Abstract
In recent years, with the application of GANs and diffusion generative network algorithms, many highly realistic synthetic images are emerging, greatly increasing the potential for misuse, and deepfakes have become a serious social concern. To cope with indistinguishable deep forgery face images, this paper proposes a novel detection network with a generative adversarial network (GAN) and transformer as the main architectures. It adds frequency domain analysis and noise detection prediction modules. In the proposed model in which GAN is used to capture local forgery, artifacts and transformers are used to model global dependencies and predict anomalies in the forged images using frequency domain and noise information; the framework enhances the detection of subtle and diverse deep forgery patterns. Experiments on benchmark datasets show that the proposed method achieves higher accuracy and robustness compared to existing methods.