ExplainFake

Problem

Deepfake detectors output a single score with no explanation, so a flagged video can't be audited by a human reviewer or trusted in a real moderation pipeline.

Solution

Undergraduate thesis project combining pretrained visual and audio encoders through cross-attention, with attention maps forming the explainability layer; evaluated within- and cross-dataset on FakeAVCeleb and AV-Deepfake1M.