FAD-Net: Fake Images Detection and Generalization Based on Frequency Domain Transformation

With the continuous development of neural network technology, the generation methods of fake images are gradually improved. More and more faked photos and face changing videos appear on major social media platforms, causing people to pay attention to their reputation security, information security, and public opinion guidance. At present, the spatial detection model has excellent results. But most of them need a large number of training sets as support. Simultaneously, the frequency detection model primarily uses complex feature extraction operations, and most of the two detection models only detect a single fake image generation method. Given the above two points, this paper makes the following work: Based on the common issues and attention mechanism of fake image generation methods on the network, FAD-Net (Frequency-domain Attention Detection Network) is designed, which is suitable for most fake image generation methods. We use the frequency domain image as the network input to train the detector. Good detection results are obtained on 11 fake image generation methods such as Deepfakes and Gan series. Compared with the best spatial detection model, FAD-Net uses a smaller training set and shorter training time to get better detection generalization, which shows the superiority of frequency information in fake image detection generalization.