HP-MIA: A novel membership inference attack scheme for high membership prediction precision

Membership Inference Attacks (MIAs) have been considered as one of the major privacy threats in recent years, especially in machine learning models. Most canonical MIAs identify whether a specific data point was presented in the confidential training set of a neural network by analyzing its output pattern on such data point. However, these methods heavily rely on overfitting and are difficult to achieve high precision. Although some recent works, such as difficulty calibration techniques, have tried to tackle this problem in a tentative manner, identifying members with high precision is still a difficult task.To address above challenge, in this paper we rethink how overfitting impacts MIA and argue that it can provide much clearer signals of non-member samples. In scenarios where the cost of launching an attack is high, such signals can avoid unnecessary attacks and reduce the attack's false positive rate. Based on our observation, we propose High-Precision MIA (HP-MIA), a novel two-stage attack scheme that leverages membership exclusion techniques to guarantee high membership prediction precision. Our empirical results have illustrated that our two-stage attack can significantly increase the number of identified members while guaranteeing high precision.