Please Tell Me More: Privacy Impact of Explainability through the Lens of Membership Inference Attack
Han Liu, Yuhao Wu, Zhiyuan Yu, Ning Zhang
IEEE Symposium on Security and Privacy 2024 · Day 3 · Continental Ballroom 6
AI review
This research meticulously demonstrates how Explainable AI (XAI) methods, intended for transparency, inadvertently amplify privacy risks by providing new, exploitable signals for membership inference attacks. The attack leverages distinct attribution maps and prediction trajectories for training members versus non-members, a clever and concerning new vector. It's a critical warning that transparency and privacy are often at odds in ML systems.