Deepfake detection based on Spatio-Temporal Fusion

To address the limitations of existing detection algorithms that utilize temporal information but are weak in extracting local information of forged traces, this paper introduces a detection algorithm for forged faces which based on spatiotemporal feature fusion. The algorithm comprises a spatial-frequency domain hybrid feature enhancement module and a frequency-aware dynamic gating unit module. Through the frequency domain, the high-frequency components of forged traces are enhanced to capture the global spatial and local frequency-domain cues. Subsequently, the frequency-aware dynamic gating unit is employed to amplify subtle cues in the temporal dimension that are difficult to discern. These cues from both dimensions are then integrated through an information fusion module to obtain a more comprehensive feature representation. Experimental results demonstrate that the proposed algorithm exhibits superior generalization performance on datasets in the forged domain.