1329-GRP DeepAudit: Auditing Multimodal Deepfake Detectors Beyond Benchmark Performance
Audio-visual deepfake detectors check both the voice and the face, and they score nearperfectly on benchmarks. A high score does not reveal what they learned: on some benchmarks, the silence at the start of a clip alone separates real from fake (Smeu et al., CVPR 2025). DeepAudit: Auditing Multimodal Deepfake Detectors Beyond Benchmark Performance Prateek Paudel, Nitin Jha A detector that only watches the video can match one that listens, yet miss every fake that changes only the voice. We built DeepAudit to find out which stream a detector’s decisions follow
Authors
- Prateek Paudel
Institutions
- Kennesaw State University (US)
Publication Details
- Journal
- DigitalCommons - Kennesaw State University (Kennesaw State University)
- Published
- 2026-11-20
- Primary Topic
- Digital Media Forensic Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00