Dual defense: Combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacks

Abstract

Membership inference (MI) attacks threaten user privacy through determining if a given data example has been used to train a target model. Existing MI defenses protect the membership privacy through preemptive exclusion of members techniques and knowledge distillation. Unfortunately, using either of these two defenses alone, the defense effect can still offers an unsatisfactory trade-off between membership privacy and utility.

Given that the defense method that directly combines these two defenses is still very limited (e.g., the test accuracy of the target model is decreased by about 40% (in our experiments)), in this work, we propose a dual defense (DD) method that includes the preemptive exclusion of high-risk member samples module and the knowledge distillation module, which thwarts the access of the resulting models to the private training data twice to mitigate MI attacks. Our defense method can be divided into two steps: the preemptive exclusion of high-risk member samples (Step 1) and the knowledge distillation to obtain the protected student model (Step 2). We propose three types of exclusions: existing MI attacks-based exclusions, sample distances of members and nonmembers-based exclusions, and mutual information value-based exclusions, to preemptively exclude the high-risk member samples. During the knowledge distillation phase, we add ground-truth labeled data to the reference dataset to decrease the protected student model's dependency on soft labels, aiming to maintain or improve its test accuracy. Extensive evaluation shows that DD significantly outperforms state-of-the-art defenses and offers a better privacy-utility trade-off. For example, DD achieves ∼100% test accuracy improvement over the distillation for membership privacy (DMP) defense for ResNet50 trained on CIFAR100. DD simultaneously achieves the reductions in the attack effectiveness (e.g., the [email protected]%FPR of enhanced MI attacks decreased by 2.10% on the ImageNet dataset, the membership advantage (MA) of risk score-based attacks decreased by 56.30%) and improvements of the target models' test accuracies (e.g., by 42.80% on CIFAR100).