Membership Inference Attacks against GANs by Leveraging Over-representation Regions

Abstract

Generative adversarial networks (GANs) have made unprecedented performance in image synthesis and play a key role in various downstream applications of computer vision. However, GAN models trained on sensitive data also pose a distinct threat to privacy. In this poster, we present a novel over-representation based membership inference attack. Unlike prior attacks against GANs which focus on the overall metrics, such as the attack accuracy, our attack aims to make inference from the high-precision perspective, which allows the adversary to concentrate on inferring a sample as a member confidently. Initial experimental results demonstrate that the adversary can achieve a high precision attack even if the overall attack accuracy is about 50% for a well-trained GAN model. Our work will raise awareness of the importance of precision when GAN owners evaluate the privacy risks of their models.