Reproduce “Estimating Model-Level Membership Inference Vulnerability Without Reference Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Euodia Dodd, Nataša Krčo, Igor Shilov, Matthew Wicker, Yves-Alexandre de Montjoye

The authors introduce a reference-free estimator that predicts a model-level LiRA audit’s performance from the target model’s training and held-out losses, avoiding extra reference-model training. It uses loss TNR when nonmember losses have a heavy high-loss tail and AUC when distributions are relatively symmetric. Across tested image configurations, RMSE is 0.036; across five GPT-2 sizes, it is 0.018. The estimator does not identify individual members.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.