Andrii Yermakov

With the increasing sophistication of deepfake technology, detecting manipulated audio-visual content has become a critical challenge. While current methods primarily focus on spatial inconsistencies, temporal signals in both audio and video provide complementary cues that can enhance detection robustness. This dissertation explores deepfake detection through the lens of temporal signal analysis, leveraging the synchronization, rhythm, and temporal coherence between audio and visual streams. The research addresses issues of generalization and out-of-distribution detection to ensure the robustness of the models across diverse datasets and manipulation techniques. The proposed methods are evaluated on real-world datasets, highlighting their potential for scalable and accurate deepfake detection in multimedia forensics.