GAN-generated fake face detection via two-stream CNN with PRNU in the wild

The rapid development of the generative adversarial networks (GANs) has made it an unprecedented success in image generation. The emergence of BigGAN, StyleGAN, and other advanced GAN makes the generated images more real and deceptive, which poses potential threats to national security, social stability, and personal privacy. In this paper, we proposed a new framework—two-stream CNN to detect GAN generated fake images, which contains RGB stream and Photo Response Non-Uniformity (PRNU) stream, respectively. In the preprocessing stage of the RGB stream, the use of random erasing enhances the diversity of samples and assists the network to pay more attention to the difference in GAN fingerprints in the image content. The PRNU stream’s construction is based on the uniqueness of PRNU features in real images and the robustness of the features to image transformation. The existence of PRNU guides the network to focus on the changes in the image pixel value itself and enhances the generalization performance of the network. Experimental results on multiple datasets show that the proposed method has apparent advantages in accuracy and generalization and is more robust to various image transformations, such as downsampling, JPEG compression, Gaussian noise, and Gaussian blur.