Attributing and detecting fake images generated by known GANs

Abstract

The quality of GAN-generated fake images has improved significantly, and recent GAN approaches, such as StyleGAN, achieve near indistinguishability from real images for the naked eye. As a result, adversaries are attracted to using GAN-generated fake images for disinformation campaigns and fraud on social networks. However, training an image generation network to produce realistic-looking samples remains a time-consuming and difficult problem, so adversaries are more likely to use published GAN models to generate fake images. In this paper, we analyze the frequency domain to attribute and detect fake images generated by a known GAN model. We derive a similarity metric on the frequency domain and develop a new approach for GAN image attribution. We conduct experiments on four trained GAN models and two real image datasets. Our results show high attribution accuracy against real images and those from other GAN models. We further analyze our method under evasion attempts and find the frequency-based approach is comparatively robust. In this paper, we analyze the frequency domain to attribute and detect fake images generated by a known GAN model. We derive a similarity metric on the frequency domain and develop a new approach for GAN image attribution. We conduct experiments on four trained GAN models and two real image datasets. Our results show high attribution accuracy against real images and those from other GAN models. We further analyze our method under evasion attempts and find the frequency-based approach is comparatively robust.