ALGANs: Enhancing membership inference attacks in federated learning with GANs and active learning

Abstract

Federated learning has received a lot of attention in recent years due to its privacy protection features. However, federated learning is susceptible to various inference attacks. Membership inference attack aims to determine whether the target data is a member or non-member of the target federated learning model, which poses a serious threat to the privacy of the training data set. Membership inference method in federated learning is dissatisfied due to a lack of attack data. Recent work shows that generative adversarial networks(GANs) can effectively enrich attack data. However, data generated by GANs lacks labels. Previous work labels data by inputting it to the target classifier model, which may be imprecise when the target model outputs ambiguous results. In this paper, to overcome the lack of attack data and the lack of labels for GANs, we propose ALGANs. ALGANs increases data diversity using GANs while applies active learning to label data generated by GANs. Membership inference attack enhanced by ALGANs has a high attack accuracy due to applying active learning to label data and extensive experimental results prove our point. We performed experiments to show that ALGAN makes membership inference attacks more threatening in federated learning.