A Frequency-Aware Three-Branch Network for Deepfake Detection

An arXiv paper proposes a frequency-aware three-branch network that jointly learns spatial and frequency features from the original image and images reconstructed from different frequency channels, to alleviate the problem that existing methods over-rely on specific frequency-domain artifacts and feature redundancy. The authors derive feature decoupling and fusion losses based on mutual information theory, guiding the model to focus on task-relevant features across channels. On six large-scale benchmark datasets, the method achieves consistent SOTA performance, and the code has been released.