Spatial-Frequency Dual-Stream Reconstruction for Deepfake Detection

The widespread usage of Deepfake technology poses a significant threat to societal security, making the detection of Deepfakes a critical area of research. In recent years, forgery detection methods based on reconstruction errors have garnered widespread attention due to their excellent performance and generalization capabilities. However, those methods often focus on spatial reconstruction errors while neglecting the potential utility of frequency-based reconstruction errors. In this paper, we propose a novel deepfake detection framework based on Spatial-Frequency Dual-stream Reconstruction (SFDR). Specifically, our approach to forgery detection utilizes both frequency reconstruction error and spatial reconstruction error to provide complementary information that enhances the detection process. In addition, during the reconstruction, we ensure the consistency of frequency content between the original genuine images and their reconstructed versions. Finally, to mitigate the adverse impact of reconstruction tasks on the performance of forgery detection, we have refined the reconstruction loss to minimize the discrepancy between the original genuine images and their reconstructed counterparts; while simultaneously maximizing the difference between manipulated images and their reconstructions. Experimental results on multiple challenging forged datasets evaluation show that our method achieves superior performance in detection and generalization ability.