Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection
The paper pairs a 100,000-video deepfake benchmark with a multi-agent detector that argues from texture, lighting, motion, and physics cues.
The authors say FaceVid-Forensics-100K spans 33 synthesis methods, including face swapping, reenactment, entire-face synthesis, and newer generators such as Seedance 2.0. Its annotations include visual observations and verdict-aligned forensic explanations generated through multi-model aggregation and conflict resolution. Their detector uses four specialized expert agents and a judge agent to reconcile the reports into a prediction and explanation. In out-of-domain tests, they report that the small open-source MLLM system beats all compared methods, including closed-source GPT and Gemini models, on the benchmark’s reported metrics. HF Daily Papers' note
The authors say FaceVid-Forensics-100K spans 33 synthesis methods, including face swapping, reenactment, entire-face synthesis, and newer generators such as Seedance 2.0. Its annotations include visual observations and verdict-aligned forensic explanations generated through multi-model aggregation and conflict resolution. Their detector uses four specialized expert agents and a judge agent to reconcile the reports into a prediction and explanation. In out-of-domain tests, they report that the small open-source MLLM system beats all compared methods, including closed-source GPT and Gemini models, on the benchmark’s reported metrics. HF Daily Papers' note
score 5