TCMabNet: A Spatio-Temporal Network for Fine-Grained Fake Video

With the rapid development of generative artificial intelligence technology, deeply forged videos proliferate in social media and public communication platforms, seriously threatening information credibility and social trust. Although several detection methods have been proposed, they mostly rely on visual coding of static frames and global average pooling to model temporal features, which are difficult to effectively capture the details of local forgeries and dynamic changes between key frames, especially when dealing with high-quality forgeries or cross-modal interferences with significantly lower accuracy. In this paper, we propose TCMabNet, a deep forgery detection architecture that combines fine-grained spatial modeling with efficient temporal modeling. The method firstly maintains the continuity of image structure through spatial block division and snake rearrangement, introduces CNN to extract local features, and combines with Transformer to construct spatial dependencies across blocks, thus enhancing the global modeling capability. In the temporal dimension, the patch token sequence is dynamically modeled using the Mamba module, focusing on the temporal dependency and change trend of key frames. We evaluate TCMabNet on three challenging datasets, namely Crafter, Gen-2 and HotShot, and the experimental results show that TCMabNet significantly outperforms the existing methods in several indexes, and the average accuracy is about 4 percentage points higher than that of the mainstream baseline model, which demonstrates good generalizability and practical value.