A predictive uncertainty curve framework for membership inference attacks

Abstract

Membership Inference Attacks (MIAs) pose a notable privacy threat by determining whether specific samples were used to train a target model. Existing approaches primarily rely on static model outputs or computationally expensive shadow models, neglecting the dynamic stability of predictions under input perturbations. We propose a novel framework that systematically integrates Temperature Scaling, MC Dropout, and FGSM, shifting the attack paradigm from point-based observation to function-based uncertainty modeling. Through this multi-perturbation lens, PUC captures fine-grained memorization dynamics, enabling a comprehensive and efficient evaluation. We introduce a specialized curve feature engineering pipeline to extract geometric and statistical indicators from continuous response curves, quantifying prediction dynamic stability into compact feature vectors that enhance discriminative power. Extensive experiments on CIFAR-10, CIFAR-100, and PathMNIST validate the framework’s effectiveness. Results demonstrate that PUC outperforms competing methods, particularly on ResNet-18, maintaining robust performance even at extremely low false positive rates. Specifically, it improves the AUC by 16.7 percentage points on CIFAR-10 and 4.2 percentage points on CIFAR-100 compared to LiRA.