Beyond Provenance: Evidence Attribution as a Framework for Synthetic Media Verification

Synthetic-media verification is often treated as a binary decision, where we decide whether the media is authentic or synthetic. We argue that verification should also identify the evidence supporting that decision, which we call evidence attribution. We propose a framework with two complementary evidence levels. Content-level evidence is derived from the media itself, including spatial, temporal, physiological, and cross-modal inconsistencies. Context-level evidence comes from the surrounding interaction and situation, including behavioural, situational-pragmatic, and narrative-semantic inconsistencies. We further present an empirical probe on FakeAVCeleb showing that a multimodal model can correctly identify manipulated media while citing evidence inconsistent with the known manipulation. This demonstrates that prediction correctness does not necessarily imply attribution correctness. Together with real-world cases where important evidence lies outside the media itself, these findings motivate synthetic-media verification based on attributable evidence across both content and context. Code is available at: https://anonymous.4open.science/r/deepfake-attribution-2FE4