FAClue: Exploring Frequency Clues by Adaptive Frequency-Attention for Deepfake Detection

Detecting fake faces produced by face forgery technologies attracts intensive attention in recent years. Deep learning approaches have shown their effectiveness in deepfake detection task. Some previous deep learning-based methods exploit forgery artifacts in spatial domain but easily overfit the specific forgery patterns. Therefore, some works utilize additional frequency domain information to obtain generalized features. We consider to improve the frequency-based methods in two aspects: 1) extracting discriminative frequency features comprehensively; 2) mining complementary features in different domains sufficiently. In this paper, we propose a dual-stream network named FAClue for deepfake detection, which extracts comprehensive frequency information to complement spatial domain features. Specifically, the FAClue consists of three main components. A Frequency-Attention Extractor (FAE) is proposed to adaptively highlight prominent frequency bands from both global and local perspectives. A RGB-Frequency Complementary Enhancement (RFCE) module is developed to mine complementary information between RGB and frequency domains in an explicit manner. A Frequency Guided Attention (FGA) module is designed to fuse different domain features and generate discriminative features for detection. Extensive experiments on three benchmark datasets demonstrate the FAClue achieves competitive performance compared with state-of-the-art methods.