Where to Focus: Central Attention-Based Face Forgery Detection

Face forgery detection in compressed images is an active area of research. However, previous frequency-based methods are subject to two limitations. One aspect to consider is that they apply the same weight to different frequency bands. Moreover, they exhibit an equal treatment of regions that contain distinct semantic information. To address these limitations above, we propose the Central Attention Network (CAN), a multi-modal architecture comprising two bright components: Adaptive Frequency Embedding (AFE) and Central Attention (CA) block. The AFE module adaptively embeds practical frequency information to enhance forged traces and minimize the impact of redundant interference. Moreover, the CA block can achieve fine-grained trace observation by concentrating on facial regions where indications of forgery frequently manifest. CAN is efficient in extracting forgery traces and robust to noise. It effectively reduces the unnecessary focus of our model on irrelevant factors. Extensive experiments on multiple datasets validate the advantages of CAN over existing state-of-the-art methods.