Generalizable face forgery detection based on adaptive spatial-frequency information mining

In the current field of face forgery detection, researchers are focused on recognizing forgery cues through the combination of frequency information and convolutional neural networks (CNN). However, existing methods often fail to capture spatial correlations with image content when extracting frequency features, making it difficult to accurately recognize highly simulated forged images. In addition, these methods perform well on homogeneous datasets, but their effectiveness decreases significantly when evaluated on cross-dataset samples. To address these issues, we propose a novel adaptive spatial-frequency information mining (ASFIM) method for generalizable face forgery detection. Specifically, the ASFIM method first processes the original RGB image through a frequency-aware learning module. This module extracts forgery frequency information closely related to the image content, which is subsequently used as input for frequency branching. Next, a spatial texture enhancement module is introduced to enable interaction between spatial and frequency features at an early stage. This approach not only strengthens the expressiveness of forgery features in the spatial domain but also provides an effective guide for recognizing forgery cues in the frequency domain. Finally, we designed the cross-domain interactive attention (CDIA) module to enhance forgery cues by deeply fusing spatial texture and frequency-aware features. Extensive experimental results demonstrate that the proposed ASFIM method outperforms various advanced methods in terms of generalization ability across challenging benchmark tests.