Multimodal DeepFake Detection via Audio-Visual Feature Fusion and Temporal Attention

With the rapid development of AI-generated content, DeepFake videos have become increasingly realistic, posing significant challenges to the authenticity of information dissemination and the security of personal identity. As a result, detecting DeepFake videos has emerged as a critical task in the field of digital media forensics. Despite extensive research efforts, existing methods still face two major limitations: (i) insufficient capacity in modeling long-range temporal dependencies, which hinders the detection of complex dynamic inconsistencies; and (ii) reliance on a single modality, making it difficult to fully exploit the complementary information between audio and visual streams. To address these issues, we propose a novel multimodal DeepFake detection framework that jointly leverages audio and visual information for video detection. Specifically, we employ a visual Transformer with temporal self-attention to model dynamic inconsistencies across sampled video frames, and adopt an audio encoder to extract discriminative features from Mel-spectrogram representations. The extracted visual and audio features are then globally aggregated and fused through a lightweight classification head. Extensive experiments demonstrate that our approach effectively captures cross-modal forgery cues and achieves strong generalization capability across diverse and challenging Deep-Fake scenarios.