knobs, we propose a\nframework that can learn the proper trigger patterns, model parameters and\npruning masks in an efficient way. Thereby achieving high trigger stealthiness,\nhigh attack success rate and high model efficiency simultaneously. Extensive\nevaluations across different datasets, including the test against the\nstate-of-the-art defense mechanisms, demonstrate the high robustness,\nstealthiness and model efficiency of RIBAC. Code is available at\nhttps://github.com/huyvnphan/ECCV2022-RIBAC\n", "date": "2022-08-22", "categories": []} {"id": "2208.10231", "title": "An anomaly detection approach for backdoored neural networks: face\n recognition as a case study", "abstract": " Backdoor attacks allow an attacker to embed functionality jeopardizing proper\nbehavior of any algorithm, machine learning or not. This hidden functionality\ncan remain inactive for normal use of the algorithm until activated by the\nattacker. Given how stealthy backdoor attacks are, consequences of these\nbackdoors could be disastrous if such networks were to be deployed for\napplications as critical as border or access control. In this paper, we propose\na novel backdoored network detection method based on the principle of anomaly\ndetection, involving access to the clean part of the training data and the\ntrained network. We highlight its promising potential when considering various\ntriggers, locations and identity pairs, without the need to make any\nassumptions on the nature of the backdoor and its setup. We test our method on\na novel dataset of backdoored networks and report detectability results with\nperfect scores.\n", "date": "2022-08-22", "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": "2208.06537", "title": "Defense against Backdoor Attacks via Identifying and Purifying Bad\n Neurons", "abstract": " The opacity of neural networks leads their vulnerability to backdoor attacks,\nwhere hidden attention of infected neurons is triggered to override normal\npredictions to the attacker-chosen ones. In this paper, we propose a novel\nbackdoor defense method to mark and purify the infected neurons in the\nbackdoored neural networks. Specifically, we first define a new metric, called\nbenign salience. By combining the first-order gradient to retain the\nconnections between neurons, benign salience can identify the infected neurons\nwith higher accuracy than the commonly used metric in backdoor defense. Then, a\nnew Adaptive Regularization (AR) mechanism is proposed to assist in purifying\nthese identified infected neurons via fine-tuning. Due to the ability to adapt\nto different magnitudes of parameters, AR can provide faster and more stable\nconvergence than the common regularization mechanism in neuron purifying.\nExtensive experimental results demonstrate that our method can erase the\nbackdoor in neural networks with negligible performance degradation.\n", "date": "2022-08-13", "categories": []} {"id": "2208.06592", "title": "Confidence Matters: Inspecting Backdoors in Deep Neural Networks via\n Distribution Transfer", "abstract": " Backdoor attacks have been shown to be a serious security threat against deep\nlearning models, and detecting whether a given model has been backdoored\nbecomes a crucial task. Existing defenses are mainly built upon the observation\nthat the backdoor trigger is usually of small size or affects the activation of\nonly a few neurons. However, the above observations are violated in many cases\nespecially for advanced backdoor attacks, hindering the performance and\napplicability of the existing defenses. In this paper, we propose a backdoor\ndefense DTInspector built upon a new observation. That is, an effective\nbackdoor attack usually requires high prediction confidence on the poisoned\ntraining samples, so as to ensure that the trained model exhibits the targeted\nbehavior with a high probability. Based on this observation, DTInspector first\nlearns a patch that could change the predictions of most high-confidence data,\nand then decides the existence of backdoor by checking the ratio of prediction\nchanges after applying the learned patch on the low-confidence data. Extensive\nevaluations on five backdoor attacks, four datasets, and three advanced\nattacking types demonstrate the effectiveness of the proposed defense.\n", "date": "2022-08-13", "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": "2207.00762", "title": "Backdoor Attack is a Devil in Federated GAN-based Medical Image\n Synthesis", "abstract": " Deep Learning-based image synthesis techniques have been applied in\nhealthcare research for generating medical images to support open research.\nTraining generative adversarial neural networks (GAN) usually requires large\namounts of training data. Federated learning (FL) provides a way of training a\ncentral model using distributed data from different medical institutions while\nkeeping raw data locally. However, FL is vulnerable to backdoor attack, an\nadversarial by poisoning training data, given the central server cannot access\nthe original data directly. Most backdoor attack strategies focus on\nclassification models and centralized domains. In this study, we propose a way\nof attacking federated GAN (FedGAN) by treating the discriminator with a\ncommonly used data poisoning strategy in backdoor attack classification models.\nWe demonstrate that adding a small trigger with size less than 0.5 percent of\nthe original image size can corrupt the FL-GAN model. Based on the proposed\nattack, we provide two effective defense strategies: global malicious detection\nand local training regularization. We show that combining the two defense\nstrategies yields a robust medical image generation.\n", "date": "2022-07-02", "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": "2207.12863", "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": "2207.10825", "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": "2207.00278", "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": "2207.05937", "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": "2207.05641", "title": "Backdoor Attacks on Crowd Counting", "abstract": " Crowd counting is a regression task that estimates the number of people in a\nscene image, which plays a vital role in a range of safety-critical\napplications, such as video surveillance, traffic monitoring and flow control.