EDFM: An enhanced dual-branch fusion model for face deepfake detection

Face deepfake technology brings serious security risks such as privacy leakage, false information dissemination, and network fraud, which need to be widely concerned and prevented.In recent years, many detection methods were proposed, among which enhancing the robustness and generalization ability of the model has always been an important topic.In this paper, we propose a novel enhanced dual-branch fusion model to improve the robustness and generalization ability of CNN-based face deepfake detector.Our method begins by enhancing the RGB high-frequency noise in the face image to extract its abnormal features, and then performs preservation fusion.Specifically, we use a deep separable convolution module to improve the model performance when extracting image features.When extracting noise features, we use a selective kernel module to adaptively extract more representative noise features by dynamically adjusting the convolution kernel.In addition, we specially design a multi-scale channel spatial attention fusion module to effectively fuse the feature information of each part, thereby reducing model overfitting and enhancing the robustness and generalization ability of the model.Finally, through comprehensive evaluation on several benchmark datasets, it is confirmed that our method has significantly improved robustness and generalization.