A Multi-Agent Framework with Adaptive Orchestration for Explainable Multi-Modal Deepfake Detection

Modern deepfakes leave forensic traces across audio, visual, and cross-modal channels, yet contemporary detectors are monolithic transformer-based architectures whose post hoc explanations struggle to isolate modality-specific artefacts. We propose a multi-agent system in which five specialised detection agents each target distinct forensic features: facial biometrics and image quality, ECAPA-TDNN speaker embeddings, cross-modal lip-sync, XceptionNet spatial analysis, and FreqNet frequency-domain analysis. Each agent is a self-contained detector with its own perception, decision and explanation. An adaptive orchestrator invokes supplementary agents when first-wave agents disagree; this escalation does not significantly improve in-distribution accuracy but adds forensic evidence, and its rate rises under domain shift. Each agent produces its own modality-appropriate explanation, and a report-generation agent synthesises these into a natural-language forensic summary. On the PolyGlotFake test set, a within-source measurement, the system reaches 99.91% accuracy at τ = 0.35, set on the validation partition, with an AUC-ROC of 1.000. On the public MAVOS-DD benchmark the system detects forged speech, including unseen voice-conversion methods (AUC-ROC 0.79), while per-agent scores locate the gap in video-only forgeries (0.44); there, escalation significantly reduces errors, and one additional agent raises the overall AUC-ROC from 0.58 to 0.74, showing the framework’s ability to ameliorate system weaknesses to unseen forgeries with additional agents.