Teacher-Student Structure for Domain Adaptation in Ensemble Audio-Visual Video Deepfake Detection

The article presents the EAV-DFD method, a deep ensemble audio-visual model enhanced by a teacher-student framework for domain adaptation in deepfake detection. It utilizes the FakeAVCeleb dataset for training and demonstrates improved AUC performance on unseen datasets (DFDC, Deepfake_TIMIT, and PolyGlotFake) with gains of 4.09%, 17.94%, and 0.5%, respectively, indicating its effectiveness in generalizing across different domains. This approach is significant for practitioners as it enhances the robustness of deepfake detection systems against variations in media content, crucial for real-world applications.