Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection
Xuechao Zou, Shun Zhang, Kai Li, Yi Zhou, Xinyu Sun, Yuhui Chen, Zhe Wu, Congyan Lang, Junliang Xing
cs.CV, cs.AI, cs.MA
2026-08-07
Open-source multi-agent deepfake detector (four forensic experts plus a judge) beats GPT-5.5 and Gemini-2.5-Pro on a 100K-video benchmark, 69.9% accuracy on unseen generators.
A new generation of video generators (Seedance 2.0, Sora-class systems) produces deepfake videos with no obvious blending boundary and no low-level artifacts frame by frame. Conventional detectors and multimodal LLMs (MLLMs), when they rely on a single model or a single analytical angle, miss these faint forgery cues and fail on generators they were not trained on.
Existing benchmarks are also dated: they cover few recent synthesis methods and mostly carry only real/fake labels, with no explainable forensic account. This paper fills both gaps with FaceVid-Forensics-100K, a 100,000-video benchmark spanning 33 synthesis methods, and ARGUS, a detection framework built entirely from small open-source MLLMs that nonetheless outperforms closed-source GPT and Gemini models on cross-generator generalization.
The core idea is to split "is this face real or fake" into four non-interfering forensic angles, each handled by a specialist agent:
Each agent outputs only what it observes, never a verdict. This separation prevents a single salient cue from biasing the whole call. A judge agent then reconciles the four reports, weighs corroborating and conflicting evidence, and produces a binary verdict plus a forensic explanation.
Training is two-stage: supervised fine-tuning (SFT) on all agents, then GRPO policy optimization to sharpen the judge while the observation agents stay frozen. The annotation pipeline is similar in spirit: five MLLMs (GPT-4o, Gemini 3.5 Flash, Qwen2.5-VL, among others) each produce observations, and DeepSeek-V4 Pro resolves conflicts per dimension to produce explanations consistent with the final verdict.
The headline test is out-of-domain (OOD): evaluated on generators unseen during training.
| Method class | Acc | Recall | F1 |
| Small vision models (best) | 64.28% | 42.97% | 45.20% |
| Closed-source MLLMs (best, Gemini-2.5-Pro) | 63.78% | 75.29% | 47.45% |
| ARGUS (no video to judge) | 67.41% | 65.00% | 51.01% |
| ARGUS (judge sees video) | 69.87% | 81.82% | 53.28% |
F1 beats the strongest single-model baseline (Gemini-2.5-Pro at 47.45%) by 5.83 points, and the whole framework is built from 7B-8B open-source models. Letting the judge see video frames lifts F1 from 51.01% to 53.28%.
Two caveats from the paper itself. On generators seen during training, the dedicated vision detector TFCU still hits 98.66% accuracy against this framework's 87.34%; the win is in generalization and interpretability, not raw in-domain accuracy. And a 53% F1 means high recall (81.82%) but plenty of false alarms, which reflects how hard unseen-generator detection genuinely is. For explanation quality, DeepSeek-V4 Pro scores the system 6.37/10 against GPT-5-mini's 6.25 for comparable methods. Per-video latency runs 5.79 to 6.13 seconds.
Two things matter for practitioners. First, interpretability: the system does not just emit a label, it gives forensic reasons ("lighting is inconsistent", "hair lacks strand-level motion"), which suits settings that need human accountability such as content moderation, forensics, and appeals. Second, small models beating large ones: a combination of 7B-8B open models out-generalizes GPT-5.5 and Gemini, which suggests that for detection tasks needing multiple perspectives, division of labor beats simply scaling up a single model.
If your workload is offline detection of known generators, a specialized small vision model is faster and more accurate. If you face a steady stream of new generators and need to justify each call, this design is worth borrowing. The 100K-video benchmark with fine-grained annotations is itself a usable asset.