\nIn this paper, we investigate the vulnerability of deep learning based crowd\ncounting models to backdoor attacks, a major security threat to deep learning.\nA backdoor attack implants a backdoor trigger into a target model via data\npoisoning so as to control the model's predictions at test time. Different from\nimage classification models on which most of existing backdoor attacks have\nbeen developed and tested, crowd counting models are regression models that\noutput multi-dimensional density maps, thus requiring different techniques to\nmanipulate.\n In this paper, we propose two novel Density Manipulation Backdoor Attacks\n(DMBA$^{-}$ and DMBA$^{+}$) to attack the model to produce arbitrarily large or\nsmall density estimations. Experimental results demonstrate the effectiveness\nof our DMBA attacks on five classic crowd counting models and four types of\ndatasets. We also provide an in-depth analysis of the unique challenges of\nbackdooring crowd counting models and reveal two key elements of effective\nattacks: 1) full and dense triggers and 2) manipulation of the ground truth\ncounts or density maps. Our work could help evaluate the vulnerability of crowd\ncounting models to potential backdoor attacks.\n", "date": "2022-07-12", "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": "2207.03895", "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": "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": "2206.09272", "title": "DECK: Model Hardening for Defending Pervasive Backdoors", "abstract": " Pervasive backdoors are triggered by dynamic and pervasive input\nperturbations. They can be intentionally injected by attackers or naturally\nexist in normally trained models. They have a different nature from the\ntraditional static and localized backdoors that can be triggered by perturbing\na small input area with some fixed pattern, e.g., a patch with solid color.\nExisting defense techniques are highly effective for traditional backdoors.\nHowever, they may not work well for pervasive backdoors, especially regarding\nbackdoor removal and model hardening. In this paper, we propose a novel model\nhardening technique against pervasive backdoors, including both natural and\ninjected backdoors. We develop a general pervasive attack based on an\nencoder-decoder architecture enhanced with a special transformation layer. The\nattack can model a wide range of existing pervasive backdoor attacks and\nquantify them by class distances. As such, using the samples derived from our\nattack in adversarial training can harden a model against these backdoor\nvulnerabilities. Our evaluation on 9 datasets with 15 model structures shows\nthat our technique can enlarge class distances by 59.65% on average with less\nthan 1% accuracy degradation and no robustness loss, outperforming five\nhardening techniques such as adversarial training, universal adversarial\ntraining, MOTH, etc. It can reduce the attack success rate of six pervasive\nbackdoor attacks from 99.06% to 1.94%, surpassing seven state-of-the-art\nbackdoor removal techniques.\n", "date": "2022-06-18", "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": "2206.07840", "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": "2206.04881", "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": "2206.01832", "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": "2206.01102", "title": "A temporal chrominance trigger for clean-label backdoor attack against\n anti-spoof rebroadcast detection", "abstract": " We propose a stealthy clean-label video backdoor attack against Deep Learning\n(DL)-based models aiming at detecting a particular class of spoofing attacks,\nnamely video rebroadcast attacks. The injected backdoor does not affect\nspoofing detection in normal conditions, but induces a misclassification in the\npresence of a specific triggering signal. The proposed backdoor relies on a\ntemporal trigger altering the average chrominance of the video sequence. The\nbackdoor signal is designed by taking into account the peculiarities of the\nHuman Visual System (HVS) to reduce the visibility of the trigger, thus\nincreasing the stealthiness of the backdoor. To force the network to look at\nthe presence of the trigger in the challenging clean-label scenario, we choose\nthe poisoned samples used for the injection of the backdoor following a\nso-called Outlier Poisoning Strategy (OPS). According to OPS, the triggering\nsignal is inserted in the training samples that the network finds more\ndifficult to classify. The effectiveness of the proposed backdoor attack and\nits generality are validated experimentally on different datasets and\nanti-spoofing rebroadcast detection architectures.\n", "date": "2022-06-02", "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": "2606.22837", "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": "2606.23361", "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": "2606.28962", "title": "FlipGuard: Defending Large Language Models Against Quantization-Conditioned Backdoor Attacks", "abstract": "Model quantization is essential for the efficient deployment of Large Language Models (LLMs), but introduces a critical vulnerability: Quantization-Conditioned Backdoor (QCB) attacks. In these attacks, malicious behaviors remain dormant in full-precision models and activate only after specific quantization distortions, bypassing standard security audits. To mitigate this, we introduce FlipGuard, a proactive defense framework that selectively perturbs model weights prior to quantization. By breaking the adversary's precise alignment between weight patterns and quantization boundaries, FlipGuard suppresses backdoor activation without requiring access to training data or trigger samples. We further propose the Defense Effectiveness Ratio (DER), a unified metric to jointly evaluate security gains, utility preservation, and computational cost. Extensive experiments across seven LLMs (including StarCoder and LLaMA-family models) and three quantization schemes (INT8, FP4, NF4) demonstrate that FlipGuard effectively neutralizes QCBs across three scenarios, i.e., vulnerable code generation, content injection, and over-refusal, achieving high security