Boosting Model Inversion Attacks with Adversarial Examples

Model inversion attacks involve reconstructing the training data of a target model, which raises serious privacy concerns for machine learning models. However, these attacks, especially learning-based methods, are likely to suffer from low attack accuracy, i.e., classification accuracy reconstructed by classifiers. Recent studies showed an alternative strategy model GAN-based optimization, can improve effectively. series reconstruct only class-representative class, whereas diverse different in each class. Hence, this paper, we propose new paradigm that achieve higher black-box setting. First, regularize process with added semantic loss function and, second, inject adversarial examples into increase diversity class-related parts (i.e., he essential features tasks) data. This scheme guides pay more attention original during reconstruction process. The experimental results show our method greatly boosts performance existing attacks. Even when no extra queries allowed, approach still shows severity threat adversaries is underestimated and robust defenses required.