Mamba-guided feature enhancement and fusion network for fake video detection

Abstract

The rapid dissemination of fake videos poses a serious threat to information security and social trust. Existing fake video detection methods mostly focus on image-text multimodal feature extraction, failing to fully utilize the inherent spatiotemporal features of videos, which limits detection performance. To address this issue, we proposed a Mamba-guided Feature Enhancement and Fusion Network (MFEFN) for fake video detection. Specifically, the MFEFN framework consists of three core components: an Intra-frame Spatial Modeling (ISM) module, a Cross-frame Temporal Interaction (CTI) module, and a Dynamic Collaborative Fusion (DCF) module. The ISM module employs parameter-shared Mamba blocks to process video frames in parallel, refining spatial features within each frame using dynamically generated scanning parameters to enhance local structural discriminability. The CTI module adopts a multi-directional 3D selective scanning mechanism to capture long-range temporal dependencies and dynamic inconsistencies across frames. The DCF module performs staged fusion, utilizing a Mamba-based mechanism for audio-visual temporal fusion and a cross-attention mechanism for text-image semantic fusion, effectively exploring inter-modal correlations to obtain more discriminative cross-modal fusion features. Extensive experiments conducted on FakeSV and FakeTT demonstrate that the proposed MFEFN framework outperforms existing state-of-the-art methods, providing a new and effective perspective for spatiotemporal modeling in fake video detection.