Reproduce “Generalized Leverage Score for Scalable Assessment of Privacy Vulnerability” 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
Valentin Dorseuil, Jamal Atif, Olivier Cappé
Researchers introduce the Generalized Leverage Score (GLS) as an efficient, influence-based method for evaluating the privacy vulnerability of individual data points in deep neural networks. This metric exhibits a Spearman correlation of 0.50-0.51 with LiRA scores and effectively identifies data points that are more susceptible to membership inference attacks, offering a scalable alternative to computationally intensive shadow model approaches.
Key insights
- In linear models, the leverage score quantifies the inherent susceptibility of a data point to membership inference, directly impacting the optimal MIA's error trade-off.
- The Generalized Leverage Score (GLS) serves as a computationally tractable proxy for individual privacy risk in deep learning, circumventing the need for expensive model retraining.
- Data points with high GLS values often correspond to geometric outliers in the model's representation space, exhibiting atypical features that contribute to increased vulnerability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.