Resisting Membership Inference Attacks by Dynamically Adjusting Loss Targets

Machine learning (ML) models are susceptible to membership inference attacks (MIAs), which aim to infer whether a particular sample was involved in model training. Previous research suggests that the difference in loss distribution between member and non-member sample is an essential factor of MIAs. In the latest mitigation strategies to reduce the loss distribution discrepancy, the model owner must manually set a loss target for the training task. However, this can be challenging due to differences in datasets and model structures. We propose a new mitigation strategy based on existing studies, which can dynamically adjust the loss target during the training process according to the model structure and dataset characteristics, to achieve a reduced loss gap. We extensively evaluated our strategy in white-box and black-box environments, respectively. Our experimental results show that our approach avoids the problem of setting loss targets and even improves the model's resistance to attacks in most cases. Specifically, the accuracy of the attacks is reduced by an average of 4.92% and 11.3% in the black-box and white-box environments, respectively.