Multi-Scale Spatial-Frequency Features Representation and Learnable Cross Modal Feature Fusion in DeepFake Detection

With the proliferation of fraudulent videos driven by DeepFake technology, DeepFake detection has become a prominent research topic. Many detection methods face generalization challenges when deep feature distributions between training and testing datasets differ. To address this, we propose a novel DeepFake detection model that extracts spatial-frequency features from forged videos across multiple scales and integrates them in a balanced manner to form a reliable and generalized representation of forgery features. Specifically, we employ cascaded MBConv and Swin Transformer modules for local and global spatial feature extraction. Additionally, we analyze spectral features, including high-frequency components in spatial and channel domains, and learnable source-agnostic features. During the fusion phase, we design a learnable feature channel adaption strategy to balance contributions from different modalities and prevent feature degradation. Extensive experiments show that our method outperforms others in both intra-dataset and cross-dataset scenarios, demonstrating strong performance and generalization capability.