Papers
arxiv:2608.06865

Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection

Published on Aug 7
· Submitted by
XavierJiezou
on Aug 10
Authors:
,
,
,
,
,
,
,
,

Abstract

The malicious use of generative artificial intelligence to create highly realistic deepfake videos raises serious ethical concerns and poses substantial challenges to AI safety. However, existing deepfake video benchmarks provide limited coverage of recent synthesis methods and generally lack reliable fine-grained textual annotations. Meanwhile, conventional detectors and multimodal large language models (MLLMs), whether operating as a single model or relying on a single analytical perspective, often fail to capture subtle forgery artifacts, limiting their generalization to emerging AI-generated methods. To address these limitations, we introduce FaceVid-Forensics-100K, a large-scale deepfake video dataset comprising 100,000 videos and spanning 33 synthesis methods across face swapping, face reenactment, and entire-face synthesis, including recent generators such as Seedance 2.0. The dataset provides fine-grained textual annotations of visual observations and verdict-consistent forensic explanations, automatically synthesized through a multi-model aggregation and conflict-resolution pipeline powered by advanced MLLMs. Building on this benchmark, we propose a multi-agent forensic reasoning framework that employs four specialized domain-expert agents to independently analyze forgery cues from four perspectives: texture, lighting, motion, and physics. A judge agent then reconciles their reports to produce a final prediction together with an explanation. Extensive evaluations on out-of-domain test sets show that, despite being composed entirely of small open-source MLLMs, our framework outperforms all methods including closed-source GPT and Gemini models and ranks first across all reported metrics on this benchmark. The project page is available at https://xavierjiezou.github.io/ARGUS/.

Community

Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection

This paper tackles a critical challenge in deepfake video detection: poor generalization to emerging generative models and the lack of fine-grained, evidence-grounded forensic reasoning. Conventional detectors and single MLLMs often rely on superficial artifacts or a single analytical perspective, failing to capture subtle, distributed forgery cues across video frames.

🧠 Motivation

Most existing deepfake detection approaches rely on opaque binary classification (real vs. fake) or single-model holistic judgments, making them vulnerable to shortcut learning and confirmation bias on unseen synthesis methods. Furthermore, distinct forgery traces—texture over-smoothing, illumination inconsistency, temporal motion jitter, and physical violations—stem from entirely different underlying flaws. Our goal is to introduce role-specialized multi-agent collaborative reasoning backed by a large-scale, finely annotated benchmark to achieve generalizable deepfake video forensics.

✨ Key Contributions

FaceVid-Forensics-100K Benchmark

We introduce a large-scale dataset of 100,000 face-centric videos spanning 33 synthesis methods (including recent SOTA generators such as Seedance 2.0). It provides fine-grained textual annotations of visual observations across four forensic dimensions (Texture, Lighting, Motion, Physics) and verdict-consistent explanations synthesized via a multi-model aggregation and conflict-resolution pipeline.

Multi-Agent Forensic Reasoning Framework (ARGUS)

We propose a collaborative multi-agent architecture where four specialized domain-expert observation agents independently examine texture, lighting, motion, and physics cues, while a central Judge agent reconciles multi-perspective evidence to produce the final authenticity verdict and forensic explanation.

📊 Experimental Results

Extensive evaluations on out-of-domain (OOD) test sets spanning 20 held-out modern video generators demonstrate that ARGUS consistently outperforms small vision models, forensics-tuned MLLMs, and closed-source foundation models (e.g., GPT-4o, Gemini 3.5-Flash), achieving top-ranking accuracy (69.87%), recall (81.82%), and F1 score (53.28%).

📦 Open-Source Resources

Sign up or log in to comment

Get this paper in your agent:

hf papers read 2608.06865
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2608.06865 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2608.06865 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2608.06865 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.