Can Forensic Detectors Identify GAN Generated Images?

Generative adversarial network (GAN) has shown its powerful capability in generating photorealistic images. Although the generated images can fool human eyes, it is not clear whether they can evade the detection of forensic detectors, which aim to identify the originality and authenticity of images. In this paper, we investigate how forensic detectors perform in differentiating between GAN generated images and real images. We consider two kinds of approaches, one is intrusive and the other is non-intrusive, based on whether the GAN architecture is needed for performing detection. We have conducted extensive experiments on a celebrity face image dataset to evaluate the effectiveness of different approaches. The results and analyses show that the intrusive approach can detect GAN generated images but with a relatively high false alarm rate. The non-intrusive approach with features extracted from a VGG network is very effective for detecting GAN generated images when the training data is sufficient, but it still faces challenge when the training data and testing data are mismatched.