An Eyes-Based Siamese Neural Network for the Detection of GAN-Generated Face Images

Abstract

Generative Adversarial Network (GAN) models are nowadays able to generate synthetic images which are visually indistinguishable from the real ones, thus raising serious concerns about the spread of fake news and the need to develop tools to distinguish fake and real images in order to preserve the trustworthiness of digital images. The most powerful current detection methods are based on Deep Learning (DL) technology. While these methods get excellent performance when tested under conditions similar to those considered for training, they often suffer from a lack of robustness and generalization ability, as they fail to detect fake images that are generated by “unseen” GAN models. A possibility to overcome this problem is to develop tools that rely on the semantic attributes of the image. In this paper, we propose a semantic-based method for distinguishing GAN-generated from real faces, that relies on the analysis of inter-eye symmetries and inconsistencies. The method resorts to the superior capabilities of similarity learning of extracting representative and robust features. More specifically, a Siamese Neural Network (SNN) is utilized to extract high-level features characterizing the inter-eye similarity, that can be used to discriminate between real and synthetic pairs of eyes. We carried out extensive experiments to assess the performance of the proposed method in both matched and mismatched conditions pertaining to the GAN type used to generate the synthetic images and the robustness of the method in presence of post-processing. The results we got are comparable, and in some cases superior, to those achieved by the best performing state-of-the-art method leveraging on the analysis of the entire face image.