Member Inference Attacks in Federated Contrastive Learning

In the past, the research community has studied privacy issues in federated learning, self-supervised learning, and deep models. However, privacy investigations into the domain of federated contrast learning are rarely exploited. Consequently, our research endeavours to unveil the potential privacy risks intrinsic to federated contrast learning. In this paper, we introduce four types of membership inference attacks to probe into and analyse the privacy protection performance of federated contrast learning models. To gain a more holistic understanding of the privacy concerns in federated contrast learning, we systematically assess the efficacy of various membership inference attacks within this realm. Simultaneously, we scrutinise the potential risks posed by these attack methods from multiple perspectives and examine their applicability in real-world settings. Through these evaluations, our objective is to furnish the academic community with a more lucid viewpoint, thereby fostering a comprehensive appreciation of the privacy safeguarding capabilities of federated contrast learning models.