ARGUS-Net: An Uncertainty-Aware Multi-Agent Framework for Multimodal Deepfake Detection With Probabilistic Evidence Fusion

Abstract

The rapid proliferation of generative models has made deepfake media increasingly realistic and accessible, posing critical threats to digital forensics and multimedia trust. Existing detection methods predominantly rely on single-modality analysis and deterministic predictions, yielding overconfident outputs and limited robustness under ambiguous and cross-dataset conditions. We introduce ARGUS-Net, an uncertainty-aware multi-agent framework for multimodal deepfake detection based on probabilistic evidence fusion. Four independent forensic agents analyzing physiological signals (rPPG), audio-visual synchronization, spectral artifacts, and CNN-based facial manipulation produce modality-specific suspicion and confidence scores. A probabilistic reasoning module fuses this multimodal evidence to estimate deepfake probability and quantify decision uncertainty via Shannon entropy. When uncertainty is high, the ARGUS-AC mechanism triggers controlled perturbations (illumination, motion, and occlusion) to re-evaluate ambiguous cases, reducing mean predictive entropy from 0.985 to 0.935. Decision-making is strictly separated from a post-hoc LLM-based forensic explanation module, ensuring interpretability without affecting the integrity of the detection process. Evaluated on Celeb-DF v2 (AUC = 0.793) and FaceForensics++ (AUC = 0.633), and further assessed on WildDeepfake under a restricted-modality external protocol, ARGUS-Net demonstrates that probabilistic multimodal reasoning combined with active uncertainty resolution improves the reliability of deepfake detection systems while remaining strongly sensitive to domain shift.