Multi-scale Wavelet Transformer for Face Forgery Detection

Currently, many face forgery detection methods aggregate spatial and frequency features to enhance the generalization ability gain promising performance under cross-dataset scenario. However, these only leverage one level information which limits their expressive ability. To overcome limitations, we propose a multi-scale wavelet transformer framework for detection. Specifically, take full advantage of multi-frequency representation, gradually representation at different stages backbone network. better fuse feature with features, frequency-based attention is designed guide extractor concentrate more on traces. Meanwhile, cross-modality proposed features. These two modules are calculated through unified block efficiency. A wide variety experiments demonstrate that method efficient effective both within cross datasets.