Blind Baselines Beat Membership Inference Attacks for Foundation Models

Debeshee Das, Jie Zhang, Florian Tramèr

Keywords: machine learning privacy, membership inference attacks, flawed evaluation dataset, foundation models

TL;DR: Membership inference attacks on foundation models like ChatGPT are not reliably evaluated and can be beaten by blind attacks independent of the model.

Abstract: Membership inference (MI) attacks try to determine if a data sample was used to train a machine learning model. For foundation models trained on unknown Web data, MI attacks are often used to detect copyrighted training materials, measure test set contamination, or audit machine unlearning. Unfortunately, we find that evaluations of MI attacks for foundation models are flawed, because they sample members and non-members from different distributions. For 8 published MI evaluation datasets, we show that blind attacks—that distinguish the member and non-member distributions without looking at any trained model—outperform state-of-the-art MI attacks. Existing evaluations thus tell us nothing about membership leakage of a foundation model’s training data.

Submission Number: 41