A method for detecting deepfake faces by integrating multi-scale deep-shallow spatial and frequency features

In recent years, deep learning technology has been excessively used in the creation of fake videos. Deepfake technology alters or replaces the facial information of the original video, synthesizes false voices, and is used to create pornographic films, fake news, political rumors, and so on, posing significant challenges to the authenticity of information and personal privacy. Currently, most face deepfake detection networks are based on single-scale and single-source detection research, and the proposed networks and methods often lack the ability to recognize unknown attack methods and generalization ability. To address these issues, this paper proposes a face deepfake detection method that integrates deep and shallow multi-scale spatial and frequency-domain features to achieve high-accuracy recognition of deepfake faces. The network is based on feature extraction from shallow and deep layers, where the shallow branch captures detailed information from low-level features to detect small-scale targets such as minor noise, while the deep branch focuses on detecting larger and higher-level fake features with a larger receptive field. Each branch adopts a method to fuse multi-scale spatial and frequency-domain features. Multi-scale features can capture information at different levels and granularities, while frequency-domain features can supplement fake traces that are difficult to detect in the spatial domain. Extensive experiments show that tthis framework outperforms most face spoofing detection networks and achieves good performance.