Reproduce “Membership Inference Attacks Beyond Overfitting” 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
Mona Khalil, Alberto Blanco-Justicia, Najeeb Jebreel, Josep Domingo-Ferrer
Machine learning models exhibiting good generalization can still leak sensitive training data via membership inference attacks, a vulnerability attributed to models memorizing spurious features of outlier training samples. A novel logit-reweighting defense reduces membership inference attack AUC by up to 4.33 percentage points for vulnerable samples with minimal impact on model utility.
Key insights
- While overfitting strongly correlates with Membership Inference Attack (MIA) vulnerability, certain training samples remain susceptible even in well-generalized models.
- Persistently vulnerable samples are often outliers within their classes, characterized by noise or ambiguity, causing the model to memorize spurious, non-generalizable features.
- Label smoothing, contrary to intuition, was found to consistently increase a model's susceptibility to MIAs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.