with negligible performance degradation.", "date": "2026-06-27", "categories": []} {"id": "2606.29451", "title": "The Platonic Defense: Backdoor Defense for Self-Supervised Encoders in the Era of Large Scale Pre-training", "abstract": "Self-supervised learning (SSL) pretrained models have become a dominant paradigm for visual representation learning, but they are vulnerable to backdoor attacks. Existing defenses struggle to defend against such attacks in a fully black-box setting because they often require access to labels, attack patterns, or training data. To tackle this issue, we propose a new attack-agnostic, model-agnostic, and modality-agnostic black-box test-time defense paradigm, called \\emph{Platonic Representation Defense}. It is inspired by the Platonic Representation Hypothesis, which suggests that large-scale independently trained encoders converge toward compatible projections of the same underlying reality. We formalize this idea as a conditional energy function defined over source representations and a set of reference representations. The energy function is trained for detection through noise-contrastive estimation and for representation purification through denoising score matching. Theoretically, the energy gap between matched and mismatched samples is lower bounded by the mutual information between source and reference representations. We demonstrate the effectiveness of our method on multiple self-supervised encoders and more than 10 attacks. The method can perform both representation detection and purification, and achieves substantial performance gains across multiple attacks. Code is available \\href{https://github.com/jsrdcht/Platonic-Representation-Defense}{here}.", "date": "2026-06-28", "categories": []} {"id": "2606.30595", "title": "Wireless Backdoor Attack and Defense for Semantic Communications over Multiple Access Channel", "abstract": "Semantic communication (SemCom) aims to preserve semantic meaning and task-oriented information beyond conventional message recovery over wireless channels. The adoption of SemCom in shared-access wireless networks introduces new vulnerabilities for multi-user semantic inference. This paper considers a SemCom system for two transmitters communicating with a common receiver over a multiple access channel. Each transmitter maps source information into latent semantic representations, while the receiver jointly reconstructs and classifies the semantic information for both transmitters. A selective over-the-air backdoor (Trojan) attack is presented in which an adversary transmits a low-power trigger waveform over the air and injects it into the shared received signal during training. By transmitting the trigger again during testing, this stealthy, low-power attack selectively manipulates the semantic inference for one transmitter while minimally affecting the inference of the other transmitter. To mitigate this vulnerability, a trigger-aware defense mechanism is developed to preserve correct semantic labels under trigger-contaminated wireless observations. The results demonstrate both the vulnerability of shared-access SemCom systems to selective over-the-air backdoor attacks and the effectiveness of trigger-aware robust training for semantic protection.", "date": "2026-06-29", "categories": []} {"id": "2606.30899", "title": "Curvature-Guided Module Localization for Low-Rank Detoxification of Backdoored Large Language Models", "abstract": "Backdoor attacks pose a serious threat to large language models (LLMs) by causing otherwise benign systems to produce attacker-specified malicious behavior when a hidden trigger is present. In this work, we study post hoc detoxification of backdoored LLMs in a practical setting where the defender has access to the poisoned model but does not wish to retrain the full network from scratch. We propose a mechanistically guided weight-space repair framework that first localizes modules involved in propagating trigger-induced behavior using activation patching and Fisher/K-FAC curvature analysis, and then applies targeted low-rank repair to only the most influential modules. We evaluate the method on poisoned variants of \\texttt{Llama-3.2-1B-Instruct} with triggers inserted at the beginning, middle, and end of otherwise benign prompts. Results show that the proposed approach substantially suppresses trigger-conditioned malicious responses while preserving benign model behavior. These findings suggest that backdoor removal in LLMs can be formulated as a localized structural repair problem rather than only a broad behavioral alignment problem.", "date": "2026-06-29", "categories": []} {"id": "2606.31309", "title": "CSO-LLM: Class Subspace Orthogonalization for Post-Training Backdoor Detection and Trigger Inversion in LLMs", "abstract": "While post-training backdoor detection and trigger inversion schemes have been developed for AIs used e.g. for images, there is a paucity of such methods for LLMs. First, the LLM input space is discrete, with up to 150,000^k k-tuples to consider with k the token-length of a putative trigger. Second, one must blacklist tokens typical of the putative target response (class) of an attack, as such tokens may give false detection signals. However, a comprehensive blacklist is not available, in general, for a given domain. We develop a highly effective detection and inversion framework for LLMs treated as classifiers. Central to our approach is class subspace orthogonalization (CSO), a novel plug-and-play paradigm for backdoor detection that serves two fundamental roles when applied to LLMs: i) it enhances both sensitivity and specificity of a baseline detector; ii) it provides a form of implicit blacklisting, as it penalizes against inclusion, in a candidate trigger, of tokens that induce signal perturbations \"in the direction of\" the putative target class of an attack. One version of our detector performs continuous optimization in token embedding space, while a companion trigger-inversion and detection method performs greedy accretion in discrete token space. Our methods give both strong detection performance and accurate inversion of ground-truth triggers on several LLM classification domains, and for several different LLM architectures.", "date": "2026-06-30", "categories": []}