Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection

Authors: Xuechao Zou, Shun Zhang, Kai Li, Yi Zhou, Xinyu Sun, Yuhui Chen, Zhe Wu, Congyan Lang, Junliang Xing

Published: 2026-08-07 06:44:14+00:00

Comment: 22 pages, 8 figures, 14 tables

AI Summary

This research introduces FaceVid-Forensics-100K, a large-scale deepfake video dataset with fine-grained textual annotations and explanations, addressing the limitations of existing benchmarks. Building on this, they propose a multi-agent forensic reasoning framework that leverages specialized agents to analyze deepfake videos from multiple perspectives (texture, lighting, motion, physics) and a judge agent to reconcile their findings, significantly improving generalization to unseen synthesis methods.

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/.

Key findings

The multi-agent framework significantly outperforms existing methods, including closed-source GPT and Gemini models, on out-of-domain test sets, achieving 69.87% accuracy, 81.82% recall, and 53.28% F1. The specialized observation agents and their collaborative reasoning, combined with policy optimization, are crucial for generalization and producing reliable, explainable deepfake detection.

Approach

A multi-agent forensic reasoning framework is proposed, consisting of four specialized observation agents (texture, lighting, motion, physics) that independently analyze video frames and generate textual observations. A central judge agent then reconciles these reports, optionally with direct access to video frames, to produce a final deepfake prediction and a coherent explanation. This framework is trained using supervised fine-tuning and group relative policy optimization.

Datasets

FaceVid-Forensics-100K (newly introduced), AIGVDBench, GenVidBench, ViF-Bench, FaceForensics++

Model(s)

Qwen2.5-VL-7B (base model for all agents), InternVL3.5-8B, DeepSeek-V4 Pro (for label aggregation), GPT-4o, Gemini 3.5-Flash, GPT-5-mini, Gemini 3.1 Pro, EDVD-LLaMA, VidGuard-R1, Skyra, VideoVeritas

Author countries

China

← Previous